From efe6a83e3046a94cda38c8be5a7801eadc21d78d Mon Sep 17 00:00:00 2001 From: Salvatore Mesoraca Date: Wed, 28 Aug 2024 10:23:02 +0200 Subject: [PATCH 01/32] ggml : fix cont with transposed tensors when one dimension is 1 (ggml/934) * ggml_cont: fix issue with transposed tensors when one dimension is 1 when using multiple threads, it is not enough to check for the tensors to be contiguous for ggml_compute_forward_dup_same_cont to work correctly. The tensors strides also need to match. Signed-off-by: Salvatore Mesoraca * Add ggml_cont tests Signed-off-by: Salvatore Mesoraca * Remove dead code it isn't possible to reach this code because all these functions are invoked by ggml_compute_forward_dup if and only if src0->type != dst->type Signed-off-by: Salvatore Mesoraca * Make ggml_compute_forward_dup_same_cont work with contiguous tensors Co-authored-by: Georgi Gerganov Signed-off-by: Salvatore Mesoraca --------- Signed-off-by: Salvatore Mesoraca Co-authored-by: Georgi Gerganov --- ggml/src/ggml.c | 26 +++++--------------------- tests/test-backend-ops.cpp | 9 +++++++++ 2 files changed, 14 insertions(+), 21 deletions(-) diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index c25f9caa3..3504e9ad3 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -8322,8 +8322,7 @@ static void ggml_compute_forward_dup_same_cont( GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); GGML_ASSERT(src0->type == dst->type); - const size_t nb00 = src0->nb[0]; - const size_t nb0 = dst->nb[0]; + const size_t nb0 = ggml_type_size(src0->type); const int ith = params->ith; // thread index const int nth = params->nth; // number of threads @@ -8337,8 +8336,8 @@ static void ggml_compute_forward_dup_same_cont( if (ie0 < ie1) { memcpy( ((char *) dst->data + ie0*nb0), - ((char *) src0->data + ie0*nb00), - (ie1 - ie0) * ggml_type_size(src0->type)); + ((char *) src0->data + ie0*nb0), + (ie1 - ie0) * nb0); } } @@ -8355,11 +8354,6 @@ static void ggml_compute_forward_dup_f16( const int ith = params->ith; // thread index const int nth = params->nth; // number of threads - if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { - ggml_compute_forward_dup_same_cont(params, dst); - return; - } - // parallelize by rows const int nr = ne01; // number of rows per thread @@ -8624,11 +8618,6 @@ static void ggml_compute_forward_dup_bf16( const int ith = params->ith; // thread index const int nth = params->nth; // number of threads - if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { - ggml_compute_forward_dup_same_cont(params, dst); - return; - } - // parallelize by rows const int nr = ne01; // number of rows per thread @@ -8980,11 +8969,6 @@ static void ggml_compute_forward_dup_f32( const int ith = params->ith; // thread index const int nth = params->nth; // number of threads - if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { - ggml_compute_forward_dup_same_cont(params, dst); - return; - } - // parallelize by rows const int nr = ne01; // number of rows per thread @@ -9294,13 +9278,13 @@ static void ggml_compute_forward_dup_bytes( GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); GGML_ASSERT(src0->type == dst->type); + GGML_TENSOR_UNARY_OP_LOCALS; + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) { ggml_compute_forward_dup_same_cont(params, dst); return; } - GGML_TENSOR_UNARY_OP_LOCALS; - const size_t type_size = ggml_type_size(src0->type); const int ith = params->ith; // thread index const int nth = params->nth; // number of threads diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index bd65e8cb3..c3536054a 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -2322,6 +2322,15 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op } test_cases.emplace_back(new test_cont()); + test_cases.emplace_back(new test_cont(GGML_TYPE_F32, {2, 1, 1 ,1})); + test_cases.emplace_back(new test_cont(GGML_TYPE_F32, {2, 1, 3 ,5})); + test_cases.emplace_back(new test_cont(GGML_TYPE_F32, {2, 3, 5 ,7})); + test_cases.emplace_back(new test_cont(GGML_TYPE_F16, {2, 1, 1 ,1})); + test_cases.emplace_back(new test_cont(GGML_TYPE_F16, {2, 1, 3 ,5})); + test_cases.emplace_back(new test_cont(GGML_TYPE_F16, {2, 3, 5 ,7})); + test_cases.emplace_back(new test_cont(GGML_TYPE_BF16, {2, 1, 1 ,1})); + test_cases.emplace_back(new test_cont(GGML_TYPE_BF16, {2, 1, 3 ,5})); + test_cases.emplace_back(new test_cont(GGML_TYPE_BF16, {2, 3, 5 ,7})); auto add_test_bin_bcast = [&](ggml_type type, std::array ne, std::array nr) { for (auto op : {ggml_add, ggml_mul, ggml_div}) { From 51d964a4efdb0e8d43f5b85a775951185f6b641e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 28 Aug 2024 17:08:03 +0300 Subject: [PATCH 02/32] cuda : mark BF16 CONT as unsupported --- ggml/src/ggml-cuda.cu | 2 ++ 1 file changed, 2 insertions(+) diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index d33988d02..af1bd0518 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -2887,7 +2887,9 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_SIN: case GGML_OP_COS: case GGML_OP_CLAMP: + return true; case GGML_OP_CONT: + return op->src[0]->type != GGML_TYPE_BF16; case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: return true; From d2d3200b385970cb2a4658fd9342a3b1212bdb46 Mon Sep 17 00:00:00 2001 From: Mengqing Cao Date: Fri, 9 Aug 2024 20:21:56 +0800 Subject: [PATCH 03/32] cann : add Ascend NPU support (whisper/2336) * enable Ascend NPU in src/whisper.cpp * sync test-backend-ops with llama.cpp --- ggml/src/ggml-cann/Doxyfile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-cann/Doxyfile b/ggml/src/ggml-cann/Doxyfile index 2b009e8f9..5b7595b53 100644 --- a/ggml/src/ggml-cann/Doxyfile +++ b/ggml/src/ggml-cann/Doxyfile @@ -32,7 +32,7 @@ DOXYFILE_ENCODING = UTF-8 # title of most generated pages and in a few other places. # The default value is: My Project. -PROJECT_NAME = "llama.cpp" +PROJECT_NAME = "whisper.cpp" # The PROJECT_NUMBER tag can be used to enter a project or revision number. This # could be handy for archiving the generated documentation or if some version @@ -44,7 +44,7 @@ PROJECT_NUMBER = # for a project that appears at the top of each page and should give viewer a # quick idea about the purpose of the project. Keep the description short. -PROJECT_BRIEF = "llama inference engine" +PROJECT_BRIEF = "Port of OpenAI's Whisper model in C/C++" # With the PROJECT_LOGO tag one can specify a logo or an icon that is included # in the documentation. The maximum height of the logo should not exceed 55 From ba1cf846edeb94fc567a9ab0f7ef705282e6d7d3 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 28 Aug 2024 18:45:01 +0300 Subject: [PATCH 04/32] cann : fix doxy (ggml/0) --- ggml/src/ggml-cann/Doxyfile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-cann/Doxyfile b/ggml/src/ggml-cann/Doxyfile index 5b7595b53..3290a4859 100644 --- a/ggml/src/ggml-cann/Doxyfile +++ b/ggml/src/ggml-cann/Doxyfile @@ -32,7 +32,7 @@ DOXYFILE_ENCODING = UTF-8 # title of most generated pages and in a few other places. # The default value is: My Project. -PROJECT_NAME = "whisper.cpp" +PROJECT_NAME = "ggml" # The PROJECT_NUMBER tag can be used to enter a project or revision number. This # could be handy for archiving the generated documentation or if some version @@ -44,7 +44,7 @@ PROJECT_NUMBER = # for a project that appears at the top of each page and should give viewer a # quick idea about the purpose of the project. Keep the description short. -PROJECT_BRIEF = "Port of OpenAI's Whisper model in C/C++" +PROJECT_BRIEF = "Tensor library for machine learning" # With the PROJECT_LOGO tag one can specify a logo or an icon that is included # in the documentation. The maximum height of the logo should not exceed 55 From dbbebcab339c7d63d3806e8f32574bb9aad9a694 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Sat, 31 Aug 2024 14:35:42 +0200 Subject: [PATCH 05/32] ggml: fix ggml_graph_cpy undefined behavior (ggml/943) --- ggml/include/ggml.h | 4 ++-- ggml/src/ggml.c | 3 ++- 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 09c72b095..6354fcf51 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -681,8 +681,8 @@ extern "C" { struct ggml_hash_set { size_t size; - ggml_bitset_t * used; - struct ggml_tensor ** keys; + ggml_bitset_t * used; // whether or not the keys are in use i.e. set + struct ggml_tensor ** keys; // actual tensors in the set, keys[i] is only defined if ggml_bitset_get(used, i) }; // computation graph diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 3504e9ad3..5f106d52a 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -19130,7 +19130,8 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) { } for (size_t i = 0; i < src->visited_hash_set.size; ++i) { - if (src->visited_hash_set.keys[i]) { + // copy all hashset keys (tensors) that are in use + if (ggml_bitset_get(src->visited_hash_set.used, i)) { ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]); } } From 202084d31d4247764fc6d6d40d2e2bda0c89a73a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Tue, 3 Sep 2024 17:21:46 +0200 Subject: [PATCH 06/32] tests: add gradient tests for all backends (ggml/932) * tests: add gradient checking to test-backend-ops * remove old comment * reorder includes * adjust SIN/COS parameters * add documentation, use supports_op if possible --- ggml/include/ggml.h | 12 +- ggml/src/ggml-backend.c | 4 + ggml/src/ggml-cuda.cu | 12 + ggml/src/ggml-cuda/cross-entropy-loss.cu | 4 +- ggml/src/ggml-cuda/sum.cu | 41 + ggml/src/ggml-cuda/sum.cuh | 5 + ggml/src/ggml-cuda/unary.cu | 29 + ggml/src/ggml-cuda/unary.cuh | 3 + ggml/src/ggml.c | 32 +- tests/test-backend-ops.cpp | 1030 ++++++++++++++++++++-- 10 files changed, 1080 insertions(+), 92 deletions(-) create mode 100644 ggml/src/ggml-cuda/sum.cu create mode 100644 ggml/src/ggml-cuda/sum.cuh diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 6354fcf51..536018b66 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -1272,7 +1272,7 @@ extern "C" { size_t nb1, size_t nb2, size_t nb3, - size_t offset); + size_t offset); // in bytes // b -> view(a,offset,nb1,nb2,3), return view(a) GGML_API struct ggml_tensor * ggml_set_inplace( @@ -1282,19 +1282,19 @@ extern "C" { size_t nb1, size_t nb2, size_t nb3, - size_t offset); + size_t offset); // in bytes GGML_API struct ggml_tensor * ggml_set_1d( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - size_t offset); + size_t offset); // in bytes GGML_API struct ggml_tensor * ggml_set_1d_inplace( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - size_t offset); + size_t offset); // in bytes // b -> view(a,offset,nb1,nb2,3), return modified a GGML_API struct ggml_tensor * ggml_set_2d( @@ -1302,7 +1302,7 @@ extern "C" { struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, - size_t offset); + size_t offset); // in bytes // b -> view(a,offset,nb1,nb2,3), return view(a) GGML_API struct ggml_tensor * ggml_set_2d_inplace( @@ -1310,7 +1310,7 @@ extern "C" { struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, - size_t offset); + size_t offset); // in bytes // a -> b, return view(b) GGML_API struct ggml_tensor * ggml_cpy( diff --git a/ggml/src/ggml-backend.c b/ggml/src/ggml-backend.c index 96b93d2d8..b5d9301a7 100644 --- a/ggml/src/ggml-backend.c +++ b/ggml/src/ggml-backend.c @@ -827,6 +827,10 @@ GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float case GGML_OP_MUL_MAT: return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type; + case GGML_OP_ROPE_BACK: + return op->src[2] == NULL && (op->op_params[2] & 4) == 0; + case GGML_OP_IM2COL_BACK: + return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32; default: return true; } diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index af1bd0518..982316f56 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -27,6 +27,7 @@ #include "ggml-cuda/rope.cuh" #include "ggml-cuda/scale.cuh" #include "ggml-cuda/softmax.cuh" +#include "ggml-cuda/sum.cuh" #include "ggml-cuda/sumrows.cuh" #include "ggml-cuda/tsembd.cuh" #include "ggml-cuda/unary.cuh" @@ -2180,6 +2181,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg ggml_cuda_dup(ctx, dst); break; case GGML_OP_ADD: + case GGML_OP_ADD1: // TODO: more efficient implementation ggml_cuda_op_add(ctx, dst); break; case GGML_OP_SUB: @@ -2196,6 +2198,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg break; case GGML_OP_UNARY: switch (ggml_get_unary_op(dst)) { + case GGML_UNARY_OP_NEG: + ggml_cuda_op_neg(ctx, dst); + break; case GGML_UNARY_OP_GELU: ggml_cuda_op_gelu(ctx, dst); break; @@ -2304,6 +2309,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_POOL_2D: ggml_cuda_op_pool2d(ctx, dst); break; + case GGML_OP_SUM: + ggml_cuda_op_sum(ctx, dst); + break; case GGML_OP_SUM_ROWS: ggml_cuda_op_sum_rows(ctx, dst); break; @@ -2748,6 +2756,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons switch (op->op) { case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { + case GGML_UNARY_OP_NEG: case GGML_UNARY_OP_GELU: case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_RELU: @@ -2877,6 +2886,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_TRANSPOSE: case GGML_OP_NORM: case GGML_OP_ADD: + case GGML_OP_ADD1: case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: @@ -2896,7 +2906,9 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_ROPE: return ggml_is_contiguous(op->src[0]); case GGML_OP_IM2COL: + return op->src[0]->type == GGML_TYPE_F16; case GGML_OP_POOL_2D: + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: case GGML_OP_ARGSORT: case GGML_OP_ACC: diff --git a/ggml/src/ggml-cuda/cross-entropy-loss.cu b/ggml/src/ggml-cuda/cross-entropy-loss.cu index a14043e70..5575a90f6 100644 --- a/ggml/src/ggml-cuda/cross-entropy-loss.cu +++ b/ggml/src/ggml-cuda/cross-entropy-loss.cu @@ -1,6 +1,6 @@ #include "common.cuh" #include "cross-entropy-loss.cuh" -#include "sumrows.cuh" +#include "sum.cuh" #include #include @@ -102,5 +102,5 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * cross_entropy_loss_f32<<>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows); // Combine results from individual blocks: - sum_rows_f32_cuda(dst_tmp.ptr, dst_d, blocks_num.x, 1, stream); + sum_f32_cuda(pool, dst_tmp.ptr, dst_d, blocks_num.x, stream); } diff --git a/ggml/src/ggml-cuda/sum.cu b/ggml/src/ggml-cuda/sum.cu new file mode 100644 index 000000000..0d5e953ee --- /dev/null +++ b/ggml/src/ggml-cuda/sum.cu @@ -0,0 +1,41 @@ +#include "sumrows.cuh" +#include "sum.cuh" + +#include + +#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) +#include +using namespace cub; +#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) + +void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) { +#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) + size_t tmp_size = 0; + DeviceReduce::Sum(nullptr, tmp_size, x, dst, ne, stream); + ggml_cuda_pool_alloc tmp_alloc(pool, tmp_size); + DeviceReduce::Sum(tmp_alloc.ptr, tmp_size, x, dst, ne, stream); +#else + // Use (inefficient) sum_rows implementation as a fallback. + // For AMD there is rocPRIM which could be used as a drop-in replacement via hipcub but this would require C++11 -> C++14. + sum_rows_f32_cuda(x, dst, ne, 1, stream); + GGML_UNUSED(pool); +#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) +} + +void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + GGML_ASSERT(ggml_is_contiguous(src0)); + + const float * src0_d = (const float *) src0->data; + float * dst_d = (float *) dst->data; + + const int64_t ne = ggml_nelements(src0); + + ggml_cuda_pool & pool = ctx.pool(); + cudaStream_t stream = ctx.stream(); + + sum_f32_cuda(pool, src0_d, dst_d, ne, stream); +} diff --git a/ggml/src/ggml-cuda/sum.cuh b/ggml/src/ggml-cuda/sum.cuh new file mode 100644 index 000000000..8cadc3736 --- /dev/null +++ b/ggml/src/ggml-cuda/sum.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream); + +void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/unary.cu b/ggml/src/ggml-cuda/unary.cu index 89abfc21d..8ac669f94 100644 --- a/ggml/src/ggml-cuda/unary.cu +++ b/ggml/src/ggml-cuda/unary.cu @@ -1,5 +1,15 @@ #include "unary.cuh" +static __global__ void neg_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + + dst[i] = -x[i]; +} + static __global__ void gelu_f32(const float * x, float * dst, const int k) { const float GELU_COEF_A = 0.044715f; const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; @@ -119,6 +129,11 @@ static __global__ void cos_f32(const float * x, float * dst, const int k) { dst[i] = cosf(x[i]); } +static void neg_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE; + neg_f32<<>>(x, dst, k); +} + static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE; gelu_f32<<>>(x, dst, k); @@ -184,6 +199,20 @@ static void cos_f32_cuda(const float * x, float * dst, const int k, cudaStream_t cos_f32<<>>(x, dst, k); } +void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(ggml_is_contiguous(src0)); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + neg_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const float * src0_d = (const float *)src0->data; diff --git a/ggml/src/ggml-cuda/unary.cuh b/ggml/src/ggml-cuda/unary.cuh index c610e996a..ed2ffc461 100644 --- a/ggml/src/ggml-cuda/unary.cuh +++ b/ggml/src/ggml-cuda/unary.cuh @@ -1,5 +1,6 @@ #include "common.cuh" +#define CUDA_NEG_BLOCK_SIZE 256 #define CUDA_GELU_BLOCK_SIZE 256 #define CUDA_SILU_BLOCK_SIZE 256 #define CUDA_TANH_BLOCK_SIZE 256 @@ -12,6 +13,8 @@ #define CUDA_SIN_BLOCK_SIZE 256 #define CUDA_COS_BLOCK_SIZE 256 +void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 5f106d52a..28ee46e04 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -5267,6 +5267,7 @@ struct ggml_tensor * ggml_concat( bool is_node = false; if (a->grad || b->grad) { + GGML_ABORT("fatal error"); // TODO: implement is_node = true; } @@ -5388,6 +5389,7 @@ struct ggml_tensor * ggml_leaky_relu( bool is_node = false; if (!inplace && (a->grad)) { + GGML_ABORT("fatal error"); // TODO: not implemented is_node = true; } @@ -5826,6 +5828,7 @@ static struct ggml_tensor * ggml_set_impl( // make a view of the destination struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + GGML_ASSERT(offset < (size_t)(1 << 30)); int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 }; ggml_set_op_params(result, params, sizeof(params)); @@ -6783,14 +6786,12 @@ struct ggml_tensor * ggml_rope_back( GGML_ASSERT(ggml_is_vector(b)); GGML_ASSERT(b->type == GGML_TYPE_I32); GGML_ASSERT(a->ne[2] == b->ne[0]); - GGML_ASSERT(c == NULL && "freq factors not implemented yet"); - - GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet"); bool is_node = false; if (a->grad) { - is_node = false; // TODO: implement backward + GGML_ASSERT(false && "backwards pass not implemented"); + is_node = false; } struct ggml_tensor * result = ggml_dup_tensor(ctx, a); @@ -6808,6 +6809,7 @@ struct ggml_tensor * ggml_rope_back( result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; + result->src[2] = c; return result; } @@ -7361,6 +7363,11 @@ struct ggml_tensor * ggml_argsort( enum ggml_sort_order order) { bool is_node = false; + if (a->grad) { + GGML_ABORT("fatal error"); // TODO: not implemented + is_node = true; + } + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne); ggml_set_op_params_i32(result, 0, (int32_t) order); @@ -10953,9 +10960,6 @@ static void ggml_compute_forward_sum_f32( return; } - assert(ggml_is_scalar(dst)); - - assert(ggml_is_scalar(dst)); assert(src0->nb[0] == sizeof(float)); @@ -18356,14 +18360,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor if (src0->grad || src1->grad) { GGML_ASSERT(src0->type == tensor->type); GGML_ASSERT(tensor->grad->type == tensor->type); - GGML_ASSERT(tensor->grad->type == src1->grad->type); + GGML_ASSERT(!src1->grad || src1->grad->type == tensor->grad->type); tensor_grad_view = ggml_view_4d(ctx, - tensor->grad, - src1->grad->ne[0], - src1->grad->ne[1], - src1->grad->ne[2], - src1->grad->ne[3], + tensor->grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], nb1, nb2, nb3, offset); } @@ -18432,9 +18432,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor memcpy(&offset, tensor->op_params, sizeof(offset)); - size_t nb1 = tensor->nb[1]; - size_t nb2 = tensor->nb[2]; - size_t nb3 = tensor->nb[3]; + size_t nb1 = tensor->nb[1]; + size_t nb2 = tensor->nb[2]; + size_t nb3 = tensor->nb[3]; if (src0->type != src0->grad->type) { // gradient is typically F32, but src0 could be other type diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index c3536054a..8f659c148 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -1,3 +1,20 @@ +// This file defines tests for various GGML ops and backends. +// For the forward pass it asserts that the results of multiple backends computing the same GGML ops are consistent. +// For the backwards pass it asserts that the gradients from backpropagation are consistent +// with the gradients obtained via the method of finite differences ("grad" mode, this is optional). +// It is also possible to check the performance ("perf" mode). +// +// this file has three sections: Section 1 does general setup, section 2 defines the GGML ops to be tested, +// and section 3 defines which tests to run. +// Quick start for adding a new GGML op: Go to section 2 and create a struct that inherits from test_case, +// then go to section 3 and add an instantiation of your struct. + + +// ############################## +// ## Section 1: General Setup ## +// ############################## + + #include #include #include @@ -5,6 +22,7 @@ #include #include #include +#include #include #include #include @@ -212,6 +230,39 @@ static double nmse(const float * a, const float * b, size_t n) { return mse_a_b / mse_a_0; } +// maximum absolute asymmetry between a and b +// asymmetry: (a - b) / (a + b) +// This is more stable than relative error if one of the values fluctuates towards zero. +// n: number of values to compare. +// expected_vals: optional vector of expected values for a. If expected_vals is not empty, filter out all comparisons where +// a does not match any of the expected values. Needed for noncontinuous gradients where the numerical calculation can fail. +static double mean_abs_asymm(const float * a, const float * b, const size_t n, const std::vector & expected_vals) { + double sum = 0.0f; + + size_t nvalid = 0; + for (size_t i = 0; i < n; i++) { + if (!expected_vals.empty()) { + bool matches_any = false; + for (const float & ev : expected_vals) { + if (fabsf(a[i] - ev) < 1e-3f) { + matches_any = true; + break; + } + } + if (!matches_any) { + continue; + } + } + + const float asymm = (a[i] - b[i]) / (a[i] + b[i]); + + sum += fabsf(asymm); + nvalid++; + } + + return sum/nvalid; +} + // utils for printing the variables of the test cases #define VAR_TO_STR(x) (#x "=" + var_to_str(x)) @@ -295,6 +346,7 @@ static bool ggml_is_view_op(enum ggml_op op) { enum test_mode { MODE_TEST, MODE_PERF, + MODE_GRAD, }; struct test_case { @@ -314,6 +366,32 @@ struct test_case { return 1e-7; } + virtual double max_maa_err() { + return 1e-4; + } + + virtual float grad_eps(){ + return 1e-1f; + } + + // If false, estimate gradient with 2 points, neglects 3rd order derivative and higher. + // If true, estimate gradient with 4 points, neglects 5th order derivative and higher. + virtual bool grad_precise(){ + return false; + } + + // Skip gradient checks if total number of gradients to be checked is larger than this (to speed up the tests). + virtual int64_t grad_nmax() { + return 10000; + } + + // No effect if empty. + // If not empty, skip all gradient checks where the numerical result does not match any of the values. + // Needed for dealing with noncontinuous gradients (e.g. ReLU) where estimation using finite differences is unreliable. + virtual std::vector grad_expect() { + return {}; + } + virtual void initialize_tensors(ggml_context * ctx) { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) { init_tensor_uniform(t); @@ -332,6 +410,7 @@ struct test_case { } ggml_cgraph * gf = nullptr; + ggml_cgraph * gb = nullptr; static const int sentinel_size = 1024; @@ -340,7 +419,7 @@ struct test_case { std::vector sentinels; void add_sentinel(ggml_context * ctx) { - if (mode == MODE_PERF) { + if (mode == MODE_PERF || mode == MODE_GRAD) { return; } ggml_tensor * sentinel = ::ggml_new_tensor_1d(ctx, GGML_TYPE_F32, sentinel_size); @@ -389,6 +468,7 @@ struct test_case { /* .no_alloc = */ true, }; ggml_context * ctx = ggml_init(params); + GGML_ASSERT(ctx); gf = ggml_new_graph(ctx); @@ -550,6 +630,7 @@ struct test_case { /* .no_alloc = */ true, }; ggml_context * ctx = ggml_init(params); + GGML_ASSERT(ctx); ggml_tensor * out = build_graph(ctx); @@ -643,8 +724,282 @@ struct test_case { return true; } + + bool eval_grad(ggml_backend_t backend, const char * op_name) { + mode = MODE_GRAD; + const std::vector expect = grad_expect(); + + ggml_init_params params = { + /* .mem_size = */ ggml_tensor_overhead()*128 + 2*ggml_graph_overhead_custom(GGML_DEFAULT_GRAPH_SIZE, true), + /* .mem_base = */ NULL, + /* .no_alloc = */ true, + }; + ggml_context * ctx = ggml_init(params); + GGML_ASSERT(ctx); + + gf = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, true); + gb = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, true); + + ggml_tensor * out = build_graph(ctx); + + if (op_name != nullptr && op_desc(out) != op_name) { + //printf(" %s: skipping\n", op_desc(out).c_str()); + ggml_free(ctx); + return true; + } + + printf(" %s(%s): ", op_desc(out).c_str(), vars().c_str()); + fflush(stdout); + + if (out->grad == nullptr) { + printf("backwards pass not supported \n"); + ggml_free(ctx); + return true; + } + if (out->type != GGML_TYPE_F32) { + ggml_free(ctx); + printf("not supported [%s->type != FP32]\n", out->name); + return true; + } + + // check if the backend supports the ops + bool supported = true; + for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { + if (!ggml_backend_supports_op(backend, t)) { + printf("not supported [%s] ", ggml_backend_name(backend)); + supported = false; + break; + } + if ((t->flags & GGML_TENSOR_FLAG_PARAM) && t->type != GGML_TYPE_F32) { + printf("not supported [%s->type != FP32] ", t->name); + supported = false; + break; + } + } + if (!supported) { + printf("\n"); + ggml_free(ctx); + return true; + } + + int64_t ngrads = 0; + for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { + if (t->flags & GGML_TENSOR_FLAG_PARAM) { + ngrads += ggml_nelements(t); + } + } + if (ngrads > grad_nmax()) { + printf("skipping large tensors for speed \n"); + ggml_free(ctx); + return true; + } + + + if (!ggml_is_scalar(out)) { + out = ggml_sum(ctx, out); + ggml_set_name(out, "sum_of_out"); + } + + ggml_build_forward_expand(gf, out); + ggml_graph_cpy(gf, gb); + ggml_build_backward_expand(ctx, gf, gb, false); + if (expect.size() != 1 || expect[0] != 0.0f) { + GGML_ASSERT(gb->n_nodes > gf->n_nodes); + for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { + GGML_ASSERT(!(t->flags & GGML_TENSOR_FLAG_PARAM) || t->grad->op != GGML_OP_NONE); + } + } + + // TODO: refactor so that this check is only needed once + for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { + if (!ggml_backend_supports_op(backend, t)) { + printf("not supported [%s] ", ggml_backend_name(backend)); + supported = false; + break; + } + if ((t->flags & GGML_TENSOR_FLAG_PARAM) && t->type != GGML_TYPE_F32) { + printf("not supported [%s->type != FP32] ", t->name); + supported = false; + break; + } + } + if (!supported) { + printf("\n"); + ggml_free(ctx); + return true; + } + + // allocate + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend); + if (buf == NULL) { + printf("failed to allocate tensors [%s] ", ggml_backend_name(backend)); + ggml_free(ctx); + return false; + } + + // randomize tensors + initialize_tensors(ctx); + + for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) { + if (!t->grad) { + continue; + } + + std::vector tmp(ggml_nelements(t->grad)); + ggml_backend_tensor_set(t->grad, tmp.data(), 0, ggml_nbytes(t->grad)); + } + + // build graphs + const float onef = 1.0f; + ggml_backend_graph_compute(backend, gf); + ggml_backend_tensor_set(out->grad, &onef, 0, ggml_nbytes(out->grad)); + ggml_backend_graph_compute(backend, gb); + + bool ok = true; + for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) { + if (!(t->flags & GGML_TENSOR_FLAG_PARAM)) { + continue; + } + + const char * bn = ggml_backend_name(backend); + const int64_t ne = ggml_nelements(t); + + std::vector ga = tensor_to_float(t->grad); + + for (int64_t i = 0; i < ne; ++i) { // gradient algebraic + // check for nans + if (!std::isfinite(ga[i])) { + printf("[%s] nonfinite gradient at index %zu (%s=%f) ", ggml_op_desc(t), i, bn, ga[i]); + ok = false; + break; + } + } + if (!ok) { + break; + } + + std::vector gn(ne); // gradient numeric + GGML_ASSERT(ga.size() == gn.size()); + + std::vector x0 = tensor_to_float(t); // original t data + GGML_ASSERT(ggml_is_scalar(out)); + GGML_ASSERT(out->type == GGML_TYPE_F32); + + const float eps = grad_eps(); + for (int64_t i = 0; i < ne; ++i) { + const float xiu = x0[i] + 1.0f*eps; // x, index i, up + const float xiuh = x0[i] + 0.5f*eps; // x, index i, up half + const float xidh = x0[i] - 0.5f*eps; // x, index i, down half + const float xid = x0[i] - 1.0f*eps; // x, index i, down + + float fu, fuh, fdh, fd; // output values for xiu, xiuh, xid, xidh + + ggml_backend_tensor_set(t, &xiu, i*sizeof(float), sizeof(float)); + ggml_backend_graph_compute(backend, gf); + ggml_backend_tensor_get(out, &fu, 0, ggml_nbytes(out)); + + ggml_backend_tensor_set(t, &xid, i*sizeof(float), sizeof(float)); + ggml_backend_graph_compute(backend, gf); + ggml_backend_tensor_get(out, &fd, 0, ggml_nbytes(out)); + + if (grad_precise()) { + ggml_backend_tensor_set(t, &xiuh, i*sizeof(float), sizeof(float)); + ggml_backend_graph_compute(backend, gf); + ggml_backend_tensor_get(out, &fuh, 0, ggml_nbytes(out)); + + ggml_backend_tensor_set(t, &xidh, i*sizeof(float), sizeof(float)); + ggml_backend_graph_compute(backend, gf); + ggml_backend_tensor_get(out, &fdh, 0, ggml_nbytes(out)); + + gn[i] = (8.0*(double)fuh + (double)fd - (8.0*(double)fdh + (double)fu)) / (6.0*(double)eps); + } else { + gn[i] = (fu - fd) / (2.0f*eps); + } + + ggml_backend_tensor_set(t, x0.data(), 0, ggml_nbytes(t)); + } + + const double err = mean_abs_asymm(gn.data(), ga.data(), gn.size(), expect); + if (err > max_maa_err()) { + printf("[%s] MAA = %.9f > %.9f ", ggml_op_desc(t), err, max_maa_err()); + ok = false; + break; + } + if (!ok) { + break; + } + } + + if (!ok) { + printf("compare failed "); + } + + ggml_backend_buffer_free(buf); + + ggml_free(ctx); + + if (ok) { + printf("\033[1;32mOK\033[0m\n"); + return true; + } + + printf("\033[1;31mFAIL\033[0m\n"); + return false; + } }; + +// ################################### +// ## Section 2: GGML Op Defintions ## +// ################################### + + +// The following is an example showing the bare minimum for creating a test for a GGML op. + +// GGML_OP_EXAMPLE +struct test_example : public test_case { + // Always define these 2 or variants thereof: + const ggml_type type; // The type of the input tensors. + const std::array ne; // The shape of the input tensors. + // For some ops it's necessary to define multiple types or shapes for the inputs. + // Or they may need additional parameters. + + // Put all parameters needed to fully define the test into one of the VARS_TO_STR macros. + // In most cases these are just the properties of the struct that you defined above. + // This is needed for info prints. + std::string vars() override { + return VARS_TO_STR2(type, ne); + } + + // Define a constructor for the struct. + // In most cases it will be sufficient to have the same arguments as the struct has properties + // and just use initializer lists. + test_example(ggml_type type = GGML_TYPE_F32, + std::array ne = {10, 5, 4, 3}) + : type(type), ne(ne) {} + + // Define how a simple GGML compute graph can be constructed for the new GGML op. + ggml_tensor * build_graph(ggml_context * ctx) override { + // Step 1: create input tensors that don't depend on any other tensors: + ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); // Setting names is optional but it's useful for debugging. + + ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(b, "b"); + + // Step 2: use the op that you want to test in the GGML compute graph. + ggml_tensor * out = ggml_add(ctx, a, b); // For this example we're just doing a simple addition. + ggml_set_name(out, "out"); + + // Step 3: return the output tensor. + return out; + } + // In order to also check the gradients for your op, add calls like ggml_set_param(ctx, a) + // immediately after you create the tensors. + // This is optional and only makes sense if a backwards pass has actually been implemented for the new op. +}; + + // GGML_OP_UNARY struct test_unary : public test_case { const ggml_unary_op op; @@ -658,20 +1013,36 @@ struct test_unary : public test_case { test_unary(ggml_unary_op op, ggml_type type = GGML_TYPE_F32, - std::array ne_a = {128, 10, 10, 10}, + std::array ne_a = {128, 2, 2, 2}, int v = 0) : op(op), type(type), ne_a(ne_a), v(v) {} ggml_tensor * build_graph(ggml_context * ctx) override { + const bool grad_supported = op == GGML_UNARY_OP_ABS || op == GGML_UNARY_OP_SGN || op == GGML_UNARY_OP_NEG || + op == GGML_UNARY_OP_STEP || op == GGML_UNARY_OP_RELU || op == GGML_UNARY_OP_SILU; + ggml_tensor * a; if (v & 1) { auto ne = ne_a; ne[0] *= 3; a = ggml_new_tensor(ctx, type, 4, ne.data()); + if (grad_supported) { + ggml_set_param(ctx, a); + } + ggml_set_name(a, "a"); + a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0); + ggml_set_name(a, "view_of_a"); } else { a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + if (grad_supported) { + ggml_set_param(ctx, a); + } + ggml_set_name(a, "a"); } + ggml_tensor * out = ggml_unary(ctx, a, op); + ggml_set_name(out, "out"); + return out; } @@ -681,6 +1052,24 @@ struct test_unary : public test_case { init_tensor_uniform(t, -150.f, 150.f); } } + + float grad_eps() override { + return 15.0f; + } + + std::vector grad_expect() override { + if (op == GGML_UNARY_OP_ABS) { + return {-1.0f, 1.0f}; + } + if (op == GGML_UNARY_OP_SGN || op == GGML_UNARY_OP_STEP) { + return {0.0f}; + } + if (op == GGML_UNARY_OP_RELU) { + return {0.0f, 1.0f}; + } + return {}; + } + }; // GGML_OP_GET_ROWS @@ -701,11 +1090,24 @@ struct test_get_rows : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * in = ggml_new_tensor_3d(ctx, type, n, m, b); + ggml_set_name(in, "in"); + ggml_tensor * rows = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, r, b); + ggml_set_name(rows, "rows"); if (v) { rows = ggml_view_2d(ctx, rows, r/2, b, rows->nb[1], 0); + ggml_set_name(rows, "view_of_rows"); } + + const bool grad_supported = ggml_is_matrix(in) && ggml_is_vector(rows); + if (grad_supported) { + ggml_set_param(ctx, in); + // rows is a constant input -> no gradients + } + ggml_tensor * out = ggml_get_rows(ctx, in, rows); + ggml_set_name(out, "out"); + return out; } @@ -741,14 +1143,21 @@ struct test_repeat : public test_case { } test_repeat(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}, + std::array ne = {10, 5, 4, 3}, std::array nr = {2, 2, 2, 2}) : type(type), ne(ne), nr(nr) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * target = ggml_new_tensor_4d(ctx, type, ne[0]*nr[0], ne[1]*nr[1], ne[2]*nr[2], ne[3]*nr[3]); + ggml_set_name(target, "target"); + ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, src); + ggml_set_name(src, "src"); + ggml_tensor * out = ggml_repeat(ctx, src, target); + ggml_set_name(out, "out"); + return out; } }; @@ -774,10 +1183,62 @@ struct test_dup : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, src); + ggml_set_name(src, "src"); + if (_use_permute) { src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]); + ggml_set_name(src, "src_permuted"); } + ggml_tensor * out = ggml_dup(ctx, src); + ggml_set_name(out, "out"); + + return out; + } +}; + +// GGML_OP_SET +struct test_set : public test_case { + const ggml_type type_src; + const ggml_type type_dst; + const std::array ne; + const int dim; + + std::string vars() override { + return VARS_TO_STR4(type_src, type_dst, ne, dim); + } + + size_t op_size(ggml_tensor * t) override { + return ggml_nbytes(t) + ggml_nbytes(t->src[0]); + } + + test_set(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32, + std::array ne = {6, 5, 4, 3}, int dim = 1) + : type_src(type_src), type_dst(type_dst), ne(ne), dim(dim) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data()); + ggml_set_param(ctx, src); + ggml_set_name(src, "src"); + + auto ne_dst = ne; + for (int i = 0; i < dim; ++i) { + ne_dst[i] *= 2; + } + ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, ne_dst.data()); + ggml_set_param(ctx, dst); + ggml_set_name(dst, "dst"); + + size_t offset = 0; + for (int i = 0; i < dim; ++i) { + offset += ((ne_dst[i] - ne[i])/2)*dst->nb[i]; + } + ggml_tensor * out = ggml_set(ctx, dst, src, + // The backwards pass requires setting a contiguous region: + src->nb[1], src->nb[2], src->nb[3], offset); + ggml_set_name(out, "out"); + return out; } }; @@ -810,11 +1271,20 @@ struct test_cpy : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data()); + ggml_set_param(ctx, src); + ggml_set_name(src, "src"); + if (_src_use_permute) { src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]); + ggml_set_name(src, "src_permuted"); } + ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, src->ne); + ggml_set_name(dst, "dst"); + ggml_tensor * out = ggml_cpy(ctx, src, dst); + ggml_set_name(out, "out"); + return out; } }; @@ -834,8 +1304,14 @@ struct test_cont : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, src); + ggml_set_name(src, "src"); + src = ggml_transpose(ctx, src); + ggml_set_name(src, "src_transposed"); + ggml_tensor * out = ggml_cont(ctx, src); + ggml_set_name(out, "out"); return out; } @@ -866,21 +1342,79 @@ struct test_bin_bcast : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0]*nr[0], ne[1]*nr[1], ne[2]*nr[2], ne[3]*nr[3]); + ggml_set_name(a, "a"); + ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(b, "b"); + + // The backwards pass supports broadcasting only for GGML_ADD: + const bool grad_supported = op == ggml_add || ggml_are_same_shape(a, b); + if (grad_supported) { + ggml_set_param(ctx, a); + ggml_set_param(ctx, b); + } + ggml_tensor * out = op(ctx, a, b); + ggml_set_name(out, "out"); + return out; } void initialize_tensors(ggml_context * ctx) override { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { - if (op == ggml_div) { - // avoid division by zero - init_tensor_uniform(t, 1.0f, 2.0f); + if (op == ggml_mul || op == ggml_div) { + // MUL and DIV have numerical issues around zero: + init_tensor_uniform(t, 0.9f, 1.1f); } else { init_tensor_uniform(t); } } } + + float grad_eps() override { + return 0.1f * (op == ggml_mul ? ne[0]*ne[1]*ne[2]*ne[3] : 1); + } + + bool grad_precise() override { + return op == ggml_div; + } + + double max_maa_err() override { + return op == ggml_add ? 1e-4 : 1e-3; + } +}; + +// GGML_OP_ADD1 +struct test_add1 : public test_case { + const ggml_type type; + const std::array ne; + + std::string vars() override { + return VARS_TO_STR2(type, ne); + } + + test_add1(ggml_type type = GGML_TYPE_F32, + std::array ne = {10, 5, 4, 3}) + : type(type), ne(ne) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + + ggml_tensor * b = ggml_new_tensor_1d(ctx, type, 1); + // ggml_set_param(ctx, b); // TODO: implement + ggml_set_name(b, "b"); + + ggml_tensor * out = ggml_add1(ctx, a, b); + ggml_set_name(out, "out"); + + return out; + } + + float grad_eps() override { + return 0.1f * ne[0]*ne[1]*ne[2]*ne[3]; + } }; // GGML_OP_SCALE @@ -900,7 +1434,12 @@ struct test_scale : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_scale(ctx, a, scale); + ggml_set_name(out, "out"); + return out; } }; @@ -916,13 +1455,17 @@ struct test_norm : public test_case { } test_norm(ggml_type type = GGML_TYPE_F32, - std::array ne = {64, 10, 10, 10}, + std::array ne = {64, 5, 4, 3}, float eps = 1e-6f) : type(type), ne(ne), eps(eps) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_norm(ctx, a, eps); + ggml_set_name(out, "out"); + return out; } }; @@ -938,15 +1481,24 @@ struct test_rms_norm : public test_case { } test_rms_norm(ggml_type type = GGML_TYPE_F32, - std::array ne = {64, 10, 10, 10}, + std::array ne = {64, 5, 4, 3}, float eps = 1e-6f) : type(type), ne(ne), eps(eps) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_rms_norm(ctx, a, eps); + ggml_set_name(out, "out"); + return out; } + + bool grad_precise() override { + return true; + } }; // GGML_OP_SSM_CONV @@ -1038,7 +1590,14 @@ struct test_mul_mat : public test_case { // C^T = A * B^T: (k, m) * (k, n) => (m, n) ggml_tensor * a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0] , bs[1]); ggml_tensor * b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]); + ggml_set_param(ctx, a); + ggml_set_param(ctx, b); + ggml_set_name(a, "a"); + ggml_set_name(b, "b"); + ggml_tensor * out = ggml_mul_mat(ctx, a, b); + ggml_set_name(out, "out"); + return out; } }; @@ -1082,12 +1641,21 @@ struct test_mul_mat_id : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { // C^T = A * B^T: (k, m) * (k, n) => (m, n) ggml_tensor * as = ggml_new_tensor_3d(ctx, type_a, k, m, n_mats); + ggml_set_name(as, "as"); + ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_mats, n); + ggml_set_name(ids, "ids"); if (n_used != n_mats) { ids = ggml_view_2d(ctx, ids, n_used, n, ids->nb[1], 0); + ggml_set_name(ids, "view_of_ids"); } + ggml_tensor * b = ggml_new_tensor_3d(ctx, type_b, k, this->b ? 