biondizzle 0c3796966d Add BF16 fallback for shared expert: dequantize NVFP4 → BF16 F.linear ...

Same approach as router gate. enable_bf16_fallback() pre-materializes
BF16 weight matrices from NVFP4 weights, then _run_impl uses F.linear
instead of NVFP4 GEMM.

2026-06-03 13:19:27 +00:00

..

__init__.py

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00

grouped_linear.py

P0 complete: Kill .item() in grouped_linear, reduce hot-path syncs

2026-06-01 22:21:12 +00:00

linear.py

P5 integration + B3 q_a_norm fused + gsa scalar fix

2026-06-02 21:20:34 +00:00

mhc.py

Fix mHC Sinkhorn kernel: remove VLA, remove Python fallback

2026-06-02 10:44:53 +00:00

moe.py

Fix two correctness bugs: compressor pos bias on KV + SwiGLU clamp ordering

2026-06-03 11:17:49 +00:00

router.py

CRITICAL FIX: runtime activation global scale to prevent E4M3 overflow

2026-06-01 14:21:16 +00:00

shared_expert.py

Add BF16 fallback for shared expert: dequantize NVFP4 → BF16 F.linear

2026-06-03 13:19:27 +00:00