[Feature]dsv4 dsa_cp support dspark by weiguihua2 · Pull Request #13249 · vllm-project/vllm-ascend

weiguihua2 · 2026-07-31T06:12:10Z

What this PR does / why we need it?

dsv4 dsa_cp supports the integration of the dspark feature.

Identify DSpark query blocks with causal=False, allowing each CP rank to use the full KV length for valid queries, no longer truncating according to causal rules.
Construct "historical SWA window + complete draft block" ori_sparse_indices for each draft token.
Expand the SWA window to window_size + draft_block_size and pass the sparse indices into the attention operator to achieve non-causal attention within the draft block.

Does this PR introduce any user-facing change?

No

How was this patch tested?

vLLM version: v0.26.0
vLLM main: vllm-project/vllm@d02df74

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

gemini-code-assist · 2026-07-31T06:12:16Z

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 integrates DSpark speculative decoding support into the DSA context-parallel framework. By modifying the attention metadata and windowing logic, the changes allow for non-causal attention within draft blocks, improving performance and accuracy for speculative decoding on Ascend hardware. The update ensures that draft tokens can access the full KV length without causal truncation, maintaining consistency across parallel ranks.

Highlights

DSpark Integration: Enabled non-causal attention support for DSpark queries within the DSA context-parallel implementation, allowing full KV length usage for draft tokens.
Metadata Updates: Extended metadata structures to include sliding window indices and custom window configurations, ensuring proper attention window management during speculative decoding.
E2E Testing: Added a new E2E test case to verify DeepSeek-V4 DSpark acceptance rates under context-parallel settings.

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.

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. ↩

github-actions · 2026-07-31T06:13:25Z

👋 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.

Tip

💡 Consider Linking a Related Issue or RFC

Your PR title contains the [Feature] tag, indicating a bug fix or new feature.

Linking a related issue or RFC in the PR description is strongly encouraged — it gives reviewers helpful context and speeds up the review. You can use any of these keywords:

Fixes #<issue_number>
Closes #<issue_number>
Resolves #<issue_number>
Refs #<rfc_or_issue_number> (for RFCs)

🙏 Thanks for helping us keep the project well-organized!

gemini-code-assist

Code Review

Suggested PR Title:

[Attention][Feature] Support DSpark speculative decoding with Context Parallel (DSA CP)

Suggested PR Summary:

### What this PR does / why we need it?
This PR adds support for DSpark speculative decoding with Context Parallel (DSA CP) in the Ascend backend. It introduces non-causal CP drafting metadata building, including computing DSpark sliding window attention (SWA) indices and sparse window parameters, and passing them to the attention forward pass.

Feedback:
In `vllm_ascend/attention/context_parallel/dsa_cp.py`, when `is_noncausal` is `True`, `local_seq_lens` is completely overwritten. Cloning it unconditionally at line 471 is redundant and introduces unnecessary memory allocation overhead. It should only be cloned when `is_noncausal` is `False`.

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

### How was this patch tested?
Added an end-to-end test `test_deepseek_v4_dsa_cp_dspark_acceptance_tp4` in `tests/e2e/pull_request/four_card/spec_decode/test_dspark_deepseekv4.py` to verify the acceptance rate of DeepSeek V4 with DSA CP and DSpark speculative decoding.

gemini-code-assist · 2026-07-31T06:15:44Z

@@ -463,21 +469,62 @@ def build_req_metadata_for_drafting(
        )
        local_query_start_loc = local_query_start_loc.clone()
        local_seq_lens = local_seq_lens.clone()


When is_noncausal is True, local_seq_lens is completely overwritten by torch.where at line 486. Therefore, cloning local_seq_lens at line 471 is redundant and leads to unnecessary memory allocation and copy overhead. We should only clone it when is_noncausal is False.

Suggested change

local_seq_lens = local_seq_lens.clone()

if not is_noncausal:

local_seq_lens = local_seq_lens.clone()

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

drslark · 2026-07-31T06:19:24Z

    cleanup_dist_env_and_memory()
+
+
+@pytest.mark.parametrize("model_name", MODELS)


I think it can be merged into test_deepseek_v4_dspark_acceptance_tp4.

I think we should guard both dsa_v1.py and dsa_cp.py. But yes, we can use same function with two different param groups.

Because these are two different and separate processes with different logic, both require monitoring.

I've thought about it, and I might be wrong.
If these two test cases are merged, it will become unclear how to skip just one of them when a problem arises.

