Skip to content

Optimizing fp8 gemm cublasLt handle usage, prevent leakage #73644

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 6 commits into from
Jun 27, 2025
+24 −25
Merged

Optimizing fp8 gemm cublasLt handle usage, prevent leakage #73644

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
327053b
Fix
co63oc Jun 24, 2025
b1c636c
Merge branch 'develop' into f47
co63oc Jun 26, 2025
9c359c4
Fix
co63oc Jun 26, 2025
cb49752
Using framework handle
A-nnonymous Jun 26, 2025
e440da6
Merge commit 'refs/pull/73581/head(co's fix)' of https://github.com/P…
A-nnonymous Jun 26, 2025
173acf8
delete profiler
A-nnonymous Jun 26, 2025
File filter

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
49 changes: 24 additions & 25 deletions paddle/phi/kernels/legacy/gpu/fp8_gemm_blockwise_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,7 +76,9 @@ cudaDataType_t ScalarTypeToCudaDataType(phi::DataType dtype) {
} \
} while (0)

void cublas_gemm_blockwise_impl(const DenseTensor& A,
template <typename Context>
void cublas_gemm_blockwise_impl(const Context& dev_ctx,
const DenseTensor& A,
const DenseTensor& A_decode_scale,
const DenseTensor& B,
const DenseTensor& B_decode_scale,
Expand Down Expand Up @@ -148,15 +150,13 @@ void cublas_gemm_blockwise_impl(const DenseTensor& A,
int lda = k, ldb = k, ldc = m, ldd = m;
float alpha = 1.0, beta = accumulate ? 1.0 : 0.0;

cublasLtHandle_t ltHandle;
PADDLE_CUDABLAS_CHECK(phi::dynload::cublasLtCreate(&ltHandle));

cublasLtHandle_t ltHandle = dev_ctx.cublaslt_handle();
// Create operation descriptor
cublasLtMatmulDesc_t operationDesc = nullptr;
PADDLE_CUDABLAS_CHECK(phi::dynload::cublasLtMatmulDescCreate(
&operationDesc, CUBLAS_COMPUTE_32F, CUDA_R_32F));

#if CUBLAS_VERSION >= 120804 && CUDA_VERSION >= 12060
#if CUBLAS_VERSION >= 120805 && CUDA_VERSION >= 12080
// Setup scaling for A and B
cublasLtMatmulMatrixScale_t A_scale_mode, B_scale_mode;
// Note: in cuBLAS term, tensor name A and B are swapped.
Expand Down Expand Up @@ -187,7 +187,7 @@ void cublas_gemm_blockwise_impl(const DenseTensor& A,
sizeof(B_scale_mode)));
#else
PADDLE_THROW(phi::errors::InvalidArgument(
"Sub-channel FP8 GEMM requires CUDA 12.8 and cuBLAS 12.8.4 or later."));
"Sub-channel FP8 GEMM requires CUDA 12.8 and cuBLAS 12.8.5 or later."));
#endif

// setup transa and transb
Expand Down Expand Up @@ -285,7 +285,6 @@ void cublas_gemm_blockwise_impl(const DenseTensor& A,
workspace->data(),
workspace_size,
stream));

// Cleanup
if (preference)
PADDLE_CUDABLAS_CHECK(
Expand All @@ -301,7 +300,6 @@ void cublas_gemm_blockwise_impl(const DenseTensor& A,
if (operationDesc)
PADDLE_CUDABLAS_CHECK(
phi::dynload::cublasLtMatmulDescDestroy(operationDesc));
if (ltHandle) PADDLE_CUDABLAS_CHECK(phi::dynload::cublasLtDestroy(ltHandle));
}

} // anonymous namespace
Expand All @@ -327,23 +325,24 @@ void Fp8GemmBlockwiseKernel(const Context& dev_ctx,
DenseTensor* output,
DenseTensor* pre_gelu_out,
DenseTensor* workspace_out) {
cublas_gemm_blockwise_impl(A,
A_scale,
B,
B_scale,
output,
bias,
pre_gelu_out,
transa,
transb,
grad,
workspace_out,
accumulate,
use_split_accumulator,
math_sm_count,
is_A_1d_scaled,
is_B_1d_scaled,
dev_ctx.stream());
cublas_gemm_blockwise_impl<Context>(dev_ctx,
A,
A_scale,
B,
B_scale,
output,
bias,
pre_gelu_out,
transa,
transb,
grad,
workspace_out,
accumulate,
use_split_accumulator,
math_sm_count,
is_A_1d_scaled,
is_B_1d_scaled,
dev_ctx.stream());
}

} // namespace phi
Expand Down
Loading