Skip to content

mega: wire block-scaled FP8 weights into the fused MoE slots#1

Draft
wseaton wants to merge 1 commit into
nv_devfrom
autoresearch/run6-blockfp8-megamoe
Draft

mega: wire block-scaled FP8 weights into the fused MoE slots#1
wseaton wants to merge 1 commit into
nv_devfrom
autoresearch/run6-blockfp8-megamoe

Conversation

@wseaton

@wseaton wseaton commented Jul 17, 2026

Copy link
Copy Markdown

Agent-authored candidate from crucible run 6 of the blockfp8 brief (recovered by hand — publish-on-keep wasn't wired yet).

Wires the existing block-scaled FP8 grouped-GEMM path into the MegaMoE fused kernel slots (GEMM1/GEMM2), scales carried through the SwiGLU requant boundary. 5 files, +193/-6.

Grading state: rung 1 (frozen reference self-check) passed; rung 2 (H200 calc_diff) crashed on a harness env bug (missing PYTHONPATH for the in-tree deep_gemm build), NOT a kernel verdict — fix is in and the candidate will be re-graded by the next run. Treat as ungraded until then.

🤖 Generated with Claude Code

@wseaton

wseaton commented Jul 17, 2026

Copy link
Copy Markdown
Author

Manual re-grade of this candidate with the rung-2 env fix (PYTHONPATH=/opt/deepgemm): PASSED the frozen calc_diff correctness terminal on H200 — {"calc_diff": 0.0, "pass": true}, full FP8 grouped-GEMM numeric suite green, ~1150-1300 TFLOPS on the perf sweep. The earlier discard was the harness env bug, not the kernel. Un-drafting is a human call.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant