diff --git a/CMakeLists.txt b/CMakeLists.txt index ffda74a70..34a897327 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -386,11 +386,6 @@ if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES if (MSVC) # TODO: arm msvc? else() - if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") - # Apple M1, M2, etc. - # Raspberry Pi 3, 4, Zero 2 (64-bit) - add_compile_options(-mcpu=native) - endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6") # Raspberry Pi 1, Zero add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access) diff --git a/README.md b/README.md index ee56988c7..e890dc9c2 100644 --- a/README.md +++ b/README.md @@ -85,6 +85,7 @@ as the main playground for developing new features for the [ggml](https://github - [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy) - [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b) - [X] [WizardLM](https://github.com/nlpxucan/WizardLM) +- [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) **Bindings:** diff --git a/convert.py b/convert.py index e340d2273..142692776 100644 --- a/convert.py +++ b/convert.py @@ -136,7 +136,7 @@ def find_n_mult(n_ff: int, n_embd: int) -> int: calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult if calc_ff == n_ff: return n_mult - return 1 + raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).") @dataclass class Params: @@ -321,6 +321,10 @@ class Tensor(metaclass=ABCMeta): @abstractmethod def permute(self, n_head: int) -> 'Tensor': ... @abstractmethod + def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ... + @abstractmethod + def part(self, n_part: int) -> 'UnquantizedTensor': ... + @abstractmethod def to_ggml(self) -> 'GGMLCompatibleTensor': ... @@ -345,6 +349,14 @@ class UnquantizedTensor(Tensor): def to_ggml(self) -> 'UnquantizedTensor': return self + def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': + r = self.ndarray.shape[0] // 3 + return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head)) + + def part(self, n_part: int) -> 'UnquantizedTensor': + r = self.ndarray.shape[0] // 3 + return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...]) + def permute(self, n_head: int) -> 'UnquantizedTensor': return UnquantizedTensor(permute(self.ndarray, n_head)) @@ -642,6 +654,19 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int) -> LazyTensor: return lazy_tensor.load().permute(n_head) return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description) +def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor: + def load() -> Tensor: + return lazy_tensor.load().permute_part(n_part, n_head) + s = lazy_tensor.shape.copy() + s[0] = s[0] // 3 + return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description) + +def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor: + def load() -> Tensor: + return lazy_tensor.load().part(n_part) + s = lazy_tensor.shape.copy() + s[0] = s[0] // 3 + return LazyTensor(load, s, lazy_tensor.data_type, 'part ' + lazy_tensor.description) def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel: out: LazyModel = {} @@ -650,11 +675,17 @@ def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel: out["output.weight"] = model["lm_head.weight"] for i in itertools.count(): - if f"model.layers.{i}.self_attn.q_proj.weight" not in model: + if f"model.layers.{i}.self_attn.q_proj.weight" in model: + out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head) + out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head) + out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"] + elif f"model.layers.{i}.self_attn.W_pack.weight" in model: + out[f"layers.{i}.attention.wq.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head) + out[f"layers.{i}.attention.wk.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head) + out[f"layers.{i}.attention.wv.weight"] = part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 2) + else: break - out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head) - out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head) - out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"] + out[f"layers.{i}.attention.wo.weight"] = model[f"model.layers.{i}.self_attn.o_proj.weight"] out[f"layers.{i}.feed_forward.w1.weight"] = model[f"model.layers.{i}.mlp.gate_proj.weight"] diff --git a/examples/common.cpp b/examples/common.cpp index 5addd10a1..3278a0643 100644 --- a/examples/common.cpp +++ b/examples/common.cpp @@ -110,7 +110,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { invalid_param = true; break; } - params.seed = std::stoi(argv[i]); + params.seed = std::stoul(argv[i]); } else if (arg == "-t" || arg == "--threads") { if (++i >= argc) { invalid_param = true; diff --git a/examples/common.h b/examples/common.h index 111caf61b..951511f8a 100644 --- a/examples/common.h +++ b/examples/common.h @@ -22,7 +22,7 @@ int32_t get_num_physical_cores(); struct gpt_params { - int32_t seed = -1; // RNG seed + uint32_t seed = -1; // RNG seed int32_t n_threads = get_num_physical_cores(); int32_t n_predict = -1; // new tokens to predict int32_t n_ctx = 512; // context size diff --git a/examples/embd-input/embd-input-lib.cpp b/examples/embd-input/embd-input-lib.cpp index 37de52ad6..570e273fc 100644 --- a/examples/embd-input/embd-input-lib.cpp +++ b/examples/embd-input/embd-input-lib.cpp @@ -210,9 +210,12 @@ llama_token sampling_id(struct MyModel* mymodel) { const char * sampling(struct MyModel * mymodel) { llama_context * ctx = mymodel->ctx; int id = sampling_id(mymodel); - std::string ret; - if (id == llama_token_eos()) ret = ""; - else ret = llama_token_to_str(ctx, id); + static std::string ret; + if (id == llama_token_eos()) { + ret = ""; + } else { + ret = llama_token_to_str(ctx, id); + } eval_id(mymodel, id); return ret.c_str(); } diff --git a/examples/embd-input/embd-input.h b/examples/embd-input/embd-input.h index 4fefabd42..efb5ba5e2 100644 --- a/examples/embd-input/embd-input.h +++ b/examples/embd-input/embd-input.h @@ -5,7 +5,6 @@ #include "llama.h" #include "build-info.h" - extern "C" { typedef struct MyModel { @@ -14,14 +13,13 @@ typedef struct MyModel { int n_past = 0; } MyModel; - struct MyModel* create_mymodel(int argc, char ** argv); bool eval_float(void* model, float* input, int N); bool