fix(models): fix MoE weight dangling reference and Qwen3-Omni model adapter compatibility

[sunnyxiaohu]

Summary

Fix critical runtime issues in MoE expert weight handling and Qwen3-Omni model adapter that caused GPTQ quantization failures.

Changes

1. Fix MoE expert weight dangling reference (angelslim/models/llm/qwen.py)

After chunk() splits gate_up_proj into gate_proj and up_proj, the resulting tensors are views sharing the same underlying storage. When del self.gate_up_proj is executed subsequently, the storage is freed, leaving gate_proj and up_proj as dangling references pointing to invalid memory.

Fix: Call .clone() on all chunked/sliced weight tensors immediately after assignment to ensure each expert owns an independent copy before the source tensors are deleted.

2. Add missing block_name attribute to Qwen3-Omni (angelslim/models/omni/qwen3_omni.py)

The GPTQ quantization flow requires self.block_name to locate transformer blocks. Qwen3-Omni only defined thinker_block_name / talker_block_name but was missing the base block_name attribute, causing AttributeError during calibration.

Fix: Add self.block_name = "thinker.model.layers" to the constructor.

3. Remove incompatible self.model.use_cache = False (angelslim/models/omni/qwen3_omni.py)

The Qwen3-Omni model object does not expose a use_cache attribute at the top level (it is configured per-component). Setting it unconditionally raised AttributeError.

Fix: Remove the incompatible assignment from model_forward().

Files Changed

• angelslim/models/llm/qwen.py — clone expert weights after chunk to prevent dangling references
• angelslim/models/omni/qwen3_omni.py — add block_name attr; remove invalid use_cache assignment