[fix] Qwen3.5 AWQ quantization startup crash

[yyj6666667]

Here's a filled-in PR description for cherry-picking the Qwen3.5 AWQ fix:

---

Motivation

Cherry-pick fix from sgl-project/sglang#20370 to resolve Qwen3.5 AWQ quantization startup crash:

RuntimeError: gptq_marlin_repack.cuh:309: size_n = 32 is not divisible by tile_n_size = 64

The upstream PR has not been merged yet. Bringing the fix in early so our fork can support Qwen3.5 AWQ models now.

Modifications

• qwen3_5.py: Added hf_to_sglang_mapper override for Qwen3_5ForConditionalGeneration and Qwen3_5MoeForConditionalGeneration so quant ignore names match correctly and in_proj_a/in_proj_b layers are not mistakenly quantized.
• model_runner.py: Added Blackwell backend guard to auto-switch from flashinfer to triton for Qwen3.5 GDN models.

All changes are identical to upstream PR #20370.

Accuracy Tests

See upstream PR for results (4B AWQ 4-bit: 74.4 on C-Eval vs 75.3 original).

Benchmarking and Profiling

N/A — bug fix only, no performance impact expected.

[gemini-code-assist]

Code Review

This pull request introduces model-specific adjustments for Qwen3.5 models on Blackwell GPUs. It includes logic to automatically switch the attention backend to triton for Qwen3.5 hybrid GDN models on Blackwell hardware and adds a WeightsMapper to ensure correct weight name mapping for quantization. I have no feedback to provide.

[Copilot]

Pull request overview

Cherry-picks an upstream fix to prevent Qwen3.5 AWQ quantized model startup crashes by ensuring quantization “ignore” module-name patterns align with Qwen3.5’s internal naming, and adds a hardware-specific backend adjustment for Blackwell GPUs when running Qwen3.5 hybrid GDN models.

Changes:

• Add hf_to_sglang_mapper overrides for Qwen3_5ForConditionalGeneration and Qwen3_5MoeForConditionalGeneration to keep quant ignore-name mapping consistent with Qwen3.5 layer names.
• In ModelRunner.model_specific_adjustment(), auto-switch attention backend from flashinfer to triton for Qwen3.5 hybrid GDN on Blackwell.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
python/sglang/srt/models/qwen3_5.py	Overrides weight-name mapping to keep AWQ quant ignore lists matching Qwen3.5 module names (avoids unintended quantization of sensitive layers).
python/sglang/srt/model_executor/model_runner.py	Adds a Blackwell-specific guard to select a compatible attention backend for Qwen3.5 hybrid GDN.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.