Compare commits

...

25 Commits

Author SHA1 Message Date
Rotem Dan f3ed38d793 Set RPATH to "@loader_path" / "$ORIGIN" to ensure executables and dynamic libraries search for dependencies in their origin directory. (#14309) 2025-07-02 18:37:16 +02:00
Aman Gupta 55a1c5a5fd CUDA: add softmax broadcast (#14475)
* CUDA: add softmax broadcast

* Pass by const ref

* Review: Use blockDims for indexing, remove designated initializers

* Add TODO for noncontigous input/output
2025-07-02 15:48:33 +03:00
Johannes Gäßler 12a81af45f CUDA: broadcasting for FlashAttention mask (#14500) 2025-07-02 15:48:33 +03:00
Jeff Bolz 8875523eb3 vulkan: support softmax/FA batch and broadcast (#14449) 2025-07-02 15:48:33 +03:00
Georgi Gerganov ec68e84c32 ggml : support bcast ggml_soft_max_ext, ggml_flash_attn_ext (#14435)
ggml-ci
2025-07-02 15:48:33 +03:00
zhouwg 307e79d33d opencl : fix possible buffer overflow in dump_tensor (#14490) 2025-07-02 14:38:10 +02:00
Georgi Gerganov d7f5f4e578 simple-chat : fix context-exceeded condition (#14494)
* simple-chat : fix context-exceeded condition

ggml-ci

* cont : fix n_ctx_used computation

ggml-ci
2025-07-02 14:12:07 +03:00
Eric Zhang c8a4e470f6 opencl : skip empty nodes on cgraph compute (#14491) 2025-07-02 13:00:04 +02:00
lhez 603e43dc91 opencl : update upscale to support align corners (#14488) 2025-07-02 09:07:42 +02:00
Sigbjørn Skjæret 611ba4b264 ci : add OpenCL to labeler workflow (#14496) 2025-07-02 09:02:51 +02:00
Eric Zhang 85841e121d github : add OpenCL backend to issue templates (#14492) 2025-07-02 08:41:35 +03:00
Björn Ganster 68b3cd6514 ggml : Callback before abort (#14481)
* Add a callback that will be called just before abort. This allows apps without a console to display a message to the user and save data if needed.

* Return previous callback to allow callback chaining

* style fixes

---------

Co-authored-by: Diego Devesa <slarengh@gmail.com>
2025-07-02 08:19:31 +03:00
Georgi Gerganov de56944147 ci : disable fast-math for Metal GHA CI (#14478)
* ci : disable fast-math for Metal GHA CI

ggml-ci

* cont : remove -g flag

ggml-ci
2025-07-01 18:04:08 +03:00
Grzegorz Grasza 1b2aaf28ac Add Vulkan images to docker.md (#14472)
Right now it's not easy to find those.
2025-07-01 15:44:11 +02:00
Chenguang Li 343b6e94b6 CANN: update aclnnGroupedMatmulV2 to aclnnGroupedMatmulV3 (#14411)
* [CANN]update to aclnnGroupedMatmulV2

Signed-off-by: noemotiovon <757486878@qq.com>

* Support MUL_MAT_ID on 310p

Signed-off-by: noemotiovon <757486878@qq.com>

* fix editorconfig

Signed-off-by: noemotiovon <757486878@qq.com>

---------

Signed-off-by: noemotiovon <757486878@qq.com>
2025-07-01 16:47:30 +08:00
Jeff Bolz 6a746cf9c4 vulkan: Split large mul_mat_id to fit in shared memory (#14451) 2025-07-01 10:43:08 +02:00
Sigbjørn Skjæret eff5e45443 add GELU_ERF (#14455) 2025-07-01 10:14:21 +02:00
Georgi Gerganov a6a47958a1 ggml : remove trailing whitespace (#0) 2025-07-01 11:06:39 +03:00
Georgi Gerganov f61c05d4b1 sync : ggml
ggml-ci
2025-07-01 11:06:39 +03:00
Acly 431b2c24f3 ggml-cpu : "align corners" for bilinear upscale/downscale (ggml/1285)
* add "align corners" mode for bilinear upscale, and allow downscaling
* add ggml_interpolate, deprecate ggml_upscale_ext, pass in align-corners as bit-flag
* test-backend-ops: replace ggml_upscale_ext with ggml_interpolate, add test cases for downscale and align-corners
2025-07-01 11:06:39 +03:00
Daniel Bevenius 497be7c01d ggml-quants : rename best_mad to best_error (ggml/1283)
This commit renames the variable `best_mad` to `best_error` in the
`make_qkx2_quants` function.

The motivation for this is that the name `best_mad` can be somewhat
confusing if mean absolute deviation (MAD) is not in use.
2025-07-01 11:06:39 +03:00
lhez 79b33b2317 opencl : add GEGLU, REGLU, SWIGLU (#14456) 2025-07-01 09:19:16 +02:00
Aman Gupta 0a5a3b5cdf Add Conv2d for CPU (#14388)
* Conv2D: Add CPU version

* Half decent

* Tiled approach for F32

* remove file

* Fix tests

* Support F16 operations

* add assert about size

* Review: further formatting fixes, add assert and use CPU version of fp32->fp16
2025-06-30 23:57:04 +08:00
Georgi Gerganov 745f11fed0 memory : correctly handle failure in apply() (#14438)
ggml-ci
2025-06-30 18:03:03 +03:00
Georgi Gerganov 5dd942de59 metal : disable fast-math for some cpy kernels (#14460)
* metal : disable fast-math for some cpy kernels

