From d0a7145ba99ed3a8bc3145aa785b5c86ffe65020 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 1 Jul 2024 02:09:34 +0300 Subject: [PATCH 01/27] flake.lock: Update (#8218) --- flake.lock | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/flake.lock b/flake.lock index 79bb3f63fdc6d0..973ff4e5675c77 100644 --- a/flake.lock +++ b/flake.lock @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1718895438, - "narHash": "sha256-k3JqJrkdoYwE3fHE6xGDY676AYmyh4U2Zw+0Bwe5DLU=", + "lastModified": 1719506693, + "narHash": "sha256-C8e9S7RzshSdHB7L+v9I51af1gDM5unhJ2xO1ywxNH8=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "d603719ec6e294f034936c0d0dc06f689d91b6c3", + "rev": "b2852eb9365c6de48ffb0dc2c9562591f652242a", "type": "github" }, "original": { From 197fe6c1d7bec6718ce901f0141b2725240f298c Mon Sep 17 00:00:00 2001 From: zhentaoyu Date: Mon, 1 Jul 2024 19:39:06 +0800 Subject: [PATCH 02/27] [SYCL] Update SYCL-Rope op and Refactor (#8157) * align with rope.cu and move sycl-op to a single file --- ggml/src/ggml-sycl.cpp | 305 +-------------------------------- ggml/src/ggml-sycl/backend.hpp | 1 + ggml/src/ggml-sycl/rope.cpp | 275 +++++++++++++++++++++++++++++ ggml/src/ggml-sycl/rope.hpp | 22 +++ 4 files changed, 300 insertions(+), 303 deletions(-) create mode 100644 ggml/src/ggml-sycl/rope.cpp create mode 100644 ggml/src/ggml-sycl/rope.hpp diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index 4a668a2c34d3ea..30d8a5b33b6133 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -978,114 +978,6 @@ static void cpy_f32_q(const char * cx, char * cdst, const int ne, cpy_blck(cx + x_offset, cdst + dst_offset); } -static float rope_yarn_ramp(const float low, const float high, const int i0) { - const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low); - return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y)); -} - -struct rope_corr_dims { - float v[4]; -}; - -// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn -// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. -static void rope_yarn( - float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, - float * cos_theta, float * sin_theta -) { - // Get n-d rotational scaling corrected for extrapolation - float theta_interp = freq_scale * theta_extrap; - float theta = theta_interp; - if (ext_factor != 0.0f) { - float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; - theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; - - // Get n-d magnitude scaling corrected for interpolation - mscale *= 1.0f + 0.1f * sycl::log(1.0f / freq_scale); - } - *cos_theta = sycl::cos(theta) * mscale; - *sin_theta = sycl::sin(theta) * mscale; -} - -// rope == RoPE == rotary positional embedding -template -static void rope( - const T * x, T * dst, int ncols, 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 sycl::nd_item<3> &item_ct1) { - const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1)); - - if (col >= ncols) { - return; - } - - const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - const int i = row*ncols + col; - const int i2 = row/p_delta_rows; - - const int p = has_pos ? pos[i2] : 0; - const float theta_base = p * dpct::pow(freq_base, -float(col) / ncols); - - float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + 1]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + 1] = x0*sin_theta + x1*cos_theta; -} - -template -static void rope_neox( - const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, - float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, - const float * freq_factors, const sycl::nd_item<3> &item_ct1) { - const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1)); - - if (col >= ncols) { - return; - } - - const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - const int ib = col / n_dims; - const int ic = col % n_dims; - - if (ib > 0) { - const int i = row*ncols + ib*n_dims + ic; - - dst[i + 0] = x[i + 0]; - dst[i + 1] = x[i + 1]; - - return; - } - - const int i = row*ncols + ib*n_dims + ic/2; - const int i2 = row/p_delta_rows; - - float cur_rot = inv_ndims * ic - ib; - - const int p = has_pos ? pos[i2] : 0; - const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f; - - const float theta_base = - p * freq_scale * dpct::pow(theta_scale, col / 2.0f)/freq_factor; - - float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + n_dims/2]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; -} - static void k_sum_rows_f32(const float * x, float * dst, const int ncols, const sycl::nd_item<3> &item_ct1) { const int row = item_ct1.get_group(1); @@ -2241,110 +2133,6 @@ static void clamp_f32_sycl(const float *x, float *dst, const float min, }); } -template -static void rope_sycl(const T *x, T *dst, int ncols, 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, queue_ptr stream) { - GGML_ASSERT(ncols % 2 == 0); - const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); - const int num_blocks_x = (ncols + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); - const sycl::range<3> block_nums(1, num_blocks_x, nrows); - if (pos == nullptr) { - /* - DPCT1049:40: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope(x, dst, ncols, pos, freq_scale, p_delta_rows, - freq_base, ext_factor, attn_factor, corr_dims, - item_ct1); - }); - } else { - /* - DPCT1049:41: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope(x, dst, ncols, pos, freq_scale, p_delta_rows, - freq_base, ext_factor, attn_factor, corr_dims, - item_ct1); - }); - } -} - -template -static void rope_neox_sycl(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, queue_ptr stream) { - GGML_ASSERT(ncols % 2 == 0); - const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); - const int num_blocks_x = (ncols + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); - const sycl::range<3> block_nums(1, num_blocks_x, nrows); - - const float theta_scale = powf(freq_base, -2.0f/n_dims); - const float inv_ndims = -1.0f / n_dims; - - if (pos == nullptr) { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - if (freq_factors == nullptr) { - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, freq_factors, - item_ct1); - }); - } else { - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, freq_factors, - item_ct1); - }); - } - } else { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - if (freq_factors == nullptr) { - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, freq_factors, item_ct1); - }); - } else { - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, freq_factors, item_ct1); - }); - } - } -} - static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols, const int nrows, queue_ptr stream) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); @@ -3461,97 +3249,6 @@ catch (sycl::exception const &exc) { std::exit(1); } -inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - const ggml_tensor * src2 = dst->src[2]; - - GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); - GGML_ASSERT(src0->type == dst->type); - - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - const int64_t ne2 = dst->ne[2]; - const int64_t nrows = ggml_nrows(src0); - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - //const int n_ctx = ((int32_t *) dst->op_params)[3]; - const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; - - // RoPE alteration for extended context - float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; - memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); - memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); - memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); - memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); - - const float * freq_factors = nullptr; - const int32_t * pos = nullptr; - if ((mode & 1) == 0) { - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(src1->ne[0] == ne2); - pos = (const int32_t *) src1_dd; - } - - const bool is_neox = mode & 2; - -#pragma message("TODO: update rope NORM mode to match NEOX mode") -#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7634") - - if (is_neox) { - pos = (const int32_t *) src1_dd; - - if (src2 != nullptr) { - freq_factors = (const float *) src2->data; - } - } else { - GGML_ASSERT(src2 == nullptr && "TODO: freq_factors not implemented for !is_neox"); - } - - rope_corr_dims corr_dims; - ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v); - - // compute - if (is_neox) { - if (src0->type == GGML_TYPE_F32) { - rope_neox_sycl( - (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, freq_factors, main_stream - ); - } else if (src0->type == GGML_TYPE_F16) { - rope_neox_sycl((const sycl::half *)src0_dd, (sycl::half *)dst_dd, - ne00, n_dims, nrows, pos, freq_scale, ne01, - freq_base, ext_factor, attn_factor, corr_dims, - freq_factors, main_stream); - } else { - GGML_ASSERT(false); - } - } else { - if (src0->type == GGML_TYPE_F32) { - rope_sycl( - (const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream - ); - } else if (src0->type == GGML_TYPE_F16) { - rope_sycl((const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, - nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream); - } else { - GGML_ASSERT(false); - } - } - - (void) src1; - (void) dst; - (void) src1_dd; -} - static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const float *src0_dd, const float *src1_dd, @@ -6241,7 +5938,9 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons case GGML_OP_CONT: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + return true; case GGML_OP_ROPE: + return ggml_is_contiguous(op->src[0]); case GGML_OP_IM2COL: case GGML_OP_POOL_2D: case GGML_OP_SUM_ROWS: diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index 2d37e271f90508..d5a63cd710cc34 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -19,5 +19,6 @@ #include "dmmv.hpp" #include "mmq.hpp" #include "mmvq.hpp" +#include "rope.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp new file mode 100644 index 00000000000000..eabf1693e2d2f4 --- /dev/null +++ b/ggml/src/ggml-sycl/rope.cpp @@ -0,0 +1,275 @@ +#include "rope.hpp" + +struct rope_corr_dims { + float v[2]; +}; + +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low); + return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * sycl::log(1.0f / freq_scale); + } + *cos_theta = sycl::cos(theta) * mscale; + *sin_theta = sycl::sin(theta) * mscale; +} + +template +static void rope_norm( + const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors, + const sycl::nd_item<3> &item_ct1) { + const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); + + if (i0 >= ne0) { + return; + } + + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i0 >= n_dims) { + const int i = row*ne0 + i0; + + dst[i + 0] = x[i + 0]; + dst[i + 1] = x[i + 1]; + + return; + } + + const int i = row*ne0 + i0; + const int i2 = row/p_delta_rows; + + const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f); + + const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f; + + float cos_theta; + float sin_theta; + + rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + 1]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + 1] = x0*sin_theta + x1*cos_theta; +} + +template +static void rope_neox( + const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors, + const sycl::nd_item<3> &item_ct1) { + const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); + + if (i0 >= ne0) { + return; + } + + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i0 >= n_dims) { + const int i = row*ne0 + i0; + + dst[i + 0] = x[i + 0]; + dst[i + 1] = x[i + 1]; + + return; + } + + const int i = row*ne0 + i0/2; + const int i2 = row/p_delta_rows; + + const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f); + + const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f; + + float cos_theta; + float sin_theta; + + rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + n_dims/2]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; +} + +template +static void rope_norm_sycl( + const T *x, T *dst, int ne0, int n_dims, int nr, 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, queue_ptr stream) { + GGML_ASSERT(ne0 % 2 == 0); + const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ne0 + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); + const sycl::range<3> block_nums(1, num_blocks_x, nr); + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + if (freq_factors == nullptr) { + /* + DPCT1049:40: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_norm(x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows, + ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, + item_ct1); + }); + } else { + /* + DPCT1049:41: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_norm(x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows, + ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, + item_ct1); + }); + } +} + +template +static void rope_neox_sycl( + const T *x, T *dst, int ne0, int n_dims, int nr, 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, queue_ptr stream) { + GGML_ASSERT(ne0 % 2 == 0); + const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ne0 + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); + const sycl::range<3> block_nums(1, num_blocks_x, nr); + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + if (freq_factors == nullptr) { + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ne0, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, freq_factors, + item_ct1); + }); + } else { + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ne0, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, freq_factors, + item_ct1); + }); + } +} + +void ggml_sycl_op_rope( + ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream) { + const ggml_tensor * src2 = dst->src[2]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + GGML_ASSERT(src0->type == dst->type); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t nr = ggml_nrows(src0); + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + //const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; + + // RoPE alteration for extended context + float freq_base; + float freq_scale; + float ext_factor; + float attn_factor; + float beta_fast; + float beta_slow; + + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + + const bool is_neox = mode & 2; + + const int32_t * pos = (const int32_t *) src1_dd; + + const float * freq_factors = nullptr; + if (src2 != nullptr) { + freq_factors = (const float *) src2->data; + } + + rope_corr_dims corr_dims; + ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v); + + // compute + if (is_neox) { + if (src0->type == GGML_TYPE_F32) { + rope_neox_sycl( + (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, freq_factors, main_stream + ); + } else if (src0->type == GGML_TYPE_F16) { + rope_neox_sycl( + (const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, freq_factors, main_stream + ); + } else { + GGML_ASSERT(false); + } + } else { + if (src0->type == GGML_TYPE_F32) { + rope_norm_sycl( + (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, freq_factors, main_stream + ); + } else if (src0->type == GGML_TYPE_F16) { + rope_norm_sycl( + (const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, freq_factors, main_stream + ); + } else { + GGML_ASSERT(false); + } + } + + (void) src1; + (void) dst; + (void) src1_dd; +} diff --git a/ggml/src/ggml-sycl/rope.hpp b/ggml/src/ggml-sycl/rope.hpp new file mode 100644 index 00000000000000..00354c3131bd77 --- /dev/null +++ b/ggml/src/ggml-sycl/rope.hpp @@ -0,0 +1,22 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_ROPE_HPP +#define GGML_SYCL_ROPE_HPP + +#include "common.hpp" + +void ggml_sycl_op_rope( + ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream); + +#endif // GGML_SYCL_ROPE_HPP From 694c59cb42d1ebd6a7d912ca65d3d7363e0f14c9 Mon Sep 17 00:00:00 2001 From: iacore <74560659+iacore@users.noreply.github.com> Date: Mon, 1 Jul 2024 11:40:58 +0000 Subject: [PATCH 03/27] Document BERT support. (#8205) * Update README.md document BERT support * Update README.md --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 99b16f6e29e329..153d837e374eaf 100644 --- a/README.md +++ b/README.md @@ -108,6 +108,7 @@ Typically finetunes of the base models below are supported as well. - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon) - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2) - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne) +- [X] [BERT](https://github.com/ggerganov/llama.cpp/pull/5423) - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/) - [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft) - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila) From 257f8e41e24b5bbfc27d9e907189a3e0cdb650d4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 1 Jul 2024 14:46:18 +0300 Subject: [PATCH 04/27] nix : remove OpenCL remnants (#8235) * nix : remove OpenCL remnants * minor : remove parentheses --- .devops/nix/package.nix | 11 ++--------- 1 file changed, 2 insertions(+), 9 deletions(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 4ee0d62cb49d9a..b75d7ff9e5bab5 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -17,18 +17,15 @@ rocmPackages, vulkan-headers, vulkan-loader, - clblast, useBlas ? builtins.all (x: !x) [ useCuda useMetalKit - useOpenCL useRocm useVulkan ] && blas.meta.available, useCuda ? config.cudaSupport, - useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL, + useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin, useMpi ? false, # Increases the runtime closure size by ~700M - useOpenCL ? false, useRocm ? config.rocmSupport, useVulkan ? false, llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake @@ -56,7 +53,6 @@ let ++ lib.optionals useCuda [ "CUDA" ] ++ lib.optionals useMetalKit [ "MetalKit" ] ++ lib.optionals useMpi [ "MPI" ] - ++ lib.optionals useOpenCL [ "OpenCL" ] ++ lib.optionals useRocm [ "ROCm" ] ++ lib.optionals useVulkan [ "Vulkan" ]; @@ -198,7 +194,6 @@ effectiveStdenv.mkDerivation ( optionals effectiveStdenv.isDarwin darwinBuildInputs ++ optionals useCuda cudaBuildInputs ++ optionals useMpi [ mpi ] - ++ optionals useOpenCL [ clblast ] ++ optionals useRocm rocmBuildInputs ++ optionals useBlas [ blas ] ++ optionals useVulkan vulkanBuildInputs; @@ -210,7 +205,6 @@ effectiveStdenv.mkDerivation ( (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true) (cmakeBool "GGML_NATIVE" false) (cmakeBool "GGML_BLAS" useBlas) - (cmakeBool "GGML_CLBLAST" useOpenCL) (cmakeBool "GGML_CUDA" useCuda) (cmakeBool "GGML_HIPBLAS" useRocm) (cmakeBool "GGML_METAL" useMetalKit) @@ -254,7 +248,6 @@ effectiveStdenv.mkDerivation ( useCuda useMetalKit useMpi - useOpenCL useRocm useVulkan ; @@ -281,7 +274,7 @@ effectiveStdenv.mkDerivation ( # Configurations we don't want even the CI to evaluate. Results in the # "unsupported platform" messages. This is mostly a no-op, because # cudaPackages would've refused to evaluate anyway. - badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin; + badPlatforms = optionals useCuda lib.platforms.darwin; # Configurations that are known to result in build failures. Can be # overridden by importing Nixpkgs with `allowBroken = true`. From 3840b6f593751a0ba636bfda73b630cd6c29d7b5 Mon Sep 17 00:00:00 2001 From: Michael Francis Date: Mon, 1 Jul 2024 07:47:04 -0400 Subject: [PATCH 05/27] nix : enable curl (#8043) Co-authored-by: Georgi Gerganov --- .devops/nix/package.nix | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index b75d7ff9e5bab5..49e9b75287b338 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -17,6 +17,7 @@ rocmPackages, vulkan-headers, vulkan-loader, + curl, useBlas ? builtins.all (x: !x) [ useCuda useMetalKit @@ -27,6 +28,7 @@ useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin, useMpi ? false, # Increases the runtime closure size by ~700M useRocm ? config.rocmSupport, + enableCurl ? true, useVulkan ? false, llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake @@ -196,13 +198,15 @@ effectiveStdenv.mkDerivation ( ++ optionals useMpi [ mpi ] ++ optionals useRocm rocmBuildInputs ++ optionals useBlas [ blas ] - ++ optionals useVulkan vulkanBuildInputs; + ++ optionals useVulkan vulkanBuildInputs + ++ optionals enableCurl [ curl ]; cmakeFlags = [ (cmakeBool "LLAMA_BUILD_SERVER" true) (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic)) (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true) + (cmakeBool "LLAMA_CURL" enableCurl) (cmakeBool "GGML_NATIVE" false) (cmakeBool "GGML_BLAS" useBlas) (cmakeBool "GGML_CUDA" useCuda) From 0ddeff10230b88f1fa9866bbe5fe0d71ba2323a0 Mon Sep 17 00:00:00 2001 From: Roni Date: Mon, 1 Jul 2024 14:48:16 +0200 Subject: [PATCH 06/27] readme : update tool list (#8209) * Added gppm to Tool list in README * Update README.md --------- Co-authored-by: Georgi Gerganov --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 153d837e374eaf..c136d4a5cb9c95 100644 --- a/README.md +++ b/README.md @@ -218,6 +218,7 @@ Unless otherwise noted these projects are open-source with permissive licensing: **Tools:** - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML +[crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption --- From 49122a873f54615626d1b49a2a39013ed4be98d5 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Mon, 1 Jul 2024 18:48:34 +0200 Subject: [PATCH 07/27] gemma2: add sliding window mask (#8227) * gemma2: add sliding window mask * fix data_swa uninitialized * better naming * add co-author Co-authored-by: Arlo Phoenix * replace list with single tensor * update * llama : minor styling * convert : add sanity check for query_pre_attn_scalar * fix small typo in README --------- Co-authored-by: Arlo Phoenix Co-authored-by: Georgi Gerganov --- README.md | 2 +- convert-hf-to-gguf.py | 6 +++ gguf-py/gguf/constants.py | 1 + gguf-py/gguf/gguf_writer.py | 3 ++ src/llama.cpp | 99 +++++++++++++++++++++++++------------ 5 files changed, 79 insertions(+), 32 deletions(-) diff --git a/README.md b/README.md index c136d4a5cb9c95..daba70717312e2 100644 --- a/README.md +++ b/README.md @@ -218,7 +218,7 @@ Unless otherwise noted these projects are