[Feature][MRV2] expand default MRv2 architecture whitelist and add dspark

[yjyang62]

What this PR does / why we need it?
This PR expands the default Model Runner V2 model and feature whitelists so more architectures, plus dspark, default to V2.
A request uses Model Runner V2 by default only when all of the following hold:

The model architecture is on the default-V2 whitelist:

• Qwen3ForCausalLM
• Qwen3MoeForCausalLM
• MiniMaxM2ForCausalLM
• DeepseekV3ForCausalLM
• DeepseekV32ForCausalLM
• GlmMoeDsaForCausalLM
• DeepseekV4ForCausalLM
• Qwen3_5MoeForCausalLM

Enabled features are on the default-V2 feature whitelist. Static eagle3 / mtp / dflash / dspark are included. LoRA and dynamic speculative decoding (num_speculative_tokens_per_batch_size) stay on V1.
Triton is available.
Otherwise the request stays on Model Runner V1. Explicit VLLM_USE_V2_MODEL_RUNNER=0/1 still overrides the whitelist.

Does this PR introduce any user-facing change?
Yes. The architectures above now default to Model Runner V2 when the conditions above are satisfied. Other models still default to V1. Explicit VLLM_USE_V2_MODEL_RUNNER=0/1 is unchanged.

How was this patch tested?
Unit tests updated: tests/ut/test_mrv2_utils.py (architecture parametrize, dspark on the feature whitelist)
CI: cpu-ut / e2e on this PR.

[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 expands the default Model Runner V2 (MRv2) support on Ascend platforms by updating the architecture and speculative method whitelists. It introduces a centralized utility module to manage MRv2 enablement, which decouples the logic from upstream validation and provides a more robust mechanism for applying configuration patches. Additionally, the PR improves the detection of ACL graph capturing to ensure compatibility with the V2 runner, preventing issues where GPU-specific flags were incorrectly interpreted in the NPU environment.

Highlights

• Expanded MRv2 Whitelist: Added support for multiple new model architectures (including Qwen3, DeepseekV3/V4, and others) and the 'dspark' speculative decoding method to the default Model Runner V2 (MRv2) whitelist.
• Centralized MRv2 Utilities: Created a new module, mrv2_utils.py, to encapsulate the logic for MRv2 enablement, decoupling it from upstream validation and providing a consistent patching mechanism.
• Robust Graph Capturing: Implemented a new helper, is_acl_full_graph_capturing, to accurately detect ACL stream capturing, preventing conflicts between GPU-style V2 context flags and NPU graph task groups.
• Configuration Patching: Standardized the application of MRv2 configuration overrides across the platform, worker processes, and engine core to ensure consistent runner behavior.

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.

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:

[Misc][Feature] Implement Ascend-owned whitelist heuristics for Model Runner V2 enablement

Suggested PR Summary:

### What this PR does / why we need it?

This PR implements Ascend-owned whitelist heuristics for Model Runner V2 (MRv2) enablement instead of relying on the upstream GPU-specific defaults, which can cause crashes on unsupported configurations. It introduces `vllm_ascend/mrv2_utils.py` to manage the whitelist of architectures (e.g., Qwen3, DeepseekV3) and speculative decoding methods (e.g., eagle3, mtp, dflash, dspark), ensuring MRv2 is only enabled when the environment is ready (non-310P and Triton available).

Additionally, it resolves an issue where GPU V2's `capturing` flag was incorrectly treated as ACL stream capture, which caused hangs. It introduces `is_acl_full_graph_capturing()` to verify actual NPU stream capture status.

Feedback on the code changes:
- Safely retrieve `model_config` and `speculative_config` using `getattr` to prevent potential `AttributeError`s.
- Add a warning log when Model Runner V2 is disabled due to an unsupported speculative decoding method to improve debuggability.

### Does this PR introduce _any_ user-facing change?

Yes, Model Runner V2 will now be enabled by default for whitelisted architectures (such as Qwen3) and supported speculative decoding methods on compatible Ascend platforms.

### How was this patch tested?

The changes are covered by new unit tests in `tests/ut/test_mrv2_utils.py` and updates to existing tests in `tests/ut/test_ascend_forward_context.py` and `tests/e2e/`.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[zouzy5137]

LGTM

[shiqiangA]

/nightly qwen3-30b-a3b-w8a8 kimi25_w4a8_step_3_7 kimi-k2.6_max_model_len acceptace_rate_deepseekv4-flash-w8a8-mtp Qwen3-235B-A22B-W8A8_piecewise_fullgraph_A3 Qwen3_32b_w8a8_ml_A3 qwen3-235b-a22b-w8a8 glm-5.2-w4a8-mtp glm-5.2-w4a8-dspark glm-5.2-w4a8c8-sfa-dcp compressor-metadata-cross-stream Qwen3-32B-QuaRot Qwen3_32b_W8A8_wl_A3 qwen3-32b-int8 qwen3-32b-int8-prefix-cache Qwen3-30B-A3B-W4A8-llm-compressor Qwen3-30B-QuaRot qwen3-30b-acc Minimax_m2.7_w8a8_A3 deepseek-v3-2-w8a8 Deepseek_R1_W8A8_Reasoning_output_A3 deepseek-r1-0528-w8a8-prefix-cache mtpx-deepseek-r1-0528-w8a8 --a3-560t
nightly command triggered.