1 : n_used, n); + ggml_set_name(b, "b"); + ggml_tensor * out = ggml_mul_mat_id(ctx, as, b, ids); + ggml_set_name(out, "out"); + return out; } @@ -1123,14 +1691,23 @@ struct test_sqr : public test_case { } test_sqr(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}) + std::array ne = {10, 5, 4, 3}) : type(type), ne(ne) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_sqr(ctx, a); + ggml_set_name(out, "out"); + return out; } + + float grad_eps() override { + return 0.1f * 0.25f*ne[0]*ne[1]*ne[2]*ne[3]; // 10% of expected value of sum. + } }; // GGML_OP_SQRT @@ -1143,21 +1720,70 @@ struct test_sqrt : public test_case { } test_sqrt(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}) + std::array ne = {10, 3, 3, 2}) : type(type), ne(ne) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_sqrt(ctx, a); + ggml_set_name(out, "out"); + return out; } void initialize_tensors(ggml_context * ctx) override { // fill with positive values for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { - init_tensor_uniform(t, 0.0f, 100.0f); + init_tensor_uniform(t, 50.0f, 100.0f); } } + + float grad_eps() override { + return 20.0f; + } + + bool grad_precise() override { + return true; + } +}; + +// GGML_OP_LOG +struct test_log : public test_case { + const ggml_type type; + const std::array ne; + + std::string vars() override { + return VARS_TO_STR2(type, ne); + } + + test_log(ggml_type type = GGML_TYPE_F32, + std::array ne = {10, 5, 4, 3}) + : type(type), ne(ne) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + + ggml_tensor * out = ggml_log(ctx, a); + ggml_set_name(out, "out"); + + return out; + } + + void initialize_tensors(ggml_context * ctx) override { + for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { + // log(1) == 0, cluster values there to keep the sum low for better precision in the backwards pass: + init_tensor_uniform(t, 0.9f, 1.1f); + } + } + + bool grad_precise() override { + return true; + } }; // GGML_OP_SIN @@ -1170,20 +1796,37 @@ struct test_sin : public test_case { } test_sin(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}) + std::array ne = {10, 2, 2, 2}) : type(type), ne(ne) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_sin(ctx, a); + ggml_set_name(out, "out"); + return out; } void initialize_tensors(ggml_context * ctx) override { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { - init_tensor_uniform(t, -100.0f, 100.0f); + init_tensor_uniform(t, -6.5f, 6.5f); // Covers interval [-2*pi, 2*pi]. } } + + double max_maa_err() override { + return 1e-3; + } + + float grad_eps() override { + return 0.2f; + } + + bool grad_precise() override { + return true; + } }; // GGML_OP_COS @@ -1196,20 +1839,37 @@ struct test_cos : public test_case { } test_cos(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}) + std::array ne = {10, 2, 2, 2}) : type(type), ne(ne) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_cos(ctx, a); + ggml_set_name(out, "out"); + return out; } void initialize_tensors(ggml_context * ctx) override { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { - init_tensor_uniform(t, -100.0f, 100.0f); + init_tensor_uniform(t, -6.5f, 6.5f); // Covers interval [-2*pi, 2*pi]. } } + + double max_maa_err() override { + return 1e-3; + } + + float grad_eps() override { + return 0.2f; + } + + bool grad_precise() override { + return true; + } }; // GGML_OP_CLAMP @@ -1224,15 +1884,27 @@ struct test_clamp : public test_case { } test_clamp(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}, + std::array ne = {10, 5, 4, 3}, float min = -0.5f, float max = 0.5f) : type(type), ne(ne), min(min), max(max) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_clamp(ctx, a, min, max); + ggml_set_name(out, "out"); + return out; } + + float grad_eps() override { + return 1e-2f; + } + + std::vector grad_expect() override { + return {0.0f, 1.0f}; + } }; // GGML_OP_DIAG_MASK_INF @@ -1246,13 +1918,18 @@ struct test_diag_mask_inf : public test_case { } test_diag_mask_inf(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}, + std::array ne = {10, 10, 3, 2}, int n_past = 5) : type(type), ne(ne), n_past(n_past) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_diag_mask_inf(ctx, a, n_past); + ggml_set_name(out, "out"); + return out; } }; @@ -1276,7 +1953,7 @@ struct test_soft_max : public test_case { } test_soft_max(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}, + std::array ne = {10, 5, 4, 3}, bool mask = false, float scale = 1.0f, float max_bias = 0.0f) @@ -1284,13 +1961,24 @@ struct test_soft_max : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * mask = nullptr; if (this->mask) { mask = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ne[0], ne[1]); + ggml_set_name(mask, "mask"); } + ggml_tensor * out = ggml_soft_max_ext(ctx, a, mask, scale, max_bias); + ggml_set_name(out, "out"); + return out; } + + bool grad_precise() override { + return true; + } }; @@ -1312,7 +2000,7 @@ struct test_rope : public test_case { } test_rope(ggml_type type = GGML_TYPE_F32, - std::array ne_a = {10, 10, 10, 1}, + std::array ne_a = {10, 5, 3, 1}, int n_dims = 10, int mode = 0, int n_ctx = 512, float fs = 1.0f, float ef = 0.0f, float af = 0.0f, bool ff = false, int v = 0) : type(type), ne_a(ne_a), n_dims(n_dims), mode(mode), n_ctx(n_ctx), fs(fs), ef(ef), af(af), ff(ff), v(v) {} @@ -1321,13 +2009,29 @@ struct test_rope : public test_case { if (v & 1) { auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3; a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0); + ggml_set_name(a, "view_of_a"); } else { a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); } + ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne_a[2]); - ggml_tensor * freq = ff ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_dims/2) : nullptr; + ggml_set_name(pos, "pos"); + + ggml_tensor * freq = nullptr; + if (ff) { + freq = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_dims/2); + ggml_set_name(freq, "freq"); + } + ggml_tensor * out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f); + ggml_set_name(out, "out"); + return out; } @@ -1350,6 +2054,14 @@ struct test_rope : public test_case { } } } + + double max_maa_err() override { + return 1e-3; + } + + bool grad_precise() override { + return true; + } }; // GGML_OP_POOL2D @@ -1381,7 +2093,12 @@ struct test_pool2d : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data()); + ggml_set_param(ctx, input); + ggml_set_name(input, "input"); + ggml_tensor * out = ggml_pool_2d(ctx, input, pool_type, k0, k1, s0, s1, p0, p1); + ggml_set_name(out, "out"); + return out; } }; @@ -1406,8 +2123,14 @@ struct test_conv_transpose_1d : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data()); + ggml_set_name(input, "input"); + ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_kernel.data()); + ggml_set_name(kernel, "kernel"); + ggml_tensor * out = ggml_conv_transpose_1d(ctx, kernel, input, s0, p0, d0); + ggml_set_name(out, "out"); + return out; } }; @@ -1446,8 +2169,15 @@ struct test_im2col : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data()); + ggml_set_param(ctx, input); + ggml_set_name(input, "input"); + ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel.data()); + ggml_set_name(kernel, "kernel"); + ggml_tensor * out = ggml_im2col(ctx, kernel, input, s0, s1, p0, p1, d0, d1, is_2D, dst_type); + ggml_set_name(out, "out"); + return out; } }; @@ -1465,8 +2195,8 @@ struct test_concat : public test_case { } test_concat(ggml_type type = GGML_TYPE_F32, - std::array ne_a = {10, 10, 10, 10}, - int64_t ne_b_d = 10, + std::array ne_a = {10, 5, 5, 5}, + int64_t ne_b_d = 5, int dim = 2, int v = 0) : type(type), ne_a(ne_a), ne_b_d(ne_b_d), dim(dim), v(v) {} @@ -1477,19 +2207,30 @@ struct test_concat : public test_case { if (v & 1) { auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3; a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); + a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0); + ggml_set_name(a, "view_of_a"); } else { a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + ggml_set_name(a, "a"); } ggml_tensor * b; if (v & 2) { auto ne = ne_b; ne[0] *= 3; ne[1] *= 2; ne[2] *= 4; b = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(b, "b"); + b = ggml_view_4d(ctx, b, ne_b[0], ne_b[1], ne_b[2], ne_b[3], b->nb[1], b->nb[2], b->nb[3], 0); + ggml_set_name(b, "view_of_b"); } else { b = ggml_new_tensor(ctx, type, 4, ne_b.data()); + ggml_set_name(b, "b"); } + ggml_tensor * out = ggml_concat(ctx, a, b, dim); + ggml_set_name(out, "out"); + return out; } }; @@ -1511,7 +2252,11 @@ struct test_argsort : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_argsort(ctx, a, order); + ggml_set_name(out, "out"); + return out; } @@ -1544,6 +2289,35 @@ struct test_argsort : public test_case { } }; +// GGML_OP_SUM +struct test_sum : public test_case { + const ggml_type type; + const std::array ne; + + std::string vars() override { + return VARS_TO_STR2(type, ne); + } + + test_sum(ggml_type type = GGML_TYPE_F32, + std::array ne = {10, 5, 4, 3}) + : type(type), ne(ne) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + + ggml_tensor * out = ggml_sum(ctx, a); + ggml_set_name(out, "out"); + + return out; + } + + float grad_eps() override { + return 0.1f * sqrtf(ne[0]*ne[1]*ne[2]*ne[3]); + } +}; + // GGML_OP_SUM_ROWS struct test_sum_rows : public test_case { const ggml_type type; @@ -1554,12 +2328,17 @@ struct test_sum_rows : public test_case { } test_sum_rows(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}) + std::array ne = {10, 5, 4, 3}) : type(type), ne(ne) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_sum_rows(ctx, a); + ggml_set_name(out, "out"); + return out; } }; @@ -1582,8 +2361,16 @@ struct test_upscale : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - if (transpose) a = ggml_transpose(ctx, a); + ggml_set_name(a, "a"); + + if (transpose) { + a = ggml_transpose(ctx, a); + ggml_set_name(a, "a_transposed"); + } + ggml_tensor * out = ggml_upscale(ctx, a, scale_factor); + ggml_set_name(out, "out"); + return out; } }; @@ -1605,7 +2392,11 @@ struct test_upscale_ext : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3]); + ggml_set_name(out, "out"); + return out; } }; @@ -1629,7 +2420,11 @@ struct test_group_norm : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_group_norm(ctx, a, num_groups, eps); + ggml_set_name(out, "out"); + return out; } }; @@ -1645,14 +2440,22 @@ struct test_acc : public test_case { } test_acc(ggml_type type = GGML_TYPE_F32, - std::array ne_a = {1024, 577, 1, 1}, - std::array ne_b = {1024, 576, 1, 1}) + std::array ne_a = {256, 17, 1, 1}, + std::array ne_b = {256, 16, 1, 1}) : type(type), ne_a(ne_a), ne_b(ne_b) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + ggml_set_param(ctx, a); + ggml_set_name(a, "a"); + ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne_b.data()); + ggml_set_param(ctx, b); + ggml_set_name(b, "b"); + ggml_tensor * out = ggml_acc(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], b->nb[1]); + ggml_set_name(out, "out"); + return out; } }; @@ -1675,7 +2478,11 @@ struct test_pad : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_pad(ctx, a, pad_0, pad_1, 0, 0); + ggml_set_name(out, "out"); + return out; } }; @@ -1697,6 +2504,8 @@ struct test_arange : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * out = ggml_arange(ctx, start, stop, step); + ggml_set_name(out, "out"); + return out; } }; @@ -1719,7 +2528,11 @@ struct test_timestep_embedding : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_timestep_embedding(ctx, a, dim, max_period); + ggml_set_name(out, "out"); + return out; } }; @@ -1735,13 +2548,17 @@ struct test_leaky_relu : public test_case { } test_leaky_relu(ggml_type type = GGML_TYPE_F32, - std::array ne_a = {10, 10, 10, 10}, + std::array ne_a = {10, 5, 4, 3}, float negative_slope = 0.1f) : type(type), ne_a(ne_a), negative_slope(negative_slope) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + ggml_set_name(a, "a"); + ggml_tensor * out = ggml_leaky_relu(ctx, a, negative_slope, true); + ggml_set_name(out, "out"); + return out; } }; @@ -1768,19 +2585,37 @@ struct test_flash_attn_ext : public test_case { return 5e-4; } - test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16) + test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, + bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16) : hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV) {} ggml_tensor * build_graph(ggml_context * ctx) override { const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV)); ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs_padded, nb, nh, 1); + ggml_set_name(q, "q"); + ggml_tensor * k = ggml_new_tensor_4d(ctx, type_KV, hs_padded, kv, nh, 1); + ggml_set_name(k, "k"); + ggml_tensor * v = ggml_new_tensor_4d(ctx, type_KV, hs_padded, kv, nh, 1); - ggml_tensor * m = mask ? ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1) : nullptr; + ggml_set_name(v, "v"); + + ggml_tensor * m = nullptr; + if (mask) { + m = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1); + ggml_set_name(m, "m"); + } + ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap); + ggml_set_name(out, "out"); + return out; } + + bool grad_precise() override { + return true; + } }; // GGML_OP_CROSS_ENTROPY_LOSS @@ -1793,15 +2628,42 @@ struct test_cross_entropy_loss : public test_case { } test_cross_entropy_loss(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}) + std::array ne = {10, 5, 4, 3}) : type(type), ne(ne) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * logits = ggml_new_tensor(ctx, type, 4, ne.data()); + ggml_set_param(ctx, logits); + ggml_set_name(logits, "logits"); + ggml_tensor * labels = ggml_new_tensor(ctx, type, 4, ne.data()); + // The labels are assumed to be constant -> no gradients. + ggml_set_name(labels, "labels"); + + // Ensure labels add up to 1: + labels = ggml_soft_max(ctx, labels); + ggml_set_name(labels, "labels_normalized"); + ggml_tensor * out = ggml_cross_entropy_loss(ctx, logits, labels); + ggml_set_name(out, "out"); + return out; } + + void initialize_tensors(ggml_context * ctx) override { + // For larger abs. diffs between logits softmax is more linear, therefore more precise num. gradients. + for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { + init_tensor_uniform(t, -100.0f, 100.0f); + } + } + + float grad_eps() override { + return 1.0f; + } + + bool grad_precise() override { + return true; + } }; enum llm_norm_type { @@ -2188,6 +3050,12 @@ struct test_falcon : public test_llm { } }; + +// ########################################### +// ## Section 3: GGML Op Test Instantiation ## +// ########################################### + + static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) { std::vector> test_cases; std::default_random_engine rng(0); @@ -2230,8 +3098,8 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op // unary ops for (int v : {0, 1}) { for (int op = 0; op < GGML_UNARY_OP_COUNT; op++) { - test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 128, 10, 10, 10 }, v)); - test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 7, 13, 19, 23 }, v)); + test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 128, 2, 2, 2 }, v)); + test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 5, 7, 11, 13 }, v)); } } @@ -2267,11 +3135,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op } } + test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32)); test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32)); test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16)); // test cases for 1D im2col - test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false)); + test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false)); test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false)); + test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false)); // sycl backend will limit task global_range < MAX_INT // test cases for 2D im2col with large input W and H (occurs in stable-diffusion) @@ -2290,13 +3160,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1)); - test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 1})); - test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {2, 1, 1, 1})); - test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 2, 1, 1})); - test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 2, 1})); - test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 2})); - test_cases.emplace_back(new test_repeat(GGML_TYPE_I32, {10, 10, 10, 10}, {2, 1, 1, 1})); - test_cases.emplace_back(new test_repeat(GGML_TYPE_I16, {10, 10, 10, 10}, {1, 1, 1, 2})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 1, 1})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, 3}, {2, 1, 1, 1})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 2, 1, 1})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 1})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 1, 2})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_I32, {10, 5, 4, 3}, {2, 1, 1, 1})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_I16, {10, 5, 4, 3}, {1, 1, 1, 2})); test_cases.emplace_back(new test_dup(GGML_TYPE_F32)); test_cases.emplace_back(new test_dup(GGML_TYPE_F16)); @@ -2309,6 +3179,10 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {0, 2, 1, 3})); test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {1, 2, 0, 3})); + for (int dim = 1; dim < GGML_MAX_DIMS; ++dim) { + test_cases.emplace_back(new test_set(GGML_TYPE_F32, GGML_TYPE_F32, {6, 5, 4, 3}, dim)); + } + for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) { for (ggml_type type_dst : all_types) { test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4})); @@ -2341,16 +3215,16 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 8, 1}, {1, 1, 1, 1}); add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1, 1}, {32, 1, 1, 1}); add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 320, 320}, {1, 1, 1, 1}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 1, 1}, {1, 1, 1, 1}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 1}, {1, 1, 1, 1}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 1, 1, 1}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {2, 1, 1, 1}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 2, 1, 1}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 1, 2, 1}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 1, 1, 2}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 1, 2, 2}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 2, 2, 2}); - add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {2, 2, 2, 2}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 1, 1}, {1, 1, 1, 1}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 1}, {1, 1, 1, 1}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 1, 1}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {2, 1, 1, 1}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 2, 1, 1}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 1}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 1, 2}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 2}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 2, 2, 2}); + add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {2, 2, 2, 2}); // stable diffusion add_test_bin_bcast(GGML_TYPE_F32, {1280, 1, 1, 1}, {1, 1, 1, 1}); @@ -2369,11 +3243,12 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op //add_test_bin_bcast(GGML_TYPE_F32, {3, 3, 2560, 1280}, {1, 1, 1, 1}); //add_test_bin_bcast(GGML_TYPE_F32, {3, 3, 2560, 1280}, {2, 1, 1, 1}); + test_cases.emplace_back(new test_add1()); test_cases.emplace_back(new test_scale()); for (float eps : {1e-6f, 1e-5f, 1e-3f, 1e-1f}) { - test_cases.emplace_back(new test_norm(GGML_TYPE_F32, {64, 10, 10, 10}, eps)); - test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 10, 10, 10}, eps)); + test_cases.emplace_back(new test_norm(GGML_TYPE_F32, {64, 5, 4, 3}, eps)); + test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 5, 4, 3}, eps)); } test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 1, 1}, {4, 1536, 1, 1})); @@ -2477,13 +3352,14 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op test_cases.emplace_back(new test_sqr()); test_cases.emplace_back(new test_sqrt()); + test_cases.emplace_back(new test_log()); test_cases.emplace_back(new test_sin()); test_cases.emplace_back(new test_cos()); test_cases.emplace_back(new test_clamp()); - test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 1, 1}, 5)); - test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 1}, 5)); - test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 10}, 5)); + test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 1, 1}, 5)); + test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 3, 1}, 5)); + test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 3, 2}, 5)); #if 0 std::uniform_int_distribution<> dist_ne1(1, 50); @@ -2527,23 +3403,23 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op for (float af : { 1.0f, 1.4245f }) { for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) { for (bool ff : {false, true}) { // freq_factors - test_cases.emplace_back(new test_rope(type, {128, 32, 10, 1}, 128, 0, 512, fs, ef, af, ff, v)); // llama 7B + test_cases.emplace_back(new test_rope(type, {128, 32, 2, 1}, 128, 0, 512, fs, ef, af, ff, v)); // llama 7B if (all) { - test_cases.emplace_back(new test_rope(type, {128, 40, 10, 1}, 128, 0, 512, fs, ef, af, ff, v)); // llama 13B - test_cases.emplace_back(new test_rope(type, {128, 52, 10, 1}, 128, 0, 512, fs, ef, af, ff, v)); // llama 30B - test_cases.emplace_back(new test_rope(type, {128, 64, 10, 1}, 128, 0, 512, fs, ef, af, ff, v)); // llama 65B + test_cases.emplace_back(new test_rope(type, {128, 40, 2, 1}, 128, 0, 512, fs, ef, af, ff, v)); // llama 13B + test_cases.emplace_back(new test_rope(type, {128, 52, 2, 1}, 128, 0, 512, fs, ef, af, ff, v)); // llama 30B + test_cases.emplace_back(new test_rope(type, {128, 64, 2, 1}, 128, 0, 512, fs, ef, af, ff, v)); // llama 65B } if (all) { - test_cases.emplace_back(new test_rope(type, { 64, 1, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 7B) - test_cases.emplace_back(new test_rope(type, { 64, 71, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 7B) - test_cases.emplace_back(new test_rope(type, { 64, 8, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 40B) - test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 20, 2, 512, fs, ef, af, ff, v)); // neox (stablelm) - test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 32, 2, 512, fs, ef, af, ff, v)); // neox (phi-2) + test_cases.emplace_back(new test_rope(type, { 64, 1, 2, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 7B) + test_cases.emplace_back(new test_rope(type, { 64, 71, 2, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 7B) + test_cases.emplace_back(new test_rope(type, { 64, 8, 2, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 40B) + test_cases.emplace_back(new test_rope(type, { 80, 32, 2, 1}, 20, 2, 512, fs, ef, af, ff, v)); // neox (stablelm) + test_cases.emplace_back(new test_rope(type, { 80, 32, 2, 1}, 32, 2, 512, fs, ef, af, ff, v)); // neox (phi-2) } - test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 40B) + test_cases.emplace_back(new test_rope(type, { 64, 128, 2, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 40B) } } @@ -2567,6 +3443,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {60, 10, 10, 10}, order)); // qwen } + test_cases.emplace_back(new test_sum()); test_cases.emplace_back(new test_sum_rows()); test_cases.emplace_back(new test_upscale()); test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true)); @@ -2609,6 +3486,18 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op #endif // run tests + if (mode == MODE_GRAD) { + size_t n_ok = 0; + for (auto & test : test_cases) { + if (test->eval_grad(backend, op_name)) { + n_ok++; + } + } + printf(" %zu/%zu tests passed\n", n_ok, test_cases.size()); + + return n_ok == test_cases.size(); + } + if (mode == MODE_TEST) { ggml_backend_t backend_cpu = ggml_backend_cpu_init(); @@ -2637,8 +3526,11 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op static void usage(char ** argv) { printf("Usage: %s [mode] [-o op] [-b backend]\n", argv[0]); - printf(" valid modes are: test (compare with CPU backend for correctness) or perf (performance evaluation)\n"); - printf(" op names are as given by ggml_op_desc()\n"); + printf(" valid modes:\n"); + printf(" - test (default, compare with CPU backend for correctness)\n"); + printf(" - perf (performance evaluation)\n"); + printf(" - grad (compare gradients from backpropagation with method of finite differences)\n"); + printf(" op names are as given by ggml_op_desc() (e.g. GGML_ADD)\n"); } int main(int argc, char ** argv) { @@ -2651,6 +3543,8 @@ int main(int argc, char ** argv) { mode = MODE_TEST; } else if (strcmp(argv[i], "perf") == 0) { mode = MODE_PERF; + } else if (strcmp(argv[i], "grad") == 0) { + mode = MODE_GRAD; } else if (strcmp(argv[i], "-o") == 0) { if (i + 1 < argc) { op_name_filter = argv[++i]; @@ -2688,7 +3582,7 @@ int main(int argc, char ** argv) { ggml_backend_t backend = ggml_backend_reg_init_backend(i, NULL); GGML_ASSERT(backend != NULL); - if (backend_filter == NULL && ggml_backend_is_cpu(backend)) { + if (backend_filter == NULL && ggml_backend_is_cpu(backend) && mode != MODE_GRAD) { printf(" Skipping CPU backend\n"); ggml_backend_free(backend); n_ok++; From 9cb9260861e464e40d38764cfb021856786c7202 Mon Sep 17 00:00:00 2001 From: Salvatore Mesoraca Date: Fri, 6 Sep 2024 14:34:07 +0200 Subject: [PATCH 07/32] vulkan: correctly report support for OP_CONT (ggml/946) test-backend-ops fails because ggml_cont aborts when invoked passing an unsupported type. This commit makes ggml_cont tests pass Signed-off-by: Salvatore Mesoraca --- ggml/src/ggml-vulkan.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp index 0b7a5b6e2..2d8e57756 100644 --- a/ggml/src/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan.cpp @@ -6602,6 +6602,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const return false; } } break; + case GGML_OP_CONT: case GGML_OP_CPY: case GGML_OP_DUP: { @@ -6642,7 +6643,6 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const case GGML_OP_COS: case GGML_OP_CLAMP: case GGML_OP_PAD: - case GGML_OP_CONT: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: case GGML_OP_ARGSORT: From 406c1a32a18807b9b5711aa2fa1859527106bc1d Mon Sep 17 00:00:00 2001 From: Salvatore Mesoraca Date: Fri, 6 Sep 2024 14:34:25 +0200 Subject: [PATCH 08/32] vulkan: add dryrun support to sin and cos ops (ggml/947) sin and cos failed test-backend-ops because they tried to dereference a context pointer that is null on dry runs. This commit prevents that segfault. Signed-off-by: Salvatore Mesoraca --- ggml/src/ggml-vulkan.cpp | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp index 2d8e57756..d6f647c89 100644 --- a/ggml/src/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan.cpp @@ -4616,7 +4616,7 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const }, dryrun); } -static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4626,10 +4626,10 @@ static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, - }); + }, dryrun); } -static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { +static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -4639,7 +4639,7 @@ static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, - }); + }, dryrun); } static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { @@ -5783,11 +5783,11 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod break; case GGML_OP_SIN: - ggml_vk_sin(ctx, compute_ctx, src0, node); + ggml_vk_sin(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_COS: - ggml_vk_cos(ctx, compute_ctx, src0, node); + ggml_vk_cos(ctx, compute_ctx, src0, node, dryrun); break; case GGML_OP_CLAMP: From 60a3107ccd791d858e12a87e6024ce918d3bf595 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 8 Sep 2024 09:38:42 +0300 Subject: [PATCH 09/32] scripts : option to increase git patch context --- scripts/sync-ggml-am.sh | 26 ++++++++++++++++++++------ 1 file changed, 20 insertions(+), 6 deletions(-) diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh index b29892565..f16336594 100755 --- a/scripts/sync-ggml-am.sh +++ b/scripts/sync-ggml-am.sh @@ -5,7 +5,7 @@ # Usage: # # $ cd /path/to/llama.cpp -# $ ./scripts/sync-ggml-am.sh -skip hash0,hash1,hash2... +# $ ./scripts/sync-ggml-am.sh -skip hash0,hash1,hash2... -C 3 # set -e @@ -25,9 +25,23 @@ lc=$(cat $SRC_LLAMA/scripts/sync-ggml.last) echo "Syncing ggml changes since commit $lc" to_skip="" -if [ "$1" == "-skip" ]; then - to_skip=$2 -fi + +# context for git patches in number of lines +ctx="8" + +while [ "$1" != "" ]; do + case $1 in + -skip ) + shift + to_skip=$1 + ;; + -C ) + shift + ctx=$1 + ;; + esac + shift +done cd $SRC_GGML @@ -52,7 +66,7 @@ while read c; do fi fi - git format-patch -k $c~1..$c --stdout -- \ + git format-patch -U${ctx} -k $c~1..$c --stdout -- \ CMakeLists.txt \ src/CMakeLists.txt \ cmake/FindSIMD.cmake \ @@ -191,7 +205,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then > ggml-src.patch.tmp mv ggml-src.patch.tmp ggml-src.patch - git am ggml-src.patch + git am -C${ctx} ggml-src.patch rm -v $SRC_LLAMA/ggml-src.patch fi From 385decbd632f70db9f1fbd93dbb2b908853e135c Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 8 Sep 2024 09:38:56 +0300 Subject: [PATCH 10/32] sync : ggml --- scripts/sync-ggml.last | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index 1e6db754f..3d2dfb413 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -28b7633d733bbeef0026570fbc61c79c5e9aa5ae +10e83a412717c20d57ba19f025248e18e43addf3 From a8768614558262b6762fc0feeddcc6f7ce4bc5fc Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 8 Sep 2024 09:57:57 +0300 Subject: [PATCH 11/32] metal : update support condition for im2col + fix warning (#0) --- ggml/src/ggml-metal.m | 3 ++- tests/test-backend-ops.cpp | 4 ++-- 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m index 91b5e61b2..f04e5af71 100644 --- a/ggml/src/ggml-metal.m +++ b/ggml/src/ggml-metal.m @@ -799,8 +799,9 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx return ctx->support_simdgroup_reduction; case GGML_OP_NORM: case GGML_OP_ROPE: - case GGML_OP_IM2COL: return true; + case GGML_OP_IM2COL: + return op->src[0]->type == GGML_TYPE_F16; case GGML_OP_POOL_1D: case GGML_OP_POOL_2D: return false; diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 8f659c148..635de01d7 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -24,6 +24,7 @@ #include #include #include +#include #include #include #include @@ -33,7 +34,6 @@ #include #include - static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) { // static RNG initialization (revisit if n_threads stops being constant) static const size_t n_threads = std::thread::hardware_concurrency(); @@ -869,7 +869,7 @@ struct test_case { for (int64_t i = 0; i < ne; ++i) { // gradient algebraic // check for nans if (!std::isfinite(ga[i])) { - printf("[%s] nonfinite gradient at index %zu (%s=%f) ", ggml_op_desc(t), i, bn, ga[i]); + printf("[%s] nonfinite gradient at index %" PRId64 " (%s=%f) ", ggml_op_desc(t), i, bn, ga[i]); ok = false; break; } From 00b02bb249406ec0123757a67a5b714c3f33a699 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Sun, 8 Sep 2024 12:12:17 +0200 Subject: [PATCH 12/32] imatrix : fix arg parser for imatrix (#9366) * imatrix : fix arg parser * beautify printing first arg --- common/common.cpp | 4 +++- examples/imatrix/imatrix.cpp | 2 +- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index c5c4d7508..9228eafca 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -600,7 +600,9 @@ std::string llama_arg::to_string() { if (args.size() == 1) { ss << arg; } else { - ss << format("%-7s", arg) << ", "; + // first arg is usually abbreviation, we need padding to make it more beautiful + auto tmp = std::string(arg) + ", "; + ss << format("%-7s", tmp.c_str()); } } else { ss << arg << (arg != args.back() ? ", " : ""); diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 302292ab2..15a3f0d14 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -577,7 +577,7 @@ int main(int argc, char ** argv) { params.logits_all = true; params.verbosity = 1; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, print_usage); + auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_IMATRIX, print_usage); if (!gpt_params_parse(argc, argv, params, options)) { return 1; } From eae597182cb61bbde0b26e7cec5999d28b9327fe Mon Sep 17 00:00:00 2001 From: slaren Date: Sun, 8 Sep 2024 12:41:51 +0200 Subject: [PATCH 13/32] llama : sanitize tokens in the upper bound (#9359) --- src/llama.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/llama.cpp b/src/llama.cpp index f590bcd3b..39e20440e 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -16077,7 +16077,7 @@ static int llama_decode_internal( } for (uint32_t i = 0; i < n_tokens_all; ++i) { - if (batch_all.token[i] < 0) { + if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= lctx.model.vocab.n_vocab) { LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch_all.token[i]); return -1; } @@ -16376,7 +16376,7 @@ static int llama_encode_internal( } for (uint32_t i = 0; i < n_tokens; ++i) { - if (batch.token[i] < 0) { + if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= lctx.model.vocab.n_vocab) { LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch.token[i]); return -1; } From 2a358fb0c4b6e917ac852aa17444cc94dd28a2a6 Mon Sep 17 00:00:00 2001 From: Neo Zhang Jianyu Date: Sun, 8 Sep 2024 19:05:29 +0800 Subject: [PATCH 14/32] [SYCL] add check malloc result on device (#9346) * add check malloc result on device * update for review comments, check all malloc_device() result --------- Co-authored-by: arthw <14088817+arthw@users.noreply.github.com> --- ggml/src/ggml-sycl.cpp | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index 0d884f89a..4f03b01e7 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -1954,6 +1954,11 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool { SYCL_CHECK( CHECK_TRY_ERROR(ptr = (void *)sycl::malloc_device( look_ahead_size, *qptr))); + if (!ptr) { + fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, look_ahead_size); + return nullptr; + } + *actual_size = look_ahead_size; pool_size += look_ahead_size; @@ -4350,6 +4355,10 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, void * dev_ptr; SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device( size, *stream))); + if (!dev_ptr) { + fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, size); + return nullptr; + } ggml_backend_sycl_buffer_context * ctx = new ggml_backend_sycl_buffer_context(buft_ctx->device, dev_ptr, buft_ctx->stream); return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size); } @@ -4570,7 +4579,11 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer, */ SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)sycl::malloc_device( size, *stream))); - + if (!buf) { + char err_buf[1024]; + snprintf(err_buf, 1023, "%s: can't malloc %lu Bytes memory on device", __func__, size); + throw std::runtime_error(err_buf); + } // set padding to 0 to avoid possible NaN values if (size > original_size) { /* From 19f4a7b296efda7c13a6b21d428b2286b5d1aa06 Mon Sep 17 00:00:00 2001 From: slaren Date: Sun, 8 Sep 2024 15:52:07 +0200 Subject: [PATCH 15/32] llama : refactor samplers internal implementation (#9370) --- src/llama-impl.h | 4 + src/llama-sampling.cpp | 1582 +++++++++++++++++++++------------------ src/llama-sampling.h | 10 - tests/test-sampling.cpp | 10 +- 4 files changed, 881 insertions(+), 725 deletions(-) diff --git a/src/llama-impl.h b/src/llama-impl.h index fa2e09e1f..87012617f 100644 --- a/src/llama-impl.h +++ b/src/llama-impl.h @@ -101,6 +101,10 @@ struct ring_buffer { } void push_back(const T & value) { + if (capacity == 0) { + throw std::runtime_error("ring buffer: capacity is zero"); + } + if (sz == capacity) { // advance the start when buffer is full first = (first + 1) % capacity; diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp index 1661d9a83..41f48ec28 100644 --- a/src/llama-sampling.cpp +++ b/src/llama-sampling.cpp @@ -13,12 +13,40 @@ #include static int llama_sample_dist(llama_token_data_array * cur_p, std::mt19937 & rng, std::vector & probs) { +#if 1 probs.resize(cur_p->size); for (size_t i = 0; i < cur_p->size; ++i) { probs[i] = cur_p->data[i].p; } std::discrete_distribution dist(probs.begin(), probs.end()); +#else + // avoid the copy with a custom iterator + #pragma GCC diagnostic push + #pragma GCC diagnostic ignored "-Wunused-local-typedefs" + + struct probs_iterator { + typedef std::input_iterator_tag iterator_category; + typedef float value_type; + typedef float * pointer; + typedef float & reference; + typedef size_t difference_type; + + const llama_token_data_array * data; + size_t i; + + bool operator==(const probs_iterator & other) const { return data + i == other.data + other.i; } + bool operator!=(const probs_iterator & other) const { return data + i != other.data + other.i; } + float operator*() const { return data->data[i].p; } + probs_iterator & operator++() { ++i; return *this; } + probs_iterator operator++(int) { probs_iterator tmp = *this; ++i; return tmp; } + }; + #pragma GCC diagnostic pop + + std::discrete_distribution dist(probs_iterator{cur_p, 0}, probs_iterator{cur_p, cur_p->size}); + + GGML_UNUSED(probs); +#endif return dist(rng); } @@ -138,301 +166,6 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) cur_p->size = k; } -static void llama_sampler_top_p_impl(llama_token_data_array * cur_p, float p, size_t min_keep) { - if (p >= 1.0f) { - return; - } - - llama_sampler_softmax_impl(cur_p); - - // Compute the cumulative probabilities - float cum_sum = 0.0f; - size_t last_idx = cur_p->size; - - for (size_t i = 0; i < cur_p->size; ++i) { - cum_sum += cur_p->data[i].p; - - // Check if the running sum is at least p or if we have kept at least min_keep tokens - // we set the last index to i+1 to indicate that the current iterate should be included in the set - if (cum_sum >= p && i + 1 >= min_keep) { - last_idx = i + 1; - break; - } - } - - // Resize the output vector to keep only the top-p tokens - cur_p->size = last_idx; -} - -static void llama_sampler_min_p_impl(llama_token_data_array * cur_p, float p, size_t min_keep) { - if (p <= 0.0f || !cur_p->size) { - return; - } - - bool min_p_applied = false; - - // if the cur_p aren't sorted, try the unsorted implementation first - if (!cur_p->sorted) { - std::vector filtered_tokens; - - float max_logit = -FLT_MAX; - for (size_t i = 0; i < cur_p->size; ++i) { - max_logit = std::max(max_logit, cur_p->data[i].logit); - } - const float min_logit = max_logit + logf(p); // min logit for p_i >= p * p_max - - for (size_t i = 0; i < cur_p->size; ++i) { - if (cur_p->data[i].logit >= min_logit) { - filtered_tokens.push_back(cur_p->data[i]); - } - } - - // if we have enough values the operation was a success - if (filtered_tokens.size() >= min_keep) { - memcpy(cur_p->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data)); - cur_p->size = filtered_tokens.size(); - min_p_applied = true; - } - } - - // if the cur_p are sorted or the unsorted implementation failed, use this implementation - if (!min_p_applied) { - // Sort the logits in descending order - if (!cur_p->sorted) { - std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) { - return a.logit > b.logit; - }); - cur_p->sorted = true; - } - - const float min_logit = cur_p->data[0].logit + logf(p); // min logit for p_i >= p * p_max - size_t i = 1; // first token always matches - - for (; i < cur_p->size; ++i) { - if (cur_p->data[i].logit < min_logit && i >= min_keep) { - break; // prob too small - } - } - - // Resize the output vector to keep only the matching tokens - cur_p->size = i; - } -} - -static void llama_sampler_tail_free_impl(llama_token_data_array * cur_p, float z, size_t min_keep) { - if (z >= 1.0f || cur_p->size <= 2) { - return; - } - - llama_sampler_softmax_impl(cur_p); - - // Compute the first and second derivatives - std::vector first_derivatives(cur_p->size - 1); - std::vector second_derivatives(cur_p->size - 2); - - for (size_t i = 0; i < first_derivatives.size(); ++i) { - first_derivatives[i] = cur_p->data[i].p - cur_p->data[i + 1].p; - } - for (size_t i = 0; i < second_derivatives.size(); ++i) { - second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1]; - } - - // Calculate absolute value of second derivatives - for (size_t i = 0; i < second_derivatives.size(); ++i) { - second_derivatives[i] = std::abs(second_derivatives[i]); - } - - // Normalize the second derivatives - { - const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); - - if (second_derivatives_sum > 1e-6f) { - for (float & value : second_derivatives) { - value /= second_derivatives_sum; - } - } else { - for (float & value : second_derivatives) { - value = 1.0f / second_derivatives.size(); - } - } - } - - float cum_sum = 0.0f; - size_t last_idx = cur_p->size; - for (size_t i = 0; i < second_derivatives.size(); ++i) { - cum_sum += second_derivatives[i]; - - // Check if the running sum is greater than z or if we have kept at least min_keep tokens - if (cum_sum > z && i >= min_keep) { - last_idx = i; - break; - } - } - - // Resize the output vector to keep only the tokens above the tail location - cur_p->size = last_idx; -} - -static void llama_sampler_typical_impl(llama_token_data_array * cur_p, float p, size_t min_keep) { - // Reference implementation: - // https://github.com/huggingface/transformers/compare/main...cimeister:typical-sampling:typical-pr - if (p >= 1.0f) { - return; - } - - // Compute the softmax of logits and calculate entropy - llama_sampler_softmax_impl(cur_p); - - float entropy = 0.0f; - for (size_t i = 0; i < cur_p->size; ++i) { - entropy += -cur_p->data[i].p * logf(cur_p->data[i].p); - } - - // Compute the absolute difference between negative log probability and entropy for each candidate - std::vector shifted_scores; - for (size_t i = 0; i < cur_p->size; ++i) { - float shifted_score = fabsf(-logf(cur_p->data[i].p) - entropy); - shifted_scores.push_back(shifted_score); - } - - // Sort tokens based on the shifted_scores and their corresponding indices - std::vector indices(cur_p->size); - std::iota(indices.begin(), indices.end(), 0); - - std::sort(indices.begin(), indices.end(), [&](size_t a, size_t b) { - return shifted_scores[a] < shifted_scores[b]; - }); - - // Compute the cumulative probabilities - float cum_sum = 0.0f; - size_t last_idx = indices.size(); - - for (size_t i = 0; i < indices.size(); ++i) { - size_t idx = indices[i]; - cum_sum += cur_p->data[idx].p; - - // Check if the running sum is greater than typical or if we have kept at least min_keep tokens - if (cum_sum > p && i >= min_keep - 1) { - last_idx = i + 1; - break; - } - } - - // Resize the output vector to keep only the locally typical tokens - std::vector cur_p_new; - for (size_t i = 0; i < last_idx; ++i) { - size_t idx = indices[i]; - cur_p_new.push_back(cur_p->data[idx]); - } - - // Replace the data in cur_p with the cur_p_new data - std::copy(cur_p_new.begin(), cur_p_new.end(), cur_p->data); - cur_p->size = cur_p_new.size(); - cur_p->sorted = false; -} - -static void llama_sampler_entropy_impl(llama_token_data_array * cur_p, float min_temp, float max_temp, float exponent_val) { - // no need to do anything if there is only one (or zero) candidates - if (cur_p->size <= 1) { - return; - } - - // Calculate maximum possible entropy - float max_entropy = -logf(1.0f / cur_p->size); - - llama_sampler_softmax_impl(cur_p); - - // Calculate entropy of the softmax probabilities - float entropy = 0.0f; - for (size_t i = 0; i < cur_p->size; ++i) { - float prob = cur_p->data[i].p; - if (prob > 0.0f) { // Ensure no log(0) - entropy -= prob * logf(prob); - } - } - - // Normalize the entropy (max_entropy cannot be 0 here because we checked cur_p->size != 1 above) - float normalized_entropy = entropy / max_entropy; - - // Map the normalized entropy to the desired temperature range using the power function - float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val); - -#ifdef DEBUG - LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp); - LLAMA_LOG_INFO("Entropy: %f\n", entropy); - LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy); - LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy); - LLAMA_LOG_INFO("Exponent: %f\n", exponent_val); - LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp); -#endif - - // Apply the dynamically calculated temperature scaling - for (size_t i = 0; i < cur_p->size; ++i) { - cur_p->data[i].logit /= dyn_temp; - } - - // Re-compute softmax probabilities after scaling logits with dynamic temperature - const double max_l_double = cur_p->data[0].logit; - - double cum_sum_double = 0.0; - for (size_t i = 0; i < cur_p->size; ++i) { - double p = exp(cur_p->data[i].logit - max_l_double); - cur_p->data[i].p = p; // Store the scaled probability - cum_sum_double += p; - } - - for (size_t i = 0; i < cur_p->size; ++i) { - cur_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities - } - -#ifdef DEBUG - // Print the updated top 25 probabilities after temperature scaling - LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n"); - for (size_t i = 0; i < 25 && i < cur_p->size; ++i) { - LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, cur_p->data[i].p * 100.0f); - } -#endif -} - -static void llama_sampler_temp_impl(llama_token_data_array * cur_p, float temp) { - for (size_t i = 0; i < cur_p->size; ++i) { - cur_p->data[i].logit /= temp; - } -} - -static void llama_sampler_grammar_impl(llama_token_data_array * cur_p, const struct llama_grammar & grammar) { - llama_grammar_apply_impl(grammar, cur_p); -} - -void llama_sampler_penalties_impl( - llama_token_data_array * cur_p, - const llama_token_cnt & token_count, - float penalty_repeat, - float penalty_freq, - float penalty_present) { - // Apply frequency and presence penalties to the cur_p - for (size_t i = 0; i < cur_p->size; ++i) { - const auto token_iter = token_count.find(cur_p->data[i].id); - if (token_iter == token_count.end()) { - continue; - } - - const int count = token_iter->second; - - // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong. - // This is common fix for this problem, which is to multiply by the penalty instead of dividing. - if (cur_p->data[i].logit <= 0) { - cur_p->data[i].logit *= penalty_repeat; - } else { - cur_p->data[i].logit /= penalty_repeat; - } - - cur_p->data[i].logit -= float(count) * penalty_freq + float(count > 0) * penalty_present; - } - - cur_p->sorted = false; -} - // llama_sampler API const char * llama_sampler_name(const struct llama_sampler * smpl) { @@ -600,17 +333,23 @@ int llama_sampler_chain_n(const struct llama_sampler * chain) { // greedy -static struct llama_sampler_i llama_sampler_greedy_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "greedy"; }, - /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * /*smpl*/, llama_token_data_array * cur_p) { - cur_p->selected = 0; - for (size_t i = 1; i < cur_p->size; ++i) { - if (cur_p->data[i].logit > cur_p->data[cur_p->selected].logit) { - cur_p->selected = i; - } +static const char * llama_sampler_greedy_name(const struct llama_sampler * /*smpl*/) { + return "greedy"; +} + +static void llama_sampler_greedy_apply(struct llama_sampler * /*smpl*/, llama_token_data_array * cur_p) { + cur_p->selected = 0; + for (size_t i = 1; i < cur_p->size; ++i) { + if (cur_p->data[i].logit > cur_p->data[cur_p->selected].logit) { + cur_p->selected = i; } - }, + } +} + +static struct llama_sampler_i llama_sampler_greedy_i = { + /* .name = */ llama_sampler_greedy_name, + /* .accept = */ nullptr, + /* .apply = */ llama_sampler_greedy_apply, /* .reset = */ nullptr, /* .clone = */ nullptr, /* .free = */ nullptr, @@ -633,30 +372,45 @@ struct llama_sampler_dist { std::vector probs; // work array }; +static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*/) { + return "dist"; +} + +static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + auto * ctx = (llama_sampler_dist *) smpl->ctx; + cur_p->selected = llama_sample_dist(cur_p, ctx->rng, ctx->probs); +} + +static struct llama_sampler * llama_sampler_dist_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_dist *) smpl->ctx; + auto * result = llama_sampler_init_dist(ctx->seed); + + // copy the state + { + auto * result_ctx = (llama_sampler_dist *) result->ctx; + + result_ctx->rng = ctx->rng; + } + + return result; +} + +static void llama_sampler_dist_reset(struct llama_sampler * smpl) { + auto * ctx = (llama_sampler_dist *) smpl->ctx; + ctx->rng = std::mt19937(ctx->seed); +} + +static void llama_sampler_dist_free(struct llama_sampler * smpl) { + delete (llama_sampler_dist *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_dist_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "dist"; }, + /* .name = */ llama_sampler_dist_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - auto * ctx = (llama_sampler_dist *) smpl->ctx; - cur_p->selected = llama_sample_dist(cur_p, ctx->rng, ctx->probs); - }, - /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_dist *) smpl->ctx; - auto * result = llama_sampler_init_dist(ctx->seed); - - // copy the state - { - auto * result_ctx = (llama_sampler_dist *) result->ctx; - - result_ctx->rng = ctx->rng; - } - - return result; - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_dist *) smpl->ctx; - }, + /* .apply = */ llama_sampler_dist_apply, + /* .reset = */ llama_sampler_dist_reset, + /* .clone = */ llama_sampler_dist_clone, + /* .free = */ llama_sampler_dist_free, }; struct llama_sampler * llama_sampler_init_dist(uint32_t seed) { @@ -672,12 +426,18 @@ struct llama_sampler * llama_sampler_init_dist(uint32_t seed) { // softmax +static const char * llama_sampler_softmax_name(const struct llama_sampler * /*smpl*/) { + return "softmax"; +} + +static void llama_sampler_softmax_apply(struct llama_sampler * /*smpl*/, llama_token_data_array * cur_p) { + llama_sampler_softmax_impl(cur_p); +} + static struct llama_sampler_i llama_sampler_softmax_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "softmax"; }, + /* .name = */ llama_sampler_softmax_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * /*smpl*/, llama_token_data_array * cur_p) { - llama_sampler_softmax_impl(cur_p); - }, + /* .apply = */ llama_sampler_softmax_apply, /* .reset = */ nullptr, /* .clone = */ nullptr, /* .free = */ nullptr, @@ -696,21 +456,31 @@ struct llama_sampler_top_k { const int32_t k; }; +static const char * llama_sampler_top_k_name(const struct llama_sampler * /*smpl*/) { + return "top-k"; +} + +static void llama_sampler_top_k_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + const auto * ctx = (llama_sampler_top_k *) smpl->ctx; + llama_sampler_top_k_impl(cur_p, ctx->k); +} + +static struct llama_sampler * llama_sampler_top_k_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_top_k *) smpl->ctx; + return llama_sampler_init_top_k(ctx->k); +} + +static void llama_sampler_top_k_free(struct llama_sampler * smpl) { + delete (llama_sampler_top_k *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_top_k_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "top-k"; }, + /* .name = */ llama_sampler_top_k_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_top_k *) smpl->ctx; - llama_sampler_top_k_impl(cur_p, ctx->k); - }, + /* .apply = */ llama_sampler_top_k_apply, /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_top_k *) smpl->ctx; - return llama_sampler_init_top_k(ctx->k); - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_top_k *) smpl->ctx; - }, + /* .clone = */ llama_sampler_top_k_clone, + /* .free = */ llama_sampler_top_k_free, }; struct llama_sampler * llama_sampler_init_top_k(int32_t k) { @@ -729,21 +499,54 @@ struct llama_sampler_top_p { const size_t min_keep; }; +static const char * llama_sampler_top_p_name(const struct llama_sampler * /*smpl*/) { + return "top-p"; +} + +static void llama_sampler_top_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + const auto * ctx = (llama_sampler_top_p *) smpl->ctx; + + if (ctx->p >= 1.0f) { + return; + } + + llama_sampler_softmax_impl(cur_p); + + // Compute the cumulative probabilities + float cum_sum = 0.0f; + size_t last_idx = cur_p->size; + + for (size_t i = 0; i < cur_p->size; ++i) { + cum_sum += cur_p->data[i].p; + + // Check if the running sum is at least p or if we have kept at least min_keep tokens + // we set the last index to i+1 to indicate that the current iterate should be included in the set + if (cum_sum >= ctx->p && i + 1 >= ctx->min_keep) { + last_idx = i + 1; + break; + } + } + + // Resize the output vector to keep only the top-p tokens + cur_p->size = last_idx; +} + +static struct llama_sampler * llama_sampler_top_p_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_top_p *) smpl->ctx; + return llama_sampler_init_top_p(ctx->p, ctx->min_keep); +} + +static void llama_sampler_top_p_free(struct llama_sampler * smpl) { + delete (llama_sampler_top_p *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_top_p_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "top-p"; }, + /* .name = */ llama_sampler_top_p_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_top_p *) smpl->ctx; - llama_sampler_top_p_impl(cur_p, ctx->p, ctx->min_keep); - }, + /* .apply = */ llama_sampler_top_p_apply, /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_top_p *) smpl->ctx; - return llama_sampler_init_top_p(ctx->p, ctx->min_keep); - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_top_p *) smpl->ctx; - }, + /* .clone = */ llama_sampler_top_p_clone, + /* .free = */ llama_sampler_top_p_free, }; struct llama_sampler * llama_sampler_init_top_p(float p, size_t min_keep) { @@ -763,21 +566,83 @@ struct llama_sampler_min_p { const size_t min_keep; }; +static const char * llama_sampler_min_p_name(const struct llama_sampler * /*smpl*/) { + return "min-p"; +} + +static void llama_sampler_min_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + const auto * ctx = (llama_sampler_min_p *) smpl->ctx; + + if (ctx->p <= 0.0f || !cur_p->size) { + return; + } + + bool min_p_applied = false; + + // if the cur_p aren't sorted, try the unsorted implementation first + if (!cur_p->sorted) { + std::vector filtered_tokens; + + float max_logit = -FLT_MAX; + for (size_t i = 0; i < cur_p->size; ++i) { + max_logit = std::max(max_logit, cur_p->data[i].logit); + } + const float min_logit = max_logit + logf(ctx->p); // min logit for p_i >= p * p_max + + for (size_t i = 0; i < cur_p->size; ++i) { + if (cur_p->data[i].logit >= min_logit) { + filtered_tokens.push_back(cur_p->data[i]); + } + } + + // if we have enough values the operation was a success + if (filtered_tokens.size() >= ctx->min_keep) { + memcpy(cur_p->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data)); + cur_p->size = filtered_tokens.size(); + min_p_applied = true; + } + } + + // if the cur_p are sorted or the unsorted implementation failed, use this implementation + if (!min_p_applied) { + // Sort the logits in descending order + if (!cur_p->sorted) { + std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) { + return a.logit > b.logit; + }); + cur_p->sorted = true; + } + + const float min_logit = cur_p->data[0].logit + logf(ctx->p); // min logit for p_i >= p * p_max + size_t i = 1; // first token always matches + + for (; i < cur_p->size; ++i) { + if (cur_p->data[i].logit < min_logit && i >= ctx->min_keep) { + break; // prob too small + } + } + + // Resize the output vector to keep only the matching tokens + cur_p->size = i; + } +} + +static struct llama_sampler * llama_sampler_min_p_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_min_p *) smpl->ctx; + return llama_sampler_init_min_p(ctx->p, ctx->min_keep); +} + +static void llama_sampler_min_p_free(struct llama_sampler * smpl) { + delete (llama_sampler_min_p *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_min_p_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "min-p"; }, + /* .name = */ llama_sampler_min_p_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_min_p *) smpl->ctx; - llama_sampler_min_p_impl(cur_p, ctx->p, ctx->min_keep); - }, + /* .apply = */ llama_sampler_min_p_apply, /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_min_p *) smpl->ctx; - return llama_sampler_init_min_p(ctx->p, ctx->min_keep); - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_min_p *) smpl->ctx; - }, + /* .clone = */ llama_sampler_min_p_clone, + /* .free = */ llama_sampler_min_p_free, }; struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) { @@ -797,21 +662,82 @@ struct llama_sampler_tail_free { const size_t min_keep; }; +static const char * llama_sampler_tail_free_name(const struct llama_sampler * /*smpl*/) { + return "tail-free"; +} + +static void llama_sampler_tail_free_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + const auto * ctx = (llama_sampler_tail_free *) smpl->ctx; + + if (ctx->z >= 1.0f || cur_p->size <= 2) { + return; + } + + llama_sampler_softmax_impl(cur_p); + + // Compute the first and second derivatives + std::vector first_derivatives(cur_p->size - 1); + std::vector second_derivatives(cur_p->size - 2); + + for (size_t i = 0; i < first_derivatives.size(); ++i) { + first_derivatives[i] = cur_p->data[i].p - cur_p->data[i + 1].p; + } + for (size_t i = 0; i < second_derivatives.size(); ++i) { + second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1]; + } + + // Calculate absolute value of second derivatives + for (size_t i = 0; i < second_derivatives.size(); ++i) { + second_derivatives[i] = std::abs(second_derivatives[i]); + } + + // Normalize the second derivatives + { + const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); + + if (second_derivatives_sum > 1e-6f) { + for (float & value : second_derivatives) { + value /= second_derivatives_sum; + } + } else { + for (float & value : second_derivatives) { + value = 1.0f / second_derivatives.size(); + } + } + } + + float cum_sum = 0.0f; + size_t last_idx = cur_p->size; + for (size_t i = 0; i < second_derivatives.size(); ++i) { + cum_sum += second_derivatives[i]; + + // Check if the running sum is greater than z or if we have kept at least min_keep tokens + if (cum_sum > ctx->z && i >= ctx->min_keep) { + last_idx = i; + break; + } + } + + // Resize the output vector to keep only the tokens above the tail location + cur_p->size = last_idx; +} + +static struct llama_sampler * llama_sampler_tail_free_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_tail_free *) smpl->ctx; + return llama_sampler_init_tail_free(ctx->z, ctx->min_keep); +} + +static void llama_sampler_tail_free_free(struct llama_sampler * smpl) { + delete (llama_sampler_tail_free *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_tail_free_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "tail-free"; }, + /* .name = */ llama_sampler_tail_free_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_tail_free *) smpl->ctx; - llama_sampler_tail_free_impl(cur_p, ctx->z, ctx->min_keep); - }, + /* .apply = */ llama_sampler_tail_free_apply, /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_tail_free *) smpl->ctx; - return llama_sampler_init_tail_free(ctx->z, ctx->min_keep); - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_tail_free *) smpl->ctx; - }, + /* .clone = */ llama_sampler_tail_free_clone, + /* .free = */ llama_sampler_tail_free_free, }; struct llama_sampler * llama_sampler_init_tail_free(float z, size_t min_keep) { @@ -831,21 +757,86 @@ struct llama_sampler_typical { const size_t min_keep; }; +static const char * llama_sampler_typical_name(const struct llama_sampler * /*smpl*/) { + return "typical"; +} + +static void llama_sampler_typical_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + const auto * ctx = (llama_sampler_typical *) smpl->ctx; + + // Reference implementation: + // https://github.com/huggingface/transformers/compare/main...cimeister:typical-sampling:typical-pr + if (ctx->p >= 1.0f) { + return; + } + + // Compute the softmax of logits and calculate entropy + llama_sampler_softmax_impl(cur_p); + + float entropy = 0.0f; + for (size_t i = 0; i < cur_p->size; ++i) { + entropy += -cur_p->data[i].p * logf(cur_p->data[i].p); + } + + // Compute the absolute difference between negative log probability and entropy for each candidate + std::vector shifted_scores; + for (size_t i = 0; i < cur_p->size; ++i) { + float shifted_score = fabsf(-logf(cur_p->data[i].p) - entropy); + shifted_scores.push_back(shifted_score); + } + + // Sort tokens based on the shifted_scores and their corresponding indices + std::vector indices(cur_p->size); + std::iota(indices.begin(), indices.end(), 0); + + std::sort(indices.begin(), indices.end(), [&](size_t a, size_t b) { + return shifted_scores[a] < shifted_scores[b]; + }); + + // Compute the cumulative probabilities + float cum_sum = 0.0f; + size_t last_idx = indices.size(); + + for (size_t i = 0; i < indices.size(); ++i) { + size_t idx = indices[i]; + cum_sum += cur_p->data[idx].p; + + // Check if the running sum is greater than typical or if we have kept at least min_keep tokens + if (cum_sum > ctx->p && i >= ctx->min_keep - 1) { + last_idx = i + 1; + break; + } + } + + // Resize the output vector to keep only the locally typical tokens + std::vector cur_p_new; + for (size_t i = 0; i < last_idx; ++i) { + size_t idx = indices[i]; + cur_p_new.push_back(cur_p->data[idx]); + } + + // Replace the data in cur_p with the cur_p_new data + std::copy(cur_p_new.begin(), cur_p_new.end(), cur_p->data); + cur_p->size = cur_p_new.size(); + cur_p->sorted = false; +} + +static struct llama_sampler * llama_sampler_typical_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_typical *) smpl->ctx; + return llama_sampler_init_typical(ctx->p, ctx->min_keep); +} + +static void llama_sampler_typical_free(struct llama_sampler * smpl) { + delete (llama_sampler_typical *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_typical_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "typical"; }, + /* .name = */ llama_sampler_typical_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_typical *) smpl->ctx; - llama_sampler_typical_impl(cur_p, ctx->p, ctx->min_keep); - }, + /* .apply = */ llama_sampler_typical_apply, /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_typical *) smpl->ctx; - return llama_sampler_init_typical(ctx->p, ctx->min_keep); - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_typical *) smpl->ctx; - }, + /* .clone = */ llama_sampler_typical_clone, + /* .free = */ llama_sampler_typical_free, }; struct llama_sampler * llama_sampler_init_typical(float p, size_t min_keep) { @@ -864,21 +855,33 @@ struct llama_sampler_temp { const float temp; }; +static const char * llama_sampler_temp_name(const struct llama_sampler * /*smpl*/) { + return "temp"; +} + +static void llama_sampler_temp_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + const auto * ctx = (llama_sampler_temp *) smpl->ctx; + for (size_t i = 0; i < cur_p->size; ++i) { + cur_p->data[i].logit /= ctx->temp; + } +} + +static struct llama_sampler * llama_sampler_temp_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_temp *) smpl->ctx; + return llama_sampler_init_temp(ctx->temp); +} + +static void llama_sampler_temp_free(struct llama_sampler * smpl) { + delete (llama_sampler_temp *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_temp_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "temp"; }, + /* .name = */ llama_sampler_temp_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_temp *) smpl->ctx; - llama_sampler_temp_impl(cur_p, ctx->temp); - }, + /* .apply = */ llama_sampler_temp_apply, /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_temp *) smpl->ctx; - return llama_sampler_init_temp(ctx->temp); - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_temp *) smpl->ctx; - }, + /* .clone = */ llama_sampler_temp_clone, + /* .free = */ llama_sampler_temp_free, }; struct llama_sampler * llama_sampler_init_temp(float temp) { @@ -898,28 +901,100 @@ struct llama_sampler_temp_ext { const float exponent; }; -static struct llama_sampler_i llama_sampler_temp_ext_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "temp-ext"; }, - /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_temp_ext *) smpl->ctx; - if (ctx->delta > 0) { - const float temp_min = std::max(0.0f, ctx->temp - ctx->delta); - const float temp_max = ctx->temp + ctx->delta; +static const char * llama_sampler_temp_ext_name(const struct llama_sampler * /*smpl*/) { + return "temp-ext"; +} - llama_sampler_entropy_impl(cur_p, temp_min, temp_max, ctx->exponent); - } else { - llama_sampler_temp_impl(cur_p, ctx->temp); +static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + const auto * ctx = (llama_sampler_temp_ext *) smpl->ctx; + if (ctx->delta > 0) { + const float min_temp = std::max(0.0f, ctx->temp - ctx->delta); + const float max_temp = ctx->temp + ctx->delta; + float exponent_val = ctx->exponent; + + // no need to do anything if there is only one (or zero) candidates + if (cur_p->size <= 1) { + return; } - }, + + // Calculate maximum possible entropy + float max_entropy = -logf(1.0f / cur_p->size); + + llama_sampler_softmax_impl(cur_p); + + // Calculate entropy of the softmax probabilities + float entropy = 0.0f; + for (size_t i = 0; i < cur_p->size; ++i) { + float prob = cur_p->data[i].p; + if (prob > 0.0f) { // Ensure no log(0) + entropy -= prob * logf(prob); + } + } + + // Normalize the entropy (max_entropy cannot be 0 here because we checked cur_p->size != 1 above) + float normalized_entropy = entropy / max_entropy; + + // Map the normalized entropy to the desired temperature range using the power function + float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val); + + #ifdef DEBUG + LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp); + LLAMA_LOG_INFO("Entropy: %f\n", entropy); + LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy); + LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy); + LLAMA_LOG_INFO("Exponent: %f\n", exponent_val); + LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp); + #endif + + // Apply the dynamically calculated temperature scaling + for (size_t i = 0; i < cur_p->size; ++i) { + cur_p->data[i].logit /= dyn_temp; + } + + // Re-compute softmax probabilities after scaling logits with dynamic temperature + const double max_l_double = cur_p->data[0].logit; + + double cum_sum_double = 0.0; + for (size_t i = 0; i < cur_p->size; ++i) { + double p = exp(cur_p->data[i].logit - max_l_double); + cur_p->data[i].p = p; // Store the scaled probability + cum_sum_double += p; + } + + for (size_t i = 0; i < cur_p->size; ++i) { + cur_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities + } + + #ifdef DEBUG + // Print the updated top 25 probabilities after temperature scaling + LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n"); + for (size_t i = 0; i < 25 && i < cur_p->size; ++i) { + LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, cur_p->data[i].p * 100.0f); + } + #endif + } else { + for (size_t i = 0; i < cur_p->size; ++i) { + cur_p->data[i].logit /= ctx->temp; + } + } +} + +static struct llama_sampler * llama_sampler_temp_ext_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_temp_ext *) smpl->ctx; + return llama_sampler_init_temp_ext(ctx->temp, ctx->delta, ctx->exponent); +} + +static void llama_sampler_temp_ext_free(struct llama_sampler * smpl) { + delete (llama_sampler_temp_ext *) smpl->ctx; +} + +static struct llama_sampler_i llama_sampler_temp_ext_i = { + /* .name = */ llama_sampler_temp_ext_name, + /* .accept = */ nullptr, + /* .apply = */ llama_sampler_temp_ext_apply, /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_temp_ext *) smpl->ctx; - return llama_sampler_init_temp_ext(ctx->temp, ctx->delta, ctx->exponent); - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_temp_ext *) smpl->ctx; - }, + /* .clone = */ llama_sampler_temp_ext_clone, + /* .free = */ llama_sampler_temp_ext_free, }; struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, float exponent) { @@ -952,65 +1027,77 @@ struct llama_sampler_mirostat { std::vector probs; }; +static const char * llama_sampler_mirostat_name(const struct llama_sampler * /*smpl*/) { + return "mirostat"; +} + +static void llama_sampler_mirostat_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + auto * ctx = (llama_sampler_mirostat *) smpl->ctx; + + llama_sampler_softmax_impl(cur_p); + + // Estimate s_hat using the most probable m tokens + float s_hat = 0.0; + float sum_ti_bi = 0.0; + float sum_ti_sq = 0.0; + for (size_t i = 0; i < size_t(ctx->m - 1) && i < cur_p->size - 1; ++i) { + float t_i = logf(float(i + 2) / float(i + 1)); + float b_i = logf(cur_p->data[i].p / cur_p->data[i + 1].p); + sum_ti_bi += t_i * b_i; + sum_ti_sq += t_i * t_i; + } + s_hat = sum_ti_bi / sum_ti_sq; + + // Compute k from the estimated s_hat and target surprise value + float epsilon_hat = s_hat - 1; + float k = powf((epsilon_hat * powf(2, ctx->mu)) / (1 - powf(ctx->n_vocab, -epsilon_hat)), 1 / s_hat); + + llama_sampler_top_k_impl(cur_p, std::max(int(k), 1)); + llama_sampler_softmax_impl(cur_p); + + const int idx = llama_sample_dist(cur_p, ctx->rng, ctx->probs); + + cur_p->selected = idx; + + float observed_surprise = -log2f(cur_p->data[idx].p); + float e = observed_surprise - ctx->tau; + + // Update mu using the learning rate and error + ctx->mu = ctx->mu - ctx->eta * e; +} + +static struct llama_sampler * llama_sampler_mirostat_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_mirostat *) smpl->ctx; + auto * result = llama_sampler_init_mirostat(ctx->n_vocab, ctx->seed, ctx->tau, ctx->eta, ctx->m); + + // copy the state + { + auto * result_ctx = (llama_sampler_mirostat *) smpl->ctx; + + result_ctx->mu = ctx->mu; + result_ctx->rng = ctx->rng; + } + + return result; +} + +static void llama_sampler_mirostat_reset(struct llama_sampler * smpl) { + auto * ctx = (llama_sampler_mirostat *) smpl->ctx; + ctx->mu = 2.0f*ctx->tau; + ctx->rng = std::mt19937(ctx->seed); +} + +static void llama_sampler_mirostat_free(struct llama_sampler * smpl) { + delete (llama_sampler_mirostat *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_mirostat_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "mirostat"; }, + /* .name = */ llama_sampler_mirostat_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - auto * ctx = (llama_sampler_mirostat *) smpl->ctx; - - llama_sampler_softmax_impl(cur_p); - - // Estimate s_hat using the most probable m tokens - float s_hat = 0.0; - float sum_ti_bi = 0.0; - float sum_ti_sq = 0.0; - for (size_t i = 0; i < size_t(ctx->m - 1) && i < cur_p->size - 1; ++i) { - float t_i = logf(float(i + 2) / float(i + 1)); - float b_i = logf(cur_p->data[i].p / cur_p->data[i + 1].p); - sum_ti_bi += t_i * b_i; - sum_ti_sq += t_i * t_i; - } - s_hat = sum_ti_bi / sum_ti_sq; - - // Compute k from the estimated s_hat and target surprise value - float epsilon_hat = s_hat - 1; - float k = powf((epsilon_hat * powf(2, ctx->mu)) / (1 - powf(ctx->n_vocab, -epsilon_hat)), 1 / s_hat); - - llama_sampler_top_k_impl(cur_p, std::max(int(k), 1)); - llama_sampler_softmax_impl(cur_p); - - const int idx = llama_sample_dist(cur_p, ctx->rng, ctx->probs); - - cur_p->selected = idx; - - float observed_surprise = -log2f(cur_p->data[idx].p); - float e = observed_surprise - ctx->tau; - - // Update mu using the learning rate and error - ctx->mu = ctx->mu - ctx->eta * e; - }, - /* .reset = */ [](struct llama_sampler * smpl) { - auto * ctx = (llama_sampler_mirostat *) smpl->ctx; - ctx->mu = 2.0f*ctx->tau; - ctx->rng = std::mt19937(ctx->seed); - }, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_mirostat *) smpl->ctx; - auto * result = llama_sampler_init_mirostat(ctx->n_vocab, ctx->seed, ctx->tau, ctx->eta, ctx->m); - - // copy the state - { - auto * result_ctx = (llama_sampler_mirostat *) smpl->ctx; - - result_ctx->mu = ctx->mu; - result_ctx->rng = ctx->rng; - } - - return result; - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_mirostat *) smpl->ctx; - }, + /* .apply = */ llama_sampler_mirostat_apply, + /* .reset = */ llama_sampler_mirostat_reset, + /* .clone = */ llama_sampler_mirostat_clone, + /* .free = */ llama_sampler_mirostat_free, }; struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t seed, float tau, float eta, int32_t m) { @@ -1044,59 +1131,71 @@ struct llama_sampler_mirostat_v2 { std::vector probs; }; +static const char * llama_sampler_mirostat_v2_name(const struct llama_sampler * /*smpl*/) { + return "mirostat-v2"; +} + +static void llama_sampler_mirostat_v2_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + auto * ctx = (llama_sampler_mirostat_v2 *) smpl->ctx; + + llama_sampler_softmax_impl(cur_p); + + // Truncate the words with surprise values greater than mu + cur_p->size = std::distance(cur_p->data, std::find_if(cur_p->data, cur_p->data + cur_p->size, [&](const llama_token_data & candidate) { + return -log2f(candidate.p) > ctx->mu; + })); + + if (cur_p->size == 0) { + cur_p->size = 1; + } + + // Normalize the probabilities of the remaining words + llama_sampler_softmax_impl(cur_p); + + const int idx = llama_sample_dist(cur_p, ctx->rng, ctx->probs); + + cur_p->selected = idx; + + float observed_surprise = -log2f(cur_p->data[idx].p); + float e = observed_surprise - ctx->tau; + + // Update mu using the learning rate and error + ctx->mu = ctx->mu - ctx->eta * e; +} + +static void llama_sampler_mirostat_v2_reset(struct llama_sampler * smpl) { + auto * ctx = (llama_sampler_mirostat_v2 *) smpl->ctx; + ctx->mu = 2.0f*ctx->tau; + ctx->rng = std::mt19937(ctx->seed); +} + +static struct llama_sampler * llama_sampler_mirostat_v2_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_mirostat_v2 *) smpl->ctx; + + auto * result = llama_sampler_init_mirostat_v2(ctx->seed, ctx->tau, ctx->eta); + + // copy the state + { + auto * result_ctx = (llama_sampler_mirostat_v2 *) result->ctx; + + result_ctx->mu = ctx->mu; + result_ctx->rng = ctx->rng; + } + + return result; +} + +static void llama_sampler_mirostat_v2_free(struct llama_sampler * smpl) { + delete (llama_sampler_mirostat_v2 *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_mirostat_v2_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "mirostat-v2"; }, + /* .name = */ llama_sampler_mirostat_v2_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - auto * ctx = (llama_sampler_mirostat_v2 *) smpl->ctx; - - llama_sampler_softmax_impl(cur_p); - - // Truncate the words with surprise values greater than mu - cur_p->size = std::distance(cur_p->data, std::find_if(cur_p->data, cur_p->data + cur_p->size, [&](const llama_token_data & candidate) { - return -log2f(candidate.p) > ctx->mu; - })); - - if (cur_p->size == 0) { - cur_p->size = 1; - } - - // Normalize the probabilities of the remaining words - llama_sampler_softmax_impl(cur_p); - - const int idx = llama_sample_dist(cur_p, ctx->rng, ctx->probs); - - cur_p->selected = idx; - - float observed_surprise = -log2f(cur_p->data[idx].p); - float e = observed_surprise - ctx->tau; - - // Update mu using the learning rate and error - ctx->mu = ctx->mu - ctx->eta * e; - }, - /* .reset = */ [](struct llama_sampler * smpl) { - auto * ctx = (llama_sampler_mirostat_v2 *) smpl->ctx; - ctx->mu = 2.0f*ctx->tau; - ctx->rng = std::mt19937(ctx->seed); - }, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_mirostat_v2 *) smpl->ctx; - - auto * result = llama_sampler_init_mirostat_v2(ctx->seed, ctx->tau, ctx->eta); - - // copy the state - { - auto * result_ctx = (llama_sampler_mirostat_v2 *) result->ctx; - - result_ctx->mu = ctx->mu; - result_ctx->rng = ctx->rng; - } - - return result; - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_mirostat_v2 *) smpl->ctx; - }, + /* .apply = */ llama_sampler_mirostat_v2_apply, + /* .reset = */ llama_sampler_mirostat_v2_reset, + /* .clone = */ llama_sampler_mirostat_v2_clone, + /* .free = */ llama_sampler_mirostat_v2_free, }; struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau, float eta) { @@ -1124,59 +1223,73 @@ struct llama_sampler_grammar { struct llama_grammar * grammar; }; -static struct llama_sampler_i llama_sampler_grammar_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "grammar"; }, - /* .accept = */ [](struct llama_sampler * smpl, llama_token token) { - const auto * ctx = (llama_sampler_grammar *) smpl->ctx; - if (ctx->grammar) { - llama_grammar_accept_impl(*ctx->grammar, token); - } - }, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - const auto * ctx = (llama_sampler_grammar *) smpl->ctx; - if (ctx->grammar) { - llama_sampler_grammar_impl(cur_p, *ctx->grammar); - } - }, - /* .reset = */ [](struct llama_sampler * smpl) { - auto * ctx = (llama_sampler_grammar *) smpl->ctx; - if (!ctx->grammar) { - return; - } +static const char * llama_sampler_grammar_name(const struct llama_sampler * /*smpl*/) { + return "grammar"; +} - auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str()); +static void llama_sampler_grammar_accept_impl(struct llama_sampler * smpl, llama_token token) { + auto * ctx = (llama_sampler_grammar *) smpl->ctx; + if (ctx->grammar) { + llama_grammar_accept_impl(*ctx->grammar, token); + } +} +static void llama_sampler_grammar_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + auto * ctx = (llama_sampler_grammar *) smpl->ctx; + if (ctx->grammar) { + llama_grammar_apply_impl(*ctx->grammar, cur_p); + } +} + +static void llama_sampler_grammar_reset(struct llama_sampler * smpl) { + auto * ctx = (llama_sampler_grammar *) smpl->ctx; + if (!ctx->grammar) { + return; + } + + auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str()); + + llama_grammar_free_impl(ctx->grammar); + ctx->grammar = grammar_new; +} + +static struct llama_sampler * llama_sampler_grammar_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_grammar *) smpl->ctx; + + auto * result = llama_sampler_init_grammar_impl(*ctx->vocab, nullptr, nullptr); + + // copy the state + { + auto * result_ctx = (llama_sampler_grammar *) result->ctx; + + if (ctx->grammar) { + result_ctx->grammar_str = ctx->grammar_str; + result_ctx->grammar_root = ctx->grammar_root; + + result_ctx->grammar = llama_grammar_clone_impl(*ctx->grammar); + } + } + + return result; +} + +static void llama_sampler_grammar_free(struct llama_sampler * smpl) { + const auto * ctx = (llama_sampler_grammar *) smpl->ctx; + + if (ctx->grammar) { llama_grammar_free_impl(ctx->grammar); - ctx->grammar = grammar_new; - }, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_grammar *) smpl->ctx; + } - auto * result = llama_sampler_init_grammar_impl(*ctx->vocab, nullptr, nullptr); + delete ctx; +} - // copy the state - { - auto * result_ctx = (llama_sampler_grammar *) result->ctx; - - if (ctx->grammar) { - result_ctx->grammar_str = ctx->grammar_str; - result_ctx->grammar_root = ctx->grammar_root; - - result_ctx->grammar = llama_grammar_clone_impl(*ctx->grammar); - } - } - - return result; - }, - /* .free = */ [](struct llama_sampler * smpl) { - const auto * ctx = (llama_sampler_grammar *) smpl->ctx; - - if (ctx->grammar) { - llama_grammar_free_impl(ctx->grammar); - } - - delete ctx; - }, +static struct llama_sampler_i llama_sampler_grammar_i = { + /* .name = */ llama_sampler_grammar_name, + /* .accept = */ llama_sampler_grammar_accept_impl, + /* .apply = */ llama_sampler_grammar_apply, + /* .reset = */ llama_sampler_grammar_reset, + /* .clone = */ llama_sampler_grammar_clone, + /* .free = */ llama_sampler_grammar_free, }; struct llama_sampler * llama_sampler_init_grammar_impl(const struct llama_vocab & vocab, const char * grammar_str, const char * grammar_root) { @@ -1222,106 +1335,144 @@ struct llama_sampler_penalties { ring_buffer prev; }; +static const char * llama_sampler_penalties_name(const struct llama_sampler * /*smpl*/) { + return "penalties"; +} + +static void llama_sampler_penalties_accept(struct llama_sampler * smpl, llama_token token) { + auto * ctx = (llama_sampler_penalties *) smpl->ctx; + if (ctx->penalty_last_n == 0) { + return; + } + + ctx->prev.push_back(token); +} + +static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + auto * ctx = (llama_sampler_penalties *) smpl->ctx; + + if (ctx->ignore_eos) { + assert(ctx->special_eos_id >= 0); + + // optimistically check if the candidates are not yet sorted/shuffled/truncated + if (cur_p->size > (size_t) ctx->special_eos_id && cur_p->data[ctx->special_eos_id].id == ctx->special_eos_id) { + cur_p->data[ctx->special_eos_id].logit = -INFINITY; + } else { + // else, search for the special EOS token + for (size_t i = 0; i < cur_p->size; ++i) { + if (cur_p->data[i].id == ctx->special_eos_id) { + cur_p->data[i].logit = -INFINITY; + break; + } + } + } + } + + if ((ctx->penalty_last_n == 0) || + (ctx->penalty_repeat == 1.0f && ctx->penalty_freq == 0.0f && ctx->penalty_present == 0.0f)) { + return; + } + + bool nl_found = false; + size_t nl_idx = 0; + float nl_logit = -INFINITY; + if (!ctx->penalize_nl) { + assert(ctx->linefeed_id >= 0); + + // optimistically check if the candidates are not yet sorted/shuffled/truncated + if (cur_p->size > (size_t) ctx->linefeed_id && cur_p->data[ctx->linefeed_id].id == ctx->linefeed_id) { + nl_found = true; + nl_idx = ctx->linefeed_id; + nl_logit = cur_p->data[ctx->linefeed_id].logit; + } else { + // else, search for the linefeed token + for (size_t i = 0; i < cur_p->size; ++i) { + if (cur_p->data[i].id == ctx->linefeed_id) { + nl_found = true; + nl_idx = i; + nl_logit = cur_p->data[i].logit; + break; + } + } + } + } + + // Create a frequency map to count occurrences of each token in last_tokens + // TODO: optimize this by maintaining the token count in the sampler context + using llama_token_cnt = std::unordered_map; + llama_token_cnt token_count; + + for (int i = 0; i < std::min(ctx->penalty_last_n, ctx->prev.size()); ++i) { + token_count[ctx->prev.rat(i)]++; + } + + // Apply frequency and presence penalties to the cur_p + for (size_t i = 0; i < cur_p->size; ++i) { + const auto token_iter = token_count.find(cur_p->data[i].id); + if (token_iter == token_count.end()) { + continue; + } + + const int count = token_iter->second; + + // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong. + // This is common fix for this problem, which is to multiply by the penalty instead of dividing. + if (cur_p->data[i].logit <= 0) { + cur_p->data[i].logit *= ctx->penalty_repeat; + } else { + cur_p->data[i].logit /= ctx->penalty_repeat; + } + + cur_p->data[i].logit -= float(count) * ctx->penalty_freq + float(count > 0) * ctx->penalty_present; + } + + cur_p->sorted = false; + + if (!ctx->penalize_nl && nl_found) { + // restore the logit of the newline token if it was penalized + cur_p->data[nl_idx].logit = nl_logit; + } +} + +static void llama_sampler_penalties_reset(struct llama_sampler * smpl) { + auto * ctx = (llama_sampler_penalties *) smpl->ctx; + ctx->prev.clear(); +} + +static struct llama_sampler * llama_sampler_penalties_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_penalties *) smpl->ctx; + auto * result = llama_sampler_init_penalties( + ctx->n_vocab, + ctx->special_eos_id, + ctx->linefeed_id, + ctx->penalty_last_n, + ctx->penalty_repeat, + ctx->penalty_freq, + ctx->penalty_present, + ctx->penalize_nl, + ctx->ignore_eos); + + // copy the state + { + auto * result_ctx = (llama_sampler_penalties *) result->ctx; + + result_ctx->prev = ctx->prev; + } + + return result; +} + +static void llama_sampler_penalties_free(struct llama_sampler * smpl) { + delete (llama_sampler_penalties *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_penalties_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "penalties"; }, - /* .accept = */ [](struct llama_sampler * smpl, llama_token token) { - auto * ctx = (llama_sampler_penalties *) smpl->ctx; - if (ctx->prev.size()) { - ctx->prev.push_back(token); - } - }, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - auto * ctx = (llama_sampler_penalties *) smpl->ctx; - - if (ctx->ignore_eos) { - assert(ctx->special_eos_id >= 0); - - // optimistically check if the candidates are not yet sorted/shuffled/truncated - if (cur_p->size > (size_t) ctx->special_eos_id && cur_p->data[ctx->special_eos_id].id == ctx->special_eos_id) { - cur_p->data[ctx->special_eos_id].logit = -INFINITY; - } else { - // else, search for the special EOS token - for (size_t i = 0; i < cur_p->size; ++i) { - if (cur_p->data[i].id == ctx->special_eos_id) { - cur_p->data[i].logit = -INFINITY; - break; - } - } - } - } - - if ((ctx->penalty_last_n == 0) || - (ctx->penalty_repeat == 1.0f && ctx->penalty_freq == 0.0f && ctx->penalty_present == 0.0f)) { - return; - } - - bool nl_found = false; - size_t nl_idx = 0; - float nl_logit = -INFINITY; - if (!ctx->penalize_nl) { - assert(ctx->linefeed_id >= 0); - - // optimistically check if the candidates are not yet sorted/shuffled/truncated - if (cur_p->size > (size_t) ctx->linefeed_id && cur_p->data[ctx->linefeed_id].id == ctx->linefeed_id) { - nl_found = true; - nl_idx = ctx->linefeed_id; - nl_logit = cur_p->data[ctx->linefeed_id].logit; - } else { - // else, search for the linefeed token - for (size_t i = 0; i < cur_p->size; ++i) { - if (cur_p->data[i].id == ctx->linefeed_id) { - nl_found = true; - nl_idx = i; - nl_logit = cur_p->data[i].logit; - break; - } - } - } - } - - // Create a frequency map to count occurrences of each token in last_tokens - // TODO: optimize this by maintaining the token count in the sampler context - llama_token_cnt token_count; - for (int i = 0; i < std::min(ctx->penalty_last_n, ctx->prev.size()); ++i) { - token_count[ctx->prev.rat(i)]++; - } - - llama_sampler_penalties_impl(cur_p, token_count, ctx->penalty_repeat, ctx->penalty_freq, ctx->penalty_present); - - if (!ctx->penalize_nl && nl_found) { - // restore the logit of the newline token if it was penalized - cur_p->data[nl_idx].logit = nl_logit; - } - }, - /* .reset = */ [](struct llama_sampler * smpl) { - auto * ctx = (llama_sampler_penalties *) smpl->ctx; - ctx->prev.clear(); - }, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_penalties *) smpl->ctx; - auto * result = llama_sampler_init_penalties( - ctx->n_vocab, - ctx->special_eos_id, - ctx->linefeed_id, - ctx->penalty_last_n, - ctx->penalty_repeat, - ctx->penalty_freq, - ctx->penalty_present, - ctx->penalize_nl, - ctx->ignore_eos); - - // copy the state - { - auto * result_ctx = (llama_sampler_penalties *) result->ctx; - - result_ctx->prev = ctx->prev; - } - - return result; - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_penalties *) smpl->ctx; - }, + /* .name = */ llama_sampler_penalties_name, + /* .accept = */ llama_sampler_penalties_accept, + /* .apply = */ llama_sampler_penalties_apply, + /* .reset = */ llama_sampler_penalties_reset, + /* .clone = */ llama_sampler_penalties_clone, + /* .free = */ llama_sampler_penalties_free, }; struct llama_sampler * llama_sampler_init_penalties( @@ -1335,11 +1486,11 @@ struct llama_sampler * llama_sampler_init_penalties( bool penalize_nl, bool ignore_eos) { if (linefeed_id == LLAMA_TOKEN_NULL) { - penalize_nl = false; + penalize_nl = true; } if (special_eos_id == LLAMA_TOKEN_NULL) { - ignore_eos = true; + ignore_eos = false; } return new llama_sampler { @@ -1369,41 +1520,50 @@ struct llama_sampler_logit_bias { std::vector to_search; }; +static const char * llama_sampler_logit_bias_name(const struct llama_sampler * /*smpl*/) { + return "logit-bias"; +} + +static void llama_sampler_logit_bias_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + auto * ctx = (llama_sampler_logit_bias *) smpl->ctx; + + ctx->to_search.clear(); + + // update the candidates that have not been shuffled in the vocabulary (i.e. idx == id) + for (const auto & lb : ctx->logit_bias) { + if (lb.token >= 0 && cur_p->size > (size_t) lb.token && cur_p->data[lb.token].id == lb.token) { + cur_p->data[lb.token].logit += lb.bias; + } else { + ctx->to_search.push_back(lb); + } + } + + // search for the remaining candidates that were not found in the previous step + for (size_t i = 0; i < cur_p->size; ++i) { + for (const auto & lb : ctx->to_search) { + if (cur_p->data[i].id == lb.token) { + cur_p->data[i].logit += lb.bias; + break; + } + } + } +} +static struct llama_sampler * llama_sampler_logit_bias_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_logit_bias *) smpl->ctx; + return llama_sampler_init_logit_bias(ctx->n_vocab, ctx->logit_bias.size(), ctx->logit_bias.data()); +} + +static void llama_sampler_logit_bias_free(struct llama_sampler * smpl) { + delete (llama_sampler_logit_bias *) smpl->ctx; +} + static struct llama_sampler_i llama_sampler_logit_bias_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "logit-bias"; }, + /* .name = */ llama_sampler_logit_bias_name, /* .accept = */ nullptr, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - auto * ctx = (llama_sampler_logit_bias *) smpl->ctx; - - ctx->to_search.clear(); - - // update the candidates that have not been shuffled in the vocabulary (i.e. idx == id) - for (const auto & lb : ctx->logit_bias) { - if (lb.token >= 0 && cur_p->size > (size_t) lb.token && cur_p->data[lb.token].id == lb.token) { - cur_p->data[lb.token].logit += lb.bias; - } else { - ctx->to_search.push_back(lb); - } - } - - // search for the remaining candidates that were not found in the previous step - for (size_t i = 0; i < cur_p->size; ++i) { - for (const auto & lb : ctx->to_search) { - if (cur_p->data[i].id == lb.token) { - cur_p->data[i].logit += lb.bias; - break; - } - } - } - }, + /* .apply = */ llama_sampler_logit_bias_apply, /* .reset = */ nullptr, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * ctx = (const llama_sampler_logit_bias *) smpl->ctx; - return llama_sampler_init_logit_bias(ctx->n_vocab, ctx->logit_bias.size(), ctx->logit_bias.data()); - }, - /* .free = */ [](struct llama_sampler * smpl) { - delete (llama_sampler_logit_bias *) smpl->ctx; - }, + /* .clone = */ llama_sampler_logit_bias_clone, + /* .free = */ llama_sampler_logit_bias_free, }; struct llama_sampler * llama_sampler_init_logit_bias( diff --git a/src/llama-sampling.h b/src/llama-sampling.h index 137c0025c..d90b14713 100644 --- a/src/llama-sampling.h +++ b/src/llama-sampling.h @@ -23,16 +23,6 @@ struct llama_sampler_chain { mutable int32_t n_sample; }; -using llama_token_cnt = std::unordered_map; - -// TODO: tmp exposed until test-sampling is fixed -void llama_sampler_penalties_impl( - llama_token_data_array * cur_p, - const llama_token_cnt & token_count, - float penalty_repeat, - float penalty_freq, - float penalty_present); - struct llama_sampler * llama_sampler_init_grammar_impl( const struct llama_vocab & vocab, const char * grammar_str, diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index cc4882d37..37400c179 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -148,15 +148,17 @@ static void test_penalties( cur.emplace_back(llama_token_data{token_id, logit, 0.0f}); } - llama_token_cnt token_count; + llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false }; + + auto * sampler = llama_sampler_init_penalties(n_vocab, LLAMA_TOKEN_NULL, LLAMA_TOKEN_NULL, last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence, false, false); + for (size_t i = 0; i < last_tokens.size(); i++) { - token_count[last_tokens[i]]++; + llama_sampler_accept(sampler, last_tokens[i]); } - llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false }; APPLY(llama_sampler_init_softmax(), &cur_p); DUMP(&cur_p); - llama_sampler_penalties_impl(&cur_p, token_count, repeat_penalty, alpha_frequency, alpha_presence); // TODO: avoid + APPLY(sampler, &cur_p); APPLY(llama_sampler_init_softmax(), &cur_p); DUMP(&cur_p); From a249843d89bd6373772eede26d7f2aa708b26600 Mon Sep 17 00:00:00 2001 From: slaren Date: Sun, 8 Sep 2024 16:44:42 +0200 Subject: [PATCH 16/32] common : restore --n-gpu-layers (#9371) --- common/common.