ggml-ci

* cont : disable for q4_1

ggml-ci

* cont : disable for iq4_nl

ggml-ci
2025-06-30 17:04:05 +03:00
50 changed files with 1374 additions and 333 deletions
@@ -40,7 +40,7 @@ body:
attributes:
label: GGML backends
description: Which GGML backends do you know to be affected?
options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
multiple: true
validations:
required: true
+1 -1
View File
@@ -42,7 +42,7 @@ body:
attributes:
label: GGML backends
description: Which GGML backends do you know to be affected?
options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
multiple: true
validations:
required: true
+5
View File
@@ -93,3 +93,8 @@ Ascend NPU:
- ggml/include/ggml-cann.h
- ggml/src/ggml-cann/**
- docs/backend/CANN.md
OpenCL:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-opencl.h
- ggml/src/ggml-opencl/**
+2 -1
View File
@@ -84,7 +84,8 @@ jobs:
-DCMAKE_BUILD_RPATH="@loader_path" \
-DLLAMA_FATAL_WARNINGS=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DGGML_METAL_EMBED_LIBRARY=OFF \
-DGGML_METAL_SHADER_DEBUG=ON \
-DGGML_RPC=ON
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+8 -2
View File
@@ -49,7 +49,8 @@ jobs:
run: |
sysctl -a
cmake -B build \
-DCMAKE_BUILD_RPATH="@loader_path" \
-DCMAKE_INSTALL_RPATH='@loader_path' \
-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-DLLAMA_FATAL_WARNINGS=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
@@ -103,7 +104,8 @@ jobs:
# Metal is disabled due to intermittent failures with Github runners not having a GPU:
# https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
cmake -B build \
-DCMAKE_BUILD_RPATH="@loader_path" \
-DCMAKE_INSTALL_RPATH='@loader_path' \
-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-DLLAMA_FATAL_WARNINGS=ON \
-DGGML_METAL=OFF \
-DGGML_RPC=ON
@@ -160,6 +162,8 @@ jobs:
id: cmake_build
run: |
cmake -B build \
-DCMAKE_INSTALL_RPATH='$ORIGIN' \
-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-DGGML_BACKEND_DL=ON \
-DGGML_NATIVE=OFF \
-DGGML_CPU_ALL_VARIANTS=ON \
@@ -211,6 +215,8 @@ jobs:
id: cmake_build
run: |
cmake -B build \
-DCMAKE_INSTALL_RPATH='$ORIGIN' \
-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-DGGML_BACKEND_DL=ON \
-DGGML_NATIVE=OFF \
-DGGML_CPU_ALL_VARIANTS=ON \
+3
View File
@@ -25,6 +25,9 @@ Additionally, there the following images, similar to the above:
- `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`)
The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
+4 -3
View File
@@ -113,15 +113,16 @@ int main(int argc, char ** argv) {
while (true) {
// check if we have enough space in the context to evaluate this batch
int n_ctx = llama_n_ctx(ctx);
int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0);
int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) + 1;
if (n_ctx_used + batch.n_tokens > n_ctx) {
printf("\033[0m\n");
fprintf(stderr, "context size exceeded\n");
exit(0);
}
if (llama_decode(ctx, batch)) {
GGML_ABORT("failed to decode\n");
int ret = llama_decode(ctx, batch);
if (ret != 0) {
GGML_ABORT("failed to decode, ret = %d\n", ret);
}
// sample the next token
+56 -8
View File
@@ -314,6 +314,13 @@
extern "C" {
#endif
// Function type used in fatal error callbacks
typedef void (*ggml_abort_callback_t)(const char * error_message);
// Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
// Returns the old callback for chaining
GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback);
GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
@@ -482,6 +489,7 @@ extern "C" {
GGML_OP_CONV_TRANSPOSE_1D,
GGML_OP_IM2COL,
GGML_OP_IM2COL_BACK,
GGML_OP_CONV_2D,
GGML_OP_CONV_2D_DW,
GGML_OP_CONV_TRANSPOSE_2D,
GGML_OP_POOL_1D,
@@ -1502,8 +1510,14 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// a [ne0, ne01, ne02, ne03]
// mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
//
// broadcast:
// ne02 % ne12 == 0
// ne03 % ne13 == 0
//
// fused soft_max(a*scale + mask*(ALiBi slope))
// mask is optional
// max_bias = 0.0f for no ALiBi
GGML_API struct ggml_tensor * ggml_soft_max_ext(
struct ggml_context * ctx,
@@ -1813,6 +1827,17 @@ extern "C" {
struct ggml_tensor * b,
int stride);
GGML_API struct ggml_tensor * ggml_conv_2d_direct(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel [KW, KH, IC, OC]
struct ggml_tensor * b, // input data [W, H, C, N]
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1
enum ggml_op_pool {
GGML_OP_POOL_MAX,
GGML_OP_POOL_AVG,
@@ -1855,6 +1880,12 @@ extern "C" {
enum ggml_scale_mode {
GGML_SCALE_MODE_NEAREST = 0,
GGML_SCALE_MODE_BILINEAR = 1,
GGML_SCALE_MODE_COUNT
};
enum ggml_scale_flag {
GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
};
// interpolate
@@ -1867,14 +1898,26 @@ extern "C" {
// interpolate
// interpolate scale to specified dimensions
GGML_API struct ggml_tensor * ggml_upscale_ext(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_upscale_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1,
int ne2,
int ne3,
enum ggml_scale_mode mode);
enum ggml_scale_mode mode),
"use ggml_interpolate instead");
// Up- or downsamples the input to the specified size.
// 2D scale modes (eg. bilinear) are applied to the first two dimensions.
GGML_API struct ggml_tensor * ggml_interpolate(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3,
uint32_t mode); // ggml_scale_mode [ | ggml_scale_flag...]
// pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
GGML_API struct ggml_tensor * ggml_pad(
@@ -1937,11 +1980,16 @@ extern "C" {
#define GGML_KQ_MASK_PAD 64
// q: [n_embd_k, n_batch, n_head, 1]
// k: [n_embd_k, n_kv, n_head_kv, 1]
// v: [n_embd_v, n_kv, n_head_kv, 1] !! not transposed !!
// mask: [n_kv, n_batch_pad, 1, 1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
// res: [n_embd_v, n_head, n_batch, 1] !! permuted !!
// q: [n_embd_k, n_batch, n_head, ne3]
// k: [n_embd_k, n_kv, n_head_kv, ne3]
// v: [n_embd_v, n_kv, n_head_kv, ne3] !! not transposed !!
// mask: [n_kv, n_batch_pad, ne32, 1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
// res: [n_embd_v, n_head, n_batch, ne3] !! permuted !!
//
// broadcast:
// n_head % n_head_kv == 0
// ne3 % ne32 == 0
//
GGML_API struct ggml_tensor * ggml_flash_attn_ext(
struct ggml_context * ctx,
struct ggml_tensor * q,
+65 -4
View File
@@ -65,7 +65,7 @@
#include <aclnnop/aclnn_eq_tensor.h>
#include <aclnnop/aclnn_gt_scalar.h>
#include <aclnnop/aclnn_pow.h>
#include <aclnnop/aclnn_grouped_matmul_v2.h>
#include <aclnnop/aclnn_grouped_matmul_v3.h>
#include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
#include <float.h>
@@ -2654,6 +2654,67 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
}
#ifdef ASCEND_310P
ggml_tensor src0_row = *src0;
ggml_tensor src1_row = *src1;
ggml_tensor dst_row = *dst;
if (src0->type == GGML_TYPE_F16) {
src0_row.type = GGML_TYPE_F32;
}
// src0_row [D, M, 1, 1] weight without permute
src0_row.ne[2] = 1;
src0_row.ne[3] = 1;
src0_row.nb[0] = ori_src0_nb[0];
src0_row.nb[1] = ori_src0_nb[1];
src0_row.nb[2] = ori_src0_nb[1];
src0_row.nb[3] = ori_src0_nb[1];
// src1_row [D, 1, 1, 1] -> input
src1_row.ne[1] = 1;
src1_row.ne[2] = 1;
src1_row.ne[3] = 1;
src1_row.nb[2] = nb11;
src1_row.nb[3] = nb11;
// dst_row [M, 1, 1, 1] -> out
dst_row.ne[1] = 1;
dst_row.ne[2] = 1;
dst_row.ne[3] = 1;
dst_row.nb[2] = nb1;
dst_row.nb[3] = nb1;
//create weight for one row
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
for (int64_t id = 0; id < n_ids; id++) {
// expert index
int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
GGML_ASSERT(i02 >= 0 && i02 < n_as);
// If B = 1 (broadcast), always use 0; otherwise, use id.
int64_t i11 = (ne11 == 1 ? 0 : id);
int64_t i12 = iid1;
int64_t i1 = id;
int64_t i2 = i12;
void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
void* dst_tmp_ptr = dst_original + i1*nb1 + i2*nb2;
src0_row.data = src0_tmp_ptr;
src1_row.data = src1_tmp_ptr;
dst_row.data = dst_tmp_ptr;
dst_row.src[0] = &src0_row;
dst_row.src[1] = &src1_row;
ggml_cann_mul_mat(ctx, &dst_row);
}
}
return;
#endif
std::vector<aclTensor*> src0_tensor_vec;
std::vector<aclTensor*> src1_tensor_vec;
std::vector<aclTensor*> dst_tensor_vec;
@@ -2701,9 +2762,9 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
}
size_t GROUP_SIZE = 128;
// GroupedMatmulV2 required tensor_list.size < 128
// GroupedMatmulV3 required tensor_list.size < 128
for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
// split and call GroupedMatmulV2
// split and call GroupedMatmulV3
size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
@@ -2713,7 +2774,7 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, src1_tensor_list, src0_tensor_list,
nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);
+6 -1
View File
@@ -2187,7 +2187,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_SQRT:
case GGML_OP_CLAMP:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
case GGML_OP_SUM_ROWS:
case GGML_OP_ARGSORT:
case GGML_OP_ACC:
@@ -2205,6 +2204,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_PAD_REFLECT_1D:
case GGML_OP_COUNT_EQUAL:
return true;
case GGML_OP_SOFT_MAX:
// TODO: support broadcast
// ref: https://github.com/ggml-org/llama.cpp/pull/14435
return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
case GGML_OP_FLASH_ATTN_EXT:{
// derived from [ggml-cuda.cu]
if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@@ -2227,6 +2230,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
// DeepSeek MLA
return false;
}
// TODO: support broadcast
// ref: https://github.com/ggml-org/llama.cpp/pull/14435
if (op->src[0]->ne[3] != 1) {
return false;
}
+10 -1
View File
@@ -1193,7 +1193,7 @@ static void ggml_compute_forward_mul_mat_one_chunk(
}
}
static void ggml_compute_forward_mul_mat(
void ggml_compute_forward_mul_mat(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
@@ -1866,6 +1866,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_im2col_back_f32(params, tensor);
} break;
case GGML_OP_CONV_2D:
{
ggml_compute_forward_conv_2d(params, tensor);
} break;
case GGML_OP_CONV_2D_DW:
{
ggml_compute_forward_conv_2d_dw(params, tensor);
@@ -2228,6 +2232,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
} break;
case GGML_OP_IM2COL:
case GGML_OP_IM2COL_BACK:
case GGML_OP_CONV_2D:
case GGML_OP_CONV_2D_DW:
case GGML_OP_CONV_TRANSPOSE_1D:
case GGML_OP_CONV_TRANSPOSE_2D:
@@ -2746,6 +2751,10 @@ struct ggml_cplan ggml_graph_plan(
GGML_ABORT("fatal error");
}
} break;
case GGML_OP_CONV_2D:
{
cur = GGML_IM2COL_WORK_SIZE;
} break;
case GGML_OP_CONV_TRANSPOSE_2D:
{
const int64_t ne00 = node->src[0]->ne[0]; // W
+251 -64
View File
@@ -3,6 +3,7 @@
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include "binary-ops.h"
#include "ggml.h"
#include "unary-ops.h"
#include "vec.h"
@@ -5231,14 +5232,17 @@ static void ggml_compute_forward_soft_max_f32(
memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
// TODO: handle transposed/permuted matrices
const int ith = params->ith;
const int nth = params->nth;
GGML_TENSOR_UNARY_OP_LOCALS
//const int64_t ne11 = src1 ? src1->ne[1] : 1;
const int64_t nb11 = src1 ? src1->nb[1] : 1;
const int64_t nb12 = src1 ? src1->nb[2] : 1;
const int64_t nb13 = src1 ? src1->nb[3] : 1;
const int64_t ne12 = src1 ? src1->ne[2] : 1;
const int64_t ne13 = src1 ? src1->ne[3] : 1;
// TODO: is this supposed to be ceil instead of floor?
// https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370
@@ -5248,68 +5252,66 @@ static void ggml_compute_forward_soft_max_f32(
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
const int nc = src0->ne[0];
const int nr = ggml_nrows(src0);
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
float * wp = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
for (int i1 = ir0; i1 < ir1; i1++) {
// ALiBi
const uint32_t h = (i1/ne01)%ne02; // head
const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
const int64_t i11 = i01;
const int64_t i12 = i02%ne12;
const int64_t i13 = i03%ne13;
float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
float * dp = (float *)((char *) dst->data + i1*dst->nb[1]);
// ALiBi
const uint32_t h = i02; // head
const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
// broadcast the mask across rows
ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL;
float * mp_f32 = src1 ? (float *)((char *) src1->data) + (i1%ne01)*ne00 : NULL;
float * sp = (float *)((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
float * dp = (float *)((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
ggml_vec_cpy_f32 (nc, wp, sp);
ggml_vec_scale_f32(nc, wp, scale);
if (mp_f32) {
if (use_f16) {
for (int i = 0; i < nc; ++i) {
wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]);
// broadcast the mask across rows
ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13) : NULL;
float * mp_f32 = src1 ? (float *)((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13) : NULL;
ggml_vec_cpy_f32 (ne00, wp, sp);
ggml_vec_scale_f32(ne00, wp, scale);
if (mp_f32) {
if (use_f16) {
for (int i = 0; i < ne00; ++i) {
wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]);
}
} else {
for (int i = 0; i < ne00; ++i) {
wp[i] += slope*mp_f32[i];
}
}
}
} else {
for (int i = 0; i < nc; ++i) {
wp[i] += slope*mp_f32[i];
#ifndef NDEBUG
for (int i = 0; i < ne00; ++i) {
//printf("p[%d] = %f\n", i, p[i]);
assert(!isnan(wp[i]));
}
#endif
float max = -INFINITY;
ggml_vec_max_f32(ne00, &max, wp);
ggml_float sum = ggml_vec_soft_max_f32(ne00, dp, wp, max);
assert(sum > 0.0);
sum = 1.0/sum;
ggml_vec_scale_f32(ne00, dp, sum);