open-source with permissive licensing: **Tools:** - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML -[crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption +- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption --- diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 3ef2f69e7c0dfa..4a7f500ff7d5ca 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -2369,6 +2369,12 @@ def set_gguf_parameters(self): self.gguf_writer.add_final_logit_softcapping( self.hparams["final_logit_softcapping"] ) + self.gguf_writer.add_sliding_window(self.hparams["sliding_window"]) + + # sanity check + attn_scalar = self.hparams["query_pre_attn_scalar"] + if attn_scalar != hparams["hidden_size"] / hparams["num_attention_heads"]: + raise ValueError("query_pre_attn_scalar must be equal to n_embd / n_head") def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unusem diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 9bfa891d5dc529..e87c58266158a1 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -66,6 +66,7 @@ class Attention: Q_LORA_RANK = "{arch}.attention.q_lora_rank" KV_LORA_RANK = "{arch}.attention.kv_lora_rank" REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" + SLIDING_WINDOW = "{arch}.attention.sliding_window" class Rope: DIMENSION_COUNT = "{arch}.rope.dimension_count" diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 1aeb0d9b08685a..75a8b2636a6a21 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -552,6 +552,9 @@ def add_kv_lora_rank(self, length: int) -> None: def add_relative_attn_buckets_count(self, value: int) -> None: self.add_uint32(Keys.Attention.REL_BUCKETS_COUNT.format(arch=self.arch), value) + def add_sliding_window(self, value: int) -> None: + self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value) + def add_pooling_type(self, value: PoolingType) -> None: self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value) diff --git a/src/llama.cpp b/src/llama.cpp index 2a4d73856fcd93..eea532f6ac2ff3 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -317,6 +317,7 @@ enum llm_kv { LLM_KV_ATTENTION_Q_LORA_RANK, LLM_KV_ATTENTION_KV_LORA_RANK, LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, + LLM_KV_ATTENTION_SLIDING_WINDOW, LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_FREQ_BASE, @@ -409,6 +410,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, + { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, @@ -2085,6 +2087,7 @@ struct llama_hparams { uint32_t n_head_kv; uint32_t n_layer; uint32_t n_rot; + uint32_t n_swa = 0; // sliding window attention (SWA) uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head uint32_t n_ff; @@ -2139,6 +2142,7 @@ struct llama_hparams { if (this->n_head_kv != other.n_head_kv) return true; if (this->n_layer != other.n_layer) return true; if (this->n_rot != other.n_rot) return true; + if (this->n_swa != other.n_swa) return true; if (this->n_embd_head_k != other.n_embd_head_k) return true; if (this->n_embd_head_v != other.n_embd_head_v) return true; if (this->n_ff != other.n_ff) return true; @@ -2649,17 +2653,18 @@ struct llama_context { void * abort_callback_data = nullptr; // input tensors - struct ggml_tensor * inp_tokens; // I32 [n_batch] - struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] - struct ggml_tensor * inp_pos; // I32 [n_batch] - struct ggml_tensor * inp_out_ids; // I32 [n_outputs] - struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] - struct ggml_tensor * inp_K_shift; // I32 [kv_size] - struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] - struct ggml_tensor * inp_cls; // I32 [n_batch] - struct ggml_tensor * inp_s_copy; // I32 [kv_size] - struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] - struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] + struct ggml_tensor * inp_tokens; // I32 [n_batch] + struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] + struct ggml_tensor * inp_pos; // I32 [n_batch] + struct ggml_tensor * inp_out_ids; // I32 [n_outputs] + struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] + struct ggml_tensor * inp_KQ_mask_swa; // F32 [kv_size, n_batch] + struct ggml_tensor * inp_K_shift; // I32 [kv_size] + struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] + struct ggml_tensor * inp_cls; // I32 [n_batch] + struct ggml_tensor * inp_s_copy; // I32 [kv_size] + struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] + struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] // control vectors struct llama_control_vector cvec; @@ -4709,6 +4714,8 @@ static void llm_load_hparams( } break; case LLM_ARCH_GEMMA2: { + hparams.n_swa = 4096; // default value of gemma 2 + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false); ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false); @@ -5419,6 +5426,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv); LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); + LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa); LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k); LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v); LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa()); @@ -7775,17 +7783,18 @@ struct llm_build_context { ctx0 = ggml_init(params); - lctx.inp_tokens = nullptr; - lctx.inp_embd = nullptr; - lctx.inp_pos = nullptr; - lctx.inp_out_ids = nullptr; - lctx.inp_KQ_mask = nullptr; - lctx.inp_K_shift = nullptr; - lctx.inp_mean = nullptr; - lctx.inp_cls = nullptr; - lctx.inp_s_copy = nullptr; - lctx.inp_s_mask = nullptr; - lctx.inp_s_seq = nullptr; + lctx.inp_tokens = nullptr; + lctx.inp_embd = nullptr; + lctx.inp_pos = nullptr; + lctx.inp_out_ids = nullptr; + lctx.inp_KQ_mask = nullptr; + lctx.inp_KQ_mask_swa = nullptr; + lctx.inp_K_shift = nullptr; + lctx.inp_mean = nullptr; + lctx.inp_cls = nullptr; + lctx.inp_s_copy = nullptr; + lctx.inp_s_mask = nullptr; + lctx.inp_s_seq = nullptr; } void free() { @@ -7804,7 +7813,6 @@ struct llm_build_context { cb(lctx.inp_K_shift, "K_shift", -1); ggml_set_input(lctx.inp_K_shift); - for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * rope_factors = build_rope_factors(il); struct ggml_tensor * tmp = @@ -7939,16 +7947,27 @@ struct llm_build_context { } struct ggml_tensor * build_inp_KQ_mask(bool causal = true) { - if (causal) { - lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - } else { - lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - } + lctx.inp_KQ_mask = causal + ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)) + : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); cb(lctx.inp_KQ_mask, "KQ_mask", -1); ggml_set_input(lctx.inp_KQ_mask); + return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask; } + struct ggml_tensor * build_inp_KQ_mask_swa(bool causal = true) { + GGML_ASSERT(hparams.n_swa > 0); + + lctx.inp_KQ_mask_swa = causal + ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)) + : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + cb(lctx.inp_KQ_mask_swa, "KQ_mask_swa", -1); + ggml_set_input(lctx.inp_KQ_mask_swa); + + return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask_swa, GGML_TYPE_F16) : lctx.inp_KQ_mask_swa; + } + struct ggml_tensor * build_inp_mean() { lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); cb(lctx.inp_mean, "inp_mean", -1); @@ -11029,9 +11048,14 @@ struct llm_build_context { struct ggml_tensor * inp_pos = build_inp_pos(); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + // gemma 2 requires different mask for layers using sliding window (SWA) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(true); + struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true); for (int il = 0; il < n_layer; ++il) { + // (il % 2) layers use SWA + struct ggml_tensor * KQ_mask_l = (il % 2 == 0) ? KQ_mask_swa : KQ_mask; + // norm cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, @@ -11067,7 +11091,7 @@ struct llm_build_context { cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, 1.0f, cb, il); } cur = llm_build_norm(ctx0, cur, hparams, @@ -12670,7 +12694,12 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); - float * data = (float *) lctx.inp_KQ_mask->data; + float * data = (float *) lctx.inp_KQ_mask->data; + float * data_swa = nullptr; + + if (lctx.inp_KQ_mask_swa) { + data_swa = (float *) lctx.inp_KQ_mask_swa->data; + } // For causal attention, use only the previous KV cells // of the correct sequence for each token of the batch. @@ -12692,6 +12721,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } data[h*(n_kv*n_tokens) + j*n_kv + i] = f; + + // may need to cut off old tokens for sliding window + if (data_swa) { + if (pos - lctx.kv_self.cells[i].pos >= (int32_t)hparams.n_swa) { + f = -INFINITY; + } + data_swa[h*(n_kv*n_tokens) + j*n_kv + i] = f; + } } } From dae57a1ebc1c9bd5693ab999e19d77c5506ae559 Mon Sep 17 00:00:00 2001 From: Mateusz Charytoniuk Date: Mon, 1 Jul 2024 19:13:22 +0200 Subject: [PATCH 08/27] readme: add Paddler to the list of projects (#8239) --- README.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/README.md b/README.md index daba70717312e2..3569b2bbb5e34c 100644 --- a/README.md +++ b/README.md @@ -220,6 +220,10 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption +**Infrastructure:** + +- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp + --- Here is a typical run using LLaMA v2 13B on M2 Ultra: From cb5fad4c6c2cbef92e9b8b63449e1cb7664e4846 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Mon, 1 Jul 2024 20:39:06 +0200 Subject: [PATCH 09/27] CUDA: refactor and optimize IQ MMVQ (#8215) * CUDA: refactor and optimize IQ MMVQ * uint -> uint32_t * __dp4a -> ggml_cuda_dp4a * remove MIN_CC_DP4A checks * change default * try CI fix --- ggml/src/ggml-common.h | 14 +- ggml/src/ggml-cuda.cu | 12 +- ggml/src/ggml-cuda/common.cuh | 76 ++- ggml/src/ggml-cuda/fattn-common.cuh | 50 +- ggml/src/ggml-cuda/mmvq.cu | 26 +- ggml/src/ggml-cuda/vecdotq.cuh | 688 +++++++++++++--------------- ggml/src/ggml-sycl/mmvq.cpp | 12 +- ggml/src/ggml-sycl/vecdotq.hpp | 21 - 8 files changed, 409 insertions(+), 490 deletions(-) diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index e8efceb760d409..c74060cc4b9919 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -106,19 +106,19 @@ typedef sycl::half2 ggml_half2; #define QR6_K 2 #define QI2_XXS (QK_K / (4*QR2_XXS)) -#define QR2_XXS 8 +#define QR2_XXS 4 #define QI2_XS (QK_K / (4*QR2_XS)) -#define QR2_XS 8 +#define QR2_XS 4 #define QI2_S (QK_K / (4*QR2_S)) -#define QR2_S 8 +#define QR2_S 4 #define QI3_XXS (QK_K / (4*QR3_XXS)) -#define QR3_XXS 8 +#define QR3_XXS 4 #define QI3_XS (QK_K / (4*QR3_XS)) -#define QR3_XS 8 +#define QR3_XS 4 #define QI1_S (QK_K / (4*QR1_S)) #define QR1_S 8 @@ -130,10 +130,10 @@ typedef sycl::half2 ggml_half2; #define QR4_NL 2 #define QI4_XS (QK_K / (4*QR4_XS)) -#define QR4_XS 8 +#define QR4_XS 2 #define QI3_S (QK_K / (4*QR3_S)) -#define QR3_S 8 +#define QR3_S 4 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 0acfda91d3e51b..649ef5a0819107 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -1882,6 +1882,11 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor bool use_mul_mat_q = ggml_is_quantized(src0->type) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32; + // if mmvq is available it's a better choice than dmmv: +#ifndef GGML_CUDA_FORCE_DMMV + use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q; +#endif // GGML_CUDA_FORCE_DMMV + bool any_gpus_with_slow_fp16 = false; if (split) { @@ -1894,22 +1899,15 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor } const int cc = ggml_cuda_info().devices[id].cc; - use_mul_mat_vec_q = use_mul_mat_vec_q && cc >= MIN_CC_DP4A; use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc); } } else { const int cc = ggml_cuda_info().devices[ctx.device].cc; - use_mul_mat_vec_q = use_mul_mat_vec_q && cc >= MIN_CC_DP4A; use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc); } - // if mmvq is available it's a better choice than dmmv: -#ifndef GGML_CUDA_FORCE_DMMV - use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q; -#endif // GGML_CUDA_FORCE_DMMV - // debug helpers //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]); //printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]); diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 8d00db6c193ff6..472f4ace1c2ad2 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -3,6 +3,7 @@ #include "ggml.h" #include "ggml-cuda.h" +#include #include #if defined(GGML_USE_HIPBLAS) @@ -268,30 +269,15 @@ static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigne return c; } -static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { -#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) - c = __builtin_amdgcn_sdot4(a, b, c, false); -#elif defined(RDNA3) - c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); -#elif defined(__gfx1010__) || defined(__gfx900__) - int tmp1; - int tmp2; - asm("\n \ - v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \ - v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \ - v_add3_u32 %0, %1, %2, %0 \n \ - v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \ - v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \ - v_add3_u32 %0, %1, %2, %0 \n \ - " - : "+v"(c), "=&v"(tmp1), "=&v"(tmp2) - : "v"(a), "v"(b) - ); -#else - const int8x4_t va = reinterpret_cast(a); - const int8x4_t vb = reinterpret_cast(b); - c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3]; -#endif +static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) { + const uint8x4_t& va = reinterpret_cast(a); + const uint8x4_t& vb = reinterpret_cast(b); + unsigned int c; + uint8x4_t& vc = reinterpret_cast(c); +#pragma unroll + for (int i = 0; i < 4; ++i) { + vc[i] = va[i] == vb[i] ? 0x00 : 0xff; + } return c; } @@ -467,8 +453,48 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half } #endif // CUDART_VERSION < 12000 +static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) { +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) + c = __builtin_amdgcn_sdot4(a, b, c, false); +#elif defined(RDNA3) + c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); +#elif defined(__gfx1010__) || defined(__gfx900__) + int tmp1; + int tmp2; + asm("\n \ + v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \ + v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \ + v_add3_u32 %0, %1, %2, %0 \n \ + v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \ + v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \ + v_add3_u32 %0, %1, %2, %0 \n \ + " + : "+v"(c), "=&v"(tmp1), "=&v"(tmp2) + : "v"(a), "v"(b) + ); +#else + const int8x4_t va = reinterpret_cast(a); + const int8x4_t vb = reinterpret_cast(b); + c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3]; +#endif + return c; + +#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + +#if __CUDA_ARCH__ >= MIN_CC_DP4A + return __dp4a(a, b, c); +#else // __CUDA_ARCH__ >= MIN_CC_DP4A + const int8_t * a8 = (const int8_t *) &a; + const int8_t * b8 = (const int8_t *) &b; + return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3]; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A + +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +} + // TODO: move to ggml-common.h -static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; +static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v); diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh index bd7993595467c2..650780fd2e4624 100644 --- a/ggml/src/ggml-cuda/fattn-common.cuh +++ b/ggml/src/ggml-cuda/fattn-common.cuh @@ -54,12 +54,11 @@ typedef float (*vec_dot_KQ_f32_t)( template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q4_0 * K_q4_0 = (const block_q4_0 *) K_c; GGML_UNUSED(Q_v); - half sum = 0.0f; + T sum = 0.0f; #pragma unroll for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) { @@ -72,7 +71,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( const int v = (get_int_from_uint8(K_q4_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; const int u = Q_q8[k_KQ_0/WARP_SIZE]; - const int sumi = __dp4a(v, u, 0); + const int sumi = ggml_cuda_dp4a(v, u, 0); #ifdef FP16_AVAILABLE if (std::is_same::value) { @@ -90,19 +89,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q4_1 * K_q4_1 = (const block_q4_1 *) K_c; GGML_UNUSED(Q_v); @@ -120,7 +111,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( const int v = (get_int_from_uint8_aligned(K_q4_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; const int u = Q_q8[k_KQ_0/WARP_SIZE]; - const int sumi = __dp4a(v, u, 0); + const int sumi = ggml_cuda_dp4a(v, u, 0); #ifdef FP16_AVAILABLE if (std::is_same::value) { @@ -142,19 +133,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q5_0 * K_q5_0 = (const block_q5_0 *) K_c; GGML_UNUSED(Q_v); @@ -179,7 +162,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( const int u = Q_q8[k_KQ_0/WARP_SIZE]; - const int sumi = __dp4a(v, u, 0); + const int sumi = ggml_cuda_dp4a(v, u, 0); #ifdef FP16_AVAILABLE if (std::is_same::value) { @@ -197,19 +180,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q5_1 * K_q5_1 = (const block_q5_1 *) K_c; GGML_UNUSED(Q_v); @@ -234,7 +209,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( const int u = Q_q8[k_KQ_0/WARP_SIZE]; - const int sumi = __dp4a(v, u, 0); + const int sumi = ggml_cuda_dp4a(v, u, 0); #ifdef FP16_AVAILABLE if (std::is_same::value) { @@ -256,19 +231,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A const block_q8_0 * K_q8_0 = (const block_q8_0 *) K_c; GGML_UNUSED(Q_v); @@ -297,13 +264,6 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( } return sum; -#else - GGML_UNUSED(K_c); - GGML_UNUSED(Q_v); - GGML_UNUSED(Q_q8); - GGML_UNUSED(Q_ds_v); - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu index e8d157169544f7..e22faf69b7287f 100644 --- a/ggml/src/ggml-cuda/mmvq.cu +++ b/ggml/src/ggml-cuda/mmvq.cu @@ -28,16 +28,22 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) static constexpr __device__ int get_vdr_mmvq(ggml_type type) { return type == GGML_TYPE_Q4_0 ? VDR_Q4_0_Q8_1_MMVQ : - type == GGML_TYPE_Q4_1 ? VDR_Q4_1_Q8_1_MMVQ : - type == GGML_TYPE_Q5_0 ? VDR_Q5_0_Q8_1_MMVQ : - type == GGML_TYPE_Q5_1 ? VDR_Q5_1_Q8_1_MMVQ : - type == GGML_TYPE_Q8_0 ? VDR_Q8_0_Q8_1_MMVQ : - type == GGML_TYPE_Q2_K ? VDR_Q2_K_Q8_1_MMVQ : - type == GGML_TYPE_Q3_K ? VDR_Q3_K_Q8_1_MMVQ : - type == GGML_TYPE_Q4_K ? VDR_Q4_K_Q8_1_MMVQ : - type == GGML_TYPE_Q5_K ? VDR_Q5_K_Q8_1_MMVQ : - type == GGML_TYPE_Q6_K ? VDR_Q6_K_Q8_1_MMVQ : - type == GGML_TYPE_IQ4_NL ? VDR_Q4_K_Q8_1_MMVQ : + type == GGML_TYPE_Q4_1 ? VDR_Q4_1_Q8_1_MMVQ : + type == GGML_TYPE_Q5_0 ? VDR_Q5_0_Q8_1_MMVQ : + type == GGML_TYPE_Q5_1 ? VDR_Q5_1_Q8_1_MMVQ : + type == GGML_TYPE_Q8_0 ? VDR_Q8_0_Q8_1_MMVQ : + type == GGML_TYPE_Q2_K ? VDR_Q2_K_Q8_1_MMVQ : + type == GGML_TYPE_Q3_K ? VDR_Q3_K_Q8_1_MMVQ : + type == GGML_TYPE_Q4_K ? VDR_Q4_K_Q8_1_MMVQ : + type == GGML_TYPE_Q5_K ? VDR_Q5_K_Q8_1_MMVQ : + type == GGML_TYPE_Q6_K ? VDR_Q6_K_Q8_1_MMVQ : + type == GGML_TYPE_IQ2_XXS ? VDR_IQ2_XXS_Q8_1_MMVQ : + type == GGML_TYPE_IQ2_XS ? VDR_IQ2_XS_Q8_1_MMVQ : + type == GGML_TYPE_IQ2_S ? VDR_IQ2_S_Q8_1_MMVQ : + type == GGML_TYPE_IQ3_XXS ? VDR_IQ3_XXS_Q8_1_MMVQ : + type == GGML_TYPE_IQ3_S ? VDR_IQ3_S_Q8_1_MMVQ : + type == GGML_TYPE_IQ4_NL ? VDR_IQ4_NL_Q8_1_MMVQ : + type == GGML_TYPE_IQ4_XS ? VDR_IQ4_XS_Q8_1_MMVQ : 1; } diff --git a/ggml/src/ggml-cuda/vecdotq.cuh b/ggml/src/ggml-cuda/vecdotq.cuh index 3b12d656616be4..3f3a87c750b211 100644 --- a/ggml/src/ggml-cuda/vecdotq.cuh +++ b/ggml/src/ggml-cuda/vecdotq.cuh @@ -1,4 +1,5 @@ #include "common.cuh" +#include static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) { const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment @@ -28,6 +29,18 @@ static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment } +static __device__ __forceinline__ int get_int_b2(const void * x, const int & i32) { + const uint16_t * x16 = (const uint16_t *) x; + + int x32 = x16[2*i32 + 0] << 0; + x32 |= x16[2*i32 + 1] << 16; + + return x32; +} + +static __device__ __forceinline__ int get_int_b4(const void * x, const int & i32) { + return ((const int *) x)[i32]; // assume at least 4 byte alignment +} // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q @@ -38,7 +51,6 @@ static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl( const int * v, const int * u, const float & d4, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -47,17 +59,14 @@ template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_imp const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; // SIMD dot product of quantized values - sumi = __dp4a(vi0, u[2*i+0], sumi); - sumi = __dp4a(vi1, u[2*i+1], sumi); + sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); + sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); } const float2 ds8f = __half22float2(ds8); // second part effectively subtracts 8 from each quant value return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q4_1_Q8_1_MMVQ 2 @@ -66,7 +75,6 @@ template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_imp template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl( const int * v, const int * u, const half2 & dm4, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -75,8 +83,8 @@ template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; // SIMD dot product of quantized values - sumi = __dp4a(vi0, u[2*i+0], sumi); - sumi = __dp4a(vi1, u[2*i+1], sumi); + sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); + sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); } #ifdef GGML_CUDA_F16 @@ -92,9 +100,6 @@ template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q5_0_Q8_1_MMVQ 2 @@ -103,7 +108,6 @@ template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl( const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -113,23 +117,20 @@ template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_imp vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 - sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values + sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 - sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values + sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values } const float2 ds8f = __half22float2(ds8); // second part effectively subtracts 16 from each quant value return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q5_1_Q8_1_MMVQ 2 @@ -138,7 +139,6 @@ template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_imp template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl( const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -148,14 +148,14 @@ template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 - sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values + sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 - sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values + sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values } #ifdef GGML_CUDA_F16 @@ -171,10 +171,6 @@ template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it return sumi*d5d8 + m5s8 / (QI5_1 / vdr); - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q8_0_Q8_1_MMVQ 2 @@ -183,31 +179,26 @@ template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp template static __device__ __forceinline__ T vec_dot_q8_0_q8_1_impl( const int * v, const int * u, const T & d8_0, const T & d8_1) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll for (int i = 0; i < vdr; ++i) { // SIMD dot product of quantized values - sumi = __dp4a(v[i], u[i], sumi); + sumi = ggml_cuda_dp4a(v[i], u[i], sumi); } return d8_0*d8_1 * ((T) sumi); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } template static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl( const int * v, const int * u, const half2 & dm8, const half2 & ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll for (int i = 0; i < vdr; ++i) { // SIMD dot product of quantized values - sumi = __dp4a(v[i], u[i], sumi); + sumi = ggml_cuda_dp4a(v[i], u[i], sumi); } #ifdef GGML_CUDA_F16 @@ -223,9 +214,6 @@ template static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it return sumi*d8d8 + m8s8 / (QI8_1 / vdr); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q2_K_Q8_1_MMVQ 1 @@ -236,7 +224,6 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq( const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales, const half2 & dm2, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -246,28 +233,24 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq( const int vi = (v >> (2*i)) & 0x03030303; - sumf_d += d8[i] * (__dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product + sumf_d += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product // fill int with 4x m int m = sc >> 4; m |= m << 8; m |= m << 16; - sumf_m += d8[i] * __dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values + sumf_m += d8[i] * ggml_cuda_dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values } const float2 dm2f = __half22float2(dm2); return dm2f.x*sumf_d - dm2f.y*sumf_m; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const half2 * dm2, const float & d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -281,8 +264,8 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( #pragma unroll for (int i = i0; i < i0 + QI8_1/2; ++i) { const int vi = (vi0 >> (2*(i % (QI8_1/2)))) & 0x03030303; - sumi_d = __dp4a(vi, u[i], sumi_d); // SIMD dot product - sumi_m = __dp4a(0x01010101, u[i], sumi_m); + sumi_d = ggml_cuda_dp4a(vi, u[i], sumi_d); // SIMD dot product + sumi_m = ggml_cuda_dp4a(0x01010101, u[i], sumi_m); } sumf_d += dm2f.x * sumi_d; @@ -290,9 +273,6 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( } return d8*(sumf_d - sumf_m); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q3_K_Q8_1_MMVQ 1 @@ -303,7 +283,6 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq( const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales, const int & scale_offset, const float & d3, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf = 0.0f; #pragma unroll @@ -326,13 +305,10 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq( const int vi = __vsubss4(vil, vih); - sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product + sumf += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * sc); // SIMD dot product } return d3 * sumf; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values @@ -340,7 +316,6 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales, const float & d3, const float & d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics int sumi = 0; #pragma unroll @@ -350,16 +325,13 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq( #pragma unroll for (int i = i0; i < i0 + QI8_1/2; ++i) { const int vi = __vsubss4((v[i/2] >> (4*(i%2))) & 0x0F0F0F0F, 0x04040404); - sumi_sc = __dp4a(vi, u[i], sumi_sc); // SIMD dot product + sumi_sc = ggml_cuda_dp4a(vi, u[i], sumi_sc); // SIMD dot product } sumi += sumi_sc * scales[i0 / (QI8_1/2)]; } return d3*d8 * sumi; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q4_K_Q8_1_MMVQ 2 @@ -370,7 +342,6 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -379,8 +350,8 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F; const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F; - const int dot1 = __dp4a(v1i, u[2*i+1], __dp4a(v0i, u[2*i+0], 0)); // SIMD dot product - const int dot2 = __dp4a(0x01010101, u[2*i+1], __dp4a(0x01010101, u[2*i+0], 0)); // sum of u + const int dot1 = ggml_cuda_dp4a(v1i, u[2*i+1], ggml_cuda_dp4a(v0i, u[2*i+0], 0)); // SIMD dot product + const int dot2 = ggml_cuda_dp4a(0x01010101, u[2*i+1], ggml_cuda_dp4a(0x01010101, u[2*i+0], 0)); // sum of u sumf_d += d8[i] * (dot1 * sc[i]); sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values @@ -389,10 +360,6 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( const float2 dm4f = __half22float2(dm4); return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values @@ -400,7 +367,6 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -410,7 +376,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( #pragma unroll for (int j = 0; j < QI8_1; ++j) { - sumi_d = __dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product + sumi_d = ggml_cuda_dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product } const float2 ds8f = __half22float2(ds8[i]); @@ -422,10 +388,6 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( const float2 dm4f = __half22float2(dm4); return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q5_K_Q8_1_MMVQ 2 @@ -436,7 +398,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc, const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -451,8 +412,8 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( const int v0i = vl0i | vh0i; const int v1i = vl1i | vh1i; - const int dot1 = __dp4a(v0i, u[2*i+0], __dp4a(v1i, u[2*i+1], 0)); // SIMD dot product - const int dot2 = __dp4a(0x01010101, u[2*i+0], __dp4a(0x01010101, u[2*i+1], 0)); // sum of u + const int dot1 = ggml_cuda_dp4a(v0i, u[2*i+0], ggml_cuda_dp4a(v1i, u[2*i+1], 0)); // SIMD dot product + const int dot2 = ggml_cuda_dp4a(0x01010101, u[2*i+0], ggml_cuda_dp4a(0x01010101, u[2*i+1], 0)); // sum of u sumf_d += d8[i] * (dot1 * sc[i]); sumf_m += d8[i] * (dot2 * m[i]); @@ -462,10 +423,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( const float2 dm5f = __half22float2(dm5); return dm5f.x*sumf_d - dm5f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values @@ -473,7 +430,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; float sumf_m = 0.0f; @@ -483,7 +439,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( #pragma unroll for (int j = 0; j < QI8_1; ++j) { - sumi_d = __dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product + sumi_d = ggml_cuda_dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product } const float2 ds8f = __half22float2(ds8[i]); @@ -495,10 +451,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( const float2 dm4f = __half22float2(dm4); return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } #define VDR_Q6_K_Q8_1_MMVQ 1 @@ -509,7 +461,6 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq( const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales, const float & d, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf = 0.0f; #pragma unroll @@ -522,13 +473,10 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq( const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32 - sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product + sumf += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * sc); // SIMD dot product } return d*sumf; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } // contiguous u/y values @@ -536,7 +484,6 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq( const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc, const float & d6, const float * __restrict__ d8) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics float sumf_d = 0.0f; #pragma unroll @@ -545,21 +492,17 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq( #pragma unroll for (int i = i0; i < i0 + 2; ++i) { - sumi_d.x = __dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product - sumi_d.x = __dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product + sumi_d.x = ggml_cuda_dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product + sumi_d.x = ggml_cuda_dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product - sumi_d.y = __dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product - sumi_d.y = __dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product + sumi_d.y = ggml_cuda_dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product + sumi_d.y = ggml_cuda_dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product } sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y); } return d6 * sumf_d; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A } static __device__ __forceinline__ float vec_dot_q4_0_q8_1( @@ -572,9 +515,9 @@ static __device__ __forceinline__ float vec_dot_q4_0_q8_1( #pragma unroll for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); + v[i] = get_int_b2(bq4_0->qs, iqs + i); + u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI4_0); } return vec_dot_q4_0_q8_1_impl(v, u, bq4_0->d, bq8_1->ds); @@ -591,9 +534,9 @@ static __device__ __forceinline__ float vec_dot_q4_1_q8_1( #pragma unroll for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1); + v[i] = get_int_b4(bq4_1->qs, iqs + i); + u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI4_1); } return vec_dot_q4_1_q8_1_impl(v, u, bq4_1->dm, bq8_1->ds); @@ -610,10 +553,10 @@ static __device__ __forceinline__ float vec_dot_q5_0_q8_1( #pragma unroll for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) { - vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i); - vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i)); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0); + vl[i] = get_int_b2(bq5_0->qs, iqs + i); + vh[i] = get_int_b2(bq5_0->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI5_0); } return vec_dot_q5_0_q8_1_impl(vl, vh, u, bq5_0->d, bq8_1->ds); @@ -630,10 +573,10 @@ static __device__ __forceinline__ float vec_dot_q5_1_q8_1( #pragma unroll for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) { - vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i); - vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i)); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1); + vl[i] = get_int_b4(bq5_1->qs, iqs + i); + vh[i] = get_int_b4(bq5_1->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI5_1); } return vec_dot_q5_1_q8_1_impl(vl, vh, u, bq5_1->dm, bq8_1->ds); @@ -649,8 +592,8 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1( #pragma unroll for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_int8(bq8_0->qs, iqs + i); - u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + v[i] = get_int_b2(bq8_0->qs, iqs + i); + u[i] = get_int_b4(bq8_1->qs, iqs + i); } return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, __low2half(bq8_1->ds)); @@ -666,13 +609,13 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1( const uint8_t * scales = bq2_K->scales + scale_offset; - const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs); + const int v = get_int_b4(bq2_K->qs, iqs); int u[QR2_K]; float d8[QR2_K]; #pragma unroll for (int i = 0; i < QR2_K; ++ i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + u[i] = get_int_b4(bq8_1[bq8_offset + i].qs, iqs % QI8_1); d8[i] = __low2float(bq8_1[bq8_offset + i].ds); } @@ -689,17 +632,17 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1( const float d = bq3_K->d; - const int vl = get_int_from_uint8(bq3_K->qs, iqs); + const int vl = get_int_b2(bq3_K->qs, iqs); // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted - const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; + const int vh = ~get_int_b2(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; int u[QR3_K]; float d8[QR3_K]; #pragma unroll for (int i = 0; i < QR3_K; ++i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + u[i] = get_int_b4(bq8_1[bq8_offset + i].qs, iqs % QI8_1); d8[i] = __low2float(bq8_1[bq8_offset + i].ds); } @@ -807,8 +750,8 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1( const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8); const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4)); - const int vl = get_int_from_uint8(bq6_K->ql, iqs); - const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; + const int vl = get_int_b2(bq6_K->ql, iqs); + const int vh = get_int_b2(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; const int8_t * scales = bq6_K->scales + scale_offset; @@ -817,335 +760,342 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1( #pragma unroll for (int i = 0; i < QR6_K; ++i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); + u[i] = get_int_b4(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds); } return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8); } +#define VDR_IQ2_XXS_Q8_1_MMVQ 2 + static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { + const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq + kbx; -#if QR2_XXS == 8 - const int ib32 = iqs; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const int8_t * q8 = bq8_1[ib32].qs; - uint32_t aux32 = q2[2] | (q2[3] << 16); + const int q2 = get_int_b2(bq2->qs, iqs); + const uint8_t * aux8 = (const uint8_t *) &q2; + const uint32_t aux32 = get_int_b2(bq2->qs, iqs + 1); + int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); - const uint8_t signs = ksigns_iq2xs[aux32 & 127]; - for (int j = 0; j < 8; ++j) { - sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - aux32 >>= 7; +#pragma unroll + for (int k0 = 0; k0 < 8; k0 += 2) { + const int * grid_pos = (const int *) (iq2xxs_grid + aux8[k0/2]); + const int signs_packed = ksigns_iq2xs[(aux32 >> (7*k0/2)) & 0x7F]; + + const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000); + const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0); + const int u0 = get_int_b4(bq8_1[iqs/2].qs, k0 + 0); + sumi = ggml_cuda_dp4a(grid0, u0, sumi); + + const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000); + const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, k0 + 1); + sumi = ggml_cuda_dp4a(grid1, u1, sumi); } - const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f; + + const int ls = aux32 >> 28; + sumi = (ls*sumi + sumi/2)/4; + const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds); return d * sumi; -#else - // iqs is 0...15 - const int ib32 = iqs/2; - const int il = iqs%2; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); - const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f; - const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; - const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; - const int8_t * q8 = bq8_1[ib32].qs + 16*il; - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); - sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); - } - return d * (sumi1 + sumi2); -#endif } +#define VDR_IQ2_XS_Q8_1_MMVQ 2 + static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq + kbx; - const int ib32 = iqs; - const uint16_t * q2 = bq2->qs + 4*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - const uint8_t ls1 = bq2->scales[ib32] & 0xf; - const uint8_t ls2 = bq2->scales[ib32] >> 4; + const int2 q2_packed = make_int2(get_int_b2(bq2->qs, iqs + 0), get_int_b2(bq2->qs, iqs + 1)); + const uint16_t * q2 = (const uint16_t *) &q2_packed; + const int ls0 = bq2->scales[iqs/2] & 0x0F; + const int ls1 = bq2->scales[iqs/2] >> 4; + + int sumi0 = 0; int sumi1 = 0; - for (int l = 0; l < 2; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); - const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); - const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]); - sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1); - sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1); - q8 += 8; - } - int sumi2 = 0; - for (int l = 2; l < 4; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); - const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); - const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]); - sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2); - sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2); - q8 += 8; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l0/2] & 0x000001FF)); + const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l0/2] >> 9)); + + const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]); + const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]); + + const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1); + + if (l0 < 4) { + sumi0 = ggml_cuda_dp4a(grid_l, u0, sumi0); + sumi0 = ggml_cuda_dp4a(grid_h, u1, sumi0); + } else { + sumi1 = ggml_cuda_dp4a(grid_l, u0, sumi1); + sumi1 = ggml_cuda_dp4a(grid_h, u1, sumi1); + } } - const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; - return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -#else - GGML_UNUSED(ksigns64); - NO_DEVICE_CODE; -#endif + const int sumi = (sumi0*ls0 + sumi1*ls1 + (sumi0 + sumi1)/2)/4; + const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds); + return d * sumi; } -// TODO +#define VDR_IQ2_S_Q8_1_MMVQ 2 + static __device__ __forceinline__ float vec_dot_iq2_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + const block_iq2_s * bq2 = (const block_iq2_s *) vbq + kbx; - const int ib32 = iqs; - const int8_t * q8 = bq8_1[ib32].qs; - const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32; - const uint8_t ls1 = bq2->scales[ib32] & 0xf; - const uint8_t ls2 = bq2->scales[ib32] >> 4; + const int qs_packed = get_int_b2(bq2->qs, iqs/2); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bq2->qh[iqs/2]; + + const int signs_packed_32 = get_int_b2(bq2->qs, QK_K/32 + iqs/2); + const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; + + const int ls0 = bq2->scales[iqs/2] & 0x0F; + const int ls1 = bq2->scales[iqs/2] >> 4; + + int sumi0 = 0; int sumi1 = 0; - for (int l = 0; l < 2; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); - const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid[1] ^ signs1, signs1); - sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1); - sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1); - q8 += 8; - } - int sumi2 = 0; - for (int l = 2; l < 4; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); - const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid[1] ^ signs1, signs1); - sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2); - sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2); - q8 += 8; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int * grid_pos = (const int *)(iq2s_grid + (qs[l0/2] | ((qh << (8-l0)) & 0x300))); + + const int signs0 = __vcmpne4(((signs_packed_8[l0/2] & 0x03) << 7) | ((signs_packed_8[l0/2] & 0x0C) << 21), 0x00000000); + const int signs1 = __vcmpne4(((signs_packed_8[l0/2] & 0x30) << 3) | ((signs_packed_8[l0/2] & 0xC0) << 17), 0x00000000); + + const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0); + const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1); + + const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1); + + if (l0 < 4) { + sumi0 = ggml_cuda_dp4a(grid_l, u0, sumi0); + sumi0 = ggml_cuda_dp4a(grid_h, u1, sumi0); + } else { + sumi1 = ggml_cuda_dp4a(grid_l, u0, sumi1); + sumi1 = ggml_cuda_dp4a(grid_h, u1, sumi1); + } } - const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; - return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -#else - GGML_UNUSED(ksigns64); - NO_DEVICE_CODE; -#endif + const int sumi = (sumi0*ls0 + sumi1*ls1 + (sumi0 + sumi1)/2)/4; + + const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds); + return d * sumi; } +#define VDR_IQ3_XXS_Q8_1_MMVQ 2 + static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq + kbx; - - const int ib32 = iqs; - const uint8_t * q3 = bq2->qs + 8*ib32; - const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - uint32_t aux32 = gas[0] | (gas[1] << 16); + + const block_iq3_xxs * bq3 = (const block_iq3_xxs *) vbq + kbx; + + const int2 q3_packed = make_int2(get_int_b2(bq3->qs, iqs), get_int_b2(bq3->qs, iqs+1)); + const uint8_t * q3 = (const uint8_t *) &q3_packed; + const uint32_t aux32 = get_int_b2(bq3->qs, QK_K/16 + iqs/2); + int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0]; - const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1]; - const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127)); - const int grid_l = __vsub4(grid1[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid2[0] ^ signs[1], signs[1]); - sumi = __dp4a(grid_l, *((int *)q8+0), sumi); - sumi = __dp4a(grid_h, *((int *)q8+1), sumi); - q8 += 8; - aux32 >>= 7; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int2 grid_pos = make_int2(iq3xxs_grid[q3[l0 + 0]], iq3xxs_grid[q3[l0 + 1]]); + + const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l0/2)) & 0x7F)); + + const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]); + const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]); + + const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1); + + sumi = ggml_cuda_dp4a(grid_l, u0, sumi); + sumi = ggml_cuda_dp4a(grid_h, u1, sumi); } - const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f; + + const int ls = aux32 >> 28; + sumi = (ls*sumi + sumi/2)/2; + const float d = __half2float(bq3->d) * __low2float(bq8_1[iqs/2].ds); return d * sumi; -#else - NO_DEVICE_CODE; -#endif } +#define VDR_IQ3_S_Q8_1_MMVQ 2 + // TODO: don't use lookup table for signs static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_iq3_s * bq2 = (const block_iq3_s *) vbq + kbx; - const int ib32 = iqs; - const uint8_t * qs = bq2->qs + 8*ib32; - const int8_t * q8 = bq8_1[ib32].qs; + const block_iq3_s * bq3 = (const block_iq3_s *) vbq + kbx; + + const int2 qs_packed = make_int2(get_int_b2(bq3->qs, iqs + 0), get_int_b2(bq3->qs, iqs + 1)); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bq3->qh[iqs/2]; + + const int signs_packed_32 = get_int_b2(bq3->signs, iqs/2); + const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32; + int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)); - const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)); - uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid1[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid2[0] ^ signs1, signs1); - sumi = __dp4a(grid_l, *((int *)q8+0), sumi); - sumi = __dp4a(grid_h, *((int *)q8+1), sumi); - q8 += 8; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int2 grid_pos = make_int2( + iq3s_grid[qs[l0 + 0] | ((qh << (8 - l0)) & 0x100)], + iq3s_grid[qs[l0 + 1] | ((qh << (7 - l0)) & 0x100)]); + + const int signs0 = __vcmpne4(((signs_packed_8[l0/2] & 0x03) << 7) | ((signs_packed_8[l0/2] & 0x0C) << 21), 0x00000000); + const int signs1 = __vcmpne4(((signs_packed_8[l0/2] & 0x30) << 3) | ((signs_packed_8[l0/2] & 0xC0) << 17), 0x00000000); + + const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); + const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); + + const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1); + + sumi = ggml_cuda_dp4a(grid_l, u0, sumi); + sumi = ggml_cuda_dp4a(grid_h, u1, sumi); } - const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds); + + sumi *= 1 + 2*((bq3->scales[iqs/4] >> ((iqs << 1) & 0x04)) & 0x0F); + + const float d = __half2float(bq3->d) * __low2float(bq8_1[iqs/2].ds); return d * sumi; -#else - NO_DEVICE_CODE; -#endif } static __device__ __forceinline__ float vec_dot_iq1_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { const block_iq1_s * bq1 = (const block_iq1_s *) vbq + kbx; - const int ib32 = iqs; + const int qs_packed = get_int_b2(bq1->qs, iqs); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + const int qh = bq1->qh[iqs]; + int sumi = 0; -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const int * q8 = (const int *)bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); - int grid0 = grid[0] & 0x0f0f0f0f; - int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; - sumi = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi)); - } -#else - const int8_t * q8 = bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); - for (int j = 0; j < 4; ++j) { - sumi += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4); - } - q8 += 8; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int grid = iq1s_grid_gpu[qs[l0/2] | (((qh >> 3*(l0/2)) & 0x07) << 8)]; + + const int grid0 = (grid >> 0) & 0x0F0F0F0F; + const int grid1 = (grid >> 4) & 0x0F0F0F0F; + + const int u0 = get_int_b4(bq8_1[iqs].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs].qs, l0 + 1); + + sumi = ggml_cuda_dp4a(grid0, u0, sumi); + sumi = ggml_cuda_dp4a(grid1, u1, sumi); } -#endif - const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA; - const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1); - const float d = d1q * __low2float (bq8_1[ib32].ds); - const float m = d1q * __high2float(bq8_1[ib32].ds); - return d * sumi + m * delta; + + const float d1q = __half2float(bq1->d) * (((qh >> 11) & 0x0E) + 1); + const float delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000); + const float2 ds = __half22float2(bq8_1[iqs].ds); + return d1q * (ds.x*sumi + ds.y*delta); } static __device__ __forceinline__ float vec_dot_iq1_m_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { + const block_iq1_m * bq1 = (const block_iq1_m *) vbq + kbx; - const int ib32 = iqs; - int sumi[2] = {0, 0}; - float sumf[2] = {0.f, 0.f}; -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const int * q8 = (const int *)bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8))); - int grid0 = grid[0] & 0x0f0f0f0f; - int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; - sumi[l/2] = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi[l/2])); - const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA; - const int sumy = __dp4a(q8[2*l+1], 0x01010101, __dp4a(q8[2*l+0], 0x01010101, 0)); - sumf[l/2] += delta*sumy; - } -#else - const int8_t * q8 = bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); + const int qs_packed = get_int_b4(bq1->qs, iqs); + const uint8_t * qs = (const uint8_t *) &qs_packed; + + int sumi[2] = {0}; + float sumf[2] = {0.0f}; +#pragma unroll + for (int l0 = 0; l0 < 8; l0 += 2) { + const int qhl = bq1->qh[2*iqs + l0/4] >> (4 * ((l0/2) % 2)); + + const int grid = iq1s_grid_gpu[qs[l0/2] | ((qhl & 0x07) << 8)]; + + const int grid0 = (grid >> 0) & 0x0F0F0F0F; + const int grid1 = (grid >> 4) & 0x0F0F0F0F; + + const int u0 = get_int_b4(bq8_1[iqs].qs, l0 + 0); + const int u1 = get_int_b4(bq8_1[iqs].qs, l0 + 1); + + sumi[l0/4] = ggml_cuda_dp4a(grid0, u0, sumi[l0/4]); + sumi[l0/4] = ggml_cuda_dp4a(grid1, u1, sumi[l0/4]); + + const float delta = -1.0f + IQ1M_DELTA - (qhl & 0x08) * (2.0f*IQ1M_DELTA/0x08); int sumy = 0; - for (int j = 0; j < 4; ++j) { - sumi[l/2] += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4); - sumy += q8[j] + q8[j+4]; - } - const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA; - sumf[l/2] += delta*sumy; - q8 += 8; + sumy = ggml_cuda_dp4a(u0, 0x01010101, sumy); + sumy = ggml_cuda_dp4a(u1, 0x01010101, sumy); + sumf[l0/4] += delta*sumy; } -#endif + + const uint16_t * sc = (const uint16_t *) bq1->scales; + iq1m_scale_t scale; - const uint16_t * sc = (const uint16_t *)bq1->scales; - scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); - const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds); - return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1)); + scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00F0) | ((sc[2] >> 4) & 0x0F00) | (sc[3] & 0xF000); + const float d = __half2float(scale.f16) * __low2float(bq8_1[iqs].ds); + + const int tmp = sc[iqs/2] >> (6*(iqs%2)); + const int sc0 = 2*((tmp >> 0) & 0x07) + 1; + const int sc1 = 2*((tmp >> 3) & 0x07) + 1; + return d * ((sumi[0] + sumf[0]) * sc0 + (sumi[1] + sumf[1]) * sc1); } -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values, - int & val1, int & val2) { - - uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32; - aux32 = q4 & 0x0f0f0f0f; - uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8); - uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8); - val1 = v1 | (v2 << 16); - aux32 = (q4 >> 4) & 0x0f0f0f0f; - v1 = values[q8[0]] | (values[q8[1]] << 8); - v2 = values[q8[2]] | (values[q8[3]] << 8); - val2 = v1 | (v2 << 16); +static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4) { + const int q0_32 = (q4 >> 0) & 0x0F0F0F0F; + const int8_t * q0_8 = (const int8_t *) &q0_32; + const char4 val0_8 = make_char4( + kvalues_iq4nl[q0_8[0]], kvalues_iq4nl[q0_8[1]], kvalues_iq4nl[q0_8[2]], kvalues_iq4nl[q0_8[3]]); + + const int q1_32 = (q4 >> 4) & 0x0F0F0F0F; + const int8_t * q1_8 = (const int8_t *) &q1_32; + const char4 val1_8 = make_char4( + kvalues_iq4nl[q1_8[0]], kvalues_iq4nl[q1_8[1]], kvalues_iq4nl[q1_8[2]], kvalues_iq4nl[q1_8[3]]); + + return make_int2(*((const int *) &val0_8), *((const int *) &val1_8)); } -#endif + +#define VDR_IQ4_NL_Q8_1_MMVQ 2 static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { - const block_iq4_nl * bq = (const block_iq4_nl *) vbq + kbx; + const block_iq4_nl * bq4 = (const block_iq4_nl *) vbq + kbx; -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs; - const int32_t * q8 = (const int32_t *)bq8_1->qs + iqs; + const int * q8 = (const int *) bq8_1->qs + iqs; - const uint8_t * values = (const uint8_t *)kvalues_iq4nl; - - int v1, v2; - int sumi1 = 0, sumi2 = 0; + int sumi = 0; +#pragma unroll for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) { - const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16); - get_int_from_table_16(aux, values, v1, v2); - sumi1 = __dp4a(v1, q8[l+0], sumi1); - sumi2 = __dp4a(v2, q8[l+4], sumi2); - } - -#else - const uint8_t * q4 = bq->qs + 4*iqs; - const int8_t * q8 = bq8_1->qs + 4*iqs; + const int aux_q4 = get_int_b2(bq4->qs, iqs + l); + const int2 v = get_int_from_table_16(aux_q4); - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) { - sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf]; - sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >> 4]; + sumi = ggml_cuda_dp4a(v.x, q8[l + 0], sumi); + sumi = ggml_cuda_dp4a(v.y, q8[l + 4], sumi); } -#endif - const float d = (float)bq->d * __low2float(bq8_1->ds); - return d * (sumi1 + sumi2); + + const float d = __half2float(bq4->d) * __low2float(bq8_1->ds); + return d * sumi; } +#define VDR_IQ4_XS_Q8_1_MMVQ 4 + static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq + kbx; - const uint8_t * values = (const uint8_t *)kvalues_iq4nl; - - // iqs is 0...7 - const int ib32 = iqs; - const int32_t * q8 = (const int *)bq8_1[ib32].qs; - const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32; - const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4); - const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds); - int v1, v2; - int sumi1 = 0, sumi2 = 0; + + int sumi = 0; +#pragma unroll for (int j = 0; j < 4; ++j) { - get_int_from_table_16(q4[j], values, v1, v2); - sumi1 = __dp4a(v1, q8[j+0], sumi1); - sumi2 = __dp4a(v2, q8[j+4], sumi2); + const int aux_q4 = get_int_b4(bq4->qs, iqs + j); + const int2 v = get_int_from_table_16(aux_q4); + + const int u0 = get_int_b4(bq8_1[iqs/4].qs, j + 0); + const int u1 = get_int_b4(bq8_1[iqs/4].qs, j + 4); + + sumi = ggml_cuda_dp4a(v.x, u0, sumi); + sumi = ggml_cuda_dp4a(v.y, u1, sumi); } - return d * (sumi1 + sumi2); -#else - return vec_dot_iq4_xs_q8_1(vbq, bq8_1, kbx, iqs); -#endif + + const int ls = ((bq4->scales_l[iqs/8] >> (iqs & 0x04)) & 0x0F) | (((bq4->scales_h >> (iqs/2)) & 0x03) << 4); + sumi *= ls - 32; + + const float d = __half2float(bq4->d) * __low2float(bq8_1[iqs/4].ds); + return d * sumi; } diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 23227649e2661e..9b751f3c67281f 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -735,7 +735,7 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_xxs_q8_1( + mul_mat_vec_q_iq2_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -760,7 +760,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_xs_q8_1( + mul_mat_vec_q_iq2_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -785,7 +785,7 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_s_q8_1( + mul_mat_vec_q_iq2_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -810,7 +810,7 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq3_xxs_q8_1( + mul_mat_vec_q_iq3_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -834,7 +834,7 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq3_s_q8_1( + mul_mat_vec_q_iq3_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); @@ -924,7 +924,7 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq4_xs_q8_1( + mul_mat_vec_q_iq4_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); }); diff --git a/ggml/src/ggml-sycl/vecdotq.hpp b/ggml/src/ggml-sycl/vecdotq.hpp index 5e2e825463cde3..d2dccade20bfd6 100644 --- a/ggml/src/ggml-sycl/vecdotq.hpp +++ b/ggml/src/ggml-sycl/vecdotq.hpp @@ -820,7 +820,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, #if QK_K == 256 const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; -#if QR2_XXS == 8 const int ib32 = iqs; const uint16_t * q2 = bq2->qs + 4*ib32; const uint8_t * aux8 = (const uint8_t *)q2; @@ -838,26 +837,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, } const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.25f; return d * sumi; -#else - // iqs is 0...15 - const int ib32 = iqs/2; - const int il = iqs%2; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); - const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * bq8_1[ib32].ds[0] * 0.25f; - const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; - const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; - const int8_t * q8 = bq8_1[ib32].qs + 16*il; - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); - sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); - } - return d * (sumi1 + sumi2); -#endif #else assert(false); return 0.f; From 5fac350b9cc49d0446fc291b9c4ad53666c77591 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Tue, 2 Jul 2024 01:07:23 +0200 Subject: [PATCH 10/27] Fix gemma2 tokenizer convert (#8244) * fix gemma2 tokenizer convert * remove scores * improve code, fix new line issue --- convert-hf-to-gguf.py | 37 +++++++++++++++++++++++++++---------- 1 file changed, 27 insertions(+), 10 deletions(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 4a7f500ff7d5ca..6833e943765f79 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -576,7 +576,19 @@ def _set_vocab_qwen(self): special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"]) special_vocab.add_to_gguf(self.gguf_writer) - def _set_vocab_sentencepiece(self): + def _set_vocab_sentencepiece(self, add_to_gguf=True): + tokens, scores, toktypes = self._create_vocab_sentencepiece() + + self.gguf_writer.add_tokenizer_model("llama") + self.gguf_writer.add_tokenizer_pre("default") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) + + def _create_vocab_sentencepiece(self): from sentencepiece import SentencePieceProcessor tokenizer_path = self.dir_model / 'tokenizer.model' @@ -638,14 +650,7 @@ def _set_vocab_sentencepiece(self): scores.append(-1000.0) toktypes.append(SentencePieceTokenTypes.UNUSED) - self.gguf_writer.add_tokenizer_model("llama") - self.gguf_writer.add_tokenizer_pre("default") - self.gguf_writer.add_token_list(tokens) - self.gguf_writer.add_token_scores(scores) - self.gguf_writer.add_token_types(toktypes) - - special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) - special_vocab.add_to_gguf(self.gguf_writer) + return tokens, scores, toktypes def _set_vocab_llama_hf(self): vocab = gguf.LlamaHfVocab(self.dir_model) @@ -2345,7 +2350,19 @@ class Gemma2Model(Model): model_arch = gguf.MODEL_ARCH.GEMMA2 def set_vocab(self): - self._set_vocab_llama_hf() + tokens, scores, toktypes = self._create_vocab_sentencepiece() + # hack: This is required so that we can properly use start/end-of-turn for chat template + for i in range(108): + # including , , + toktypes[i] = SentencePieceTokenTypes.CONTROL + self.gguf_writer.add_tokenizer_model("llama") + self.gguf_writer.add_tokenizer_pre("default") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) self.gguf_writer.add_add_space_prefix(False) def set_gguf_parameters(self): From d08c20eddedb24515a3212e2de66bdff41a26b8c Mon Sep 17 00:00:00 2001 From: luoyu-intel Date: Tue, 2 Jul 2024 02:16:00 +0000 Subject: [PATCH 11/27] [SYCL] Fix the sub group size of Intel (#8106) * use warp_size macro for all sycl kernels * fix mask of permute_sub_group_by_xor * fix rms_norm with correct warp number * fix rms_norm_f32/group_norm_f32 * move norm to norm.cpp file * fix quantize bug * fix mmvq's batch size --- ggml/src/CMakeLists.txt | 4 +- ggml/src/ggml-sycl.cpp | 472 +++------------------------------ ggml/src/ggml-sycl/backend.hpp | 1 + ggml/src/ggml-sycl/common.hpp | 55 ++++ ggml/src/ggml-sycl/dmmv.cpp | 44 +-- ggml/src/ggml-sycl/mmvq.cpp | 113 ++++---- ggml/src/ggml-sycl/norm.cpp | 370 ++++++++++++++++++++++++++ ggml/src/ggml-sycl/norm.hpp | 35 +++ ggml/src/ggml-sycl/presets.hpp | 2 +- 9 files changed, 587 insertions(+), 509 deletions(-) create mode 100644 ggml/src/ggml-sycl/norm.cpp create mode 100644 ggml/src/ggml-sycl/norm.hpp diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index d0f4097d8cd0c8..a18198f1693e59 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -486,9 +486,11 @@ if (GGML_SYCL) add_compile_options(-I./) #include DPCT set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing") - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3") if (GGML_SYCL_TARGET STREQUAL "NVIDIA") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda") + add_compile_definitions(GGML_SYCL_WARP_SIZE=32) + else() + add_compile_definitions(GGML_SYCL_WARP_SIZE=16) endif() file(GLOB GGML_HEADERS_SYCL "ggml-sycl/*.hpp") diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index 30d8a5b33b6133..76bad57e2320b3 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -74,51 +74,6 @@ typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const gg const float *src1_dd, float *dst_dd, const queue_ptr &main_stream); -static __dpct_inline__ float warp_reduce_sum(float x, - const sycl::nd_item<3> &item_ct1) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - /* - DPCT1096:98: The right-most dimension of the work-group used in the SYCL - kernel that calls this function may be less than "32". The function - "dpct::permute_sub_group_by_xor" may return an unexpected result on the - CPU device. Modify the size of the work-group to ensure that the value - of the right-most dimension is a multiple of "32". - */ - x += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), x, mask); - } - return x; -} - -static __dpct_inline__ sycl::float2 -warp_reduce_sum(sycl::float2 a, const sycl::nd_item<3> &item_ct1) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - a.x() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.x(), - mask); - a.y() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.y(), - mask); - } - return a; -} - -static __dpct_inline__ float warp_reduce_max(float x, - const sycl::nd_item<3> &item_ct1) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - /* - DPCT1096:97: The right-most dimension of the work-group used in the SYCL - kernel that calls this function may be less than "32". The function - "dpct::permute_sub_group_by_xor" may return an unexpected result on the - CPU device. Modify the size of the work-group to ensure that the value - of the right-most dimension is a multiple of "32". - */ - x = sycl::fmax(x, dpct::permute_sub_group_by_xor( - item_ct1.get_sub_group(), x, mask)); - } - return x; -} - static __dpct_inline__ float op_repeat(const float a, const float b) { return b; GGML_UNUSED(a); @@ -336,47 +291,6 @@ static void sqr_f32(const float * x, float * dst, const int k, dst[i] = x[i] * x[i]; } -static void norm_f32(const float * x, float * dst, const int ncols, const float eps, - const sycl::nd_item<3> &item_ct1, sycl::float2 *s_sum, int block_size) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - const int tid = item_ct1.get_local_id(2); - - sycl::float2 mean_var = sycl::float2(0.f, 0.f); - - for (int col = tid; col < ncols; col += block_size) { - const float xi = x[row*ncols + col]; - mean_var.x() += xi; - mean_var.y() += xi * xi; - } - - // sum up partial sums - mean_var = warp_reduce_sum(mean_var, item_ct1); - if (block_size > WARP_SIZE) { - - int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = mean_var; - } - /* - DPCT1118:0: SYCL group functions and algorithms must be encountered in - converged control flow. You may need to adjust the code. - */ - item_ct1.barrier(sycl::access::fence_space::local_space); - mean_var = s_sum[lane_id]; - mean_var = warp_reduce_sum(mean_var, item_ct1); - } - - const float mean = mean_var.x() / ncols; - const float var = mean_var.y() / ncols - mean * mean; - const float inv_std = sycl::rsqrt(var + eps); - - for (int col = tid; col < ncols; col += block_size) { - dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std; - } -} - static void concat_f32(const float *x,const float *y, float *dst, const int ne0, const int ne02, const sycl::nd_item<3> &item_ct1) { int nidx = item_ct1.get_local_id(2) + @@ -444,126 +358,11 @@ static void pad_f32(const float *x, float *dst, const int ne0, const int ne00, } } -static void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps, - const sycl::nd_item<3> &item_ct1, float *s_sum, int block_size) { - int start = item_ct1.get_group(2) * group_size; - int end = start + group_size; - - start += item_ct1.get_local_id(2); - - if (end >= ne_elements) { - end = ne_elements; - } - - float tmp = 0.0f; // partial sum for thread in warp - - for (int j = start; j < end; j += block_size) { - tmp += x[j]; - } - - tmp = warp_reduce_sum(tmp, item_ct1); - if (block_size > WARP_SIZE) { - - int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; - } - /* - DPCT1118:1: SYCL group functions and algorithms must be encountered in - converged control flow. You may need to adjust the code. - */ - /* - DPCT1065:54: Consider replacing sycl::nd_item::barrier() with - sycl::nd_item::barrier(sycl::access::fence_space::local_space) for - better performance if there is no access to global memory. - */ - item_ct1.barrier(); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp, item_ct1); - } - - float mean = tmp / group_size; - tmp = 0.0f; - - for (int j = start; j < end; j += block_size) { - float xi = x[j] - mean; - dst[j] = xi; - tmp += xi * xi; - } - - tmp = warp_reduce_sum(tmp, item_ct1); - if (block_size > WARP_SIZE) { - - int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; - } - /* - DPCT1118:2: SYCL group functions and algorithms must be encountered in - converged control flow. You may need to adjust the code. - */ - /* - DPCT1065:55: Consider replacing sycl::nd_item::barrier() with - sycl::nd_item::barrier(sycl::access::fence_space::local_space) for - better performance if there is no access to global memory. - */ - item_ct1.barrier(); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp, item_ct1); - } - - float variance = tmp / group_size; - float scale = sycl::rsqrt(variance + eps); - for (int j = start; j < end; j += block_size) { - dst[j] *= scale; - } -} - -static void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps, - const sycl::nd_item<3> &item_ct1, float *s_sum, int block_size) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - const int tid = item_ct1.get_local_id(2); - - float tmp = 0.0f; // partial sum for thread in warp - - for (int col = tid; col < ncols; col += block_size) { - const float xi = x[row*ncols + col]; - tmp += xi * xi; - } - - // sum up partial sums - tmp = warp_reduce_sum(tmp, item_ct1); - if (block_size > WARP_SIZE) { - - int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; - } - /* - DPCT1118:3: SYCL group functions and algorithms must be encountered in - converged control flow. You may need to adjust the code. - */ - item_ct1.barrier(sycl::access::fence_space::local_space); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp, item_ct1); - } - - const float mean = tmp / ncols; - const float scale = sycl::rsqrt(mean + eps); - - for (int col = tid; col < ncols; col += block_size) { - dst[row*ncols + col] = scale * x[row*ncols + col]; - } -} - +template static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded, const sycl::nd_item<3> &item_ct1) { - const int ix = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); + const int ix = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2)) * QUANT_BLOCK_TILE; if (ix >= kx_padded) { return; @@ -578,23 +377,39 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int ib = i_padded / QK8_1; // block index const int iqs = i_padded % QK8_1; // quant index - - const float xi = ix < kx ? x[iy*kx + ix] : 0.0f; - float amax = sycl::fabs((float)xi); - float sum = xi; - + typedef sycl::vec TC; + typedef sycl::vec TQ; + TC zeros; + TQ qzeros; #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - amax = sycl::fmax(amax, dpct::permute_sub_group_by_xor( - item_ct1.get_sub_group(), amax, mask)); - sum += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), sum, mask); + for (int i = 0; i < QUANT_BLOCK_TILE; i++) + { + zeros[i] = 0.f; + qzeros[i] = 0; + } + const TC xi = ix < kx ? *(TC *)&x[iy * kx + ix] : zeros; + float sum = xi[0]; + float amax = sycl::fabs(xi[0]); +#pragma unroll + for (int i = 1; i < QUANT_BLOCK_TILE; i++) + { + sum += xi[i]; + amax = sycl::fmax(sycl::fabs(xi[i]), amax); } + sum = warp_reduce_sum(sum, item_ct1); + amax = warp_reduce_max(amax, item_ct1); const float d = amax / 127; - const int8_t q = amax == 0.0f ? 