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 9228eafca..a9c7813f7 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1656,7 +1656,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example } )); add_opt(llama_arg( - {"-ngl", "--gpu-layers"}, "N", + {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N", "number of layers to store in VRAM", [](gpt_params & params, int value) { params.n_gpu_layers = value; @@ -1667,7 +1667,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example } ).set_env("LLAMA_ARG_N_GPU_LAYERS")); add_opt(llama_arg( - {"-ngld", "--gpu-layers-draft"}, "N", + {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N", "number of layers to store in VRAM for the draft model", [](gpt_params & params, int value) { params.n_gpu_layers_draft = value; From 3f7ccfd649abc83d059b5462221ac14de4ede6b7 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Sun, 8 Sep 2024 18:08:55 +0200 Subject: [PATCH 17/32] common : bring back missing args, add env var duplication check (#9375) * common : bring back missing args * move duplication check to test-arg-parser * add check for duplicated env var * correct default values --- common/common.cpp | 109 +++++++++++++++++++++++++------------ common/common.h | 1 - examples/infill/infill.cpp | 5 -- tests/test-arg-parser.cpp | 25 ++++++++- 4 files changed, 99 insertions(+), 41 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index a9c7813f7..916b1731e 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -673,17 +673,8 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example * - if LLAMA_EXAMPLE_* is set (other than COMMON), we only show the option in the corresponding example * - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example */ - std::unordered_set seen_args; auto add_opt = [&](llama_arg arg) { if (arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) { - // make sure there is no argument duplications - for (const auto & a : arg.args) { - if (seen_args.find(a) == seen_args.end()) { - seen_args.insert(a); - } else { - throw std::runtime_error(format("found duplicated argument in source code: %s", a)); - } - } options.push_back(std::move(arg)); } }; @@ -790,8 +781,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example add_opt(llama_arg( {"-C", "--cpu-mask"}, "M", "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")", - [](gpt_params & params, const std::string & value) { - std::string mask = value; + [](gpt_params & params, const std::string & mask) { params.cpuparams.mask_valid = true; if (!parse_cpu_mask(mask, params.cpuparams.cpumask)) { throw std::invalid_argument("invalid cpumask"); @@ -801,8 +791,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example add_opt(llama_arg( {"-Cr", "--cpu-range"}, "lo-hi", "range of CPUs for affinity. Complements --cpu-mask", - [](gpt_params & params, const std::string & value) { - std::string range = value; + [](gpt_params & params, const std::string & range) { params.cpuparams.mask_valid = true; if (!parse_cpu_range(range, params.cpuparams.cpumask)) { throw std::invalid_argument("invalid range"); @@ -816,6 +805,16 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.cpuparams.strict_cpu = std::stoul(value); } )); + add_opt(llama_arg( + {"--prio"}, "N", + format("set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority), + [](gpt_params & params, int prio) { + if (prio < 0 || prio > 3) { + throw std::invalid_argument("invalid value"); + } + params.cpuparams.priority = (enum ggml_sched_priority) prio; + } + )); add_opt(llama_arg( {"--poll"}, "<0...100>", format("use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll), @@ -826,8 +825,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example add_opt(llama_arg( {"-Cb", "--cpu-mask-batch"}, "M", "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)", - [](gpt_params & params, const std::string & value) { - std::string mask = value; + [](gpt_params & params, const std::string & mask) { params.cpuparams_batch.mask_valid = true; if (!parse_cpu_mask(mask, params.cpuparams_batch.cpumask)) { throw std::invalid_argument("invalid cpumask"); @@ -837,8 +835,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example add_opt(llama_arg( {"-Crb", "--cpu-range-batch"}, "lo-hi", "ranges of CPUs for affinity. Complements --cpu-mask-batch", - [](gpt_params & params, const std::string & value) { - std::string range = value; + [](gpt_params & params, const std::string & range) { params.cpuparams_batch.mask_valid = true; if (!parse_cpu_range(range, params.cpuparams_batch.cpumask)) { throw std::invalid_argument("invalid range"); @@ -852,6 +849,16 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.cpuparams_batch.strict_cpu = value; } )); + add_opt(llama_arg( + {"--prio-batch"}, "N", + format("set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams_batch.priority), + [](gpt_params & params, int prio) { + if (prio < 0 || prio > 3) { + throw std::invalid_argument("invalid value"); + } + params.cpuparams_batch.priority = (enum ggml_sched_priority) prio; + } + )); add_opt(llama_arg( {"--poll-batch"}, "<0|1>", "use polling to wait for work (default: same as --poll)", @@ -862,8 +869,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example add_opt(llama_arg( {"-Cd", "--cpu-mask-draft"}, "M", "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)", - [](gpt_params & params, const std::string & value) { - std::string mask = value; + [](gpt_params & params, const std::string & mask) { params.draft_cpuparams.mask_valid = true; if (!parse_cpu_mask(mask, params.draft_cpuparams.cpumask)) { throw std::invalid_argument("invalid cpumask"); @@ -873,8 +879,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example add_opt(llama_arg( {"-Crd", "--cpu-range-draft"}, "lo-hi", "Ranges of CPUs for affinity. Complements --cpu-mask-draft", - [](gpt_params & params, const std::string & value) { - std::string range = value; + [](gpt_params & params, const std::string & range) { params.draft_cpuparams.mask_valid = true; if (!parse_cpu_range(range, params.draft_cpuparams.cpumask)) { throw std::invalid_argument("invalid range"); @@ -888,6 +893,16 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.draft_cpuparams.strict_cpu = value; } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--prio-draft"}, "N", + format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams.priority), + [](gpt_params & params, int prio) { + if (prio < 0 || prio > 3) { + throw std::invalid_argument("invalid value"); + } + params.draft_cpuparams.priority = (enum ggml_sched_priority) prio; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); add_opt(llama_arg( {"--poll-draft"}, "<0|1>", "Use polling to wait for draft model work (default: same as --poll])", @@ -895,11 +910,20 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.draft_cpuparams.poll = value; } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-Cbd", "--cpu-mask-batch-draft"}, "M", + "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)", + [](gpt_params & params, const std::string & mask) { + params.draft_cpuparams_batch.mask_valid = true; + if (!parse_cpu_mask(mask, params.draft_cpuparams_batch.cpumask)) { + throw std::invalid_argument("invalid cpumask"); + } + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); add_opt(llama_arg( {"-Crbd", "--cpu-range-batch-draft"}, "lo-hi", "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)", - [](gpt_params & params, const std::string & value) { - std::string range = value; + [](gpt_params & params, const std::string & range) { params.draft_cpuparams_batch.mask_valid = true; if (!parse_cpu_range(range, params.draft_cpuparams_batch.cpumask)) { throw std::invalid_argument("invalid cpumask"); @@ -913,6 +937,16 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.draft_cpuparams_batch.strict_cpu = value; } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--prio-batch-draft"}, "N", + format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams_batch.priority), + [](gpt_params & params, int prio) { + if (prio < 0 || prio > 3) { + throw std::invalid_argument("invalid value"); + } + params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) prio; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); add_opt(llama_arg( {"--poll-batch-draft"}, "<0|1>", "Use polling to wait for draft model work (default: --poll-draft)", @@ -1124,21 +1158,21 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example [](gpt_params & params) { params.interactive = true; } - ).set_examples({LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(llama_arg( {"-if", "--interactive-first"}, format("run in interactive mode and wait for input right away (default: %s)", params.interactive_first ? "true" : "false"), [](gpt_params & params) { params.interactive_first = true; } - ).set_examples({LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(llama_arg( {"-mli", "--multiline-input"}, "allows you to write or paste multiple lines without ending each in '\\'", [](gpt_params & params) { params.multiline_input = true; } - ).set_examples({LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(llama_arg( {"--in-prefix-bos"}, "prefix BOS to user inputs, preceding the `--in-prefix` string", @@ -1146,7 +1180,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.input_prefix_bos = true; params.enable_chat_template = false; } - ).set_examples({LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(llama_arg( {"--in-prefix"}, "STRING", "string to prefix user inputs with (default: empty)", @@ -1154,7 +1188,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.input_prefix = value; params.enable_chat_template = false; } - ).set_examples({LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(llama_arg( {"--in-suffix"}, "STRING", "string to suffix after user inputs with (default: empty)", @@ -1162,7 +1196,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.input_suffix = value; params.enable_chat_template = false; } - ).set_examples({LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(llama_arg( {"--no-warmup"}, "skip warming up the model with an empty run", @@ -1499,7 +1533,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example } )); add_opt(llama_arg( - {"--all-logits"}, + {"--perplexity", "--all-logits"}, format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"), [](gpt_params & params) { params.logits_all = true; @@ -1554,6 +1588,13 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example params.kl_divergence = true; } ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--save-all-logits", "--kl-divergence-base"}, "FNAME", + "set logits file", + [](gpt_params & params, const std::string & value) { + params.logits_file = value; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); add_opt(llama_arg( {"--ppl-stride"}, "N", format("stride for perplexity calculation (default: %d)", params.ppl_stride), @@ -1802,7 +1843,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example [](gpt_params & params, const std::string & value) { params.model_alias = value; } - ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL")); + ).set_examples({LLAMA_EXAMPLE_SERVER})); add_opt(llama_arg( {"-m", "--model"}, "FNAME", ex == LLAMA_EXAMPLE_EXPORT_LORA @@ -1890,7 +1931,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example } ).set_examples({LLAMA_EXAMPLE_PASSKEY})); add_opt(llama_arg( - {"-o", "--output"}, "FNAME", + {"-o", "--output", "--output-file"}, "FNAME", format("output file (default: '%s')", ex == LLAMA_EXAMPLE_EXPORT_LORA ? params.lora_outfile.c_str() @@ -1932,7 +1973,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example } ).set_examples({LLAMA_EXAMPLE_IMATRIX})); add_opt(llama_arg( - {"--chunk"}, "N", + {"--chunk", "--from-chunk"}, "N", format("start processing the input from chunk N (default: %d)", params.i_chunk), [](gpt_params & params, int value) { params.i_chunk = value; @@ -2057,7 +2098,7 @@ std::vector gpt_params_parser_init(gpt_params & params, llama_example } ).set_examples({LLAMA_EXAMPLE_SERVER})); add_opt(llama_arg( - {"--timeout"}, "N", + {"-to", "--timeout"}, "N", format("server read/write timeout in seconds (default: %d)", params.timeout_read), [](gpt_params & params, int value) { params.timeout_read = value; diff --git a/common/common.h b/common/common.h index d7c08f20a..5b945ac02 100644 --- a/common/common.h +++ b/common/common.h @@ -211,7 +211,6 @@ struct gpt_params { bool use_mlock = false; // use mlock to keep model in memory bool verbose_prompt = false; // print prompt tokens before generation bool display_prompt = true; // print prompt before generation - bool infill = false; // use infill mode bool dump_kv_cache = false; // dump the KV cache contents for debugging purposes bool no_kv_offload = false; // disable KV offloading bool warmup = true; // warmup run diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index d06071377..87abb761f 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -306,11 +306,6 @@ int main(int argc, char ** argv) { LOG_TEE("\n\n"); LOG_TEE("\n##### Infill mode #####\n\n"); - if (params.infill) { - printf("\n************\n"); - printf("no need to specify '--infill', always running infill\n"); - printf("************\n\n"); - } if (params.interactive) { const char *control_message; if (params.multiline_input) { diff --git a/tests/test-arg-parser.cpp b/tests/test-arg-parser.cpp index 8852bfc7e..9ad91acc0 100644 --- a/tests/test-arg-parser.cpp +++ b/tests/test-arg-parser.cpp @@ -1,6 +1,7 @@ #include #include #include +#include #undef NDEBUG #include @@ -13,7 +14,29 @@ int main(void) { printf("test-arg-parser: make sure there is no duplicated arguments in any examples\n\n"); for (int ex = 0; ex < LLAMA_EXAMPLE_COUNT; ex++) { try { - gpt_params_parser_init(params, (enum llama_example)ex); + auto options = gpt_params_parser_init(params, (enum llama_example)ex); + std::unordered_set seen_args; + std::unordered_set seen_env_vars; + for (const auto & opt : options) { + // check for args duplications + for (const auto & arg : opt.args) { + if (seen_args.find(arg) == seen_args.end()) { + seen_args.insert(arg); + } else { + fprintf(stderr, "test-arg-parser: found different handlers for the same argument: %s", arg); + exit(1); + } + } + // check for env var duplications + if (opt.env) { + if (seen_env_vars.find(opt.env) == seen_env_vars.end()) { + seen_env_vars.insert(opt.env); + } else { + fprintf(stderr, "test-arg-parser: found different handlers for the same env var: %s", opt.env); + exit(1); + } + } + } } catch (std::exception & e) { printf("%s\n", e.what()); assert(false); From e079bffb6678e1b5ded21c719600bedd7175e726 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 8 Sep 2024 22:01:02 +0300 Subject: [PATCH 18/32] cuda : fix FA Q src index (1 -> 0) (#9374) --- ggml/src/ggml-cuda/fattn.cu | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu index f87f33b3e..f28a19d40 100644 --- a/ggml/src/ggml-cuda/fattn.cu +++ b/ggml/src/ggml-cuda/fattn.cu @@ -152,7 +152,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g } \ static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - ggml_tensor * Q = dst->src[1]; + ggml_tensor * Q = dst->src[0]; ggml_tensor * K = dst->src[1]; ggml_tensor * V = dst->src[2]; @@ -227,7 +227,7 @@ static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, gg } \ static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - ggml_tensor * Q = dst->src[1]; + ggml_tensor * Q = dst->src[0]; ggml_tensor * K = dst->src[1]; ggml_tensor * V = dst->src[2]; From daa9623ab051a8162ae750b150b9522571b55f21 Mon Sep 17 00:00:00 2001 From: Markus Tavenrath Date: Sun, 8 Sep 2024 21:43:48 +0200 Subject: [PATCH 19/32] Overlap cmdbuffer creation and cmdbuffer execution in Vulkan backend by submitting smaller cmdbuffers early. (#9118) * Overlap cmdbuffer creation and cmdbuffer execution in Vulkan backend by submitting smaller cmdbuffers early. * fix compile issues * Fix issues where the last submit wasn't executed or handled properly. * remove trailing whitespace * Repair GGML_VULKAN_CHECK_RESULTS * Increase submit counter only if actual work has been submitted and increase submit count to 100. * Fix some nodes are not checked with GGML_VULKAN_CHECK_RESULTS enabled. --- ggml/src/ggml-vulkan.cpp | 132 ++++++++++++++++++++++++--------------- 1 file changed, 81 insertions(+), 51 deletions(-) diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp index d6f647c89..83737c1d9 100644 --- a/ggml/src/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan.cpp @@ -787,6 +787,9 @@ static vk_submission ggml_vk_create_submission(vk_device& device, vk_queue& q, s static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) { if (ctx->seqs.empty()) { + if (fence) { + ctx->q->queue.submit({}, fence); + } return; } VK_LOG_DEBUG("ggml_vk_submit(" << ctx << ", " << fence << ")"); @@ -5658,11 +5661,15 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { } } -static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, bool last_node, bool dryrun){ +static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_tensor* tensor, int tensor_idx, bool use_fence); + +// Returns true if node has enqueued work into the queue, false otherwise +// If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution. +static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool submit){ ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra; if (ggml_is_empty(node) || extra == nullptr) { - return; + return false; } VK_LOG_DEBUG("ggml_vk_build_graph(" << node << ", " << ggml_op_name(node->op) << ")"); @@ -5679,7 +5686,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_NONE: - return; + return false; case GGML_OP_UNARY: switch (ggml_get_unary_op(node)) { case GGML_UNARY_OP_SILU: @@ -5689,7 +5696,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_UNARY_OP_TANH: break; default: - return; + return false; } break; case GGML_OP_REPEAT: @@ -5726,7 +5733,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod default: std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl; GGML_ABORT("fatal error"); - return; + return false; } vk_context compute_ctx; @@ -5826,7 +5833,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod ggml_vk_unary(ctx, compute_ctx, src0, node, dryrun); break; default: - return; + return false; } break; case GGML_OP_DIAG_MASK_INF: @@ -5870,11 +5877,11 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod break; default: - return; + return false; } if (dryrun) { - return; + return false; } ctx->tensor_ctxs[node_idx] = compute_ctx; @@ -5885,14 +5892,34 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod last_node = true; #endif - if (last_node) { + if (submit || last_node) { ggml_vk_ctx_end(compute_ctx); - compute_ctx->exit_tensor_idx = node_idx; + + // TODO probably it'd be better to pass a exit_node flag to ggml_vk_compute_forward + if (last_node) { + compute_ctx->exit_tensor_idx = node_idx_begin; + } + else { + compute_ctx->exit_tensor_idx = -1; + } + ctx->compute_ctx.reset(); + + bool ok = ggml_vk_compute_forward(ctx, node_begin, node_idx_begin, false); + if (!ok) { + if (node->op == GGML_OP_UNARY) { + std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast(node->op_params[0])) << ")" << std::endl; + } + else { + std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl; + } + } + } + return true; } -static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx){ +static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx, bool use_fence = true){ ggml_tensor_extra_gpu * extra = nullptr; switch (tensor->op) { @@ -5960,40 +5987,38 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * VK_LOG_DEBUG("ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")"); -#ifdef GGML_VULKAN_CHECK_RESULTS - ggml_vk_check_results_0(tensor); -#endif - vk_context subctx = ctx->tensor_ctxs[tensor_idx].lock(); -#ifdef GGML_VULKAN_PERF - std::chrono::steady_clock::time_point start; -#endif // GGML_VULKAN_PERF + // always wait for the GPU work to be done for the last submit + if (tensor_idx == subctx->exit_tensor_idx) { + use_fence = true; + } // Only run if ctx hasn't been submitted yet if (!subctx->seqs.empty()) { +#ifdef GGML_VULKAN_CHECK_RESULTS + ggml_vk_check_results_0(tensor); + use_fence = true; +#endif + // Do staging buffer copies for (auto& cpy : subctx->in_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); } -#ifdef GGML_VULKAN_PERF - start = std::chrono::steady_clock::now(); -#endif // GGML_VULKAN_PERF + ggml_vk_submit(subctx, use_fence ? ctx->fence : vk::Fence{}); - ggml_vk_submit(subctx, ctx->fence); + if (use_fence) { + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences"); + + ctx->device->device.resetFences({ ctx->fence }); + } +#ifdef GGML_VULKAN_CHECK_RESULTS + ggml_vk_check_results_1(tensor); +#endif } if (tensor_idx == subctx->exit_tensor_idx) { - VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences"); - -#ifdef GGML_VULKAN_PERF - auto duration = std::chrono::duration_cast(std::chrono::steady_clock::now() - start); - ctx->device->perf_logger->log_timing(tensor, duration.count()); -#endif // GGML_VULKAN_PERF - - ctx->device->device.resetFences({ ctx->fence }); - // Do staging buffer copies for (auto& cpy : subctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); @@ -6482,7 +6507,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; for (int i = 0; i < cgraph->n_nodes; i++) { - ggml_vk_build_graph(ctx, cgraph->nodes[i], i, 0, true); + ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false); } ggml_vk_preallocate_buffers(ctx); ggml_pipeline_allocate_descriptor_sets(ctx->device); @@ -6497,31 +6522,36 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen // Reserve tensor context space for all nodes ctx->tensor_ctxs.resize(cgraph->n_nodes); - for (int i = 0; i < cgraph->n_nodes; i++) { - ggml_vk_build_graph(ctx, cgraph->nodes[i], i, i == last_node, false); - } + bool first_node_in_batch = true; // true if next node will be first node in a batch + int submit_node_idx = 0; // index to first node in a batch + // submit work every submit_count node to overlap CPU cmdbuffer generation with GPU execution + constexpr int submit_count = 100; + int submitted_nodes = 0; for (int i = 0; i < cgraph->n_nodes; i++) { - ggml_tensor * node = cgraph->nodes[i]; - - if (ggml_vk_is_empty(node)) { - continue; + if (first_node_in_batch) { + submit_node_idx = i; } - bool ok = ggml_vk_compute_forward(ctx, node, i); - if (!ok) { - if (node->op == GGML_OP_UNARY) { - std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast(node->op_params[0])) << ")" << std::endl; - } else { - std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl; + bool submit = (submitted_nodes >= submit_count) || (i == last_node); + + + bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit); + + if (enqueued) { + ++submitted_nodes; + +#ifndef GGML_VULKAN_CHECK_RESULTS + if (first_node_in_batch) { + first_node_in_batch = false; } - } -#ifdef GGML_VULKAN_CHECK_RESULTS - else { - ggml_vk_check_results_1(node); - } #endif - GGML_ASSERT(ok); + } + + if (submit) { + first_node_in_batch = true; + submitted_nodes = 0; + } } #ifdef GGML_VULKAN_PERF From b2e89a327457179a34eae4d7de0d412ed945679c Mon Sep 17 00:00:00 2001 From: Dan Johansson <164997844+eddnjjn@users.noreply.github.com> Date: Mon, 9 Sep 2024 09:02:45 +0200 Subject: [PATCH 20/32] Arm AArch64: Documentation updates (#9321) * Arm AArch64: Documentation updates * Update docs/build.md to include information on how to enable the Arm optimized gemm/gemv kernels * Update examples/quantize/README.md with information on the Q4_0_4_4, Q4_0_4_8 and Q4_0_8_8 formats * Add newline to the end of docs/build.md --- docs/build.md | 6 ++++++ examples/quantize/README.md | 2 ++ 2 files changed, 8 insertions(+) diff --git a/docs/build.md b/docs/build.md index 152d46d6f..faa0ecfa4 100644 --- a/docs/build.md +++ b/docs/build.md @@ -380,3 +380,9 @@ For detailed info, such as model/device supports, CANN install, please refer to ### Android To read documentation for how to build on Android, [click here](./android.md) + +### Arm CPU optimized mulmat kernels + +Llama.cpp includes a set of optimized mulmat kernels for the Arm architecture, leveraging Arm® Neonâ„¢, int8mm and SVE instructions. These kernels are enabled at build time through the appropriate compiler cpu-type flags, such as `-DCMAKE_C_FLAGS=-march=armv8.2a+i8mm+sve`. Note that these optimized kernels require the model to be quantized into one of the formats: `Q4_0_4_4` (Arm Neon), `Q4_0_4_8` (int8mm) or `Q4_0_8_8` (SVE). The SVE mulmat kernel specifically requires a vector width of 256 bits. When running on devices with a different vector width, it is recommended to use the `Q4_0_4_8` (int8mm) or `Q4_0_4_4` (Arm Neon) formats for better performance. Refer to [examples/quantize/README.md](../examples/quantize/README.md) for more information on the quantization formats. + +To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_LLAMAFILE=OFF` (`cmake`). diff --git a/examples/quantize/README.md b/examples/quantize/README.md index 5d1e11c67..704f0d56b 100644 --- a/examples/quantize/README.md +++ b/examples/quantize/README.md @@ -54,6 +54,8 @@ As the models are currently fully loaded into memory, you will need adequate dis Several quantization methods are supported. They differ in the resulting model disk size and inference speed. +The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interleaved variants of the `Q4_0` format, providing a data layout that is better suited for specific implementations of optimized mulmat kernels. Since these formats differ only in data layout, they have the same quantized size as the `Q4_0` format. + *(outdated)* | Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 | From 54f376d0b92c6ff6feb1fa2ef8ed2022348100ba Mon Sep 17 00:00:00 2001 From: Radoslav Gerganov Date: Mon, 9 Sep 2024 11:04:39 +0300 Subject: [PATCH 21/32] rpc : update README [no ci] (#9320) Update README with instructions how to offload model layers to both local and remote devices --- examples/rpc/README.md | 29 ++++++++++++----------------- 1 file changed, 12 insertions(+), 17 deletions(-) diff --git a/examples/rpc/README.md b/examples/rpc/README.md index adedc8909..36ff4a330 100644 --- a/examples/rpc/README.md +++ b/examples/rpc/README.md @@ -10,20 +10,21 @@ This can be used for distributed LLM inference with `llama.cpp` in the following ```mermaid flowchart TD - rpcb---|TCP|srva - rpcb---|TCP|srvb - rpcb-.-|TCP|srvn + rpcb<-->|TCP|srva + rpcb<-->|TCP|srvb + rpcb<-.->|TCP|srvn subgraph hostn[Host N] - srvn[rpc-server]-.-backend3["Backend (CUDA,Metal,etc.)"] + srvn[rpc-server]<-.->backend3["Backend (CUDA,Metal,etc.)"] end subgraph hostb[Host B] - srvb[rpc-server]---backend2["Backend (CUDA,Metal,etc.)"] + srvb[rpc-server]<-->backend2["Backend (CUDA,Metal,etc.)"] end subgraph hosta[Host A] - srva[rpc-server]---backend["Backend (CUDA,Metal,etc.)"] + srva[rpc-server]<-->backend["Backend (CUDA,Metal,etc.)"] end subgraph host[Main Host] - ggml[llama.cpp]---rpcb[RPC backend] + local["Backend (CUDA,Metal,etc.)"]<-->ggml[llama-cli] + ggml[llama-cli]<-->rpcb[RPC backend] end style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5 ``` @@ -62,17 +63,11 @@ $ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052 This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device. -On the main host build `llama.cpp` only with `-DGGML_RPC=ON`: - -```bash -mkdir build-rpc -cd build-rpc -cmake .. -DGGML_RPC=ON -cmake --build . --config Release -``` - -Finally, use the `--rpc` option to specify the host and port of each `rpc-server`: +On the main host build `llama.cpp` for the local backend and add `-DGGML_RPC=ON` to the build options. +Finally, when running `llama-cli`, use the `--rpc` option to specify the host and port of each `rpc-server`: ```bash $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99 ``` + +This way you can offload model layers to both local and remote devices. \ No newline at end of file From 5ed087573e1f326cfa70e29c1895d074a7a1a00c Mon Sep 17 00:00:00 2001 From: Antonis Makropoulos Date: Mon, 9 Sep 2024 14:21:38 +0300 Subject: [PATCH 22/32] readme : add LLMUnity to UI projects (#9381) * add LLMUnity to UI projects * add newline to examples/rpc/README.md to fix editorconfig-checker unit test --- README.md | 1 + examples/rpc/README.md | 3 ++- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index e30ab0c8c..8cbfdda6b 100644 --- a/README.md +++ b/README.md @@ -163,6 +163,7 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT) - [AIKit](https://github.com/sozercan/aikit) (MIT) - [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL) +- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT) *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)* diff --git a/examples/rpc/README.md b/examples/rpc/README.md index 36ff4a330..312bb634d 100644 --- a/examples/rpc/README.md +++ b/examples/rpc/README.md @@ -70,4 +70,5 @@ Finally, when running `llama-cli`, use the `--rpc` option to specify the host an $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99 ``` -This way you can offload model layers to both local and remote devices. \ No newline at end of file +This way you can offload model layers to both local and remote devices. + From 8e6e2fbe1458ac91387266241262294a964d6b95 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Mon, 9 Sep 2024 14:22:53 +0200 Subject: [PATCH 23/32] CUDA: fix variable name conflict for Windows build (#9382) --- ggml/src/ggml-cuda/sum.cu | 12 +++++++----- 1 file changed, 7 insertions(+), 5 deletions(-) diff --git a/ggml/src/ggml-cuda/sum.cu b/ggml/src/ggml-cuda/sum.cu index 0d5e953ee..21da63509 100644 --- a/ggml/src/ggml-cuda/sum.cu +++ b/ggml/src/ggml-cuda/sum.cu @@ -1,13 +1,15 @@ +#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) +// On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh. +// For this reason CUB must be included BEFORE anything else. +#include +using namespace cub; +#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) + #include "sumrows.cuh" #include "sum.cuh" #include -#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) -#include -using namespace cub; -#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) - void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) { #if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) size_t tmp_size = 0; From 38ca6f644bd48301e9caa80f9913c22e70a8fd1b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 9 Sep 2024 15:51:37 +0300 Subject: [PATCH 24/32] readme : update hot topics --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 8cbfdda6b..c945e125c 100644 --- a/README.md +++ b/README.md @@ -17,7 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ## Hot topics -- *add hot topics here* +- Huggingface GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor) ---- From 5fb5e24811cb01d48b482c15a974bfbd9f433e1d Mon Sep 17 00:00:00 2001 From: slaren Date: Mon, 9 Sep 2024 17:10:46 +0200 Subject: [PATCH 25/32] llama : minor sampling refactor (2) (#9386) --- examples/batched.swift/Sources/main.swift | 2 - examples/batched/batched.cpp | 2 - examples/gritlm/gritlm.cpp | 1 - .../llama/src/main/cpp/llama-android.cpp | 2 - .../llama.cpp.swift/LibLlama.swift | 2 - examples/passkey/passkey.cpp | 2 - examples/save-load-state/save-load-state.cpp | 6 - examples/server/server.cpp | 2 +- examples/simple/simple.cpp | 2 - include/llama.h | 11 +- src/llama-sampling.cpp | 194 ++++++++++-------- tests/test-sampling.cpp | 2 +- 12 files changed, 115 insertions(+), 113 deletions(-) diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift index 4bc2bbf2c..9f7c49492 100644 --- a/examples/batched.swift/Sources/main.swift +++ b/examples/batched.swift/Sources/main.swift @@ -140,8 +140,6 @@ while n_cur <= n_len { let new_token_id = llama_sampler_sample(smpl, context, i_batch[i]) - llama_sampler_accept(smpl, new_token_id) - // is it an end of stream? -> mark the stream as finished if llama_token_is_eog(model, new_token_id) || n_cur == n_len { i_batch[i] = -1 diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index f5f309022..615d6f0f5 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -172,8 +172,6 @@ int main(int argc, char ** argv) { const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]); - llama_sampler_accept(smpl, new_token_id); - // is it an end of generation? -> mark the stream as finished if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) { i_batch[i] = -1; diff --git a/examples/gritlm/gritlm.cpp b/examples/gritlm/gritlm.cpp index e1efbf573..6f060e2dc 100644 --- a/examples/gritlm/gritlm.cpp +++ b/examples/gritlm/gritlm.cpp @@ -121,7 +121,6 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std llama_decode(ctx, bat); llama_token token = llama_sampler_sample(smpl, ctx, bat.n_tokens - 1); - llama_sampler_accept(smpl, token); if (token == eos_token) { break; diff --git a/examples/llama.android/llama/src/main/cpp/llama-android.cpp b/examples/llama.android/llama/src/main/cpp/llama-android.cpp index 06ec160c2..f611809c6 100644 --- a/examples/llama.android/llama/src/main/cpp/llama-android.cpp +++ b/examples/llama.android/llama/src/main/cpp/llama-android.cpp @@ -414,8 +414,6 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop( // sample the most likely token const auto new_token_id = llama_sampler_sample(sampler, context, -1); - llama_sampler_accept(sampler, new_token_id); - const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value); if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) { return nullptr; diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index 92f61fe83..dcd9803a2 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -152,8 +152,6 @@ actor LlamaContext { new_token_id = llama_sampler_sample(sampling, context, batch.n_tokens - 1) - llama_sampler_accept(sampling, new_token_id) - if llama_token_is_eog(model, new_token_id) || n_cur == n_len { print("\n") is_done = true diff --git a/examples/passkey/passkey.cpp b/examples/passkey/passkey.cpp index 76d235c2c..271ef3a98 100644 --- a/examples/passkey/passkey.cpp +++ b/examples/passkey/passkey.cpp @@ -220,8 +220,6 @@ int main(int argc, char ** argv) { { const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1); - llama_sampler_accept(smpl, new_token_id); - // is it an end of generation? if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) { LOG_TEE("\n"); diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index b54ec3bd8..e17ab0ed0 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -74,8 +74,6 @@ int main(int argc, char ** argv) { auto next_token = llama_sampler_sample(smpl, ctx, -1); auto next_token_str = llama_token_to_piece(ctx, next_token); - llama_sampler_accept(smpl, next_token); - printf("%s", next_token_str.c_str()); result0 += next_token_str; @@ -132,8 +130,6 @@ int main(int argc, char ** argv) { auto next_token = llama_sampler_sample(smpl2, ctx2, -1); auto next_token_str = llama_token_to_piece(ctx2, next_token); - llama_sampler_accept(smpl2, next_token); - printf("%s", next_token_str.c_str()); result1 += next_token_str; @@ -222,8 +218,6 @@ int main(int argc, char ** argv) { auto next_token = llama_sampler_sample(smpl3, ctx3, -1); auto next_token_str = llama_token_to_piece(ctx3, next_token); - llama_sampler_accept(smpl3, next_token); - printf("%s", next_token_str.c_str()); result2 += next_token_str; diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 9ab8f8ca6..de3ea313c 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -613,7 +613,7 @@ struct server_context { gpt_params params; - llama_batch batch; + llama_batch batch = {}; bool clean_kv_cache = true; bool add_bos_token = true; diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp index a53cef547..d040172a5 100644 --- a/examples/simple/simple.cpp +++ b/examples/simple/simple.cpp @@ -118,8 +118,6 @@ int main(int argc, char ** argv) { { const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1); - llama_sampler_accept(smpl, new_token_id); - // is it an end of generation? if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) { LOG_TEE("\n"); diff --git a/include/llama.h b/include/llama.h index 6334fc30d..93b3e6e85 100644 --- a/include/llama.h +++ b/include/llama.h @@ -1127,15 +1127,16 @@ extern "C" { int32_t n_logit_bias, const llama_logit_bias * logit_bias); - // Shorthand for: + /// @details Sample and accept a token from the idx-th output of the last evaluation // + // Shorthand for: // const auto * logits = llama_get_logits_ith(ctx, idx); // llama_token_data_array cur_p = { ... init from logits ... }; // llama_sampler_apply(smpl, &cur_p); - // return cur_p.data[cur_p.selected].id; - // - // At this point, this is mostly a convenience function. - // + // auto token = cur_p.data[cur_p.selected].id; + // llama_sampler_accept(smpl, token); + // return token; + // Returns the sampled token LLAMA_API llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx); // TODO: extend in the future diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp index 41f48ec28..6f448b80c 100644 --- a/src/llama-sampling.cpp +++ b/src/llama-sampling.cpp @@ -8,49 +8,44 @@ #include #include #include +#include #include #include #include -static int llama_sample_dist(llama_token_data_array * cur_p, std::mt19937 & rng, std::vector & probs) { -#if 1 - probs.resize(cur_p->size); - for (size_t i = 0; i < cur_p->size; ++i) { - probs[i] = cur_p->data[i].p; - } - - std::discrete_distribution dist(probs.begin(), probs.end()); -#else - // avoid the copy with a custom iterator +static int llama_sample_dist(llama_token_data_array * cur_p, std::mt19937 & rng) { + // iterator for the probabilities +#ifdef __GNUC__ #pragma GCC diagnostic push #pragma GCC diagnostic ignored "-Wunused-local-typedefs" +#endif struct probs_iterator { typedef std::input_iterator_tag iterator_category; typedef float value_type; typedef float * pointer; typedef float & reference; - typedef size_t difference_type; + typedef ptrdiff_t difference_type; - const llama_token_data_array * data; - size_t i; + const llama_token_data * data; - bool operator==(const probs_iterator & other) const { return data + i == other.data + other.i; } - bool operator!=(const probs_iterator & other) const { return data + i != other.data + other.i; } - float operator*() const { return data->data[i].p; } - probs_iterator & operator++() { ++i; return *this; } - probs_iterator operator++(int) { probs_iterator tmp = *this; ++i; return tmp; } + bool operator==(const probs_iterator & other) const { return data == other.data; } + bool operator!=(const probs_iterator & other) const { return data != other.data; } + const float & operator*() const { return data->p; } + probs_iterator & operator++() { ++data; return *this; } + probs_iterator operator++(int) { probs_iterator tmp = *this; ++data; return tmp; } }; + +#ifdef __GNUC__ #pragma GCC diagnostic pop - - std::discrete_distribution dist(probs_iterator{cur_p, 0}, probs_iterator{cur_p, cur_p->size}); - - GGML_UNUSED(probs); #endif + std::discrete_distribution dist(probs_iterator{cur_p->data}, probs_iterator{cur_p->data + cur_p->size}); + return dist(rng); } +/* static void llama_log_softmax(float * array, size_t size) { float max_l = *std::max_element(array, array + size); float sum = 0.f; @@ -64,6 +59,7 @@ static void llama_log_softmax(float * array, size_t size) { array[i] = logf(array[i] / sum); } } +*/ static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) { GGML_ASSERT(cur_p->size > 0); @@ -231,67 +227,92 @@ llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_conte cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f}; } - llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false }; + llama_token_data_array cur_p = { + /* .data = */ cur.data(), + /* .size = */ cur.size(), + /* .selected = */ -1, + /* .sorted = */ false, + }; llama_sampler_apply(smpl, &cur_p); - return cur_p.data[cur_p.selected].id; + GGML_ASSERT(cur_p.selected >= 0 && cur_p.selected < (int32_t) cur_p.size); + + auto token = cur_p.data[cur_p.selected].id; + + llama_sampler_accept(smpl, token); + + return token; } // sampler chain +static const char * llama_sampler_chain_name(const struct llama_sampler * /*smpl*/) { + return "chain"; +} + +static void llama_sampler_chain_accept(struct llama_sampler * smpl, llama_token token) { + auto * chain = (llama_sampler_chain *) smpl->ctx; + + time_meas tm(chain->t_sample_us, chain->params.no_perf); + + for (auto * smpl : chain->samplers) { + llama_sampler_accept(smpl, token); + } + + chain->n_sample++; +} + +static void llama_sampler_chain_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + auto * chain = (llama_sampler_chain *) smpl->ctx; + + time_meas tm(chain->t_sample_us, chain->params.no_perf); + + for (auto * smpl : chain->samplers) { + llama_sampler_apply(smpl, cur_p); + } +} + +static void llama_sampler_chain_reset(struct llama_sampler * smpl) { + auto * chain = (llama_sampler_chain *) smpl->ctx; + + for (auto * smpl : chain->samplers) { + llama_sampler_reset(smpl); + } + + chain->t_sample_us = 0; + chain->n_sample = 0; +} + +static struct llama_sampler * llama_sampler_chain_clone(const struct llama_sampler * smpl) { + const auto * chain_src = (const llama_sampler_chain *) smpl->ctx; + + auto * result = llama_sampler_chain_init(chain_src->params); + + for (auto * smpl : chain_src->samplers) { + llama_sampler_chain_add(result, llama_sampler_clone(smpl)); + } + + return result; +} + +static void llama_sampler_chain_free(struct llama_sampler * smpl) { + auto * chain = (llama_sampler_chain *) smpl->ctx; + + for (auto * smpl : chain->samplers) { + llama_sampler_free(smpl); + } + + delete chain; +} + static struct llama_sampler_i llama_sampler_chain_i = { - /* .name = */ [](const struct llama_sampler * /*smpl*/) { return "chain"; }, - /* .accept = */ [](struct llama_sampler * smpl, llama_token token) { - auto * chain = (llama_sampler_chain *) smpl->ctx; - - time_meas tm(chain->t_sample_us, chain->params.no_perf); - - for (auto * smpl : chain->samplers) { - llama_sampler_accept(smpl, token); - } - - chain->n_sample++; - }, - /* .apply = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) { - auto * chain = (llama_sampler_chain *) smpl->ctx; - - time_meas tm(chain->t_sample_us, chain->params.no_perf); - - for (auto * smpl : chain->samplers) { - llama_sampler_apply(smpl, cur_p); - } - }, - /* .reset = */ [](struct llama_sampler * smpl) { - auto * chain = (llama_sampler_chain *) smpl->ctx; - - for (auto * smpl : chain->samplers) { - llama_sampler_reset(smpl); - } - - chain->t_sample_us = 0; - chain->n_sample = 0; - }, - /* .clone = */ [](const struct llama_sampler * smpl) { - const auto * chain_src = (const llama_sampler_chain *) smpl->ctx; - - auto * result = llama_sampler_chain_init(chain_src->params); - - for (auto * smpl : chain_src->samplers) { - llama_sampler_chain_add(result, llama_sampler_clone(smpl)); - } - - return result; - }, - /* .free = */ [](struct llama_sampler * smpl) { - auto * chain = (llama_sampler_chain *) smpl->ctx; - - for (auto * smpl : chain->samplers) { - llama_sampler_free(smpl); - } - - delete chain; - }, + /* .name = */ llama_sampler_chain_name, + /* .accept = */ llama_sampler_chain_accept, + /* .apply = */ llama_sampler_chain_apply, + /* .reset = */ llama_sampler_chain_reset, + /* .clone = */ llama_sampler_chain_clone, + /* .free = */ llama_sampler_chain_free, }; struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params) { @@ -368,8 +389,6 @@ struct llama_sampler_dist { const uint32_t seed; std::mt19937 rng; - - std::vector probs; // work array }; static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*/) { @@ -378,7 +397,7 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl* static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { auto * ctx = (llama_sampler_dist *) smpl->ctx; - cur_p->selected = llama_sample_dist(cur_p, ctx->rng, ctx->probs); + cur_p->selected = llama_sample_dist(cur_p, ctx->rng); } static struct llama_sampler * llama_sampler_dist_clone(const struct llama_sampler * smpl) { @@ -419,7 +438,6 @@ struct llama_sampler * llama_sampler_init_dist(uint32_t seed) { /* .ctx = */ new llama_sampler_dist { /* .seed = */ seed, /* .rng = */ std::mt19937(seed), - /* .probs = */ {}, }, }; } @@ -1023,8 +1041,6 @@ struct llama_sampler_mirostat { float mu; std::mt19937 rng; - - std::vector probs; }; static const char * llama_sampler_mirostat_name(const struct llama_sampler * /*smpl*/) { @@ -1055,7 +1071,7 @@ static void llama_sampler_mirostat_apply(struct llama_sampler * smpl, llama_toke llama_sampler_top_k_impl(cur_p, std::max(int(k), 1)); llama_sampler_softmax_impl(cur_p); - const int idx = llama_sample_dist(cur_p, ctx->rng, ctx->probs); + const int idx = llama_sample_dist(cur_p, ctx->rng); cur_p->selected = idx; @@ -1111,7 +1127,6 @@ struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t see /* .m = */ m, /* .mu = */ 2.0f*tau, /* .rng = */ std::mt19937(seed), - /* .probs = */ {}, }, }; } @@ -1127,8 +1142,6 @@ struct llama_sampler_mirostat_v2 { float mu; std::mt19937 rng; - - std::vector probs; }; static const char * llama_sampler_mirostat_v2_name(const struct llama_sampler * /*smpl*/) { @@ -1152,7 +1165,7 @@ static void llama_sampler_mirostat_v2_apply(struct llama_sampler * smpl, llama_t // Normalize the probabilities of the remaining words llama_sampler_softmax_impl(cur_p); - const int idx = llama_sample_dist(cur_p, ctx->rng, ctx->probs); + const int idx = llama_sample_dist(cur_p, ctx->rng); cur_p->selected = idx; @@ -1207,7 +1220,6 @@ struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau, /* .eta = */ eta, /* .mu = */ 2.0f*tau, /* .rng = */ std::mt19937(seed), - /* .probs = */ {}, }, }; } @@ -1527,6 +1539,10 @@ static const char * llama_sampler_logit_bias_name(const struct llama_sampler * / static void llama_sampler_logit_bias_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { auto * ctx = (llama_sampler_logit_bias *) smpl->ctx; + if (ctx->logit_bias.empty()) { + return; + } + ctx->to_search.clear(); // update the candidates that have not been shuffled in the vocabulary (i.e. idx == id) @@ -1538,6 +1554,10 @@ static void llama_sampler_logit_bias_apply(struct llama_sampler * smpl, llama_to } } + if (ctx->to_search.empty()) { + return; + } + // search for the remaining candidates that were not found in the previous step for (size_t i = 0; i < cur_p->size; ++i) { for (const auto & lb : ctx->to_search) { diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index 37400c179..d738b7a45 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -245,7 +245,7 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler } } - printf("Sampler queue %3s OK with n_vocab=%05ld top_k=%05d top_p=%f min_p=%f\n", + printf("Sampler queue %3s OK with n_vocab=%05zu top_k=%05d top_p=%f min_p=%f\n", samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p); } From 5fac4d57643b1de8e9ab746f14d2fc4e319ae0c2 Mon Sep 17 00:00:00 2001 From: Prashant Vithule <119530321+Vithulep@users.noreply.github.com> Date: Mon, 9 Sep 2024 21:07:18 +0530 Subject: [PATCH 26/32] ggml : vector length agnostic SVE support (#9290) * Implemented vector length agnostic SVE using switch case for 512-bit, 256-bit, 128-bit vector lengths * Implemented vector length agnostic SVE using switch case for 512-bit, 256-bit, 128-bit vector lengths * Removed WhiteSpaces * ggml : style changes + fix 512-bit nb loop check - fix local scope in switch cases - consistent predicate names - empty lines when necessary - opening braces, spaces - const-correctness - add asserts * Update ggml/src/ggml-quants.c Co-authored-by: Georgi Gerganov --------- Co-authored-by: Georgi Gerganov --- ggml/src/ggml-quants.c | 284 ++++++++++++++++++++++++++++++++++------- 1 file changed, 239 insertions(+), 45 deletions(-) diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 8c31e2cca..322c85d2a 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -4003,42 +4003,141 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r float sumf = 0; #if defined(__ARM_FEATURE_SVE) - if (ggml_sve_cnt_b == QK8_0) { - const svbool_t ptrueh = svptrue_pat_b8(SV_VL16); - const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh); + svfloat32_t sumv0 = svdup_n_f32(0.0f); + svfloat32_t sumv1 = svdup_n_f32(0.0f); - svfloat32_t sumv0 = svdup_n_f32(0.0f); - svfloat32_t sumv1 = svdup_n_f32(0.0f); + const int vector_length = ggml_sve_cnt_b*8; - for (; ib + 1 < nb; ib += 2) { - const block_q4_0 * restrict x0 = &x[ib + 0]; - const block_q4_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + // VLA Implementation using switch case + switch (vector_length) { + case 128: + { + // predicate for activating higher lanes for 4 float32 elements + const svbool_t ph4 = svptrue_pat_b32(SV_VL4); - // load x - const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs); - const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs); + for (; ib + 1 < nb; ib += 2) { + const block_q4_0 * restrict x0 = &x[ib + 0]; + const block_q4_0 * restrict x1 = &x[ib + 1]; + const block_q8_0 * restrict y0 = &y[ib + 0]; + const block_q8_0 * restrict y1 = &y[ib + 1]; - // 4-bit -> 8-bit - const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04)); - const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04)); + // load x + const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs); + const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs); - // sub 8 - const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8); - const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8); + // 4-bit -> 8-bit + const svint8_t qx0l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx0r, 0x0F)); + const svint8_t qx0h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx0r, 0x04)); + const svint8_t qx1l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx1r, 0x0F)); + const svint8_t qx1h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx1r, 0x04)); - // load y - const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); - const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); + // sub 8 + const svint8_t qx0ls = svsub_n_s8_x(svptrue_b8(), qx0h, 8); + const svint8_t qx0hs = svsub_n_s8_x(svptrue_b8(), qx0l, 8); + const svint8_t qx1ls = svsub_n_s8_x(svptrue_b8(), qx1h, 8); + const svint8_t qx1hs = svsub_n_s8_x(svptrue_b8(), qx1l, 8); - // dot product - sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); - } + // load y + const svint8_t qy0h = svld1_s8(svptrue_b8(), y0->qs); + const svint8_t qy0l = svld1_s8(svptrue_b8(), y0->qs + 16); + const svint8_t qy1h = svld1_s8(svptrue_b8(), y1->qs); + const svint8_t qy1l = svld1_s8(svptrue_b8(), y1->qs + 16); - sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); + // dot product + sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4, + svdot_s32(svdup_n_s32(0), qx0ls, qy0l), + svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4, + svdot_s32(svdup_n_s32(0), qx1ls, qy1l), + svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); + } break; + case 256: + { + // predicate for activating higher lanes for 16 int8 elements + const svbool_t ph16 = svptrue_pat_b8(SV_VL16); + // predicate for activating lower lanes for 16 int8 elements + const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16); + + for (; ib + 1 < nb; ib += 2) { + const block_q4_0 * restrict x0 = &x[ib + 0]; + const block_q4_0 * restrict x1 = &x[ib + 1]; + const block_q8_0 * restrict y0 = &y[ib + 0]; + const block_q8_0 * restrict y1 = &y[ib + 1]; + + // load x + const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs); + const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs); + + // 4-bit -> 8-bit + const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04)); + const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04)); + + // sub 8 + const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8); + const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8); + + // load y + const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); + const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); + + // dot product + sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), + svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), + svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); + } break; + case 