Qwen3-0.6B x86 v1 q4_k path fails due to lm_head name mismatch

[kkkzbh]

Summary

The documented x86 Qwen3-0.6B v1 q4_k path currently appears to fail before inference because the recommended v1 model file uses lm_head.weight, while the current Qwen3 FA2 implementation tries to load lm_head_out.weight.

Documented path

docs/cpu_backend/x86/index.rst recommends downloading the pre-converted model from the mllmTeam HuggingFace organization:

wget https://huggingface.co/mllmTeam/qwen-3-0.6b-mllm/blob/main/qwen-3-0.6b-q4_k.mllm

and then running it with -mv v1:

/path/to/build/bin/mllm-qwen3-runner \
  -m /path/to/model/qwen-3-0.6b-q4_k.mllm \
  -mv v1 \
  -t /path/to/tokenizer/tokenizer.json \
  -c /path/to/config/config_0.6B_w4a32_kai.json

The README mllm v1 model table also points Qwen3-0.6B CPU FP32/INT4 users to mllmTeam/qwen-3-0.6b-mllm.

What I found

I parsed the descriptor of the downloaded v1 .mllm file using the upstream v1 file layout. The file is a valid v1 model file and contains lm_head.weight, but not lm_head_out.weight:

magic 20012
param_count 311
has lm_head.weight true
has lm_head_out.weight false

However, the current Qwen3 FA2 implementation registers lm_head_out when tie_word_embeddings is enabled:

lm_head_ = reg<nn::Linear>("lm_head_out", cfg.hidden_size, cfg.vocab_size, false, cfg.linear_impl_type);

examples/qwen3/config_0.6B_w4a32_kai.json sets:

"tie_word_embeddings": true

so the runner tries to load lm_head_out.weight, which is not present in the documented v1 model file.

Possible cause

This looks like a mismatch between the older v1 pre-converted Qwen3-0.6B artifact and the newer Qwen3/KAI naming convention. The quantization config maps lm_head.weight to lm_head_out.weight, but the documented v1 model artifact still uses the old name.

Question

What is the preferred direction for fixing this path?

• keep backward compatibility with the documented v1 artifact by accepting lm_head.weight as a fallback for tied embeddings;
• update the x86 docs to point to a v2/newly converted artifact;
• or re-publish the pre-converted model with the lm_head_out.weight name?

I found a separate x86 FP16/Q4_K dequantization issue while investigating this path, and I will submit that as a separate PR because it affects the common x86 quantization code rather than only Qwen3.