Skip to content

perf(scale): optimize FP8 scale conversion and support E5M3 - #112

Merged
jinzhen-lin merged 5 commits into
mainfrom
perf/scale-conversion
Oct 5, 2026
Merged

perf(scale): optimize FP8 scale conversion and support E5M3#112
jinzhen-lin merged 5 commits into
mainfrom
perf/scale-conversion

Conversation

@jinzhen-lin

@jinzhen-lin jinzhen-lin commented Oct 4, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator
  • Support E8M0 scales with FP16 by rebiasing to E5M0, adding a tensorwise secondary scale only when needed, and preserving the high end when the range cannot fit.
  • Use packed dequantization for E4M3/E5M2 scales before SM89 and cvt instructions on SM89 and newer.
  • Add unsigned E5M3 scales using existing dequantization paths.
  • Read software-dequantized group scales in place and avoid packed permutations before FP32 conversion.

jinzhen-lin and others added 5 commits October 4, 2026 18:19
@jinzhen-lin @codex
perf(scale): optimize FP8 scale conversion and support E5M3
fd52f61 
Co-authored-by: Codex (GPT-6) <codex@openai.com>
Signed-off-by: jinzhen.ljz <jinzhen.ljz@antgroup.com>
@jinzhen-lin @codex
perf(scale): vectorize scale dequantization without special-value bra…
b663ce1 
…nches

Co-authored-by: Codex (GPT-6) <codex@openai.com>
Signed-off-by: jinzhen.ljz <jinzhen.ljz@antgroup.com>
@jinzhen-lin @codex
perf(scale): compensate exponent bias before widening BF16 scales
2fe9a38 
Disable fast math in Humming kernels and apply packed BF16 exponent compensation before conversion to FP32.

Co-authored-by: Codex (GPT-6) <codex@openai.com>
Signed-off-by: jinzhen.ljz <jinzhen.ljz@antgroup.com>
@jinzhen-lin @codex
merge(main): preserve dequantized scale order and avoid permutations
a04adee 
Integrate PR #107 with scale conversion and restore FP32 scale order through register indices.

Co-authored-by: Codex (GPT-6) <codex@openai.com>
Signed-off-by: jinzhen.ljz <jinzhen.ljz@antgroup.com>
@jinzhen-lin
merge(main): integrate CI fixes into scale conversion
4e59116 
Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
@jinzhen-lin
jinzhen-lin merged commit a2ac79d into main Oct 5, 2026
5 checks passed
@jinzhen-lin
jinzhen-lin deleted the perf/scale-conversion branch October 5, 2026 10:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@jinzhen-lin