0 : sycl::round(xi / d); + TQ q = qzeros; + if (amax != 0.0f) + { +#pragma unroll + for (int i = 0; i < QUANT_BLOCK_TILE; i++) { + q[i] = sycl::round(xi[i] / d); + } + } - y[ib].qs[iqs] = q; + *(TQ *)&y[ib].qs[iqs] = q; if (iqs > 0) { return; @@ -728,7 +543,7 @@ static void mul_mat_p021_f16_f32( // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -781,7 +596,7 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -1643,99 +1458,6 @@ static void sqr_f32_sycl(const float *x, float *dst, const int k, }); } -static void norm_f32_sycl(const float *x, float *dst, const int ncols, - const int nrows, const float eps, - queue_ptr stream) { - GGML_ASSERT(ncols % WARP_SIZE == 0); - if (ncols < 1024) { - const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1( - sycl::range<1>(32), cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - const int work_group_size = get_work_group_size(stream->get_device()); - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:17: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1( - sycl::range<1>(32), cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - -static void group_norm_f32_sycl(const float *x, float *dst, - const int num_groups, const int group_size, - const int ne_elements, queue_ptr stream) { - static const float eps = 1e-6f; - if (group_size < 1024) { - const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - const float eps_ct4 = eps; - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - group_norm_f32( - x, dst, group_size, ne_elements, eps_ct4, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - const int work_group_size = get_work_group_size(stream->get_device()); - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:18: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - const float eps_ct4 = eps; - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - group_norm_f32(x, dst, group_size, ne_elements, - eps_ct4, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - static void concat_f32_sycl(const float *x, const float *y, float *dst, const int ne0, int ne1, int ne2, int ne02, queue_ptr stream) { @@ -1777,64 +1499,22 @@ static void pad_f32_sycl(const float *x, float *dst, const int ne00, }); } -static void rms_norm_f32_sycl(const float *x, float *dst, const int ncols, - const int nrows, const float eps, - queue_ptr stream) { - GGML_ASSERT(ncols % WARP_SIZE == 0); - // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); - if (ncols < 1024) { - const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - rms_norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - const int work_group_size = get_work_group_size(stream->get_device()); - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:19: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - rms_norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx, const int ky, const int kx_padded, queue_ptr stream) { const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE; const sycl::range<3> num_blocks(1, ky, block_num_x); - const sycl::range<3> block_size(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE); + int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE; + static_assert(QK8_1 % WARP_SIZE == 0); + const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE); { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); stream->parallel_for( sycl::nd_range<3>(num_blocks * block_size, block_size), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - quantize_q8_1(x, vy, kx, kx_padded, item_ct1); + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + quantize_q8_1(x, vy, kx, kx_padded, item_ct1); }); } } @@ -1854,7 +1534,7 @@ static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_p021_f16_f32(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y, item_ct1); }); @@ -1874,7 +1554,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl( stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_nc_f16_f32(vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y / nchannels_x, item_ct1); @@ -2139,7 +1819,7 @@ static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols, const sycl::range<3> block_nums(1, nrows, 1); stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { k_sum_rows_f32(x, dst, ncols, item_ct1); }); } @@ -2220,7 +1900,7 @@ static void soft_max_f32_submitter(const float * x, const float * mask, float * cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { soft_max_f32(x, mask, dst, ncols_par, nrows_y, scale, max_bias, m0, m1, n_head_log2, item_ct1, @@ -2400,12 +2080,6 @@ static inline int get_sycl_env(const char *env_name, int default_val) { return user_number; } -static inline int get_work_group_size(const sycl::device& device) { - dpct::device_info prop; - dpct::get_device_info(prop, device); - return prop.get_max_work_group_size(); -} - static void ggml_check_sycl() try { static bool initialized = false; @@ -2964,45 +2638,6 @@ inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor (void) src1_dd; } -inline void ggml_sycl_op_norm(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ne00 = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -inline void ggml_sycl_op_group_norm(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int num_groups = dst->op_params[0]; - int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); - group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - inline void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const float *src0_dd, const float *src1_dd, @@ -3066,28 +2701,6 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor (void) src1_dd; } -inline void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ne00 = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - static int64_t get_row_rounding(ggml_type type, const std::array & tensor_split) { int64_t min_compute_capability = INT_MAX; int64_t max_compute_capability = INT_MIN; @@ -4273,7 +3886,6 @@ bool ggml_sycl_supports_dmmv(enum ggml_type type) { static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer); - int64_t min_compute_capability = INT_MAX; if (split) { diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index d5a63cd710cc34..3afa3391938f2c 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -20,5 +20,6 @@ #include "mmq.hpp" #include "mmvq.hpp" #include "rope.hpp" +#include "norm.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index e01f91633a4bff..dfd4a7c2c606bb 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -295,5 +295,60 @@ struct ggml_backend_sycl_context { } }; +// common host functions + +static inline int get_work_group_size(const sycl::device& device) { + dpct::device_info prop; + dpct::get_device_info(prop, device); + return prop.get_max_work_group_size(); +} + + +// common device functions + +static __dpct_inline__ float warp_reduce_sum(float x, + const sycl::nd_item<3>& item_ct1) { +#pragma unroll + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + /* + DPCT1096:98: The right-most dimension of the work-group used in the SYCL + kernel that calls this function may be less than "32". The function + "dpct::permute_sub_group_by_xor" may return an unexpected result on the + CPU device. Modify the size of the work-group to ensure that the value + of the right-most dimension is a multiple of "32". + */ + x += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), x, mask); + } + return x; +} + +static __dpct_inline__ sycl::float2 +warp_reduce_sum(sycl::float2 a, const sycl::nd_item<3>& item_ct1) { +#pragma unroll + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + a.x() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.x(), + mask); + a.y() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.y(), + mask); + } + return a; +} + +static __dpct_inline__ float warp_reduce_max(float x, + const sycl::nd_item<3>& item_ct1) { +#pragma unroll + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + /* + DPCT1096:97: The right-most dimension of the work-group used in the SYCL + kernel that calls this function may be less than "32". The function + "dpct::permute_sub_group_by_xor" may return an unexpected result on the + CPU device. Modify the size of the work-group to ensure that the value + of the right-most dimension is a multiple of "32". + */ + x = sycl::fmax(x, dpct::permute_sub_group_by_xor( + item_ct1.get_sub_group(), x, mask)); + } + return x; +} #endif // GGML_SYCL_COMMON_HPP diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp index 3a87d3ef8e45cd..927819281fd0a0 100644 --- a/ggml/src/ggml-sycl/dmmv.cpp +++ b/ggml/src/ggml-sycl/dmmv.cpp @@ -76,7 +76,7 @@ static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -104,7 +104,7 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, nrows, item_ct1); }); @@ -227,7 +227,7 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -346,7 +346,7 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -499,7 +499,7 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -633,7 +633,7 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -748,7 +748,7 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -774,7 +774,7 @@ static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -795,7 +795,7 @@ static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -816,7 +816,7 @@ static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -837,7 +837,7 @@ static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -858,7 +858,7 @@ static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -873,10 +873,10 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, 32); + const sycl::range<3> block_dims(1, ny, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -889,10 +889,10 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, 32); + const sycl::range<3> block_dims(1, ny, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -905,10 +905,10 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, 32); + const sycl::range<3> block_dims(1, ny, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -918,10 +918,10 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, const int nrows, dpct::queue_ptr stream) { GGML_ASSERT(ncols % QK_K == 0); - const sycl::range<3> block_dims(1, 1, 32); + const sycl::range<3> block_dims(1, 1, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); }); } @@ -934,10 +934,10 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, 32); + const sycl::range<3> block_dims(1, ny, WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); }); } diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 9b751f3c67281f..3fbc4dd606bbed 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -37,7 +37,7 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_ // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -85,7 +85,7 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -133,7 +133,7 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -181,7 +181,7 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -229,7 +229,7 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -277,7 +277,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -325,7 +325,7 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -373,7 +373,7 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -421,7 +421,7 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -470,7 +470,7 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -495,7 +495,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -519,7 +519,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -543,7 +543,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -567,7 +567,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -591,7 +591,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -615,7 +615,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -639,7 +639,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -663,7 +663,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -687,7 +687,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -711,7 +711,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -734,7 +734,7 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -759,7 +759,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -784,7 +784,7 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -809,7 +809,7 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq3_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -833,7 +833,7 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq3_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -858,7 +858,7 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq1_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -879,7 +879,7 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq1_m_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -901,7 +901,7 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq4_nl_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -923,7 +923,7 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq4_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -936,7 +936,7 @@ void ggml_sycl_op_mul_mat_vec_q( const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, float *dst_dd_i, const int64_t row_low, const int64_t row_high, - const int64_t src1_ncols, const int64_t src1_padded_row_size, + const int64_t src1_ncols, const int64_t src1_padded_col_size, const dpct::queue_ptr &stream) { const int64_t ne10 = src1->ne[0]; @@ -948,77 +948,80 @@ void ggml_sycl_op_mul_mat_vec_q( int id; SYCL_CHECK( CHECK_TRY_ERROR(id = get_current_device_id())); - + const size_t q8_1_ts = sizeof(block_q8_1); + const size_t q8_1_bs = QK8_1; // the main device has a larger memory buffer to hold the results from all GPUs // nrows_dst == nrows of the matrix that the kernel writes into const int64_t nrows_dst = id == ctx.device ? ne00 : row_diff; - - switch (src0->type) { + for (int i = 0; i < src1_ncols; i++) + { + const size_t src1_ddq_i_offset = i * src1_padded_col_size * q8_1_ts / q8_1_bs; + const char* src1_ddq_i_bs = src1_ddq_i + src1_ddq_i_offset; + float* dst_dd_i_bs = dst_dd_i + i * dst->ne[0]; + switch (src0->type) { case GGML_TYPE_Q4_0: - mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q4_1: - mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q5_0: - mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q5_1: - mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q8_0: - mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q2_K: - mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q3_K: - mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q4_K: - mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q5_K: - mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_Q6_K: - mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ1_S: - mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ1_M: - mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ2_XXS: - mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ2_XS: - mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ2_S: - mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ3_XXS: - mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ3_S: - mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ4_NL: - mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; case GGML_TYPE_IQ4_XS: - mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); break; default: GGML_ASSERT(false); break; + } } - (void) src1; (void) dst; (void) src1_ddf_i; - (void) src1_ncols; - (void) src1_padded_row_size; } diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp new file mode 100644 index 00000000000000..a77f7852ccecd9 --- /dev/null +++ b/ggml/src/ggml-sycl/norm.cpp @@ -0,0 +1,370 @@ +#include "norm.hpp" + +static void norm_f32(const float* x, float* dst, const int ncols, const float eps, + const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + const int tid = item_ct1.get_local_id(2); + + const int nthreads = item_ct1.get_local_range(2); + const int nwarps = nthreads / WARP_SIZE; + assert(nwarps % WARP_SIZE == 0); + sycl::float2 mean_var = sycl::float2(0.f, 0.f); + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row * ncols + col]; + mean_var.x() += xi; + mean_var.y() += xi * xi; + } + + // sum up partial sums + mean_var = warp_reduce_sum(mean_var, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = mean_var; + } + /* + DPCT1118:0: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + item_ct1.barrier(sycl::access::fence_space::local_space); + mean_var = 0.f; + int nreduce = nwarps / WARP_SIZE; + for (size_t i = 0; i < nreduce; i += 1) + { + mean_var += s_sum[lane_id + i * WARP_SIZE]; + } + mean_var = warp_reduce_sum(mean_var, item_ct1); + } + + const float mean = mean_var.x() / ncols; + const float var = mean_var.y() / ncols - mean * mean; + const float inv_std = sycl::rsqrt(var + eps); + + for (int col = tid; col < ncols; col += block_size) { + dst[row * ncols + col] = (x[row * ncols + col] - mean) * inv_std; + } +} + +static void group_norm_f32(const float* x, float* dst, const int group_size, const int ne_elements, const float eps, + const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) { + int start = item_ct1.get_group(2) * group_size; + int end = start + group_size; + const int nthreads = item_ct1.get_local_range(2); + const int nwarps = nthreads / WARP_SIZE; + assert(nwarps % WARP_SIZE == 0); + start += item_ct1.get_local_id(2); + + if (end >= ne_elements) { + end = ne_elements; + } + + float tmp = 0.0f; // partial sum for thread in warp + + for (int j = start; j < end; j += block_size) { + tmp += x[j]; + } + + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:1: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:54: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + tmp = 0.f; + int nreduce = nwarps / WARP_SIZE; + for (size_t i = 0; i < nreduce; i += 1) + { + tmp += s_sum[lane_id + i * WARP_SIZE]; + } + tmp = warp_reduce_sum(tmp, item_ct1); + } + + float mean = tmp / group_size; + tmp = 0.0f; + + for (int j = start; j < end; j += block_size) { + float xi = x[j] - mean; + dst[j] = xi; + tmp += xi * xi; + } + + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:2: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:55: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp, item_ct1); + } + + float variance = tmp / group_size; + float scale = sycl::rsqrt(variance + eps); + for (int j = start; j < end; j += block_size) { + dst[j] *= scale; + } +} + +static void rms_norm_f32(const float* x, float* dst, const int ncols, const float eps, + const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + const int tid = item_ct1.get_local_id(2); + const int nthreads = item_ct1.get_local_range(2); + const int nwarps = nthreads / WARP_SIZE; + assert(nwarps % WARP_SIZE == 0); + float tmp = 0.0f; // partial sum for thread in warp + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row * ncols + col]; + tmp += xi * xi; + } + + // sum up partial sums + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:3: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + item_ct1.barrier(sycl::access::fence_space::local_space); + int nreduce = nwarps / WARP_SIZE; + tmp = 0.f; + for (size_t i = 0; i < nreduce; i += 1) + { + tmp += s_sum[lane_id + i * WARP_SIZE]; + } + tmp = warp_reduce_sum(tmp, item_ct1); + } + + const float mean = tmp / ncols; + const float scale = sycl::rsqrt(mean + eps); + + for (int col = tid; col < ncols; col += block_size) { + dst[row * ncols + col] = scale * x[row * ncols + col]; + } +} + +static void norm_f32_sycl(const float* x, float* dst, const int ncols, + const int nrows, const float eps, + queue_ptr stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + if (ncols < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler& cgh) { + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + norm_f32(x, dst, ncols, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); + } + else { + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:17: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor s_sum_acc_ct1( + sycl::range<1>(work_group_size / WARP_SIZE), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +static void group_norm_f32_sycl(const float* x, float* dst, + const int num_groups, const int group_size, + const int ne_elements, queue_ptr stream) { + static const float eps = 1e-6f; + if (group_size < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler& cgh) { + const float eps_ct4 = eps; + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + group_norm_f32( + x, dst, group_size, ne_elements, eps_ct4, item_ct1, + nullptr, WARP_SIZE); + }); + }); + } + else { + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:18: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), + cgh); + + const float eps_ct4 = eps; + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + group_norm_f32(x, dst, group_size, ne_elements, + eps_ct4, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, + const int nrows, const float eps, + queue_ptr stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); + if (ncols < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler& cgh) { + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + rms_norm_f32(x, dst, ncols, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); + } + else { + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:19: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), + cgh); + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(WARP_SIZE)]] { + rms_norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1, + ggml_tensor* dst, const float* src0_dd, + const float* src1_dd, float* dst_dd, + const queue_ptr& main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); + + (void)src1; + (void)dst; + (void)src1_dd; +} + +void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst, + const float* src0_dd, const float* src1_dd, + float* dst_dd, + const queue_ptr& main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + int num_groups = dst->op_params[0]; + int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); + group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream); + + (void)src1; + (void)dst; + (void)src1_dd; +} + +void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst, + const float* src0_dd, const float* src1_dd, + float* dst_dd, + const queue_ptr& main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); + + (void)src1; + (void)dst; + (void)src1_dd; +} diff --git a/ggml/src/ggml-sycl/norm.hpp b/ggml/src/ggml-sycl/norm.hpp new file mode 100644 index 00000000000000..a9ad9156fa33e0 --- /dev/null +++ b/ggml/src/ggml-sycl/norm.hpp @@ -0,0 +1,35 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_NORM_HPP +#define GGML_SYCL_NORM_HPP + +#include "common.hpp" + +void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1, + ggml_tensor* dst, const float* src0_dd, + const float* src1_dd, float* dst_dd, + const queue_ptr& main_stream); + +void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst, + const float* src0_dd, const float* src1_dd, + float* dst_dd, + const queue_ptr& main_stream); + +void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, + const ggml_tensor* src1, ggml_tensor* dst, + const float* src0_dd, const float* src1_dd, + float* dst_dd, + const queue_ptr& main_stream); + +#endif // GGML_SYCL_NORM_HPP diff --git a/ggml/src/ggml-sycl/presets.hpp b/ggml/src/ggml-sycl/presets.hpp index fe9d41770b76a4..c09c75dc7c73c1 100644 --- a/ggml/src/ggml-sycl/presets.hpp +++ b/ggml/src/ggml-sycl/presets.hpp @@ -16,7 +16,7 @@ #define GGML_SYCL_MAX_STREAMS 8 #define GGML_SYCL_MAX_BUFFERS 256 -#define WARP_SIZE 32 +#define WARP_SIZE GGML_SYCL_WARP_SIZE #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses #define SYCL_GELU_BLOCK_SIZE 256 From a9f3b102157ba992cfe058909b7f6e1906d2d647 Mon Sep 17 00:00:00 2001 From: luoyu-intel Date: Tue, 2 Jul 2024 04:50:07 +0000 Subject: [PATCH 12/27] [SYCL] Fix win build conflict of math library (#8230) * fix win build conflict of math library * fix the condition: !(win32 & SYCL) * revert warp_size=16 --- CMakePresets.json | 1 + ggml/src/CMakeLists.txt | 6 ++++-- 2 files changed, 5 insertions(+), 2 deletions(-) diff --git a/CMakePresets.json b/CMakePresets.json index d69bc03447ae9a..bdad38952d3cbe 100644 --- a/CMakePresets.json +++ b/CMakePresets.json @@ -19,6 +19,7 @@ "cacheVariables": { "CMAKE_EXPORT_COMPILE_COMMANDS": "ON", "CMAKE_CXX_COMPILER": "icx", + "CMAKE_C_COMPILER": "cl", "GGML_SYCL": "ON", "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.." } diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index a18198f1693e59..08b71d410d82e2 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -490,7 +490,7 @@ if (GGML_SYCL) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda") add_compile_definitions(GGML_SYCL_WARP_SIZE=32) else() - add_compile_definitions(GGML_SYCL_WARP_SIZE=16) + add_compile_definitions(GGML_SYCL_WARP_SIZE=32) endif() file(GLOB GGML_HEADERS_SYCL "ggml-sycl/*.hpp") @@ -1168,7 +1168,9 @@ target_link_libraries(ggml PRIVATE Threads::Threads ${GGML_EXTRA_LIBS}) find_library(MATH_LIBRARY m) if (MATH_LIBRARY) - target_link_libraries(ggml PRIVATE ${MATH_LIBRARY}) + if (NOT WIN32 OR NOT GGML_SYCL) + target_link_libraries(ggml PRIVATE ${MATH_LIBRARY}) + endif() endif() if (BUILD_SHARED_LIBS) From 0e0590adab9f367b15ae2bf090a6d24f9df47ff1 Mon Sep 17 00:00:00 2001 From: slaren Date: Tue, 2 Jul 2024 08:39:38 +0200 Subject: [PATCH 13/27] cuda : update supports_op for matrix multiplication (#8245) --- ggml/src/ggml-cuda.cu | 47 ++++++++++++++++++++++++-------------- tests/test-backend-ops.cpp | 1 + 2 files changed, 31 insertions(+), 17 deletions(-) diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 649ef5a0819107..1c9ccc8a15e54e 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -2711,27 +2711,40 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: { - struct ggml_tensor * a; - struct ggml_tensor * b; + struct ggml_tensor * a = op->src[0]; if (op->op == GGML_OP_MUL_MAT) { - a = op->src[0]; - b = op->src[1]; - } else { - a = op->src[2]; - b = op->src[1]; - } - if (a->ne[3] != b->ne[3]) { - return false; - } - ggml_type a_type = a->type; - if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS || - a_type == GGML_TYPE_IQ1_S || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S || - a_type == GGML_TYPE_IQ1_M || a_type == GGML_TYPE_IQ2_S || a_type == GGML_TYPE_IQ4_XS) { - if (b->ne[1] == 1 && ggml_nrows(b) > 1) { + struct ggml_tensor * b = op->src[1]; + if (a->ne[3] != b->ne[3]) { return false; } } - return true; + switch (a->type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_Q8_K: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + return true; + default: + return false; + } } break; case GGML_OP_GET_ROWS: { diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index f74c0db475e2e2..2bb71ac03817fd 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -2052,6 +2052,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M, GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS, + GGML_TYPE_BF16, }; // unary ops From 023b8807e10bc3ade24a255f01c1ad2a01bb4228 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Tue, 2 Jul 2024 08:40:49 +0200 Subject: [PATCH 14/27] convert-hf : print output file name when completed (#8181) * convert-hf : print output file name when completed This commit adds the output file name to the log message when the conversion is completed. The motivation for this change is that when `--outfile` option is not specified it migth not be obvious where the output file is written. With this change the output of running the script will be something like the following: ```console INFO:hf-to-gguf:Model successfully exported to models/gemma-2-9b-it.gguf. ``` Signed-off-by: Daniel Bevenius * squash! convert-hf : print output file name when completed Updates the output of to support printing the directory if the output is split into multiple files. Also the output file name is now retrieved from the model_instance object. Signed-off-by: Daniel Bevenius * squash! convert-hf : print output file name when completed Use parent attribute of Path object and string interpolation. Signed-off-by: Daniel Bevenius * squash! convert-hf : print output file name when completed Use os.sep instead of hardcoding the path separator. Signed-off-by: Daniel Bevenius --------- Signed-off-by: Daniel Bevenius --- convert-hf-to-gguf.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 6833e943765f79..05fd70171de384 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -3120,7 +3120,8 @@ def main() -> None: "auto": gguf.LlamaFileType.GUESSED, } - if args.use_temp_file and (args.split_max_tensors > 0 or args.split_max_size != "0"): + is_split = args.split_max_tensors > 0 or args.split_max_size != "0" + if args.use_temp_file and is_split: logger.error("Error: Cannot use temp file when splitting") sys.exit(1) @@ -3157,11 +3158,12 @@ def main() -> None: if args.vocab_only: logger.info("Exporting model vocab...") model_instance.write_vocab() - logger.info("Model vocab successfully exported.") + logger.info(f"Model vocab successfully exported to {model_instance.fname_out}") else: logger.info("Exporting model...") model_instance.write() - logger.info("Model successfully exported.") + out_path = f"{model_instance.fname_out.parent}{os.sep}" if is_split else model_instance.fname_out + logger.info(f"Model successfully exported to {out_path}") if __name__ == '__main__': From 968967376dc2c018d29f897c4883d335bbf384fb Mon Sep 17 00:00:00 2001 From: Faisal Zaghloul Date: Tue, 2 Jul 2024 10:36:00 -0400 Subject: [PATCH 15/27] Add `JAIS` model(s) (#8118) * Add `JAIS` model(s) * cleanup * address review comments * remove hack * un-hardcode max-alibi-bias * minor tweaks --------- Co-authored-by: fmz --- convert-hf-to-gguf-update.py | 1 + convert-hf-to-gguf.py | 93 ++++++++++++++++++ gguf-py/gguf/constants.py | 14 +++ gguf-py/gguf/tensor_mapping.py | 19 ++-- include/llama.h | 1 + src/llama.cpp | 169 +++++++++++++++++++++++++++++++++ 6 files changed, 288 insertions(+), 9 deletions(-) diff --git a/convert-hf-to-gguf-update.py b/convert-hf-to-gguf-update.py index 2758214fa8730b..944e9d15abdcc4 100755 --- a/convert-hf-to-gguf-update.py +++ b/convert-hf-to-gguf-update.py @@ -86,6 +86,7 @@ class TOKENIZER_TYPE(IntEnum): {"name": "poro-chat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", }, {"name": "jina-v2-code", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", }, {"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B + {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, ] diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 05fd70171de384..6add27cbb11a6d 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -490,6 +490,9 @@ def get_vocab_base_pre(self, tokenizer) -> str: if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee": # ref: https://huggingface.co/LumiOpen/Viking-7B res = "viking" + if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901": + # ref: https://huggingface.co/core42/jais-13b + res = "jais" if res is None: logger.warning("\n") @@ -2965,6 +2968,96 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter return [(self.map_tensor_name(name), data_torch)] +@Model.register("JAISLMHeadModel") +class JaisModel(Model): + model_arch = gguf.MODEL_ARCH.JAIS + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # SwigLU activation + assert self.hparams["activation_function"] == "swiglu" + # ALiBi position embedding + assert self.hparams["position_embedding_type"] == "alibi" + + # Embeddings scale + self.embeddings_scale = 1.0 + # note: For some JAIS flavors, output is tied to (same as) wte in original model + self.output_is_wte = False + if 'mup_embeddings_scale' in self.hparams: + self.output_is_wte = True # Hack (?) + self.embeddings_scale = self.hparams['mup_embeddings_scale'] + elif 'embeddings_scale' in self.hparams: + self.embeddings_scale = self.hparams['embeddings_scale'] + else: + assert False + + self.width_scale = 1.0 + if 'mup_output_alpha' in self.hparams: + assert 'mup_width_scale' in self.hparams + self.width_scale = self.hparams['mup_output_alpha'] * self.hparams['mup_width_scale'] + elif 'width_scale' in self.hparams: + self.width_scale = self.hparams['width_scale'] + else: + assert False + + self.max_alibi_bias = 8.0 + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_block_count(self.hparams["n_layer"]) + self.gguf_writer.add_context_length(self.hparams["n_positions"]) + self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) + self.gguf_writer.add_feed_forward_length(self.hparams["n_inner"]) + self.gguf_writer.add_head_count(self.hparams["n_head"]) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + + tensors: list[tuple[str, Tensor]] = [] + + # we don't need these + if name.endswith((".attn.bias")): + return tensors + + if name.endswith(("relative_pe.slopes")): + # Calculate max ALiBi bias (this is the inverse of the ALiBi calculation) + # Some other models has max_alibi_bias spelled out explicitly in the hyperparams, + # but Jais's PyTorch model simply precalculates the slope values and places them + # in relative_pes.slopes + n_head_closest_log2 = 2 ** math.floor(math.log2(self.hparams["n_head"])) + first_val = float(data_torch._data[0]) + self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2) + + return tensors + + if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_fc2.weight")): + data_torch = data_torch.transpose(1, 0) + + new_name = self.map_tensor_name(name) + + if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD): + tensors.append((new_name, data_torch * self.embeddings_scale)) + if self.output_is_wte: + tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale)) + elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT): + assert not self.output_is_wte + tensors.append((new_name, data_torch * self.width_scale)) + else: + tensors.append((new_name, data_torch)) + + return tensors + + def write_tensors(self): + super().write_tensors() + self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias) + + ###### CONVERSION LOGIC ###### diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index e87c58266158a1..419f10cee74a7a 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -164,6 +164,7 @@ class MODEL_ARCH(IntEnum): DEEPSEEK2 = auto() BITNET = auto() T5 = auto() + JAIS = auto() class MODEL_TENSOR(IntEnum): @@ -288,6 +289,7 @@ class MODEL_TENSOR(IntEnum): MODEL_ARCH.DEEPSEEK2: "deepseek2", MODEL_ARCH.BITNET: "bitnet", MODEL_ARCH.T5: "t5", + MODEL_ARCH.JAIS: "jais", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -954,6 +956,18 @@ class MODEL_TENSOR(IntEnum): MODEL_TENSOR.ENC_FFN_UP, MODEL_TENSOR.ENC_OUTPUT_NORM, ], + MODEL_ARCH.JAIS: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_UP, + ], # TODO } diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 0bed439397bcdb..20e28423b9a99c 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -10,7 +10,7 @@ class TensorNameMap: # Token embeddings MODEL_TENSOR.TOKEN_EMBD: ( "gpt_neox.embed_in", # gptneox - "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx + "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais "transformer.word_embeddings", # falcon "word_embeddings", # bloom "model.embed_tokens", # llama-hf @@ -49,7 +49,7 @@ class TensorNameMap: # Output MODEL_TENSOR.OUTPUT: ( "embed_out", # gptneox - "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx + "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais "output", # llama-pth bloom internlm2 "word_embeddings_for_head", # persimmon "lm_head.linear", # phi2 @@ -58,7 +58,7 @@ class TensorNameMap: # Output norm MODEL_TENSOR.OUTPUT_NORM: ( "gpt_neox.final_layer_norm", # gptneox - "transformer.ln_f", # gpt2 gpt-j falcon + "transformer.ln_f", # gpt2 gpt-j falcon jais "model.norm", # llama-hf baichuan internlm2 "norm", # llama-pth "transformer.norm_f", # mpt dbrx @@ -81,7 +81,7 @@ class TensorNameMap: # Attention norm MODEL_TENSOR.ATTN_NORM: ( "gpt_neox.layers.{bid}.input_layernorm", # gptneox - "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen + "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen jais "transformer.blocks.{bid}.norm_1", # mpt "transformer.h.{bid}.input_layernorm", # falcon7b "h.{bid}.input_layernorm", # bloom @@ -109,7 +109,7 @@ class TensorNameMap: # Attention query-key-value MODEL_TENSOR.ATTN_QKV: ( "gpt_neox.layers.{bid}.attention.query_key_value", # gptneox - "transformer.h.{bid}.attn.c_attn", # gpt2 qwen + "transformer.h.{bid}.attn.c_attn", # gpt2 qwen jais "transformer.blocks.{bid}.attn.Wqkv", # mpt "transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv", # dbrx "transformer.h.{bid}.self_attention.query_key_value", # falcon @@ -160,7 +160,7 @@ class TensorNameMap: # Attention output MODEL_TENSOR.ATTN_OUT: ( "gpt_neox.layers.{bid}.attention.dense", # gptneox - "transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen + "transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen jais "transformer.blocks.{bid}.attn.out_proj", # mpt "transformer.h.{bid}.self_attention.dense", # falcon "h.{bid}.self_attention.dense", # bloom @@ -202,7 +202,7 @@ class TensorNameMap: # Feed-forward norm MODEL_TENSOR.FFN_NORM: ( "gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox - "transformer.h.{bid}.ln_2", # gpt2 refact qwen + "transformer.h.{bid}.ln_2", # gpt2 refact qwen jais "h.{bid}.post_attention_layernorm", # bloom "transformer.blocks.{bid}.norm_2", # mpt "model.layers.{bid}.post_attention_layernorm", # llama-hf @@ -239,7 +239,7 @@ class TensorNameMap: # Feed-forward up MODEL_TENSOR.FFN_UP: ( "gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox - "transformer.h.{bid}.mlp.c_fc", # gpt2 + "transformer.h.{bid}.mlp.c_fc", # gpt2 jais "transformer.blocks.{bid}.ffn.up_proj", # mpt "transformer.h.{bid}.mlp.dense_h_to_4h", # falcon "h.{bid}.mlp.dense_h_to_4h", # bloom @@ -285,6 +285,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.gate_proj", # llama-hf refact "layers.{bid}.feed_forward.w1", # llama-pth "transformer.h.{bid}.mlp.w2", # qwen + "transformer.h.{bid}.mlp.c_fc2", # jais "model.layers.layers.{bid}.mlp.gate_proj", # plamo "model.layers.{bid}.feed_forward.w1", # internlm2 "encoder.layers.{bid}.mlp.fc12", # nomic-bert @@ -308,7 +309,7 @@ class TensorNameMap: # Feed-forward down MODEL_TENSOR.FFN_DOWN: ( "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox - "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen + "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen jais "transformer.blocks.{bid}.ffn.down_proj", # mpt "transformer.h.{bid}.mlp.dense_4h_to_h", # falcon "h.