#ifndef NDEBUG
for (int i = 0; i < ne00; ++i) {
assert(!isnan(dp[i]));
assert(!isinf(dp[i]));
}
#endif
}
}
#ifndef NDEBUG
for (int i = 0; i < nc; ++i) {
//printf("p[%d] = %f\n", i, p[i]);
assert(!isnan(wp[i]));
}
#endif
float max = -INFINITY;
ggml_vec_max_f32(nc, &max, wp);
ggml_float sum = ggml_vec_soft_max_f32(nc, dp, wp, max);
assert(sum > 0.0);
sum = 1.0/sum;
ggml_vec_scale_f32(nc, dp, sum);
#ifndef NDEBUG
for (int i = 0; i < nc; ++i) {
assert(!isnan(dp[i]));
assert(!isinf(dp[i]));
}
#endif
}
}
@@ -6545,6 +6547,186 @@ void ggml_compute_forward_im2col_back_f32(
}
}
static void ggml_call_mul_mat(ggml_type type, const ggml_compute_params * params, int64_t m, int64_t n, int64_t k,
void * a, void * b, float * c) {
const ggml_type_traits * traits = ggml_get_type_traits(type);
struct ggml_tensor src1 = {};
src1.type = type;
src1.ne[0] = k;
src1.ne[1] = m;
src1.ne[2] = 1;
src1.ne[3] = 1;
src1.nb[0] = traits->type_size;
src1.nb[1] = k * traits->type_size;
src1.nb[2] = src1.nb[1];
src1.nb[3] = src1.nb[2];
src1.data = a;
struct ggml_tensor src0 = {};
src0.type = type;
src0.ne[0] = k;
src0.ne[1] = n;
src0.ne[2] = 1;
src0.ne[3] = 1;
src0.nb[0] = traits->type_size;
src0.nb[1] = k * traits->type_size;
src0.nb[2] = src0.nb[1];
src0.nb[3] = src0.nb[2];
src0.data = b;
struct ggml_tensor dst = {};
dst.ne[0] = n;
dst.ne[1] = m;
dst.ne[2] = 1;
dst.ne[3] = 1;
dst.nb[0] = sizeof(float);
dst.nb[1] = n * sizeof(float);
dst.nb[2] = dst.nb[1];
dst.nb[3] = dst.nb[2];
dst.data = c;
dst.src[0] = &src0;
dst.src[1] = &src1;
ggml_compute_forward_mul_mat(params, &dst);
}
// ggml_compute_forward_conv_2d
static void ggml_compute_forward_conv_2d_impl(const ggml_compute_params * params,
const ggml_tensor * kernel, // [KW, KH, IC, OC]
const ggml_tensor * src, // [W, H, C, N]
ggml_tensor * dst, // [OW, OH, OC, N]
ggml_type kernel_type) {
GGML_ASSERT(ggml_is_contiguous(kernel));
GGML_ASSERT(kernel_type == GGML_TYPE_F16 || kernel_type == GGML_TYPE_F32);
GGML_ASSERT(kernel->type == kernel_type);
const ggml_type_traits * traits = ggml_get_type_traits(kernel_type);
const int32_t stride_x = dst->op_params[0];
const int32_t stride_y = dst->op_params[1];
const int32_t pad_x = dst->op_params[2];
const int32_t pad_y = dst->op_params[3];
const int32_t dilation_x = dst->op_params[4];
const int32_t dilation_y = dst->op_params[5];
const int64_t c_in = src->ne[2];
const int64_t c_out = kernel->ne[3];
GGML_ASSERT(c_in == kernel->ne[2]);
const int64_t src_w = src->ne[0];
const int64_t src_h = src->ne[1];
const int64_t knl_w = kernel->ne[0];
const int64_t knl_h = kernel->ne[1];
const int64_t dst_w = dst->ne[0];
const int64_t dst_h = dst->ne[1];
const float * src_data = (float *) src->data;
void * knl_data = kernel->data;
float * dst_data = (float *) dst->data;
const int64_t knl_n = knl_w * knl_h * c_in;
const int64_t patch_total = dst->ne[3] * dst_w * dst_h;
const int64_t space_per_patch = knl_n * traits->type_size + c_out * sizeof(float);
const int64_t batch_size = params->wsize / space_per_patch;
const int64_t patches_per_batch = batch_size > 8 ? (batch_size / 8) * 8 : batch_size;
const int64_t batch_n = (patch_total + patches_per_batch - 1) / patches_per_batch;
GGML_ASSERT(patches_per_batch > 0 && batch_size >= 1);
void * tmp = params->wdata;
for (int64_t batch_i = 0; batch_i < batch_n; ++batch_i) {
const int64_t patch_start_batch = batch_i * patches_per_batch;
const int64_t patch_end_batch = std::min(patch_start_batch + patches_per_batch,
patch_total);
const int64_t patch_n = patch_end_batch - patch_start_batch;
const int64_t patch_per_thread = (patch_n + params->nth - 1) / params->nth;
const int64_t patch_start = patch_start_batch + params->ith * patch_per_thread;
const int64_t patch_end = std::min(patch_start + patch_per_thread, patch_end_batch);
//im2col for a patch
for (int64_t p = patch_start; p < patch_end; ++p) {
const int64_t batch_n = p / (dst_w * dst_h);
const int64_t src_x = (p / dst_w) % dst_h;
const int64_t src_y = p % dst_w;
const float * src_base = (const float *)((const char *)src_data + batch_n * src->nb[3]);
char * dst_row = (char *) tmp + (p % patches_per_batch) * knl_n * traits->type_size;
for (int64_t ic = 0; ic < c_in; ++ic) {
for (int64_t ky = 0; ky < knl_h; ++ky) {
for (int64_t kx = 0; kx < knl_w; ++kx) {
const int64_t sy = src_x * stride_y + ky * dilation_y - pad_y;
const int64_t sx = src_y * stride_x + kx * dilation_x - pad_x;
int64_t dst_idx = ic * (knl_h * knl_w) + ky * knl_w + kx;
float src_val;
if (sy < 0 || sy >= src_h || sx < 0 || sx >= src_w) {
src_val = 0.0f;
} else {
const float * src_ptr = (const float *)((const char *)src_base + sx * src->nb[0] + sy * src->nb[1] + ic * src->nb[2]);
src_val = *src_ptr;
}
char * element_ptr = dst_row + dst_idx * traits->type_size;
if (kernel_type == GGML_TYPE_F32) {
*(float *) element_ptr = src_val;
} else if (kernel_type == GGML_TYPE_F16) {
*(ggml_fp16_t *) element_ptr = GGML_CPU_FP32_TO_FP16(src_val);
}
}
}
}
} // patches handled by this thread
ggml_barrier(params->threadpool);
float * gemm_output = (float *) ((char *) tmp + patches_per_batch * knl_n * traits->type_size);
GGML_ASSERT(gemm_output + patch_n * c_out <= (float*)tmp + params->wsize);
// GEMM: patches[patch_n, knl_n] × kernel[knl_n, c_out] = output[patch_n, c_out]
ggml_call_mul_mat(kernel_type, params, patch_n, c_out, knl_n, tmp, knl_data, gemm_output);
ggml_barrier(params->threadpool);
//permute back [OC, N, OH, OW] to [N, OC, OH, OW]
const int64_t permute_per_thread = (patch_n + params->nth - 1) / params->nth;
const int64_t permute_start = params->ith * permute_per_thread;
const int64_t permute_end = std::min(permute_start + permute_per_thread, patch_n);
for (int64_t i = permute_start; i < permute_end; ++i) {
const int64_t p = patch_start_batch + i;
const int64_t batch_n = p / (dst_w * dst_h);
const int64_t dst_y = (p / dst_w) % dst_h;
const int64_t dst_x = p % dst_w;
for (int64_t oc = 0; oc < c_out; ++oc) {
const float value = gemm_output[i * c_out + oc];
float * dst_ptr = (float *)((char *)dst_data + dst_x * dst->nb[0] + dst_y * dst->nb[1] + oc * dst->nb[2] + batch_n * dst->nb[3]);
*dst_ptr = value;
}
}
}
}
void ggml_compute_forward_conv_2d(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
ggml_compute_forward_conv_2d_impl(params, src0, src1, dst, src0->type);
}
// ggml_compute_forward_conv_transpose_2d
void ggml_compute_forward_conv_transpose_2d(
@@ -7095,12 +7277,13 @@ static void ggml_compute_forward_upscale_f32(
GGML_TENSOR_UNARY_OP_LOCALS
const float sf0 = (float)ne0/src0->ne[0];
const float sf1 = (float)ne1/src0->ne[1];
const float sf2 = (float)ne2/src0->ne[2];
const float sf3 = (float)ne3/src0->ne[3];
float sf0 = (float)ne0/src0->ne[0];
float sf1 = (float)ne1/src0->ne[1];
float sf2 = (float)ne2/src0->ne[2];
float sf3 = (float)ne3/src0->ne[3];
const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
const int32_t mode_flags = ggml_get_op_params_i32(dst, 0);
const ggml_scale_mode mode = (ggml_scale_mode) (mode_flags & 0xFF);
if (mode == GGML_SCALE_MODE_NEAREST) {
for (int64_t i3 = 0; i3 < ne3; i3++) {
@@ -7121,8 +7304,12 @@ static void ggml_compute_forward_upscale_f32(
}
}
} else if (mode == GGML_SCALE_MODE_BILINEAR) {
// setting a pixel offset of 0 would replicate the behavior of pytorch interpolate with align_corners=True
const float pixel_offset = 0.5f;
float pixel_offset = 0.5f;
if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
pixel_offset = 0.0f;
sf0 = (float)(ne0 - 1) / (src0->ne[0] - 1);
sf1 = (float)(ne1 - 1) / (src0->ne[1] - 1);
}
for (int64_t i3 = 0; i3 < ne3; i3++) {
const int64_t i03 = i3 / sf3;
@@ -7580,7 +7767,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
ggml_type const k_vec_dot_type = ggml_get_type_traits_cpu(k->type)->vec_dot_type;
ggml_type const k_vec_dot_type = ggml_get_type_traits_cpu(k->type)->vec_dot_type;
ggml_from_float_t const q_to_vec_dot = ggml_get_type_traits_cpu(k_vec_dot_type)->from_float;
ggml_vec_dot_t const kq_vec_dot = ggml_get_type_traits_cpu(k->type)->vec_dot;
ggml_to_float_t const v_to_float = ggml_get_type_traits(v->type)->to_float;
@@ -7612,7 +7799,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
memset(VKQ32, 0, DV*sizeof(float));
}
const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL;
const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq3%mask->ne[2])*mask->nb[2]) : NULL;
// k indices
const int ik3 = iq3 / rk3;
+5
View File
@@ -20,6 +20,9 @@
static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
// Work buffer size for im2col operations in CONV2D
#define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
#ifdef __cplusplus
extern "C" {
#endif
@@ -65,6 +68,7 @@ void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struc
void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -107,6 +111,7 @@ void ggml_compute_forward_custom(const struct ggml_compute_params * params, stru
void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
+4 -1
View File
@@ -32,7 +32,9 @@ typedef void (* fattn_kernel_t)(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@@ -851,7 +853,8 @@ void launch_fattn(
scale, max_bias, m0, m1, n_head_log2, logit_softcap,
Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
K->ne[0], K->ne[1], K->ne[2], K->ne[3],
mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0,
mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0,
mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0,
Q->nb[1], Q->nb[2], Q->nb[3],
nb11, nb12, nb13,
nb21, nb22, nb23,
+8 -4
View File
@@ -1223,7 +1223,9 @@ static __global__ void flash_attn_ext_f16(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@@ -1288,7 +1290,8 @@ static __global__ void flash_attn_ext_f16(
const float2 * Q_f2 = (const float2 *) (Q + nb02* channel*ncols2);
const half2 * K_h2 = (const half2 *) (K + nb12*(channel*ncols2 / gqa_ratio));
const half2 * mask_h2 = ncols2 > 1 || mask ? (const half2 *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
const half2 * mask_h2 = ncols2 == 1 && !mask ? nullptr :
(const half2 *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
float2 * dstk = ((float2 *) dst) + channel*(ncols2 * DV/2);
const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@@ -1327,7 +1330,8 @@ static __global__ void flash_attn_ext_f16(
const float2 * Q_f2 = (const float2 *) (Q + nb02* channel*ncols2);
const half2 * K_h2 = (const half2 *) (K + nb12*(channel*ncols2 / gqa_ratio));
const half2 * mask_h2 = ncols2 > 1 || mask ? (const half2 *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
const half2 * mask_h2 = ncols2 == 1 && !mask ? nullptr :
(const half2 *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
float2 * dstk = ((float2 *) dst) + channel*(ncols2 * DV/2);
const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@@ -1348,8 +1352,8 @@ static __global__ void flash_attn_ext_f16(
GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+6 -4
View File
@@ -6,7 +6,7 @@
template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
__launch_bounds__(nwarps*WARP_SIZE, 2)
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f16(
const char * __restrict__ Q,
@@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f16(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@@ -64,7 +66,7 @@ static __global__ void flash_attn_tile_ext_f16(
const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.z + nb01*ic0);
const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.z / gqa_ratio));
const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) mask + ne11*ic0;
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const int stride_KV2 = nb11 / sizeof(half2);
@@ -288,8 +290,8 @@ static __global__ void flash_attn_tile_ext_f16(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+6 -4
View File
@@ -6,7 +6,7 @@
template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
__launch_bounds__(nwarps*WARP_SIZE, 2)
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f32(
const char * __restrict__ Q,
@@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f32(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@@ -58,8 +60,8 @@ static __global__ void flash_attn_tile_ext_f32(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@@ -76,7 +78,7 @@ static __global__ void flash_attn_tile_ext_f32(
const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.z + nb01*ic0);
const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.z / gqa_ratio));
const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) mask + ne11*ic0;
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const int stride_KV2 = nb11 / sizeof(half2);
+5 -3
View File
@@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f16(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@@ -68,7 +70,7 @@ static __global__ void flash_attn_vec_ext_f16(
K += nb12*(blockIdx.z / gqa_ratio);
V += nb22*(blockIdx.z / gqa_ratio);
const half * maskh = (const half *) mask + ne11*ic0;
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
const half slopeh = __float2half(slopef);
@@ -342,8 +344,8 @@ static __global__ void flash_attn_vec_ext_f16(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+6 -3
View File
@@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f32(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@@ -51,8 +53,8 @@ static __global__ void flash_attn_vec_ext_f32(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@@ -79,7 +81,8 @@ static __global__ void flash_attn_vec_ext_f32(
