biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

NVFP4-1.1: fix cute.math.fmax -> cute.arch.fmax (correct CuTeDSL API)

Browse Source
This commit is contained in:
biondizzle
2026-05-28 03:48:51 +00:00
parent 60790564f0
commit 6f94925491
2 changed files with 8 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/unit/test_nvfp4_1_1_quant.py
View File

@@ -48,8 +48,8 @@ def fp4_quant_test_kernel(
ptr = input_bf16.iterator + i * cutlass.Int32(2) # BF16=2 bytes
bf16_val = cute.arch.load(ptr, cutlass.BFloat16)
v = bf16_val.to(cutlass.Float32) / gs
a = cute.math.fmax(v, cutlass.Float32(0.0) - v)
amax = cute.math.fmax(amax, a)
a = cute.arch.fmax(v, cutlass.Float32(0.0) - v)
amax = cute.arch.fmax(amax, a)
# Block scale
bsf_f32 = amax / cutlass.Float32(6.0)