I'll think about it over again.

pisceskkk · 2026-07-31T06:49:16Z

            input_positions = common_attn_metadata.positions[:num_input_tokens].long()
            self.common_ratio_to_sas_metadata["input_positions"] = input_positions
-            cos, sin = get_cos_and_sin_dsa(input_positions, use_cache=not has_prefill)
+            cos, sin = get_cos_and_sin_dsa(input_positions, use_cache=self.num_prefills == 0)


we can set has_prefill = self.num_prefills > 0 before this line, and use not has_prefill here.

pisceskkk · 2026-07-31T06:52:50Z

+            local_query_lens = local_query_start_loc[1:] - local_query_start_loc[:-1]
+            local_seq_lens = torch.where(


should we make local_query_lens and local_seq_lens persistent?

pisceskkk · 2026-07-31T06:53:44Z

            seq_lens=self.seq_lens_cpu[:num_reqs],
        )
+
+        if is_noncausal:


maybe we should move this branch into _build_local_token_metadata

pisceskkk · 2026-07-31T06:56:59Z

    cleanup_dist_env_and_memory()
+
+
+@pytest.mark.parametrize("model_name", MODELS)


I think we should guard both dsa_v1.py and dsa_cp.py. But yes, we can use same function with two different param groups.

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

yiz-liu

Use parametrize instead of creating a new case

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

weiguihua2 · 2026-08-01T02:27:33Z

Use parametrize instead of creating a new case

done

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

weiguihua2 · 2026-08-01T03:00:20Z

/cherry-pick releases/v0.25.1rc
[Bot]: cherry-pick branch has been refreshed. Existing PR: #13316

@weiguihua2

Cherry-pick of PR vllm-project#13249 by @weiguihua2

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

drslark

LGTM.
Thanks for your efforts.

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

weiguihua2 · 2026-08-01T07:31:46Z

/cherry-pick releases/v0.26.0rc
[Bot]: cherry-pick completed successfully. New PR created: #13322

…13320) ### What this PR does / why we need it? Cherry-pick of PR #13249 onto releases/v0.25.1rc. Original PR: #13249 dsv4 dsa_cp supports the integration of the dspark feature. Identify DSpark query blocks with causal=False, allowing each CP rank to use the full KV length for valid queries, no longer truncating according to causal rules. Construct "historical SWA window + complete draft block" ori_sparse_indices for each draft token. Expand the SWA window to window_size + draft_block_size and pass the sparse indices into the attention operator to achieve non-causal attention within the draft block. ### Does this PR introduce _any_ user-facing change? NO ### How was this patch tested? - vLLM version: v0.25.1 - vLLM main: vllm-project/vllm@fe784ff --------- Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

@weiguihua2

…(from #13249) (#13322) Cherry-pick of PR #13249 onto `releases/v0.26.0rc`. Original PR: #13249 Original author: @weiguihua2 --- ### What this PR does / why we need it? dsv4 dsa_cp supports the integration of the dspark feature. 1. Identify DSpark query blocks with causal=False, allowing each CP rank to use the full KV length for valid queries, no longer truncating according to causal rules. 2. Construct "historical SWA window + complete draft block" ori_sparse_indices for each draft token. 3. Expand the SWA window to window_size + draft_block_size and pass the sparse indices into the attention operator to achieve non-causal attention within the draft block. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? - vLLM version: v0.26.0 - vLLM main: vllm-project/vllm@d02df74 Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

### What this PR does / why we need it? dsv4 dsa_cp supports the integration of the dspark feature. 1. Identify DSpark query blocks with causal=False, allowing each CP rank to use the full KV length for valid queries, no longer truncating according to causal rules. 2. Construct "historical SWA window + complete draft block" ori_sparse_indices for each draft token. 3. Expand the SWA window to window_size + draft_block_size and pass the sparse indices into the attention operator to achieve non-causal attention within the draft block. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? - vLLM version: v0.26.0 - vLLM main: vllm-project/vllm@d02df74 --------- Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

