[Kernel] Add nvfp4 gemm flashinfer backends (#22346)

Signed-off-by: Julien Lin <jullin@nvidia.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
nvjullin
2025-08-14 16:03:55 -04:00
committed by GitHub
co-authored by mgoin
parent b8ff05361a
commit 279a5f31b3
9 changed files with 369 additions and 39 deletions
@@ -65,9 +65,12 @@ def test_nvfp4_gemm(
b_global_scale = ((FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX) /
torch.amax(b_dtype.flatten(), dim=-1)).to(torch.float32)
alpha = 1. / (a_global_scale * b_global_scale)
# ops.scaled_fp4_quant returns swizzled scales, while weights
# from checkpoints are in linear scales.
a_fp4, a_scale_interleaved = ops.scaled_fp4_quant(a_dtype, a_global_scale)
b_fp4, b_scale_interleaved = ops.scaled_fp4_quant(b_dtype, b_global_scale)
# get_ref_results unswizzles the scales internally.
expected_out = get_ref_results(a_fp4, b_fp4, a_scale_interleaved,
b_scale_interleaved, a_global_scale,
b_global_scale, m, n, dtype, block_size,