Support multistream of shared experts in FusedMoE #997
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
sdmyzlp
force-pushed
the
br_infer_opt
branch
8 times, most recently
from
6ed1be1 to
db76771
Compare
|
You can run |
sdmyzlp
force-pushed
the
br_infer_opt
branch
4 times, most recently
from
d6286aa to
52ba3d3
Compare
Done, with mypy error code |
|
@sdmyzlp Can you upload the profiling graph on this part? so we can have a more intuisive perspective on this PR. |
vllm_ascend/ops/fused_moe.py
Outdated
| @@ -83,11 +85,20 @@ def fused_experts_with_mc2( | |||
| } | |||
| kwargs.update(stage1_kwargs) | |||
|
|
|||
| if shared_experts is not None: | |||
There was a problem hiding this comment.
Have you tried to launch this after the dispatch op? If the dispatch will block enough time on the first stream, launch this after it seems can better overlap the execution part over host side right?
There was a problem hiding this comment.
Fixed, with all secondary-stream operations launched after the corresponding main-stream operation.
By the way, this patch only means to implement multi-stream shared experts for graph mode decode, operations will still be executed sequentially otherwise. One may extend npu_switch_stream / npu_wait_tensor in future to support eager mode multi-stream functionality.
sdmyzlp
force-pushed
the
br_infer_opt
branch
4 times, most recently
from
e2542aa to
7b28633
Compare
Described expected overlap using ascii in commit message, I have trouble uploading screenshot of the profiling. |
76 tasks
sdmyzlp
force-pushed
the
br_infer_opt
branch
2 times, most recently
from
ec0553e to
1ef0f68
Compare
sdmyzlp
force-pushed
the
br_infer_opt
branch
2 times, most recently
from
41af4d9 to
e4fe832
Compare
|
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
sdmyzlp
force-pushed
the
br_infer_opt
branch
2 times, most recently
from
dbfdfe8 to
fb0cdf2
Compare
ApsarasX
approved these changes
Jun 9, 2025
|
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
AscendW8A8DynamicLinearMethod is integrated into CustomDeepseekV2MLP in a very awkward way, causing scattered quantization operations all over the model scripts. Refactor to solve this problem. Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
The model of chosen is vllm-ascend/DeepSeek-V2-Lite-W8A8. Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
1. Concentrate the usage of `enable_multistream_moe` to one single place, and sink the computation of shared experts to `self.experts()` when multistream MoE is enabled, regardless of decode or prefill. 2. Move computation of shared experts out of `apply_mlp`. Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
Helps to unite pathes where multistream is turned on or off. Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
As the replicated version of MergedColumnParallelLinear, aiming at removing TP communication of DeepSeek-V2's `gate_up_proj` linear. Also, with replicated weight, the chunked input hidden_states can be used by shared experts. Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
With the expected overlaping being: ``` | shared gate_up | shared act | | shared down | | dispatch | routed gate_up, act, down | combine | ``` Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
ganyi1996ppo
added a commit
that referenced
this pull request
Jun 11, 2025
This reverts commit 7bdc606.
raindaywhu
pushed a commit
to raindaywhu/vllm-ascend
that referenced
this pull request
Jun 16, 2025
Contains on vllm-project#1111 for completeness. <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> Implement multi-stream parallelism for MoE layers with shared experts, where computation of shared experts will be overlapped with expert token dispatch and combine. Also, when multi-stream is enabled, weights of shared experts will be force to replicate across all cards, regardless of any tensor parallelism configurations, to avoid AllReduce operations. With the expected overlaping being: ``` | shared gate_up | shared act | | shared down | | dispatch | routed gate_up, act, down | combine | ``` <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> No. <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Tested on 1x16 910 node, with tailored 2 layer DSKv2. <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> --------- Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
momo609
pushed a commit
to momo609/vllm-ascend
that referenced
this pull request
Jun 17, 2025
Contains on vllm-project#1111 for completeness. <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? Implement multi-stream parallelism for MoE layers with shared experts, where computation of shared experts will be overlapped with expert token dispatch and combine. Also, when multi-stream is enabled, weights of shared experts will be force to replicate across all cards, regardless of any tensor parallelism configurations, to avoid AllReduce operations. With the expected overlaping being: ``` | shared gate_up | shared act | | shared down | | dispatch | routed gate_up, act, down | combine | ``` <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> ### Does this PR introduce _any_ user-facing change? No. <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> ### How was this patch tested? Tested on 1x16 910 node, with tailored 2 layer DSKv2. <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> --------- Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
momo609
pushed a commit
to momo609/vllm-ascend
that referenced
this pull request
Jun 17, 2025
Contains on vllm-project#1111 for completeness. <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? Implement multi-stream parallelism for MoE layers with shared experts, where computation of shared experts will be overlapped with expert token dispatch and combine. Also, when multi-stream is enabled, weights of shared experts will be force to replicate across all cards, regardless of any tensor parallelism configurations, to avoid AllReduce operations. With the expected overlaping being: ``` | shared gate_up | shared act | | shared down | | dispatch | routed gate_up, act, down | combine | ``` <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> ### Does this PR introduce _any_ user-facing change? No. <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> ### How was this patch tested? Tested on 1x16 910 node, with tailored 2 layer DSKv2. <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> --------- Signed-off-by: sdmyzlp <lrwei2@petalmail.com> Signed-off-by: wangxiaoxin (A) <wangxiaoxin7@huawei.com>
momo609
pushed a commit
to momo609/vllm-ascend
that referenced
this pull request
Jun 17, 2025
Contains on vllm-project#1111 for completeness. <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? Implement multi-stream parallelism for MoE layers with shared experts, where computation of shared experts will be overlapped with expert token dispatch and combine. Also, when multi-stream is enabled, weights of shared experts will be force to replicate across all cards, regardless of any tensor parallelism configurations, to avoid AllReduce operations. With the expected overlaping being: ``` | shared gate_up | shared act | | shared down | | dispatch | routed gate_up, act, down | combine | ``` <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> ### Does this PR introduce _any_ user-facing change? No. <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> ### How was this patch tested? Tested on 1x16 910 node, with tailored 2 layer DSKv2. <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> --------- Signed-off-by: sdmyzlp <lrwei2@petalmail.com> Signed-off-by: wangxiaoxin (A) <wangxiaoxin7@huawei.com>
shiyuan680
pushed a commit
to raindaywhu/vllm-ascend
that referenced
this pull request
Jul 7, 2025
Contains on vllm-project#1111 for completeness. <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> Implement multi-stream parallelism for MoE layers with shared experts, where computation of shared experts will be overlapped with expert token dispatch and combine. Also, when multi-stream is enabled, weights of shared experts will be force to replicate across all cards, regardless of any tensor parallelism configurations, to avoid AllReduce operations. With the expected overlaping being: ``` | shared gate_up | shared act | | shared down | | dispatch | routed gate_up, act, down | combine | ``` <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> No. <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Tested on 1x16 910 node, with tailored 2 layer DSKv2. <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> --------- Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Contains on #1111 for completeness.
What this PR does / why we need it?
Implement multi-stream parallelism for MoE layers with shared experts, where computation of shared experts will be overlapped with expert token dispatch and combine. Also, when multi-stream is enabled, weights of shared experts will be force to replicate across all cards, regardless of any tensor parallelism configurations, to avoid AllReduce operations.
With the expected overlaping being:
Does this PR introduce any user-facing change?
No.
How was this patch tested?
Tested on 1x16 910 node, with tailored 2 layer DSKv2.