@weiguihua2

…(from vllm-project#13249) (vllm-project#13322) Cherry-pick of PR vllm-project#13249 onto `releases/v0.26.0rc`. Original PR: vllm-project#13249 Original author: @weiguihua2 --- ### What this PR does / why we need it? dsv4 dsa_cp supports the integration of the dspark feature. 1. Identify DSpark query blocks with causal=False, allowing each CP rank to use the full KV length for valid queries, no longer truncating according to causal rules. 2. Construct "historical SWA window + complete draft block" ori_sparse_indices for each draft token. 3. Expand the SWA window to window_size + draft_block_size and pass the sparse indices into the attention operator to achieve non-causal attention within the draft block. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? - vLLM version: v0.26.0 - vLLM main: vllm-project/vllm@d02df74 Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

@weiguihua2

…(from vllm-project#13249) (vllm-project#13322) Cherry-pick of PR vllm-project#13249 onto `releases/v0.26.0rc`. Original PR: vllm-project#13249 Original author: @weiguihua2 --- ### What this PR does / why we need it? dsv4 dsa_cp supports the integration of the dspark feature. 1. Identify DSpark query blocks with causal=False, allowing each CP rank to use the full KV length for valid queries, no longer truncating according to causal rules. 2. Construct "historical SWA window + complete draft block" ori_sparse_indices for each draft token. 3. Expand the SWA window to window_size + draft_block_size and pass the sparse indices into the attention operator to achieve non-causal attention within the draft block. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? - vLLM version: v0.26.0 - vLLM main: vllm-project/vllm@d02df74 Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

### What this PR does / why we need it? This PR enables MTP=3 speculative decoding in the DSA‑CP path for both eager and graph modes. It introduces full ACL graph support for the MTP draft steps, including pre‑allocated metadata buffers that allow deterministic graph capture and replay when enable_dsa_cp is true. ### Changes summary #### MTP > 1 support in DSA‑CP ~Previously build_for_drafting only handled MTP=1 and lacked the CPU‑side sequence length path, which caused tensor dimension errors for any MTP > 1.~ Supported by #13249. This PR updates build_for_drafting to populate CPU sequence lengths, so that all downstream metadata builders can correctly process multiple draft steps. #### Full ACL graph support for MTP draft steps ACL graph capture and replay require stable tensor addresses across invocations. To guarantee this, the PR pre‑allocates per‑step buffers for all draft‑step metadata during builder initialization, and pads seq_lens_cpu to match the graph‑dispatched batch size, ensuring deterministic addresses throughout replay. In addition, a static `update_graph_params` method was added, so that the graph dispatch can correctly invoke the DSA-CP attention backend during graph replay. ~#### Depends on PR #12193.~ ~#12193 also fixes metadata mismatch issues in the DSA-CP builder, and this PR is directly based on the refactored code. Cherry-picking the commits onto main without #12193 causes runtime errors under concurrent requests.~ ~Only the top commits (after Commits on Jul 21, 2026) belong to this PR. Please review by focusing on the top commits. Once it is merged, I'll rebase onto main quickly.`~ ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? - Added an E2E accuracy test in `tests/e2e/pull_request/four_card/test_deepseek_v4.py` with exact `expected_token_ids` assertion to guard against regressions in the DSA-CP + MTP=3 + full graph path. - GMS8K accuracy evaluation for MTP+eager and MTP+graph is attached below. ``` mtp3+eager +---------+-----------+----------+----------+-------+---------+---------+ | Model | Dataset | Metric | Subset | Num | Score | Cat.0 | +=========+===========+==========+==========+=======+=========+=========+ | dsv4-f | gsm8k | mean_acc | main | 1319 | 0.9666 | default | +---------+-----------+----------+----------+-------+---------+---------+ mtp3+full graph +---------+-----------+----------+----------+-------+---------+---------+ | Model | Dataset | Metric | Subset | Num | Score | Cat.0 | +=========+===========+==========+==========+=======+=========+=========+ | dsv4-f | gsm8k | mean_acc | main | 1319 | 0.9659 | default | +---------+-----------+----------+----------+-------+---------+---------+ ``` - vLLM main: vllm-project/vllm@e6bfe03 --------- Signed-off-by: frankie <wangyongsheng686@gmail.com> Co-authored-by: frankie <wangyongsheng686@gmail.com>

