[Bugfix] Fix Qwen3 MoE GPTQ inference

[Isotr0py]

Purpose

• Fix [Bug]: PR #22017 break Qwen3 MoE GPTQ loading #23467

Test Plan

python examples/offline_inference/basic/generate.py --model Qwen/Qwen3-30B-A3B-GPTQ-Int4/ --max-model-len 4096 --enforce-eager

Test Result

Processed prompts: 100%|███████████████████████████████████████████████████████████████████████| 4/4 [00:00<00:00,  7.06it/s, est. speed input: 38.87 toks/s, output: 113.07 toks/s]
--------------------------------------------------
Prompt: 'Hello, my name is'
Generated text: ' John, and I am 12 years old. I am currently in the'
--------------------------------------------------
Prompt: 'The president of the United States is'
Generated text: ' the commander-in-chief of the United States military. The president is also the head'
--------------------------------------------------
Prompt: 'The capital of France is'
Generated text: ' Paris. The capital of Germany is Berlin. The capital of the United Kingdom is'
--------------------------------------------------
Prompt: 'The future of AI is'
Generated text: ' not just about the technology itself, but also the way it is integrated into our'
--------------------------------------------------

(Optional) Documentation Update

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[gemini-code-assist]

Code Review

This pull request introduces a fix for GPTQ inference on Qwen3 MoE models. The core change is to prevent the quantization of the 'gate' layer within the MoE blocks when using GPTQ or GPTQ-Marlin quantization. This is achieved by adding a check that passes a None quantization configuration to the gate's ReplicatedLinear layer for these specific quantization methods. This is a common practice for MoE models where the gating mechanism is sensitive and often kept at a higher precision. The changes are well-contained and appear correct for the stated purpose. I have not found any high or critical severity issues in this pull request.