512: + { + // predicate for activating higher lanes for 32 int8 elements + const svbool_t ph32 = svptrue_pat_b8(SV_VL32); + + // predicate for activating higher lanes for 16 int8 elements + const svbool_t ph16 = svptrue_pat_b8(SV_VL16); + // predicate for activating lower lanes for 16 int8 elements from first 32 int8 activated lanes + const svbool_t pl16 = svnot_b_z(ph32, ph16); + + for (; ib + 1 < nb; ib += 2) { + const block_q4_0 * restrict x0 = &x[ib + 0]; + const block_q4_0 * restrict x1 = &x[ib + 1]; + const block_q8_0 * restrict y0 = &y[ib + 0]; + const block_q8_0 * restrict y1 = &y[ib + 1]; + + // load x + const svuint8_t qx0r = svld1rq_u8(ph32, x0->qs); + const svuint8_t qx1r = svld1rq_u8(ph32, x1->qs); + + // 4-bit -> 8-bit + const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04)); + const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04)); + + // sub 8 + const svint8_t qx0s = svsub_n_s8_x(ph32, qx0, 8); + const svint8_t qx1s = svsub_n_s8_x(ph32, qx1, 8); + + // load y + const svint8_t qy0 = svld1_s8(ph32, y0->qs); + const svint8_t qy1 = svld1_s8(ph32, y1->qs); + + // dot product + sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32, + svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32, + svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1)); + } break; + default: + assert(false && "Unsupported vector length"); + break; } + #elif defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -5488,29 +5587,124 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r float sumf = 0; #if defined(__ARM_FEATURE_SVE) - if (ggml_sve_cnt_b == QK8_0) { - svfloat32_t sumv0 = svdup_n_f32(0.0f); - svfloat32_t sumv1 = svdup_n_f32(0.0f); + svfloat32_t sumv0 = svdup_n_f32(0.0f); + svfloat32_t sumv1 = svdup_n_f32(0.0f); - for (; ib + 1 < nb; ib += 2) { - const block_q8_0 * restrict x0 = &x[ib + 0]; - const block_q8_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const int vector_length = ggml_sve_cnt_b*8; - // load x - const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs); - const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs); + //VLA Implemenation for SVE + switch (vector_length) { + case 128: + { + // predicate for activating lanes for 16 Int8 elements + const svbool_t ph16 = svptrue_pat_b8 (SV_VL16); + const svbool_t pl16 = svptrue_pat_b32(SV_VL4); - // load y - const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); - const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); + for (; ib + 1 < nb; ib += 2) { + const block_q8_0 * restrict x0 = &x[ib + 0]; + const block_q8_0 * restrict x1 = &x[ib + 1]; + const block_q8_0 * restrict y0 = &y[ib + 0]; + const block_q8_0 * restrict y1 = &y[ib + 1]; - sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); - } + // load x + const svint8_t qx0_0 = svld1_s8(ph16, x0->qs); + const svint8_t qx0_1 = svld1_s8(ph16, x0->qs+16); + const svint8_t qx1_0 = svld1_s8(ph16, x1->qs); + const svint8_t qx1_1 = svld1_s8(ph16, x1->qs+16); - sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); + // load y + const svint8_t qy0_0 = svld1_s8(ph16, y0->qs); + const svint8_t qy0_1 = svld1_s8(ph16, y0->qs+16); + const svint8_t qy1_0 = svld1_s8(ph16, y1->qs); + const svint8_t qy1_1 = svld1_s8(ph16, y1->qs+16); + + sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16, + svdot_s32(svdup_n_s32(0), qx0_0, qy0_0), + svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16, + svdot_s32(svdup_n_s32(0), qx1_0, qy1_0), + svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1)); + } break; + case 256: + { + //printf("sve256"); + for (; ib + 1 < nb; ib += 2) { + const block_q8_0 * restrict x0 = &x[ib + 0]; + const block_q8_0 * restrict x1 = &x[ib + 1]; + const block_q8_0 * restrict y0 = &y[ib + 0]; + const block_q8_0 * restrict y1 = &y[ib + 1]; + + // load x + const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs); + const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs); + + // load y + const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); + const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); + + sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), + svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), + svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); + } break; + case 512: + { + // predicate for activating high 256 bit + const svbool_t ph32 = svptrue_pat_b8(SV_VL32); + // predicate for activating low 256 bit + const svbool_t pl32 = svnot_b_z(svptrue_b8(), ph32); + + // predicate for activating high lanes for 8 float32 elements + const svbool_t ph8 = svptrue_pat_b32(SV_VL8); + // predicate for activating low lanes for 8 float32 elements + const svbool_t pl8 = svnot_b_z(svptrue_b32(), ph8); + + svfloat32_t sumv00 = svdup_n_f32(0.0f); + + for (; ib + 1 < nb; ib += 2) { + const block_q8_0 * restrict x0 = &x[ib + 0]; + const block_q8_0 * restrict x1 = &x[ib + 1]; + const block_q8_0 * restrict y0 = &y[ib + 0]; + const block_q8_0 * restrict y1 = &y[ib + 1]; + + //load 32 int8_t in first half of vector and put another 32 int8_t in second vector lower bits + // and add them to make one 64 element vector + // load x + const svint8_t qx_32 = svld1_s8(ph32, x0->qs); + svint8_t qx_64 = svld1_s8(pl32, x0->qs + 2); + + qx_64 = svadd_s8_x(svptrue_b8(), qx_32, qx_64); + + // load y + const svint8_t qy_32 = svld1_s8(ph32, y0->qs); + svint8_t qy_64 = svld1_s8(pl32, y0->qs + 2); + + qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64); + + // scale creation + const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d); + const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d); + + // duplicate deq1 in first half of vector and deq2 in second half of vector + const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2); + + const svfloat32_t sumvt = svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx_64, qy_64)); + + sumv00 = svmla_f32_m(svptrue_b32(), sumv00, sumvt, temp); + } + + sumf = svaddv_f32(svptrue_b32(), sumv00); + break; + } + default: + assert(false && "Unsupported vector length"); + break; } #elif defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); From 293bebe0773c907c0c866213856eeba41b035df1 Mon Sep 17 00:00:00 2001 From: Radoslav Gerganov Date: Mon, 9 Sep 2024 18:40:10 +0300 Subject: [PATCH 27/32] rpc : fix segfault with nkvo (#9389) * rpc : fix nkvo * rpc : buf_size must not be static ref: #9337 --------- Co-authored-by: slaren --- ggml/src/ggml-cuda.cu | 6 +++++- ggml/src/ggml-rpc.cpp | 18 ++++++++++-------- ggml/src/ggml.c | 2 +- 3 files changed, 16 insertions(+), 10 deletions(-) diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 982316f56..d53de4edd 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -2552,7 +2552,11 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t for (int i = 0; i < cgraph->n_nodes; i++) { ggml_tensor * node = cgraph->nodes[i]; - if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) { + if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) { + continue; + } + + if (node->src[0] && node->src[0]->buffer && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) { use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture #ifndef NDEBUG GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__); diff --git a/ggml/src/ggml-rpc.cpp b/ggml/src/ggml-rpc.cpp index 8f9d0a460..9c600c7ca 100644 --- a/ggml/src/ggml-rpc.cpp +++ b/ggml/src/ggml-rpc.cpp @@ -883,15 +883,17 @@ ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rp } result->buffer = reinterpret_cast(tensor->buffer); if (result->buffer && buffers.find(result->buffer) == buffers.end()) { - return nullptr; + result->buffer = nullptr; } - // require that the tensor data does not go beyond the buffer end - uint64_t tensor_size = (uint64_t) ggml_nbytes(result); - uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer); - uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer); - GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow - GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size); + if (result->buffer) { + // require that the tensor data does not go beyond the buffer end + uint64_t tensor_size = (uint64_t) ggml_nbytes(result); + uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer); + uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer); + GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow + GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size); + } result->op = (ggml_op) tensor->op; for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) { @@ -1060,7 +1062,7 @@ bool rpc_server::graph_compute(const std::vector & input, std::vector ctx->mem_size) { GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", - __func__, cur_end + size_needed, ctx->mem_size); + __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size); assert(false); return NULL; } From bfe76d4a17228bfd1565761f203123bc4914771b Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Mon, 9 Sep 2024 23:36:09 +0200 Subject: [PATCH 28/32] common : move arg parser code to `arg.cpp` (#9388) * common : move arg parser to arg.cpp * better categorize args * add cmake * missing climits * missing cstdarg * common : more explicit includes * fix build * refactor gpt_params_parse * update server readme * fix test --------- Co-authored-by: Georgi Gerganov --- Makefile | 6 + common/CMakeLists.txt | 2 + common/arg.cpp | 1994 +++++++++++++++++ common/arg.h | 77 + common/common.cpp | 1929 ---------------- common/common.h | 214 +- common/sampling.cpp | 5 +- common/sampling.h | 54 +- examples/batched-bench/batched-bench.cpp | 4 +- examples/batched/batched.cpp | 4 +- .../cvector-generator/cvector-generator.cpp | 4 +- examples/cvector-generator/pca.hpp | 5 +- examples/embedding/embedding.cpp | 4 +- examples/eval-callback/eval-callback.cpp | 4 +- examples/export-lora/export-lora.cpp | 4 +- examples/gen-docs/gen-docs.cpp | 5 +- examples/gritlm/gritlm.cpp | 4 +- examples/imatrix/imatrix.cpp | 4 +- examples/infill/infill.cpp | 6 +- examples/llava/llava-cli.cpp | 10 +- examples/llava/minicpmv-cli.cpp | 7 +- examples/lookahead/lookahead.cpp | 5 +- examples/lookup/lookup-create.cpp | 10 +- examples/lookup/lookup-stats.cpp | 8 +- examples/lookup/lookup.cpp | 7 +- examples/main/main.cpp | 7 +- examples/parallel/parallel.cpp | 5 +- examples/passkey/passkey.cpp | 4 +- examples/perplexity/perplexity.cpp | 14 +- examples/retrieval/retrieval.cpp | 4 +- examples/save-load-state/save-load-state.cpp | 4 +- examples/server/README.md | 19 +- examples/server/server.cpp | 5 +- examples/simple/simple.cpp | 4 +- examples/speculative/speculative.cpp | 7 +- tests/test-arg-parser.cpp | 42 +- 36 files changed, 2281 insertions(+), 2210 deletions(-) create mode 100644 common/arg.cpp create mode 100644 common/arg.h diff --git a/Makefile b/Makefile index 6053bc17b..97ef37c0e 100644 --- a/Makefile +++ b/Makefile @@ -925,6 +925,7 @@ OBJ_LLAMA = \ OBJ_COMMON = \ common/common.o \ + common/arg.o \ common/console.o \ common/ngram-cache.o \ common/sampling.o \ @@ -1157,6 +1158,11 @@ common/common.o: \ include/llama.h $(CXX) $(CXXFLAGS) -c $< -o $@ +common/arg.o: \ + common/arg.cpp \ + common/arg.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + common/sampling.o: \ common/sampling.cpp \ common/sampling.h \ diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt index 2c72793b8..22fd99689 100644 --- a/common/CMakeLists.txt +++ b/common/CMakeLists.txt @@ -54,6 +54,8 @@ add_library(${TARGET} STATIC base64.hpp common.h common.cpp + arg.h + arg.cpp sampling.h sampling.cpp console.h diff --git a/common/arg.cpp b/common/arg.cpp new file mode 100644 index 000000000..c5134be51 --- /dev/null +++ b/common/arg.cpp @@ -0,0 +1,1994 @@ +#include "arg.h" + +#include "sampling.h" + +#include +#include +#include +#include +#include +#include +#include +#include + +#include "json-schema-to-grammar.h" + +using json = nlohmann::ordered_json; + +llama_arg & llama_arg::set_examples(std::initializer_list examples) { + this->examples = std::move(examples); + return *this; +} + +llama_arg & llama_arg::set_env(const char * env) { + help = help + "\n(env: " + env + ")"; + this->env = env; + return *this; +} + +llama_arg & llama_arg::set_sparam() { + is_sparam = true; + return *this; +} + +bool llama_arg::in_example(enum llama_example ex) { + return examples.find(ex) != examples.end(); +} + +bool llama_arg::get_value_from_env(std::string & output) { + if (env == nullptr) return false; + char * value = std::getenv(env); + if (value) { + output = value; + return true; + } + return false; +} + +bool llama_arg::has_value_from_env() { + return env != nullptr && std::getenv(env); +} + +static std::vector break_str_into_lines(std::string input, size_t max_char_per_line) { + std::vector result; + std::istringstream iss(input); + std::string line; + auto add_line = [&](const std::string& l) { + if (l.length() <= max_char_per_line) { + result.push_back(l); + } else { + std::istringstream line_stream(l); + std::string word, current_line; + while (line_stream >> word) { + if (current_line.length() + !current_line.empty() + word.length() > max_char_per_line) { + if (!current_line.empty()) result.push_back(current_line); + current_line = word; + } else { + current_line += (!current_line.empty() ? " " : "") + word; + } + } + if (!current_line.empty()) result.push_back(current_line); + } + }; + while (std::getline(iss, line)) { + add_line(line); + } + return result; +} + +std::string llama_arg::to_string() { + // params for printing to console + const static int n_leading_spaces = 40; + const static int n_char_per_line_help = 70; // TODO: detect this based on current console + std::string leading_spaces(n_leading_spaces, ' '); + + std::ostringstream ss; + for (const auto arg : args) { + if (arg == args.front()) { + if (args.size() == 1) { + ss << arg; + } else { + // first arg is usually abbreviation, we need padding to make it more beautiful + auto tmp = std::string(arg) + ", "; + auto spaces = std::string(std::max(0, 7 - (int)tmp.size()), ' '); + ss << tmp << spaces; + } + } else { + ss << arg << (arg != args.back() ? ", " : ""); + } + } + if (value_hint) ss << " " << value_hint; + if (value_hint_2) ss << " " << value_hint_2; + if (ss.tellp() > n_leading_spaces - 3) { + // current line is too long, add new line + ss << "\n" << leading_spaces; + } else { + // padding between arg and help, same line + ss << std::string(leading_spaces.size() - ss.tellp(), ' '); + } + const auto help_lines = break_str_into_lines(help, n_char_per_line_help); + for (const auto & line : help_lines) { + ss << (&line == &help_lines.front() ? "" : leading_spaces) << line << "\n"; + } + return ss.str(); +} + +// +// utils +// + +#ifdef __GNUC__ +#ifdef __MINGW32__ +#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) +#else +#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) +#endif +#else +#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) +#endif + +LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2) +static std::string format(const char * fmt, ...) { + va_list ap; + va_list ap2; + va_start(ap, fmt); + va_copy(ap2, ap); + int size = vsnprintf(NULL, 0, fmt, ap); + GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT + std::vector buf(size + 1); + int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2); + GGML_ASSERT(size2 == size); + va_end(ap2); + va_end(ap); + return std::string(buf.data(), size); +} + +static void gpt_params_handle_model_default(gpt_params & params) { + if (!params.hf_repo.empty()) { + // short-hand to avoid specifying --hf-file -> default it to --model + if (params.hf_file.empty()) { + if (params.model.empty()) { + throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n"); + } + params.hf_file = params.model; + } else if (params.model.empty()) { + params.model = fs_get_cache_file(string_split(params.hf_file, '/').back()); + } + } else if (!params.model_url.empty()) { + if (params.model.empty()) { + auto f = string_split(params.model_url, '#').front(); + f = string_split(f, '?').front(); + params.model = fs_get_cache_file(string_split(f, '/').back()); + } + } else if (params.model.empty()) { + params.model = DEFAULT_MODEL_PATH; + } +} + +// +// CLI argument parsing functions +// + +static bool gpt_params_parse_ex(int argc, char ** argv, gpt_params_context & ctx_arg) { + std::string arg; + const std::string arg_prefix = "--"; + gpt_params & params = ctx_arg.params; + gpt_sampler_params & sparams = params.sparams; + + std::unordered_map arg_to_options; + for (auto & opt : ctx_arg.options) { + for (const auto & arg : opt.args) { + arg_to_options[arg] = &opt; + } + } + + // handle environment variables + for (auto & opt : ctx_arg.options) { + std::string value; + if (opt.get_value_from_env(value)) { + try { + if (opt.handler_void && (value == "1" || value == "true")) { + opt.handler_void(params); + } + if (opt.handler_int) { + opt.handler_int(params, std::stoi(value)); + } + if (opt.handler_string) { + opt.handler_string(params, value); + continue; + } + } catch (std::exception & e) { + throw std::invalid_argument(format( + "error while handling environment variable \"%s\": %s\n\n", opt.env, e.what())); + } + } + } + + // handle command line arguments + auto check_arg = [&](int i) { + if (i+1 >= argc) { + throw std::invalid_argument("expected value for argument"); + } + }; + + for (int i = 1; i < argc; i++) { + const std::string arg_prefix = "--"; + + std::string arg = argv[i]; + if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) { + std::replace(arg.begin(), arg.end(), '_', '-'); + } + if (arg_to_options.find(arg) == arg_to_options.end()) { + throw std::invalid_argument(format("error: invalid argument: %s", arg.c_str())); + } + auto opt = *arg_to_options[arg]; + if (opt.has_value_from_env()) { + fprintf(stderr, "warn: %s environment variable is set, but will be overwritten by command line argument %s\n", opt.env, arg.c_str()); + } + try { + if (opt.handler_void) { + opt.handler_void(params); + continue; + } + + // arg with single value + check_arg(i); + std::string val = argv[++i]; + if (opt.handler_int) { + opt.handler_int(params, std::stoi(val)); + continue; + } + if (opt.handler_string) { + opt.handler_string(params, val); + continue; + } + + // arg with 2 values + check_arg(i); + std::string val2 = argv[++i]; + if (opt.handler_str_str) { + opt.handler_str_str(params, val, val2); + continue; + } + } catch (std::exception & e) { + throw std::invalid_argument(format( + "error while handling argument \"%s\": %s\n\n" + "usage:\n%s\n\nto show complete usage, run with -h", + arg.c_str(), e.what(), arg_to_options[arg]->to_string().c_str())); + } + } + + postprocess_cpu_params(params.cpuparams, nullptr); + postprocess_cpu_params(params.cpuparams_batch, ¶ms.cpuparams); + postprocess_cpu_params(params.draft_cpuparams, ¶ms.cpuparams); + postprocess_cpu_params(params.draft_cpuparams_batch, ¶ms.cpuparams_batch); + + if (params.prompt_cache_all && (params.interactive || params.interactive_first)) { + throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); + } + + gpt_params_handle_model_default(params); + + if (params.escape) { + string_process_escapes(params.prompt); + string_process_escapes(params.input_prefix); + string_process_escapes(params.input_suffix); + for (auto & antiprompt : params.antiprompt) { + string_process_escapes(antiprompt); + } + } + + if (!params.kv_overrides.empty()) { + params.kv_overrides.emplace_back(); + params.kv_overrides.back().key[0] = 0; + } + + if (sparams.seed == LLAMA_DEFAULT_SEED) { + sparams.seed = time(NULL); + } + + return true; +} + +static void gpt_params_print_usage(gpt_params_context & ctx_arg) { + auto print_options = [](std::vector & options) { + for (llama_arg * opt : options) { + printf("%s", opt->to_string().c_str()); + } + }; + + std::vector common_options; + std::vector sparam_options; + std::vector specific_options; + for (auto & opt : ctx_arg.options) { + // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example + if (opt.is_sparam) { + sparam_options.push_back(&opt); + } else if (opt.in_example(ctx_arg.ex)) { + specific_options.push_back(&opt); + } else { + common_options.push_back(&opt); + } + } + printf("----- common params -----\n\n"); + print_options(common_options); + printf("\n\n----- sampling params -----\n\n"); + print_options(sparam_options); + // TODO: maybe convert enum llama_example to string + printf("\n\n----- example-specific params -----\n\n"); + print_options(specific_options); +} + +bool gpt_params_parse(int argc, char ** argv, gpt_params & params, llama_example ex, void(*print_usage)(int, char **)) { + auto ctx_arg = gpt_params_parser_init(params, ex, print_usage); + const gpt_params params_org = ctx_arg.params; // the example can modify the default params + + try { + if (!gpt_params_parse_ex(argc, argv, ctx_arg)) { + ctx_arg.params = params_org; + return false; + } + if (ctx_arg.params.usage) { + gpt_params_print_usage(ctx_arg); + if (ctx_arg.print_usage) { + ctx_arg.print_usage(argc, argv); + } + exit(0); + } + } catch (const std::invalid_argument & ex) { + fprintf(stderr, "%s\n", ex.what()); + ctx_arg.params = params_org; + return false; + } + + return true; +} + +gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, void(*print_usage)(int, char **)) { + gpt_params_context ctx_arg(params); + ctx_arg.print_usage = print_usage; + ctx_arg.ex = ex; + + std::string sampler_type_chars; + std::string sampler_type_names; + for (const auto & sampler : params.sparams.samplers) { + sampler_type_chars += gpt_sampler_type_to_chr(sampler); + sampler_type_names += gpt_sampler_type_to_str(sampler) + ";"; + } + sampler_type_names.pop_back(); + + + /** + * filter options by example + * rules: + * - all examples inherit options from LLAMA_EXAMPLE_COMMON + * - if LLAMA_EXAMPLE_* is set (other than COMMON), we only show the option in the corresponding example + * - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example + */ + auto add_opt = [&](llama_arg arg) { + if (arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) { + ctx_arg.options.push_back(std::move(arg)); + } + }; + + + add_opt(llama_arg( + {"-h", "--help", "--usage"}, + "print usage and exit", + [](gpt_params & params) { + params.usage = true; + } + )); + add_opt(llama_arg( + {"--version"}, + "show version and build info", + [](gpt_params &) { + fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT); + fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET); + exit(0); + } + )); + add_opt(llama_arg( + {"-v", "--verbose"}, + "print verbose information", + [](gpt_params & params) { + params.verbosity = 1; + } + )); + add_opt(llama_arg( + {"--verbosity"}, "N", + format("set specific verbosity level (default: %d)", params.verbosity), + [](gpt_params & params, int value) { + params.verbosity = value; + } + )); + add_opt(llama_arg( + {"--verbose-prompt"}, + format("print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false"), + [](gpt_params & params) { + params.verbose_prompt = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--no-display-prompt"}, + format("don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false"), + [](gpt_params & params) { + params.display_prompt = false; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"-co", "--color"}, + format("colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false"), + [](gpt_params & params) { + params.use_color = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL})); + add_opt(llama_arg( + {"-t", "--threads"}, "N", + format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads), + [](gpt_params & params, int value) { + params.cpuparams.n_threads = value; + if (params.cpuparams.n_threads <= 0) { + params.cpuparams.n_threads = std::thread::hardware_concurrency(); + } + } + ).set_env("LLAMA_ARG_THREADS")); + add_opt(llama_arg( + {"-tb", "--threads-batch"}, "N", + "number of threads to use during batch and prompt processing (default: same as --threads)", + [](gpt_params & params, int value) { + params.cpuparams_batch.n_threads = value; + if (params.cpuparams_batch.n_threads <= 0) { + params.cpuparams_batch.n_threads = std::thread::hardware_concurrency(); + } + } + )); + add_opt(llama_arg( + {"-td", "--threads-draft"}, "N", + "number of threads to use during generation (default: same as --threads)", + [](gpt_params & params, int value) { + params.draft_cpuparams.n_threads = value; + if (params.draft_cpuparams.n_threads <= 0) { + params.draft_cpuparams.n_threads = std::thread::hardware_concurrency(); + } + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-tbd", "--threads-batch-draft"}, "N", + "number of threads to use during batch and prompt processing (default: same as --threads-draft)", + [](gpt_params & params, int value) { + params.draft_cpuparams_batch.n_threads = value; + if (params.draft_cpuparams_batch.n_threads <= 0) { + params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency(); + } + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-C", "--cpu-mask"}, "M", + "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")", + [](gpt_params & params, const std::string & mask) { + params.cpuparams.mask_valid = true; + if (!parse_cpu_mask(mask, params.cpuparams.cpumask)) { + throw std::invalid_argument("invalid cpumask"); + } + } + )); + add_opt(llama_arg( + {"-Cr", "--cpu-range"}, "lo-hi", + "range of CPUs for affinity. Complements --cpu-mask", + [](gpt_params & params, const std::string & range) { + params.cpuparams.mask_valid = true; + if (!parse_cpu_range(range, params.cpuparams.cpumask)) { + throw std::invalid_argument("invalid range"); + } + } + )); + add_opt(llama_arg( + {"--cpu-strict"}, "<0|1>", + format("use strict CPU placement (default: %u)\n", (unsigned) params.cpuparams.strict_cpu), + [](gpt_params & params, const std::string & value) { + params.cpuparams.strict_cpu = std::stoul(value); + } + )); + add_opt(llama_arg( + {"--prio"}, "N", + format("set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority), + [](gpt_params & params, int prio) { + if (prio < 0 || prio > 3) { + throw std::invalid_argument("invalid value"); + } + params.cpuparams.priority = (enum ggml_sched_priority) prio; + } + )); + add_opt(llama_arg( + {"--poll"}, "<0...100>", + format("use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll), + [](gpt_params & params, const std::string & value) { + params.cpuparams.poll = std::stoul(value); + } + )); + add_opt(llama_arg( + {"-Cb", "--cpu-mask-batch"}, "M", + "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)", + [](gpt_params & params, const std::string & mask) { + params.cpuparams_batch.mask_valid = true; + if (!parse_cpu_mask(mask, params.cpuparams_batch.cpumask)) { + throw std::invalid_argument("invalid cpumask"); + } + } + )); + add_opt(llama_arg( + {"-Crb", "--cpu-range-batch"}, "lo-hi", + "ranges of CPUs for affinity. Complements --cpu-mask-batch", + [](gpt_params & params, const std::string & range) { + params.cpuparams_batch.mask_valid = true; + if (!parse_cpu_range(range, params.cpuparams_batch.cpumask)) { + throw std::invalid_argument("invalid range"); + } + } + )); + add_opt(llama_arg( + {"--cpu-strict-batch"}, "<0|1>", + "use strict CPU placement (default: same as --cpu-strict)", + [](gpt_params & params, int value) { + params.cpuparams_batch.strict_cpu = value; + } + )); + add_opt(llama_arg( + {"--prio-batch"}, "N", + format("set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams_batch.priority), + [](gpt_params & params, int prio) { + if (prio < 0 || prio > 3) { + throw std::invalid_argument("invalid value"); + } + params.cpuparams_batch.priority = (enum ggml_sched_priority) prio; + } + )); + add_opt(llama_arg( + {"--poll-batch"}, "<0|1>", + "use polling to wait for work (default: same as --poll)", + [](gpt_params & params, int value) { + params.cpuparams_batch.poll = value; + } + )); + add_opt(llama_arg( + {"-Cd", "--cpu-mask-draft"}, "M", + "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)", + [](gpt_params & params, const std::string & mask) { + params.draft_cpuparams.mask_valid = true; + if (!parse_cpu_mask(mask, params.draft_cpuparams.cpumask)) { + throw std::invalid_argument("invalid cpumask"); + } + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-Crd", "--cpu-range-draft"}, "lo-hi", + "Ranges of CPUs for affinity. Complements --cpu-mask-draft", + [](gpt_params & params, const std::string & range) { + params.draft_cpuparams.mask_valid = true; + if (!parse_cpu_range(range, params.draft_cpuparams.cpumask)) { + throw std::invalid_argument("invalid range"); + } + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--cpu-strict-draft"}, "<0|1>", + "Use strict CPU placement for draft model (default: same as --cpu-strict)", + [](gpt_params & params, int value) { + params.draft_cpuparams.strict_cpu = value; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--prio-draft"}, "N", + format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams.priority), + [](gpt_params & params, int prio) { + if (prio < 0 || prio > 3) { + throw std::invalid_argument("invalid value"); + } + params.draft_cpuparams.priority = (enum ggml_sched_priority) prio; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--poll-draft"}, "<0|1>", + "Use polling to wait for draft model work (default: same as --poll])", + [](gpt_params & params, int value) { + params.draft_cpuparams.poll = value; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-Cbd", "--cpu-mask-batch-draft"}, "M", + "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)", + [](gpt_params & params, const std::string & mask) { + params.draft_cpuparams_batch.mask_valid = true; + if (!parse_cpu_mask(mask, params.draft_cpuparams_batch.cpumask)) { + throw std::invalid_argument("invalid cpumask"); + } + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-Crbd", "--cpu-range-batch-draft"}, "lo-hi", + "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)", + [](gpt_params & params, const std::string & range) { + params.draft_cpuparams_batch.mask_valid = true; + if (!parse_cpu_range(range, params.draft_cpuparams_batch.cpumask)) { + throw std::invalid_argument("invalid cpumask"); + } + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--cpu-strict-batch-draft"}, "<0|1>", + "Use strict CPU placement for draft model (default: --cpu-strict-draft)", + [](gpt_params & params, int value) { + params.draft_cpuparams_batch.strict_cpu = value; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--prio-batch-draft"}, "N", + format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams_batch.priority), + [](gpt_params & params, int prio) { + if (prio < 0 || prio > 3) { + throw std::invalid_argument("invalid value"); + } + params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) prio; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--poll-batch-draft"}, "<0|1>", + "Use polling to wait for draft model work (default: --poll-draft)", + [](gpt_params & params, int value) { + params.draft_cpuparams_batch.poll = value; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"--draft"}, "N", + format("number of tokens to draft for speculative decoding (default: %d)", params.n_draft), + [](gpt_params & params, int value) { + params.n_draft = value; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP})); + add_opt(llama_arg( + {"-ps", "--p-split"}, "N", + format("speculative decoding split probability (default: %.1f)", (double)params.p_split), + [](gpt_params & params, const std::string & value) { + params.p_split = std::stof(value); + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-lcs", "--lookup-cache-static"}, "FNAME", + "path to static lookup cache to use for lookup decoding (not updated by generation)", + [](gpt_params & params, const std::string & value) { + params.lookup_cache_static = value; + } + ).set_examples({LLAMA_EXAMPLE_LOOKUP})); + add_opt(llama_arg( + {"-lcd", "--lookup-cache-dynamic"}, "FNAME", + "path to dynamic lookup cache to use for lookup decoding (updated by generation)", + [](gpt_params & params, const std::string & value) { + params.lookup_cache_dynamic = value; + } + ).set_examples({LLAMA_EXAMPLE_LOOKUP})); + add_opt(llama_arg( + {"-c", "--ctx-size"}, "N", + format("size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx), + [](gpt_params & params, int value) { + params.n_ctx = value; + } + ).set_env("LLAMA_ARG_CTX_SIZE")); + add_opt(llama_arg( + {"-n", "--predict", "--n-predict"}, "N", + format("number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict), + [](gpt_params & params, int value) { + params.n_predict = value; + } + ).set_env("LLAMA_ARG_N_PREDICT")); + add_opt(llama_arg( + {"-b", "--batch-size"}, "N", + format("logical maximum batch size (default: %d)", params.n_batch), + [](gpt_params & params, int value) { + params.n_batch = value; + } + ).set_env("LLAMA_ARG_BATCH")); + add_opt(llama_arg( + {"-ub", "--ubatch-size"}, "N", + format("physical maximum batch size (default: %d)", params.n_ubatch), + [](gpt_params & params, int value) { + params.n_ubatch = value; + } + ).set_env("LLAMA_ARG_UBATCH")); + add_opt(llama_arg( + {"--keep"}, "N", + format("number of tokens to keep from the initial prompt (default: %d, -1 = all)", params.n_keep), + [](gpt_params & params, int value) { + params.n_keep = value; + } + )); + add_opt(llama_arg( + {"--chunks"}, "N", + format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks), + [](gpt_params & params, int value) { + params.n_chunks = value; + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_RETRIEVAL})); + add_opt(llama_arg( + {"-fa", "--flash-attn"}, + format("enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled"), + [](gpt_params & params) { + params.flash_attn = true; + } + ).set_env("LLAMA_ARG_FLASH_ATTN")); + add_opt(llama_arg( + {"-p", "--prompt"}, "PROMPT", + ex == LLAMA_EXAMPLE_MAIN + ? "prompt to start generation with\nif -cnv is set, this will be used as system prompt" + : "prompt to start generation with", + [](gpt_params & params, const std::string & value) { + params.prompt = value; + } + )); + add_opt(llama_arg( + {"-f", "--file"}, "FNAME", + "a file containing the prompt (default: none)", + [](gpt_params & params, const std::string & value) { + std::ifstream file(value); + if (!file) { + throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); + } + // store the external file name in params + params.prompt_file = value; + std::copy(std::istreambuf_iterator(file), std::istreambuf_iterator(), back_inserter(params.prompt)); + if (!params.prompt.empty() && params.prompt.back() == '\n') { + params.prompt.pop_back(); + } + } + )); + add_opt(llama_arg( + {"--in-file"}, "FNAME", + "an input file (repeat to specify multiple files)", + [](gpt_params & params, const std::string & value) { + std::ifstream file(value); + if (!file) { + throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); + } + params.in_files.push_back(value); + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX})); + add_opt(llama_arg( + {"-bf", "--binary-file"}, "FNAME", + "binary file containing the prompt (default: none)", + [](gpt_params & params, const std::string & value) { + std::ifstream file(value, std::ios::binary); + if (!file) { + throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); + } + // store the external file name in params + params.prompt_file = value; + std::ostringstream ss; + ss << file.rdbuf(); + params.prompt = ss.str(); + fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), value.c_str()); + } + )); + add_opt(llama_arg( + {"-e", "--escape"}, + format("process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false"), + [](gpt_params & params) { + params.escape = true; + } + )); + add_opt(llama_arg( + {"--no-escape"}, + "do not process escape sequences", + [](gpt_params & params) { + params.escape = false; + } + )); + add_opt(llama_arg( + {"-ptc", "--print-token-count"}, "N", + format("print token count every N tokens (default: %d)", params.n_print), + [](gpt_params & params, int value) { + params.n_print = value; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--prompt-cache"}, "FNAME", + "file to cache prompt state for faster startup (default: none)", + [](gpt_params & params, const std::string & value) { + params.path_prompt_cache = value; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--prompt-cache-all"}, + "if specified, saves user input and generations to cache as well\n", + [](gpt_params & params) { + params.prompt_cache_all = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--prompt-cache-ro"}, + "if specified, uses the prompt cache but does not update it", + [](gpt_params & params) { + params.prompt_cache_ro = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"-r", "--reverse-prompt"}, "PROMPT", + "halt generation at PROMPT, return control in interactive mode\n", + [](gpt_params & params, const std::string & value) { + params.antiprompt.emplace_back(value); + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"-sp", "--special"}, + format("special tokens output enabled (default: %s)", params.special ? "true" : "false"), + [](gpt_params & params) { + params.special = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"-cnv", "--conversation"}, + format( + "run in conversation mode:\n" + "- does not print special tokens and suffix/prefix\n" + "- interactive mode is also enabled\n" + "(default: %s)", + params.conversation ? "true" : "false" + ), + [](gpt_params & params) { + params.conversation = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"-i", "--interactive"}, + format("run in interactive mode (default: %s)", params.interactive ? "true" : "false"), + [](gpt_params & params) { + params.interactive = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"-if", "--interactive-first"}, + format("run in interactive mode and wait for input right away (default: %s)", params.interactive_first ? "true" : "false"), + [](gpt_params & params) { + params.interactive_first = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"-mli", "--multiline-input"}, + "allows you to write or paste multiple lines without ending each in '\\'", + [](gpt_params & params) { + params.multiline_input = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--in-prefix-bos"}, + "prefix BOS to user inputs, preceding the `--in-prefix` string", + [](gpt_params & params) { + params.input_prefix_bos = true; + params.enable_chat_template = false; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--in-prefix"}, "STRING", + "string to prefix user inputs with (default: empty)", + [](gpt_params & params, const std::string & value) { + params.input_prefix = value; + params.enable_chat_template = false; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--in-suffix"}, "STRING", + "string to suffix after user inputs with (default: empty)", + [](gpt_params & params, const std::string & value) { + params.input_suffix = value; + params.enable_chat_template = false; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--no-warmup"}, + "skip warming up the model with an empty run", + [](gpt_params & params) { + params.warmup = false; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(llama_arg( + {"--spm-infill"}, + format( + "use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: %s)", + params.spm_infill ? "enabled" : "disabled" + ), + [](gpt_params & params) { + params.spm_infill = true; + } + ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_INFILL})); + add_opt(llama_arg( + {"--samplers"}, "SAMPLERS", + format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()), + [](gpt_params & params, const std::string & value) { + const auto sampler_names = string_split(value, ';'); + params.sparams.samplers = gpt_sampler_types_from_names(sampler_names, true); + } + ).set_sparam()); + add_opt(llama_arg( + {"-s", "--seed"}, "SEED", + format("RNG seed (default: %d, use random seed for < 0)", params.sparams.seed), + [](gpt_params & params, const std::string & value) { + params.sparams.seed = std::stoul(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--sampling-seq"}, "SEQUENCE", + format("simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str()), + [](gpt_params & params, const std::string & value) { + params.sparams.samplers = gpt_sampler_types_from_chars(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--ignore-eos"}, + "ignore end of stream token and continue generating (implies --logit-bias EOS-inf)", + [](gpt_params & params) { + params.sparams.ignore_eos = true; + } + ).set_sparam()); + add_opt(llama_arg( + {"--penalize-nl"}, + format("penalize newline tokens (default: %s)", params.sparams.penalize_nl ? "true" : "false"), + [](gpt_params & params) { + params.sparams.penalize_nl = true; + } + ).set_sparam()); + add_opt(llama_arg( + {"--temp"}, "N", + format("temperature (default: %.1f)", (double)params.sparams.temp), + [](gpt_params & params, const std::string & value) { + params.sparams.temp = std::stof(value); + params.sparams.temp = std::max(params.sparams.temp, 0.0f); + } + ).set_sparam()); + add_opt(llama_arg( + {"--top-k"}, "N", + format("top-k sampling (default: %d, 0 = disabled)", params.sparams.top_k), + [](gpt_params & params, int value) { + params.sparams.top_k = value; + } + ).set_sparam()); + add_opt(llama_arg( + {"--top-p"}, "N", + format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sparams.top_p), + [](gpt_params & params, const std::string & value) { + params.sparams.top_p = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--min-p"}, "N", + format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sparams.min_p), + [](gpt_params & params, const std::string & value) { + params.sparams.min_p = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--tfs"}, "N", + format("tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)params.sparams.tfs_z), + [](gpt_params & params, const std::string & value) { + params.sparams.tfs_z = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--typical"}, "N", + format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sparams.typ_p), + [](gpt_params & params, const std::string & value) { + params.sparams.typ_p = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--repeat-last-n"}, "N", + format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sparams.penalty_last_n), + [](gpt_params & params, int value) { + params.sparams.penalty_last_n = value; + params.sparams.n_prev = std::max(params.sparams.n_prev, params.sparams.penalty_last_n); + } + ).set_sparam()); + add_opt(llama_arg( + {"--repeat-penalty"}, "N", + format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sparams.penalty_repeat), + [](gpt_params & params, const std::string & value) { + params.sparams.penalty_repeat = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--presence-penalty"}, "N", + format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sparams.penalty_present), + [](gpt_params & params, const std::string & value) { + params.sparams.penalty_present = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--frequency-penalty"}, "N", + format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sparams.penalty_freq), + [](gpt_params & params, const std::string & value) { + params.sparams.penalty_freq = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--dynatemp-range"}, "N", + format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sparams.dynatemp_range), + [](gpt_params & params, const std::string & value) { + params.sparams.dynatemp_range = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--dynatemp-exp"}, "N", + format("dynamic temperature exponent (default: %.1f)", (double)params.sparams.dynatemp_exponent), + [](gpt_params & params, const std::string & value) { + params.sparams.dynatemp_exponent = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--mirostat"}, "N", + format("use Mirostat sampling.