{bid}.mlp.dense_4h_to_h", # bloom diff --git a/include/llama.h b/include/llama.h index cafeafb85dbc7f..c5b61829204285 100644 --- a/include/llama.h +++ b/include/llama.h @@ -89,6 +89,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, LLAMA_VOCAB_PRE_TYPE_PORO = 15, LLAMA_VOCAB_PRE_TYPE_VIKING = 16, + LLAMA_VOCAB_PRE_TYPE_JAIS = 17, }; // note: these values should be synchronized with ggml_rope diff --git a/src/llama.cpp b/src/llama.cpp index eea532f6ac2ff3..73f52435a503ef 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -228,6 +228,7 @@ enum llm_arch { LLM_ARCH_DEEPSEEK2, LLM_ARCH_BITNET, LLM_ARCH_T5, + LLM_ARCH_JAIS, LLM_ARCH_UNKNOWN, }; @@ -269,6 +270,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_DEEPSEEK2, "deepseek2" }, { LLM_ARCH_BITNET, "bitnet" }, { LLM_ARCH_T5, "t5" }, + { LLM_ARCH_JAIS, "jais" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -1236,6 +1238,21 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_JAIS, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -2035,6 +2052,7 @@ enum e_model { MODEL_410M, MODEL_0_5B, MODEL_1B, + MODEL_1_3B, MODEL_1_4B, MODEL_2B, MODEL_2_8B, @@ -4276,6 +4294,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_410M: return "410M"; case MODEL_0_5B: return "0.5B"; case MODEL_1B: return "1B"; + case MODEL_1_3B: return "1.3B"; case MODEL_1_4B: return "1.4B"; case MODEL_2B: return "2B"; case MODEL_2_8B: return "2.8B"; @@ -4898,6 +4917,18 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_JAIS: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1_3B; break; + case 40: model.type = e_model::MODEL_13B; break; + /* TODO: add variants */ + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -5129,6 +5160,9 @@ static void llm_load_vocab( } else if ( tokenizer_pre == "viking") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_VIKING; + } else if ( + tokenizer_pre == "jais") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -6962,6 +6996,44 @@ static bool llm_load_tensors( layer.ffn_up_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "scale", i), {1}); } } break; + case LLM_ARCH_JAIS: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // Output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); + layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}); + + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_gate_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}); + + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -12354,6 +12426,97 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_jais() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/float(n_embd_head), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = llm_build_norm(ctx0, inpL, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } }; static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { @@ -12585,6 +12748,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_bitnet(); } break; + case LLM_ARCH_JAIS: + { + result = llm.build_jais(); + } break; default: GGML_ASSERT(false); } @@ -13947,6 +14114,7 @@ struct llm_tokenizer_bpe { break; case LLAMA_VOCAB_PRE_TYPE_GPT2: case LLAMA_VOCAB_PRE_TYPE_OLMO: + case LLAMA_VOCAB_PRE_TYPE_JAIS: regex_exprs = { "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", }; @@ -17826,6 +17994,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_MAMBA: case LLM_ARCH_JINA_BERT_V2: case LLM_ARCH_T5: + case LLM_ARCH_JAIS: return LLAMA_ROPE_TYPE_NONE; // use what we call a normal RoPE, operating on pairs of consecutive head values From 07a3fc0608a68c0c93a5fbfa9c58f4c9ec64cb81 Mon Sep 17 00:00:00 2001 From: Clint Herron Date: Tue, 2 Jul 2024 12:18:10 -0400 Subject: [PATCH 16/27] Removes multiple newlines at the end of files that is breaking the editorconfig step of CI. (#8258) --- .github/ISSUE_TEMPLATE/config.yml | 2 -- common/common.h | 1 - examples/embedding/README.md | 1 - examples/infill/infill.cpp | 1 - examples/lookup/README.md | 1 - examples/main-cmake-pkg/.gitignore | 1 - examples/main-cmake-pkg/CMakeLists.txt | 1 - examples/server-embd.py | 1 - examples/server/tests/features/passkey.feature | 1 - examples/server/themes/buttons-top/index.html | 1 - examples/server/themes/wild/index.html | 1 - examples/sycl/run-llama2.sh | 1 - examples/sycl/win-build-sycl.bat | 1 - examples/sycl/win-run-llama2.bat | 2 -- ggml/include/ggml-metal.h | 1 - ggml/src/ggml-cuda/cpy.cu | 1 - ggml/src/ggml-metal.metal | 1 - ggml/src/ggml-quants.h | 1 - ggml/src/ggml-vulkan-shaders.hpp | 1 - scripts/pod-llama.sh | 1 - src/unicode-data.cpp | 1 - tests/test-rope.cpp | 1 - 22 files changed, 24 deletions(-) diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml index c88134dbb644a5..eb8c4b472df4c4 100644 --- a/.github/ISSUE_TEMPLATE/config.yml +++ b/.github/ISSUE_TEMPLATE/config.yml @@ -9,5 +9,3 @@ contact_links: - name: Want to contribute? url: https://github.com/ggerganov/llama.cpp/wiki/contribute about: Head to the contribution guide page of the wiki for areas you can help with - - diff --git a/common/common.h b/common/common.h index 627b7ed854757e..65c0ef81adf7ca 100644 --- a/common/common.h +++ b/common/common.h @@ -459,4 +459,3 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha void yaml_dump_non_result_info( FILE * stream, const gpt_params & params, const llama_context * lctx, const std::string & timestamp, const std::vector & prompt_tokens, const char * model_desc); - diff --git a/examples/embedding/README.md b/examples/embedding/README.md index 86df1895878465..e3705b45476772 100644 --- a/examples/embedding/README.md +++ b/examples/embedding/README.md @@ -58,4 +58,3 @@ The above command will output space-separated float values. ```powershell embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null ``` - diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index ca71dd687f30e0..0e682154d5f6be 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -659,4 +659,3 @@ int main(int argc, char ** argv) { return 0; } - diff --git a/examples/lookup/README.md b/examples/lookup/README.md index 5bfb0de9360414..71c345c037a2fb 100644 --- a/examples/lookup/README.md +++ b/examples/lookup/README.md @@ -10,4 +10,3 @@ More info: https://github.com/ggerganov/llama.cpp/pull/4484 https://github.com/ggerganov/llama.cpp/issues/4226 - diff --git a/examples/main-cmake-pkg/.gitignore b/examples/main-cmake-pkg/.gitignore index e32c11c7f4653c..67c01d64cb7ab2 100644 --- a/examples/main-cmake-pkg/.gitignore +++ b/examples/main-cmake-pkg/.gitignore @@ -48,4 +48,3 @@ build*/ out/ tmp/ - diff --git a/examples/main-cmake-pkg/CMakeLists.txt b/examples/main-cmake-pkg/CMakeLists.txt index a97ded3653f0cb..3b38db292320f8 100644 --- a/examples/main-cmake-pkg/CMakeLists.txt +++ b/examples/main-cmake-pkg/CMakeLists.txt @@ -30,4 +30,3 @@ target_include_directories(${TARGET} PRIVATE ${_common_path}) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) - diff --git a/examples/server-embd.py b/examples/server-embd.py index 118e042716c020..a9a36a44ccac5d 100644 --- a/examples/server-embd.py +++ b/examples/server-embd.py @@ -31,4 +31,3 @@ async def main(): embedding2 = np.array(result[j]) similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2)) print(f"Similarity between {i} and {j}: {similarity:.2f}") - diff --git a/examples/server/tests/features/passkey.feature b/examples/server/tests/features/passkey.feature index 1bde7aab8bab0d..6a5a84e6a19417 100644 --- a/examples/server/tests/features/passkey.feature +++ b/examples/server/tests/features/passkey.feature @@ -52,4 +52,3 @@ Feature: Passkey / Self-extend with context shift #| TheBloke/Llama-2-7B-GGUF | llama-2-7b.Q2_K.gguf | 4096 | 3 | 16384 | 512 | 4 | 512 | 500 | 300 | 1234 | 5 | 1234 | #| TheBloke/Mixtral-8x7B-v0.1-GGUF | mixtral-8x7b-v0.1.Q2_K.gguf | 32768 | 2 | 16384 | 512 | 4 | 512 | 500 | 100 | 0987 | 5 | 0 # 987 | - diff --git a/examples/server/themes/buttons-top/index.html b/examples/server/themes/buttons-top/index.html index 6af30d307a4b5b..8334bcde5049cc 100644 --- a/examples/server/themes/buttons-top/index.html +++ b/examples/server/themes/buttons-top/index.html @@ -1054,4 +1054,3 @@

llama.cpp

- diff --git a/examples/server/themes/wild/index.html b/examples/server/themes/wild/index.html index 772e716cdb2e07..8361c577494d72 100644 --- a/examples/server/themes/wild/index.html +++ b/examples/server/themes/wild/index.html @@ -1058,4 +1058,3 @@ - diff --git a/examples/sycl/run-llama2.sh b/examples/sycl/run-llama2.sh index da0e4aaba688c8..111366fb036a51 100755 --- a/examples/sycl/run-llama2.sh +++ b/examples/sycl/run-llama2.sh @@ -34,4 +34,3 @@ fi #use multiple GPUs with same max compute units #ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 - diff --git a/examples/sycl/win-build-sycl.bat b/examples/sycl/win-build-sycl.bat index cdae5a52855a21..17dd1ff5c169ec 100644 --- a/examples/sycl/win-build-sycl.bat +++ b/examples/sycl/win-build-sycl.bat @@ -31,4 +31,3 @@ exit /B 0 :ERROR echo comomand error: %errorlevel% exit /B %errorlevel% - diff --git a/examples/sycl/win-run-llama2.bat b/examples/sycl/win-run-llama2.bat index 1d4d7d2cdcb6fa..f0385cdf0783e6 100644 --- a/examples/sycl/win-run-llama2.bat +++ b/examples/sycl/win-run-llama2.bat @@ -7,5 +7,3 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:" .\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0 - - diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h index e7543ae795d284..6c3226c37e0ef4 100644 --- a/ggml/include/ggml-metal.h +++ b/ggml/include/ggml-metal.h @@ -63,4 +63,3 @@ GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend); #ifdef __cplusplus } #endif - diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu index 12d741f017d3b4..3db57034b488d9 100644 --- a/ggml/src/ggml-cuda/cpy.cu +++ b/ggml/src/ggml-cuda/cpy.cu @@ -487,4 +487,3 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) { GGML_ASSERT(false); } } - diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal index e2796fd6012811..c3503479b35bac 100644 --- a/ggml/src/ggml-metal.metal +++ b/ggml/src/ggml-metal.metal @@ -6537,4 +6537,3 @@ template [[host_name("kernel_mul_mv_id_iq3_s_f32")]] kernel kernel_mul_mv_id_t template [[host_name("kernel_mul_mv_id_iq2_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; - diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h index 4d436a8f06b3e5..30983b8728fa24 100644 --- a/ggml/src/ggml-quants.h +++ b/ggml/src/ggml-quants.h @@ -130,4 +130,3 @@ void iq3xs_free_impl(int grid_size); #ifdef __cplusplus } #endif - diff --git a/ggml/src/ggml-vulkan-shaders.hpp b/ggml/src/ggml-vulkan-shaders.hpp index 01ff66f71fcf04..f0c4c6baf592b2 100644 --- a/ggml/src/ggml-vulkan-shaders.hpp +++ b/ggml/src/ggml-vulkan-shaders.hpp @@ -144954,4 +144954,3 @@ unsigned char sum_rows_f32_data[] = { }; const uint64_t sum_rows_f32_len = 2112; - diff --git a/scripts/pod-llama.sh b/scripts/pod-llama.sh index 586d6ea18af01e..0d6d4032d8a9e5 100644 --- a/scripts/pod-llama.sh +++ b/scripts/pod-llama.sh @@ -210,4 +210,3 @@ fi # more benches #GGML_CUDA=1 make -j && ./llama-batched-bench ./models/codellama-7b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1 #GGML_CUDA=1 make -j && ./llama-batched-bench ./models/codellama-13b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1 - diff --git a/src/unicode-data.cpp b/src/unicode-data.cpp index 4a939898b367f5..02bdf782380fe7 100644 --- a/src/unicode-data.cpp +++ b/src/unicode-data.cpp @@ -7030,4 +7030,3 @@ const std::vector unicode_ranges_nfd = { // start, last, nfd {0x02FA1C, 0x02FA1C, 0x009F3B}, {0x02FA1D, 0x02FA1D, 0x02A600}, }; - diff --git a/tests/test-rope.cpp b/tests/test-rope.cpp index f0895ffaad6a10..8159e276af617b 100644 --- a/tests/test-rope.cpp +++ b/tests/test-rope.cpp @@ -218,4 +218,3 @@ int main(int /*argc*/, const char ** /*argv*/) { return 0; } - From 3e2618bc7bf9e9fbf58c32cc3c8dd7d5df1de27e Mon Sep 17 00:00:00 2001 From: Clint Herron Date: Tue, 2 Jul 2024 13:19:56 -0400 Subject: [PATCH 17/27] Adding step to `clean` target to remove legacy binary names to reduce upgrade / migration confusion arising from #7809. (#8257) --- Makefile | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/Makefile b/Makefile index 8ae4f1dc4ede3a..2730d8b602093f 100644 --- a/Makefile +++ b/Makefile @@ -62,6 +62,11 @@ TEST_TARGETS = \ tests/test-tokenizer-1-bpe \ tests/test-tokenizer-1-spm +# Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned +LEGACY_TARGETS = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ + simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \ + retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm + # Deprecation aliases ifdef LLAMA_CUBLAS $(error LLAMA_CUBLAS is removed. Use GGML_CUDA instead.) @@ -1086,6 +1091,7 @@ clean: rm -vrf ggml/src/ggml-cuda/template-instances/*.o rm -rvf $(BUILD_TARGETS) rm -rvf $(TEST_TARGETS) + rm -rvf $(LEGACY_TARGETS) find examples pocs -type f -name "*.o" -delete # From a27152b602b369e76f85b7cb7b872a321b7218f7 Mon Sep 17 00:00:00 2001 From: MistApproach <98988043+MistApproach@users.noreply.github.com> Date: Tue, 2 Jul 2024 22:56:46 +0200 Subject: [PATCH 18/27] fix: add missing short command line argument -mli for multiline-input (#8261) --- common/common.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/common/common.cpp b/common/common.cpp index 5a0d0ee0381239..2c05a4d4a17c17 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -757,7 +757,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.cache_type_v = argv[++i]; return true; } - if (arg == "--multiline-input") { + if (arg == "-mli" || arg == "--multiline-input") { params.multiline_input = true; return true; } From fadde6713506d9e6c124f5680ab8c7abebe31837 Mon Sep 17 00:00:00 2001 From: AidanBeltonS Date: Wed, 3 Jul 2024 02:55:34 +0100 Subject: [PATCH 19/27] Dequant improvements rebase (#8255) * Single load for half2 * Store scales in local mem * Vec load quantized values --- ggml/src/ggml-sycl/common.hpp | 6 ++++++ ggml/src/ggml-sycl/convert.cpp | 7 +++++-- ggml/src/ggml-sycl/dequantize.hpp | 30 +++++++++++++++++++----------- 3 files changed, 30 insertions(+), 13 deletions(-) diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index dfd4a7c2c606bb..476d847ca575e3 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -351,4 +351,10 @@ static __dpct_inline__ float warp_reduce_max(float x, return x; } +// Helper for vec loading aligned data +template +inline sycl::vec vec_aligned_load(const Tp* aligned_ptr) { + return *reinterpret_cast*>(aligned_ptr); +} + #endif // GGML_SYCL_COMMON_HPP diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp index ce9de2b42b7220..a15271b516fa80 100644 --- a/ggml/src/ggml-sycl/convert.cpp +++ b/ggml/src/ggml-sycl/convert.cpp @@ -152,12 +152,15 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor scale_local_acc(sycl::range<1>(12), cgh); + cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_K(vx, y, item_ct1); + dequantize_block_q4_K(vx, y, scale_local_acc.get_pointer(), item_ct1); }); + }); } } diff --git a/ggml/src/ggml-sycl/dequantize.hpp b/ggml/src/ggml-sycl/dequantize.hpp index b6080d83a33eb8..ed8ad098bcb2fd 100644 --- a/ggml/src/ggml-sycl/dequantize.hpp +++ b/ggml/src/ggml-sycl/dequantize.hpp @@ -293,7 +293,8 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri #if QK_K == 256 static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { if (j < 4) { - d = q[j] & 63; m = q[j + 4] & 63; + d = q[j] & 63; + m = q[j + 4] & 63; } else { d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); @@ -303,7 +304,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8 template static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1) { + uint8_t* scales_local, const sycl::nd_item<3> &item_ct1) { const block_q4_K * x = (const block_q4_K *) vx; const int i = item_ct1.get_group(2); @@ -318,19 +319,26 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri dst_t * y = yy + i*QK_K + 64*il + n*ir; - const float dall = x[i].dm[0]; - const float dmin = x[i].dm[1]; + const sycl::half2 dm = x[i].dm; + const float dall = dm[0]; + const float dmin = dm[1]; - const uint8_t * q = x[i].qs + 32*il + n*ir; + if (tid < 12) + scales_local[tid] = x[i].scales[tid]; + item_ct1.barrier(sycl::access::fence_space::local_space); uint8_t sc, m; - get_scale_min_k4(is + 0, x[i].scales, sc, m); - const float d1 = dall * sc; const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[i].scales, sc, m); - const float d2 = dall * sc; const float m2 = dmin * m; + get_scale_min_k4(is + 0, scales_local, sc, m); + const float d1 = dall * sc; + const float m1 = dmin * m; + get_scale_min_k4(is + 1, scales_local, sc, m); + const float d2 = dall * sc; + const float m2 = dmin * m; + + sycl::vec q_vec = vec_aligned_load(x[i].qs + 32*il + n*ir); for (int l = 0; l < n; ++l) { - y[l + 0] = d1 * (q[l] & 0xF) - m1; - y[l +32] = d2 * (q[l] >> 4) - m2; + y[l + 0] = d1 * (q_vec[l] & 0xF) - m1; + y[l +32] = d2 * (q_vec[l] >> 4) - m2; } #else const int tid = item_ct1.get_local_id(2); From f8d6a23804f3798ff2869da68c1223b618df09ec Mon Sep 17 00:00:00 2001 From: Judd Date: Wed, 3 Jul 2024 20:40:16 +0800 Subject: [PATCH 20/27] fix typo (#8267) Co-authored-by: Judd --- ggml/src/ggml.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index f5502afbe98b36..bc91ac3a726abd 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -5312,7 +5312,7 @@ void ggml_mul_mat_set_prec( as -> [cols, rows, n_expert] ids -> [n_experts_used, n_tokens] (i32) b -> [cols, n_expert_used, n_tokens] - c -> [cols, n_expert_used, n_tokens] + c -> [rows, n_expert_used, n_tokens] in b, n_experts_used can be broadcasted to match the n_expert_used of ids From 916248af1f3c16abd7408de848e025da095c621c Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Wed, 3 Jul 2024 16:01:54 +0200 Subject: [PATCH 21/27] fix phi 3 conversion (#8262) --- convert-hf-to-gguf.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 6add27cbb11a6d..d01aed224d6009 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -1942,7 +1942,7 @@ def set_gguf_parameters(self): if len(rope_scaling_type) == 0: raise KeyError('Missing the required key rope_scaling.type') - if rope_scaling_type == 'su': + if rope_scaling_type == 'su' or rope_scaling_type == 'longrope': attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0 elif rope_scaling_type == 'yarn': attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0 From 5f2d4e60e202aabee10051e6615bb821e51787be Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 3 Jul 2024 19:33:31 +0200 Subject: [PATCH 22/27] ppl : fix n_seq_max for perplexity (#8277) * ppl : fix n_seq_max for perplexity * use 1 seq for kl_divergence --- examples/perplexity/perplexity.cpp | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index efde8dfdff47b7..dbe445391736cf 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1991,6 +1991,12 @@ int main(int argc, char ** argv) { params.n_batch = std::min(params.n_batch, n_kv); } else { params.n_batch = std::min(params.n_batch, params.n_ctx); + if (params.kl_divergence) { + params.n_parallel = 1; + } else { + // ensure there's at least enough seq_ids for HellaSwag + params.n_parallel = std::max(4, params.n_parallel); + } } if (params.ppl_stride > 0) { @@ -2015,9 +2021,6 @@ int main(int argc, char ** argv) { llama_model * model; llama_context * ctx; - // ensure there's at least enough seq_ids for HellaSwag - params.n_parallel = std::max(4, params.n_parallel); - // load the model and apply lora adapter, if any std::tie(model, ctx) = llama_init_from_gpt_params(params); if (model == NULL) { From d23287f122c34ebef368742116d53a0ccb2041ee Mon Sep 17 00:00:00 2001 From: Daniele <57776841+daniandtheweb@users.noreply.github.com> Date: Wed, 3 Jul 2024 23:02:58 +0000 Subject: [PATCH 23/27] Define and optimize RDNA1 (#8085) --- ggml/src/ggml-cuda/common.cuh | 4 ++++ ggml/src/ggml-cuda/mmq.cuh | 10 +++++++--- 2 files changed, 11 insertions(+), 3 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 472f4ace1c2ad2..4ff06b8719d378 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -227,6 +227,10 @@ typedef float2 dfloat2; #define RDNA2 #endif +#if defined(__gfx1010__) || defined(__gfx1012__) +#define RDNA1 +#endif + #ifndef __has_builtin #define __has_builtin(x) 0 #endif diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index 1396e7a753ac34..deaed066f7c908 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -60,12 +60,16 @@ static constexpr __device__ int get_mmq_x_max_device() { } static constexpr int get_mmq_y_host(const int cc) { - return int8_mma_available(cc) || cc >= CC_VOLTA ? 128 : 64; + return cc >= CC_OFFSET_AMD ? (cc == CC_RDNA1 ? 64 : 128) : (cc >= CC_VOLTA ? 128 : 64); } static constexpr __device__ int get_mmq_y_device() { #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA1) + return 64; +#else return 128; +#endif // defined RDNA1 #else #if __CUDA_ARCH__ >= CC_VOLTA return 128; @@ -2259,9 +2263,9 @@ static __device__ void mul_mat_q_process_tile( template #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) +#if defined(RDNA3) || defined(RDNA2) || defined(RDNA1) __launch_bounds__(WARP_SIZE*nwarps, 2) -#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(RDNA3) || defined(RDNA2) || defined(RDNA1) #else #if __CUDA_ARCH__ >= CC_VOLTA __launch_bounds__(WARP_SIZE*nwarps, 1) From f619024764e72261f14d7c31d892b8fb976603b4 Mon Sep 17 00:00:00 2001 From: AidanBeltonS Date: Thu, 4 Jul 2024 02:07:19 +0100 Subject: [PATCH 24/27] [SYCL] Remove unneeded semicolons (#8280) --- ggml/src/ggml-sycl/dpct/helper.hpp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp index 1ff297218c6853..5e98660dc888d2 100644 --- a/ggml/src/ggml-sycl/dpct/helper.hpp +++ b/ggml/src/ggml-sycl/dpct/helper.hpp @@ -255,7 +255,7 @@ namespace dpct void set_pitch(size_t pitch) { _pitch = pitch; } size_t get_x() { return _x; } - void set_x(size_t x) { _x = x; }; + void set_x(size_t x) { _x = x; } size_t get_y() { return _y; } void set_y(size_t y) { _y = y; } @@ -1056,7 +1056,7 @@ namespace dpct #error "Only support Windows and Linux." #endif next_free = mapped_address_space; - }; + } public: using buffer_id_t = int; @@ -1077,7 +1077,7 @@ namespace dpct #else #error "Only support Windows and Linux." #endif - }; + } mem_mgr(const mem_mgr &) = delete; mem_mgr &operator=(const mem_mgr &) = delete; From 20fc3804bfb727074bc270b6eacb60af8d0bf7d4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 4 Jul 2024 10:41:03 +0300 Subject: [PATCH 25/27] convert : fix gemma v1 tokenizer convert (#8248) ggml-ci --- convert-hf-to-gguf-update.py | 5 ++++- convert-hf-to-gguf.py | 3 +++ models/ggml-vocab-bert-bge.gguf.inp | 4 ++++ models/ggml-vocab-bert-bge.gguf.out | 2 ++ models/ggml-vocab-command-r.gguf.inp | 4 ++++ models/ggml-vocab-command-r.gguf.out | 2 ++ models/ggml-vocab-deepseek-coder.gguf.inp | 4 ++++ models/ggml-vocab-deepseek-coder.gguf.out | 2 ++ models/ggml-vocab-deepseek-llm.gguf.inp | 4 ++++ models/ggml-vocab-deepseek-llm.gguf.out | 2 ++ models/ggml-vocab-falcon.gguf.inp | 4 ++++ models/ggml-vocab-falcon.gguf.out | 2 ++ models/ggml-vocab-gpt-2.gguf.inp | 4 ++++ models/ggml-vocab-gpt-2.gguf.out | 2 ++ models/ggml-vocab-llama-bpe.gguf.inp | 6 ++++-- models/ggml-vocab-llama-bpe.gguf.out | 3 ++- models/ggml-vocab-llama-spm.gguf.inp | 4 ++++ models/ggml-vocab-llama-spm.gguf.out | 2 ++ models/ggml-vocab-mpt.gguf.inp | 4 ++++ models/ggml-vocab-mpt.gguf.out | 2 ++ models/ggml-vocab-phi-3.gguf.inp | 4 ++++ models/ggml-vocab-phi-3.gguf.out | 2 ++ models/ggml-vocab-qwen2.gguf.inp | 4 ++++ models/ggml-vocab-qwen2.gguf.out | 2 ++ models/ggml-vocab-refact.gguf.inp | 4 ++++ models/ggml-vocab-refact.gguf.out | 2 ++ models/ggml-vocab-starcoder.gguf.inp | 4 ++++ models/ggml-vocab-starcoder.gguf.out | 2 ++ 28 files changed, 85 insertions(+), 4 deletions(-) diff --git a/convert-hf-to-gguf-update.py b/convert-hf-to-gguf-update.py index 944e9d15abdcc4..9eb406cb878686 100755 --- a/convert-hf-to-gguf-update.py +++ b/convert-hf-to-gguf-update.py @@ -86,6 +86,8 @@ class TOKENIZER_TYPE(IntEnum): {"name": "poro-chat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", }, {"name": "jina-v2-code", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", }, {"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B + {"name": "gemma", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", }, + {"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", }, {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, ] @@ -273,7 +275,8 @@ def get_vocab_base_pre(self, tokenizer) -> str: "3333333", "33333333", "333333333", - # "Cửa Việt", # llama-bpe fails on this + "Cửa Việt", # llama-bpe fails on this + " discards", chktxt, ] diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index d01aed224d6009..bae14558992c8f 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -2316,6 +2316,8 @@ def set_vocab(self): special_vocab._set_special_token("eot", 107) special_vocab.add_to_gguf(self.gguf_writer) + self.gguf_writer.add_add_space_prefix(False) + def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] @@ -2366,6 +2368,7 @@ def set_vocab(self): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) + self.gguf_writer.add_add_space_prefix(False) def set_gguf_parameters(self): diff --git a/models/ggml-vocab-bert-bge.gguf.inp b/models/ggml-vocab-bert-bge.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-bert-bge.gguf.inp +++ b/models/ggml-vocab-bert-bge.