eval_tokens(void* model, std::vector tokens); bool eval_id(struct MyModel* mymodel, int id); bool eval_string(struct MyModel* mymodel, const char* str); -const char* sampling(struct MyModel* mymodel); +const char * sampling(struct MyModel* mymodel); llama_token sampling_id(struct MyModel* mymodel); void free_mymodel(struct MyModel* mymodel); diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 3cd5bb794..2b7eb39c5 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -24,11 +24,11 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); - if (params.seed < 0) { + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } - fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); + fprintf(stderr, "%s: seed = %u\n", __func__, params.seed); std::mt19937 rng(params.seed); if (params.random_prompt) { diff --git a/examples/main/README.md b/examples/main/README.md index 9ba1eb384..375386130 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -242,7 +242,7 @@ Example usage: `--logit-bias 29905-inf` ### RNG Seed -- `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed). +- `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, -1 = random seed). The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run. diff --git a/examples/main/main.cpp b/examples/main/main.cpp index bcdc98d61..3a171925b 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -94,11 +94,11 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); - if (params.seed < 0) { + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } - fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); + fprintf(stderr, "%s: seed = %u\n", __func__, params.seed); std::mt19937 rng(params.seed); if (params.random_prompt) { diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index f8a6cb516..dd54ed3c4 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -136,11 +136,11 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); - if (params.seed < 0) { + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } - fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); + fprintf(stderr, "%s: seed = %u\n", __func__, params.seed); std::mt19937 rng(params.seed); if (params.random_prompt) { diff --git a/examples/server/README.md b/examples/server/README.md index fa95c0044..ba4b2fec9 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -152,7 +152,7 @@ node . `mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1). - `seed`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed). + `seed`: Set the random number generator (RNG) seed (default: -1, -1 = random seed). `ignore_eos`: Ignore end of stream token and continue generating (default: false). diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp index a05881d16..c50eeb343 100644 --- a/examples/train-text-from-scratch/train-text-from-scratch.cpp +++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp @@ -2671,7 +2671,8 @@ struct train_params { const char * fn_checkpoint_out; const char * fn_model_out; - int seed; + uint32_t seed; + int n_ctx; int n_embd; int n_mult; @@ -2768,7 +2769,7 @@ void train_print_usage(int /*argc*/, char ** argv, const struct train_params * p fprintf(stderr, " --checkpoint-in FNAME path from which to load training checkpoint (default '%s')\n", params->fn_checkpoint_in); fprintf(stderr, " --checkpoint-out FNAME path to save training checkpoint (default '%s')\n", params->fn_checkpoint_out); fprintf(stderr, " --model-out FNAME path to save ggml model (default '%s')\n", params->fn_model_out); - fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for < 0)\n"); + fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for -1)\n"); fprintf(stderr, " -c N, --ctx N Context size used during training (default %d)\n", params->n_ctx); fprintf(stderr, " --embd N Embedding size used for new models (default %d)\n", params->n_embd); fprintf(stderr, " --mult N Mult size used for new models, influences feedforward size. (default %d)\n", params->n_mult); @@ -3034,10 +3035,10 @@ int main(int argc, char ** argv) { return 1; } - if (params.seed < 0) { + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } - printf("%s: seed: %d\n", __func__, params.seed); + printf("%s: seed: %u\n", __func__, params.seed); srand(params.seed); struct llama_context_params llama_params = llama_context_default_params(); diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 4e0d3dbde..50df20edd 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -214,6 +214,11 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2"); #endif +struct ggml_tensor_extra_gpu { + void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors + cudaEvent_t events[GGML_CUDA_MAX_DEVICES]; // events for synchronizing multiple GPUs +}; + static __global__ void add_f32(const float * x, const float * y, float * dst, const int k) { const int i = blockDim.x*blockIdx.x + threadIdx.x; @@ -1970,7 +1975,6 @@ inline void ggml_cuda_op_add( } else { GGML_ASSERT(false); } - CUDA_CHECK(cudaGetLastError()); (void) src1; (void) dst; @@ -2002,7 +2006,6 @@ inline void ggml_cuda_op_mul( // compute mul_f32_cuda(src0_ddf_i01, src1_ddf_i01, dst_ddf_i01, ne00, ne10, cudaStream_main); - CUDA_CHECK(cudaGetLastError()); } (void) dst; @@ -2023,7 +2026,6 @@ inline void ggml_cuda_op_silu( // compute silu_f32_cuda(src0_ddf_i, dst_ddf_i, ne00*i01_diff, cudaStream_main); - CUDA_CHECK(cudaGetLastError()); (void) src1; (void) dst; @@ -2046,7 +2048,6 @@ inline void ggml_cuda_op_rms_norm( // compute rms_norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main); - CUDA_CHECK(cudaGetLastError()); (void) src1; (void) dst; @@ -2125,7 +2126,6 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec( GGML_ASSERT(false); break; } - CUDA_CHECK(cudaGetLastError()); #ifdef GGML_CUDA_DMMV_F16 if (src1_convert_f16) { @@ -2202,7 +2202,6 @@ inline void ggml_cuda_op_rope( // compute rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p, theta_scale, cudaStream_main); - CUDA_CHECK(cudaGetLastError()); (void) dst; (void) src0_ddq_i; @@ -2226,7 +2225,6 @@ inline void ggml_cuda_op_diag_mask_inf( // compute diag_mask_inf_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_past, cudaStream_main); - CUDA_CHECK(cudaGetLastError()); (void) dst; (void) src0_ddq_i; @@ -2248,7 +2246,6 @@ inline void