Skip to content

feat: support mxfp4 & mxfp8 entrypoint for blackwell cutedsl dense gemm #2660

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
flashinfer-bot merged 8 commits into from
Mar 6, 2026
Merged

feat: support mxfp4 & mxfp8 entrypoint for blackwell cutedsl dense gemm #2660

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
73ab0bc
feat: Add mxfp4 blockscale dense gemm support
b8zhong Feb 28, 2026
c52ce25
fix: adjust alpha calculation for fp4 gemm
b8zhong Feb 28, 2026
32f22d8
more
b8zhong Feb 28, 2026
fd9d432
more
b8zhong Feb 28, 2026
949c488
more
b8zhong Feb 28, 2026
6608b2f
more
b8zhong Mar 1, 2026
0703627
more
b8zhong Mar 5, 2026
59de382
more
b8zhong Mar 6, 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
4 changes: 2 additions & 2 deletions benchmarks/routines/flashinfer_benchmark_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -368,8 +368,8 @@ def dtype_str_to_torch_dtype(dtype_str):
"8.6": [],
"8.9": [],
"9.0": [],
"10.0": ["cutlass"],
"10.3": ["cutlass"],
"10.0": ["cutlass", "cute-dsl"],
"10.3": ["cutlass", "cute-dsl"],
"11.0": ["cutlass"],
"12.0": [],
},
Expand Down
17 changes: 11 additions & 6 deletions benchmarks/routines/gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1071,6 +1071,13 @@ def testMmFp4(args):
print(f"[VVERBOSE] {mat2_fp4.dtype = }")

alpha = 1.0 / (global_sf_input * global_sf_mat2) if use_nvfp4 else None
# TODO: for MXFP4, we don't need a global scale, we should change the compile interface to make
# alpha optional.
alpha_for_cute_dsl_mxfp4 = (
Copy link
Copy Markdown
Collaborator

@nv-yunzheq nv-yunzheq Mar 4, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is there a reason why mxfp4 cutedsl backend have to be using a device tensor with value 1.0?

Copy link
Copy Markdown
Contributor Author

@b8zhong b8zhong Mar 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

When I removed it, I encountered some compilation relating to make_fake_compact_tensor(cutlass.Float32, (1,)), as I believe they still share the exact same code path. To avoid touching too much cute-dsl which I am not super familiar with, I left it as passing it 1.0 for now. But I also think it can be removed.

Copy link
Copy Markdown
Collaborator

@nv-yunzheq nv-yunzheq Mar 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you add a comment something like TODO so that in the future people could understand why you add this and we should get it removed in the future?

torch.tensor([1.0], dtype=torch.float32, device=device)
if not use_nvfp4
else None
)
# Completed preparing inputs. Now programmatically filter backends
block_size = 16 if use_nvfp4 else 32
backends_to_remove = []
Expand All @@ -1091,7 +1098,7 @@ def testMmFp4(args):
b=mat2_fp4.T if backend != "trtllm" else mat2_fp4_trtllm.T,
a_descale=input_inv_s,
b_descale=mat2_inv_s.T if backend != "trtllm" else mat2_inv_s_trtllm.T,
alpha=alpha,
alpha=(alpha_for_cute_dsl_mxfp4 if (backend == "cute-dsl") else alpha),
out_dtype=res_dtype,
block_size=16
if use_nvfp4
Expand Down Expand Up @@ -1129,7 +1136,7 @@ def run_backend(
b=mat2_fp4.T if backend != "trtllm" else mat2_fp4_trtllm.T,
a_descale=input_inv_s,
b_descale=mat2_inv_s.T if backend != "trtllm" else mat2_inv_s_trtllm.T,
alpha=alpha,
alpha=(alpha_for_cute_dsl_mxfp4 if (backend == "cute-dsl") else alpha),
out_dtype=res_dtype,
block_size=block_size,
use_8x4_sf_layout=not use_128x4_sf_layout,
Expand Down Expand Up @@ -1289,9 +1296,7 @@ def testMmMxfp8(args):
res_dtype = args.out_dtype
is_cuda_graph_compatible = not args.no_cuda_graph
run_refcheck = args.refcheck
autotune_supported_backends = [
"cutlass",
]
autotune_supported_backends = ["cutlass", "cute-dsl", "auto"]
res = []

backends = filter_backends_by_compute_capability(backends, args.routine, device)
Expand Down Expand Up @@ -1344,7 +1349,7 @@ def testMmMxfp8(args):
print(f"[VVERBOSE] {mat2_scale.dtype = }")

def run_backend(backend, input_mxfp8, mat2_mxfp8, input_scale, mat2_scale):
if backend == "cutlass":
if backend in ["cutlass", "cute-dsl", "auto"]:
return flashinfer.gemm.mm_mxfp8(
a=input_mxfp8,
b=mat2_mxfp8.t(), # mm_mxfp8 expects b.t()
Expand Down
Loading
Loading