Q += nb02* blockIdx.z + nb01*ic0;
K += nb12*(blockIdx.z / gqa_ratio);
V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
const half * maskh = (const half *) mask + ne11*ic0;
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+8 -6
View File
@@ -46,7 +46,9 @@ static __global__ void flash_attn_ext_f16(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@@ -94,11 +96,11 @@ static __global__ void flash_attn_ext_f16(
constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
const float * Q_f = (const float *) (Q + nb02* blockIdx.z + nb01*ic0);
const half * K_h = (const half *) (K + nb12*(blockIdx.z / gqa_ratio));
const half * V_h = (const half *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) mask + (nb31/sizeof(half))* ic0;
const half2 * mask2 = (const half2 *) mask + (nb31/sizeof(half))*(ic0/2);
const float * Q_f = (const float *) (Q + nb02* blockIdx.z + nb01*ic0);
const half * K_h = (const half *) (K + nb12*(blockIdx.z / gqa_ratio));
const half * V_h = (const half *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const half2 * mask2 = (const half2 *) maskh;
const int stride_Q = nb01 / sizeof(float);
const int stride_KV = nb11 / sizeof(half);
@@ -440,7 +442,7 @@ static __global__ void flash_attn_ext_f16(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
GGML_UNUSED(ne31); GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+3
View File
@@ -3327,6 +3327,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_CONT:
return op->src[0]->type != GGML_TYPE_BF16;
case GGML_OP_DIAG_MASK_INF:
return true;
case GGML_OP_SOFT_MAX:
return true;
case GGML_OP_SOFT_MAX_BACK: {
@@ -3375,6 +3376,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
if (op->src[0]->ne[0] == 192) {
return false;
}
// TODO: support broadcast
// ref: https://github.com/ggml-org/llama.cpp/pull/14435
if (op->src[0]->ne[3] != 1) {
return false;
}
+91 -28
View File
@@ -13,6 +13,29 @@ __device__ float __forceinline__ t2f32<half>(half val) {
return __half2float(val);
}
struct soft_max_params {
int64_t nheads;
uint32_t n_head_log2;
int64_t ncols;
int64_t nrows_x;
int64_t nrows_y;
int64_t ne00;
int64_t ne01;
int64_t ne02;
int64_t ne03;
int64_t nb11;
int64_t nb12;
int64_t nb13;
int64_t ne12;
int64_t ne13;
float scale;
float max_bias;
float m0;
float m1;
};
// When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
// As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
#ifdef __clang__
@@ -21,16 +44,24 @@ __device__ float __forceinline__ t2f32<half>(half val) {
#endif // __clang__
template <bool use_shared, int ncols_template, int block_size_template, typename T>
static __global__ void soft_max_f32(
const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y,
const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
const float * x, const T * mask, float * dst, const soft_max_params p) {
const int ncols = ncols_template == 0 ? p.ncols : ncols_template;
const int tid = threadIdx.x;
const int rowx = blockIdx.x;
const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
const int64_t i03 = blockIdx.z;
const int64_t i02 = blockIdx.y;
const int64_t i01 = blockIdx.x;
//TODO: noncontigous inputs/outputs
const int rowx = blockIdx.x + blockIdx.y * gridDim.x + blockIdx.z * gridDim.x * gridDim.y;
const int64_t i11 = i01;
const int64_t i12 = i02 % p.ne12;
const int64_t i13 = i03 % p.ne13;
x += int64_t(rowx)*ncols;
mask += int64_t(rowy)*ncols * (mask != nullptr);
mask += (i11*p.nb11 + i12*p.nb12 + i13*p.nb13) / sizeof(T) * (mask != nullptr);
dst += int64_t(rowx)*ncols;
const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
@@ -38,7 +69,7 @@ static __global__ void soft_max_f32(
const int warp_id = threadIdx.x / WARP_SIZE;
const int lane_id = threadIdx.x % WARP_SIZE;
const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
const float slope = get_alibi_slope(p.max_bias, i02, p.n_head_log2, p.m0, p.m1);
extern __shared__ float data_soft_max_f32[];
float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
@@ -55,7 +86,7 @@ static __global__ void soft_max_f32(
break;
}
const float val = x[col]*scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
const float val = x[col]*p.scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
vals[col] = val;
max_val = max(max_val, val);
@@ -151,63 +182,60 @@ static __global__ void soft_max_back_f32(
}
template<typename T>
static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const soft_max_params & params, cudaStream_t stream) {
int nth = WARP_SIZE;
const int64_t ncols_x = params.ncols;
while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
const dim3 block_dims(nth, 1, 1);
const dim3 block_nums(nrows_x, 1, 1);
const dim3 block_nums(params.ne01, params.ne02, params.ne03);
const size_t nbytes_shared = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
const uint32_t n_head = nrows_x/nrows_y;
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
// FIXME: this limit could be raised by ~2-4x on Ampere or newer
if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
switch (ncols_x) {
case 32:
soft_max_f32<true, 32, 32><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
case 64:
soft_max_f32<true, 64, 64><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
case 128:
soft_max_f32<true, 128, 128><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
case 256:
soft_max_f32<true, 256, 256><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
case 512:
soft_max_f32<true, 512, 512><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
case 1024:
soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
case 2048:
soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
case 4096:
soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
default:
soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
(x, mask, dst, params);
break;
}
} else {
const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, params);
}
}
@@ -235,10 +263,11 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
const int64_t ne00 = src0->ne[0];
const int64_t nrows_x = ggml_nrows(src0);
const int64_t nrows_y = src0->ne[1];
const int64_t ne00 = src0->ne[0];
float scale = 1.0f;
float max_bias = 0.0f;
@@ -247,10 +276,44 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
const int64_t nb11 = src1 ? src1->nb[1] : 1;
const int64_t nb12 = src1 ? src1->nb[2] : 1;
const int64_t nb13 = src1 ? src1->nb[3] : 1;
const int64_t ne12 = src1 ? src1->ne[2] : 1;
const int64_t ne13 = src1 ? src1->ne[3] : 1;
const uint32_t n_head = src0->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
soft_max_params params = {};
params.nheads = src0->ne[2];
params.n_head_log2 = n_head_log2;
params.ncols = ne00;
params.nrows_x = nrows_x;
params.nrows_y = nrows_y;
params.ne00 = src0->ne[0];
params.ne01 = src0->ne[1];
params.ne02 = src0->ne[2];
params.ne03 = src0->ne[3];
params.nb11 = nb11;
params.nb12 = nb12;
params.nb13 = nb13;
params.ne12 = ne12;
params.ne13 = ne13;
params.scale = scale;
params.max_bias = max_bias;
params.m0 = m0;
params.m1 = m1;
if (use_f16) {
soft_max_f32_cuda(src0_d, (const half *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
soft_max_f32_cuda(src0_d, (const half *) src1_d, dst_d, params, stream);
} else {
soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, params, stream);
}
}
+4 -2
View File
@@ -71,7 +71,9 @@ else()
# note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
# note: unfortunately, we have to call it default.metallib instead of ggml.metallib
# ref: https://github.com/ggerganov/whisper.cpp/issues/1720
set(XC_FLAGS -fno-fast-math -fno-inline -g)
# note: adding -g causes segmentation fault during compile
#set(XC_FLAGS -fno-fast-math -fno-inline -g)
set(XC_FLAGS -fno-fast-math -fno-inline)
else()
set(XC_FLAGS -O3)
endif()
@@ -90,7 +92,7 @@ else()
add_custom_command(
OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
xcrun -sdk macosx metallib - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
xcrun -sdk macosx metallib - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
DEPENDS ggml-metal.metal ${METALLIB_COMMON}
+17 -3
View File
@@ -229,7 +229,9 @@ typedef struct {
uint64_t nb21;
uint64_t nb22;
uint64_t nb23;
int32_t ne32;
uint64_t nb31;
uint64_t nb32;
int32_t ne1;
int32_t ne2;
float scale;
@@ -461,9 +463,21 @@ typedef struct {
} ggml_metal_kargs_sum_rows;
typedef struct {
int64_t ne00;
int64_t ne01;
int64_t ne02;
int32_t ne00;
int32_t ne01;
int32_t ne02;
uint64_t nb01;
uint64_t nb02;
uint64_t nb03;
int32_t ne11;
int32_t ne12;
int32_t ne13;
uint64_t nb11;
uint64_t nb12;
uint64_t nb13;
uint64_t nb1;
uint64_t nb2;
uint64_t nb3;
float scale;
float max_bias;
float m0;
+17 -6
View File
@@ -1725,7 +1725,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_OP_MEAN:
case GGML_OP_SOFT_MAX:
case GGML_OP_GROUP_NORM:
return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
return has_simdgroup_reduction && ggml_is_contiguous_rows(op->src[0]);
case GGML_OP_RMS_NORM:
case GGML_OP_L2_NORM:
return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0]));
@@ -2644,10 +2644,7 @@ static bool ggml_metal_encode_node(
memcpy(&scale, ((const int32_t *) dst->op_params) + 0, sizeof(scale));
memcpy(&max_bias, ((const int32_t *) dst->op_params) + 1, sizeof(max_bias));
const int64_t nrows_x = ggml_nrows(src0);
const int64_t nrows_y = src0->ne[1];
const uint32_t n_head = nrows_x/nrows_y;
const uint32_t n_head = src0->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
@@ -2707,6 +2704,18 @@ static bool ggml_metal_encode_node(
/*.ne00 =*/ ne00,
/*.ne01 =*/ ne01,
/*.ne02 =*/ ne02,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.ne11 =*/ ne11,
/*.ne12 =*/ ne12,
/*.ne13 =*/ ne13,
/*.nb11 =*/ nb11,
/*.nb12 =*/ nb12,
/*.nb13 =*/ nb13,
/*.nb1 =*/ nb1,
/*.nb2 =*/ nb2,
/*.nb3 =*/ nb3,
/*.scale =*/ scale,
/*.max_bias =*/ max_bias,
/*.m0 =*/ m0,
@@ -2726,7 +2735,7 @@ static bool ggml_metal_encode_node(
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_DIAG_MASK_INF:
{
@@ -4979,7 +4988,9 @@ static bool ggml_metal_encode_node(
/*.nb21 =*/ nb21,
/*.nb22 =*/ nb22,
/*.nb23 =*/ nb23,
/*.ne32 =*/ ne32,
/*.nb31 =*/ nb31,
/*.nb32 =*/ nb32,
/*.ne1 =*/ ne1,
/*.ne2 =*/ ne2,
/*.scale =*/ scale,
+43 -32
View File
@@ -138,6 +138,7 @@ void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
}
void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
#pragma METAL fp math_mode(safe)
float min = FLT_MAX;
float max = -FLT_MAX;
@@ -203,6 +204,7 @@ void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
}
void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
#pragma METAL fp math_mode(safe)
float max = src[0];
float min = src[0];
@@ -239,6 +241,7 @@ void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
}
void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
#pragma METAL fp math_mode(safe)
float amax = 0.0f; // absolute max
float max = 0.0f;
@@ -1317,24 +1320,28 @@ kernel void kernel_soft_max(
device char * dst,
constant ggml_metal_kargs_soft_max & args,
threadgroup float * buf [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint ntg[[threads_per_threadgroup]]) {
const int64_t i03 = (tgpig) / (args.ne02*args.ne01);
const int64_t i02 = (tgpig - i03*args.ne02*args.ne01) / args.ne01;
const int64_t i01 = (tgpig - i03*args.ne02*args.ne01 - i02*args.ne01);
uint3 tptg[[threads_per_threadgroup]]) {
const int32_t i03 = tgpig.z;
const int32_t i02 = tgpig.y;
const int32_t i01 = tgpig.x;
device const float * psrc0 = (device const float *) src0 + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00);
device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*args.ne00 : nullptr;
device float * pdst = (device float *) dst + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00);
const int32_t i13 = i03%args.ne13;
const int32_t i12 = i02%args.ne12;
const int32_t i11 = i01;
device const float * psrc0 = (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
device const T * pmask = src1 != src0 ? (device const T * ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
device float * pdst = (device float *) (dst + i01*args.nb1 + i02*args.nb2 + i03*args.nb3);
float slope = 1.0f;
// ALiBi
if (args.max_bias > 0.0f) {
const int64_t h = i02;
const int32_t h = i02;
const float base = h < args.n_head_log2 ? args.m0 : args.m1;
const int exp = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
@@ -1345,13 +1352,13 @@ kernel void kernel_soft_max(
// parallel max
float lmax = -INFINITY;
for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
}
// find the max value in the block
float max_val = simd_max(lmax);
if (ntg > N_SIMDWIDTH) {
if (tptg.x > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = -INFINITY;
}
@@ -1370,7 +1377,7 @@ kernel void kernel_soft_max(
// parallel sum
float lsum = 0.0f;
for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
lsum += exp_psrc0;
pdst[i00] = exp_psrc0;
@@ -1382,7 +1389,7 @@ kernel void kernel_soft_max(
float sum = simd_sum(lsum);
if (ntg > N_SIMDWIDTH) {
if (tptg.x > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = 0.0f;
}
@@ -1401,7 +1408,7 @@ kernel void kernel_soft_max(
const float inv_sum = 1.0f/sum;
for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
pdst[i00] *= inv_sum;
}
}
@@ -1413,23 +1420,27 @@ kernel void kernel_soft_max_4(
device char * dst,
constant ggml_metal_kargs_soft_max & args,
threadgroup float * buf [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint ntg[[threads_per_threadgroup]]) {
const int64_t i03 = (tgpig) / (args.ne02*args.ne01);