• Nightly-A3-560T: workflow run

[shiqiangA]

/nightly kimi-k2-thinking Qwq_32b_A3 MiniMax-M3-BF16-A3 Qwen3.8-27B-w8a8-A3 qwen_3_6_27b_w8a8_A3 Qwen3.5-397B-A17B-w8a8-mtp Qwen3.5-27B-w8a8-A3 Qwen3.5-122B-A10B-W8A8-A3 qwen3-vl-235b-a22b-instruct-w8a8 Qwen3-32B-W8A8C8-A3 MiniMax-M3-W8A8-A3 Kimi-K2.6-w4a8-A3 kimi-k2.5 glm-5.1-w8a8-prefill-mc2 glm-4.7-w8a8 DeepSeek-V4-Flash-W8A8-A3 Deepseek_R1_W8A8_A3 rejection-sample gemma4-31b-dense gemma4 qwen3-30b-a3b-w8a8-a2-performance qwen3-32b-int8 Qwen3-Next-80B-A3B-Instruct Qwen3-ASR-1.7B Qwen3-8B qwen3-30b-a3b-bf16-a2-performance multi-node-qwen3-235b-dp Qwen3.5-397B-A17B-w4a8-mtp Qwen3.5-27B-w8a8-A2 qwen3-vl-32b-instruct-w8a8 MiniMax-M2.5-w8a8-QuaRot-A2 multi-node-Kimi-K2.5-W4A8-A2 multi-node-GLM-5.1-w8a8-A2 --a3-560t
nightly command triggered.

• Nightly-A3-560T: workflow run

[shiqiangA]

/nightly qwen3-30b-a3b-w8a8 kimi25_w4a8_step_3_7 kimi-k2.6_max_model_len acceptace_rate_deepseekv4-flash-w8a8-mtp Qwen3-235B-A22B-W8A8_piecewise_fullgraph_A3 Qwen3_32b_w8a8_ml_A3 qwen3-235b-a22b-w8a8 glm-5.2-w4a8-mtp glm-5.2-w4a8-dspark glm-5.2-w4a8c8-sfa-dcp compressor-metadata-cross-stream Qwen3-32B-QuaRot Qwen3_32b_W8A8_wl_A3 qwen3-32b-int8 qwen3-32b-int8-prefix-cache Qwen3-30B-A3B-W4A8-llm-compressor Qwen3-30B-QuaRot qwen3-30b-acc Minimax_m2.7_w8a8_A3 deepseek-v3-2-w8a8 Deepseek_R1_W8A8_Reasoning_output_A3 deepseek-r1-0528-w8a8-prefix-cache mtpx-deepseek-r1-0528-w8a8 --a3-560t
nightly command triggered.

• Nightly-A3-560T: workflow run

[shiqiangA]

/nightly kimi-k2-thinking Qwq_32b_A3 MiniMax-M3-BF16-A3 Qwen3.8-27B-w8a8-A3 qwen_3_6_27b_w8a8_A3 Qwen3.5-397B-A17B-w8a8-mtp Qwen3.5-27B-w8a8-A3 Qwen3.5-122B-A10B-W8A8-A3 qwen3-vl-235b-a22b-instruct-w8a8 Qwen3-32B-W8A8C8-A3 MiniMax-M3-W8A8-A3 Kimi-K2.6-w4a8-A3 kimi-k2.5 glm-5.1-w8a8-prefill-mc2 glm-4.7-w8a8 DeepSeek-V4-Flash-W8A8-A3 Deepseek_R1_W8A8_A3 rejection-sample gemma4-31b-dense gemma4 qwen3-30b-a3b-w8a8-a2-performance qwen3-32b-int8 Qwen3-Next-80B-A3B-Instruct Qwen3-ASR-1.7B Qwen3-8B qwen3-30b-a3b-bf16-a2-performance multi-node-qwen3-235b-dp Qwen3.5-397B-A17B-w4a8-mtp Qwen3.5-27B-w8a8-A2 qwen3-vl-32b-instruct-w8a8 MiniMax-M2.5-w8a8-QuaRot-A2 multi-node-Kimi-K2.5-W4A8-A2 multi-node-GLM-5.1-w8a8-A2 --a3-560t
nightly command triggered.

• Nightly-A3-560T: workflow run

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[shiqiangA]

/nightly QWEN3_235B_PD DeepSeek-V4-Pro-w4a8-1M-PD multi-node-deepseek-v3.2-W8A8-EP QWEN3_235B_PD_3_5K_1_5k multi-node-dpsk3.2-2node multi-node-GLM-5.1-w8a8-A3 multi-node-GLM-5.2-w8a8-A3 multi-node-qwenw8a8-2node-eplb multi-node-GLM-5.1-W8A8C8-A3_128k_90_50 multi-node-GLM-5.1-W8A8C8-MTP-A3_198k_function multi-node-deepseek-v3.1 Minimax_m2.7_in128k_1k_prefix90_tpot50 DeepSeek-V4-flash-w8a8-PD-prefix