refactor: abstract graph mode support into platform interface #25161
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
yiz-liu
requested review from
ProExpertProg,
WoosukKwon,
hmellor,
houseroad,
jikunshang,
mgoin,
robertgshaw2-redhat,
simon-mo,
tlrmchlsmth,
yewentao256 and
youkaichao
as code owners
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request effectively refactors the platform-specific graph mode support into a unified support_graph_mode interface method. The changes in vllm/config/__init__.py, vllm/platforms/cuda.py, vllm/platforms/rocm.py, and vllm/platforms/interface.py are clean and improve modularity. However, there is a logical contradiction in the implementation for the XPU platform. I've left a specific comment with a suggestion to resolve it.
zhuohan123
approved these changes
Sep 18, 2025
|
This pull request has merge conflicts that must be resolved before it can be |
yewentao256
approved these changes
Sep 19, 2025
There was a problem hiding this comment.
Great cleanup, thanks! Can we just wait for #24281 to land first (it's time sensitive) and then rebase this PR on top of that one?
Introduces a `support_graph_mode` method to the `Platform` interface to centralize the logic for determining if a backend supports graph execution. This change replaces hardcoded checks for CUDA-like or XPU platforms with a single call to the new interface method. This improves modularity and simplifies adding graph mode support for future hardware backends. Signed-off-by: Yizhou Liu <[email protected]>
Renames the platform method to more accurately reflect that it checks for static graph support, such as CUDA graphs. Updates the XPU platform to correctly report that it does not support static graphs. The runtime fallback for `cudagraph_mode` on XPU is also replaced with an assertion to enforce correct configuration. Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
force-pushed
the
support-oot-graphmode
branch
from
a9706b5 to
73270a4
Compare
github-actions
bot
added
the
ready
wangxiyuan
pushed a commit
to vllm-project/vllm-ascend
that referenced
this pull request
Sep 22, 2025
Note: This depends on [vLLM #25161](vllm-project/vllm#25161) and the torch\_npu release from September 30. ### What this PR does / why we need it? This pull request adds `FULL_DECODE_ONLY` mode for GQA/MHA models (MLA models like DeepSeek V3/R1 are not included). Key improvements include: * **Reduced dispatch latency:** By replaying the entire model execution graph at once, we cut overhead compared with multiple smaller replays. * **Stabilized multi-device performance:** Captureing the whole model as one static graph also mitigates the dispatch fluctuations across devices. * **Stream/resource savings:** Consolidating graph captures frees up streams, allowing more graphs to be captured. **Known issues:** 1. `_npu_paged_attention` currently manages its own workspace in `torch_npu`, which can deadlock when synchronizing during graph replay — we’re working on a fix. There may be other corner cases. This PR is the first in a planned series; we’ll continue to iterate and address remaining issues in follow-ups. This is essentially a port of #1503 and #1677, but includes two major changes: 1. Let `graph_dispatcher` decide the graph mode instead of hard-coding it in the backend, which decouples Full Graph and Piecewise Graph and could make it possible to remove dynamo. 2. Adapt to the new `attn_group` logic, but leave a small hack in `update_graph_params`; multi-attention models may or may not be fully supported yet. ### Does this PR introduce _any_ user-facing change? ```python compilation_config={ "cudagraph_mode": "FULL_DECODE_ONLY", }, ``` ### How was this patch tested? Tests included. - vLLM version: v0.10.2 - vLLM main: vllm-project/vllm@9607d5e --------- Signed-off-by: Yizhou Liu <[email protected]>
Mercykid-bash
pushed a commit
to Mercykid-bash/vllm-ascend
that referenced
this pull request
Sep 22, 2025