### What this PR does / why we need it? This PR enables MTP=3 speculative decoding in the DSA‑CP path for both eager and graph modes. It introduces full ACL graph support for the MTP draft steps, including pre‑allocated metadata buffers that allow deterministic graph capture and replay when enable_dsa_cp is true. ### Changes summary #### MTP > 1 support in DSA‑CP ~Previously build_for_drafting only handled MTP=1 and lacked the CPU‑side sequence length path, which caused tensor dimension errors for any MTP > 1.~ Supported by vllm-project#13249. This PR updates build_for_drafting to populate CPU sequence lengths, so that all downstream metadata builders can correctly process multiple draft steps. #### Full ACL graph support for MTP draft steps ACL graph capture and replay require stable tensor addresses across invocations. To guarantee this, the PR pre‑allocates per‑step buffers for all draft‑step metadata during builder initialization, and pads seq_lens_cpu to match the graph‑dispatched batch size, ensuring deterministic addresses throughout replay. In addition, a static `update_graph_params` method was added, so that the graph dispatch can correctly invoke the DSA-CP attention backend during graph replay. ~#### Depends on PR vllm-project#12193.~ ~vllm-project#12193 also fixes metadata mismatch issues in the DSA-CP builder, and this PR is directly based on the refactored code. Cherry-picking the commits onto main without vllm-project#12193 causes runtime errors under concurrent requests.~ ~Only the top commits (after Commits on Jul 21, 2026) belong to this PR. Please review by focusing on the top commits. Once it is merged, I'll rebase onto main quickly.`~ ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? - Added an E2E accuracy test in `tests/e2e/pull_request/four_card/test_deepseek_v4.py` with exact `expected_token_ids` assertion to guard against regressions in the DSA-CP + MTP=3 + full graph path. - GMS8K accuracy evaluation for MTP+eager and MTP+graph is attached below. ``` mtp3+eager +---------+-----------+----------+----------+-------+---------+---------+ | Model | Dataset | Metric | Subset | Num | Score | Cat.0 | +=========+===========+==========+==========+=======+=========+=========+ | dsv4-f | gsm8k | mean_acc | main | 1319 | 0.9666 | default | +---------+-----------+----------+----------+-------+---------+---------+ mtp3+full graph +---------+-----------+----------+----------+-------+---------+---------+ | Model | Dataset | Metric | Subset | Num | Score | Cat.0 | +=========+===========+==========+==========+=======+=========+=========+ | dsv4-f | gsm8k | mean_acc | main | 1319 | 0.9659 | default | +---------+-----------+----------+----------+-------+---------+---------+ ``` - vLLM main: vllm-project/vllm@e6bfe03 --------- Signed-off-by: frankie <wangyongsheng686@gmail.com> Co-authored-by: frankie <wangyongsheng686@gmail.com>

@weiguihua2

…(from vllm-project#13249) (vllm-project#13322) Cherry-pick of PR vllm-project#13249 onto `releases/v0.26.0rc`. Original PR: vllm-project#13249 Original author: @weiguihua2 --- ### What this PR does / why we need it? dsv4 dsa_cp supports the integration of the dspark feature. 1. Identify DSpark query blocks with causal=False, allowing each CP rank to use the full KV length for valid queries, no longer truncating according to causal rules. 2. Construct "historical SWA window + complete draft block" ori_sparse_indices for each draft token. 3. Expand the SWA window to window_size + draft_block_size and pass the sparse indices into the attention operator to achieve non-causal attention within the draft block. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? - vLLM version: v0.26.0 - vLLM main: vllm-project/vllm@d02df74 Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

