[#17723][fix] Preserve rowwise FP8 Qwen lm_head weights

[mihai-chiorean]

Summary

Keep an explicit FP8_PER_CHANNEL_PER_TOKEN Qwen3.5/3.8 lm_head entry quantized instead of forcing the head to bf16. Reject unsupported FP8 layouts where fallback would cast the FP8 weight and discard its scale.

Related to #17723.

Motivation

unsloth/Qwen3.8-27B-NVFP4 stores lm_head.weight as E4M3 plus a rowwise scale. The previous Qwen normalization always excluded lm_head from quantization; the generic loader then assigned the E4M3 tensor to a bf16 parameter and ignored its scale, corrupting logits.

Safety

The quantized path is enabled only for the rowwise FP8 format demonstrated by the checkpoint. Tied embeddings, lm_head TP under attention-DP, missing/non-divisible vocabulary metadata, and unsupported FP8 formats raise instead of silently losing scale information. Other quantized algorithms retain their existing fallback behavior. Existing W4A16 NVFP4 behavior is preserved.

Ordinary attention-DP without lm_head TP is allowed because LMHead is TP-less in that mode; this matches DecoderModelForCausalLM._resolve_lm_head_quant_config.

Testing

• pre-commit run --files on all changed files
• Focused head tests included in a 200-pass integrated Spark run
• Real SM121 load constructed LMHead as FP8_PER_CHANNEL_PER_TOKEN with weight shape (248320, 5120) and E4M3 dtype
• Full model generated coherent output both without MTP and with the checkpoint MTP layer

Dependencies

Qwen3.8 receives the explicit rowwise-FP8 head entry from #17725.