const int64_t i02 = (tgpig - i03*args.ne02*args.ne01) / args.ne01;
const int64_t i01 = (tgpig - i03*args.ne02*args.ne01 - i02*args.ne01);
uint3 tptg[[threads_per_threadgroup]]) {
const int32_t i03 = tgpig.z;
const int32_t i02 = tgpig.y;
const int32_t i01 = tgpig.x;
device const float4 * psrc4 = (device const float4 *) src0 + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00)/4;
device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*args.ne00/4 : nullptr;
device float4 * pdst4 = (device float4 *) dst + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00)/4;
const int32_t i13 = i03%args.ne13;
const int32_t i12 = i02%args.ne12;
const int32_t i11 = i01;
device const float4 * psrc4 = (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
device const T * pmask = src1 != src0 ? (device const T * ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
device float4 * pdst4 = (device float4 *) (dst + i01*args.nb1 + i02*args.nb2 + i03*args.nb3);
float slope = 1.0f;
if (args.max_bias > 0.0f) {
const int64_t h = i02;
const int32_t h = i02;
const float base = h < args.n_head_log2 ? args.m0 : args.m1;
const int exp = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
@@ -1440,14 +1451,14 @@ kernel void kernel_soft_max_4(
// parallel max
float4 lmax4 = -INFINITY;
for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
}
const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
float max_val = simd_max(lmax);
if (ntg > N_SIMDWIDTH) {
if (tptg.x > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = -INFINITY;
}
@@ -1466,7 +1477,7 @@ kernel void kernel_soft_max_4(
// parallel sum
float4 lsum4 = 0.0f;
for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
lsum4 += exp_psrc4;
pdst4[i00] = exp_psrc4;
@@ -1480,7 +1491,7 @@ kernel void kernel_soft_max_4(
float sum = simd_sum(lsum);
if (ntg > N_SIMDWIDTH) {
if (tptg.x > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = 0.0f;
}
@@ -1499,7 +1510,7 @@ kernel void kernel_soft_max_4(
const float inv_sum = 1.0f/sum;
for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
pdst4[i00] *= inv_sum;
}
}
@@ -3773,7 +3784,7 @@ kernel void kernel_flash_attn_ext(
// load the mask in shared memory
#pragma unroll(Q)
for (short j = 0; j < Q; ++j) {
device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31);
device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31 + (iq3%args.ne32)*args.nb32);
const float m = pm[ic + tiisg];
@@ -4259,7 +4270,7 @@ kernel void kernel_flash_attn_ext_vec(
const bool has_mask = mask != q;
// pointer to the mask
device const half * pm = (device const half *) (mask + iq1*args.nb31);
device const half * pm = (device const half *) (mask + iq1*args.nb31 + (iq3%args.ne32)*args.nb32);
float slope = 1.0f;
+1
View File
@@ -65,6 +65,7 @@ set(GGML_OPENCL_KERNELS
gemv_noshuffle_general
gemv_noshuffle
get_rows
glu
group_norm
im2col_f32
im2col_f16
+160 -24
View File
@@ -351,6 +351,7 @@ struct ggml_backend_opencl_context {
cl_program program_gemv_noshuffle_general;
cl_program program_gemv_noshuffle;
cl_program program_get_rows;
cl_program program_glu;
cl_program program_im2col_f16;
cl_program program_im2col_f32;
cl_program program_mul_mat_Ab_Bi_8x4;
@@ -401,6 +402,8 @@ struct ggml_backend_opencl_context {
cl_kernel kernel_relu;
cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
cl_kernel kernel_clamp;
cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu,
kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16;
cl_kernel kernel_norm;
cl_kernel kernel_rms_norm;
cl_kernel kernel_group_norm;
@@ -738,6 +741,27 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
GGML_LOG_CONT(".");
}
// glu
{
#ifdef GGML_OPENCL_EMBED_KERNELS
const std::string kernel_src {
#include "glu.cl.h"
};
#else
const std::string kernel_src = read_file("glu.cl");
#endif
backend_ctx->program_glu =
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
CL_CHECK((backend_ctx->kernel_geglu = clCreateKernel(backend_ctx->program_glu, "kernel_geglu", &err), err));
CL_CHECK((backend_ctx->kernel_reglu = clCreateKernel(backend_ctx->program_glu, "kernel_reglu", &err), err));
CL_CHECK((backend_ctx->kernel_swiglu = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu", &err), err));
CL_CHECK((backend_ctx->kernel_geglu_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_f16", &err), err));
CL_CHECK((backend_ctx->kernel_reglu_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_reglu_f16", &err), err));
CL_CHECK((backend_ctx->kernel_swiglu_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_f16", &err), err));
GGML_LOG_CONT(".");
}
// get_rows
{
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2163,7 +2187,7 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
// dependencies.
sync_with_other_backends(backend);
if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue;
}
@@ -2242,6 +2266,15 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
default:
return false;
}
case GGML_OP_GLU:
switch (ggml_get_glu_op(op)) {
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_SWIGLU:
return ggml_is_contiguous_1(op->src[0]) && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
default:
return false;
}
case GGML_OP_CLAMP:
return op->src[0]->type == GGML_TYPE_F32;
case GGML_OP_SOFT_MAX:
@@ -3166,7 +3199,7 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
// Open file and dump.
char fname[512];
sprintf(fname, "./tensor-dumps/%s.txt", tensor->name);
snprintf(fname, sizeof(fname), "./tensor-dumps/%s.txt", tensor->name);
FILE * f = fopen(fname, "w");
if (!f) {
printf("Failed to open %s\n", fname);
@@ -4420,7 +4453,8 @@ static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, gg
ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
const int mode_flags = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
const ggml_scale_mode mode = (ggml_scale_mode) (mode_flags & 0xFF);
cl_kernel kernel = nullptr;
if (mode == GGML_SCALE_MODE_NEAREST) {
@@ -4451,18 +4485,22 @@ static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, gg
const cl_ulong nb02 = src0->nb[2];
const cl_ulong nb03 = src0->nb[3];
const int ne00_src = src0->ne[0];
const int ne01_src = src0->ne[1];
const int ne00 = src0->ne[0];
const int ne01 = src0->ne[1];
const int ne02 = src0->ne[2];
const int ne03 = src0->ne[3];
const int ne10_dst = dst->ne[0];
const int ne11_dst = dst->ne[1];
const int ne12_dst = dst->ne[2];
const int ne13_dst = dst->ne[3];
const int ne0 = dst->ne[0];
const int ne1 = dst->ne[1];
const int ne2 = dst->ne[2];
const int ne3 = dst->ne[3];
const float sf0 = (float)dst->ne[0] / src0->ne[0];
const float sf1 = (float)dst->ne[1] / src0->ne[1];
const float sf2 = (float)dst->ne[2] / src0->ne[2];
const float sf3 = (float)dst->ne[3] / src0->ne[3];
float sf0 = (float)ne0 / ne00;
float sf1 = (float)ne1 / ne01;
float sf2 = (float)ne2 / ne02;
float sf3 = (float)ne3 / ne03;
float pixel_offset = 0.5f;
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra_src0->data_device));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &off_src0));
@@ -4474,29 +4512,36 @@ static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, gg
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &nb03));
if (mode == GGML_SCALE_MODE_NEAREST) {
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne10_dst));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne11_dst));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne12_dst));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne13_dst));
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne0));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne1));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne2));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne3));
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float), &sf0));
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float), &sf1));
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float), &sf2));
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float), &sf3));
} else if (mode == GGML_SCALE_MODE_BILINEAR) {
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne00_src));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne01_src));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne10_dst));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne11_dst));
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int), &ne12_dst));
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int), &ne13_dst));
if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
sf0 = (float)(ne0 - 1) / (ne00 - 1);
sf1 = (float)(ne1 - 1) / (ne01 - 1);
pixel_offset = 0.0f;
}
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne0));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne1));
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int), &ne2));
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int), &ne3));
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float), &sf0));
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float), &sf1));
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(float), &sf2));
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(float), &sf3));
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float), &pixel_offset));
}
size_t dst_total_elements = (size_t)ne10_dst * ne11_dst * ne12_dst * ne13_dst;
size_t dst_total_elements = (size_t)ne0 * ne1 * ne2 * ne3;
if (dst_total_elements == 0) {
return;
}
@@ -6143,6 +6188,91 @@ static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, c
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
}
static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src0);
GGML_ASSERT(src0->extra);
GGML_ASSERT(dst);
GGML_ASSERT(dst->extra);
GGML_ASSERT(ggml_is_contiguous_1(src0));
if (src1) {
GGML_ASSERT(src1);
GGML_ASSERT(src1->extra);
GGML_ASSERT(ggml_are_same_shape(src0, src1));
}
ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
cl_kernel kernel;
switch (ggml_get_glu_op(dst)) {
case GGML_GLU_OP_GEGLU:
if (dst->type == GGML_TYPE_F32) {
kernel = backend_ctx->kernel_geglu;
} else {
kernel = backend_ctx->kernel_geglu_f16;
}
break;
case GGML_GLU_OP_REGLU:
if (dst->type == GGML_TYPE_F32) {
kernel = backend_ctx->kernel_reglu;
} else {
kernel = backend_ctx->kernel_reglu_f16;
}
break;
case GGML_GLU_OP_SWIGLU:
if (dst->type == GGML_TYPE_F32) {
kernel = backend_ctx->kernel_swiglu;
} else {
kernel = backend_ctx->kernel_swiglu_f16;
}
break;
default:
GGML_ABORT("Unsupported glu op");
}
ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
ggml_tensor_extra_cl * extra1 = src1 ? (ggml_tensor_extra_cl *)src1->extra : nullptr;
cl_ulong offset0 = extra0->offset + src0->view_offs;
cl_ulong offsetd = extrad->offset + dst->view_offs;
cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
const int ne0 = dst->ne[0];
const cl_ulong nb01 = src0->nb[1];
const cl_ulong nb11 = src1 ? src1->nb[1] : nb01;
const cl_ulong nb1 = dst->nb[1];
const int swp = ((const int32_t *) dst->op_params)[1];
const int ne00_off = src1 ? 0 : (swp ? ne0 : 0);
const int ne10_off = src1 ? 0 : (swp ? 0 : ne0);
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), src1 ? &extra1->data_device : &extra0->data_device));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extrad->data_device));
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &nb01));
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &nb11));
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne0));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb1));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne00_off));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne10_off));
const size_t nrows = ggml_nrows(src0);
size_t nth = 512;
size_t global_work_size[] = {nrows*nth, 1, 1};
size_t local_work_size[] = {nth, 1, 1};
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
}
//------------------------------------------------------------------------------
// Op offloading
//------------------------------------------------------------------------------
@@ -6244,6 +6374,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
default:
return false;
} break;
case GGML_OP_GLU:
if (!any_on_device) {
return false;
}
func = ggml_cl_glu;
break;
case GGML_OP_CLAMP:
if (!any_on_device) {
return false;
+201
View File
@@ -0,0 +1,201 @@
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#define GELU_COEF_A 0.044715f
#define SQRT_2_OVER_PI 0.79788456080286535587989211986876f
//------------------------------------------------------------------------------
// geglu
//------------------------------------------------------------------------------
kernel void kernel_geglu(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb01,
ulong nb11,
int ne0,
ulong nb1,
int ne00_off,
int ne10_off
) {
src0 = (global char*)((global char*)src0 + offset0);
src1 = (global char*)((global char*)src1 + offset1);
dst = (global char*)((global char*)dst + offsetd);
global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1);
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const float x0 = src0_row[i0];
const float x1 = src1_row[i0];
const float gelu = 0.5f*x0*(1.0f + tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
dst_row[i0] = gelu*x1;
}
}
kernel void kernel_geglu_f16(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb01,
ulong nb11,
int ne0,
ulong nb1,
int ne00_off,
int ne10_off
) {
src0 = (global char*)((global char*)src0 + offset0);
src1 = (global char*)((global char*)src1 + offset1);