ggml_cuda_op_soft_max( // compute soft_max_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main); - CUDA_CHECK(cudaGetLastError()); (void) src1; (void) dst; @@ -2344,10 +2341,11 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm size_t src1_asf[GGML_CUDA_MAX_DEVICES] = {0}; size_t dst_asf[GGML_CUDA_MAX_DEVICES] = {0}; - // if multiple GPUs are used they need to wait for the main GPU to finish + // if multiple devices are used they need to wait for the main device + // here an event is recorded that signifies that the main device has finished calculating the input data if (split && g_device_count > 1) { CUDA_CHECK(cudaSetDevice(g_main_device)); - CUDA_CHECK(cudaDeviceSynchronize()); + CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device], g_cudaStreams_main[g_main_device])); } for (int id = 0; id < g_device_count; ++id) { @@ -2373,6 +2371,12 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm int64_t row_diff = row_high - row_low; cudaSetDevice(id); + cudaStream_t cudaStream_main = g_cudaStreams_main[id]; + + // wait for main GPU data if necessary + if (split && id != g_main_device) { + CUDA_CHECK(cudaStreamWaitEvent(cudaStream_main, src0_extra->events[g_main_device])); + } if (src0_on_device && src0_is_contiguous) { if (src0_is_f32) { @@ -2448,8 +2452,6 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm } const int64_t i11 = i13*ne12 + i12; - cudaStream_t cudaStream_main = g_cudaStreams_main[id]; - // for split tensors the data begins at i0 == i0_offset_low char * src0_ddq_i = src0_ddq[id] + (i0 - i0_offset_low)*src0_stride*src0_ts/src0_bs; float * src0_ddf_i = src0_ddf[id] + (i0 - i0_offset_low)*src0_stride; @@ -2509,6 +2511,7 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm // do the computation op(src0, src1, dst, src0_ddq_i, src0_ddf_i, src1_ddf_i, dst_ddf_i, i02, i01_low, i01_high, i11, cudaStream_main); + CUDA_CHECK(cudaGetLastError()); // copy dst to host or other device if necessary if (!dst_on_device) { @@ -2538,6 +2541,11 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_ddf_i, dst_stride*sizeof(float), kind, cudaStream_main)); } } + + // signify to main device that other device is done + if (split && g_device_count > 1 && id != g_main_device) { + CUDA_CHECK(cudaEventRecord(src0_extra->events[id], cudaStream_main)); + } } } } @@ -2549,7 +2557,6 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm } CUDA_CHECK(cudaSetDevice(id)); - CUDA_CHECK(cudaDeviceSynchronize()); if (src0_asq[id] > 0) { ggml_cuda_pool_free(src0_ddq[id], src0_asq[id]); @@ -2564,6 +2571,21 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm ggml_cuda_pool_free(dst_ddf[id], dst_asf[id]); } } + + // main device waits for all other devices to be finished + if (split && g_device_count > 1) { + CUDA_CHECK(cudaSetDevice(g_main_device)); + for (int id = 0; id < g_device_count; ++id) { + if (id != g_main_device) { + CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams_main[g_main_device], src0_extra->events[id])); + } + } + } + + if (dst->backend == GGML_BACKEND_CPU) { + CUDA_CHECK(cudaSetDevice(g_main_device)); + CUDA_CHECK(cudaDeviceSynchronize()); + } } void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -2803,6 +2825,10 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) { cudaMemcpy(buf, buf_host, size, cudaMemcpyHostToDevice); extra->data_device[id] = buf; + + if (backend == GGML_BACKEND_GPU_SPLIT) { + CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id], cudaEventDisableTiming)); + } } tensor->extra = extra; @@ -2816,12 +2842,15 @@ void ggml_cuda_free_data(struct ggml_tensor * tensor) { ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; for (int id = 0; id < g_device_count; ++id) { - if (extra->data_device[id] == nullptr) { - continue; + if (extra->data_device[id] != nullptr) { + CUDA_CHECK(cudaSetDevice(id)); + CUDA_CHECK(cudaFree(extra->data_device[id])); } - CUDA_CHECK(cudaSetDevice(id)); - CUDA_CHECK(cudaFree(extra->data_device[id])); + if (extra->events[id] != nullptr) { + CUDA_CHECK(cudaSetDevice(id)); + CUDA_CHECK(cudaEventDestroy(extra->events[id])); + } } delete extra; diff --git a/ggml-cuda.h b/ggml-cuda.h index 7a65a3558..3c1e8deb6 100644 --- a/ggml-cuda.h +++ b/ggml-cuda.h @@ -8,10 +8,6 @@ extern "C" { #define GGML_CUDA_MAX_DEVICES 16 -struct ggml_tensor_extra_gpu { - void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors -}; - void ggml_init_cublas(void); void ggml_cuda_set_tensor_split(const float * tensor_split); diff --git a/ggml-metal.m b/ggml-metal.m index 7551231b9..fd69c41fe 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -202,7 +202,9 @@ struct ggml_metal_context * ggml_metal_init(void) { void ggml_metal_free(struct ggml_metal_context * ctx) { fprintf(stderr, "%s: deallocating\n", __func__); - + for (int i = 0; i < ctx->n_buffers; ++i) { + [ctx->buffers[i].metal release]; + } free(ctx); } diff --git a/ggml-opencl.cpp b/ggml-opencl.cpp index 95f4cec6d..fed4ffb0c 100644 --- a/ggml-opencl.cpp +++ b/ggml-opencl.cpp @@ -21,11 +21,19 @@ #define CL_DMMV_BLOCK_SIZE 32 +#ifndef K_QUANTS_PER_ITERATION +#define K_QUANTS_PER_ITERATION 1 +#else +static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2"); +#endif + #define MULTILINE_QUOTE(...) #__VA_ARGS__ static std::string program_source = MULTILINE_QUOTE( typedef char int8_t; typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; typedef int int32_t; typedef uint uint32_t; @@ -175,7 +183,9 @@ void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float *v0 = vload_half(0, &x[ib + 0]); *v1 = vload_half(0, &x[ib + 1]); } +); +static std::string k_quants_source = MULTILINE_QUOTE( inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8_t *m) { if (j < 4) @@ -199,7 +209,7 @@ __kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __globa const int is = 8 * n + l / 16; const uint8_t q = x[i].qs[32 * n + l]; - __global float *y = yy + i * 256 + 128 * n; + __global float *y = yy + i * QK_K + 128 * n; const float dall = vload_half(0, &x[i].d); const float dmin = vload_half(0, &x[i].dmin); @@ -231,7 +241,7 @@ __kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __globa float d_all = vload_half(0, &x[i].d); float dl = d_all * (us - 32); - __global float *y = yy + i * 256 + 128 * n + 32 * j; + __global float *y = yy + i * QK_K + 128 * n + 32 * j; const __global uint8_t *q = x[i].qs + 32 * n; const __global uint8_t *hm = x[i].hmask; @@ -248,7 +258,7 @@ __kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __globa const int is = 2 * il; const int n = 4; - __global float *y = yy + i * 256 + 64 * il + n * ir; + __global float *y = yy + i * QK_K + 64 * il + n * ir; const float dall = vload_half(0, &x[i].d); const float dmin = vload_half(0, &x[i].dmin); @@ -277,7 +287,7 @@ __kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __globa const int ir = tid % 16; const int is = 2 * il; - __global float *y = yy + i * 256 + 64 * il + 2 * ir; + __global float *y = yy + i * QK_K + 64 * il + 2 * ir; const float dall = vload_half(0, &x[i].d); const float dmin = vload_half(0, &x[i].dmin); @@ -309,7 +319,7 @@ __kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __globa const int il = tid - 32 * ip; const int is = 8 * ip + il / 16; - __global float *y = yy + i * 256 + 128 * ip + il; + __global float *y = yy + i * QK_K + 128 * ip + il; const float d = vload_half(0, &x[i].d); @@ -323,161 +333,383 @@ __kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __globa y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32); } +__kernel void dequantize_mul_mat_vec_q2_K(__global const struct block_q2_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) { -void vec_dot_q2_K(__global const struct block_q2_K* x, const int ib, const int iqs, const __global float *yy, float *result) { + const int row = get_group_id(0); - int n = iqs / 128; - int r = iqs - 128 * n; - int l = r / 8; + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; - __global const float *y = yy + 128 * n + l; - __global const uint8_t *q = x[ib].qs + 32 * n + l; - __global const uint8_t *s = x[ib].scales + 8 * n; + __global const struct block_q2_K * x = xx + ib0; - const float dall = vload_half(0, &x[ib].d); - const float dmin = vload_half(0, &x[ib].dmin); + const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...15 + const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0,1 - float sum = y[ 0] * (dall * ((s[0] & 0xF) * ((q[ 0] >> 0) & 3)) - dmin * (s[0] >> 4)) - + y[ 32] * (dall * ((s[2] & 0xF) * ((q[ 0] >> 2) & 3)) - dmin * (s[2] >> 4)) - + y[ 64] * (dall * ((s[4] & 0xF) * ((q[ 0] >> 4) & 3)) - dmin * (s[4] >> 4)) - + y[ 96] * (dall * ((s[6] & 0xF) * ((q[ 0] >> 6) & 3)) - dmin * (s[6] >> 4)) - + y[ 16] * (dall * ((s[1] & 0xF) * ((q[16] >> 0) & 3)) - dmin * (s[1] >> 4)) - + y[ 48] * (dall * ((s[3] & 0xF) * ((q[16] >> 2) & 3)) - dmin * (s[3] >> 4)) - + y[ 80] * (dall * ((s[5] & 0xF) * ((q[16] >> 4) & 3)) - dmin * (s[5] >> 4)) - + y[112] * (dall * ((s[7] & 0xF) * ((q[16] >> 6) & 3)) - dmin * (s[7] >> 4)); + const int step = 16/K_QUANTS_PER_ITERATION; - *result = sum; -} + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0...15 or 0...7 -void vec_dot_q3_K(__global const struct block_q3_K* x, const int ib, const int iqs, const __global float *yy, float *result) { + const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2 + const int q_offset = 32*im + l0; + const int s_offset = 8*im; + const int y_offset = 128*im + l0; - const uint32_t kmask1 = 0x03030303; - const uint32_t kmask2 = 0x0f0f0f0f; + tmp[16 * ix + tid] = 0; - uint32_t aux[3]; - uint32_t utmp[4]; + uint32_t aux[4]; + const uint8_t * d = (const uint8_t *)aux; + const uint8_t * m = (const uint8_t *)(aux + 2); - int n = iqs/128; - int r = iqs - 128*n; - int l = r/8; + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - __global const float * y = yy + 128*n + l; - __global const uint8_t * q = x[ib].qs + 32*n + l; - __global const uint8_t * hm = x[ib].hmask + l; - const int8_t * s = (const int8_t *)utmp + 8*n; + __global const float * y = yy + i * QK_K + y_offset; + __global const uint8_t * q = x[i].qs + q_offset; - aux[0] = x[ib].scales[0] | x[ib].scales[1] << 8 | x[ib].scales[2] << 16 | x[ib].scales[3] << 24; - aux[1] = x[ib].scales[4] | x[ib].scales[5] << 8 | x[ib].scales[6] << 16 | x[ib].scales[7] << 24; - aux[2] = x[ib].scales[8] | x[ib].scales[9] << 8 | x[ib].scales[10] << 16 | x[ib].scales[11] << 24; + const float dall = vload_half(0, &x[i].d); + const float dmin = vload_half(0, &x[i].dmin); - utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4); - utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4); - utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4); - utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4); + __global const uint32_t * a = (__global const uint32_t *)(x[i].scales + s_offset); + aux[0] = a[0] & 0x0f0f0f0f; + aux[1] = a[1] & 0x0f0f0f0f; + aux[2] = (a[0] >> 4) & 0x0f0f0f0f; + aux[3] = (a[1] >> 4) & 0x0f0f0f0f; - const float dall = vload_half(0, &x[ib].d); - const uint8_t m = 1 << (4*n); + float sum1 = 0, sum2 = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3) + + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3) + + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3) + + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3) + + y[l+16] * d[1] * ((q[l+16] >> 0) & 3) + + y[l+48] * d[3] * ((q[l+16] >> 2) & 3) + + y[l+80] * d[5] * ((q[l+16] >> 4) & 3) + +y[l+112] * d[7] * ((q[l+16] >> 6) & 3); + sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6] + + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7]; - float sum = y[ 0] * (s[0] - 32) * (((q[ 0] >> 0) & 3) - (hm[ 0] & (m << 0) ? 0 : 4)) - + y[ 32] * (s[2] - 32) * (((q[ 0] >> 2) & 3) - (hm[ 0] & (m << 1) ? 0 : 4)) - + y[ 64] * (s[4] - 32) * (((q[ 0] >> 4) & 3) - (hm[ 0] & (m << 2) ? 0 : 4)) - + y[ 96] * (s[6] - 32) * (((q[ 0] >> 6) & 3) - (hm[ 0] & (m << 3) ? 0 : 4)) - + y[ 16] * (s[1] - 32) * (((q[16] >> 0) & 3) - (hm[16] & (m << 0) ? 0 : 4)) - + y[ 48] * (s[3] - 32) * (((q[16] >> 2) & 3) - (hm[16] & (m << 1) ? 0 : 4)) - + y[ 80] * (s[5] - 32) * (((q[16] >> 4) & 3) - (hm[16] & (m << 2) ? 0 : 4)) - + y[112] * (s[7] - 32) * (((q[16] >> 6) & 3) - (hm[16] & (m << 3) ? 