biondizzle b0b5113467 Fix weight mapper: compressor → attn.compressor (not mla_attn), quant weights_proj ...

- The compressor is on attn.compressor (not attn.mla_attn.compressor)
- weights_proj in indexer is NVFP4-quantized in our checkpoint

2026-05-19 03:20:41 +00:00

..

fused_moe

Eliminate weight copies: pass stacked checkpoint tensors directly

2026-05-19 02:16:43 +00:00

layers

Proper NVFP4 integration: quantized compressor/indexer + mapper fixes

2026-05-18 23:20:13 +00:00

deepseek_v4_attention.py

Fix weight mapper: compressor → attn.compressor (not mla_attn), quant weights_proj

2026-05-19 03:20:41 +00:00

deepseek_v4.py

Fix weight mapper: compressor → attn.compressor (not mla_attn), quant weights_proj

2026-05-19 03:20:41 +00:00

kernel.py

Add cutedsl to MoEBackend type in kernel config

2026-05-18 22:38:41 +00:00

register_cutedsl_kernel.py

Register CuTeDSL as proper NvFp4LinearKernel for NVFP4 linear layers

2026-05-19 00:44:44 +00:00