From 471d8170bcea85db5ed87579cada83e077c79b44 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 16 May 2024 13:23:04 +0300 Subject: [PATCH] ggml : update ggml_rope_ext API to support freq. factors --- examples/finetune/finetune.cpp | 4 +- .../train-text-from-scratch.cpp | 4 +- ggml-cuda/rope.cu | 14 +- ggml.c | 146 ++++++++------ ggml.h | 49 +++-- llama.cpp | 184 +++++++++--------- 6 files changed, 225 insertions(+), 176 deletions(-) diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp index 22743b1bf..992426c1b 100644 --- a/examples/finetune/finetune.cpp +++ b/examples/finetune/finetune.cpp @@ -563,8 +563,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs( // not capturing these, to silcence warnings const int rope_mode = 0; - return ggml_rope_custom(ctx, - t, KQ_pos, n_rot, rope_mode, n_ctx, 0, + return ggml_rope_ext(ctx, + t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f ); }; 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 587418cc7..45bdfa8f5 100644 --- a/examples/train-text-from-scratch/train-text-from-scratch.cpp +++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp @@ -301,8 +301,8 @@ static struct ggml_tensor * llama_build_train_graphs( // not capturing these, to silcence warnings const int rope_mode = 0; - return ggml_rope_custom( - ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f + return ggml_rope_ext( + ctx, t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f ); }; diff --git a/ggml-cuda/rope.cu b/ggml-cuda/rope.cu index 8d11d9d14..e7968817f 100644 --- a/ggml-cuda/rope.cu +++ b/ggml-cuda/rope.cu @@ -166,7 +166,7 @@ static void rope_cuda( template static void rope_neox_cuda( const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, - float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float* freq_factors, cudaStream_t stream + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream ) { GGML_ASSERT(ncols % 2 == 0); const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); @@ -233,14 +233,14 @@ static void rope_cuda_f32( static void rope_neox_cuda_f16( const half * x, half * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, - float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float* freq_factors, cudaStream_t stream) { + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) { rope_neox_cuda(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream); } static void rope_neox_cuda_f32( const float * x, float * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, - float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float* freq_factors, cudaStream_t stream + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream ) { rope_neox_cuda(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream); @@ -288,16 +288,10 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const bool is_glm = mode & 4; if (is_neox) { - // TODO: move these asserts to ggml.c - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(src1->ne[0] == ne2); pos = (const int32_t *) src1_d; if (src2 != nullptr) { - // TODO: move these asserts to ggml.c - GGML_ASSERT(src2->type == GGML_TYPE_F32); - GGML_ASSERT(src2->ne[0] >= n_dims / 2); - freq_factors = (const float*) src2->data; + freq_factors = (const float *) src2->data; } } else { GGML_ASSERT(src2 == nullptr && "TODO: freq_factors not implemented for mode 1"); diff --git a/ggml.c b/ggml.c index fa4efccee..37b16b7a9 100644 --- a/ggml.c +++ b/ggml.c @@ -6231,6 +6231,7 @@ static struct ggml_tensor * ggml_rope_impl( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, + struct ggml_tensor * c, int n_dims, int mode, int n_ctx, @@ -6248,6 +6249,11 @@ static struct ggml_tensor * ggml_rope_impl( GGML_ASSERT(b->type == GGML_TYPE_I32); GGML_ASSERT(a->ne[2] == b->ne[0]); + if (c) { + GGML_ASSERT(c->type == GGML_TYPE_F32); + GGML_ASSERT(c->ne[0] >= n_dims / 2); + } + bool is_node = false; if (a->grad) { @@ -6271,17 +6277,11 @@ static struct ggml_tensor * ggml_rope_impl( result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; + result->src[2] = c; return result; } -struct ggml_tensor * ggml_rope_with_freq_factors( - struct ggml_tensor* rope_tensor, - struct ggml_tensor* freq_factors) { - rope_tensor->src[2] = freq_factors; - return rope_tensor; -} - struct ggml_tensor * ggml_rope( struct ggml_context * ctx, struct ggml_tensor * a, @@ -6290,7 +6290,7 @@ struct ggml_tensor * ggml_rope( int mode, int n_ctx) { return ggml_rope_impl( - ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, false + ctx, a, b, NULL, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, false ); } @@ -6302,7 +6302,49 @@ struct ggml_tensor * ggml_rope_inplace( int mode, int n_ctx) { return ggml_rope_impl( - ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, true + ctx, a, b, NULL, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, true + ); +} + +struct ggml_tensor * ggml_rope_ext( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int n_dims, + int mode, + int n_ctx, + int n_orig_ctx, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow) { + return ggml_rope_impl( + ctx, a, b, c, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, false + ); +} + +struct ggml_tensor * ggml_rope_ext_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int n_dims, + int mode, + int n_ctx, + int