Skip to content

Commit

Permalink
Merge branch 'master' into sycl-gemm-dispatch
Browse files Browse the repository at this point in the history
  • Loading branch information
@airMeng
airMeng authored May 28, 2024
2 parents 9ef4671 + 852aafb commit bcadd61
Show file tree
Hide file tree
Showing 7 changed files with 197 additions and 65 deletions.
8 changes: 6 additions & 2 deletions CMakePresets.json
View file
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
{
{
"version": 4,
"configurePresets": [
{
Expand Down Expand Up @@ -40,6 +40,10 @@

{ "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "release" ] },
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "release", "static" ] }
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "release", "static" ] },

{ "name": "x64-windows-msvc-debug" , "inherits": [ "base", "debug" ] },
{ "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] },
{ "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] }
]
}
3 changes: 3 additions & 0 deletions Makefile
View file
Original file line number Diff line number Diff line change
Expand Up @@ -441,6 +441,9 @@ endif # JETSON_EOL_MODULE_DETECT
ifdef LLAMA_DEBUG
MK_NVCCFLAGS += -lineinfo
endif # LLAMA_DEBUG
ifdef LLAMA_CUDA_DEBUG
MK_NVCCFLAGS += --device-debug
endif # LLAMA_CUDA_DEBUG
ifdef LLAMA_CUDA_NVCC
NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC)
else
Expand Down
33 changes: 24 additions & 9 deletions ggml-cuda.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -119,6 +119,20 @@ int ggml_cuda_get_device() {
return id;
}

static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
ggml_cuda_set_device(device);
#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
auto res = hipMallocManaged(ptr, size);
if (res == hipSuccess) {
// if error we "need" to know why...
CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
}
return res;
#else
return cudaMalloc(ptr, size);
#endif
}

static ggml_cuda_device_info ggml_cuda_init() {
#ifdef __HIP_PLATFORM_AMD__
// Workaround for a rocBLAS bug when using multiple graphics cards:
Expand Down Expand Up @@ -271,7 +285,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
size_t look_ahead_size = (size_t) (1.05 * size);
look_ahead_size = 256 * ((look_ahead_size + 255)/256);
ggml_cuda_set_device(device);
CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
*actual_size = look_ahead_size;
pool_size += look_ahead_size;
#ifdef DEBUG_CUDA_MALLOC
Expand Down Expand Up @@ -537,7 +551,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe
size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0

void * dev_ptr;
cudaError_t err = cudaMalloc(&dev_ptr, size);
cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device);
if (err != cudaSuccess) {
// clear the error
cudaGetLastError();
Expand Down Expand Up @@ -798,7 +812,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_bu
// currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
ggml_cuda_set_device(id);
char * buf;
CUDA_CHECK(cudaMalloc(&buf, size));
CUDA_CHECK(ggml_cuda_device_malloc((void**)&buf, size, id));

// set padding to 0 to avoid possible NaN values
if (size > original_size) {
Expand Down Expand Up @@ -2510,9 +2524,9 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t

bool use_cuda_graph = true;
bool cuda_graph_update_required = false;
// pointer to CUDA cpy kernel, which is required to identify
// vector of pointers to CUDA cpy kernels, which are required to identify
// kernel parameters which need updated in the graph for each token
void * ggml_cuda_cpy_fn_ptr = nullptr;
std::vector<void *> ggml_cuda_cpy_fn_ptrs;

if (cuda_ctx->cuda_graph->graph == nullptr) {
if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
Expand Down Expand Up @@ -2588,9 +2602,10 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
if (node->op == GGML_OP_CPY) {
// store the copy op parameter which changes with each token.
cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
if (ggml_cuda_cpy_fn_ptr == nullptr) {
// store a pointer to the copy op CUDA kernel to identify it later
ggml_cuda_cpy_fn_ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
// store a pointer to each copy op CUDA kernel to identify it later
void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
ggml_cuda_cpy_fn_ptrs.push_back(ptr);
}
}

