biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
2,400 Commits 1 Branch 17 Tags
master
Commit Graph

3 Commits

Author SHA1 Message Date
biondizzle
df05289d6f CUDA graph: Fix remaining sync violations from B200 detector run 2 
1. grouped_linear.py: Remove conditional host read of GPU tensor
   - 'if group_offsets[0] != 0' reads GPU value on host → sync
   - Fix: unconditionally update offsets every call (GPU-only multiply)

2. test_cuda_graph_readiness.py: Use pinned CPU buffers for token transfer
   - dec_tid_buf[0] = python_int → CPU→GPU sync
   - Fix: write to pinned CPU buffer, then copy_ (async, graph-capturable)

3. Add dsv4/decode/cuda_graph_decoder.py (skeleton)
 2026-06-03 17:20:34 +00:00
biondizzle
caac8ae108 Fix syntax error: 'is not not None' -> 'is not None' 2026-06-03 16:34:33 +00:00
biondizzle
ba68212fa7 Add CUDA graph readiness detector (Section A of GETTING_CUDAGRAPH_READY.md) 
- Grep for Section B sync patterns in hot path files
- Method 1: run decode forward with torch.cuda.set_sync_debug_mode('error')
- Method 2: attempt CUDA graph capture of L0 decode step
- Full model load + prefill + warmup before detection
- Results saved to /tmp/cuda_graph_readiness_results.json
 2026-06-03 16:34:15 +00:00