…m-project#2128) Note: This depends on [vLLM #25161](vllm-project/vllm#25161) and the torch\_npu release from September 30. ### What this PR does / why we need it? This pull request adds `FULL_DECODE_ONLY` mode for GQA/MHA models (MLA models like DeepSeek V3/R1 are not included). Key improvements include: * **Reduced dispatch latency:** By replaying the entire model execution graph at once, we cut overhead compared with multiple smaller replays. * **Stabilized multi-device performance:** Captureing the whole model as one static graph also mitigates the dispatch fluctuations across devices. * **Stream/resource savings:** Consolidating graph captures frees up streams, allowing more graphs to be captured. **Known issues:** 1. `_npu_paged_attention` currently manages its own workspace in `torch_npu`, which can deadlock when synchronizing during graph replay — we’re working on a fix. There may be other corner cases. This PR is the first in a planned series; we’ll continue to iterate and address remaining issues in follow-ups. This is essentially a port of vllm-project#1503 and vllm-project#1677, but includes two major changes: 1. Let `graph_dispatcher` decide the graph mode instead of hard-coding it in the backend, which decouples Full Graph and Piecewise Graph and could make it possible to remove dynamo. 2. Adapt to the new `attn_group` logic, but leave a small hack in `update_graph_params`; multi-attention models may or may not be fully supported yet. ### Does this PR introduce _any_ user-facing change? ```python compilation_config={ "cudagraph_mode": "FULL_DECODE_ONLY", }, ``` ### How was this patch tested? Tests included. - vLLM version: v0.10.2 - vLLM main: vllm-project/vllm@9607d5e --------- Signed-off-by: Yizhou Liu <[email protected]> Signed-off-by: Che Ruan <[email protected]>
Mercykid-bash
pushed a commit
to Mercykid-bash/vllm-ascend
that referenced
this pull request
Sep 22, 2025
…m-project#2128) Note: This depends on [vLLM #25161](vllm-project/vllm#25161) and the torch\_npu release from September 30. ### What this PR does / why we need it? This pull request adds `FULL_DECODE_ONLY` mode for GQA/MHA models (MLA models like DeepSeek V3/R1 are not included). Key improvements include: * **Reduced dispatch latency:** By replaying the entire model execution graph at once, we cut overhead compared with multiple smaller replays. * **Stabilized multi-device performance:** Captureing the whole model as one static graph also mitigates the dispatch fluctuations across devices. * **Stream/resource savings:** Consolidating graph captures frees up streams, allowing more graphs to be captured. **Known issues:** 1. `_npu_paged_attention` currently manages its own workspace in `torch_npu`, which can deadlock when synchronizing during graph replay — we’re working on a fix. There may be other corner cases. This PR is the first in a planned series; we’ll continue to iterate and address remaining issues in follow-ups. This is essentially a port of vllm-project#1503 and vllm-project#1677, but includes two major changes: 1. Let `graph_dispatcher` decide the graph mode instead of hard-coding it in the backend, which decouples Full Graph and Piecewise Graph and could make it possible to remove dynamo. 2. Adapt to the new `attn_group` logic, but leave a small hack in `update_graph_params`; multi-attention models may or may not be fully supported yet. ### Does this PR introduce _any_ user-facing change? ```python compilation_config={ "cudagraph_mode": "FULL_DECODE_ONLY", }, ``` ### How was this patch tested? Tests included. - vLLM version: v0.10.2 - vLLM main: vllm-project/vllm@9607d5e --------- Signed-off-by: Yizhou Liu <[email protected]> Signed-off-by: Che Ruan <[email protected]>
5 tasks
FeiDaLI
pushed a commit
to FeiDaLI/vllm
that referenced
this pull request
Sep 25, 2025
…roject#25161) Signed-off-by: Yizhou Liu <[email protected]>
charlifu
pushed a commit
to ROCm/vllm
that referenced
this pull request
Sep 25, 2025
…roject#25161) Signed-off-by: Yizhou Liu <[email protected]> Signed-off-by: charlifu <[email protected]>
yewentao256
pushed a commit
that referenced
this pull request
Oct 3, 2025
Signed-off-by: Yizhou Liu <[email protected]> Signed-off-by: yewentao256 <[email protected]>
xuebwang-amd
pushed a commit
to xuebwang-amd/vllm
that referenced
this pull request
Oct 10, 2025
…roject#25161) Signed-off-by: Yizhou Liu <[email protected]> Signed-off-by: xuebwang-amd <[email protected]>
choprahetarth
pushed a commit
to Tandemn-Labs/vllm
that referenced
this pull request
Oct 11, 2025
…roject#25161) Signed-off-by: Yizhou Liu <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Purpose
Introduces a
support_graph_modemethod to thePlatforminterface to centralize the logic for determining if a backend supports graph execution.This change replaces hardcoded checks for CUDA-like or XPU platforms with a single call to the new interface method. This improves modularity and simplifies adding graph mode support for future hardware backends.
Test Plan
No further tests needed.
Test Result
None
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.