Skip to content

[Bugfix][ROCM] Fix the MXFP8 block scale exponent #53110

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
AndreasKaratzas merged 2 commits into from
Aug 22, 2026
Merged

[Bugfix][ROCM] Fix the MXFP8 block scale exponent #53110

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
61c8a85
[Bugfix][Quantization] Fix the MXFP8 block scale exponent
stefankoncarevic Aug 19, 2026
6f54e26
Merge branch 'main' into fix-mxfp8-block-scale
stefankoncarevic Aug 21, 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
11 changes: 9 additions & 2 deletions vllm/model_executor/layers/quantization/utils/mxfp8_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -107,6 +107,8 @@ def _kernel(
ssm,
ssk,
BLOCK_M: tl.constexpr,
FP8_MAX: tl.constexpr,
TINY: tl.constexpr,
):
pid_m = tl.program_id(0)
pid_b = tl.program_id(1) # which 32-element block along K
Expand All @@ -118,8 +120,11 @@ def _kernel(
mask=m_mask[:, None],
other=0.0,
).to(tl.float32)
amax = tl.maximum(tl.max(tl.abs(x), axis=1), 1e-30) # [BLOCK_M]
sb = tl.floor(tl.log2(amax)) + 127.0
# Mirror _mxfp8_e4m3_quantize_torch: the scale has to put the block amax
# at the top of the e4m3 range rather than at 1.0, or small elements of
# the block end up in the subnormals.
amax = tl.maximum(tl.max(tl.abs(x), axis=1), TINY) # [BLOCK_M]
sb = tl.ceil(tl.log2(amax / FP8_MAX)) + 127.0
sb = tl.minimum(tl.maximum(sb, 0.0), 254.0)
descale = tl.exp2(sb - 127.0)
xq = (x / descale[:, None]).to(xq_ptr.dtype.element_ty)
Expand Down Expand Up @@ -167,6 +172,8 @@ def _mxfp8_e4m3_quantize_triton(
scales.stride(0),
scales.stride(1),
BLOCK_M=BLOCK_M,
FP8_MAX=float(torch.finfo(MXFP8_VALUE_DTYPE).max),
TINY=float(torch.finfo(torch.float32).tiny),
)
return xq, scales

Expand Down
Loading