convert: read GGUF (any ggml quant level) as a conversion source

[giveen]

Summary

• Adds tools/convert/sources/gguf.py: GGUFSource/HFAliasSource let a recipe source some or all logical parameters from a GGUF file (F16 through low-bit ggml/I-quants) instead of the full Safetensors checkpoint, via the existing "Read another source" recipe mechanism — no changes needed to model.py/methods.py/recipe.py.
• Covers Qwen3.5's hybrid GDN layers (qwen35_linear_attention_name_map), the MTP extra layer (qwen35_mtp_name_map), and the vision tower + qwen3vl_merger projector exported as a separate mmproj-*.gguf (qwen35_vision_name_map), plus a generic standard_dense_name_map for ordinary attention/MLP/norm tensors. DFlash2 needs no new code — it already reads natively via the existing --source dflash2=PATH Safetensors path.
• Every name/shape/reorder map was checked end-to-end against real checkpoint/GGUF pairs (dequantized GGUF values compared back to the original HF tensors, not just names/shapes) before landing. That caught two wrong assumptions along the way: the GDN head reorder needs swapped arguments to invert (calling it twice with the same arguments does not round-trip except in the square case), and llama.cpp's MTP export does not apply the same norm.weight + 1 shift the base model's layers get.
• Fixes GGUFSource to decode only the GGUF-native row range a read actually needs, instead of dequantizing the whole tensor on every call. GGUF's raw tensor data is already row-structured (ReaderTensor.data.shape[0] always matches the tensor's outer dimension), so this is exact, not approximate. recipe.prepare()'s 1-element preflight probe was fully dequantizing every tensor once for the probe and once again for the real write; this was roughly doubling conversion time and is now near-free (a 1.27B-element tensor's preflight-style read went from ~3.5s to ~0.0002s in testing).
• tools/convert/__main__.py: --source NAME=path.gguf now opens as a GGUFSource automatically (suffix-detected), same as any other named source.
• docs/weight-conversion.md: new "Convert from a GGUF quant instead of the full checkpoint" section with a worked override-recipe example, plus a subsection on MTP/vision/DFlash2.
• tests/convert/test_gguf.py: synthetic-fixture coverage (built with the gguf package's own writer, no external model files needed) for row-partial reads, name-map construction, the GDN reorder's swapped-argument inverse, the norm shift, and the vision patch-embed composite reconstruction.

Test plan

• python3 -m pytest tests/convert/ — 42 passed (14 new)
• Manually validated GGUFSource/HFAliasSource/all four name-map functions against real Qwen3.8-27B checkpoint/GGUF pairs (Q5_0, IQ2_M, Q8_0) across text, GDN, MTP, and vision tensors, comparing dequantized values back to the HF originals
• End-to-end: converted a real checkpoint from Q5_0/IQ2_M/Q8_0 GGUF sources (text-only and full text,vision,mtp,dflash2) and loaded each resulting .ninfer in ninfer-serve, including a real image through the vision path and both --spec dflash2 and --spec mtp speculative decoding (draft acceptance and correct, baseline-matching output confirmed for both)

🤖 Generated with Claude Code