[CI][xlite] Match xlite backend with the current main and add e2e MoE model test coverage

[SijieFu]

What this PR does / why we need it?

This PR brings the xlite backend in line with the current main and vllm==0.24.0, refactors xlite model adapter/docking with its backend implementation, re-enables and adds e2e coverage for the xlite graph mode, and bumps the pinned xlite dependency to the latest dev release.

Refer to PR #12019 for the context.

More specifically,

1. The earliest commit completes the main2main update for xlite;
2. The second commit refactors xlite model adapters to remove duplicated model weight extraction;
3. The third commit adds a two-card e2e test coverage using an MoE model, to catch future main2main discrepancies;
4. The last commit bumps the xlite package to its latest dev vesion.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

e2e test

The following tests were run locally on Atlas A3 and passed:

python tests/e2e/pull_request/one_card/test_xlite.py
python tests/e2e/pull_request/two_card/test_xlite.py

main2main update + refactoring

We benchmarked model accuracies with multiple models using aisbench. The batch_aisbench.py script was used with 208 questions from the ceval dataset:

# on one Atlas A3
python batch_aisbench.py "/root/benchmark" \
--num-prompts 4 \
--model-dir "/mnt/sdb/models" \
--models "Qwen3-32B-w8a8-nopdmix" "Qwen3-30B-A3B" "Qwen3-VL-32B-Instruct" "MiniMax-M2.7-w8a8-QuaRot" "GLM-4.7-W8A8-floatmtp" \
--tps 2 2 2 4 8 \
--dps 1 1 1 1 1 \
--eps 0 1 0 1 1 \
-MNS 32 \
-MML 8192 \
--xlite 2 1 0 \
--broadcast-xlite

For most models, no accuracy regression was observed, except:

1. Qwen3-VL-32B-Instruct served with xlite decode-only mode showed degraded accuracy - upon further investigation, the model's outputs contained random texts for certain questions, likely a previously existing issue (e.g., PR [BugFix][xlite] Performance improvement in xlite full mode by avoiding redundant memory allocation during profiling runs #11380).
2. MiniMax-M2.7-w8a8-QuaRot served with xlite decode-only + aclgraph for prefill and aclgraph threw an AttributionError during serving, likely not related to xlite since xlite full mode worked correctly.

Model	Dataset	TP	EP	DP	Backend	Metric	Accuracy	Error
Qwen3-32B-w8a8-nopdmix	ceval-weighted	2	N	1	xlite full	weighted_average	88.78
Qwen3-32B-w8a8-nopdmix	ceval-weighted	2	N	1	xlite decode-only	weighted_average	88.63
Qwen3-32B-w8a8-nopdmix	ceval-weighted	2	N	1	aclgraph	weighted_average	87.09

Model	Dataset	TP	EP	DP	Backend	Metric	Accuracy	Error
Qwen3-30B-A3B	ceval-weighted	2	Y	1	xlite full	weighted_average	87.74
Qwen3-30B-A3B	ceval-weighted	2	Y	1	xlite decode-only	weighted_average	85.68
Qwen3-30B-A3B	ceval-weighted	2	Y	1	aclgraph	weighted_average	86.13

Model	Dataset	TP	EP	DP	Backend	Metric	Accuracy	Error
Qwen3-VL-32B-Instruct	ceval-weighted	2	N	1	xlite full	weighted_average	85.46
Qwen3-VL-32B-Instruct	ceval-weighted	2	N	1	xlite decode-only	weighted_average	77.84
Qwen3-VL-32B-Instruct	ceval-weighted	2	N	1	aclgraph	weighted_average	88.52

Model	Dataset	TP	EP	DP	Backend	Metric	Accuracy	Error
MiniMax-M2.7-w8a8-QuaRot	ceval-weighted	4	Y	1	xlite full	weighted_average	84.16
MiniMax-M2.7-w8a8-QuaRot	ceval-weighted	4	Y	1	xlite decode-only	weighted_average	0.00	AttributeError: 'AscendMoERunner' object has no attribute 'maybe_all_reduce_tensor_model_parallel'
MiniMax-M2.7-w8a8-QuaRot	ceval-weighted	4	Y	1	aclgraph	weighted_average	0.00	AttributeError: 'AscendMoERunner' object has no attribute 'maybe_all_reduce_tensor_model_parallel'

Model	Dataset	TP	EP	DP	Backend	Metric	Accuracy	Error
GLM-4.7-W8A8-floatmtp	ceval-weighted	8	Y	1	xlite full	weighted_average	87.41
GLM-4.7-W8A8-floatmtp	ceval-weighted	8	Y	1	xlite decode-only	weighted_average	87.82
GLM-4.7-W8A8-floatmtp	ceval-weighted	8	Y	1	aclgraph	weighted_average	87.74

• vLLM version: v0.24.0
• vLLM main: vllm-project/vllm@85c09e9

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request aligns the xlite backend with the latest main branch and vLLM version. It introduces a significant refactoring of the model adapter hierarchy, centralizing weight loading logic to improve maintainability. Additionally, the PR expands test coverage by adding new e2e tests for MoE models and updating existing test configurations to ensure stability across different hardware setups.

Highlights

• Backend Synchronization and Dependency Update: Synchronized the xlite backend with the current main branch and vLLM 0.24.0, and bumped the xlite dependency to 0.1.0rc12.dev210.
• Adapter Architecture Refactoring: Consolidated model adapter logic into a new StandardXliteModel base class to reduce code duplication and streamline weight extraction.
• Enhanced Test Coverage: Added e2e test coverage for MoE models on two NPU cards and updated existing test markers to improve CI reliability.

New Features

🧠 You can now enable Memory (public preview) to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Ignored Files

• Ignored by pattern: .github/workflows/** (1)
  • .github/workflows/scripts/test_config.yaml

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize the Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counterproductive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

Suggested PR Title:

[Ops][Misc] Refactor Xlite model adapters and add MoE multi-card tests

Suggested PR Summary:

### What this PR does / why we need it?
This pull request refactors the xlite integration by consolidating MoE adapter weight-loading logic into a unified StandardXliteModel base class, upgrading the xlite dependency, and adding two-card MoE e2e tests. However, the refactoring introduces critical bugs where conditionally assigned attributes (mha_q_norm, mha_k_norm, and mha_qkv_bias) are accessed immediately after the condition block, which will raise AttributeError crashes when the conditions are not met.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
Tested with updated single-card and new two-card e2e tests for MoE serving.