### What this PR does / why we need it? This PR enables MTP=3 speculative decoding in the DSA‑CP path for both eager and graph modes. It introduces full ACL graph support for the MTP draft steps, including pre‑allocated metadata buffers that allow deterministic graph capture and replay when enable_dsa_cp is true. ### Changes summary #### MTP > 1 support in DSA‑CP ~Previously build_for_drafting only handled MTP=1 and lacked the CPU‑side sequence length path, which caused tensor dimension errors for any MTP > 1.~ Supported by vllm-project#13249. This PR updates build_for_drafting to populate CPU sequence lengths, so that all downstream metadata builders can correctly process multiple draft steps. #### Full ACL graph support for MTP draft steps ACL graph capture and replay require stable tensor addresses across invocations. To guarantee this, the PR pre‑allocates per‑step buffers for all draft‑step metadata during builder initialization, and pads seq_lens_cpu to match the graph‑dispatched batch size, ensuring deterministic addresses throughout replay. In addition, a static `update_graph_params` method was added, so that the graph dispatch can correctly invoke the DSA-CP attention backend during graph replay. ~#### Depends on PR vllm-project#12193.~ ~vllm-project#12193 also fixes metadata mismatch issues in the DSA-CP builder, and this PR is directly based on the refactored code. Cherry-picking the commits onto main without vllm-project#12193 causes runtime errors under concurrent requests.~ ~Only the top commits (after Commits on Jul 21, 2026) belong to this PR. Please review by focusing on the top commits. Once it is merged, I'll rebase onto main quickly.`~ ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? - Added an E2E accuracy test in `tests/e2e/pull_request/four_card/test_deepseek_v4.py` with exact `expected_token_ids` assertion to guard against regressions in the DSA-CP + MTP=3 + full graph path. - GMS8K accuracy evaluation for MTP+eager and MTP+graph is attached below. ``` mtp3+eager +---------+-----------+----------+----------+-------+---------+---------+ | Model | Dataset | Metric | Subset | Num | Score | Cat.0 | +=========+===========+==========+==========+=======+=========+=========+ | dsv4-f | gsm8k | mean_acc | main | 1319 | 0.9666 | default | +---------+-----------+----------+----------+-------+---------+---------+ mtp3+full graph +---------+-----------+----------+----------+-------+---------+---------+ | Model | Dataset | Metric | Subset | Num | Score | Cat.0 | +=========+===========+==========+==========+=======+=========+=========+ | dsv4-f | gsm8k | mean_acc | main | 1319 | 0.9659 | default | +---------+-----------+----------+----------+-------+---------+---------+ ``` - vLLM main: vllm-project/vllm@e6bfe03 --------- Signed-off-by: frankie <wangyongsheng686@gmail.com> Co-authored-by: frankie <wangyongsheng686@gmail.com>

### What this PR does / why we need it? This PR enables MTP=3 speculative decoding in the DSA‑CP path for both eager and graph modes. It introduces full ACL graph support for the MTP draft steps, including pre‑allocated metadata buffers that allow deterministic graph capture and replay when enable_dsa_cp is true. ### Changes summary #### MTP > 1 support in DSA‑CP ~Previously build_for_drafting only handled MTP=1 and lacked the CPU‑side sequence length path, which caused tensor dimension errors for any MTP > 1.~ Supported by vllm-project#13249. This PR updates build_for_drafting to populate CPU sequence lengths, so that all downstream metadata builders can correctly process multiple draft steps. #### Full ACL graph support for MTP draft steps ACL graph capture and replay require stable tensor addresses across invocations. To guarantee this, the PR pre‑allocates per‑step buffers for all draft‑step metadata during builder initialization, and pads seq_lens_cpu to match the graph‑dispatched batch size, ensuring deterministic addresses throughout replay. In addition, a static `update_graph_params` method was added, so that the graph dispatch can correctly invoke the DSA-CP attention backend during graph replay. ~#### Depends on PR vllm-project#12193.~ ~vllm-project#12193 also fixes metadata mismatch issues in the DSA-CP builder, and this PR is directly based on the refactored code. Cherry-picking the commits onto main without vllm-project#12193 causes runtime errors under concurrent requests.~ ~Only the top commits (after Commits on Jul 21, 2026) belong to this PR. Please review by focusing on the top commits. Once it is merged, I'll rebase onto main quickly.`~ ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? - Added an E2E accuracy test in `tests/e2e/pull_request/four_card/test_deepseek_v4.py` with exact `expected_token_ids` assertion to guard against regressions in the DSA-CP + MTP=3 + full graph path. - GMS8K accuracy evaluation for MTP+eager and MTP+graph is attached below. ``` mtp3+eager +---------+-----------+----------+----------+-------+---------+---------+ | Model | Dataset | Metric | Subset | Num | Score | Cat.0 | +=========+===========+==========+==========+=======+=========+=========+ | dsv4-f | gsm8k | mean_acc | main | 1319 | 0.9666 | default | +---------+-----------+----------+----------+-------+---------+---------+ mtp3+full graph +---------+-----------+----------+----------+-------+---------+---------+ | Model | Dataset | Metric | Subset | Num | Score | Cat.0 | +=========+===========+==========+==========+=======+=========+=========+ | dsv4-f | gsm8k | mean_acc | main | 1319 | 0.9659 | default | +---------+-----------+----------+----------+-------+---------+---------+ ``` - vLLM main: vllm-project/vllm@e6bfe03 --------- Signed-off-by: frankie <wangyongsheng686@gmail.com> Co-authored-by: frankie <wangyongsheng686@gmail.com> Signed-off-by: like-0517 <ithwlike@126.com>