dst = (global char*)((global char*)dst + offsetd);
global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
global half * dst_row = (global half *) ((global char *) dst + get_group_id(0)*nb1);
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const half x0 = src0_row[i0];
const half x1 = src1_row[i0];
const half gelu = 0.5f*x0*(1.0f + tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
dst_row[i0] = gelu*x1;
}
}
//------------------------------------------------------------------------------
// reglu
//------------------------------------------------------------------------------
kernel void kernel_reglu(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb01,
ulong nb11,
int ne0,
ulong nb1,
int ne00_off,
int ne10_off
) {
src0 = (global char*)((global char*)src0 + offset0);
src1 = (global char*)((global char*)src1 + offset1);
dst = (global char*)((global char*)dst + offsetd);
global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1);
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const float x0 = src0_row[i0];
const float x1 = src1_row[i0];
dst_row[i0] = x0*x1*(x0 > 0.0f);
}
}
kernel void kernel_reglu_f16(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb01,
ulong nb11,
int ne0,
ulong nb1,
int ne00_off,
int ne10_off
) {
src0 = (global char*)((global char*)src0 + offset0);
src1 = (global char*)((global char*)src1 + offset1);
dst = (global char*)((global char*)dst + offsetd);
global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
global half * dst_row = (global half *) ((global char *) dst + get_group_id(0)*nb1);
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const half x0 = src0_row[i0];
const half x1 = src1_row[i0];
dst_row[i0] = x0*x1*(x0 > 0.0f);
}
}
//------------------------------------------------------------------------------
// swiglu
//------------------------------------------------------------------------------
kernel void kernel_swiglu(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb01,
ulong nb11,
int ne0,
ulong nb1,
int ne00_off,
int ne10_off
) {
src0 = (global char*)((global char*)src0 + offset0);
src1 = (global char*)((global char*)src1 + offset1);
dst = (global char*)((global char*)dst + offsetd);
global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1);
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const float x0 = src0_row[i0];
const float x1 = src1_row[i0];
const float silu = x0 / (1.0f + exp(-x0));
dst_row[i0] = silu*x1;
}
}
kernel void kernel_swiglu_f16(
global char * src0,
ulong offset0,
global char * src1,
ulong offset1,
global char * dst,
ulong offsetd,
ulong nb01,
ulong nb11,
int ne0,
ulong nb1,
int ne00_off,
int ne10_off
) {
src0 = (global char*)((global char*)src0 + offset0);
src1 = (global char*)((global char*)src1 + offset1);
dst = (global char*)((global char*)dst + offsetd);
global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
global half * dst_row = (global half *) ((global char *) dst + get_group_id(0)*nb1);
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
const half x0 = src0_row[i0];
const half x1 = src1_row[i0];
const half silu = x0 / (1.0f + exp(-x0));
dst_row[i0] = silu*x1;
}
}
+2 -3
View File
@@ -60,7 +60,8 @@ kernel void kernel_upscale_bilinear(
float sf0,
float sf1,
float sf2,
float sf3
float sf3,
float pixel_offset
) {
global const char * src_base = (global const char *)p_src0 + off_src0;
global float * dst_base = (global float *)((global char *)p_dst + off_dst);
@@ -80,8 +81,6 @@ kernel void kernel_upscale_bilinear(
int i02_src = (int)(i12_dst / sf2);
int i03_src = (int)(i13_dst / sf3);
const float pixel_offset = 0.5f;
float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
long y0_src = (long)floor(y_src_f);
long y1_src = y0_src + 1;
+6 -6
View File
@@ -568,14 +568,14 @@ static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, co
}
float iscale = nmax/(max - min);
float scale = 1/iscale;
float best_mad = 0;
float best_error = 0;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale*(x[i] - min));
L[i] = MAX(0, MIN(nmax, l));
float diff = scale * L[i] + min - x[i];
diff = use_mad ? fabsf(diff) : diff * diff;
float w = weights[i];
best_mad += w * diff;
best_error += w * diff;
}
if (nstep < 1) {
*the_min = -min;
@@ -601,18 +601,18 @@ static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, co
this_min = 0;
this_scale = sum_xl / sum_l2;
}
float mad = 0;
float cur_error = 0;
for (int i = 0; i < n; ++i) {
float diff = this_scale * Laux[i] + this_min - x[i];
diff = use_mad ? fabsf(diff) : diff * diff;
float w = weights[i];
mad += w * diff;
cur_error += w * diff;
}
if (mad < best_mad) {
if (cur_error < best_error) {
for (int i = 0; i < n; ++i) {
L[i] = Laux[i];
}
best_mad = mad;
best_error = cur_error;
scale = this_scale;
min = this_min;
}
+8 -2
View File
@@ -4395,9 +4395,15 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
return true;
case GGML_OP_CONT:
return op->src[0]->type != GGML_TYPE_BF16;
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
return true;
// TODO: support batching
if (op->src[0]->ne[3] != 1) {
return false;
}
// TODO: support broadcast
// ref: https://github.com/ggml-org/llama.cpp/pull/14435
return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_ROPE:
case GGML_OP_IM2COL:
return true;
+73 -15
View File
@@ -431,6 +431,7 @@ struct vk_device_struct {
// [src/dst 0=fp32,1=fp16]
vk_pipeline pipeline_gelu[2];
vk_pipeline pipeline_gelu_erf[2];
vk_pipeline pipeline_gelu_quick[2];
vk_pipeline pipeline_silu[2];
vk_pipeline pipeline_relu[2];
@@ -632,6 +633,7 @@ struct vk_flash_attn_push_constants {
uint32_t nev2;
uint32_t nev3;
uint32_t nem1;
uint32_t nem2;
uint32_t nb01;
uint32_t nb02;
@@ -642,7 +644,6 @@ struct vk_flash_attn_push_constants {
uint32_t nb21;
uint32_t nb22;
uint32_t nb23;
uint32_t nb31;
float scale;
float max_bias;
@@ -657,6 +658,7 @@ struct vk_flash_attn_push_constants {
uint32_t split_kv;
uint32_t k_num;
};
static_assert(sizeof(vk_flash_attn_push_constants) <= 128, "sizeof(vk_flash_attn_push_constants) must be <= 128");
struct vk_op_push_constants {
uint32_t KX;
@@ -755,6 +757,14 @@ struct vk_op_rope_push_constants {
struct vk_op_soft_max_push_constants {
uint32_t KX;
uint32_t KY;
uint32_t ne00;
uint32_t ne01;
uint32_t ne02;
uint32_t ne12;
uint32_t ne13;
uint32_t nb11;
uint32_t nb12;
uint32_t nb13;
float scale;
float max_bias;
float m0;
@@ -2761,6 +2771,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
CREATE_UNARY(gelu)
CREATE_UNARY(gelu_erf)
CREATE_UNARY(gelu_quick)
CREATE_UNARY(silu)
CREATE_UNARY(relu)
@@ -5964,7 +5975,30 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
if (src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
} else {
ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
// Split based on number of ids, to fit in shared memory
const uint32_t nei0 = (uint32_t)src2->ne[0];
const uint32_t nei1 = (uint32_t)src2->ne[1];
GGML_ASSERT(nei0 <= 4096);
const uint32_t split_size = std::min(nei1, 4096u / nei0);
ggml_tensor src1_copy = *src1;
ggml_tensor src2_copy = *src2;
ggml_tensor dst_copy = *dst;
for (uint32_t token_start = 0; token_start < nei1; token_start += split_size) {
const uint32_t n_tokens = std::min(split_size, nei1 - token_start);
src1_copy.view_offs = src1->view_offs + token_start * src1_copy.nb[2];
src2_copy.view_offs = src2->view_offs + token_start * src2_copy.nb[1];
dst_copy.view_offs = dst->view_offs + token_start * dst_copy.nb[2];
src1_copy.ne[2] = n_tokens;
src2_copy.ne[1] = n_tokens;
dst_copy.ne[2] = n_tokens;
ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, &src1_copy, &src2_copy, &dst_copy, dryrun);
}
}
}
@@ -6015,7 +6049,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
GGML_TENSOR_LOCALS(size_t, nb, dst, nb)
const uint32_t nem1 = mask ? mask->ne[1] : 0;
const uint32_t nbm1 = mask ? mask->nb[1] : 0;
const uint32_t nem2 = mask ? mask->ne[2] : 0;
const uint32_t D = neq0;
uint32_t N = neq1;
@@ -6178,7 +6212,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
// Try to use split_k when KV is large enough to be worth the overhead
if (workgroups_x == 1 && shader_core_count > 0 && KV >= 512) {
// Try to run two workgroups per SM.
split_k = ctx->device->shader_core_count * 2 / workgroups_y;
split_k = ctx->device->shader_core_count * 2 / (workgroups_y * workgroups_z);
if (split_k > 1) {
// Try to evenly split KV into split_k chunks, but it needs to be a multiple
// of "align", so recompute split_k based on that.
@@ -6188,9 +6222,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
}
}
// Reserve space for split_k temporaries. For each split, we need to store the O matrix (D x ne1)
// and the per-row m and L values (ne1 rows).
const uint64_t split_k_size = split_k > 1 ? (D * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k : 0;
// Reserve space for split_k temporaries. For each split x batch, we need to store the O matrix (D x ne1)
// and the per-row m and L values (ne1 rows). We store all the matrices first, followed by the rows.
const uint64_t split_k_size = split_k > 1 ? (D * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne3 : 0;
if (split_k_size > ctx->device->max_memory_allocation_size) {
GGML_ABORT("Requested preallocation size is too large");
}
@@ -6282,11 +6316,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
(uint32_t)neq2, (uint32_t)neq3,
(uint32_t)nek2, (uint32_t)nek3,
(uint32_t)nev2, (uint32_t)nev3,
nem1,
nem1, nem2,
q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
nbm1,
scale, max_bias, logit_softcap,
mask != nullptr, n_head_log2, m0, m1,
gqa_ratio, split_kv, split_k };
@@ -6309,13 +6342,13 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
ggml_vk_sync_buffers(subctx);
const std::array<uint32_t, 3> pc2 = { D, (uint32_t)ne1, split_k };
const std::array<uint32_t, 4> pc2 = { D, (uint32_t)ne1, (uint32_t)ne3, split_k };
ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
{
vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE},
vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
},
pc2, { (uint32_t)ne1, 1, 1 });
pc2, { (uint32_t)ne1, 1, (uint32_t)ne3 });
} else {
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
{
@@ -6481,6 +6514,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_silu[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_GELU:
return ctx->device->pipeline_gelu[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_GELU_ERF:
return ctx->device->pipeline_gelu_erf[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_GELU_QUICK:
return ctx->device->pipeline_gelu_quick[dst->type == GGML_TYPE_F16];
case GGML_UNARY_OP_RELU:
@@ -7639,7 +7674,13 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
const uint32_t nrows_x = (uint32_t)ggml_nrows(src0);
const uint32_t nrows_y = (uint32_t)src0->ne[1];
const uint32_t n_head_kv = nrows_x/nrows_y;
const uint32_t ne12 = src1 ? (uint32_t)(src1->ne[2]) : 0u;
const uint32_t ne13 = src1 ? (uint32_t)(src1->ne[3]) : 0u;
const uint32_t nb11 = src1 ? (uint32_t)(src1->nb[1] / src1->nb[0]) : 0u;
const uint32_t nb12 = src1 ? (uint32_t)(src1->nb[2] / src1->nb[0]) : 0u;
const uint32_t nb13 = src1 ? (uint32_t)(src1->nb[3] / src1->nb[0]) : 0u;
const uint32_t n_head_kv = src0->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
@@ -7648,6 +7689,9 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX, {
ncols,
src1 != nullptr ? nrows_y : (uint32_t)0,
(uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],
ne12, ne13,
nb11, nb12, nb13,
scale, max_bias,
m0, m1,
n_head_log2,
@@ -8827,6 +8871,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
switch (ggml_get_unary_op(node)) {
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_GELU_ERF:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_TANH:
@@ -9072,6 +9117,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
switch (ggml_get_unary_op(node)) {
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_GELU_ERF:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_TANH:
@@ -9289,6 +9335,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
switch (ggml_get_unary_op(tensor)) {
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_GELU_ERF:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_TANH:
@@ -10095,6 +10142,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_GELU_ERF:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_RELU:
@@ -10127,9 +10175,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
ggml_type src0_type = op->src[0]->type;
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
const vk_device& device = ggml_vk_get_device(ctx->device);
if (op->op == GGML_OP_MUL_MAT_ID && !device->mul_mat_id_s[src0_type] && !device->mul_mat_id_m[src0_type] && !device->mul_mat_id_l[src0_type]) {
// If there's not enough shared memory for row_ids and the result tile, fallback to CPU
return false;
if (op->op == GGML_OP_MUL_MAT_ID) {
if (!device->mul_mat_id_s[src0_type] && !device->mul_mat_id_m[src0_type] && !device->mul_mat_id_l[src0_type]) {
// If there's not enough shared memory for row_ids and the result tile, fallback to CPU
return false;
}
// Check against size of shared memory variable
if (op->src[2]->ne[0] > 4096) {
return false;
}
}
switch (src0_type) {
case GGML_TYPE_F32:
@@ -10369,6 +10423,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_SCALE:
case GGML_OP_PAD:
case GGML_OP_DIAG_MASK_INF:
return true;
case GGML_OP_SOFT_MAX:
case GGML_OP_SOFT_MAX_BACK:
case GGML_OP_ARGSORT:
@@ -10835,6 +10890,9 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
