[autoresearch] Goal: fold block-scaled FP8 into DeepGEMM's MegaMoE (the H200-now subset)

[wseaton]

Autoresearch run 20260722T150510Z-goal-fold-block-scaled-fp8-into-deepgemm.

Goal: Goal: fold block-scaled FP8 into DeepGEMM's MegaMoE (the H200-now subset)

Gate: correctness

Kept candidates:

• iter 1: # Candidate: Block-scaled FP8 MegaMoE on SM90 (H200) ## Summary Added an SM90 (Hopper/H200) execution path for `fp8_fp
• iter 2: # Candidate: Block-scaled FP8 MegaMoE on SM90 — grouped GEMM path ## Summary Replaced the pure-Python per-expert dequa
• iter 3: # Candidate: Block-scaled FP8 MegaMoE on SM90 — SwiGLU path optimization ## Summary Optimized the SM90 FP8 MegaMoE pat
• iter 4: # Candidate: Block-scaled FP8 MegaMoE on SM90 — fused SwiGLU+requant ## Summary Optimized the SM90 FP8 MegaMoE path by
• iter 7: # Candidate: Block-scaled FP8 MegaMoE on SM90 — direct FP8 gather with per-32→per-128 rescaling ## Summary Optimized t
• iter 8: # Candidate: Block-scaled FP8 MegaMoE on SM90 — BF16 SwiGLU to reduce memory traffic ## Summary Optimized the SM90 FP8
• iter 10: # Candidate: Block-scaled FP8 MegaMoE on SM90 — eliminate redundant float casts in SwiGLU+requant ## Summary Optimized

Model claude-opus-4-6 · cost $69.50 · 40203s.

🤖 opened by crucible autoresearch (draft — review and steer).