\nTop K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n" + "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sparams.mirostat), + [](gpt_params & params, int value) { + params.sparams.mirostat = value; + } + ).set_sparam()); + add_opt(llama_arg( + {"--mirostat-lr"}, "N", + format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sparams.mirostat_eta), + [](gpt_params & params, const std::string & value) { + params.sparams.mirostat_eta = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"--mirostat-ent"}, "N", + format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sparams.mirostat_tau), + [](gpt_params & params, const std::string & value) { + params.sparams.mirostat_tau = std::stof(value); + } + ).set_sparam()); + add_opt(llama_arg( + {"-l", "--logit-bias"}, "TOKEN_ID(+/-)BIAS", + "modifies the likelihood of token appearing in the completion,\n" + "i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n" + "or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'", + [](gpt_params & params, const std::string & value) { + std::stringstream ss(value); + llama_token key; + char sign; + std::string value_str; + try { + if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) { + const float bias = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f); + params.sparams.logit_bias.push_back({key, bias}); + } else { + throw std::invalid_argument("invalid input format"); + } + } catch (const std::exception&) { + throw std::invalid_argument("invalid input format"); + } + } + ).set_sparam()); + add_opt(llama_arg( + {"--grammar"}, "GRAMMAR", + format("BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", params.sparams.grammar.c_str()), + [](gpt_params & params, const std::string & value) { + params.sparams.grammar = value; + } + ).set_sparam()); + add_opt(llama_arg( + {"--grammar-file"}, "FNAME", + "file to read grammar from", + [](gpt_params & params, const std::string & value) { + std::ifstream file(value); + if (!file) { + throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); + } + std::copy( + std::istreambuf_iterator(file), + std::istreambuf_iterator(), + std::back_inserter(params.sparams.grammar) + ); + } + ).set_sparam()); + add_opt(llama_arg( + {"-j", "--json-schema"}, "SCHEMA", + "JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead", + [](gpt_params & params, const std::string & value) { + params.sparams.grammar = json_schema_to_grammar(json::parse(value)); + } + ).set_sparam()); + add_opt(llama_arg( + {"--pooling"}, "{none,mean,cls,last}", + "pooling type for embeddings, use model default if unspecified", + [](gpt_params & params, const std::string & value) { + /**/ if (value == "none") { params.pooling_type = LLAMA_POOLING_TYPE_NONE; } + else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; } + else if (value == "cls") { params.pooling_type = LLAMA_POOLING_TYPE_CLS; } + else if (value == "last") { params.pooling_type = LLAMA_POOLING_TYPE_LAST; } + else { throw std::invalid_argument("invalid value"); } + } + ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); + add_opt(llama_arg( + {"--attention"}, "{causal,non,causal}", + "attention type for embeddings, use model default if unspecified", + [](gpt_params & params, const std::string & value) { + /**/ if (value == "causal") { params.attention_type = LLAMA_ATTENTION_TYPE_CAUSAL; } + else if (value == "non-causal") { params.attention_type = LLAMA_ATTENTION_TYPE_NON_CAUSAL; } + else { throw std::invalid_argument("invalid value"); } + } + ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); + add_opt(llama_arg( + {"--rope-scaling"}, "{none,linear,yarn}", + "RoPE frequency scaling method, defaults to linear unless specified by the model", + [](gpt_params & params, const std::string & value) { + /**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; } + else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; } + else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; } + else { throw std::invalid_argument("invalid value"); } + } + )); + add_opt(llama_arg( + {"--rope-scale"}, "N", + "RoPE context scaling factor, expands context by a factor of N", + [](gpt_params & params, const std::string & value) { + params.rope_freq_scale = 1.0f / std::stof(value); + } + )); + add_opt(llama_arg( + {"--rope-freq-base"}, "N", + "RoPE base frequency, used by NTK-aware scaling (default: loaded from model)", + [](gpt_params & params, const std::string & value) { + params.rope_freq_base = std::stof(value); + } + )); + add_opt(llama_arg( + {"--rope-freq-scale"}, "N", + "RoPE frequency scaling factor, expands context by a factor of 1/N", + [](gpt_params & params, const std::string & value) { + params.rope_freq_scale = std::stof(value); + } + )); + add_opt(llama_arg( + {"--yarn-orig-ctx"}, "N", + format("YaRN: original context size of model (default: %d = model training context size)", params.yarn_orig_ctx), + [](gpt_params & params, int value) { + params.yarn_orig_ctx = value; + } + )); + add_opt(llama_arg( + {"--yarn-ext-factor"}, "N", + format("YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor), + [](gpt_params & params, const std::string & value) { + params.yarn_ext_factor = std::stof(value); + } + )); + add_opt(llama_arg( + {"--yarn-attn-factor"}, "N", + format("YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor), + [](gpt_params & params, const std::string & value) { + params.yarn_attn_factor = std::stof(value); + } + )); + add_opt(llama_arg( + {"--yarn-beta-slow"}, "N", + format("YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow), + [](gpt_params & params, const std::string & value) { + params.yarn_beta_slow = std::stof(value); + } + )); + add_opt(llama_arg( + {"--yarn-beta-fast"}, "N", + format("YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast), + [](gpt_params & params, const std::string & value) { + params.yarn_beta_fast = std::stof(value); + } + )); + add_opt(llama_arg( + {"-gan", "--grp-attn-n"}, "N", + format("group-attention factor (default: %d)", params.grp_attn_n), + [](gpt_params & params, int value) { + params.grp_attn_n = value; + } + )); + add_opt(llama_arg( + {"-gaw", "--grp-attn-w"}, "N", + format("group-attention width (default: %.1f)", (double)params.grp_attn_w), + [](gpt_params & params, int value) { + params.grp_attn_w = value; + } + )); + add_opt(llama_arg( + {"-dkvc", "--dump-kv-cache"}, + "verbose print of the KV cache", + [](gpt_params & params) { + params.dump_kv_cache = true; + } + )); + add_opt(llama_arg( + {"-nkvo", "--no-kv-offload"}, + "disable KV offload", + [](gpt_params & params) { + params.no_kv_offload = true; + } + )); + add_opt(llama_arg( + {"-ctk", "--cache-type-k"}, "TYPE", + format("KV cache data type for K (default: %s)", params.cache_type_k.c_str()), + [](gpt_params & params, const std::string & value) { + // TODO: get the type right here + params.cache_type_k = value; + } + )); + add_opt(llama_arg( + {"-ctv", "--cache-type-v"}, "TYPE", + format("KV cache data type for V (default: %s)", params.cache_type_v.c_str()), + [](gpt_params & params, const std::string & value) { + // TODO: get the type right here + params.cache_type_v = value; + } + )); + add_opt(llama_arg( + {"--perplexity", "--all-logits"}, + format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"), + [](gpt_params & params) { + params.logits_all = true; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--hellaswag"}, + "compute HellaSwag score over random tasks from datafile supplied with -f", + [](gpt_params & params) { + params.hellaswag = true; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--hellaswag-tasks"}, "N", + format("number of tasks to use when computing the HellaSwag score (default: %zu)", params.hellaswag_tasks), + [](gpt_params & params, int value) { + params.hellaswag_tasks = value; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--winogrande"}, + "compute Winogrande score over random tasks from datafile supplied with -f", + [](gpt_params & params) { + params.winogrande = true; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--winogrande-tasks"}, "N", + format("number of tasks to use when computing the Winogrande score (default: %zu)", params.winogrande_tasks), + [](gpt_params & params, int value) { + params.winogrande_tasks = value; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--multiple-choice"}, + "compute multiple choice score over random tasks from datafile supplied with -f", + [](gpt_params & params) { + params.multiple_choice = true; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--multiple-choice-tasks"}, "N", + format("number of tasks to use when computing the multiple choice score (default: %zu)", params.multiple_choice_tasks), + [](gpt_params & params, int value) { + params.multiple_choice_tasks = value; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--kl-divergence"}, + "computes KL-divergence to logits provided via --kl-divergence-base", + [](gpt_params & params) { + params.kl_divergence = true; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--save-all-logits", "--kl-divergence-base"}, "FNAME", + "set logits file", + [](gpt_params & params, const std::string & value) { + params.logits_file = value; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--ppl-stride"}, "N", + format("stride for perplexity calculation (default: %d)", params.ppl_stride), + [](gpt_params & params, int value) { + params.ppl_stride = value; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"--ppl-output-type"}, "<0|1>", + format("output type for perplexity calculation (default: %d)", params.ppl_output_type), + [](gpt_params & params, int value) { + params.ppl_output_type = value; + } + ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); + add_opt(llama_arg( + {"-dt", "--defrag-thold"}, "N", + format("KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold), + [](gpt_params & params, const std::string & value) { + params.defrag_thold = std::stof(value); + } + ).set_env("LLAMA_ARG_DEFRAG_THOLD")); + add_opt(llama_arg( + {"-np", "--parallel"}, "N", + format("number of parallel sequences to decode (default: %d)", params.n_parallel), + [](gpt_params & params, int value) { + params.n_parallel = value; + } + )); + add_opt(llama_arg( + {"-ns", "--sequences"}, "N", + format("number of sequences to decode (default: %d)", params.n_sequences), + [](gpt_params & params, int value) { + params.n_sequences = value; + } + ).set_examples({LLAMA_EXAMPLE_PARALLEL})); + add_opt(llama_arg( + {"-cb", "--cont-batching"}, + format("enable continuous batching (a.k.a dynamic batching) (default: %s)", params.cont_batching ? "enabled" : "disabled"), + [](gpt_params & params) { + params.cont_batching = true; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CONT_BATCHING")); + add_opt(llama_arg( + {"-nocb", "--no-cont-batching"}, + "disable continuous batching", + [](gpt_params & params) { + params.cont_batching = false; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING")); + add_opt(llama_arg( + {"--mmproj"}, "FILE", + "path to a multimodal projector file for LLaVA. see examples/llava/README.md", + [](gpt_params & params, const std::string & value) { + params.mmproj = value; + } + ).set_examples({LLAMA_EXAMPLE_LLAVA})); + add_opt(llama_arg( + {"--image"}, "FILE", + "path to an image file. use with multimodal models. Specify multiple times for batching", + [](gpt_params & params, const std::string & value) { + params.image.emplace_back(value); + } + ).set_examples({LLAMA_EXAMPLE_LLAVA})); +#ifdef GGML_USE_RPC + add_opt(llama_arg( + {"--rpc"}, "SERVERS", + "comma separated list of RPC servers", + [](gpt_params & params, const std::string & value) { + params.rpc_servers = value; + } + )); +#endif + add_opt(llama_arg( + {"--mlock"}, + "force system to keep model in RAM rather than swapping or compressing", + [](gpt_params & params) { + params.use_mlock = true; + } + )); + add_opt(llama_arg( + {"--no-mmap"}, + "do not memory-map model (slower load but may reduce pageouts if not using mlock)", + [](gpt_params & params) { + params.use_mmap = false; + } + )); + add_opt(llama_arg( + {"--numa"}, "TYPE", + "attempt optimizations that help on some NUMA systems\n" + "- distribute: spread execution evenly over all nodes\n" + "- isolate: only spawn threads on CPUs on the node that execution started on\n" + "- numactl: use the CPU map provided by numactl\n" + "if run without this previously, it is recommended to drop the system page cache before using this\n" + "see https://github.com/ggerganov/llama.cpp/issues/1437", + [](gpt_params & params, const std::string & value) { + /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } + else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } + else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; } + else { throw std::invalid_argument("invalid value"); } + } + )); + add_opt(llama_arg( + {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N", + "number of layers to store in VRAM", + [](gpt_params & params, int value) { + params.n_gpu_layers = value; + if (!llama_supports_gpu_offload()) { + fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers option will be ignored\n"); + fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); + } + } + ).set_env("LLAMA_ARG_N_GPU_LAYERS")); + add_opt(llama_arg( + {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N", + "number of layers to store in VRAM for the draft model", + [](gpt_params & params, int value) { + params.n_gpu_layers_draft = value; + if (!llama_supports_gpu_offload()) { + fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n"); + fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); + } + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-sm", "--split-mode"}, "{none,layer,row}", + "how to split the model across multiple GPUs, one of:\n" + "- none: use one GPU only\n" + "- layer (default): split layers and KV across GPUs\n" + "- row: split rows across GPUs", + [](gpt_params & params, const std::string & value) { + std::string arg_next = value; + if (arg_next == "none") { + params.split_mode = LLAMA_SPLIT_MODE_NONE; + } else if (arg_next == "layer") { + params.split_mode = LLAMA_SPLIT_MODE_LAYER; + } + else if (arg_next == "row") { +#ifdef GGML_USE_SYCL + fprintf(stderr, "warning: The split mode value:[row] is not supported by llama.cpp with SYCL. It's developing.\nExit!\n"); + exit(1); +#endif // GGML_USE_SYCL + params.split_mode = LLAMA_SPLIT_MODE_ROW; + } + else { + throw std::invalid_argument("invalid value"); + } +#ifndef GGML_USE_CUDA_SYCL_VULKAN + fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting the split mode has no effect.\n"); +#endif // GGML_USE_CUDA_SYCL_VULKAN + } + )); + add_opt(llama_arg( + {"-ts", "--tensor-split"}, "N0,N1,N2,...", + "fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1", + [](gpt_params & params, const std::string & value) { + std::string arg_next = value; + + // split string by , and / + const std::regex regex{ R"([,/]+)" }; + std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 }; + std::vector split_arg{ it, {} }; + if (split_arg.size() >= llama_max_devices()) { + throw std::invalid_argument( + format("got %d input configs, but system only has %d devices", (int)split_arg.size(), (int)llama_max_devices()) + ); + } + for (size_t i = 0; i < llama_max_devices(); ++i) { + if (i < split_arg.size()) { + params.tensor_split[i] = std::stof(split_arg[i]); + } else { + params.tensor_split[i] = 0.0f; + } + } +#ifndef GGML_USE_CUDA_SYCL_VULKAN + fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting a tensor split has no effect.\n"); +#endif // GGML_USE_CUDA_SYCL_VULKAN + } + )); + add_opt(llama_arg( + {"-mg", "--main-gpu"}, "INDEX", + format("the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: %d)", params.main_gpu), + [](gpt_params & params, int value) { + params.main_gpu = value; +#ifndef GGML_USE_CUDA_SYCL_VULKAN + fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting the main GPU has no effect.\n"); +#endif // GGML_USE_CUDA_SYCL_VULKAN + } + )); + add_opt(llama_arg( + {"--check-tensors"}, + format("check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false"), + [](gpt_params & params) { + params.check_tensors = true; + } + )); + add_opt(llama_arg( + {"--override-kv"}, "KEY=TYPE:VALUE", + "advanced option to override model metadata by key. may be specified multiple times.\n" + "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false", + [](gpt_params & params, const std::string & value) { + if (!string_parse_kv_override(value.c_str(), params.kv_overrides)) { + throw std::runtime_error(format("error: Invalid type for KV override: %s\n", value.c_str())); + } + } + )); + add_opt(llama_arg( + {"--lora"}, "FNAME", + "path to LoRA adapter (can be repeated to use multiple adapters)", + [](gpt_params & params, const std::string & value) { + params.lora_adapters.push_back({ std::string(value), 1.0 }); + } + // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA})); + add_opt(llama_arg( + {"--lora-scaled"}, "FNAME", "SCALE", + "path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters)", + [](gpt_params & params, const std::string & fname, const std::string & scale) { + params.lora_adapters.push_back({ fname, std::stof(scale) }); + } + // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA})); + add_opt(llama_arg( + {"--control-vector"}, "FNAME", + "add a control vector\nnote: this argument can be repeated to add multiple control vectors", + [](gpt_params & params, const std::string & value) { + params.control_vectors.push_back({ 1.0f, value, }); + } + )); + add_opt(llama_arg( + {"--control-vector-scaled"}, "FNAME", "SCALE", + "add a control vector with user defined scaling SCALE\n" + "note: this argument can be repeated to add multiple scaled control vectors", + [](gpt_params & params, const std::string & fname, const std::string & scale) { + params.control_vectors.push_back({ std::stof(scale), fname }); + } + )); + add_opt(llama_arg( + {"--control-vector-layer-range"}, "START", "END", + "layer range to apply the control vector(s) to, start and end inclusive", + [](gpt_params & params, const std::string & start, const std::string & end) { + params.control_vector_layer_start = std::stoi(start); + params.control_vector_layer_end = std::stoi(end); + } + )); + add_opt(llama_arg( + {"-a", "--alias"}, "STRING", + "set alias for model name (to be used by REST API)", + [](gpt_params & params, const std::string & value) { + params.model_alias = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"-m", "--model"}, "FNAME", + ex == LLAMA_EXAMPLE_EXPORT_LORA + ? std::string("model path from which to load base model") + : format( + "model path (default: `models/$filename` with filename from `--hf-file` " + "or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH + ), + [](gpt_params & params, const std::string & value) { + params.model = value; + } + ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL")); + add_opt(llama_arg( + {"-md", "--model-draft"}, "FNAME", + "draft model for speculative decoding (default: unused)", + [](gpt_params & params, const std::string & value) { + params.model_draft = value; + } + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + add_opt(llama_arg( + {"-mu", "--model-url"}, "MODEL_URL", + "model download url (default: unused)", + [](gpt_params & params, const std::string & value) { + params.model_url = value; + } + ).set_env("LLAMA_ARG_MODEL_URL")); + add_opt(llama_arg( + {"-hfr", "--hf-repo"}, "REPO", + "Hugging Face model repository (default: unused)", + [](gpt_params & params, const std::string & value) { + params.hf_repo = value; + } + ).set_env("LLAMA_ARG_HF_REPO")); + add_opt(llama_arg( + {"-hff", "--hf-file"}, "FILE", + "Hugging Face model file (default: unused)", + [](gpt_params & params, const std::string & value) { + params.hf_file = value; + } + ).set_env("LLAMA_ARG_HF_FILE")); + add_opt(llama_arg( + {"-hft", "--hf-token"}, "TOKEN", + "Hugging Face access token (default: value from HF_TOKEN environment variable)", + [](gpt_params & params, const std::string & value) { + params.hf_token = value; + } + ).set_env("HF_TOKEN")); + add_opt(llama_arg( + {"--context-file"}, "FNAME", + "file to load context from (repeat to specify multiple files)", + [](gpt_params & params, const std::string & value) { + std::ifstream file(value, std::ios::binary); + if (!file) { + throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); + } + params.context_files.push_back(value); + } + ).set_examples({LLAMA_EXAMPLE_RETRIEVAL})); + add_opt(llama_arg( + {"--chunk-size"}, "N", + format("minimum length of embedded text chunks (default: %d)", params.chunk_size), + [](gpt_params & params, int value) { + params.chunk_size = value; + } + ).set_examples({LLAMA_EXAMPLE_RETRIEVAL})); + add_opt(llama_arg( + {"--chunk-separator"}, "STRING", + format("separator between chunks (default: '%s')", params.chunk_separator.c_str()), + [](gpt_params & params, const std::string & value) { + params.chunk_separator = value; + } + ).set_examples({LLAMA_EXAMPLE_RETRIEVAL})); + add_opt(llama_arg( + {"--junk"}, "N", + format("number of times to repeat the junk text (default: %d)", params.n_junk), + [](gpt_params & params, int value) { + params.n_junk = value; + } + ).set_examples({LLAMA_EXAMPLE_PASSKEY})); + add_opt(llama_arg( + {"--pos"}, "N", + format("position of the passkey in the junk text (default: %d)", params.i_pos), + [](gpt_params & params, int value) { + params.i_pos = value; + } + ).set_examples({LLAMA_EXAMPLE_PASSKEY})); + add_opt(llama_arg( + {"-o", "--output", "--output-file"}, "FNAME", + format("output file (default: '%s')", + ex == LLAMA_EXAMPLE_EXPORT_LORA + ? params.lora_outfile.c_str() + : ex == LLAMA_EXAMPLE_CVECTOR_GENERATOR + ? params.cvector_outfile.c_str() + : params.out_file.c_str()), + [](gpt_params & params, const std::string & value) { + params.out_file = value; + params.cvector_outfile = value; + params.lora_outfile = value; + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA})); + add_opt(llama_arg( + {"-ofreq", "--output-frequency"}, "N", + format("output the imatrix every N iterations (default: %d)", params.n_out_freq), + [](gpt_params & params, int value) { + params.n_out_freq = value; + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX})); + add_opt(llama_arg( + {"--save-frequency"}, "N", + format("save an imatrix copy every N iterations (default: %d)", params.n_save_freq), + [](gpt_params & params, int value) { + params.n_save_freq = value; + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX})); + add_opt(llama_arg( + {"--process-output"}, + format("collect data for the output tensor (default: %s)", params.process_output ? "true" : "false"), + [](gpt_params & params) { + params.process_output = true; + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX})); + add_opt(llama_arg( + {"--no-ppl"}, + format("do not compute perplexity (default: %s)", params.compute_ppl ? "true" : "false"), + [](gpt_params & params) { + params.compute_ppl = false; + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX})); + add_opt(llama_arg( + {"--chunk", "--from-chunk"}, "N", + format("start processing the input from chunk N (default: %d)", params.i_chunk), + [](gpt_params & params, int value) { + params.i_chunk = value; + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX})); + add_opt(llama_arg( + {"-pps"}, + format("is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false"), + [](gpt_params & params) { + params.is_pp_shared = true; + } + ).set_examples({LLAMA_EXAMPLE_BENCH})); + add_opt(llama_arg( + {"-npp"}, "n0,n1,...", + "number of prompt tokens", + [](gpt_params & params, const std::string & value) { + auto p = string_split(value, ','); + params.n_pp.insert(params.n_pp.end(), p.begin(), p.end()); + } + ).set_examples({LLAMA_EXAMPLE_BENCH})); + add_opt(llama_arg( + {"-ntg"}, "n0,n1,...", + "number of text generation tokens", + [](gpt_params & params, const std::string & value) { + auto p = string_split(value, ','); + params.n_tg.insert(params.n_tg.end(), p.begin(), p.end()); + } + ).set_examples({LLAMA_EXAMPLE_BENCH})); + add_opt(llama_arg( + {"-npl"}, "n0,n1,...", + "number of parallel prompts", + [](gpt_params & params, const std::string & value) { + auto p = string_split(value, ','); + params.n_pl.insert(params.n_pl.end(), p.begin(), p.end()); + } + ).set_examples({LLAMA_EXAMPLE_BENCH})); + add_opt(llama_arg( + {"--embd-normalize"}, "N", + format("normalisation for embendings (default: %d) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)", params.embd_normalize), + [](gpt_params & params, int value) { + params.embd_normalize = value; + } + ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); + add_opt(llama_arg( + {"--embd-output-format"}, "FORMAT", + "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix", + [](gpt_params & params, const std::string & value) { + params.embd_out = value; + } + ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); + add_opt(llama_arg( + {"--embd-separator"}, "STRING", + "separator of embendings (default \\n) for example \"<#sep#>\"", + [](gpt_params & params, const std::string & value) { + params.embd_sep = value; + } + ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); + add_opt(llama_arg( + {"--host"}, "HOST", + format("ip address to listen (default: %s)", params.hostname.c_str()), + [](gpt_params & params, const std::string & value) { + params.hostname = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_HOST")); + add_opt(llama_arg( + {"--port"}, "PORT", + format("port to listen (default: %d)", params.port), + [](gpt_params & params, int value) { + params.port = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_PORT")); + add_opt(llama_arg( + {"--path"}, "PATH", + format("path to serve static files from (default: %s)", params.public_path.c_str()), + [](gpt_params & params, const std::string & value) { + params.public_path = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--embedding", "--embeddings"}, + format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"), + [](gpt_params & params) { + params.embedding = true; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS")); + add_opt(llama_arg( + {"--api-key"}, "KEY", + "API key to use for authentication (default: none)", + [](gpt_params & params, const std::string & value) { + params.api_keys.push_back(value); + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_API_KEY")); + add_opt(llama_arg( + {"--api-key-file"}, "FNAME", + "path to file containing API keys (default: none)", + [](gpt_params & params, const std::string & value) { + std::ifstream key_file(value); + if (!key_file) { + throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); + } + std::string key; + while (std::getline(key_file, key)) { + if (!key.empty()) { + params.api_keys.push_back(key); + } + } + key_file.close(); + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--ssl-key-file"}, "FNAME", + "path to file a PEM-encoded SSL private key", + [](gpt_params & params, const std::string & value) { + params.ssl_file_key = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--ssl-cert-file"}, "FNAME", + "path to file a PEM-encoded SSL certificate", + [](gpt_params & params, const std::string & value) { + params.ssl_file_cert = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"-to", "--timeout"}, "N", + format("server read/write timeout in seconds (default: %d)", params.timeout_read), + [](gpt_params & params, int value) { + params.timeout_read = value; + params.timeout_write = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--threads-http"}, "N", + format("number of threads used to process HTTP requests (default: %d)", params.n_threads_http), + [](gpt_params & params, int value) { + params.n_threads_http = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP")); + add_opt(llama_arg( + {"-spf", "--system-prompt-file"}, "FNAME", + "set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications", + [](gpt_params & params, const std::string & value) { + std::ifstream file(value); + if (!file) { + throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); + } + std::string system_prompt; + std::copy( + std::istreambuf_iterator(file), + std::istreambuf_iterator(), + std::back_inserter(system_prompt) + ); + params.system_prompt = system_prompt; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--log-format"}, "{text, json}", + "log output format: json or text (default: json)", + [](gpt_params & params, const std::string & value) { + if (value == "json") { + params.log_json = true; + } else if (value == "text") { + params.log_json = false; + } else { + throw std::invalid_argument("invalid value"); + } + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--metrics"}, + format("enable prometheus compatible metrics endpoint (default: %s)", params.endpoint_metrics ? "enabled" : "disabled"), + [](gpt_params & params) { + params.endpoint_metrics = true; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_METRICS")); + add_opt(llama_arg( + {"--no-slots"}, + format("disables slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled"), + [](gpt_params & params) { + params.endpoint_slots = false; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_ENDPOINT_SLOTS")); + add_opt(llama_arg( + {"--slot-save-path"}, "PATH", + "path to save slot kv cache (default: disabled)", + [](gpt_params & params, const std::string & value) { + params.slot_save_path = value; + // if doesn't end with DIRECTORY_SEPARATOR, add it + if (!params.slot_save_path.empty() && params.slot_save_path[params.slot_save_path.size() - 1] != DIRECTORY_SEPARATOR) { + params.slot_save_path += DIRECTORY_SEPARATOR; + } + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--chat-template"}, "JINJA_TEMPLATE", + "set custom jinja chat template (default: template taken from model's metadata)\n" + "if suffix/prefix are specified, template will be disabled\n" + "only commonly used templates are accepted:\nhttps://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template", + [](gpt_params & params, const std::string & value) { + if (!llama_chat_verify_template(value)) { + throw std::runtime_error(format( + "error: the supplied chat template is not supported: %s\n" + "note: llama.cpp does not use jinja parser, we only support commonly used templates\n", + value.c_str() + )); + } + params.chat_template = value; + } + ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE")); + add_opt(llama_arg( + {"-sps", "--slot-prompt-similarity"}, "SIMILARITY", + format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity), + [](gpt_params & params, const std::string & value) { + params.slot_prompt_similarity = std::stof(value); + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--lora-init-without-apply"}, + format("load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"), + [](gpt_params & params) { + params.lora_init_without_apply = true; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(llama_arg( + {"--simple-io"}, + "use basic IO for better compatibility in subprocesses and limited consoles", + [](gpt_params & params) { + params.simple_io = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL})); + add_opt(llama_arg( + {"-ld", "--logdir"}, "LOGDIR", + "path under which to save YAML logs (no logging if unset)", + [](gpt_params & params, const std::string & value) { + params.logdir = value; + + if (params.logdir.back() != DIRECTORY_SEPARATOR) { + params.logdir += DIRECTORY_SEPARATOR; + } + } + )); + add_opt(llama_arg( + {"--positive-file"}, "FNAME", + format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()), + [](gpt_params & params, const std::string & value) { + params.cvector_positive_file = value; + } + ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); + add_opt(llama_arg( + {"--negative-file"}, "FNAME", + format("negative prompts file, one prompt per line (default: '%s')", params.cvector_negative_file.c_str()), + [](gpt_params & params, const std::string & value) { + params.cvector_negative_file = value; + } + ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); + add_opt(llama_arg( + {"--pca-batch"}, "N", + format("batch size used for PCA. Larger batch runs faster, but uses more memory (default: %d)", params.n_pca_batch), + [](gpt_params & params, int value) { + params.n_pca_batch = value; + } + ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); + add_opt(llama_arg( + {"--pca-iter"}, "N", + format("number of iterations used for PCA (default: %d)", params.n_pca_iterations), + [](gpt_params & params, int value) { + params.n_pca_iterations = value; + } + ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); + add_opt(llama_arg( + {"--method"}, "{pca, mean}", + "dimensionality reduction method to be used (default: pca)", + [](gpt_params & params, const std::string & value) { + /**/ if (value == "pca") { params.cvector_dimre_method = DIMRE_METHOD_PCA; } + else if (value == "mean") { params.cvector_dimre_method = DIMRE_METHOD_MEAN; } + else { throw std::invalid_argument("invalid value"); } + } + ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); + add_opt(llama_arg( + {"--output-format"}, "{md,jsonl}", + "output format for batched-bench results (default: md)", + [](gpt_params & params, const std::string & value) { + /**/ if (value == "jsonl") { params.batched_bench_output_jsonl = true; } + else if (value == "md") { params.batched_bench_output_jsonl = false; } + else { std::invalid_argument("invalid value"); } + } + ).set_examples({LLAMA_EXAMPLE_BENCH})); +#ifndef LOG_DISABLE_LOGS + // TODO: make this looks less weird + add_opt(llama_arg( + {"--log-test"}, + "Log test", + [](gpt_params &) { log_param_single_parse("--log-test"); } + )); + add_opt(llama_arg( + {"--log-disable"}, + "Log disable", + [](gpt_params &) { log_param_single_parse("--log-disable"); } + )); + add_opt(llama_arg( + {"--log-enable"}, + "Log enable", + [](gpt_params &) { log_param_single_parse("--log-enable"); } + )); + add_opt(llama_arg( + {"--log-new"}, + "Log new", + [](gpt_params &) { log_param_single_parse("--log-new"); } + )); + add_opt(llama_arg( + {"--log-append"}, + "Log append", + [](gpt_params &) { log_param_single_parse("--log-append"); } + )); + add_opt(llama_arg( + {"--log-file"}, "FNAME", + "Log file", + [](gpt_params &, const std::string & value) { log_param_pair_parse(false, "--log-file", value); } + )); +#endif // LOG_DISABLE_LOGS + + return ctx_arg; +} + diff --git a/common/arg.h b/common/arg.h new file mode 100644 index 000000000..413de2c88 --- /dev/null +++ b/common/arg.h @@ -0,0 +1,77 @@ +#pragma once + +#include "common.h" + +#include +#include +#include + +// +// CLI argument parsing +// + +struct llama_arg { + std::set examples = {LLAMA_EXAMPLE_COMMON}; + std::vector args; + const char * value_hint = nullptr; // help text or example for arg value + const char * value_hint_2 = nullptr; // for second arg value + const char * env = nullptr; + std::string help; + bool is_sparam = false; // is current arg a sampling param? + void (*handler_void) (gpt_params & params) = nullptr; + void (*handler_string) (gpt_params & params, const std::string &) = nullptr; + void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr; + void (*handler_int) (gpt_params & params, int) = nullptr; + + llama_arg( + const std::initializer_list & args, + const char * value_hint, + const std::string & help, + void (*handler)(gpt_params & params, const std::string &) + ) : args(args), value_hint(value_hint), help(help), handler_string(handler) {} + + llama_arg( + const std::initializer_list & args, + const char * value_hint, + const std::string & help, + void (*handler)(gpt_params & params, int) + ) : args(args), value_hint(value_hint), help(help), handler_int(handler) {} + + llama_arg( + const std::initializer_list & args, + const std::string & help, + void (*handler)(gpt_params & params) + ) : args(args), help(help), handler_void(handler) {} + + // support 2 values for arg + llama_arg( + const std::initializer_list & args, + const char * value_hint, + const char * value_hint_2, + const std::string & help, + void (*handler)(gpt_params & params, const std::string &, const std::string &) + ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {} + + llama_arg & set_examples(std::initializer_list examples); + llama_arg & set_env(const char * env); + llama_arg & set_sparam(); + bool in_example(enum llama_example ex); + bool get_value_from_env(std::string & output); + bool has_value_from_env(); + std::string to_string(); +}; + +struct gpt_params_context { + enum llama_example ex = LLAMA_EXAMPLE_COMMON; + gpt_params & params; + std::vector options; + void(*print_usage)(int, char **) = nullptr; + gpt_params_context(gpt_params & params) : params(params) {} +}; + +// parse input arguments from CLI +// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message) +bool gpt_params_parse(int argc, char ** argv, gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr); + +// function to be used by test-arg-parser +gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr); diff --git a/common/common.cpp b/common/common.cpp index 916b1731e..5395eaa0e 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -25,7 +25,6 @@ #include #include #include -#include #if defined(__APPLE__) && defined(__MACH__) #include @@ -272,53 +271,6 @@ bool set_process_priority(enum ggml_sched_priority prio) { // CLI argument parsing // -#ifdef __GNUC__ -#ifdef __MINGW32__ -#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) -#else -#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) -#endif -#else -#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) -#endif - -LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2) -static std::string format(const char * fmt, ...) { - va_list ap; - va_list ap2; - va_start(ap, fmt); - va_copy(ap2, ap); - int size = vsnprintf(NULL, 0, fmt, ap); - GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT - std::vector buf(size + 1); - int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2); - GGML_ASSERT(size2 == size); - va_end(ap2); - va_end(ap); - return std::string(buf.data(), size); -} - -static void gpt_params_handle_model_default(gpt_params & params) { - if (!params.hf_repo.empty()) { - // short-hand to avoid specifying --hf-file -> default it to --model - if (params.hf_file.empty()) { - if (params.model.empty()) { - throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n"); - } - params.hf_file = params.model; - } else if (params.model.empty()) { - params.model = fs_get_cache_file(string_split(params.hf_file, '/').back()); - } - } else if (!params.model_url.empty()) { - if (params.model.empty()) { - auto f = string_split(params.model_url, '#').front(); - f = string_split(f, '?').front(); - params.model = fs_get_cache_file(string_split(f, '/').back()); - } - } else if (params.model.empty()) { - params.model = DEFAULT_MODEL_PATH; - } -} void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) { int32_t n_set = 0; @@ -344,150 +296,6 @@ void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) } } -bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params, std::vector & options) { - std::string arg; - const std::string arg_prefix = "--"; - gpt_sampler_params & sparams = params.sparams; - - std::unordered_map arg_to_options; - for (auto & opt : options) { - for (const auto & arg : opt.args) { - arg_to_options[arg] = &opt; - } - } - - // handle environment variables - for (auto & opt : options) { - std::string value; - if (opt.get_value_from_env(value)) { - try { - if (opt.handler_void && (value == "1" || value == "true")) { - opt.handler_void(params); - } - if (opt.handler_int) { - opt.handler_int(params, std::stoi(value)); - } - if (opt.handler_string) { - opt.handler_string(params, value); - continue; - } - } catch (std::exception & e) { - throw std::invalid_argument(format( - "error while handling environment variable \"%s\": %s\n\n", opt.env, e.what())); - } - } - } - - // handle command line arguments - auto check_arg = [&](int i) { - if (i+1 >= argc) { - throw std::invalid_argument("expected value for argument"); - } - }; - - for (int i = 1; i < argc; i++) { - const std::string arg_prefix = "--"; - - std::string arg = argv[i]; - if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) { - std::replace(arg.begin(), arg.end(), '_', '-'); - } - if (arg_to_options.find(arg) == arg_to_options.end()) { - throw std::invalid_argument(format("error: invalid argument: %s", arg.c_str())); - } - auto opt = *arg_to_options[arg]; - if (opt.has_value_from_env()) { - fprintf(stderr, "warn: %s environment variable is set, but will be overwritten by command line argument %s\n", opt.env, arg.c_str()); - } - try { - if (opt.handler_void) { - opt.handler_void(params); - continue; - } - - // arg with single value - check_arg(i); - std::string val = argv[++i]; - if (opt.handler_int) { - opt.handler_int(params, std::stoi(val)); - continue; - } - if (opt.handler_string) { - opt.handler_string(params, val); - continue; - } - - // arg with 2 values - check_arg(i); - std::string val2 = argv[++i]; - if (opt.handler_str_str) { - opt.handler_str_str(params, val, val2); - continue; - } - } catch (std::exception & e) { - throw std::invalid_argument(format( - "error while handling argument \"%s\": %s\n\n" - "usage:\n%s\n\nto show complete usage, run with -h", - arg.c_str(), e.what(), arg_to_options[arg]->to_string().c_str())); - } - } - - postprocess_cpu_params(params.cpuparams, nullptr); - postprocess_cpu_params(params.cpuparams_batch, ¶ms.cpuparams); - postprocess_cpu_params(params.draft_cpuparams, ¶ms.cpuparams); - postprocess_cpu_params(params.draft_cpuparams_batch, ¶ms.cpuparams_batch); - - if (params.prompt_cache_all && (params.interactive || params.interactive_first)) { - throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); - } - - gpt_params_handle_model_default(params); - - if (params.escape) { - string_process_escapes(params.prompt); - string_process_escapes(params.input_prefix); - string_process_escapes(params.input_suffix); - for (auto & antiprompt : params.antiprompt) { - string_process_escapes(antiprompt); - } - } - - if (!params.kv_overrides.empty()) { - params.kv_overrides.emplace_back(); - params.kv_overrides.back().key[0] = 0; - } - - if (sparams.seed == LLAMA_DEFAULT_SEED) { - sparams.seed = time(NULL); - } - - return true; -} - -bool gpt_params_parse(int argc, char ** argv, gpt_params & params, std::vector & options) { - const auto params_org = params; // the example can modify the default params - - try { - if (!gpt_params_parse_ex(argc, argv, params, options)) { - params = params_org; - return false; - } - if (params.usage) { - gpt_params_print_usage(params, options); - if (params.print_usage) { - params.print_usage(argc, argv); - } - exit(0); - } - } catch (const std::invalid_argument & ex) { - fprintf(stderr, "%s\n", ex.what()); - params = params_org; - return false; - } - - return true; -} - bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) { size_t dash_loc = range.find('-'); if (dash_loc == std::string::npos) { @@ -561,1743 +369,6 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD return true; } -static std::vector break_str_into_lines(std::string input, size_t max_char_per_line) { - std::vector result; - std::istringstream iss(input); - std::string line; - auto add_line = [&](const std::string& l) { - if (l.length() <= max_char_per_line) { - result.push_back(l); - } else { - std::istringstream line_stream(l); - std::string word, current_line; - while (line_stream >> word) { - if (current_line.length() + !current_line.empty() + word.length() > max_char_per_line) { - if (!current_line.empty()) result.push_back(current_line); - current_line = word; - } else { - current_line += (!current_line.empty() ? " " : "") + word; - } - } - if (!current_line.empty()) result.push_back(current_line); - } - }; - while (std::getline(iss, line)) { - add_line(line); - } - return result; -} - -std::string llama_arg::to_string() { - // params for printing to console - const static int n_leading_spaces = 40; - const static int n_char_per_line_help = 70; // TODO: detect this based on current console - std::string leading_spaces(n_leading_spaces, ' '); - - std::ostringstream ss; - for (const auto arg : args) { - if (arg == args.front()) { - if (args.size() == 1) { - ss << arg; - } else { - // first arg is usually abbreviation, we need padding to make it more beautiful - auto tmp = std::string(arg) + ", "; - ss << format("%-7s", tmp.c_str()); - } - } else { - ss << arg << (arg != args.back() ? ", " : ""); - } - } - if (value_hint) ss << " " << value_hint; - if (value_hint_2) ss << " " << value_hint_2; - if (ss.tellp() > n_leading_spaces - 3) { - // current line is too long, add new line - ss << "\n" << leading_spaces; - } else { - // padding between arg and help, same line - ss << std::string(leading_spaces.size() - ss.tellp(), ' '); - } - const auto help_lines = break_str_into_lines(help, n_char_per_line_help); - for (const auto & line : help_lines) { - ss << (&line == &help_lines.front() ? "" : leading_spaces) << line << "\n"; - } - return ss.str(); -} - -void gpt_params_print_usage(gpt_params & params, std::vector & options) { - auto print_options = [](std::vector & options) { - for (llama_arg * opt : options) { - printf("%s", opt->to_string().c_str()); - } - }; - - std::vector common_options; - std::vector specific_options; - for (auto & opt : options) { - // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example - if (opt.in_example(params.curr_ex)) { - specific_options.push_back(&opt); - } else { - common_options.push_back(&opt); - } - } - printf("----- common options -----\n\n"); - print_options(common_options); - // TODO: maybe convert enum llama_example to string - printf("\n\n----- example-specific options -----\n\n"); - print_options(specific_options); -} - -std::vector gpt_params_parser_init(gpt_params & params, llama_example ex) { - return gpt_params_parser_init(params, ex, nullptr); -} - -std::vector gpt_params_parser_init(gpt_params & params, llama_example ex, std::function print_usage) { - std::vector options; - params.print_usage = print_usage; - params.curr_ex = ex; - - std::string sampler_type_chars; - std::string sampler_type_names; - for (const auto & sampler : params.sparams.samplers) { - sampler_type_chars += gpt_sampler_type_to_chr(sampler); - sampler_type_names += gpt_sampler_type_to_str(sampler) + ";"; - } - sampler_type_names.pop_back(); - - - /** - * filter options by example - * rules: - * - all examples inherit options from LLAMA_EXAMPLE_COMMON - * - if LLAMA_EXAMPLE_* is set (other than COMMON), we only show the option in the corresponding example - * - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example - */ - auto add_opt = [&](llama_arg arg) { - if (arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) { - options.push_back(std::move(arg)); - } - }; - - - add_opt(llama_arg( - {"-h", "--help", "--usage"}, - "print usage and exit", - [](gpt_params & params) { - params.usage = true; - } - )); - add_opt(llama_arg( - {"--version"}, - "show version and build info", - [](gpt_params &) { - fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT); - fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET); - exit(0); - } - )); - add_opt(llama_arg( - {"-v", "--verbose"}, - "print verbose information", - [](gpt_params & params) { - params.verbosity = 1; - } - )); - add_opt(llama_arg( - {"--verbosity"}, "N", - format("set specific verbosity level (default: %d)", params.verbosity), - [](gpt_params & params, int value) { - params.verbosity = value; - } - )); - add_opt(llama_arg( - {"--verbose-prompt"}, - format("print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false"), - [](gpt_params & params) { - params.verbose_prompt = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--no-display-prompt"}, - format("don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false"), - [](gpt_params & params) { - params.display_prompt = false; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"-co", "--color"}, - format("colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false"), - [](gpt_params & params) { - params.use_color = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL})); - add_opt(llama_arg( - {"-s", "--seed"}, "SEED", - format("RNG seed (default: %d, use random seed for < 0)", params.sparams.seed), - [](gpt_params & params, const std::string & value) { - params.sparams.seed = std::stoul(value); - } - )); - add_opt(llama_arg( - {"-t", "--threads"}, "N", - format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads), - [](gpt_params & params, int value) { - params.cpuparams.n_threads = value; - if (params.cpuparams.n_threads <= 0) { - params.cpuparams.n_threads = std::thread::hardware_concurrency(); - } - } - ).set_env("LLAMA_ARG_THREADS")); - add_opt(llama_arg( - {"-tb", "--threads-batch"}, "N", - "number of threads to use during batch and prompt processing (default: same as --threads)", - [](gpt_params & params, int value) { - params.cpuparams_batch.n_threads = value; - if (params.cpuparams_batch.n_threads <= 0) { - params.cpuparams_batch.n_threads = std::thread::hardware_concurrency(); - } - } - )); - add_opt(llama_arg( - {"-td", "--threads-draft"}, "N", - "number of threads to use during generation (default: same as --threads)", - [](gpt_params & params, int value) { - params.draft_cpuparams.n_threads = value; - if (params.draft_cpuparams.n_threads <= 0) { - params.draft_cpuparams.n_threads = std::thread::hardware_concurrency(); - } - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-tbd", "--threads-batch-draft"}, "N", - "number of threads to use during batch and prompt processing (default: same as --threads-draft)", - [](gpt_params & params, int value) { - params.draft_cpuparams_batch.n_threads = value; - if (params.draft_cpuparams_batch.n_threads <= 0) { - params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency(); - } - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-C", "--cpu-mask"}, "M", - "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")", - [](gpt_params & params, const std::string & mask) { - params.cpuparams.mask_valid = true; - if (!parse_cpu_mask(mask, params.cpuparams.cpumask)) { - throw std::invalid_argument("invalid cpumask"); - } - } - )); - add_opt(llama_arg( - {"-Cr", "--cpu-range"}, "lo-hi", - "range of CPUs for affinity. Complements --cpu-mask", - [](gpt_params & params, const std::string & range) { - params.cpuparams.mask_valid = true; - if (!parse_cpu_range(range, params.cpuparams.cpumask)) { - throw std::invalid_argument("invalid range"); - } - } - )); - add_opt(llama_arg( - {"--cpu-strict"}, "<0|1>", - format("use strict CPU placement (default: %u)\n", (unsigned) params.cpuparams.strict_cpu), - [](gpt_params & params, const std::string & value) { - params.cpuparams.strict_cpu = std::stoul(value); - } - )); - add_opt(llama_arg( - {"--prio"}, "N", - format("set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority), - [](gpt_params & params, int prio) { - if (prio < 0 || prio > 3) { - throw std::invalid_argument("invalid value"); - } - params.cpuparams.priority = (enum ggml_sched_priority) prio; - } - )); - add_opt(llama_arg( - {"--poll"}, "<0...100>", - format("use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll), - [](gpt_params & params, const std::string & value) { - params.cpuparams.poll = std::stoul(value); - } - )); - add_opt(llama_arg( - {"-Cb", "--cpu-mask-batch"}, "M", - "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)", - [](gpt_params & params, const std::string & mask) { - params.cpuparams_batch.mask_valid = true; - if (!parse_cpu_mask(mask, params.cpuparams_batch.cpumask)) { - throw std::invalid_argument("invalid cpumask"); - } - } - )); - add_opt(llama_arg( - {"-Crb", "--cpu-range-batch"}, "lo-hi", - "ranges of CPUs for affinity. Complements --cpu-mask-batch", - [](gpt_params & params, const std::string & range) { - params.cpuparams_batch.mask_valid = true; - if (!parse_cpu_range(range, params.cpuparams_batch.cpumask)) { - throw std::invalid_argument("invalid range"); - } - } - )); - add_opt(llama_arg( - {"--cpu-strict-batch"}, "<0|1>", - "use strict CPU placement (default: same as --cpu-strict)", - [](gpt_params & params, int value) { - params.cpuparams_batch.strict_cpu = value; - } - )); - add_opt(llama_arg( - {"--prio-batch"}, "N", - format("set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams_batch.priority), - [](gpt_params & params, int prio) { - if (prio < 0 || prio > 3) { - throw std::invalid_argument("invalid value"); - } - params.cpuparams_batch.priority = (enum ggml_sched_priority) prio; - } - )); - add_opt(llama_arg( - {"--poll-batch"}, "<0|1>", - "use polling to wait for work (default: same as --poll)", - [](gpt_params & params, int value) { - params.cpuparams_batch.poll = value; - } - )); - add_opt(llama_arg( - {"-Cd", "--cpu-mask-draft"}, "M", - "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)", - [](gpt_params & params, const std::string & mask) { - params.draft_cpuparams.mask_valid = true; - if (!parse_cpu_mask(mask, params.draft_cpuparams.cpumask)) { - throw std::invalid_argument("invalid cpumask"); - } - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-Crd", "--cpu-range-draft"}, "lo-hi", - "Ranges of CPUs for affinity. Complements --cpu-mask-draft", - [](gpt_params & params, const std::string & range) { - params.draft_cpuparams.mask_valid = true; - if (!parse_cpu_range(range, params.draft_cpuparams.cpumask)) { - throw std::invalid_argument("invalid range"); - } - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"--cpu-strict-draft"}, "<0|1>", - "Use strict CPU placement for draft model (default: same as --cpu-strict)", - [](gpt_params & params, int value) { - params.draft_cpuparams.strict_cpu = value; - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"--prio-draft"}, "N", - format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams.priority), - [](gpt_params & params, int prio) { - if (prio < 0 || prio > 3) { - throw std::invalid_argument("invalid value"); - } - params.draft_cpuparams.priority = (enum ggml_sched_priority) prio; - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"--poll-draft"}, "<0|1>", - "Use polling to wait for draft model work (default: same as --poll])", - [](gpt_params & params, int value) { - params.draft_cpuparams.poll = value; - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-Cbd", "--cpu-mask-batch-draft"}, "M", - "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)", - [](gpt_params & params, const std::string & mask) { - params.draft_cpuparams_batch.mask_valid = true; - if (!parse_cpu_mask(mask, params.draft_cpuparams_batch.cpumask)) { - throw std::invalid_argument("invalid cpumask"); - } - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-Crbd", "--cpu-range-batch-draft"}, "lo-hi", - "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)", - [](gpt_params & params, const std::string & range) { - params.draft_cpuparams_batch.mask_valid = true; - if (!parse_cpu_range(range, params.draft_cpuparams_batch.cpumask)) { - throw std::invalid_argument("invalid cpumask"); - } - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"--cpu-strict-batch-draft"}, "<0|1>", - "Use strict CPU placement for draft model (default: --cpu-strict-draft)", - [](gpt_params & params, int value) { - params.draft_cpuparams_batch.strict_cpu = value; - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"--prio-batch-draft"}, "N", - format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams_batch.priority), - [](gpt_params & params, int prio) { - if (prio < 0 || prio > 3) { - throw std::invalid_argument("invalid value"); - } - params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) prio; - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"--poll-batch-draft"}, "<0|1>", - "Use polling to wait for draft model work (default: --poll-draft)", - [](gpt_params & params, int value) { - params.draft_cpuparams_batch.poll = value; - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"--draft"}, "N", - format("number of tokens to draft for speculative decoding (default: %d)", params.n_draft), - [](gpt_params & params, int value) { - params.n_draft = value; - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-ps", "--p-split"}, "N", - format("speculative decoding split probability (default: %.1f)", (double)params.p_split), - [](gpt_params & params, const std::string & value) { - params.p_split = std::stof(value); - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-lcs", "--lookup-cache-static"}, "FNAME", - "path to static lookup cache to use for lookup decoding (not updated by generation)", - [](gpt_params & params, const std::string & value) { - params.lookup_cache_static = value; - } - )); - add_opt(llama_arg( - {"-lcd", "--lookup-cache-dynamic"}, "FNAME", - "path to dynamic lookup cache to use for lookup decoding (updated by generation)", - [](gpt_params & params, const std::string & value) { - params.lookup_cache_dynamic = value; - } - )); - add_opt(llama_arg( - {"-c", "--ctx-size"}, "N", - format("size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx), - [](gpt_params & params, int value) { - params.n_ctx = value; - } - ).set_env("LLAMA_ARG_CTX_SIZE")); - add_opt(llama_arg( - {"-n", "--predict", "--n-predict"}, "N", - format("number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict), - [](gpt_params & params, int value) { - params.n_predict = value; - } - ).set_env("LLAMA_ARG_N_PREDICT")); - add_opt(llama_arg( - {"-b", "--batch-size"}, "N", - format("logical maximum batch size (default: %d)", params.n_batch), - [](gpt_params & params, int value) { - params.n_batch = value; - } - ).set_env("LLAMA_ARG_BATCH")); - add_opt(llama_arg( - {"-ub", "--ubatch-size"}, "N", - format("physical maximum batch size (default: %d)", params.n_ubatch), - [](gpt_params & params, int value) { - params.n_ubatch = value; - } - ).set_env("LLAMA_ARG_UBATCH")); - add_opt(llama_arg( - {"--keep"}, "N", - format("number of tokens to keep from the initial prompt (default: %d, -1 = all)", params.n_keep), - [](gpt_params & params, int value) { - params.n_keep = value; - } - )); - add_opt(llama_arg( - {"--chunks"}, "N", - format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks), - [](gpt_params & params, int value) { - params.n_chunks = value; - } - )); - add_opt(llama_arg( - {"-fa", "--flash-attn"}, - format("enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled"), - [](gpt_params & params) { - params.flash_attn = true; - } - ).set_env("LLAMA_ARG_FLASH_ATTN")); - add_opt(llama_arg( - {"-p", "--prompt"}, "PROMPT", - ex == LLAMA_EXAMPLE_MAIN - ? "prompt to start generation with\nif -cnv is set, this will be used as system prompt" - : "prompt to start generation with", - [](gpt_params & params, const std::string & value) { - params.prompt = value; - } - )); - add_opt(llama_arg( - {"-f", "--file"}, "FNAME", - "a file containing the prompt (default: none)", - [](gpt_params & params, const std::string & value) { - std::ifstream file(value); - if (!file) { - throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); - } - // store the external file name in params - params.prompt_file = value; - std::copy(std::istreambuf_iterator(file), std::istreambuf_iterator(), back_inserter(params.prompt)); - if (!params.prompt.empty() && params.prompt.back() == '\n') { - params.prompt.pop_back(); - } - } - )); - add_opt(llama_arg( - {"--in-file"}, "FNAME", - "an input file (repeat to specify multiple files)", - [](gpt_params & params, const std::string & value) { - std::ifstream file(value); - if (!file) { - throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); - } - params.in_files.push_back(value); - } - )); - add_opt(llama_arg( - {"-bf", "--binary-file"}, "FNAME", - "binary file containing the prompt (default: none)", - [](gpt_params & params, const std::string & value) { - std::ifstream file(value, std::ios::binary); - if (!file) { - throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); - } - // store the external file name in params - params.prompt_file = value; - std::ostringstream ss; - ss << file.rdbuf(); - params.prompt = ss.str(); - fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), value.c_str()); - } - )); - add_opt(llama_arg( - {"-e", "--escape"}, - format("process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false"), - [](gpt_params & params) { - params.escape = true; - } - )); - add_opt(llama_arg( - {"--no-escape"}, - "do not process escape sequences", - [](gpt_params & params) { - params.escape = false; - } - )); - add_opt(llama_arg( - {"-ptc", "--print-token-count"}, "N", - format("print token count every N tokens (default: %d)", params.n_print), - [](gpt_params & params, int value) { - params.n_print = value; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--prompt-cache"}, "FNAME", - "file to cache prompt state for faster startup (default: none)", - [](gpt_params & params, const std::string & value) { - params.path_prompt_cache = value; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--prompt-cache-all"}, - "if specified, saves user input and generations to cache as well\n", - [](gpt_params & params) { - params.prompt_cache_all = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--prompt-cache-ro"}, - "if specified, uses the prompt cache but does not update it", - [](gpt_params & params) { - params.prompt_cache_ro = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"-r", "--reverse-prompt"}, "PROMPT", - "halt generation at PROMPT, return control in interactive mode\n", - [](gpt_params & params, const std::string & value) { - params.antiprompt.emplace_back(value); - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"-sp", "--special"}, - format("special tokens output enabled (default: %s)", params.special ? "true" : "false"), - [](gpt_params & params) { - params.special = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"-cnv", "--conversation"}, - format( - "run in conversation mode:\n" - "- does not print special tokens and suffix/prefix\n" - "- interactive mode is also enabled\n" - "(default: %s)", - params.conversation ? "true" : "false" - ), - [](gpt_params & params) { - params.conversation = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"-i", "--interactive"}, - format("run in interactive mode (default: %s)", params.interactive ? "true" : "false"), - [](gpt_params & params) { - params.interactive = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"-if", "--interactive-first"}, - format("run in interactive mode and wait for input right away (default: %s)", params.interactive_first ? "true" : "false"), - [](gpt_params & params) { - params.interactive_first = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"-mli", "--multiline-input"}, - "allows you to write or paste multiple lines without ending each in '\\'", - [](gpt_params & params) { - params.multiline_input = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--in-prefix-bos"}, - "prefix BOS to user inputs, preceding the `--in-prefix` string", - [](gpt_params & params) { - params.input_prefix_bos = true; - params.enable_chat_template = false; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--in-prefix"}, "STRING", - "string to prefix user inputs with (default: empty)", - [](gpt_params & params, const std::string & value) { - params.input_prefix = value; - params.enable_chat_template = false; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--in-suffix"}, "STRING", - "string to suffix after user inputs with (default: empty)", - [](gpt_params & params, const std::string & value) { - params.input_suffix = value; - params.enable_chat_template = false; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--no-warmup"}, - "skip warming up the model with an empty run", - [](gpt_params & params) { - params.warmup = false; - } - ).set_examples({LLAMA_EXAMPLE_MAIN})); - add_opt(llama_arg( - {"--spm-infill"}, - format( - "use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: %s)", - params.spm_infill ? "enabled" : "disabled" - ), - [](gpt_params & params) { - params.spm_infill = true; - } - ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_INFILL})); - add_opt(llama_arg( - {"--samplers"}, "SAMPLERS", - format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()), - [](gpt_params & params, const std::string & value) { - const auto sampler_names = string_split(value, ';'); - params.sparams.samplers = gpt_sampler_types_from_names(sampler_names, true); - } - )); - add_opt(llama_arg( - {"--sampling-seq"}, "SEQUENCE", - format("simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str()), - [](gpt_params & params, const std::string & value) { - params.sparams.samplers = gpt_sampler_types_from_chars(value); - } - )); - add_opt(llama_arg( - {"--ignore-eos"}, - "ignore end of stream token and continue generating (implies --logit-bias EOS-inf)", - [](gpt_params & params) { - params.sparams.ignore_eos = true; - } - )); - add_opt(llama_arg( - {"--penalize-nl"}, - format("penalize newline tokens (default: %s)", params.sparams.penalize_nl ? "true" : "false"), - [](gpt_params & params) { - params.sparams.penalize_nl = true; - } - )); - add_opt(llama_arg( - {"--temp"}, "N", - format("temperature (default: %.1f)", (double)params.sparams.temp), - [](gpt_params & params, const std::string & value) { - params.sparams.temp = std::stof(value); - params.sparams.temp = std::max(params.sparams.temp, 0.0f); - } - )); - add_opt(llama_arg( - {"--top-k"}, "N", - format("top-k sampling (default: %d, 0 = disabled)", params.sparams.top_k), - [](gpt_params & params, int value) { - params.sparams.top_k = value; - } - )); - add_opt(llama_arg( - {"--top-p"}, "N", - format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sparams.top_p), - [](gpt_params & params, const std::string & value) { - params.sparams.top_p = std::stof(value); - } - )); - add_opt(llama_arg( - {"--min-p"}, "N", - format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sparams.min_p), - [](gpt_params & params, const std::string & value) { - params.sparams.min_p = std::stof(value); - } - )); - add_opt(llama_arg( - {"--tfs"}, "N", - format("tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)params.sparams.tfs_z), - [](gpt_params & params, const std::string & value) { - params.sparams.tfs_z = std::stof(value); - } - )); - add_opt(llama_arg( - {"--typical"}, "N", - format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sparams.typ_p), - [](gpt_params & params, const std::string & value) { - params.sparams.typ_p = std::stof(value); - } - )); - add_opt(llama_arg( - {"--repeat-last-n"}, "N", - format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sparams.penalty_last_n), - [](gpt_params & params, int value) { - params.sparams.penalty_last_n = value; - params.sparams.n_prev = std::max(params.sparams.n_prev, params.sparams.penalty_last_n); - } - )); - add_opt(llama_arg( - {"--repeat-penalty"}, "N", - format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sparams.penalty_repeat), - [](gpt_params & params, const std::string & value) { - params.sparams.penalty_repeat = std::stof(value); - } - )); - add_opt(llama_arg( - {"--presence-penalty"}, "N", - format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sparams.penalty_present), - [](gpt_params & params, const std::string & value) { - params.sparams.penalty_present = std::stof(value); - } - )); - add_opt(llama_arg( - {"--frequency-penalty"}, "N", - format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sparams.penalty_freq), - [](gpt_params & params, const std::string & value) { - params.sparams.penalty_freq = std::stof(value); - } - )); - add_opt(llama_arg( - {"--dynatemp-range"}, "N", - format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sparams.dynatemp_range), - [](gpt_params & params, const std::string & value) { - params.sparams.dynatemp_range = std::stof(value); - } - )); - add_opt(llama_arg( - {"--dynatemp-exp"}, "N", - format("dynamic temperature exponent (default: %.1f)", (double)params.sparams.dynatemp_exponent), - [](gpt_params & params, const std::string & value) { - params.sparams.dynatemp_exponent = std::stof(value); - } - )); - add_opt(llama_arg( - {"--mirostat"}, "N", - format("use Mirostat sampling.\nTop K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n" - "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sparams.mirostat), - [](gpt_params & params, int value) { - params.sparams.mirostat = value; - } - )); - add_opt(llama_arg( - {"--mirostat-lr"}, "N", - format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sparams.mirostat_eta), - [](gpt_params & params, const std::string & value) { - params.sparams.mirostat_eta = std::stof(value); - } - )); - add_opt(llama_arg( - {"--mirostat-ent"}, "N", - format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sparams.mirostat_tau), - [](gpt_params & params, const std::string & value) { - params.sparams.mirostat_tau = std::stof(value); - } - )); - add_opt(llama_arg( - {"-l", "--logit-bias"}, "TOKEN_ID(+/-)BIAS", - "modifies the likelihood of token appearing in the completion,\n" - "i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n" - "or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'", - [](gpt_params & params, const std::string & value) { - std::stringstream ss(value); - llama_token key; - char sign; - std::string value_str; - try { - if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) { - const float bias = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f); - params.sparams.logit_bias.push_back({key, bias}); - } else { - throw std::invalid_argument("invalid input format"); - } - } catch (const std::exception&) { - throw std::invalid_argument("invalid input format"); - } - } - )); - add_opt(llama_arg( - {"--grammar"}, "GRAMMAR", - format("BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", params.sparams.grammar.c_str()), - [](gpt_params & params, const std::string & value) { - params.sparams.grammar = value; - } - )); - add_opt(llama_arg( - {"--grammar-file"}, "FNAME", - "file to read grammar from", - [](gpt_params & params, const std::string & value) { - std::ifstream file(value); - if (!file) { - throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); - } - std::copy( - std::istreambuf_iterator(file), - std::istreambuf_iterator(), - std::back_inserter(params.sparams.grammar) - ); - } - )); - add_opt(llama_arg( - {"-j", "--json-schema"}, "SCHEMA", - "JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead", - [](gpt_params & params, const std::string & value) { - params.sparams.grammar = json_schema_to_grammar(json::parse(value)); - } - )); - add_opt(llama_arg( - {"--pooling"}, "{none,mean,cls,last}", - "pooling type for embeddings, use model default if unspecified", - [](gpt_params & params, const std::string & value) { - /**/ if (value == "none") { params.pooling_type = LLAMA_POOLING_TYPE_NONE; } - else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; } - else if (value == "cls") { params.pooling_type = LLAMA_POOLING_TYPE_CLS; } - else if (value == "last") { params.pooling_type = LLAMA_POOLING_TYPE_LAST; } - else { throw std::invalid_argument("invalid value"); } - } - ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); - add_opt(llama_arg( - {"--attention"}, "{causal,non,causal}", - "attention type for embeddings, use model default if unspecified", - [](gpt_params & params, const std::string & value) { - /**/ if (value == "causal") { params.attention_type = LLAMA_ATTENTION_TYPE_CAUSAL; } - else if (value == "non-causal") { params.attention_type = LLAMA_ATTENTION_TYPE_NON_CAUSAL; } - else { throw std::invalid_argument("invalid value"); } - } - ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); - add_opt(llama_arg( - {"--rope-scaling"}, "{none,linear,yarn}", - "RoPE frequency scaling method, defaults to linear unless specified by the model", - [](gpt_params & params, const std::string & value) { - /**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; } - else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; } - else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; } - else { throw std::invalid_argument("invalid value"); } - } - )); - add_opt(llama_arg( - {"--rope-scale"}, "N", - "RoPE context scaling factor, expands context by a factor of N", - [](gpt_params & params, const std::string & value) { - params.rope_freq_scale = 1.0f / std::stof(value); - } - )); - add_opt(llama_arg( - {"--rope-freq-base"}, "N", - "RoPE base frequency, used by NTK-aware scaling (default: loaded from model)", - [](gpt_params & params, const std::string & value) { - params.rope_freq_base = std::stof(value); - } - )); - add_opt(llama_arg( - {"--rope-freq-scale"}, "N", - "RoPE frequency scaling factor, expands context by a factor of 1/N", - [](gpt_params & params, const std::string & value) { - params.rope_freq_scale = std::stof(value); - } - )); - add_opt(llama_arg( - {"--yarn-orig-ctx"}, "N", - format("YaRN: original context size of model (default: %d = model training context size)", params.yarn_orig_ctx), - [](gpt_params & params, int value) { - params.yarn_orig_ctx = value; - } - )); - add_opt(llama_arg( - {"--yarn-ext-factor"}, "N", - format("YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor), - [](gpt_params & params, const std::string & value) { - params.yarn_ext_factor = std::stof(value); - } - )); - add_opt(llama_arg( - {"--yarn-attn-factor"}, "N", - format("YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor), - [](gpt_params & params, const std::string & value) { - params.yarn_attn_factor = std::stof(value); - } - )); - add_opt(llama_arg( - {"--yarn-beta-slow"}, "N", - format("YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow), - [](gpt_params & params, const std::string & value) { - params.yarn_beta_slow = std::stof(value); - } - )); - add_opt(llama_arg( - {"--yarn-beta-fast"}, "N", - format("YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast), - [](gpt_params & params, const std::string & value) { - params.yarn_beta_fast = std::stof(value); - } - )); - add_opt(llama_arg( - {"-gan", "--grp-attn-n"}, "N", - format("group-attention factor (default: %d)", params.grp_attn_n), - [](gpt_params & params, int value) { - params.grp_attn_n = value; - } - )); - add_opt(llama_arg( - {"-gaw", "--grp-attn-w"}, "N", - format("group-attention width (default: %.1f)", (double)params.grp_attn_w), - [](gpt_params & params, int value) { - params.grp_attn_w = value; - } - )); - add_opt(llama_arg( - {"-dkvc", "--dump-kv-cache"}, - "verbose print of the KV cache", - [](gpt_params & params) { - params.dump_kv_cache = true; - } - )); - add_opt(llama_arg( - {"-nkvo", "--no-kv-offload"}, - "disable KV offload", - [](gpt_params & params) { - params.no_kv_offload = true; - } - )); - add_opt(llama_arg( - {"-ctk", "--cache-type-k"}, "TYPE", - format("KV cache data type for K (default: %s)", params.cache_type_k.c_str()), - [](gpt_params & params, const std::string & value) { - // TODO: get the type right here - params.cache_type_k = value; - } - )); - add_opt(llama_arg( - {"-ctv", "--cache-type-v"}, "TYPE", - format("KV cache data type for V (default: %s)", params.cache_type_v.c_str()), - [](gpt_params & params, const std::string & value) { - // TODO: get the type right here - params.cache_type_v = value; - } - )); - add_opt(llama_arg( - {"--perplexity", "--all-logits"}, - format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"), - [](gpt_params & params) { - params.logits_all = true; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--hellaswag"}, - "compute HellaSwag score over random tasks from datafile supplied with -f", - [](gpt_params & params) { - params.hellaswag = true; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--hellaswag-tasks"}, "N", - format("number of tasks to use when computing the HellaSwag score (default: %zu)", params.hellaswag_tasks), - [](gpt_params & params, int value) { - params.hellaswag_tasks = value; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--winogrande"}, - "compute Winogrande score over random tasks from datafile supplied with -f", - [](gpt_params & params) { - params.winogrande = true; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--winogrande-tasks"}, "N", - format("number of tasks to use when computing the Winogrande score (default: %zu)", params.winogrande_tasks), - [](gpt_params & params, int value) { - params.winogrande_tasks = value; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--multiple-choice"}, - "compute multiple choice score over random tasks from datafile supplied with -f", - [](gpt_params & params) { - params.multiple_choice = true; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--multiple-choice-tasks"}, "N", - format("number of tasks to use when computing the multiple choice score (default: %zu)", params.multiple_choice_tasks), - [](gpt_params & params, int value) { - params.multiple_choice_tasks = value; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--kl-divergence"}, - "computes KL-divergence to logits provided via --kl-divergence-base", - [](gpt_params & params) { - params.kl_divergence = true; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--save-all-logits", "--kl-divergence-base"}, "FNAME", - "set logits file", - [](gpt_params & params, const std::string & value) { - params.logits_file = value; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--ppl-stride"}, "N", - format("stride for perplexity calculation (default: %d)", params.ppl_stride), - [](gpt_params & params, int value) { - params.ppl_stride = value; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"--ppl-output-type"}, "<0|1>", - format("output type for perplexity calculation (default: %d)", params.ppl_output_type), - [](gpt_params & params, int value) { - params.ppl_output_type = value; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); - add_opt(llama_arg( - {"-dt", "--defrag-thold"}, "N", - format("KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold), - [](gpt_params & params, const std::string & value) { - params.defrag_thold = std::stof(value); - } - ).set_env("LLAMA_ARG_DEFRAG_THOLD")); - add_opt(llama_arg( - {"-np", "--parallel"}, "N", - format("number of parallel sequences to decode (default: %d)", params.n_parallel), - [](gpt_params & params, int value) { - params.n_parallel = value; - } - )); - add_opt(llama_arg( - {"-ns", "--sequences"}, "N", - format("number of sequences to decode (default: %d)", params.n_sequences), - [](gpt_params & params, int value) { - params.n_sequences = value; - } - )); - add_opt(llama_arg( - {"-cb", "--cont-batching"}, - format("enable continuous batching (a.k.a dynamic batching) (default: %s)", params.cont_batching ? "enabled" : "disabled"), - [](gpt_params & params) { - params.cont_batching = true; - } - ).set_env("LLAMA_ARG_CONT_BATCHING")); - add_opt(llama_arg( - {"-nocb", "--no-cont-batching"}, - "disable continuous batching", - [](gpt_params & params) { - params.cont_batching = false; - } - ).set_env("LLAMA_ARG_NO_CONT_BATCHING")); - add_opt(llama_arg( - {"--mmproj"}, "FILE", - "path to a multimodal projector file for LLaVA. see examples/llava/README.md", - [](gpt_params & params, const std::string & value) { - params.mmproj = value; - } - ).set_examples({LLAMA_EXAMPLE_LLAVA})); - add_opt(llama_arg( - {"--image"}, "FILE", - "path to an image file. use with multimodal models. Specify multiple times for batching", - [](gpt_params & params, const std::string & value) { - params.image.emplace_back(value); - } - ).set_examples({LLAMA_EXAMPLE_LLAVA})); -#ifdef GGML_USE_RPC - add_opt(llama_arg( - {"--rpc"}, "SERVERS", - "comma separated list of RPC servers", - [](gpt_params & params, const std::string & value) { - params.rpc_servers = value; - } - )); -#endif - add_opt(llama_arg( - {"--mlock"}, - "force system to keep model in RAM rather than swapping or compressing", - [](gpt_params & params) { - params.use_mlock = true; - } - )); - add_opt(llama_arg( - {"--no-mmap"}, - "do not memory-map model (slower load but may reduce pageouts if not using mlock)", - [](gpt_params & params) { - params.use_mmap = false; - } - )); - add_opt(llama_arg( - {"--numa"}, "TYPE", - "attempt optimizations that help on some NUMA systems\n" - "- distribute: spread execution evenly over all nodes\n" - "- isolate: only spawn threads on CPUs on the node that execution started on\n" - "- numactl: use the CPU map provided by numactl\n" - "if run without this previously, it is recommended to drop the system page cache before using this\n" - "see https://github.com/ggerganov/llama.cpp/issues/1437", - [](gpt_params & params, const std::string & value) { - /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } - else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } - else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; } - else { throw std::invalid_argument("invalid value"); } - } - )); - add_opt(llama_arg( - {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N", - "number of layers to store in VRAM", - [](gpt_params & params, int value) { - params.n_gpu_layers = value; - if (!llama_supports_gpu_offload()) { - fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers option will be ignored\n"); - fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); - } - } - ).set_env("LLAMA_ARG_N_GPU_LAYERS")); - add_opt(llama_arg( - {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N", - "number of layers to store in VRAM for the draft model", - [](gpt_params & params, int value) { - params.n_gpu_layers_draft = value; - if (!llama_supports_gpu_offload()) { - fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n"); - fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); - } - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-sm", "--split-mode"}, "{none,layer,row}", - "how to split the model across multiple GPUs, one of:\n" - "- none: use one GPU only\n" - "- layer (default): split layers and KV across GPUs\n" - "- row: split rows across GPUs", - [](gpt_params & params, const std::string & value) { - std::string arg_next = value; - if (arg_next == "none") { - params.split_mode = LLAMA_SPLIT_MODE_NONE; - } else if (arg_next == "layer") { - params.split_mode = LLAMA_SPLIT_MODE_LAYER; - } - else if (arg_next == "row") { -#ifdef GGML_USE_SYCL - fprintf(stderr, "warning: The split mode value:[row] is not supported by llama.cpp with SYCL. It's developing.\nExit!\n"); - exit(1); -#endif // GGML_USE_SYCL - params.split_mode = LLAMA_SPLIT_MODE_ROW; - } - else { - throw std::invalid_argument("invalid value"); - } -#ifndef GGML_USE_CUDA_SYCL_VULKAN - fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting the split mode has no effect.\n"); -#endif // GGML_USE_CUDA_SYCL_VULKAN - } - )); - add_opt(llama_arg( - {"-ts", "--tensor-split"}, "N0,N1,N2,...", - "fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1", - [](gpt_params & params, const std::string & value) { - std::string arg_next = value; - - // split string by , and / - const std::regex regex{ R"([,/]+)" }; - std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 }; - std::vector split_arg{ it, {} }; - if (split_arg.size() >= llama_max_devices()) { - throw std::invalid_argument( - format("got %d input configs, but system only has %d devices", (int)split_arg.size(), (int)llama_max_devices()) - ); - } - for (size_t i = 0; i < llama_max_devices(); ++i) { - if (i < split_arg.size()) { - params.tensor_split[i] = std::stof(split_arg[i]); - } else { - params.tensor_split[i] = 0.0f; - } - } -#ifndef GGML_USE_CUDA_SYCL_VULKAN - fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting a tensor split has no effect.\n"); -#endif // GGML_USE_CUDA_SYCL_VULKAN - } - )); - add_opt(llama_arg( - {"-mg", "--main-gpu"}, "INDEX", - format("the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: %d)", params.main_gpu), - [](gpt_params & params, int value) { - params.main_gpu = value; -#ifndef GGML_USE_CUDA_SYCL_VULKAN - fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting the main GPU has no effect.\n"); -#endif // GGML_USE_CUDA_SYCL_VULKAN - } - )); - add_opt(llama_arg( - {"--check-tensors"}, - format("check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false"), - [](gpt_params & params) { - params.check_tensors = true; - } - )); - add_opt(llama_arg( - {"--override-kv"}, "KEY=TYPE:VALUE", - "advanced option to override model metadata by key. may be specified multiple times.\n" - "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false", - [](gpt_params & params, const std::string & value) { - if (!string_parse_kv_override(value.c_str(), params.kv_overrides)) { - throw std::runtime_error(format("error: Invalid type for KV override: %s\n", value.c_str())); - } - } - )); - add_opt(llama_arg( - {"--lora"}, "FNAME", - "path to LoRA adapter (can be repeated to use multiple adapters)", - [](gpt_params & params, const std::string & value) { - params.lora_adapters.push_back({ std::string(value), 1.0 }); - } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA})); - add_opt(llama_arg( - {"--lora-scaled"}, "FNAME", "SCALE", - "path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters)", - [](gpt_params & params, const std::string & fname, const std::string & scale) { - params.lora_adapters.push_back({ fname, std::stof(scale) }); - } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA})); - add_opt(llama_arg( - {"--control-vector"}, "FNAME", - "add a control vector\nnote: this argument can be repeated to add multiple control vectors", - [](gpt_params & params, const std::string & value) { - params.control_vectors.push_back({ 1.0f, value, }); - } - )); - add_opt(llama_arg( - {"--control-vector-scaled"}, "FNAME", "SCALE", - "add a control vector with user defined scaling SCALE\n" - "note: this argument can be repeated to add multiple scaled control vectors", - [](gpt_params & params, const std::string & fname, const std::string & scale) { - params.control_vectors.push_back({ std::stof(scale), fname }); - } - )); - add_opt(llama_arg( - {"--control-vector-layer-range"}, "START", "END", - "layer range to apply the control vector(s) to, start and end inclusive", - [](gpt_params & params, const std::string & start, const std::string & end) { - params.control_vector_layer_start = std::stoi(start); - params.control_vector_layer_end = std::stoi(end); - } - )); - add_opt(llama_arg( - {"-a", "--alias"}, "STRING", - "set alias for model name (to be used by REST API)", - [](gpt_params & params, const std::string & value) { - params.model_alias = value; - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"-m", "--model"}, "FNAME", - ex == LLAMA_EXAMPLE_EXPORT_LORA - ? std::string("model path from which to load base model") - : format( - "model path (default: `models/$filename` with filename from `--hf-file` " - "or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH - ), - [](gpt_params & params, const std::string & value) { - params.model = value; - } - ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL")); - add_opt(llama_arg( - {"-md", "--model-draft"}, "FNAME", - "draft model for speculative decoding (default: unused)", - [](gpt_params & params, const std::string & value) { - params.model_draft = value; - } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); - add_opt(llama_arg( - {"-mu", "--model-url"}, "MODEL_URL", - "model download url (default: unused)", - [](gpt_params & params, const std::string & value) { - params.model_url = value; - } - ).set_env("LLAMA_ARG_MODEL_URL")); - add_opt(llama_arg( - {"-hfr", "--hf-repo"}, "REPO", - "Hugging Face model repository (default: unused)", - [](gpt_params & params, const std::string & value) { - params.hf_repo = value; - } - ).set_env("LLAMA_ARG_HF_REPO")); - add_opt(llama_arg( - {"-hff", "--hf-file"}, "FILE", - "Hugging Face model file (default: unused)", - [](gpt_params & params, const std::string & value) { - params.hf_file = value; - } - ).set_env("LLAMA_ARG_HF_FILE")); - add_opt(llama_arg( - {"-hft", "--hf-token"}, "TOKEN", - "Hugging Face access token (default: value from HF_TOKEN environment variable)", - [](gpt_params & params, const std::string & value) { - params.hf_token = value; - } - ).set_env("HF_TOKEN")); - add_opt(llama_arg( - {"--context-file"}, "FNAME", - "file to load context from (repeat to specify multiple files)", - [](gpt_params & params, const std::string & value) { - std::ifstream file(value, std::ios::binary); - if (!file) { - throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); - } - params.context_files.push_back(value); - } - ).set_examples({LLAMA_EXAMPLE_RETRIEVAL})); - add_opt(llama_arg( - {"--chunk-size"}, "N", - format("minimum length of embedded text chunks (default: %d)", params.chunk_size), - [](gpt_params & params, int value) { - params.chunk_size = value; - } - ).set_examples({LLAMA_EXAMPLE_RETRIEVAL})); - add_opt(llama_arg( - {"--chunk-separator"}, "STRING", - format("separator between chunks (default: '%s')", params.chunk_separator.c_str()), - [](gpt_params & params, const std::string & value) { - params.chunk_separator = value; - } - ).set_examples({LLAMA_EXAMPLE_RETRIEVAL})); - add_opt(llama_arg( - {"--junk"}, "N", - format("number of times to repeat the junk text (default: %d)", params.n_junk), - [](gpt_params & params, int value) { - params.n_junk = value; - } - ).set_examples({LLAMA_EXAMPLE_PASSKEY})); - add_opt(llama_arg( - {"--pos"}, "N", - format("position of the passkey in the junk text (default: %d)", params.i_pos), - [](gpt_params & params, int value) { - params.i_pos = value; - } - ).set_examples({LLAMA_EXAMPLE_PASSKEY})); - add_opt(llama_arg( - {"-o", "--output", "--output-file"}, "FNAME", - format("output file (default: '%s')", - ex == LLAMA_EXAMPLE_EXPORT_LORA - ? params.lora_outfile.c_str() - : ex == LLAMA_EXAMPLE_CVECTOR_GENERATOR - ? params.cvector_outfile.c_str() - : params.out_file.c_str()), - [](gpt_params & params, const std::string & value) { - params.out_file = value; - params.cvector_outfile = value; - params.lora_outfile = value; - } - ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA})); - add_opt(llama_arg( - {"-ofreq", "--output-frequency"}, "N", - format("output the imatrix every N iterations (default: %d)", params.n_out_freq), - [](gpt_params & params, int value) { - params.n_out_freq = value; - } - ).set_examples({LLAMA_EXAMPLE_IMATRIX})); - add_opt(llama_arg( - {"--save-frequency"}, "N", - format("save an imatrix copy every N iterations (default: %d)", params.n_save_freq), - [](gpt_params & params, int value) { - params.n_save_freq = value; - } - ).set_examples({LLAMA_EXAMPLE_IMATRIX})); - add_opt(llama_arg( - {"--process-output"}, - format("collect data for the output tensor (default: %s)", params.process_output ? "true" : "false"), - [](gpt_params & params) { - params.process_output = true; - } - ).set_examples({LLAMA_EXAMPLE_IMATRIX})); - add_opt(llama_arg( - {"--no-ppl"}, - format("do not compute perplexity (default: %s)", params.compute_ppl ? "true" : "false"), - [](gpt_params & params) { - params.compute_ppl = false; - } - ).set_examples({LLAMA_EXAMPLE_IMATRIX})); - add_opt(llama_arg( - {"--chunk", "--from-chunk"}, "N", - format("start processing the input from chunk N (default: %d)", params.i_chunk), - [](gpt_params & params, int value) { - params.i_chunk = value; - } - ).set_examples({LLAMA_EXAMPLE_IMATRIX})); - add_opt(llama_arg( - {"-pps"}, - format("is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false"), - [](gpt_params & params) { - params.is_pp_shared = true; - } - ).set_examples({LLAMA_EXAMPLE_BENCH})); - add_opt(llama_arg( - {"-npp"}, "n0,n1,...", - "number of prompt tokens", - [](gpt_params & params, const std::string & value) { - auto p = string_split(value, ','); - params.n_pp.insert(params.n_pp.end(), p.begin(), p.end()); - } - ).set_examples({LLAMA_EXAMPLE_BENCH})); - add_opt(llama_arg( - {"-ntg"}, "n0,n1,...", - "number of text generation tokens", - [](gpt_params & params, const std::string & value) { - auto p = string_split(value, ','); - params.n_tg.insert(params.n_tg.end(), p.begin(), p.end()); - } - ).set_examples({LLAMA_EXAMPLE_BENCH})); - add_opt(llama_arg( - {"-npl"}, "n0,n1,...", - "number of parallel prompts", - [](gpt_params & params, const std::string & value) { - auto p = string_split(value, ','); - params.n_pl.insert(params.n_pl.end(), p.begin(), p.end()); - } - ).set_examples({LLAMA_EXAMPLE_BENCH})); - add_opt(llama_arg( - {"--embd-normalize"}, "N", - format("normalisation for embendings (default: %d) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)", params.embd_normalize), - [](gpt_params & params, int value) { - params.embd_normalize = value; - } - ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); - add_opt(llama_arg( - {"--embd-output-format"}, "FORMAT", - "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix", - [](gpt_params & params, const std::string & value) { - params.embd_out = value; - } - ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); - add_opt(llama_arg( - {"--embd-separator"}, "STRING", - "separator of embendings (default \\n) for example \"<#sep#>\"", - [](gpt_params & params, const std::string & value) { - params.embd_sep = value; - } - ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); - add_opt(llama_arg( - {"--host"}, "HOST", - format("ip address to listen (default: %s)", params.hostname.c_str()), - [](gpt_params & params, const std::string & value) { - params.hostname = value; - } - ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_HOST")); - add_opt(llama_arg( - {"--port"}, "PORT", - format("port to listen (default: %d)", params.port), - [](gpt_params & params, int value) { - params.port = value; - } - ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_PORT")); - add_opt(llama_arg( - {"--path"}, "PATH", - format("path to serve static files from (default: %s)", params.public_path.c_str()), - [](gpt_params & params, const std::string & value) { - params.public_path = value; - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--embedding", "--embeddings"}, - format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"), - [](gpt_params & params) { - params.embedding = true; - } - ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS")); - add_opt(llama_arg( - {"--api-key"}, "KEY", - "API key to use for authentication (default: none)", - [](gpt_params & params, const std::string & value) { - params.api_keys.push_back(value); - } - ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_API_KEY")); - add_opt(llama_arg( - {"--api-key-file"}, "FNAME", - "path to file containing API keys (default: none)", - [](gpt_params & params, const std::string & value) { - std::ifstream key_file(value); - if (!key_file) { - throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); - } - std::string key; - while (std::getline(key_file, key)) { - if (!key.empty()) { - params.api_keys.push_back(key); - } - } - key_file.close(); - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--ssl-key-file"}, "FNAME", - "path to file a PEM-encoded SSL private key", - [](gpt_params & params, const std::string & value) { - params.ssl_file_key = value; - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--ssl-cert-file"}, "FNAME", - "path to file a PEM-encoded SSL certificate", - [](gpt_params & params, const std::string & value) { - params.ssl_file_cert = value; - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"-to", "--timeout"}, "N", - format("server