case GGML_UNARY_OP_GELU:
tensor_clone = ggml_gelu(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_GELU_ERF:
tensor_clone = ggml_gelu_erf(ggml_ctx, src_clone[0]);
break;
case GGML_UNARY_OP_GELU_QUICK:
tensor_clone = ggml_gelu_quick(ggml_ctx, src_clone[0]);
break;
@@ -99,6 +99,10 @@ void main() {
uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
#endif
uint32_t m_offset = 0;
if (p.nem2 != 1) {
m_offset = (iq3 % p.nem2) * p.nem1 * KV;
}
[[dont_unroll]]
for (uint32_t j = start_j; j < end_j; ++j) {
@@ -150,7 +154,7 @@ void main() {
uint32_t c = (idx + tid) % Bc;
uint32_t r = (idx + tid) / Bc;
if (idx + tid < Bc * Br) {
masksh[c][r] = float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]);
masksh[c][r] = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
}
}
barrier();
@@ -277,7 +281,7 @@ void main() {
// If there is split_k, then the split_k resolve shader does the final
// division by L. Store the intermediate O value and per-row m and L values.
if (p.k_num > 1) {
uint32_t o_offset = D * p.ne1 * split_k_index;
uint32_t o_offset = D * p.ne1 * (split_k_index + iq3 * p.k_num);
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
if (r < N) {
@@ -289,7 +293,7 @@ void main() {
}
}
o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
o_offset = D * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
if (r < N) {
perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -311,7 +315,7 @@ void main() {
}
}
uint32_t o_offset = iq3*p.ne2*p.ne1;
uint32_t o_offset = iq3*p.ne2*p.ne1*D;
if (p.gqa_ratio > 1) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
@@ -24,6 +24,7 @@ layout (push_constant) uniform parameter {
uint32_t nev2;
uint32_t nev3;
uint32_t nem1;
uint32_t nem2;
uint32_t nb01;
uint32_t nb02;
@@ -34,7 +35,6 @@ layout (push_constant) uniform parameter {
uint32_t nb21;
uint32_t nb22;
uint32_t nb23;
uint32_t nb31;
float scale;
float max_bias;
@@ -123,6 +123,10 @@ void main() {
uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
#endif
uint32_t m_offset = 0;
if (p.nem2 != 1) {
m_offset = (iq3 % p.nem2) * p.nem1 * KV;
}
[[dont_unroll]]
for (uint32_t j = start_j; j < end_j; ++j) {
@@ -181,7 +185,7 @@ void main() {
uint32_t c = (idx + tid) % Bc;
uint32_t r = (idx + tid) / Bc;
if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]));
sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]));
}
}
barrier();
@@ -300,7 +304,7 @@ void main() {
// If there is split_k, then the split_k resolve shader does the final
// division by L. Store the intermediate O value and per-row m and L values.
if (p.k_num > 1) {
uint32_t o_offset = D * p.ne1 * split_k_index;
uint32_t o_offset = D * p.ne1 * (split_k_index + iq3 * p.k_num);
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
if (tile_row(r) < N) {
@@ -312,7 +316,7 @@ void main() {
}
}
o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
o_offset = D * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
if (tile_row(r) < N) {
perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -334,7 +338,7 @@ void main() {
}
}
uint32_t o_offset = iq3*p.ne2*p.ne1;
uint32_t o_offset = iq3*p.ne2*p.ne1*D;
if (p.gqa_ratio > 1) {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -130,6 +130,11 @@ void main() {
coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
}
uint32_t m_offset = 0;
if (p.nem2 != 1) {
m_offset = (iq3 % p.nem2) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
}
[[dont_unroll]]
for (uint32_t j = start_j; j < end_j; ++j) {
@@ -155,7 +160,7 @@ void main() {
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
coopMatLoadTensorNV(mv, data_m, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
}
@@ -229,10 +234,10 @@ void main() {
if (p.k_num > 1) {
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
uint32_t o_offset = D * p.ne1 * split_k_index;
uint32_t o_offset = D * p.ne1 * (split_k_index + iq3 * p.k_num);
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
o_offset = D * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
return;
@@ -250,7 +255,7 @@ void main() {
O = Ldiag*O;
uint32_t o_offset = iq3*p.ne2*p.ne1;
uint32_t o_offset = iq3*p.ne2*p.ne1*D;
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
if (p.gqa_ratio > 1) {
@@ -12,6 +12,7 @@ layout (binding = 1) writeonly buffer D {float data_d[];};
layout (push_constant) uniform parameter {
uint D;
uint N;
uint ne3;
uint k_num;
} p;
@@ -19,13 +20,14 @@ void main() {
// Each workgroup handles a row
const uint n = gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
const uint iq3 = gl_WorkGroupID.z;
uint D = p.D;
uint N = p.N;
uint k_num = p.k_num;
uint l_offset = D * N * k_num + n;
uint m_offset = D * N * k_num + N + n;
uint l_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + n;
uint m_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + N + n;
uint lm_stride = N * 2;
// Compute the max m value for the row
@@ -49,11 +51,11 @@ void main() {
for (uint d = tid; d < D; d += BLOCK_SIZE) {
float O = 0.0;
[[unroll]] for (uint k = 0; k < k_num; ++k) {
uint o_offset = D * N * k + D * n + d;
uint o_offset = D * N * (k + iq3 * k_num) + D * n + d;
float m = data_a[m_offset + k * lm_stride];
O += exp(m - m_max) * data_a[o_offset];
}
O *= L;
data_d[D * n + d] = O;
data_d[iq3 * D * N + D * n + d] = O;
}
}
@@ -0,0 +1,39 @@
#version 450
#include "generic_head.comp"
#include "types.comp"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
void main() {
// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
// ref: https://www.johndcook.com/blog/python_erf/
const float p_erf = 0.3275911f;
const float a1_erf = 0.254829592f;
const float a2_erf = -0.284496736f;
const float a3_erf = 1.421413741f;
const float a4_erf = -1.453152027f;
const float a5_erf = 1.061405429f;
const float SQRT_2_INV = 0.70710678118654752440084436210484f;
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (i >= p.KX) {
return;
}
const float a = float(data_a[i]);
const float a_div_sqr2 = a * SQRT_2_INV;
const float sign_x = sign(a_div_sqr2);
const float x = abs(a_div_sqr2);
const float t = 1.0f / (1.0f + p_erf * x);
const float y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
const float erf_approx = sign_x * y;
data_d[i] = D_TYPE(0.5f * a * (1.0f + erf_approx));
}
@@ -6,6 +6,14 @@ layout (push_constant) uniform parameter
{
uint KX;
uint KY;
uint ne00;
uint ne01;
uint ne02;
uint ne12;
uint ne13;
uint nb11;
uint nb12;
uint nb13;
float scale;
float max_bias;
float m0;
@@ -31,7 +39,15 @@ shared FLOAT_TYPE vals[BLOCK_SIZE];
void soft_max(uint num_iters) {
const uint tid = gl_LocalInvocationID.x;
const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
const uint rowy = (p.KY > 0) ? (rowx % p.KY) : 0;
const uint32_t i03 = rowx / (p.ne01 * p.ne02);
const uint32_t i02 = (rowx - i03 * p.ne01 * p.ne02) / p.ne01;
const uint32_t i01 = rowx % p.ne01;
uint rowy_start = 0;
if (p.KY > 0) {
rowy_start = i01 * p.nb11 + (i02 % p.ne12) * p.nb12 + (i03 % p.ne13) * p.nb13;
}
if (rowx >= p.nrows_x) {
return;
@@ -41,7 +57,7 @@ void soft_max(uint num_iters) {
// ALiBi
if (p.max_bias > 0.0f) {
const uint h = rowx/p.KY; // head index
const uint h = (rowx / p.ne01) % p.ne02; // head index
const float base = h < p.n_head_log2 ? p.m0 : p.m1;
const uint exp = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
@@ -67,7 +83,7 @@ void soft_max(uint num_iters) {
FLOAT_TYPE b = FLOAT_TYPE(0);
if (p.KY > 0 && col < p.KX) {
b = data_b[rowy * p.KX + col];
b = data_b[rowy_start + col];
}
FLOAT_TYPE v = a * p.scale + slope * b;
@@ -111,7 +127,7 @@ void soft_max(uint num_iters) {
if (idx < DATA_CACHE_SIZE) {
val = exp(data_cache[idx] - max_val);
} else {
val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy_start + col]) : FLOAT_TYPE(0.0f)) - max_val);
}
sum += val;
if (idx < DATA_CACHE_SIZE) {
@@ -574,6 +574,8 @@ void process_shaders() {
string_to_spv("gelu_f16", "gelu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("gelu_erf_f16", "gelu_erf.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("gelu_erf_f32", "gelu_erf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("gelu_quick_f16", "gelu_quick.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("silu_f16", "silu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
+93 -23
View File
@@ -202,19 +202,34 @@ void ggml_print_backtrace(void) {
}
#endif
static ggml_abort_callback_t g_abort_callback = NULL;
// Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback) {
ggml_abort_callback_t ret_val = g_abort_callback;
g_abort_callback = callback;
return ret_val;
}
void ggml_abort(const char * file, int line, const char * fmt, ...) {
fflush(stdout);
fprintf(stderr, "%s:%d: ", file, line);
char message[2048];
int offset = snprintf(message, sizeof(message), "%s:%d: ", file, line);
va_list args;
va_start(args, fmt);
vfprintf(stderr, fmt, args);
vsnprintf(message + offset, sizeof(message) - offset, fmt, args);
va_end(args);
fprintf(stderr, "\n");
if (g_abort_callback) {
g_abort_callback(message);
} else {
// default: print error and backtrace to stderr
fprintf(stderr, "%s\n", message);
ggml_print_backtrace();
}
ggml_print_backtrace();
abort();
}
@@ -945,6 +960,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"CONV_TRANSPOSE_1D",
"IM2COL",
"IM2COL_BACK",
"CONV_2D",
"CONV_2D_DW",
"CONV_TRANSPOSE_2D",
"POOL_1D",
@@ -986,7 +1002,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"GLU",
};
static_assert(GGML_OP_COUNT == 85, "GGML_OP_COUNT != 85");
static_assert(GGML_OP_COUNT == 86, "GGML_OP_COUNT != 86");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@@ -1044,6 +1060,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"conv_transpose_1d(x)",
"im2col(x)",
"im2col_back(x)",
"conv_2d(x)",
"conv_2d_dw(x)",
"conv_transpose_2d(x)",
"pool_1d(x)",
@@ -1085,7 +1102,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"glu(x)",
};
static_assert(GGML_OP_COUNT == 85, "GGML_OP_COUNT != 85");
static_assert(GGML_OP_COUNT == 86, "GGML_OP_COUNT != 86");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@@ -3649,9 +3666,11 @@ static struct ggml_tensor * ggml_soft_max_impl(
if (mask) {
GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(mask));
GGML_ASSERT(ggml_is_matrix(mask));
GGML_ASSERT(ggml_is_3d(mask));
GGML_ASSERT(mask->ne[0] == a->ne[0]);
GGML_ASSERT(mask->ne[1] >= a->ne[1]);
GGML_ASSERT(a->ne[2]%mask->ne[2] == 0);
GGML_ASSERT(a->ne[3]%mask->ne[3] == 0);
}
if (max_bias > 0.0f) {
@@ -4291,6 +4310,44 @@ struct ggml_tensor * ggml_conv_2d_dw_direct(
return result;
}
// ggml_conv_2d_direct
struct ggml_tensor * ggml_conv_2d_direct(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel [KW, KH, IC, OC]
struct ggml_tensor * b, // input data [W, H, C, N]
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1) {// dilation dimension 1
GGML_ASSERT(a->ne[2] == b->ne[2]);
//GGML_ASSERT(a->type == b->type);
int64_t ne[4];
ne[0] = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
ne[1] = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
ne[2] = a->ne[3];
ne[3] = b->ne[3];
struct ggml_tensor * result = ggml_new_tensor(ctx, b->type, 4, ne);
ggml_set_op_params_i32(result, 0, s0);
ggml_set_op_params_i32(result, 1, s1);
ggml_set_op_params_i32(result, 2, p0);
ggml_set_op_params_i32(result, 3, p1);
ggml_set_op_params_i32(result, 4, d0);
ggml_set_op_params_i32(result, 5, d1);
result->op = GGML_OP_CONV_2D;
result->src[0] = a;
result->src[1] = b;
return result;
}
// ggml_conv_transpose_2d_p0
static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {
@@ -4407,24 +4464,21 @@ struct ggml_tensor * ggml_pool_2d_back(
return result;
}
// ggml_upscale
// ggml_upscale / ggml_interpolate
static struct ggml_tensor * ggml_upscale_impl(
static struct ggml_tensor * ggml_interpolate_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1,
int ne2,
int ne3,
enum ggml_scale_mode mode) {
GGML_ASSERT(a->ne[0] <= ne0);
GGML_ASSERT(a->ne[1] <= ne1);
GGML_ASSERT(a->ne[2] <= ne2);
GGML_ASSERT(a->ne[3] <= ne3);
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3,
uint32_t mode) {
GGML_ASSERT((mode & 0xFF) < GGML_SCALE_MODE_COUNT);
struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
ggml_set_op_params_i32(result, 0, mode);
ggml_set_op_params_i32(result, 0, (int32_t)mode);
result->op = GGML_OP_UPSCALE;
result->src[0] = a;
@@ -4437,7 +4491,8 @@ struct ggml_tensor * ggml_upscale(
struct ggml_tensor * a,
int scale_factor,
enum ggml_scale_mode mode) {
return ggml_upscale_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3], mode);
GGML_ASSERT(scale_factor > 1);
return ggml_interpolate_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3], mode);
}
struct ggml_tensor * ggml_upscale_ext(
@@ -4448,7 +4503,18 @@ struct ggml_tensor * ggml_upscale_ext(
int ne2,
int ne3,
enum ggml_scale_mode mode) {
return ggml_upscale_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
return ggml_interpolate_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
}
struct ggml_tensor * ggml_interpolate(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3,
uint32_t mode) {
return ggml_interpolate_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
}
// ggml_pad
@@ -4625,13 +4691,17 @@ struct ggml_tensor * ggml_flash_attn_ext(
GGML_ASSERT(ggml_can_mul_mat(k, q));
// TODO: check if vT can be multiplied by (k*qT)
GGML_ASSERT(q->ne[3] == k->ne[3]);
GGML_ASSERT(q->ne[3] == v->ne[3]);
if (mask) {
GGML_ASSERT(ggml_is_contiguous(mask));
GGML_ASSERT(mask->ne[2] == 1);
GGML_ASSERT(mask->ne[3] == 1);