Expand Down Expand Up @@ -2720,7 +2735,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
if (!cuda_graph_update_required) { // on update steps, the live parameters will already be captured
int k = 0;
for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
if (cuda_ctx->cuda_graph->params[i].func == ggml_cuda_cpy_fn_ptr) {
if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
Expand Down
5 changes: 0 additions & 5 deletions ggml-cuda/common.cuh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -79,13 +79,8 @@
#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
#define cudaHostUnregister hipHostUnregister
#define cudaLaunchHostFunc hipLaunchHostFunc
#ifdef GGML_HIP_UMA
#define cudaMalloc hipMallocManaged
#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
#else
#define cudaMalloc hipMalloc
#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
#endif
#define cudaMemcpy hipMemcpy
#define cudaMemcpyAsync hipMemcpyAsync
#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
Expand Down
68 changes: 58 additions & 10 deletions ggml-metal.m
View file
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,10 @@
GGML_METAL_KERNEL_TYPE_MUL_ROW,
GGML_METAL_KERNEL_TYPE_DIV,
GGML_METAL_KERNEL_TYPE_DIV_ROW,
GGML_METAL_KERNEL_TYPE_REPEAT_F32,
GGML_METAL_KERNEL_TYPE_REPEAT_F16,
GGML_METAL_KERNEL_TYPE_REPEAT_I32,
GGML_METAL_KERNEL_TYPE_REPEAT_I16,
GGML_METAL_KERNEL_TYPE_SCALE,
GGML_METAL_KERNEL_TYPE_SCALE_4,
GGML_METAL_KERNEL_TYPE_CLAMP,
Expand Down Expand Up @@ -184,9 +188,9 @@
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
//GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,
//GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
Expand Down Expand Up @@ -485,6 +489,10 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32, repeat_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16, repeat_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32, repeat_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16, repeat_i16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true);
Expand Down Expand Up @@ -634,9 +642,9 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true);
Expand Down Expand Up @@ -746,6 +754,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
case GGML_OP_ACC:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_REPEAT:
case GGML_OP_SCALE:
case GGML_OP_CLAMP:
case GGML_OP_SQR:
Expand All @@ -770,6 +779,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
case GGML_OP_LEAKY_RELU:
return true;
case GGML_OP_FLASH_ATTN_EXT:
if (op->src[0]->ne[0] == 256) {
return false;
}
return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
Expand Down Expand Up @@ -976,8 +988,6 @@ static enum ggml_status ggml_metal_graph_compute(
switch (dst->op) {
case GGML_OP_CONCAT:
{
const int64_t nb = ne00;

id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline;

[encoder setComputePipelineState:pipeline];
Expand Down Expand Up @@ -1008,7 +1018,6 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26];
[encoder setBytes:&nb length:sizeof(nb) atIndex:27];

const int nth = MIN(1024, ne0);

Expand All @@ -1018,11 +1027,14 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_OP_MUL:
case GGML_OP_DIV:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
GGML_ASSERT(src1t == GGML_TYPE_F32);

const size_t offs = 0;

bool bcast_row = false;

int64_t nb = ne00;
int64_t nb = ne00; // used by the "row" kernels

id<MTLComputePipelineState> pipeline = nil;

Expand Down Expand Up @@ -1091,6 +1103,42 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
}
} break;
case GGML_OP_REPEAT:
{
id<MTLComputePipelineState> pipeline;

switch (src0t) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break;
case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break;
default: GGML_ASSERT(false);
}

[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17];

const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);

[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ACC:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
Expand Down Expand Up @@ -2573,7 +2621,7 @@ static enum ggml_status ggml_metal_graph_compute(
case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
//case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
Expand All @@ -2586,7 +2634,7 @@ static enum ggml_status ggml_metal_graph_compute(

switch (ne00) {
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break;
case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
//case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
Expand Down
51 changes: 49 additions & 2 deletions ggml-metal.metal
View file
Original file line number Diff line number Diff line change
Expand Up @@ -168,6 +168,53 @@ kernel void kernel_div(
}
}

template<typename T>
kernel void kernel_repeat(
device const char * src0,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i3 = tgpig.z;
const int64_t i2 = tgpig.y;
const int64_t i1 = tgpig.x;

const int64_t i03 = i3 % ne03;
const int64_t i02 = i2 % ne02;
const int64_t i01 = i1 % ne01;

device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
device char * dst_ptr = dst + i3*nb3 + i2*nb2 + i1*nb1 ;

for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
const int i00 = i0 % ne00;
*((device T *)(dst_ptr + i0*nb0)) = *((device T *)(src0_ptr + i00*nb00));
}
}

typedef decltype(kernel_repeat<float>) kernel_repeat_t;

template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;

// assumption: src1 is a row
// broadcast src1 into src0
kernel void kernel_add_row(
Expand Down Expand Up @@ -2418,7 +2465,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f
template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<96>;
template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<112>;
template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128>;
template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>;
//template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>;

template<int64_t D, int64_t Q = 1, int64_t C = 32> // head size, queries per threadgroup, cache items per threadgroup
kernel void kernel_flash_attn_ext_vec_f16(
Expand Down Expand Up @@ -2696,7 +2743,7 @@ kernel void kernel_flash_attn_ext_vec_f16(
}

template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<128>;
template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>;
//template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>;

kernel void kernel_cpy_f16_f16(
device const half * src0,
Expand Down
Loading

0 comments on commit bcadd61

Please sign in to comment.