Summary

Three small fixes in convert_hf_to_gguf.py to make --mistral-format work
end-to-end on Pixtral 12B (2409) consolidated weights. Without these,
conversion crashes before writing the GGUF; with them, F16 -> Q4_K_M -> mtmd
inference produces correct image output. No inference-side changes required --
tools/mtmd/clip.cpp already reads the bias tensors when present.

Bugs

1. LlamaModel.__init__ crashes on mistral-format input (line ~2867)

ModelBase.load_hparams(self.dir_model, is_mistral_format=False) is called
unconditionally to read architectures for origin_hf_arch (used downstream
to detect SmolVLM2 etc.). Mistral consolidated layouts have no config.json,
so this raises FileNotFoundError for any --mistral-format run.

Fix: skip the HF-only lookup when self.is_mistral_format is True and set
self.origin_hf_arch = None. The field is only consulted by HF-specific
branches, so None is safe in the mistral path.

2. PixtralModel.set_gguf_parameters requires mm_projector_id (line ~13321)

self.find_vparam([mm_projector_id]) is mandatory, but Pixtral 12B 2409
uses a plain linear projector and ships no mm_projector_id in params.json
-- only the newer Mistral Small 3.1 sets it to patch_merge. Result: KeyError
on every Pixtral 12B conversion.

Fix: pass optional=True. The body inside the if only runs when the value
equals patch_merge, so None short-circuits correctly and the existing
patch-merge path is unaffected.

3. PixtralModel.map_tensor_name rejects adapter biases (line ~13330)

Only .weight is mapped for vision_language_adapter.w_in / w_out; .bias
falls through to super().map_tensor_name and raises. Pixtral 12B 2409's
consolidated.safetensors ships both biases.

Fix: add the two .bias branches mapping to mm.1.bias / mm.2.bias. The
inference side already supports this -- see tools/mtmd/clip.cpp:2143-2148,
which loads mm.1.bias / mm.2.bias with the optional flag set. The GGUF
writer was simply refusing tensors the runtime is happy to consume.

Reproduction

python convert_hf_to_gguf.py \
    /path/to/pixtral-12b-2409 \
    --mistral-format --outtype f16 \
    --outfile pixtral-12b-2409-f16.gguf

against the official Mistral release of Pixtral-12B-2409 (consolidated
safetensors layout, no config.json). Pre-patch this fails at
LlamaModel.__init__ (FileNotFoundError on config.json). After bypassing
that it fails at set_gguf_parameters (KeyError mm_projector_id). After
bypassing that it fails inside map_tensor_name on
vision_language_adapter.w_in.bias.

Verification

• F16 GGUF written successfully from consolidated.safetensors +
  params.json + tekken.json.
• Q4_K_M produced via llama-quantize, loads cleanly.
• llama-mtmd-cli smoke test on a test image returned correct,
  image-grounded output. CUDA inference at ~47 tok/s eval on RTX 3090.
• No upstream files touched besides convert_hf_to_gguf.py.

Inference side

tools/mtmd/clip.cpp:2143-2148 already calls
get_tensor(TN_MM_INP_PROJ_B, /*optional=*/true) and
get_tensor(TN_MM_OUTP_PROJ_B, /*optional=*/true) for mm.1.bias and
mm.2.bias, so the converter change makes existing runtime behavior
reachable. No C++ change is included or needed.