GGML_ASSERT(mask->ne[2] == q->ne[3]);
GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
"the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
//GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
GGML_ASSERT(q->ne[3] % mask->ne[2] == 0);
}
if (max_bias > 0.0f) {
+1 -1
View File
@@ -1 +1 @@
9e4bee1c5afc2d677a5b32ecb90cbdb483e81fff
67ad436cb653ac1ef0986f9fb0c6191ec828d1ed
+1 -1
View File
@@ -246,7 +246,7 @@ bool llama_kv_cache_unified_iswa_context::next() {
}
bool llama_kv_cache_unified_iswa_context::apply() {
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
assert(!llama_memory_status_is_fail(status));
bool res = true;
+1 -1
View File
@@ -1776,7 +1776,7 @@ bool llama_kv_cache_unified_context::next() {
}
bool llama_kv_cache_unified_context::apply() {
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
assert(!llama_memory_status_is_fail(status));
// no ubatches -> this is a KV cache update
if (ubatches.empty()) {
+1 -1
View File
@@ -218,7 +218,7 @@ bool llama_memory_hybrid_context::next() {
}
bool llama_memory_hybrid_context::apply() {
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
assert(!llama_memory_status_is_fail(status));
bool res = true;
+9 -1
View File
@@ -1071,7 +1071,15 @@ bool llama_memory_recurrent_context::next() {
}
bool llama_memory_recurrent_context::apply() {
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
assert(!llama_memory_status_is_fail(status));
// no ubatches -> this is an update
if (ubatches.empty()) {
// recurrent cache never performs updates
assert(status == LLAMA_MEMORY_STATUS_NO_UPDATE);
return true;
}
mem->find_slot(ubatches[i_next]);
+17
View File
@@ -40,3 +40,20 @@ llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_me
// if either status has an update, then the combined status has an update
return has_update ? LLAMA_MEMORY_STATUS_SUCCESS : LLAMA_MEMORY_STATUS_NO_UPDATE;
}
bool llama_memory_status_is_fail(llama_memory_status status) {
switch (status) {
case LLAMA_MEMORY_STATUS_SUCCESS:
case LLAMA_MEMORY_STATUS_NO_UPDATE:
{
return false;
}
case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
{
return true;
}
}
return false;
}
+3
View File
@@ -31,6 +31,9 @@ enum llama_memory_status {
// useful for implementing hybrid memory types (e.g. iSWA)
llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1);
// helper function for checking if a memory status indicates a failure
bool llama_memory_status_is_fail(llama_memory_status status);
// the interface for managing the memory context during batch processing
// this interface is implemented per memory type. see:
// - llama_kv_cache_unified_context
+69 -52
View File
@@ -2685,11 +2685,12 @@ struct test_soft_max : public test_case {
const std::array<int64_t, 4> ne;
const bool mask;
const ggml_type m_prec;
const std::array<int64_t, 2> nr23; // broadcast only dims 2 and 3
const float scale;
const float max_bias;
std::string vars() override {
return VARS_TO_STR6(type, ne, mask, m_prec, scale, max_bias);
return VARS_TO_STR7(type, ne, mask, m_prec, nr23, scale, max_bias);
}
// the 1024 test with bias occasionally fails:
@@ -2702,18 +2703,19 @@ struct test_soft_max : public test_case {
std::array<int64_t, 4> ne = {10, 5, 4, 3},
bool mask = false,
ggml_type m_prec = GGML_TYPE_F32,
std::array<int64_t, 2> nr23 = {1, 1},
float scale = 1.0f,
float max_bias = 0.0f)
: type(type), ne(ne), mask(mask), m_prec(m_prec), scale(scale), max_bias(max_bias) {}
: type(type), ne(ne), mask(mask), m_prec(m_prec), nr23(nr23), scale(scale), max_bias(max_bias) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2]*nr23[0], ne[3]*nr23[1]);
ggml_set_param(a);
ggml_set_name(a, "a");
ggml_tensor * mask = nullptr;
if (this->mask) {
mask = ggml_new_tensor_2d(ctx, m_prec, ne[0], ne[1]);
mask = ggml_new_tensor_4d(ctx, m_prec, ne[0], ne[1], ne[2], ne[3]);
ggml_set_name(mask, "mask");
}
@@ -3296,28 +3298,28 @@ struct test_upscale : public test_case {
}
};
// GGML_OP_UPSCALE (ext)
struct test_upscale_ext : public test_case {
// GGML_OP_UPSCALE (via ggml_interpolate)
struct test_interpolate : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
const std::array<int64_t, 4> ne_tgt;
const ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST;
const uint32_t mode = GGML_SCALE_MODE_NEAREST;
std::string vars() override {
return VARS_TO_STR4(type, ne, ne_tgt, mode);
}
test_upscale_ext(ggml_type type = GGML_TYPE_F32,
test_interpolate(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {2, 5, 7, 11},
std::array<int64_t, 4> ne_tgt = {5, 7, 11, 13},
ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST)
uint32_t mode = GGML_SCALE_MODE_NEAREST)
: type(type), ne(ne), ne_tgt(ne_tgt), mode(mode) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_set_name(a, "a");
ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3], mode);
ggml_tensor * out = ggml_interpolate(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3], mode);
ggml_set_name(out, "out");
return out;
@@ -3544,7 +3546,7 @@ struct test_flash_attn_ext : public test_case {
const int64_t hsk; // K head size
const int64_t hsv; // V head size
const int64_t nh; // num heads
const int64_t nr; // repeat in Q, tests for grouped-query attention
const std::array<int64_t, 2> nr23; // repeat in dim 2 and 3, tests for grouped-query attention
const int64_t kv; // kv size
const int64_t nb; // batch size
@@ -3558,7 +3560,7 @@ struct test_flash_attn_ext : public test_case {
std::array<int32_t, 4> permute;
std::string vars() override {
return VARS_TO_STR12(hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
return VARS_TO_STR12(hsk, hsv, nh, nr23, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
}
double max_nmse_err() override {
@@ -3569,13 +3571,13 @@ struct test_flash_attn_ext : public test_case {
GGML_UNUSED(t);
// Just counting matmul costs:
// Q*K^T is nb x hsk x kv, P*V is nb x kv x hsv, per head
return 2 * nh*nr * nb * (hsk + hsv) * kv;
return (2 * nh*nr23[0] * nb * (hsk + hsv) * kv)*nr23[1];
}
test_flash_attn_ext(int64_t hsk = 128, int64_t hsv = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
test_flash_attn_ext(int64_t hsk = 128, int64_t hsv = 128, int64_t nh = 32, std::array<int64_t, 2> nr23 = {1, 1}, int64_t kv = 96, int64_t nb = 8,
bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
ggml_type type_KV = GGML_TYPE_F16, std::array<int32_t, 4> permute = {0, 1, 2, 3})
: hsk(hsk), hsv(hsv), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
: hsk(hsk), hsv(hsv), nh(nh), nr23(nr23), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
const int64_t hsk_padded = GGML_PAD(hsk, ggml_blck_size(type_KV));
@@ -3594,18 +3596,18 @@ struct test_flash_attn_ext : public test_case {
return t;
};
ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr, 1);
ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr23[0], nr23[1]);
ggml_set_name(q, "q");
ggml_tensor * k = create_permuted(type_KV, hsk_padded, kv, nh, 1);
ggml_tensor * k = create_permuted(type_KV, hsk_padded, kv, nh, nr23[1]);
ggml_set_name(k, "k");
ggml_tensor * v = create_permuted(type_KV, hsv_padded, kv, nh, 1);
ggml_tensor * v = create_permuted(type_KV, hsv_padded, kv, nh, nr23[1]);
ggml_set_name(v, "v");
ggml_tensor * m = nullptr;
if (mask) {
m = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1);
m = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), nr23[1], 1);
ggml_set_name(m, "m");
}
@@ -4623,6 +4625,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
// this case is verified (pass) in Intel(R) Data Center GPU Max 1100 (sycl backend) and NV A30 (cuda backend)
// test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 512, 262144, 9216, {1, 1}, {1, 1}));
// test large experts*tokens
for (bool b : {false, true}) {
test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, b, 32, 1024, 16));
}
for (ggml_type type_a : base_types) {
for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
for (int n_mats : {4, 8}) {
@@ -4709,26 +4716,31 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
for (int64_t ne1 : {16, 1024}) {
if (mask) {
for (ggml_type m_prec : {GGML_TYPE_F32, GGML_TYPE_F16}) {
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, m_prec, scale, max_bias));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, scale, max_bias));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, m_prec, {1, 1}, scale, max_bias));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, {1, 1}, scale, max_bias));
if (ne0 <= 32 && ne1 <= 32) {
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, m_prec, {3, 1}, scale, max_bias));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, {2, 3}, scale, max_bias));
}
}
} else {
/* The precision of mask here doesn't matter as boolean mask is false */
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, GGML_TYPE_F32, scale, max_bias));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, GGML_TYPE_F32, scale, max_bias));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, GGML_TYPE_F32, {1, 1}, scale, max_bias));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, GGML_TYPE_F32, {1, 1}, scale, max_bias));
}
}
}
}
}
}
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, GGML_TYPE_F32, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, GGML_TYPE_F16, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, GGML_TYPE_F32, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F32, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F16, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F32, 0.1f, 8.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F16, 0.1f, 8.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, GGML_TYPE_F32, {1, 1}, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, GGML_TYPE_F16, {1, 1}, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, GGML_TYPE_F32, {1, 1}, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F32, {1, 1}, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F16, {1, 1}, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F32, {1, 1}, 0.1f, 8.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F16, {1, 1}, 0.1f, 8.0f));
for (float max_bias : {0.0f, 8.0f}) {
for (float scale : {1.0f, 0.1f}) {
@@ -4799,8 +4811,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR}) {
test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode));
test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode, true));
test_cases.emplace_back(new test_upscale_ext(GGML_TYPE_F32, {2, 5, 7, 11}, {5, 7, 11, 13}, mode));
test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {2, 5, 7, 11}, {5, 7, 11, 13}, mode));
test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {5, 7, 11, 13}, {2, 5, 7, 11}, mode));
}
test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {2, 5, 7, 11}, {5, 7, 11, 13}, GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS));
test_cases.emplace_back(new test_sum());
test_cases.emplace_back(new test_sum_rows());
@@ -4826,20 +4840,23 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
for (float logit_softcap : {0.0f, 10.0f}) {
if (hsk != 128 && logit_softcap != 0.0f) continue;
for (int nh : { 4, }) {
for (int nr : { 1, 4, 16 }) {
if (nr == 16 && hsk != 128) continue;
for (int kv : { 512, 1024, }) {
if (nr != 1 && kv != 512) continue;
for (int nb : { 1, 3, 32, 35, }) {
for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
test_cases.emplace_back(new test_flash_attn_ext(
hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
// run fewer test cases permuted
if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
for (int nr3 : { 1, 3, }) {
if (hsk > 64 && nr3 > 1) continue; // skip broadcast for large head sizes
for (int nr2 : { 1, 4, 16 }) {
if (nr2 == 16 && hsk != 128) continue;
for (int kv : { 512, 1024, }) {
if (nr2 != 1 && kv != 512) continue;
for (int nb : { 1, 3, 32, 35, }) {
for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
test_cases.emplace_back(new test_flash_attn_ext(
hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
// run fewer test cases permuted
if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
test_cases.emplace_back(new test_flash_attn_ext(
hsk, hsv, nh, {nr2, nr3}, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
}
}
}
}
@@ -4883,13 +4900,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {8192, 512, 2, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {3072, 512, 2, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {1024, 1024, 10, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 1024, 10, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {256, 256, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {64, 64, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 64, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {1024, 1024, 10, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 1024, 10, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {256, 256, 20, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {64, 64, 20, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 64, 20, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32, 10, 1, 1}));
test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
@@ -4921,7 +4938,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
for (int kv : { 4096, 8192, 16384, }) {
for (int hs : { 64, 128, }) {
for (int nr : { 1, 4, }) {
test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, nr, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, {nr, 1}, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
}
}
}