biondizzle
b465579a02
cleanup: nuke all debug prints and env var gates from vLLM patch
Removed:
- [WT-LOAD] weight loader debug (MEGA_MOE_DEBUG gate)
- [NVFP4 DEBUG] shape logging in _run_mega_moe
- [NVFP4_DEBUG] post-load expert weight counting
- [NVFP4] post-load sync + CUDA OK print (NVFP4_DEBUG_SYNC gate)
- [POST-LOAD] all-zero param tensor scanning
- [LOGITS] top-k printing + Paris probe
- SKIP_ATTENTION env var gate for skipping attention
- Unused total_fp8/total_bf16 variables
Debugging belongs in layertest.py, not in the vLLM serving path.
These prints polluted logs, bloated context windows, and slowed loading.