Files
nvfp4-megamoe-kernel/vllm/patches/deepseek_v4.py
biondizzle 10c14ddb49 Fix NVFP4 mapper: layer norms, hc params, indexer path, q_a_norm
- input_layernorm → attn_norm, post_attention_layernorm → ffn_norm
- hc_head.fn/base/scale → hc_head_fn/base/scale
- attn_hc/ffn_hc → hc_attn/hc_ffn (dot to underscore)
- q_a_norm → q_norm, sinks → attn_sink
- Indexer params: self_attn.compressor.indexer → attn.indexer
  (not attn.mla_attn.compressor.indexer)
2026-05-19 00:24:26 +00:00

71 KiB