Skip to content

[None][perf] fp8 block scale quant fusion in SM90 Cutlass MoE #16849

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

Open
amukkara wants to merge 3 commits into
base: main
Choose a base branch
from
+519 −40
Open

[None][perf] fp8 block scale quant fusion in SM90 Cutlass MoE #16849

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0250211
Fuse pre-FC2 FP8 1x128 activation quant into doActivation kernel
amukkara Jul 22, 2026
4806536
Fuse pre-FC1 FP8 1x128 activation quant into the row-expansion kernel
amukkara Jul 23, 2026
4dd38b4
Build a single FP8 block-scale runner
amukkara Jul 23, 2026
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
17 changes: 16 additions & 1 deletion cpp/tensorrt_llm/kernels/cutlass_kernels/fp8_blockscale_gemm/fp8_blockscale_gemm.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2025-2026, NVIDIA CORPORATION. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
Expand All @@ -21,6 +21,7 @@
#include <cstdint>
#include <cuda_fp8.h>
#include <cuda_runtime_api.h>
#include <type_traits>
#include <vector>

// non-persistent-cooperative GEMM
Expand Down Expand Up @@ -55,6 +56,10 @@ class CutlassFp8BlockScaleGemmRunnerInterface
float const* scales_b = nullptr)
= 0;

virtual int64_t getActScaleLeadingDim() const = 0;

virtual bool isActivationPrequantized() const = 0;

virtual void strideBatchGemm(__nv_bfloat16* mat_d, int ld_d, int stride_d, __nv_fp8_e4m3* mat_a, int ld_a,
int stride_a, __nv_fp8_e4m3* mat_b, int ld_b, int stride_b, int num_problems, int shape_m, int shape_n,
int shape_k, cudaStream_t stream, float* scales_a, int stride_scales_a, float* scales_b)
Expand Down Expand Up @@ -117,6 +122,16 @@ class CutlassFp8BlockScaleGemmRunner : public CutlassFp8BlockScaleGemmRunnerInte
size_t num_problems, size_t shape_n, size_t shape_k, cudaStream_t stream, float const* scales_a = nullptr,
float const* scales_b = nullptr) override;

int64_t getActScaleLeadingDim() const override
{
return max_shape_m_32_align_padded_;
}

bool isActivationPrequantized() const override
{
return std::is_same_v<ElementA, __nv_fp8_e4m3>;
}

void strideBatchGemm(__nv_bfloat16* mat_d, int ld_d, int stride_d, __nv_fp8_e4m3* mat_a, int ld_a, int stride_a,
__nv_fp8_e4m3* mat_b, int ld_b, int stride_b, int num_problems, int shape_m, int shape_n, int shape_k,
cudaStream_t stream, float* scales_a, int stride_scales_a, float* scales_b) override;
Expand Down
Loading
Loading