n_orig_ctx, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow) { + return ggml_rope_impl( + ctx, a, b, c, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, true ); } @@ -6321,7 +6363,7 @@ struct ggml_tensor * ggml_rope_custom( float beta_fast, float beta_slow) { return ggml_rope_impl( - ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, + ctx, a, b, NULL, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, false ); } @@ -6341,27 +6383,18 @@ struct ggml_tensor * ggml_rope_custom_inplace( float beta_fast, float beta_slow) { return ggml_rope_impl( - ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, + ctx, a, b, NULL, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, true ); } -struct ggml_tensor * ggml_rope_xpos_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int n_dims, - float base, - bool down) { - return ggml_rope_impl(ctx, a, b, n_dims, 0, 0, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, base, down, true); -} - // ggml_rope_back struct ggml_tensor * ggml_rope_back( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, + struct ggml_tensor * c, int n_dims, int mode, int n_ctx, @@ -6377,6 +6410,7 @@ 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"); @@ -18407,6 +18441,7 @@ static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct gg static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set zero_table) { struct ggml_tensor * src0 = tensor->src[0]; struct ggml_tensor * src1 = tensor->src[1]; + struct ggml_tensor * src2 = tensor->src[2]; switch (tensor->op) { case GGML_OP_DUP: @@ -18935,23 +18970,22 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src0->grad = ggml_add_or_set(ctx, src0->grad, - ggml_rope_with_freq_factors( - ggml_rope_back(ctx, - tensor->grad, - src1, - n_dims, - mode, - n_ctx, - n_orig_ctx, - freq_base, - freq_scale, - ext_factor, - attn_factor, - beta_fast, - beta_slow, - xpos_base, - xpos_down), - tensor->src[2]), + ggml_rope_back(ctx, + tensor->grad, + src1, + src2, + n_dims, + mode, + n_ctx, + n_orig_ctx, + freq_base, + freq_scale, + ext_factor, + attn_factor, + beta_fast, + beta_slow, + xpos_base, + xpos_down), zero_table); } } break; @@ -18976,24 +19010,23 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src0->grad = ggml_add_or_set(ctx, src0->grad, - ggml_rope_with_freq_factors( - ggml_rope_impl(ctx, - tensor->grad, - src1, - n_dims, - mode, - n_ctx, - n_orig_ctx, - freq_base, - freq_scale, - ext_factor, - attn_factor, - beta_fast, - beta_slow, - xpos_base, - xpos_down, - false), - tensor->src[2]), + ggml_rope_impl(ctx, + tensor->grad, + src1, + src2, + n_dims, + mode, + n_ctx, + n_orig_ctx, + freq_base, + freq_scale, + ext_factor, + attn_factor, + beta_fast, + beta_slow, + xpos_base, + xpos_down, + false), zero_table); } } break; @@ -19062,7 +19095,6 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor masked); } - struct ggml_tensor * src2 = tensor->src[2]; const int64_t elem_q = ggml_nelements(src0); const int64_t elem_k = ggml_nelements(src1); const int64_t elem_v = ggml_nelements(src2); diff --git a/ggml.h b/ggml.h index 774757101..35ac9110c 100644 --- a/ggml.h +++ b/ggml.h @@ -1465,6 +1465,7 @@ extern "C" { // if mode & 4 == 1, ChatGLM style // // b is an int32 vector with size a->ne[2], it contains the positions + // c is freq factors (e.g. phi3-128k), (optional) GGML_API struct ggml_tensor * ggml_rope( struct ggml_context * ctx, struct ggml_tensor * a, @@ -1483,10 +1484,11 @@ extern "C" { int n_ctx); // custom RoPE - GGML_API struct ggml_tensor * ggml_rope_custom( + GGML_API struct ggml_tensor * ggml_rope_ext( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, + struct ggml_tensor * c, int n_dims, int mode, int n_ctx, @@ -1499,7 +1501,23 @@ extern "C" { float beta_slow); // in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_rope_custom_inplace( + GGML_API struct ggml_tensor * ggml_rope_ext_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int n_dims, + int mode, + int n_ctx, + int n_orig_ctx, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow); + + GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -1512,20 +1530,28 @@ extern "C" { float ext_factor, float attn_factor, float beta_fast, - float beta_slow); + float beta_slow), + "use ggml_rope_ext instead"); - // compute correction dims for YaRN RoPE scaling - GGML_CALL void ggml_rope_yarn_corr_dims( - int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]); - - // xPos RoPE, in-place, returns view(a) - GGML_API struct ggml_tensor * ggml_rope_xpos_inplace( + GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom_inplace( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, int n_dims, - float base, - bool down); + int mode, + int n_ctx, + int n_orig_ctx, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow), + "use ggml_rope_ext_inplace instead"); + + // compute correction dims for YaRN RoPE scaling + GGML_CALL void ggml_rope_yarn_corr_dims( + int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]); // rotary position embedding backward, i.e compute dx from dy // a - dy @@ -1533,6 +1559,7 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, + struct ggml_tensor * c, int n_dims, int mode, int n_ctx, diff --git a/llama.cpp b/llama.cpp index a30593087..a08941af1 100644 --- a/llama.cpp +++ b/llama.cpp @@ -6883,17 +6883,15 @@ struct