[bug] Qwen3 FP8 worse than BF16

[rhmukundan]

Problem

Issue:
Performance difference observed where FP8-MX underperforms compared to BF16 in Qwen3 235B and 30B model on 26.02 release.

Affected Chips/Configurations:
Qwen3 235B: B300, B200, H100
Qwen3 30B: GB300, GB200, B200, H100

Minimal repro

Launch the Qwen3 30B and 235B via the `setup_experiment.py`

Expected behavior

FP8 should be faster than BF16

Affected area

area:perf

Regression?

Yes

Environment

No response

Logs

[malay-nagda]

@rhmukundan there is no other duplicate GitHub issue for this, right?

[dingqingy-nv]

@rhmukundan Duplicate of #2842. Please use existing issues. Closed