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

Clean up L0 B1 verify noise (gate on VERBOSE), update FINAL_STRETCH.md

Browse Source 
Batched prefill + T>128 chunking now complete. All dangling items in
FINAL_STRETCH.md are marked done.
This commit is contained in:
biondizzle
2026-06-03 08:12:54 +00:00
parent 5e09be08af
commit 019a3a34b7
2 changed files with 7 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
FINAL_STRETCH.md
View File

@@ -91,6 +91,6 @@ Native Blackwell FP8 GEMM via tcgen05 for CSA Lightning Indexer scoring. No PyTo
# PART D — Dangling TODOS
- Batched Prefill: ✅ DONE (T=1..128, mixed FP8/BF16 kernel)
- Need to wire prefill into single_shot_inference.py (replace T=1 token-by-token prefill)
- Need T>128 support (split into multiple launches)
- Batched Prefill: ✅ DONE (T=1..128, mixed FP8/BF16 kernel, chunked for T>128)
- Prefill wired into single_shot_inference.py: ✅ DONE (chunked batched prefill replaces T=1 token-by-token)
- T>128 support: ✅ DONE (splits into multiple launches of ≤128 tokens each)