read/write timeout in seconds (default: %d)", params.timeout_read), - [](gpt_params & params, int value) { - params.timeout_read = value; - params.timeout_write = value; - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--threads-http"}, "N", - format("number of threads used to process HTTP requests (default: %d)", params.n_threads_http), - [](gpt_params & params, int value) { - params.n_threads_http = value; - } - ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP")); - add_opt(llama_arg( - {"-spf", "--system-prompt-file"}, "FNAME", - "set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications", - [](gpt_params & params, const std::string & value) { - std::ifstream file(value); - if (!file) { - throw std::runtime_error(format("error: failed to open file '%s'\n", value.c_str())); - } - std::string system_prompt; - std::copy( - std::istreambuf_iterator(file), - std::istreambuf_iterator(), - std::back_inserter(system_prompt) - ); - params.system_prompt = system_prompt; - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--log-format"}, "{text, json}", - "log output format: json or text (default: json)", - [](gpt_params & params, const std::string & value) { - if (value == "json") { - params.log_json = true; - } else if (value == "text") { - params.log_json = false; - } else { - throw std::invalid_argument("invalid value"); - } - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--metrics"}, - format("enable prometheus compatible metrics endpoint (default: %s)", params.endpoint_metrics ? "enabled" : "disabled"), - [](gpt_params & params) { - params.endpoint_metrics = true; - } - ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_METRICS")); - add_opt(llama_arg( - {"--no-slots"}, - format("disables slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled"), - [](gpt_params & params) { - params.endpoint_slots = false; - } - ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_ENDPOINT_SLOTS")); - add_opt(llama_arg( - {"--slot-save-path"}, "PATH", - "path to save slot kv cache (default: disabled)", - [](gpt_params & params, const std::string & value) { - params.slot_save_path = value; - // if doesn't end with DIRECTORY_SEPARATOR, add it - if (!params.slot_save_path.empty() && params.slot_save_path[params.slot_save_path.size() - 1] != DIRECTORY_SEPARATOR) { - params.slot_save_path += DIRECTORY_SEPARATOR; - } - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--chat-template"}, "JINJA_TEMPLATE", - "set custom jinja chat template (default: template taken from model's metadata)\n" - "if suffix/prefix are specified, template will be disabled\n" - "only commonly used templates are accepted:\nhttps://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template", - [](gpt_params & params, const std::string & value) { - if (!llama_chat_verify_template(value)) { - throw std::runtime_error(format( - "error: the supplied chat template is not supported: %s\n" - "note: llama.cpp does not use jinja parser, we only support commonly used templates\n", - value.c_str() - )); - } - params.chat_template = value; - } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE")); - add_opt(llama_arg( - {"-sps", "--slot-prompt-similarity"}, "SIMILARITY", - format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity), - [](gpt_params & params, const std::string & value) { - params.slot_prompt_similarity = std::stof(value); - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--lora-init-without-apply"}, - format("load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"), - [](gpt_params & params) { - params.lora_init_without_apply = true; - } - ).set_examples({LLAMA_EXAMPLE_SERVER})); - add_opt(llama_arg( - {"--simple-io"}, - "use basic IO for better compatibility in subprocesses and limited consoles", - [](gpt_params & params) { - params.simple_io = true; - } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL})); - add_opt(llama_arg( - {"-ld", "--logdir"}, "LOGDIR", - "path under which to save YAML logs (no logging if unset)", - [](gpt_params & params, const std::string & value) { - params.logdir = value; - - if (params.logdir.back() != DIRECTORY_SEPARATOR) { - params.logdir += DIRECTORY_SEPARATOR; - } - } - )); - add_opt(llama_arg( - {"--positive-file"}, "FNAME", - format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()), - [](gpt_params & params, const std::string & value) { - params.cvector_positive_file = value; - } - ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); - add_opt(llama_arg( - {"--negative-file"}, "FNAME", - format("negative prompts file, one prompt per line (default: '%s')", params.cvector_negative_file.c_str()), - [](gpt_params & params, const std::string & value) { - params.cvector_negative_file = value; - } - ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); - add_opt(llama_arg( - {"--pca-batch"}, "N", - format("batch size used for PCA. Larger batch runs faster, but uses more memory (default: %d)", params.n_pca_batch), - [](gpt_params & params, int value) { - params.n_pca_batch = value; - } - ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); - add_opt(llama_arg( - {"--pca-iter"}, "N", - format("number of iterations used for PCA (default: %d)", params.n_pca_iterations), - [](gpt_params & params, int value) { - params.n_pca_iterations = value; - } - ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); - add_opt(llama_arg( - {"--method"}, "{pca, mean}", - "dimensionality reduction method to be used (default: pca)", - [](gpt_params & params, const std::string & value) { - /**/ if (value == "pca") { params.cvector_dimre_method = DIMRE_METHOD_PCA; } - else if (value == "mean") { params.cvector_dimre_method = DIMRE_METHOD_MEAN; } - else { throw std::invalid_argument("invalid value"); } - } - ).set_examples({LLAMA_EXAMPLE_CVECTOR_GENERATOR})); - add_opt(llama_arg( - {"--output-format"}, "{md,jsonl}", - "output format for batched-bench results (default: md)", - [](gpt_params & params, const std::string & value) { - /**/ if (value == "jsonl") { params.batched_bench_output_jsonl = true; } - else if (value == "md") { params.batched_bench_output_jsonl = false; } - else { std::invalid_argument("invalid value"); } - } - ).set_examples({LLAMA_EXAMPLE_BENCH})); -#ifndef LOG_DISABLE_LOGS - // TODO: make this looks less weird - add_opt(llama_arg( - {"--log-test"}, - "Log test", - [](gpt_params &) { log_param_single_parse("--log-test"); } - )); - add_opt(llama_arg( - {"--log-disable"}, - "Log disable", - [](gpt_params &) { log_param_single_parse("--log-disable"); } - )); - add_opt(llama_arg( - {"--log-enable"}, - "Log enable", - [](gpt_params &) { log_param_single_parse("--log-enable"); } - )); - add_opt(llama_arg( - {"--log-new"}, - "Log new", - [](gpt_params &) { log_param_single_parse("--log-new"); } - )); - add_opt(llama_arg( - {"--log-append"}, - "Log append", - [](gpt_params &) { log_param_single_parse("--log-append"); } - )); - add_opt(llama_arg( - {"--log-file"}, "FNAME", - "Log file", - [](gpt_params &, const std::string & value) { log_param_pair_parse(false, "--log-file", value); } - )); -#endif // LOG_DISABLE_LOGS - - return options; -} - std::string gpt_params_get_system_info(const gpt_params & params) { std::ostringstream os; diff --git a/common/common.h b/common/common.h index 5b945ac02..23babdd09 100644 --- a/common/common.h +++ b/common/common.h @@ -4,20 +4,11 @@ #include "llama.h" -#include "sampling.h" - #define LOG_NO_FILE_LINE_FUNCTION #include "log.h" -#include #include #include -#include -#include -#include -#include -#include -#include #ifdef _WIN32 #define DIRECTORY_SEPARATOR '\\' @@ -56,11 +47,20 @@ struct llama_control_vector_load_info; // CPU utils // +struct cpu_params { + int n_threads = -1; + bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask. + bool mask_valid = false; // Default: any CPU + enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime) + bool strict_cpu = false; // Use strict CPU placement + uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling) +}; + int32_t cpu_get_num_physical_cores(); int32_t cpu_get_num_math(); // -// CLI argument parsing +// Common params // enum llama_example { @@ -78,28 +78,71 @@ enum llama_example { LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_LLAVA, + LLAMA_EXAMPLE_LOOKUP, + LLAMA_EXAMPLE_PARALLEL, LLAMA_EXAMPLE_COUNT, }; +enum gpt_sampler_type { + GPT_SAMPLER_TYPE_NONE = 0, + GPT_SAMPLER_TYPE_TOP_K = 1, + GPT_SAMPLER_TYPE_TOP_P = 2, + GPT_SAMPLER_TYPE_MIN_P = 3, + GPT_SAMPLER_TYPE_TFS_Z = 4, + GPT_SAMPLER_TYPE_TYPICAL_P = 5, + GPT_SAMPLER_TYPE_TEMPERATURE = 6, +}; + // dimensionality reduction methods, used by cvector-generator enum dimre_method { DIMRE_METHOD_PCA, DIMRE_METHOD_MEAN, }; -struct cpu_params { - int n_threads = -1; - bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask. - bool mask_valid = false; // Default: any CPU - enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime) - bool strict_cpu = false; // Use strict CPU placement - uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling) +// sampler parameters +struct gpt_sampler_params { + uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler + + int32_t n_prev = 64; // number of previous tokens to remember + int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens. + int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens + int32_t top_k = 40; // <= 0 to use vocab size + float top_p = 0.95f; // 1.0 = disabled + float min_p = 0.05f; // 0.0 = disabled + float tfs_z = 1.00f; // 1.0 = disabled + float typ_p = 1.00f; // typical_p, 1.0 = disabled + float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities + float dynatemp_range = 0.00f; // 0.0 = disabled + float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler + int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size) + float penalty_repeat = 1.00f; // 1.0 = disabled + float penalty_freq = 0.00f; // 0.0 = disabled + float penalty_present = 0.00f; // 0.0 = disabled + int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0 + float mirostat_tau = 5.00f; // target entropy + float mirostat_eta = 0.10f; // learning rate + bool penalize_nl = false; // consider newlines as a repeatable token + bool ignore_eos = false; + + std::vector samplers = { + GPT_SAMPLER_TYPE_TOP_K, + GPT_SAMPLER_TYPE_TFS_Z, + GPT_SAMPLER_TYPE_TYPICAL_P, + GPT_SAMPLER_TYPE_TOP_P, + GPT_SAMPLER_TYPE_MIN_P, + GPT_SAMPLER_TYPE_TEMPERATURE + }; + + std::string grammar; // optional BNF-like grammar to constrain sampling + + std::vector logit_bias; // logit biases to apply + + // print the parameters into a string + std::string print() const; }; struct gpt_params { - enum llama_example curr_ex = LLAMA_EXAMPLE_COMMON; - int32_t n_predict = -1; // new tokens to predict int32_t n_ctx = 0; // context size int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS) @@ -143,23 +186,23 @@ struct gpt_params { struct gpt_sampler_params sparams; - std::string model = ""; // model path - std::string model_draft = ""; // draft model for speculative decoding - std::string model_alias = "unknown"; // model alias - std::string model_url = ""; // model url to download - std::string hf_token = ""; // HF token - std::string hf_repo = ""; // HF repo - std::string hf_file = ""; // HF file - std::string prompt = ""; - std::string prompt_file = ""; // store the external prompt file name - std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state - std::string input_prefix = ""; // string to prefix user inputs with - std::string input_suffix = ""; // string to suffix user inputs with - std::string logdir = ""; // directory in which to save YAML log files - std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding - std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding - std::string logits_file = ""; // file for saving *all* logits - std::string rpc_servers = ""; // comma separated list of RPC servers + std::string model = ""; // model path // NOLINT + std::string model_draft = ""; // draft model for speculative decoding // NOLINT + std::string model_alias = "unknown"; // model alias // NOLINT + std::string model_url = ""; // model url to download // NOLINT + std::string hf_token = ""; // HF token // NOLINT + std::string hf_repo = ""; // HF repo // NOLINT + std::string hf_file = ""; // HF file // NOLINT + std::string prompt = ""; // NOLINT + std::string prompt_file = ""; // store the external prompt file name // NOLINT + std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state // NOLINT + std::string input_prefix = ""; // string to prefix user inputs with // NOLINT + std::string input_suffix = ""; // string to suffix user inputs with // NOLINT + std::string logdir = ""; // directory in which to save YAML log files // NOLINT + std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding // NOLINT + std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding // NOLINT + std::string logits_file = ""; // file for saving *all* logits // NOLINT + std::string rpc_servers = ""; // comma separated list of RPC servers // NOLINT std::vector in_files; // all input files std::vector antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts) @@ -189,7 +232,6 @@ struct gpt_params { bool kl_divergence = false; // compute KL divergence - std::function print_usage = nullptr; // print example-specific usage and example bool usage = false; // print usage bool use_color = false; // use color to distinguish generations and inputs bool special = false; // enable special token output @@ -220,7 +262,7 @@ struct gpt_params { std::string cache_type_v = "f16"; // KV cache data type for the V // multimodal models (see examples/llava) - std::string mmproj = ""; // path to multimodal projector + std::string mmproj = ""; // path to multimodal projector // NOLINT std::vector image; // path to image file(s) // embedding @@ -236,15 +278,15 @@ struct gpt_params { int n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool) std::string hostname = "127.0.0.1"; - std::string public_path = ""; - std::string chat_template = ""; - std::string system_prompt = ""; + std::string public_path = ""; // NOLINT + std::string chat_template = ""; // NOLINT + std::string system_prompt = ""; // NOLINT bool enable_chat_template = true; std::vector api_keys; - std::string ssl_file_key = ""; - std::string ssl_file_cert = ""; + std::string ssl_file_key = ""; // NOLINT + std::string ssl_file_cert = ""; // NOLINT bool endpoint_slots = true; bool endpoint_metrics = false; @@ -299,92 +341,6 @@ struct gpt_params { bool batched_bench_output_jsonl = false; }; -struct llama_arg { - std::set examples = {LLAMA_EXAMPLE_COMMON}; - std::vector args; - const char * value_hint = nullptr; // help text or example for arg value - const char * value_hint_2 = nullptr; // for second arg value - const char * env = nullptr; - std::string help; - void (*handler_void) (gpt_params & params) = nullptr; - void (*handler_string) (gpt_params & params, const std::string &) = nullptr; - void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr; - void (*handler_int) (gpt_params & params, int) = nullptr; - - llama_arg( - const std::initializer_list & args, - const char * value_hint, - const std::string & help, - void (*handler)(gpt_params & params, const std::string &) - ) : args(args), value_hint(value_hint), help(help), handler_string(handler) {} - - llama_arg( - const std::initializer_list & args, - const char * value_hint, - const std::string & help, - void (*handler)(gpt_params & params, int) - ) : args(args), value_hint(value_hint), help(help), handler_int(handler) {} - - llama_arg( - const std::initializer_list & args, - const std::string & help, - void (*handler)(gpt_params & params) - ) : args(args), help(help), handler_void(handler) {} - - // support 2 values for arg - llama_arg( - const std::initializer_list & args, - const char * value_hint, - const char * value_hint_2, - const std::string & help, - void (*handler)(gpt_params & params, const std::string &, const std::string &) - ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {} - - llama_arg & set_examples(std::initializer_list examples) { - this->examples = std::move(examples); - return *this; - } - - llama_arg & set_env(const char * env) { - help = help + "\n(env: " + env + ")"; - this->env = env; - return *this; - } - - bool in_example(enum llama_example ex) { - return examples.find(ex) != examples.end(); - } - - bool get_value_from_env(std::string & output) const { - if (env == nullptr) return false; - char * value = std::getenv(env); - if (value) { - output = value; - return true; - } - return false; - } - - bool has_value_from_env() const { - return env != nullptr && std::getenv(env); - } - - std::string to_string(); -}; - -// initialize list of options (arguments) that can be used by the current example -std::vector gpt_params_parser_init(gpt_params & params, llama_example ex); -// optionally, we can provide "print_usage" to print example usage -std::vector gpt_params_parser_init(gpt_params & params, llama_example ex, std::function print_usage); - -// parse input arguments from CLI -// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message) -bool gpt_params_parse (int argc, char ** argv, gpt_params & params, std::vector & options); -bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params, std::vector & options); - -// print full usage message; it will be called internally by gpt_params_parse() if "-h" is set -void gpt_params_print_usage(gpt_params & params, std::vector & options); - std::string gpt_params_get_system_info(const gpt_params & params); bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]); diff --git a/common/sampling.cpp b/common/sampling.cpp index 7806b77e0..21b956462 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -2,6 +2,9 @@ #include "common.h" +#include +#include + // the ring buffer works similarly to std::deque, but with a fixed capacity // TODO: deduplicate with llama-impl.h template @@ -420,7 +423,7 @@ std::vector gpt_sampler_types_from_names(const std::vector gpt_sampler_types_from_chars(const std::string & chars) { - std::unordered_map sampler_name_map { + std::unordered_map sampler_name_map = { { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_K), GPT_SAMPLER_TYPE_TOP_K }, { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TFS_Z), GPT_SAMPLER_TYPE_TFS_Z }, { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TYPICAL_P), GPT_SAMPLER_TYPE_TYPICAL_P }, diff --git a/common/sampling.h b/common/sampling.h index 654e0c513..0a4461fab 100644 --- a/common/sampling.h +++ b/common/sampling.h @@ -2,61 +2,11 @@ #include "llama.h" +#include "common.h" + #include #include -enum gpt_sampler_type { - GPT_SAMPLER_TYPE_NONE = 0, - GPT_SAMPLER_TYPE_TOP_K = 1, - GPT_SAMPLER_TYPE_TOP_P = 2, - GPT_SAMPLER_TYPE_MIN_P = 3, - GPT_SAMPLER_TYPE_TFS_Z = 4, - GPT_SAMPLER_TYPE_TYPICAL_P = 5, - GPT_SAMPLER_TYPE_TEMPERATURE = 6, -}; - -// sampling parameters -struct gpt_sampler_params { - uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler - - int32_t n_prev = 64; // number of previous tokens to remember - int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens. - int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens - int32_t top_k = 40; // <= 0 to use vocab size - float top_p = 0.95f; // 1.0 = disabled - float min_p = 0.05f; // 0.0 = disabled - float tfs_z = 1.00f; // 1.0 = disabled - float typ_p = 1.00f; // typical_p, 1.0 = disabled - float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities - float dynatemp_range = 0.00f; // 0.0 = disabled - float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler - int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size) - float penalty_repeat = 1.00f; // 1.0 = disabled - float penalty_freq = 0.00f; // 0.0 = disabled - float penalty_present = 0.00f; // 0.0 = disabled - int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0 - float mirostat_tau = 5.00f; // target entropy - float mirostat_eta = 0.10f; // learning rate - bool penalize_nl = false; // consider newlines as a repeatable token - bool ignore_eos = false; - - std::vector samplers = { - GPT_SAMPLER_TYPE_TOP_K, - GPT_SAMPLER_TYPE_TFS_Z, - GPT_SAMPLER_TYPE_TYPICAL_P, - GPT_SAMPLER_TYPE_TOP_P, - GPT_SAMPLER_TYPE_MIN_P, - GPT_SAMPLER_TYPE_TEMPERATURE - }; - - std::string grammar; // optional BNF-like grammar to constrain sampling - - std::vector logit_bias; // logit biases to apply - - // print the parameters into a string - std::string print() const; -}; - // gpt_sampler extends llama_sampler with additional functionality: // // - grammar support diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index f3b0c433b..a91e7f4bd 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -37,8 +38,7 @@ static void print_usage(int, char ** argv) { int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_BENCH, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) { return 1; } diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index 615d6f0f5..5d32153fe 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -18,8 +19,7 @@ int main(int argc, char ** argv) { params.prompt = "Hello my name is"; params.n_predict = 32; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) { return 1; } diff --git a/examples/cvector-generator/cvector-generator.cpp b/examples/cvector-generator/cvector-generator.cpp index 0795175a1..569b6c38f 100644 --- a/examples/cvector-generator/cvector-generator.cpp +++ b/examples/cvector-generator/cvector-generator.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" #include "ggml.h" @@ -388,8 +389,7 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) { int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) { return 1; } diff --git a/examples/cvector-generator/pca.hpp b/examples/cvector-generator/pca.hpp index 6ec3141af..05c66856c 100644 --- a/examples/cvector-generator/pca.hpp +++ b/examples/cvector-generator/pca.hpp @@ -12,12 +12,9 @@ #include #include +#include #include -#include #include -#include -#include -#include #define DEBUG_POS 5 diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 630f7c1c7..da7c79253 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -79,8 +80,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_EMBEDDING); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) { return 1; } diff --git a/examples/eval-callback/eval-callback.cpp b/examples/eval-callback/eval-callback.cpp index 881111ffd..bc7203143 100644 --- a/examples/eval-callback/eval-callback.cpp +++ b/examples/eval-callback/eval-callback.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" #include "ggml.h" @@ -144,8 +145,7 @@ int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) { return 1; } diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp index 544e7fff6..ff324926a 100644 --- a/examples/export-lora/export-lora.cpp +++ b/examples/export-lora/export-lora.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "ggml.h" #include "ggml-alloc.h" @@ -401,8 +402,7 @@ static void print_usage(int, char ** argv) { int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) { return 1; } diff --git a/examples/gen-docs/gen-docs.cpp b/examples/gen-docs/gen-docs.cpp index 8b1dafd63..b6d4725fd 100644 --- a/examples/gen-docs/gen-docs.cpp +++ b/examples/gen-docs/gen-docs.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include @@ -9,11 +10,11 @@ static void export_md(std::string fname, llama_example ex) { std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc); gpt_params params; - auto options = gpt_params_parser_init(params, ex); + auto ctx_arg = gpt_params_parser_init(params, ex); file << "| Argument | Explanation |\n"; file << "| -------- | ----------- |\n"; - for (auto & opt : options) { + for (auto & opt : ctx_arg.options) { file << "| `"; // args for (const auto & arg : opt.args) { diff --git a/examples/gritlm/gritlm.cpp b/examples/gritlm/gritlm.cpp index 6f060e2dc..14c715202 100644 --- a/examples/gritlm/gritlm.cpp +++ b/examples/gritlm/gritlm.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -153,8 +154,7 @@ static std::string gritlm_instruction(const std::string & instruction) { int main(int argc, char * argv[]) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) { return 1; } diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 15a3f0d14..032a90136 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -577,8 +578,7 @@ int main(int argc, char ** argv) { params.logits_all = true; params.verbosity = 1; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_IMATRIX, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) { return 1; } diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 87abb761f..9a527e244 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -1,6 +1,7 @@ +#include "arg.h" #include "common.h" - #include "console.h" +#include "sampling.h" #include "llama.h" #include @@ -105,8 +106,7 @@ int main(int argc, char ** argv) { gpt_params params; g_params = ¶ms; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_INFILL); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) { return 1; } diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index 5845d0106..e9108a9bd 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -1,11 +1,12 @@ -#include "ggml.h" +#include "arg.h" +#include "base64.hpp" #include "log.h" #include "common.h" +#include "sampling.h" #include "clip.h" #include "llava.h" #include "llama.h" - -#include "base64.hpp" +#include "ggml.h" #include #include @@ -278,8 +279,7 @@ int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_LLAVA, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) { return 1; } diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp index 57e7d42c5..3475bbce5 100644 --- a/examples/llava/minicpmv-cli.cpp +++ b/examples/llava/minicpmv-cli.cpp @@ -1,9 +1,11 @@ -#include "ggml.h" +#include "arg.h" #include "log.h" #include "common.h" +#include "sampling.h" #include "clip.h" #include "llava.h" #include "llama.h" +#include "ggml.h" #include #include @@ -253,8 +255,7 @@ int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, show_additional_info); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, show_additional_info)) { return 1; } diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp index 5027a483a..de8b792f2 100644 --- a/examples/lookahead/lookahead.cpp +++ b/examples/lookahead/lookahead.cpp @@ -1,4 +1,6 @@ +#include "arg.h" #include "common.h" +#include "sampling.h" #include "llama.h" #include @@ -36,8 +38,7 @@ struct ngram_container { int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) { return 1; } diff --git a/examples/lookup/lookup-create.cpp b/examples/lookup/lookup-create.cpp index 795b06c88..33287c02c 100644 --- a/examples/lookup/lookup-create.cpp +++ b/examples/lookup/lookup-create.cpp @@ -1,7 +1,8 @@ -#include "ggml.h" -#include "llama.h" +#include "arg.h" #include "common.h" #include "ngram-cache.h" +#include "ggml.h" +#include "llama.h" #include #include @@ -13,8 +14,7 @@ int main(int argc, char ** argv){ gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) { return 1; } @@ -40,4 +40,6 @@ int main(int argc, char ** argv){ fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str()); llama_ngram_cache_save(ngram_cache, params.lookup_cache_static); + + return 0; } diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp index 93299ef8b..f299d68a9 100644 --- a/examples/lookup/lookup-stats.cpp +++ b/examples/lookup/lookup-stats.cpp @@ -1,8 +1,9 @@ -#include "ggml.h" +#include "arg.h" #include "common.h" -#include "llama.h" #include "log.h" #include "ngram-cache.h" +#include "llama.h" +#include "ggml.h" #include #include @@ -15,8 +16,7 @@ int main(int argc, char ** argv){ gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) { return 1; } diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp index 9ac7f6b47..fff44a499 100644 --- a/examples/lookup/lookup.cpp +++ b/examples/lookup/lookup.cpp @@ -1,7 +1,9 @@ +#include "arg.h" #include "ggml.h" -#include "llama.h" #include "common.h" #include "ngram-cache.h" +#include "sampling.h" +#include "llama.h" #include #include @@ -12,8 +14,7 @@ int main(int argc, char ** argv){ gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) { return 1; } diff --git a/examples/main/main.cpp b/examples/main/main.cpp index ef2158842..b986a865a 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -1,6 +1,7 @@ +#include "arg.h" #include "common.h" - #include "console.h" +#include "sampling.h" #include "llama.h" #include @@ -138,9 +139,7 @@ static std::string chat_add_and_format(struct llama_model * model, std::vector @@ -100,8 +102,7 @@ int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) { return 1; } diff --git a/examples/passkey/passkey.cpp b/examples/passkey/passkey.cpp index 271ef3a98..d3d5ab46f 100644 --- a/examples/passkey/passkey.cpp +++ b/examples/passkey/passkey.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -19,8 +20,7 @@ int main(int argc, char ** argv) { params.n_keep = 32; params.i_pos = -1; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_PASSKEY, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) { return 1; } diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 570ee8aeb..c7d617988 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1,18 +1,19 @@ +#include "arg.h" #include "common.h" #include "llama.h" +#include +#include #include #include #include #include +#include +#include +#include #include #include -#include -#include #include -#include -#include -#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data @@ -1967,8 +1968,7 @@ int main(int argc, char ** argv) { params.n_ctx = 512; params.logits_all = true; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_PERPLEXITY); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) { return 1; } diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp index dd8a82e6e..7a360b731 100644 --- a/examples/retrieval/retrieval.cpp +++ b/examples/retrieval/retrieval.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -111,8 +112,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_RETRIEVAL, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) { return 1; } diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index e17ab0ed0..0117d9357 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -10,8 +11,7 @@ int main(int argc, char ** argv) { params.prompt = "The quick brown fox"; params.sparams.seed = 1234; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) { return 1; } diff --git a/examples/server/README.md b/examples/server/README.md index ed1201ba8..79196e9c1 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -23,36 +23,32 @@ The project is under active development, and we are [looking for feedback and co | `--version` | show version and build info | | `-v, --verbose` | print verbose information | | `--verbosity N` | set specific verbosity level (default: 0) | -| `--verbose-prompt` | print a verbose prompt before generation (default: false) | -| `--no-display-prompt` | don't print prompt at generation (default: false) | -| `-s, --seed SEED` | RNG seed (default: -1, use random seed for < 0) | | `-t, --threads N` | number of threads to use during generation (default: -1)
(env: LLAMA_ARG_THREADS) | | `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) | | `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") | | `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask | | `--cpu-strict <0\|1>` | use strict CPU placement (default: 0)
| +| `--prio N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)
| | `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50)
| | `-Cb, --cpu-mask-batch M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask) | | `-Crb, --cpu-range-batch lo-hi` | ranges of CPUs for affinity. Complements --cpu-mask-batch | | `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) | +| `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)
| | `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) | -| `-lcs, --lookup-cache-static FNAME` | path to static lookup cache to use for lookup decoding (not updated by generation) | -| `-lcd, --lookup-cache-dynamic FNAME` | path to dynamic lookup cache to use for lookup decoding (updated by generation) | | `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)
(env: LLAMA_ARG_CTX_SIZE) | | `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)
(env: LLAMA_ARG_N_PREDICT) | | `-b, --batch-size N` | logical maximum batch size (default: 2048)
(env: LLAMA_ARG_BATCH) | | `-ub, --ubatch-size N` | physical maximum batch size (default: 512)
(env: LLAMA_ARG_UBATCH) | | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) | -| `--chunks N` | max number of chunks to process (default: -1, -1 = all) | | `-fa, --flash-attn` | enable Flash Attention (default: disabled)
(env: LLAMA_ARG_FLASH_ATTN) | | `-p, --prompt PROMPT` | prompt to start generation with | | `-f, --file FNAME` | a file containing the prompt (default: none) | -| `--in-file FNAME` | an input file (repeat to specify multiple files) | | `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) | | `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) | | `--no-escape` | do not process escape sequences | | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) | -| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'
(default: top_k;tfs_z;typical_p;top_p;min_p;temperature) | +| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'
(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) | +| `-s, --seed SEED` | RNG seed (default: -1, use random seed for < 0) | | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) | | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) | | `--penalize-nl` | penalize newline tokens (default: false) | @@ -92,13 +88,12 @@ The project is under active development, and we are [looking for feedback and co | `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16) | | `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)
(env: LLAMA_ARG_DEFRAG_THOLD) | | `-np, --parallel N` | number of parallel sequences to decode (default: 1) | -| `-ns, --sequences N` | number of sequences to decode (default: 1) | | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)
(env: LLAMA_ARG_CONT_BATCHING) | | `-nocb, --no-cont-batching` | disable continuous batching
(env: LLAMA_ARG_NO_CONT_BATCHING) | | `--mlock` | force system to keep model in RAM rather than swapping or compressing | | `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock) | | `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggerganov/llama.cpp/issues/1437 | -| `-ngl, --gpu-layers N` | number of layers to store in VRAM
(env: LLAMA_ARG_N_GPU_LAYERS) | +| `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM
(env: LLAMA_ARG_N_GPU_LAYERS) | | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:
- none: use one GPU only
- layer (default): split layers and KV across GPUs
- row: split rows across GPUs | | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1 | | `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0) | @@ -109,7 +104,7 @@ The project is under active development, and we are [looking for feedback and co | `--control-vector FNAME` | add a control vector
note: this argument can be repeated to add multiple control vectors | | `--control-vector-scaled FNAME SCALE` | add a control vector with user defined scaling SCALE
note: this argument can be repeated to add multiple scaled control vectors | | `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive | -| `-a, --alias STRING` | set alias for model name (to be used by REST API)
(env: LLAMA_ARG_MODEL) | +| `-a, --alias STRING` | set alias for model name (to be used by REST API) | | `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)
(env: LLAMA_ARG_MODEL) | | `-mu, --model-url MODEL_URL` | model download url (default: unused)
(env: LLAMA_ARG_MODEL_URL) | | `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)
(env: LLAMA_ARG_HF_REPO) | @@ -123,7 +118,7 @@ The project is under active development, and we are [looking for feedback and co | `--api-key-file FNAME` | path to file containing API keys (default: none) | | `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key | | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate | -| `--timeout N` | server read/write timeout in seconds (default: 600) | +| `-to, --timeout N` | server read/write timeout in seconds (default: 600) | | `--threads-http N` | number of threads used to process HTTP requests (default: -1)
(env: LLAMA_ARG_THREADS_HTTP) | | `-spf, --system-prompt-file FNAME` | set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications | | `--log-format {text, json}` | log output format: json or text (default: json) | diff --git a/examples/server/server.cpp b/examples/server/server.cpp index de3ea313c..7495821f9 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1,6 +1,8 @@ #include "utils.hpp" +#include "arg.h" #include "common.h" +#include "sampling.h" #include "json-schema-to-grammar.h" #include "llama.h" @@ -2423,8 +2425,7 @@ int main(int argc, char ** argv) { // own arguments required by this example gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_SERVER); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SERVER)) { return 1; } diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp index d040172a5..3fdc04394 100644 --- a/examples/simple/simple.cpp +++ b/examples/simple/simple.cpp @@ -1,3 +1,4 @@ +#include "arg.h" #include "common.h" #include "llama.h" @@ -18,8 +19,7 @@ int main(int argc, char ** argv) { params.prompt = "Hello my name is"; params.n_predict = 32; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, print_usage); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) { return 1; } diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index 8f29b5a2c..214e4932b 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -1,11 +1,13 @@ +#include "arg.h" #include "common.h" +#include "sampling.h" #include "llama.h" -#include #include #include #include #include +#include #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 100 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5 @@ -27,8 +29,7 @@ struct seq_draft { int main(int argc, char ** argv) { gpt_params params; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_SPECULATIVE); - if (!gpt_params_parse(argc, argv, params, options)) { + if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) { return 1; } diff --git a/tests/test-arg-parser.cpp b/tests/test-arg-parser.cpp index 9ad91acc0..f26707910 100644 --- a/tests/test-arg-parser.cpp +++ b/tests/test-arg-parser.cpp @@ -1,3 +1,6 @@ +#include "arg.h" +#include "common.h" + #include #include #include @@ -6,18 +9,16 @@ #undef NDEBUG #include -#include "common.h" - int main(void) { gpt_params params; printf("test-arg-parser: make sure there is no duplicated arguments in any examples\n\n"); for (int ex = 0; ex < LLAMA_EXAMPLE_COUNT; ex++) { try { - auto options = gpt_params_parser_init(params, (enum llama_example)ex); + auto ctx_arg = gpt_params_parser_init(params, (enum llama_example)ex); std::unordered_set seen_args; std::unordered_set seen_env_vars; - for (const auto & opt : options) { + for (const auto & opt : ctx_arg.options) { // check for args duplications for (const auto & arg : opt.args) { if (seen_args.find(arg) == seen_args.end()) { @@ -52,40 +53,51 @@ int main(void) { }; std::vector argv; - auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON); printf("test-arg-parser: test invalid usage\n\n"); + // missing value argv = {"binary_name", "-m"}; - assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); + // wrong value (int) argv = {"binary_name", "-ngl", "hello"}; - assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); + // wrong value (enum) argv = {"binary_name", "-sm", "hello"}; - assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); + + // non-existence arg in specific example (--draft cannot be used outside llama-speculative) + argv = {"binary_name", "--draft", "123"}; + assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SERVER)); printf("test-arg-parser: test valid usage\n\n"); argv = {"binary_name", "-m", "model_file.gguf"}; - assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(params.model == "model_file.gguf"); argv = {"binary_name", "-t", "1234"}; - assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(params.cpuparams.n_threads == 1234); argv = {"binary_name", "--verbose"}; - assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(params.verbosity == 1); argv = {"binary_name", "-m", "abc.gguf", "--predict", "6789", "--batch-size", "9090"}; - assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(params.model == "abc.gguf"); assert(params.n_predict == 6789); assert(params.n_batch == 9090); + // --draft cannot be used outside llama-speculative + argv = {"binary_name", "--draft", "123"}; + assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SPECULATIVE)); + assert(params.n_draft == 123); + // skip this part on windows, because setenv is not supported #ifdef _WIN32 printf("test-arg-parser: skip on windows build\n"); @@ -94,12 +106,12 @@ int main(void) { setenv("LLAMA_ARG_THREADS", "blah", true); argv = {"binary_name"}; - assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(false == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); setenv("LLAMA_ARG_MODEL", "blah.gguf", true); setenv("LLAMA_ARG_THREADS", "1010", true); argv = {"binary_name"}; - assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(params.model == "blah.gguf"); assert(params.cpuparams.n_threads == 1010); @@ -109,7 +121,7 @@ int main(void) { setenv("LLAMA_ARG_MODEL", "blah.gguf", true); setenv("LLAMA_ARG_THREADS", "1010", true); argv = {"binary_name", "-m", "overwritten.gguf"}; - assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, options)); + assert(true == gpt_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON)); assert(params.model == "overwritten.gguf"); assert(params.cpuparams.n_threads == 1010); #endif // _WIN32 From fb3f2498156b3140e2050ec9c7bf61372f63ff56 Mon Sep 17 00:00:00 2001 From: slaren Date: Tue, 10 Sep 2024 08:23:33 +0200 Subject: [PATCH 29/32] make : do not run llama-gen-docs when building (#9399) --- Makefile | 1 - 1 file changed, 1 deletion(-) diff --git a/Makefile b/Makefile index 97ef37c0e..c12bc61f4 100644 --- a/Makefile +++ b/Makefile @@ -1454,7 +1454,6 @@ llama-gen-docs: examples/gen-docs/gen-docs.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - ./llama-gen-docs libllava.a: examples/llava/llava.cpp \ examples/llava/llava.h \ From 0b4ac75772b744bb0a0d674927587621d1057884 Mon Sep 17 00:00:00 2001 From: Molly Sophia Date: Tue, 10 Sep 2024 15:02:30 +0800 Subject: [PATCH 30/32] RWKV v6: Add time_mix_decay_w1/w2 in quant exclusion list (#9387) Signed-off-by: Molly Sophia --- convert_hf_to_gguf.py | 2 ++ src/llama.cpp | 2 ++ 2 files changed, 4 insertions(+) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 0a9bbc829..ca473244e 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -302,6 +302,8 @@ class Model: gguf.MODEL_TENSOR.TIME_MIX_FIRST, gguf.MODEL_TENSOR.TIME_MIX_W1, gguf.MODEL_TENSOR.TIME_MIX_W2, + gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1, + gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2, ) ) or not new_name.endswith(".weight") diff --git a/src/llama.cpp b/src/llama.cpp index 39e20440e..ee27cbd1c 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -17530,6 +17530,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s quantize &= name.find("time_mix_first.weight") == std::string::npos; quantize &= name.find("time_mix_w1.weight") == std::string::npos; quantize &= name.find("time_mix_w2.weight") == std::string::npos; + quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos; + quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos; // do not quantize relative position bias (T5) quantize &= name.find("attn_rel_b.weight") == std::string::npos; From 83008b7cfe90ad89d0c0ed2c2424fd75edc25ac1 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Tue, 10 Sep 2024 09:03:21 +0200 Subject: [PATCH 31/32] llama : update llm_build_copy_mask_state comment [no ci] (#9385) This commit updates the comment, which seems to contain a typo or be an outdated comment, in the copy_mask_state function changing the variable n_rs to n_kv. I believe this change is correct and what the comment wants to convey is to copy the states that are not going to be used in the upcoming processing, which are the tokens states from n_seqs up to the number of possible token states n_kv. --- src/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/llama.cpp b/src/llama.cpp index ee27cbd1c..40db03517 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -9258,7 +9258,7 @@ static struct ggml_tensor * llm_build_copy_mask_state( // FIXME: zero-out NANs? states = ggml_mul(ctx, states, state_mask); - // copy states which won't be changed further (between n_seqs and n_rs) + // copy states which won't be changed further (between n_seqs and n_kv) ggml_build_forward_expand(graph, ggml_cpy(ctx, ggml_view_1d(ctx, states, n_state*(n_kv - n_seqs), n_seqs*n_state*ggml_element_size(states)), From 00ba2ff78100e187ae17987bacd1c916211718b2 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 10 Sep 2024 10:17:03 +0300 Subject: [PATCH 32/32] metal : fix compile warning with GGML_METAL_NDEBUG (#0) --- ggml/src/ggml-metal.m | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m index f04e5af71..6d8a7c898 100644 --- a/ggml/src/ggml-metal.m +++ b/ggml/src/ggml-metal.m @@ -17,8 +17,8 @@ #define GGML_METAL_LOG_WARN(...) #define GGML_METAL_LOG_ERROR(...) #else -#define GGML_METAL_LOG_INFO(...) ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__) -#define GGML_METAL_LOG_WARN(...) ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__) +#define GGML_METAL_LOG_INFO(...) ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__) +#define GGML_METAL_LOG_WARN(...) ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__) #define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) #endif @@ -3039,8 +3039,7 @@ static enum ggml_status ggml_metal_graph_compute( if (status != MTLCommandBufferStatusCompleted) { GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status); if (status == MTLCommandBufferStatusError) { - NSString * error_code = [command_buffer error].localizedDescription; - GGML_METAL_LOG_INFO("error: %s\n", [error_code UTF8String]); + GGML_METAL_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]); } return GGML_STATUS_FAILED;