[Copilot]

Overview

Nemotron Labs 3 Puzzle conversion was failing in convert_hf_to_gguf.py after weight indexing because the Puzzle path assumed only mtp.layers.{0,1}.* tensors and did not robustly handle top-level mtp.* tensors.
This change makes MTP remapping resilient to actual Puzzle checkpoint naming and to variable MTP layer counts from config.

• MTP name handling

  • Added explicit remap support for non-layer tensors:
    • mtp.fc → model.layers.{bid}.eh_proj
    • mtp.pre_fc_norm_embedding → model.layers.{bid}.enorm
    • mtp.pre_fc_norm_hidden → model.layers.{bid}.hnorm
    • mtp.norm → model.layers.{bid}.shared_head.norm
  • Applies shared/top-level MTP tensors across all configured MTP blocks.

• Dynamic MTP block indexing

  • Replaced fixed bid in (0, 1) assumption with bounds from len(self.mtp_block_configs).
  • Keeps layer-local remapping for eh_proj/enorm/hnorm/final_layernorm and routes remaining mtp.layers.{bid}.* tensors through the existing backbone mapping path.

• Failure semantics

  • Improved assertion text for invalid MTP block index values to make malformed tensor names/config mismatches easier to diagnose.

# New non-layer remap path (conceptual)
remapper = {
    "mtp.fc": "model.layers.{bid}.eh_proj",
    "mtp.pre_fc_norm_embedding": "model.layers.{bid}.enorm",
    "mtp.pre_fc_norm_hidden": "model.layers.{bid}.hnorm",
    "mtp.norm": "model.layers.{bid}.shared_head.norm",
}

Additional information

The warning about trust_remote_code=True during config loading is expected for this model family and not the conversion failure root cause; the crash path was in Puzzle MTP tensor remapping during export.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES — AI was used assistively to draft and refine targeted converter changes; human review/ownership remains with the submitter. Please ensure compliance with AGENTS.md and CONTRIBUTING.md before submission.