biondizzle 188ecae47f CUDA graph: Eliminate per-step allocations in graph-captured code paths ...

- gemm_runner.py: Add out= parameter to run_nvfp4_grouped_gemm and
  run_fused_swiglu_grouped_gemm to accept pre-allocated output buffers
- quantize.py: Replace torch.zeros_like/torch.zeros with scalar 0.0 in
  torch.where() calls (graph-capturable, no memory allocation)
- Both fixes prevent 'Disallowed operation during CUDA stream capture'
  errors during graph capture

2026-06-03 21:30:24 +00:00

19 KiB

Raw Blame History

View Raw