biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add PART A diagnostic tests: compressor + KV cache + FMHA at production scale

Browse Source
This commit is contained in:
biondizzle
2026-06-03 04:13:53 +00:00
parent 75288bd12f
commit 04cf8ca848
3 changed files with 457 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
CORRRECTNESS_BACKLOG.md
View File

@@ -98,3 +98,10 @@ Let me check what seq_len the FMHA is seeing. At L1 during prefill of the first
```
SO SINCE WE HAD TO TOUCH FMHA ANYWAY IN PART B. WE DID THAT FIRST AND TRIED TO GET THAT CORRECT BEFORE WE REVISTED THIS ISSUE!!!
### UPDATE (2026-06-03): FMHA accuracy fixed by B1 mixed FP8 decode kernel
- Per-layer FMHA cos is now 0.999993+ across all 5 tested layers (was 0.679 at L1)
- The old BF16 decode path had a subtle V-matrix layout issue; B1 kernel with FP8/BF16 native storage eliminates it
- Decode output is STILL degenerate (loops on capital/Capitalization) despite correct FMHA
- The issue is NOT in the FMHA — it's in another part of the pipeline (mHC, compression, KV gathering, or RoPE)
- We will revisit this after completing the remaining FINAL_STRETCH items