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

auto: pre-test commit

Browse Source
This commit is contained in:
biondizzle
2026-06-03 09:01:02 +00:00
parent 3c06fd5591
commit 572bdd2840
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
TEMP/CROSS_REFERENCE.md
View File

@@ -1,6 +1,6 @@
# vLLM Independent-Reference Cross-Checks (DSV4)
**Purpose:** catch the architectural bugs your cosine backlog is blind to. Every check in `CORRECTNESS_BACKLOG.md` compares your kernel to your *own* PyTorch reference — if both share a wrong assumption (RoPE position, RoPE style, norm ordering, compression overlap), they agree at cos 0.9999 while both being wrong. vLLM's DSV4 is an **independent** implementation; it can't share your blind spots. Use it as the tiebreaker.
**Purpose:** catch the architectural bugs your cosine backlog is blind to. Every check in `/home/openclaw/dev/nvfp4-megamoe-kernel/DEGENERATION_TESTS.md` and `/home/openclaw/dev/nvfp4-megamoe-kernel/CORRECTNESS_BACKLOG.md` compares your kernel to your *own* PyTorch reference — if both share a wrong assumption (RoPE position, RoPE style, norm ordering, compression overlap), they agree at cos 0.9999 while both being wrong. vLLM's DSV4 is an **independent** implementation; it can't share your blind spots. Use it as the tiebreaker.
**How to use this:** these are mostly *static code diffs*, not GPU runs — fast. For each check, find the line in `single_shot_inference.py` (or `dsv4/`), find the cited line in the vLLM tree, and record match / mismatch + the fix.