gguf: Q5_0 weights are unsupported — llama-quantize's Q4_K_M of small models does not load #675

Found by the conversion pipeline of #623.

llama-quantize ... Q4_K_M keeps some tensors of small models in Q5_0. For Gemma 3 270M this includes blk.*.attn_q.weight. geist refuses the load:

geist_model_load failed: transformer: 'blk.0.attn_q.weight' has unsupported dtype Q5_0

The same file runs in llama.cpp (perplexity 46.0 on tools/data/ppl_alice.txt). A user who quantizes a small model with the standard tool therefore cannot run it in geist.

Proposal: support GGML_TYPE_Q5_0 (32-element blocks: fp16 d, 4 bytes of high bits, 16 bytes of nibbles) at least through the reference dequant + cpu_scalar linear, with the fast paths following the Q4_0/Q5_K kernels.

Done when:

• The Q4_K_M of Gemma 3 270M and of Qwen3 0.6B from tools/convert_hf.py convert --quant q4_k_m load.
• The perplexity gate passes for both.
• A dequant unit test against the GGML reference layout exists.

🤖 Generated with Claude Code