0 : 4)); + } + tmp[16 * ix + tid] += dall * sum1 - dmin * sum2; - *result = sum * dall; - -} - -void vec_dot_q4_K(__global const struct block_q4_K* x, const int ib, const int iqs, const __global float *yy, float *result) { - - const int j = iqs / 64; // j is in 0...3 - const int ir = (iqs - 64*j)/2; // ir is in 0...28 in steps of 4 - const int is = 2*j; // is is in 0...6 in steps of 2 - - __global const float * y = yy + 64*j + ir; - __global const uint8_t * q = x[ib].qs + 32*j + ir; - - const float dall = vload_half(0, &x[ib].d); - const float dmin = vload_half(0, &x[ib].dmin); - - uint8_t sc, m; - get_scale_min_k4(is + 0, x[ib].scales, &sc, &m); - const float d1 = dall * sc; - const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[ib].scales, &sc, &m); - const float d2 = dall * sc; - const float m2 = dmin * m; - - float sum = 0; - for (int k = 0; k < 4; ++k) { - sum += y[k + 0] * (d1 * (q[k] & 0xF) - m1); - sum += y[k + 32] * (d2 * (q[k] >> 4) - m2); } - *result = sum; + // sum up partial sums and write back result + barrier(CLK_LOCAL_MEM_FENCE); + for (int s=16; s>0; s>>=1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + barrier(CLK_LOCAL_MEM_FENCE); + } + if (tid == 0) { + dst[row] = tmp[0]; + } } -void vec_dot_q5_K(__global const struct block_q5_K* x, const int ib, const int iqs, const __global float *yy, float *result) { +__kernel void dequantize_mul_mat_vec_q3_K(__global const struct block_q3_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) { + const uint16_t kmask1 = 0x0303; + const uint16_t kmask2 = 0x0f0f; - const int j = iqs / 64; - const int ir = (iqs - 64*j)/2; - const int is = 2*j; + const int row = get_group_id(0); - __global const float * y = yy + 64*j + ir; - __global const uint8_t * ql = x[ib].qs + 32*j + ir; - __global const uint8_t * qh = x[ib].qh + ir; + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; - const float dall = vload_half(0, &x[ib].d); - const float dmin = vload_half(0, &x[ib].dmin); + __global const struct block_q3_K * x = xx + ib0; - uint8_t sc, m; - get_scale_min_k4(is + 0, x[ib].scales, &sc, &m); - const float d1 = dall * sc; - const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[ib].scales, &sc, &m); - const float d2 = dall * sc; - const float m2 = dmin * m; + const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop + const int step = 16/K_QUANTS_PER_ITERATION; + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0....15 or 0...7 + + const uint8_t m = 1 << (4*im); + + const int l0 = n*in; // 0...15 or 0...14 in steps of 2 + const int q_offset = 32*im + l0; + const int y_offset = 128*im + l0; + + uint16_t utmp[4]; + const int8_t * s = (const int8_t *)utmp; + + const uint16_t s_shift = 4*im; + + tmp[16 * ix + tid] = 0; + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + __global const float * y = yy + i * QK_K + y_offset; + __global const uint8_t * q = x[i].qs + q_offset; + __global const uint8_t * h = x[i].hmask + l0; + + __global const uint16_t * a = (__global const uint16_t *)x[i].scales; + utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4); + utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4); + utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4); + utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4); + + const float d = vload_half(0, &x[i].d); + + float sum = 0; + for (int l = 0; l < n; ++l) { + sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4)) + + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4)) + + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4)) + + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4)); + sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4)) + + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4)) + + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4)) + + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4)); + } + tmp[16 * ix + tid] += d * sum; - uint8_t hm = 1 << is; - float sum = 0; - for (int k = 0; k < 4; ++k) { - sum += y[k + 0] * (d1 * ((ql[k] & 0xF) + (qh[k] & hm ? 16 : 0)) - m1); } - hm <<= 1; - for (int k = 0; k < 4; ++k) { - sum += y[k + 32] * (d2 * ((ql[k] >> 4) + (qh[k] & hm ? 16 : 0)) - m2); - } - *result = sum; + // sum up partial sums and write back result + barrier(CLK_LOCAL_MEM_FENCE); + for (int s=16; s>0; s>>=1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + barrier(CLK_LOCAL_MEM_FENCE); + } + if (tid == 0) { + dst[row] = tmp[0]; + } } -void vec_dot_q6_K(__global const struct block_q6_K* x, const int ib, const int iqs, const __global float *yy, float *result) { +__kernel void dequantize_mul_mat_vec_q4_K(__global const struct block_q4_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) { + //to rename it later, just to test now + const uint16_t kmask1 = 0x3f3f; + const uint16_t kmask2 = 0x0f0f; + const uint16_t kmask3 = 0xc0c0; - const int ip = iqs / 128; // 0 or 1 - const int il = (iqs - 128*ip)/8; // 0...15 - const int is = 8*ip; + const int row = get_group_id(0); + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; - __global const float * y = yy + 128*ip + il; + const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...15 + const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; - const float d = vload_half(0, &x[ib].d); + const int step = 8/K_QUANTS_PER_ITERATION; - __global const uint8_t * ql = x[ib].ql + 64*ip + il; - __global const uint8_t * qh = x[ib].qh + 32*ip + il; - __global const int8_t * sc = x[ib].scales + is; + const int il = tid/step; // 0...3 + const int ir = tid - step*il;// 0...3 + const int n = 2*K_QUANTS_PER_ITERATION; - *result = y[ 0] * d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh[ 0] >> 0) & 3) << 4)) - 32) - + y[ 32] * d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh[ 0] >> 2) & 3) << 4)) - 32) - + y[ 64] * d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh[ 0] >> 4) & 3) << 4)) - 32) - + y[ 96] * d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh[ 0] >> 6) & 3) << 4)) - 32) - + y[ 16] * d * sc[1] * ((int8_t)((ql[16] & 0xF) | (((qh[16] >> 0) & 3) << 4)) - 32) - + y[ 48] * d * sc[3] * ((int8_t)((ql[48] & 0xF) | (((qh[16] >> 2) & 3) << 4)) - 32) - + y[ 80] * d * sc[5] * ((int8_t)((ql[16] >> 4) | (((qh[16] >> 4) & 3) << 4)) - 32) - + y[112] * d * sc[7] * ((int8_t)((ql[48] >> 4) | (((qh[16] >> 6) & 3) << 4)) - 32); + const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 + const int in = il%2; + const int l0 = n*(2*ir + in); + const int q_offset = 32*im + l0; + const int y_offset = 64*im + l0; + + uint16_t aux[4]; + const uint8_t * sc = (const uint8_t *)aux; + + __global const struct block_q4_K * x = xx + ib0; + + tmp[16 * ix + tid] = 0; + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + __global const uint8_t * q1 = x[i].qs + q_offset; + __global const uint8_t * q2 = q1 + 64; + __global const float * y1 = yy + i*QK_K + y_offset; + __global const float * y2 = y1 + 128; + + const float dall = vload_half(0, &x[i].d); + const float dmin = vload_half(0, &x[i].dmin); + + __global const uint16_t * a = (__global const uint16_t *)x[i].scales; + aux[0] = a[im+0] & kmask1; + aux[1] = a[im+2] & kmask1; + aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); + aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + + float4 s = (float4)(0.f); + float smin = 0; + for (int l = 0; l < n; ++l) { + s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4); + s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4); + smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; + } + tmp[16 * ix + tid] += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin; + + } + + // sum up partial sums and write back result + barrier(CLK_LOCAL_MEM_FENCE); + for (int s=16; s>0; s>>=1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + barrier(CLK_LOCAL_MEM_FENCE); + } + if (tid == 0) { + dst[row] = tmp[0]; + } +} + +__kernel void dequantize_mul_mat_vec_q5_K(__global const struct block_q5_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) { + + const uint16_t kmask1 = 0x3f3f; + const uint16_t kmask2 = 0x0f0f; + const uint16_t kmask3 = 0xc0c0; + + const int row = get_group_id(0); + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const int tid = get_local_id(0)/2; // 0...15 + const int ix = get_local_id(0)%2; + + const int il = tid/4; // 0...3 + const int ir = tid - 4*il;// 0...3 + const int n = 2; + + const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 + const int in = il%2; + + const int l0 = n*(2*ir + in); + const int q_offset = 32*im + l0; + const int y_offset = 64*im + l0; + + const uint8_t hm1 = 1 << (2*im); + const uint8_t hm2 = hm1 << 4; + + uint16_t aux[4]; + const uint8_t * sc = (const uint8_t *)aux; + + __global const struct block_q5_K * x = xx + ib0; + + tmp[16 * ix + tid] = 0; + + for (int i = ix; i < num_blocks_per_row; i += 2) { + + __global const uint8_t * ql1 = x[i].qs + q_offset; + __global const uint8_t * ql2 = ql1 + 64; + __global const uint8_t * qh = x[i].qh + l0; + __global const float * y1 = yy + i*QK_K + y_offset; + __global const float * y2 = y1 + 128; + + const float dall = vload_half(0, &x[i].d); + const float dmin = vload_half(0, &x[i].dmin); + + __global const uint16_t * a = (__global const uint16_t *)x[i].scales; + aux[0] = a[im+0] & kmask1; + aux[1] = a[im+2] & kmask1; + aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); + aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + + float4 sum = (float4)(0.f); + float smin = 0; + for (int l = 0; l < n; ++l) { + sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0)) + + y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0)); + sum.y += y1[l+32] * ((ql1[l+ 0] >> 4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0)) + + y1[l+48] * ((ql1[l+16] >> 4) + (qh[l+16] & (hm1 << 1) ? 16 : 0)); + sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0)) + + y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0)); + sum.w += y2[l+32] * ((ql2[l+ 0] >> 4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0)) + + y2[l+48] * ((ql2[l+16] >> 4) + (qh[l+16] & (hm2 << 1) ? 16 : 0)); + smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3] + + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7]; + } + tmp[16 * ix + tid] += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin; + + } + + // sum up partial sums and write back result + barrier(CLK_LOCAL_MEM_FENCE); + for (int s=16; s>0; s>>=1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + barrier(CLK_LOCAL_MEM_FENCE); + } + if (tid == 0) { + dst[row] = tmp[0]; + } +} + +__kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx, __local float* tmp, __global const float * yy, __global float * dst, const int ncols) { + + const int row = get_group_id(0); + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + __global const struct block_q6_K * x = xx + ib0; + + const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0, 1 + + const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0...15 or 0...7 + +#if K_QUANTS_PER_ITERATION == 1 + const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 + const int is = 0; +#else + const int l0 = 4 * in; // 0, 4, 8, ..., 28 + const int is = in / 4; +#endif + const int ql_offset = 64*im + l0; + const int qh_offset = 32*im + l0; + const int s_offset = 8*im + is; + const int y_offset = 128*im + l0; + + tmp[16 * ix + tid] = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + __global const float * y = yy + i * QK_K + y_offset; + __global const uint8_t * ql = x[i].ql + ql_offset; + __global const uint8_t * qh = x[i].qh + qh_offset; + __global const int8_t * s = x[i].scales + s_offset; + + const float d = vload_half(0, &x[i].d); + +#if K_QUANTS_PER_ITERATION == 1 + float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32) + + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32) + + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32) + + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32) + + y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32) + + y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32) + + y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32) + +y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32); + tmp[16 * ix + tid] += sum; +#else + float sum = 0; + for (int l = 0; l < 4; ++l) { + sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32) + + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32) + + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32) + + y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32); + } + tmp[16 * ix + tid] += sum; +#endif + + } + + // sum up partial sums and write back result + barrier(CLK_LOCAL_MEM_FENCE); + for (int s=16; s>0; s>>=1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + barrier(CLK_LOCAL_MEM_FENCE); + } + if (tid == 0) { + dst[row] = tmp[0]; + } } ); @@ -549,44 +781,6 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float } ); -std::string dequant_mul_mat_vec_k_template = MULTILINE_QUOTE( -__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) { - const int block_size = get_local_size(0); - const int row = get_group_id(0); - const int tid = get_local_id(0); - - const int iter_stride = 256; - const int vals_per_iter = iter_stride / block_size; - const int num_blocks_per_row = ncols / 256; - const int ib0 = row*num_blocks_per_row; - - tmp[tid] = 0; - - for (int i = 0; i < ncols; i += iter_stride) { - const int col = i + vals_per_iter*tid; - const int ib = ib0 + col/256; // x block index - const int iqs = col%256; // x quant index - const int iybs = col - col%256; // y block start index - - // dequantize - float v; - DOT_KERNEL(x, ib, iqs, y + iybs, &v); - tmp[tid] += v; - } - - // sum up partial sums and write back result - barrier(CLK_LOCAL_MEM_FENCE); - for (int s=block_size/2; s>0; s>>=1) { - if (tid < s) { - tmp[tid] += tmp[tid + s]; - } - barrier(CLK_LOCAL_MEM_FENCE); - } - if (tid == 0) { - dst[row] = tmp[0]; - } -} -); std::string mul_template = MULTILINE_QUOTE( __kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) { @@ -649,18 +843,6 @@ std::array mul_str_values = { "mul_f32", "float" }; -std::array dmmv_k_str_keys = { - "KERNEL_NAME", "X_TYPE", "DOT_KERNEL" -}; - -std::array dmmv_k_str_values = { - "dequantize_mul_mat_vec_q2_K", "struct block_q2_K", "vec_dot_q2_K", - "dequantize_mul_mat_vec_q3_K", "struct block_q3_K", "vec_dot_q3_K", - "dequantize_mul_mat_vec_q4_K", "struct block_q4_K", "vec_dot_q4_K", - "dequantize_mul_mat_vec_q5_K", "struct block_q5_K", "vec_dot_q5_K", - "dequantize_mul_mat_vec_q6_K", "struct block_q6_K", "vec_dot_q6_K", -}; - std::string& replace(std::string& s, const std::string& from, const std::string& to) { size_t pos = 0; while ((pos = s.find(from, pos)) != std::string::npos) { @@ -673,6 +855,7 @@ std::string& replace(std::string& s, const std::string& from, const std::string& std::string generate_kernels() { std::stringstream src; src << program_source << '\n'; + src << k_quants_source << '\n'; for (size_t i = 0; i < dequant_str_values.size(); i += dequant_str_keys.size()) { std::string dequant_kernel = dequant_template; std::string dmmv_kernel = dequant_mul_mat_vec_template; @@ -690,13 +873,6 @@ std::string generate_kernels() { } src << mul_kernel << '\n'; } - for (size_t i = 0; i < dmmv_k_str_values.size(); i += dmmv_k_str_keys.size()) { - std::string dmmv_k_kernel = dequant_mul_mat_vec_k_template; - for (size_t j = 0; j < dmmv_k_str_keys.size(); j++) { - replace(dmmv_k_kernel, dmmv_k_str_keys[j], dmmv_k_str_values[i + j]); - } - src << dmmv_k_kernel << '\n'; - } return src.str(); } @@ -729,10 +905,11 @@ static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, co exit(1); } - const char* compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math " - "-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1"; + std::string compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math " + "-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1 " + "-DQK_K=256 -DK_QUANTS_PER_ITERATION=" + std::to_string(K_QUANTS_PER_ITERATION); - err = clBuildProgram(p, 0, NULL, compile_opts, NULL, NULL); + err = clBuildProgram(p, 0, NULL, compile_opts.c_str(), NULL, NULL); if(err < 0) { clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size); diff --git a/ggml.c b/ggml.c index 684caaa37..75cc44baa 100644 --- a/ggml.c +++ b/ggml.c @@ -3846,6 +3846,40 @@ static_assert(GGML_OP_COUNT == 64, "GGML_OP_COUNT != 64"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN"); +// WARN: +// Mis-confguration can lead to problem that's hard to reason about: +// * At best it crash or talks nosense. +// * At worst it talks slightly difference but hard to perceive. +// +// An op has to enable INIT or FINALIZE when any of it's branch needs that pass. +// Take care about compile options (e.g., GGML_USE_xxx). +static bool GGML_OP_HAS_INIT [GGML_OP_COUNT] = { 0 }; +static bool GGML_OP_HAS_FINALIZE[GGML_OP_COUNT] = { 0 }; +static void ggml_setup_op_has_task_pass(void) { + { // INIT + bool * I = GGML_OP_HAS_INIT; + + I[GGML_OP_ACC ] = true; + I[GGML_OP_MUL_MAT ] = true; + I[GGML_OP_OUT_PROD ] = true; + I[GGML_OP_SET ] = true; + I[GGML_OP_GET_ROWS_BACK ] = true; + I[GGML_OP_DIAG_MASK_INF ] = true; + I[GGML_OP_DIAG_MASK_ZERO ] = true; + I[GGML_OP_CONV_1D_S1_PH ] = true; + I[GGML_OP_CONV_1D_S2_PH ] = true; + I[GGML_OP_CONV_2D_SK_P0 ] = true; + I[GGML_OP_FLASH_ATTN_BACK ] = true; + I[GGML_OP_CROSS_ENTROPY_LOSS ] = true; + } + + { // FINALIZE + bool * F = GGML_OP_HAS_FINALIZE; + + F[GGML_OP_CROSS_ENTROPY_LOSS ] = true; + } +} + // // ggml context // @@ -4267,6 +4301,8 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { ggml_cl_init(); #endif + ggml_setup_op_has_task_pass(); + is_first_call = false; } @@ -16791,9 +16827,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { if (node_n != -1) { /* FINALIZE */ struct ggml_tensor * node = state->shared->cgraph->nodes[node_n]; - params.nth = node->n_tasks; - ggml_compute_forward(¶ms, node); - ggml_graph_compute_perf_stats_node(node, state->shared); + if (GGML_OP_HAS_FINALIZE[node->op]) { + params.nth = node->n_tasks; + ggml_compute_forward(¶ms, node); + ggml_graph_compute_perf_stats_node(node, state->shared); + } } // distribute new work or execute it direct if 1T @@ -16805,10 +16843,13 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { state->shared->perf_node_start_cycles = ggml_perf_cycles(); state->shared->perf_node_start_time_us = ggml_perf_time_us(); + params.nth = node->n_tasks; + /* INIT */ - params.type = GGML_TASK_INIT; - params.nth = node->n_tasks; - ggml_compute_forward(¶ms, node); + if (GGML_OP_HAS_INIT[node->op]) { + params.type = GGML_TASK_INIT; + ggml_compute_forward(¶ms, node); + } if (node->n_tasks == 1) { // TODO: maybe push node_n to the atomic but if other threads see n_tasks is 1, @@ -16816,9 +16857,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { params.type = GGML_TASK_COMPUTE; ggml_compute_forward(¶ms, node); - params.type = GGML_TASK_FINALIZE; - ggml_compute_forward(¶ms, node); - ggml_graph_compute_perf_stats_node(node, state->shared); + if (GGML_OP_HAS_FINALIZE[node->op]) { + params.type = GGML_TASK_FINALIZE; + ggml_compute_forward(¶ms, node); + ggml_graph_compute_perf_stats_node(node, state->shared); + } } else { break; } diff --git a/ggml.h b/ggml.h index 459913222..11b51f8bd 100644 --- a/ggml.h +++ b/ggml.h @@ -444,6 +444,9 @@ extern "C" { // compute types + + // NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled. + // This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995. enum ggml_task_type { GGML_TASK_INIT = 0, GGML_TASK_COMPUTE, diff --git a/llama.cpp b/llama.cpp index ef80b4e8b..a869bbac8 100644 --- a/llama.cpp +++ b/llama.cpp @@ -66,6 +66,7 @@ enum e_model { MODEL_65B, }; +static const size_t kB = 1024; static const size_t MB = 1024*1024; // computed for n_ctx == 2048 @@ -129,6 +130,34 @@ static const std::map & MEM_REQ_EVAL() return k_sizes; } +// amount of VRAM needed per batch size to hold temporary results +// the values for 3b and 65b are not derived from testing but instead chosen conservatively +static const std::map & VRAM_REQ_SCRATCH_BASE() +{ + static std::map k_sizes = { + { MODEL_3B, 512ull * kB }, + { MODEL_7B, 512ull * kB }, + { MODEL_13B, 640ull * kB }, + { MODEL_30B, 768ull * kB }, + { MODEL_65B, 1536ull * kB }, + }; + return k_sizes; +} + +// amount of VRAM needed per batch size and context to hold temporary results +// the values for 3b and 65b are not derived from testing but instead chosen conservatively +static const std::map & VRAM_REQ_SCRATCH_PER_CONTEXT() +{ + static std::map k_sizes = { + { MODEL_3B, 128ull }, + { MODEL_7B, 128ull }, + { MODEL_13B, 160ull }, + { MODEL_30B, 208ull }, + { MODEL_65B, 416ull }, + }; + return k_sizes; +} + // default hparams (LLaMA 7B) struct llama_hparams { uint32_t n_vocab = 32000; @@ -253,7 +282,13 @@ struct llama_model { struct llama_context { llama_context(const llama_model & model, const llama_vocab & vocab) : model(model), vocab(vocab), t_load_us(model.t_load_us), t_start_us(model.t_start_us) {} - +#ifdef GGML_USE_METAL + ~llama_context() { + if (ctx_metal) { + ggml_metal_free(ctx_metal); + } + } +#endif std::mt19937 rng; bool has_evaluated_once = false; @@ -777,7 +812,7 @@ static bool kv_cache_init( struct llama_context_params llama_context_default_params() { struct llama_context_params result = { - /*.seed =*/ -1, + /*.seed =*/ LLAMA_DEFAULT_SEED, /*.n_ctx =*/ 512, /*.n_batch =*/ 512, /*.gpu_layers =*/ 0, @@ -1112,11 +1147,14 @@ static void llama_model_load_internal( fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__); ggml_cuda_set_scratch_size(0); // disable scratch } else { - vram_scratch = n_batch * MB; + const size_t vram_scratch_base = VRAM_REQ_SCRATCH_BASE().at(model.type); + const size_t vram_scratch_per_context = VRAM_REQ_SCRATCH_PER_CONTEXT().at(model.type); + vram_scratch = n_batch * (vram_scratch_base + n_ctx * vram_scratch_per_context); ggml_cuda_set_scratch_size(vram_scratch); if (n_gpu_layers > 0) { - fprintf(stderr, "%s: allocating batch_size x 1 MB = %zd MB VRAM for the scratch buffer\n", - __func__, vram_scratch / MB); + fprintf(stderr, "%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n", + __func__, vram_scratch_base / kB, vram_scratch_per_context, + (vram_scratch + MB - 1) / MB); // round up } } #endif // GGML_USE_CUBLAS @@ -2541,7 +2579,7 @@ struct llama_context * llama_new_context_with_model( llama_context * ctx = new llama_context(*model, model->vocab); - if (params.seed < 0) { + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } @@ -2974,8 +3012,8 @@ int llama_get_kv_cache_token_count(const struct llama_context * ctx) { #define LLAMA_MAX_RNG_STATE (64*1024) -void llama_set_rng_seed(struct llama_context * ctx, int seed) { - if (seed < 0) { +void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed) { + if (seed == LLAMA_DEFAULT_SEED) { seed = time(NULL); } ctx->rng.seed(seed); @@ -3219,7 +3257,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { return nread; } -bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { +static bool llama_load_session_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { llama_file file(path_session, "rb"); // sanity checks @@ -3273,6 +3311,15 @@ bool llama_load_session_file(struct llama_context * ctx, const char * path_sessi return true; } +bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + try { + return llama_load_session_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); + } catch (const std::exception & err) { + fprintf(stderr, "error loading session file: %s\n", err.what()); + return false; + } +} + bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { llama_file file(path_session, "wb"); diff --git a/llama.h b/llama.h index c2f2e5331..5bb1964bd 100644 --- a/llama.h +++ b/llama.h @@ -46,6 +46,8 @@ #define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN #define LLAMA_SESSION_VERSION 1 +#define LLAMA_DEFAULT_SEED 0xFFFFFFFF + #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) // Defined when llama.cpp is compiled with support for offloading model layers to GPU. #define LLAMA_SUPPORTS_GPU_OFFLOAD @@ -81,11 +83,11 @@ extern "C" { typedef void (*llama_progress_callback)(float progress, void *ctx); struct llama_context_params { - int seed; // RNG seed, -1 for random - int n_ctx; // text context - int n_batch; // prompt processing batch size - int n_gpu_layers; // number of layers to store in VRAM - int main_gpu; // the GPU that is used for scratch and small tensors + uint32_t seed; // RNG seed, -1 for random + int32_t n_ctx; // text context + int32_t n_batch; // prompt processing batch size + int32_t n_gpu_layers; // number of layers to store in VRAM + int32_t main_gpu; // the GPU that is used for scratch and small tensors float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs // called with a progress value between 0 and 1, pass NULL to disable llama_progress_callback progress_callback; @@ -196,7 +198,7 @@ extern "C" { LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx); // Sets the current rng seed. - LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed); + LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed); // Returns the maximum size in bytes of the state (rng, logits, embedding // and kv_cache) - will often be smaller after compacting tokens