llm_build_context { for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * tmp = // we rotate only the first n_rot dimensions - ggml_rope_custom_inplace(ctx0, + ggml_rope_ext_inplace(ctx0, ggml_view_3d(ctx0, kv_self.k_l[il], n_embd_head_k, n_head_kv, n_ctx, ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k), ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), 0), - lctx.inp_K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + lctx.inp_K_shift, lctx.freq_factors, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); - tmp = ggml_rope_with_freq_factors(tmp, lctx.freq_factors); - cb(tmp, "K_shifted", il); ggml_build_forward_expand(gf, tmp); } @@ -7117,15 +7115,15 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -7247,13 +7245,13 @@ struct llm_build_context { switch (model.type) { case MODEL_7B: - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -7359,15 +7357,15 @@ struct llm_build_context { struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); cb(Vcur, "Vcur", il); - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -7480,14 +7478,14 @@ struct llm_build_context { Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); // using mode = 2 for neox mode - Qcur = ggml_rope_custom( - ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); @@ -7603,15 +7601,15 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -7755,15 +7753,15 @@ struct llm_build_context { cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -8108,15 +8106,15 @@ struct llm_build_context { cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -8548,15 +8546,15 @@ struct llm_build_context { } - Qcur = ggml_rope_custom( - ctx0, Qcur, inp_pos, + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, Kcur, inp_pos, + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -8668,14 +8666,14 @@ struct llm_build_context { Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); // using mode = 2 for neox mode - Qcur = ggml_rope_custom( - ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); @@ -8779,15 +8777,15 @@ struct llm_build_context { Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); cb(Vcur, "Vcur", il); - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -8893,15 +8891,15 @@ struct llm_build_context { Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); cb(Vcur, "Vcur", il); - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -9045,8 +9043,8 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Qcur = ggml_rope_custom( - ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); @@ -9056,8 +9054,8 @@ struct llm_build_context { Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head))); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); @@ -9167,21 +9165,19 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Qcur = ggml_rope_custom( - ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, freq_factors, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); - Qcur = ggml_rope_with_freq_factors(Qcur, freq_factors); cb(Qcur, "Qcur", il); Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx, + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, freq_factors, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); - Kcur = ggml_rope_with_freq_factors(Kcur, freq_factors); cb(Kcur, "Kcur", il); cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, @@ -9285,14 +9281,14 @@ struct llm_build_context { struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); cb(Vcur, "Vcur", il); - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens), inp_pos, nullptr, n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, nullptr, n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); @@ -9493,15 +9489,15 @@ struct llm_build_context { cb(tmpk, "tmpk", il); cb(Vcur, "Vcur", il); - struct ggml_tensor * Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos, + struct ggml_tensor * Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - struct ggml_tensor * Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, + struct ggml_tensor * Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -9609,15 +9605,15 @@ struct llm_build_context { // cb(Vcur, "Vcur", il); // } - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -9726,15 +9722,15 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -9856,15 +9852,15 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -9976,8 +9972,8 @@ struct llm_build_context { struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); cb(Vcur, "Vcur", il); - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr, n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cb(Qcur, "Qcur", il); @@ -9985,8 +9981,8 @@ struct llm_build_context { Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); cb(Qcur, "Qcur_scaled", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr, n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); @@ -10096,15 +10092,15 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -10386,15 +10382,15 @@ struct llm_build_context { cb(Kcur, "Kcur", il); } - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); @@ -10517,15 +10513,15 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } - Qcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow ); cb(Qcur, "Qcur", il); - Kcur = ggml_rope_custom( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow );