[0.9.1][2/N][Feat] Restore paged attention kernel in Full Graph for performence #1677
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
changed the title
wangxiyuan
changed the title
yiz-liu
force-pushed
the
pa
branch
3 times, most recently
from
054b2db to
c40c808
Compare
Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
changed the title
ganyi1996ppo
approved these changes
Jul 11, 2025
yiz-liu
added a commit
to yiz-liu/vllm-ascend
that referenced
this pull request
Jul 31, 2025
…erformence (vllm-project#1677) Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay. Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
added a commit
to yiz-liu/vllm-ascend
that referenced
this pull request
Aug 1, 2025
…erformence (vllm-project#1677) Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay. Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
added a commit
to yiz-liu/vllm-ascend
that referenced
this pull request
Aug 11, 2025
…erformence (vllm-project#1677) Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay. Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
added a commit
to yiz-liu/vllm-ascend
that referenced
this pull request
Aug 11, 2025
…erformence (vllm-project#1677) Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay. Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
added a commit
to yiz-liu/vllm-ascend
that referenced
this pull request
Aug 12, 2025
…erformence (vllm-project#1677) Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay. Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
added a commit
to yiz-liu/vllm-ascend
that referenced
this pull request
Aug 12, 2025
…erformence (vllm-project#1677) Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay. Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
added a commit
to yiz-liu/vllm-ascend
that referenced
this pull request
Aug 13, 2025
…erformence (vllm-project#1677) Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay. Signed-off-by: Yizhou Liu <[email protected]>
yiz-liu
added a commit
to yiz-liu/vllm-ascend
that referenced
this pull request
Aug 15, 2025
…erformence (vllm-project#1677) Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay. Signed-off-by: Yizhou Liu <[email protected]>
wangxiyuan
pushed a commit
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]>
Angazenn
pushed a commit
to Angazenn/vllm-ascend
that referenced
this pull request
Oct 21, 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]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
What this PR does / why we need it?
Rectified the performance regression wherein the FIA kernel underperformed the PA kernel by enabling dynamic updates of PA parameters during graph replay.
Does this PR introduce any user-facing change?
How was this patch tested?