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-bert-bge.gguf.out b/models/ggml-vocab-bert-bge.gguf.out index e4a76cdb07d3f2..82d4ed1c13654e 100644 --- a/models/ggml-vocab-bert-bge.gguf.out +++ b/models/ggml-vocab-bert-bge.gguf.out @@ -40,4 +40,6 @@ 21211 22394 22394 21211 22394 22394 2509 21211 22394 22394 22394 + 12731 2050 19710 + 5860 18117 100 1006 3671 1007 100 1006 3674 7861 29147 2483 9530 16280 23854 1007 100 100 1017 3943 21211 21211 2509 21211 22394 21211 22394 2509 21211 22394 22394 21211 22394 22394 2509 1017 1012 1017 1017 1012 1012 1017 1017 1012 1012 1012 1017 100 1029 1855 100 100 6207 100 100 14677 23632 22203 1811 1995 1011 1011 1011 1011 1011 1011 1027 1027 1027 1027 1027 1027 1027 1192 15290 29754 14150 1192 10260 1181 29755 29436 29741 10260 16856 29747 23925 10325 1005 1005 1005 1005 1005 1005 1036 1036 1036 1036 1036 1036 1036 1000 1000 1000 1000 1012 1012 1012 1012 1012 1012 999 999 999 999 999 999 1029 1029 1029 1029 1029 1029 1045 1005 2310 2042 1005 2409 2002 1005 1055 2045 1010 1005 2128 2017 2469 1029 1005 1049 2025 2469 1045 1005 2222 2191 2009 1010 1005 1040 2017 2066 2070 5572 1029 2057 1005 2310 1037 1005 2222 diff --git a/models/ggml-vocab-command-r.gguf.inp b/models/ggml-vocab-command-r.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-command-r.gguf.inp +++ b/models/ggml-vocab-command-r.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-command-r.gguf.out b/models/ggml-vocab-command-r.gguf.out index cc4277daa1d257..939b9dc30a63ec 100644 --- a/models/ggml-vocab-command-r.gguf.out +++ b/models/ggml-vocab-command-r.gguf.out @@ -40,4 +40,6 @@ 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 + 42 30719 12584 + 3642 4388 127731 51628 205 57788 18494 97469 126134 206 2226 256 230 1737 18258 16 80503 122 35927 2226 242 112 57462 1737 54457 223165 106230 2096 16 48389 11254 107 255 2226 107 255 228 26 228 26 26 228 26 26 26 228 26 26 26 26 228 26 26 26 26 26 228 26 26 26 26 26 26 228 26 26 26 26 26 26 26 228 26 26 26 26 26 26 26 26 228 26 21 26 228 26 2271 26 228 26 3834 26 182018 230 174833 38111 249 86325 241 38111 245 86325 232 38111 252 38111 123 38111 261 165 24629 38111 261 38111 103 174833 38111 235 188568 231 5691 12081 13336 2648 29325 14315 24 26 24 27 24 28 24 5123 18372 8391 158343 3512 40071 2196 3236 8750 1764 37097 41168 29721 32797 25646 3802 4975 4975 116167 57178 10251 154048 27292 1767 5125 2632 2155 91 2378 1919 1914 2782 19 2155 3354 1933 5470 38 2155 52 2068 5470 1767 4961 3059 1894 19 2155 43 1933 3026 2725 23186 38 2930 14 20676 1671 14 83 51 diff --git a/models/ggml-vocab-deepseek-coder.gguf.inp b/models/ggml-vocab-deepseek-coder.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-deepseek-coder.gguf.inp +++ b/models/ggml-vocab-deepseek-coder.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-deepseek-coder.gguf.out b/models/ggml-vocab-deepseek-coder.gguf.out index 9ccc560d694ca3..a43e3f0f115d29 100644 --- a/models/ggml-vocab-deepseek-coder.gguf.out +++ b/models/ggml-vocab-deepseek-coder.gguf.out @@ -40,4 +40,6 @@ 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 + 34 155 119 242 64 24297 155 119 216 83 + 1607 2539 185 207 185 185 207 185 185 185 207 12405 459 22758 185 243 185 315 185 251 185 730 185 10047 235 209 334 8760 8 12394 233 114 350 222 10047 221 104 169 116 224 334 4684 3909 992 24330 262 29651 612 8 207 156 237 214 12394 99 234 10047 99 234 207 18 207 18 18 207 18 18 18 207 18 18 18 18 207 18 18 18 18 18 207 18 18 18 18 18 18 207 18 18 18 18 18 18 18 207 18 18 18 18 18 18 18 18 207 18 13 18 207 18 524 18 207 18 1202 18 207 155 239 209 155 239 114 155 239 228 155 240 220 155 239 224 155 240 211 155 239 231 155 239 115 155 239 240 155 240 210 155 239 240 155 239 95 155 239 114 155 239 214 10047 233 210 3015 19100 608 9413 2668 16 18 16 19 16 20 16 1393 169 121 239 18155 374 17194 28 2861 6478 616 2251 14994 31269 4191 6 4686 4686 10252 3358 3358 3409 524 15330 3023 15031 5668 303 6 312 798 651 83 839 362 6 82 741 11 651 1369 340 2037 30 651 44 441 2037 303 6 642 1098 359 11 651 35 340 833 738 10860 30 998 6 10709 245 6 75 43 diff --git a/models/ggml-vocab-deepseek-llm.gguf.inp b/models/ggml-vocab-deepseek-llm.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-deepseek-llm.gguf.inp +++ b/models/ggml-vocab-deepseek-llm.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-deepseek-llm.gguf.out b/models/ggml-vocab-deepseek-llm.gguf.out index fd94b896d24e7f..d31ac1cc6696eb 100644 --- a/models/ggml-vocab-deepseek-llm.gguf.out +++ b/models/ggml-vocab-deepseek-llm.gguf.out @@ -40,4 +40,6 @@ 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 + 34 32555 242 64 23708 32555 216 83 + 1763 2550 185 207 185 185 207 185 185 185 207 11969 486 22504 185 243 185 300 185 251 185 663 185 10044 95300 334 8754 8 33701 114 350 222 10044 221 104 46713 334 34732 996 24250 262 80923 8 207 37103 214 12356 99 234 10044 99 234 207 18 207 18 18 207 18 18 18 207 18 18 18 18 207 18 18 18 18 18 207 18 18 18 18 18 18 207 18 18 18 18 18 18 18 207 18 18 18 18 18 18 18 18 207 18 13 18 207 18 526 18 207 18 1204 18 207 71374 209 71374 114 71374 228 155 240 220 71374 224 155 240 211 71374 231 71374 115 71374 240 155 240 210 71374 240 71374 95 71374 114 71374 214 71899 210 3025 19017 612 9407 2681 16 18 16 19 16 20 16 1398 68940 239 78827 55170 76659 620 91754 31116 36804 4885 4885 10897 4390 4390 41047 15278 3033 14986 5675 304 6 313 803 655 33326 362 6 82 745 11 655 1374 340 2049 30 655 44 441 2049 304 6 647 1099 359 11 655 35 340 837 742 10842 30 1003 6 10699 245 6 75 43 diff --git a/models/ggml-vocab-falcon.gguf.inp b/models/ggml-vocab-falcon.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-falcon.gguf.inp +++ b/models/ggml-vocab-falcon.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-falcon.gguf.out b/models/ggml-vocab-falcon.gguf.out index 209b04cdaf3302..1ab70fe70ba7e2 100644 --- a/models/ggml-vocab-falcon.gguf.out +++ b/models/ggml-vocab-falcon.gguf.out @@ -40,4 +40,6 @@ 22287 22287 30 22287 22287 3138 22287 22287 22287 + 46 19768 239 76 9634 19768 213 95 + 1080 1502 1212 4824 1001 1212 192 204 663 49453 2069 742 561 1501 193 2571 232 206 204 19 11003 20 8196 126 283 219 48778 116 13392 204 19 51831 732 63209 1741 7955 522 20 22438 211 3346 111 231 2571 111 231 204 30 204 3138 204 22287 204 22287 30 204 22287 3138 204 22287 22287 204 22287 22287 30 204 22287 22287 3138 204 30 25 30 204 30 513 30 204 30 951 30 27171 236 206 38154 126 38154 225 167 237 217 38154 221 167 237 208 38154 228 38154 127 38154 237 167 237 207 38154 237 38154 107 38154 126 38154 211 20589 207 204 42 50087 123 2727 20300 32022 133 234 17419 30137 28 7858 181 133 236 204 37057 2228 10666 5052 133 6207 151 215 150 134 5052 133 6279 5052 223 151 216 49679 123 53110 47043 7795 204 7544 7544 7544 8543 8543 17593 3513 3513 12844 51520 17664 4247 295 18 298 650 204 18 95 693 332 18 94 629 23 204 18 1553 299 1310 42 204 18 56 416 1310 295 18 567 717 334 23 204 18 47 299 606 596 6696 42 703 18 16139 241 18 87 55 diff --git a/models/ggml-vocab-gpt-2.gguf.inp b/models/ggml-vocab-gpt-2.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-gpt-2.gguf.inp +++ b/models/ggml-vocab-gpt-2.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-gpt-2.gguf.out b/models/ggml-vocab-gpt-2.gguf.out index 78430f0d31fdcd..88217d3fa787fb 100644 --- a/models/ggml-vocab-gpt-2.gguf.out +++ b/models/ggml-vocab-gpt-2.gguf.out @@ -40,4 +40,6 @@ 24840 20370 24840 24840 24840 2091 20370 + 34 157 119 255 64 16049 157 119 229 83 + 1221 1371 198 220 628 220 628 198 220 197 220 197 197 220 197 198 220 220 198 220 220 220 198 220 220 220 220 198 220 220 220 220 220 198 8582 248 222 357 11265 8 30325 114 447 235 8582 234 104 37929 357 48101 795 13210 271 1673 36686 515 8 14519 227 12520 99 247 8582 99 247 513 4747 23460 513 20370 23460 2091 23460 20370 23460 24840 23460 2091 20370 513 13 18 513 492 18 513 986 18 28053 252 222 157 252 114 157 252 241 157 253 233 157 252 237 157 253 224 157 252 244 157 252 115 157 252 253 157 253 223 157 252 253 157 252 95 157 252 114 157 252 227 47249 223 5633 22755 239 46349 111 28839 101 18040 32432 98 43291 1485 1415 24309 25465 171 121 252 40103 1421 18604 12466 121 16843 141 231 15166 12466 121 16142 12466 239 141 232 30143 140 111 16142 21169 21727 31583 18849 705 39115 6 33153 15506 63 15931 15931 16317 13896 3228 9805 3548 314 1053 587 705 44040 339 338 612 11 705 2200 345 1654 30 705 44 407 1654 314 1183 787 340 11 705 35 345 588 617 8887 30 775 6 26979 257 6 75 43 diff --git a/models/ggml-vocab-llama-bpe.gguf.inp b/models/ggml-vocab-llama-bpe.gguf.inp index 9380bf355202ab..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-llama-bpe.gguf.inp +++ b/models/ggml-vocab-llama-bpe.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ @@ -104,5 +108,3 @@ __ggml_vocab_test__ 🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL __ggml_vocab_test__ - Việt -__ggml_vocab_test__ diff --git a/models/ggml-vocab-llama-bpe.gguf.out b/models/ggml-vocab-llama-bpe.gguf.out index 1f3607fb6a3785..bb1fe229c60710 100644 --- a/models/ggml-vocab-llama-bpe.gguf.out +++ b/models/ggml-vocab-llama-bpe.gguf.out @@ -40,5 +40,6 @@ 8765 8765 18 8765 8765 1644 8765 8765 8765 + 34 91163 101798 + 2624 2402 198 4815 15073 66597 8004 1602 2355 79772 11187 9468 248 222 320 8416 8 27623 114 102470 9468 234 104 31643 320 36773 100166 98634 8 26602 227 11410 99 247 9468 99 247 220 18 220 1644 220 8765 220 8765 18 220 8765 1644 220 8765 8765 220 8765 8765 18 220 8765 8765 1644 220 18 13 18 220 18 497 18 220 18 1131 18 220 21549 222 98629 241 45358 233 21549 237 45358 224 21549 244 21549 115 21549 253 45358 223 21549 253 21549 95 98629 227 76460 223 949 37046 101067 19000 23182 102301 9263 18136 16 36827 21909 56560 54337 19175 102118 13373 64571 34694 3114 112203 80112 3436 106451 14196 14196 74694 3089 3089 29249 17523 3001 27708 7801 358 3077 1027 364 83 820 568 596 1070 11 364 793 499 2771 30 364 44 539 2771 358 3358 1304 433 11 364 35 499 1093 1063 15600 30 1226 6 43712 264 64966 43 - 101798 diff --git a/models/ggml-vocab-llama-spm.gguf.inp b/models/ggml-vocab-llama-spm.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-llama-spm.gguf.inp +++ b/models/ggml-vocab-llama-spm.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-llama-spm.gguf.out b/models/ggml-vocab-llama-spm.gguf.out index 9c3327cb543807..1c3b0a2c979c51 100644 --- a/models/ggml-vocab-llama-spm.gguf.out +++ b/models/ggml-vocab-llama-spm.gguf.out @@ -40,4 +40,6 @@ 29871 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29941 + 315 228 190 176 29874 10630 30529 29873 + 29871 2313 3163 29871 13 29871 13 13 29871 13 13 13 29871 12 29871 12 12 29871 12 13 259 13 1678 13 268 13 418 13 243 162 157 131 313 8945 29897 29871 243 162 155 185 30722 243 162 143 174 30598 313 20787 953 3848 275 16125 630 29897 29871 31681 29871 243 162 169 156 243 162 169 156 29871 29941 29871 29941 29941 29871 29941 29941 29941 29871 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29871 29941 29889 29941 29871 29941 636 29941 29871 29941 856 29941 29871 31849 31324 31934 228 162 142 228 161 146 228 162 133 228 161 153 228 161 186 31708 228 162 132 31708 228 161 165 31324 228 161 136 243 162 155 132 1577 30672 31522 30505 11548 31041 30732 29896 29941 29896 29946 29896 29945 29896 30408 30739 448 23648 2751 25512 1538 4851 665 1386 29713 1305 14550 4907 11120 16159 16159 16159 15945 15945 3045 636 6824 6824 6824 8773 8773 8773 306 29915 345 1063 525 29873 1025 540 29915 29879 727 29892 525 1525 366 1854 29973 525 29924 451 1854 306 29915 645 1207 372 29892 525 29928 366 763 777 23429 29973 1334 29915 29963 29872 263 29915 29880 29931 diff --git a/models/ggml-vocab-mpt.gguf.inp b/models/ggml-vocab-mpt.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-mpt.gguf.inp +++ b/models/ggml-vocab-mpt.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-mpt.gguf.out b/models/ggml-vocab-mpt.gguf.out index d8d0fe90900bb7..d4a877d1c2641f 100644 --- a/models/ggml-vocab-mpt.gguf.out +++ b/models/ggml-vocab-mpt.gguf.out @@ -40,4 +40,6 @@ 26409 20084 26409 26409 26409 1610 20084 + 36 6829 244 66 17721 35177 85 + 1262 2196 586 1744 33525 186 209 623 28910 187 50276 187 50275 187 50274 187 50273 187 14931 237 211 313 6320 10 49042 116 325 224 14931 223 106 171 118 226 313 34263 802 13511 261 32147 456 10 3384 239 216 22692 101 236 14931 101 236 495 5922 30057 495 20084 495 26409 30057 20084 495 26409 1610 495 26409 20084 495 15 20 495 537 20 495 1051 20 209 18081 211 18081 116 18081 230 39936 222 18081 226 39936 213 18081 233 18081 117 18081 242 39936 212 18081 242 18081 97 18081 116 18081 216 14931 235 212 3736 15367 41197 13610 19934 41869 21275 1012 1047 18795 40120 20422 241 16081 6877 12880 11514 1068 8713 38177 13396 3415 9925 12559 10453 1389 42011 35033 34842 11202 9739 9739 33021 18963 4672 25561 8220 309 1849 644 686 42618 344 434 627 13 686 1848 368 2119 32 686 46 417 2119 309 1833 1056 352 13 686 37 368 751 690 10331 32 844 8 31516 247 8 77 45 diff --git a/models/ggml-vocab-phi-3.gguf.inp b/models/ggml-vocab-phi-3.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-phi-3.gguf.inp +++ b/models/ggml-vocab-phi-3.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-phi-3.gguf.out b/models/ggml-vocab-phi-3.gguf.out index 9c3327cb543807..1c3b0a2c979c51 100644 --- a/models/ggml-vocab-phi-3.gguf.out +++ b/models/ggml-vocab-phi-3.gguf.out @@ -40,4 +40,6 @@ 29871 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29941 + 315 228 190 176 29874 10630 30529 29873 + 29871 2313 3163 29871 13 29871 13 13 29871 13 13 13 29871 12 29871 12 12 29871 12 13 259 13 1678 13 268 13 418 13 243 162 157 131 313 8945 29897 29871 243 162 155 185 30722 243 162 143 174 30598 313 20787 953 3848 275 16125 630 29897 29871 31681 29871 243 162 169 156 243 162 169 156 29871 29941 29871 29941 29941 29871 29941 29941 29941 29871 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29871 29941 29941 29941 29941 29941 29941 29941 29941 29871 29941 29889 29941 29871 29941 636 29941 29871 29941 856 29941 29871 31849 31324 31934 228 162 142 228 161 146 228 162 133 228 161 153 228 161 186 31708 228 162 132 31708 228 161 165 31324 228 161 136 243 162 155 132 1577 30672 31522 30505 11548 31041 30732 29896 29941 29896 29946 29896 29945 29896 30408 30739 448 23648 2751 25512 1538 4851 665 1386 29713 1305 14550 4907 11120 16159 16159 16159 15945 15945 3045 636 6824 6824 6824 8773 8773 8773 306 29915 345 1063 525 29873 1025 540 29915 29879 727 29892 525 1525 366 1854 29973 525 29924 451 1854 306 29915 645 1207 372 29892 525 29928 366 763 777 23429 29973 1334 29915 29963 29872 263 29915 29880 29931 diff --git a/models/ggml-vocab-qwen2.gguf.inp b/models/ggml-vocab-qwen2.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-qwen2.gguf.inp +++ b/models/ggml-vocab-qwen2.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-qwen2.gguf.out b/models/ggml-vocab-qwen2.gguf.out index 401a510e86f3a1..4ab275396631d3 100644 --- a/models/ggml-vocab-qwen2.gguf.out +++ b/models/ggml-vocab-qwen2.gguf.out @@ -40,4 +40,6 @@ 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 + 34 90063 128324 + 2560 2347 198 4710 14731 65497 7847 1572 2303 78672 10947 145836 320 8252 8 26525 114 378 235 149921 30543 320 35673 99066 97534 8 25521 227 11162 99 247 149955 220 18 220 18 18 220 18 18 18 220 18 18 18 18 220 18 18 18 18 18 220 18 18 18 18 18 18 220 18 18 18 18 18 18 18 220 18 18 18 18 18 18 18 18 220 18 13 18 220 18 496 18 220 18 1112 18 220 146394 97529 241 44258 233 146568 44258 224 147603 20879 115 146280 44258 223 146280 147272 97529 227 144534 937 104100 18493 22377 99257 16 18 16 19 16 20 16 35727 21216 55460 53237 18658 14144 1456 13073 63471 33594 3038 133178 79012 3355 4605 4605 13874 13874 73594 3014 3014 28149 17085 2928 26610 7646 358 3003 1012 364 83 813 566 594 1052 11 364 787 498 2704 30 364 44 537 2704 358 3278 1281 432 11 364 35 498 1075 1045 15243 30 1205 6 42612 264 63866 43 diff --git a/models/ggml-vocab-refact.gguf.inp b/models/ggml-vocab-refact.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-refact.gguf.inp +++ b/models/ggml-vocab-refact.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-refact.gguf.out b/models/ggml-vocab-refact.gguf.out index 06b15c090c0f8a..46d8b4aec7e194 100644 --- a/models/ggml-vocab-refact.gguf.out +++ b/models/ggml-vocab-refact.gguf.out @@ -40,4 +40,6 @@ 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 + 53 33934 83 33217 17102 102 + 1214 12258 334 719 8878 202 10885 4222 16104 28570 203 3807 253 227 308 4382 27 18458 133 46113 44967 123 13868 308 12565 19775 33071 40824 733 27 41889 5945 118 252 3807 118 252 225 37 225 37 37 225 37 37 37 225 37 37 37 37 225 37 37 37 37 37 225 37 37 37 37 37 37 225 37 37 37 37 37 37 37 225 37 37 37 37 37 37 37 37 225 37 32 37 225 37 497 37 225 37 1179 37 225 14574 227 14574 133 14574 246 30457 238 14574 242 30457 229 14574 249 14574 134 14574 258 30457 228 14574 258 14574 114 14574 133 14574 232 36628 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838 20921 16623 13028 8372 1039 9446 40242 13852 2053 8949 12531 1520 10700 5881 9592 13299 914 31753 31359 9163 3202 35472 10397 439 4763 2583 330 102 1455 938 1182 2017 30 330 613 844 3654 49 330 63 646 3654 439 4621 1930 561 30 330 54 844 2124 1629 35993 49 2688 25 7709 312 25 94 62 diff --git a/models/ggml-vocab-starcoder.gguf.inp b/models/ggml-vocab-starcoder.gguf.inp index 0a89107c60d7f6..5b4aeb31ac9c68 100644 --- a/models/ggml-vocab-starcoder.gguf.inp +++ b/models/ggml-vocab-starcoder.gguf.inp @@ -91,6 +91,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-starcoder.gguf.out b/models/ggml-vocab-starcoder.gguf.out index ccb55c7feeef80..9ce2698a97d7bc 100644 --- a/models/ggml-vocab-starcoder.gguf.out +++ b/models/ggml-vocab-starcoder.gguf.out @@ -40,4 +40,6 @@ 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 + 72 34269 102 33245 17234 121 + 1236 12266 353 736 8886 221 10883 4238 16101 28540 222 3822 272 246 327 4434 46 18445 152 46030 45022 142 13878 327 12585 19884 33773 40920 751 46 41839 5954 137 271 3822 137 271 244 56 244 56 56 244 56 56 56 244 56 56 56 56 244 56 56 56 56 56 244 56 56 56 56 56 56 244 56 56 56 56 56 56 56 244 56 56 56 56 56 56 56 56 244 56 51 56 244 56 516 56 244 56 1198 56 244 14566 246 14566 152 14566 265 30428 257 14566 261 30428 248 14566 268 14566 153 14566 277 30428 247 14566 277 14566 133 14566 152 14566 251 36570 247 1037 4995 13379 2924 9515 17823 54 56 54 57 54 58 54 11904 47892 20895 16625 13047 8389 1059 9504 40216 13858 2073 8983 12571 1539 10721 5918 9643 13298 932 31723 31330 9221 3226 35426 10400 457 4783 2602 349 121 1477 957 1200 2038 49 349 632 863 3673 68 349 82 666 3673 457 4650 1949 580 49 349 73 863 2144 1649 35941 68 2726 44 7728 331 44 113 81 From 402d6feffa0572d1c7a957901b6d1702bd188484 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Thu, 4 Jul 2024 12:50:57 +0200 Subject: [PATCH 26/27] llama : suppress unref var in Windows MSVC (#8150) * llama : suppress unref var in Windows MSVC This commit suppresses two warnings that are currently generated for src/llama.cpp when building on Windows MSVC ```console C:\llama.cpp\src\llama.cpp(14349,45): warning C4101: 'ex': unreferenced local variable [C:\llama.cpp\build\src\llama.vcxproj] C:\llama.cpp\src\llama.cpp(19285,44): warning C4101: 'e': unreferenced local variable [C:\llama.cpp\build\src\llama.vcxproj] ``` * Update src/llama.cpp --------- Co-authored-by: Georgi Gerganov --- src/llama.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/llama.cpp b/src/llama.cpp index 73f52435a503ef..3d131b325ef209 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -14781,7 +14781,7 @@ struct llm_tokenizer_ugm { size_t prefix_offset = input_offset; unicode_cpt_from_utf8(input, prefix_offset); return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset }; - } catch(std::invalid_argument & ex) { + } catch (std::invalid_argument & /*ex*/) { // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER return { "\xEF\xBF\xBD", 3, 1 }; } @@ -19725,7 +19725,7 @@ static std::string llama_decode_text(const std::string & text) { const auto utf8 = unicode_cpt_to_utf8(cpt); try { decoded_text += unicode_utf8_to_byte(utf8); - } catch (const std::out_of_range & e) { + } catch (const std::out_of_range & /*e*/) { decoded_text += "[UNK_BYTE_0x"; for (const auto c : utf8) { decoded_text += format("%02x", (uint8_t) c); From f8c4c0738d72d2162736edd72dd5db8b269adca1 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Thu, 4 Jul 2024 12:53:42 +0200 Subject: [PATCH 27/27] tests : add _CRT_SECURE_NO_WARNINGS for WIN32 (#8231) This commit adds the compile definition `_CRT_SECURE_NO_WARNINGS` to the root cmake subproject. The motivation for this is that currently the following warnings are displayed when compiling the tests and common cmake subprojects: ```console test-llama-grammar.cpp C:\llama.cpp\src\.\llama.cpp(1406,77): warning C4996: 'strerror': This function or variable may be unsafe. Consider using strerror_s instead. To disable deprecation, use _CRT_SECURE_NO_WARNINGS. See online help for details. [C:\llama.cpp\build\tests\test-llama-grammar.vcxproj] ... ``` This compile definition is currently set for the `src` subproject and this change moves into the root cmake project so that it is applied to all cmake subprojects. --- CMakeLists.txt | 4 ++++ src/CMakeLists.txt | 2 -- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index e3a0cc369e364e..d95414d710f566 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -42,6 +42,10 @@ endif() option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT}) +if (WIN32) + add_compile_definitions(_CRT_SECURE_NO_WARNINGS) +endif() + # # option list # diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index ccb607e56d336b..c2049df79c212b 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -1,7 +1,5 @@ # TODO: should not use this if (WIN32) - add_compile_definitions(_CRT_SECURE_NO_WARNINGS) - if (BUILD_SHARED_LIBS) set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON) endif()