Skip to content

MiniMax M3 GGUF fails to load: missing indexer.head_count metadata #8360

Description

@leftbankvines

Summary

Unsloth Studio on Apple Silicon cannot load the cached unsloth/MiniMax-M3-GGUF UD-Q5_K_XL variant.

Environment

  • macOS, Apple M3 Ultra, 512 GB unified memory
  • Unsloth Studio package: 2026.8.10
  • Bundled llama.cpp: b10333 (8231608b0)
  • Model: unsloth/MiniMax-M3-GGUF, UD-Q5_K_XL
  • First shard: MiniMax-M3-UD-Q5_K_XL-00001-of-00008.gguf

Studio recognizes the model, reads its metadata, and reduces the requested 1,048,576-token context to 832,000 to fit memory. It then exits before allocating the model:

llama_model_load: error loading model: error loading model hyperparameters:
key not found in model: minimax-m3.attention.indexer.head_count

llama_model_load_from_file_impl: failed to load model
common_fit_params: encountered an error while trying to fit params to free device memory: failed to load model

The Studio UI consequently reports only:

Failed to load model: llama-server failed to start.
Check that the GGUF file is valid and you have enough memory.

This is not a memory-pressure problem: the fatal error is the missing MiniMax M3 sparse-attention metadata key. KV cache is set to f16; the server fails before KV-cache or speculative-decoding behavior can affect startup.

The GGUF appears to predate the current M3 metadata requirements. Could you please confirm whether the published M3 GGUF variants need to be regenerated with minimax-m3.attention.indexer.head_count, or whether Studio/llama.cpp should retain backward compatibility with the existing artifacts?

I can provide the full llama-server log if useful.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions