Skip to content

CUDA: fix FA occupancy, optimize tile kernel #15982

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Sep 17, 2025
+353 −245
Merged

CUDA: fix FA occupancy, optimize tile kernel #15982

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions ggml/src/ggml-cuda/common.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,6 +75,8 @@
#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
#define GGML_CUDA_CC_IS_GCN(cc) (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1)
#define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
#define GGML_CUDA_CC_IS_CDNA1(cc) (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
#define GGML_CUDA_CC_IS_CDNA2(cc) (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
#define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)

// Moore Threads
Expand Down Expand Up @@ -325,6 +327,20 @@ static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
#endif // defined(GGML_USE_HIP) && (defined(__GFX9__) || defined(__GFX8__))
}

// Maximum number of bytes that can be copied in a single instruction.
static constexpr __device__ int ggml_cuda_get_max_cpy_bytes() {
#ifdef GGML_USE_HIP
return 16;
#else
#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
return 16;
#else
return 8;
#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
#endif // GGML_USE_HIP
}


[[noreturn]]
static __device__ void no_device_code(
const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
Expand Down
13 changes: 2 additions & 11 deletions ggml/src/ggml-cuda/fattn-common.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -647,9 +647,7 @@ static __global__ void flash_attn_stream_k_fixup(
}

template<int D> // D == head size
#if !defined(GGML_USE_HIP)
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIP)
static __global__ void flash_attn_combine_results(
const float * __restrict__ VKQ_parts,
const float2 * __restrict__ VKQ_meta,
Expand Down Expand Up @@ -692,10 +690,7 @@ static __global__ void flash_attn_combine_results(
float VKQ_numerator = 0.0f;
float VKQ_denominator = 0.0f;
for (int l = 0; l < parallel_blocks; ++l) {
const float diff = meta[l].x - kqmax;
float KQ_max_scale = expf(diff);
const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
*((uint32_t *) &KQ_max_scale) &= ftz_mask;
const float KQ_max_scale = expf(meta[l].x - kqmax);

VKQ_numerator += KQ_max_scale * VKQ_parts[l*D + tid];
VKQ_denominator += KQ_max_scale * meta[l].y;
Expand Down Expand Up @@ -836,11 +831,10 @@ void launch_fattn(
CUDA_CHECK(cudaGetLastError());
}

int parallel_blocks = 1;

const dim3 block_dim(warp_size, nwarps, 1);
int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
int parallel_blocks = max_blocks_per_sm;

dim3 blocks_num;
if (stream_k) {
Expand All @@ -862,9 +856,6 @@ void launch_fattn(
GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.

// parallel_blocks should be at least large enough to achieve max. occupancy for a single wave:
parallel_blocks = std::max((nsm * max_blocks_per_sm) / ntiles_total, 1);

// parallel_blocks must not be larger than what the tensor size allows:
parallel_blocks = std::min(parallel_blocks, ntiles_KQ);

Expand Down
Loading
Loading