diff options
Diffstat (limited to 'ggml/src')
| -rw-r--r-- | ggml/src/ggml-backend.c | 4 | ||||
| -rw-r--r-- | ggml/src/ggml-cuda.cu | 10 | ||||
| -rw-r--r-- | ggml/src/ggml-cuda/fattn-new-mma.cu | 8 |
3 files changed, 18 insertions, 4 deletions
diff --git a/ggml/src/ggml-backend.c b/ggml/src/ggml-backend.c index 410ab9e5..e191c2d9 100644 --- a/ggml/src/ggml-backend.c +++ b/ggml/src/ggml-backend.c @@ -35,7 +35,7 @@ size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) { return SIZE_MAX; } -GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) { +GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) { // get_alloc_size is optional, defaults to ggml_nbytes if (buft->iface.get_alloc_size) { size_t size = buft->iface.get_alloc_size(buft, tensor); @@ -114,7 +114,7 @@ size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) { return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer)); } -size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { +size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor) { return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor); } diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index cf17cbc1..9c8c91f4 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -2101,13 +2101,19 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer); + // If src0 is a temporary compute buffer it may have some padding that needs to be cleared for mul_mat_vec_q or mul_mat_q. + // But if src0 is also a view of another tensor then this cannot be done safely because it may overwrite valid tensor data. + // Therefore, in such cases use cuBLAS. + const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE + && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src; + bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1; - bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) + bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE; - bool use_mul_mat_q = ggml_is_quantized(src0->type) + bool use_mul_mat_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32; // if mmvq is available it's a better choice than dmmv: diff --git a/ggml/src/ggml-cuda/fattn-new-mma.cu b/ggml/src/ggml-cuda/fattn-new-mma.cu index 5aeca3c4..6178b3e5 100644 --- a/ggml/src/ggml-cuda/fattn-new-mma.cu +++ b/ggml/src/ggml-cuda/fattn-new-mma.cu @@ -1093,6 +1093,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( } } + // do we really need this? + __syncthreads(); + // Write back combined meta data: #pragma unroll for (int imeta = 0; imeta < nmeta; ++imeta) { @@ -1112,6 +1115,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols; dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs); } + } else if (np > 1) { + // Warps with threadIdx.y % np == 0 execute a __syncthreads() in the if branch. + // Therefore, all other warps also need to execute a __syncthreads(). + // Otherwise the points at which warps synchronize with each other would become misaligned. + __syncthreads(); } #pragma unroll |