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

..

_archive

Cleanup Step 2: Archive Lineage P code, fix broken imports

2026-06-02 19:27:07 +00:00

cache

Cleanup Step 2: Archive Lineage P code, fix broken imports

2026-06-02 19:27:07 +00:00

kernels

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

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

layers

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

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

loader

Cleanup Step 2: Archive Lineage P code, fix broken imports

2026-06-02 19:27:07 +00:00

model

Cleanup Step 2: Archive Lineage P code, fix broken imports

2026-06-02 19:27:07 +00:00

ops

Fix non-contiguous tensor in quantize_nvfp4_gpu_fused (T>1 prefill)

2026-06-03 07:56:19 +00:00

reference

Update PyTorch reference with official DSV4 encoding + batched prefill

2026-06-03 12:19:38 +00:00

__init__.py

Restructure: cutedsl/ -> dsv4/ with proper layering

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