The MTP (multi-token-prediction) block was inadvertently NVFP4-quantized by the export pipeline. This reverts it to the base model's native precision and updates the metadata accordingly.
- model-00064-of-00064.safetensors: MTP layer (shard 64) reverted to native format — FP8 dense + native-FP4 routed experts, matching the base DeepSeek-V4-Pro release.
- model.safetensors.index.json: MTP tensor keys restored to the native set; metadata.total_size recomputed.
- hf_quant_config.json: removed MTP routed-expert entries from quantized_layers; added `mtp.*.ffn.experts.*` to exclude_modules.
- README.md: removed the now-merged vLLM PR #42209 requirement note.
Main transformer layers (0–60) are unchanged and remain NVFP4.
Drop FlashInfer MoE env vars from vLLM deploy command
