biondizzle e46b615873 Fix dense router BF16 dispatch for CUDA graph capture ...

- Run GEMM in BF16 (not FP32) during graph capture — Blackwell tensor cores
  handle BF16 natively; FP32 GEMM triggers cudaErrorStreamCaptureUnsupported
- Pre-transpose W_gate to [E, H] at load time — avoids .T view during capture
- Convert only logits output to FP32 for sqrt(softplus) numerical stability
- This fixes the graph capture failure at layer 0 Graph B

2026-06-04 05:50:13 +00:00

..

__init__.py

Wire NVFP4 fused router kernel into e2e single-shot pipeline

2026-06-01 09:47:48 +00:00

_activation_topk.py

Rewrite NVFP4 fused router kernel: MoE-style epilogue replaces broken SMEM merge

2026-06-01 09:59:34 +00:00

dense_router_decode.py

Fix dense router BF16 dispatch for CUDA graph capture

2026-06-04 05:50:13 +00:00