[Feature]dsv4 dsa_cp support dspark

84d2856

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

weiguihua2 requested review from wangxiyuan, weijinqian0 and whx-sjtu as code owners July 31, 2026 06:12

github-actions Bot added the module:tests label Jul 31, 2026

gemini-code-assist Bot reviewed Jul 31, 2026

View reviewed changes

[Feature]dsv4 dsa_cp support dspark

29342c9

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

drslark suggested changes Jul 31, 2026

View reviewed changes

pisceskkk reviewed Jul 31, 2026

View reviewed changes

weiguihua2 added 3 commits July 31, 2026 16:56

[Feature]dsv4 dsa_cp support dspark

1158ffd

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

[Feature]dsv4 dsa_cp support dspark

d543db8

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

[Feature]dsv4 dsa_cp support dspark

7f61e51

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

yiz-liu reviewed Jul 31, 2026

View reviewed changes

yiz-liu added the ready label Jul 31, 2026

weiguihua2 added 2 commits July 31, 2026 18:09

[Feature]dsv4 dsa_cp support dspark

3d05815

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

[Feature]dsv4 dsa_cp support dspark

38a8468

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

[Feature]dsv4 dsa_cp support dspark

ee6bfb4

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

vllm-ascend-ci mentioned this pull request Aug 1, 2026

[Cherry-pick][releases/v0.25.1rc][Feature]dsv4 dsa_cp support dspark (from #13249) #13316

Closed

vllm-ascend-ci pushed a commit to vllm-ascend-ci/vllm-ascend that referenced this pull request Aug 1, 2026

[Feature]dsv4 dsa_cp support dspark

f8b6658

Cherry-pick of PR vllm-project#13249 by @weiguihua2

pisceskkk approved these changes Aug 1, 2026

View reviewed changes

[Feature]dsv4 dsa_cp support dspark

1b21fa6

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

drslark approved these changes Aug 1, 2026

View reviewed changes

weiguihua2 mentioned this pull request Aug 1, 2026

[Cherry-pick][releases/v0.25.1rc][Feature]dsv4 dsa_cp support dspark #13320

Merged

weiguihua2 added 2 commits August 1, 2026 13:16

[Feature]dsv4 dsa_cp support dspark

1e2c193

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

[Feature]dsv4 dsa_cp support dspark

02d11f5

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

vllm-ascend-ci mentioned this pull request Aug 1, 2026

[Cherry-pick][releases/v0.26.0rc][Feature]dsv4 dsa_cp support dspark (from #13249) #13322

Merged

zzzzwwjj approved these changes Aug 1, 2026

View reviewed changes

zzzzwwjj merged commit fa3b1bd into vllm-project:main Aug 1, 2026
46 checks passed

Csrayz mentioned this pull request Aug 3, 2026

[Feature] Support dsv4 mtp graph for dsa_cp #12599

Merged

linfeng-yuan mentioned this pull request Sep 24, 2026

[Release]: Release checklist for v0.27.1rc1 #16288

Open

38 of 49 tasks

	local_seq_lens = local_seq_lens.clone()
	if not is_noncausal:
	local_seq_lens = local_seq_lens.clone()

		cleanup_dist_env_and_memory()


		@pytest.mark.parametrize("model_name", MODELS)

		local_query_lens = local_query_start_loc[1:] - local_query_start_loc[:-1]
		local_seq_lens = torch.where(

Conversation

weiguihua2 commented Jul 31, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this PR does / why we need it?

Does this PR introduce any user-facing change?

How was this patch tested?

Uh oh!

gemini-code-assist Bot commented Jul 31, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

github-actions Bot commented Jul 31, 2026

💡 Consider Linking a Related Issue or RFC

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Jul 31, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

yiz-liu left a comment

Choose a reason for hiding this comment

Uh oh!

weiguihua2 commented Aug 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

weiguihua2 commented Aug 1, 2026 • edited by vllm-ascend-ci Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

drslark left a comment

Choose a reason for hiding this comment

Uh oh!

weiguihua2 commented Aug 1, 2026 • edited by vllm-ascend-ci Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

weiguihua2 commented Jul 31, 2026 •

edited by github-actions Bot

Loading

weiguihua2 commented Aug 1, 2026 •

edited

Loading

weiguihua2 commented Aug 1, 2026 •

edited by vllm-ascend-ci

Loading

weiguihua2 commented Aug 1, 2026 •

edited by vllm-ascend-ci

Loading