8000 cuda : add batched cuBLAS GEMM for faster attention by ggerganov · Pull Request #3749 · ggml-org/llama.cpp · GitHub
[go: up one dir, main page]
Skip to content

cuda : add batched cuBLAS GEMM for faster attention #3749

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 10 commits into from
Oct 24, 2023
+193 −13
Merged

cuda : add batched cuBLAS GEMM for faster attention #3749

Changes from 1 commit
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
8fb1be6
cmake : add helper for faster CUDA builds
ggerganov Oct 23, 2023
6a30bf3
batched : add NGL arg
ggerganov Oct 23, 2023
8d8d54f
ggml : skip nops in compute_forward
ggerganov Oct 23, 2023
84d4ca0
cuda : minor indentation
ggerganov Oct 23, 2023
c13fcfb
cuda : batched cuBLAS GEMMs for src0 F16 and src1 F32 (attention ops)
ggerganov Oct 23, 2023
878aa4f
Apply suggestions from code review
KerfuffleV2 Oct 23, 2023
d415669
cuda : add ROCm / hipBLAS cublasGemmBatchedEx define
ggerganov Oct 23, 2023
3d297c1
cuda : add cublasGemmStridedBatchedEx for non-broadcasted cases
ggerganov Oct 24, 2023
27c34c0
cuda : reduce mallocs in cublasGemmBatchedEx branch
ggerganov Oct 24, 2023
d798a17
cuda : add TODO for calling cublas from kernel + using mem pool
ggerganov Oct 24, 2023
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
Prev Previous commit
Next Next commit
cuda : add cublasGemmStridedBatchedEx for non-broadcasted cases
  • Loading branch information
@ggerganov
ggerganov committed Oct 24, 2023
commit 3d297c1a3073bcf62c0bde46d4eea0caeb56160f
23 changes: 18 additions & 5 deletions 8000 ggml-cuda.cu
View file
Open in desktop
5D7C
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@
#define cublasCreate hipblasCreate
#define cublasGemmEx hipblasGemmEx
#define cublasGemmBatchedEx hipblasGemmBatchedEx
#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
#define cublasHandle_t hipblasHandle_t
#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
#define cublasSetStream hipblasSetStream
Expand Down Expand Up @@ -7125,17 +7126,29 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
CUBLAS_CHECK(
cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
&alpha_f16, (char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , CUDA_R_16F, nb01/sizeof(half),
(char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float),
&beta_f16, (char *) dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2, CUDA_R_16F, ne01,
&alpha_f16, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , CUDA_R_16F, nb01/sizeof(half),
(const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float),
&beta_f16, ( char *) dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2, CUDA_R_16F, ne01,
CUBLAS_COMPUTE_16F,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
}
}
}
#else
// use cublasGemmBatchedEx
{
if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
// use cublasGemmStridedBatchedEx
CUBLAS_CHECK(
cublasGemmStridedBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
&alpha_f16, (const char *) src0_as_f16, CUDA_R_16F, nb01/sizeof(half), src0->nb[2]/sizeof(half), // strideA
(const char *) src1_as_f16, CUDA_R_16F, nb11/sizeof(float), src1->nb[2]/sizeof(float), // strideB
&beta_f16, ( char *) dst_f16, CUDA_R_16F, ne01, dst->nb[2]/sizeof(float), // strideC
ne12*ne13,
CUBLAS_COMPUTE_16F,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
} else {
// use cublasGemmBatchedEx
const int ne23 = ne12*ne13;

// TODO: avoid this alloc
Expand Down
0