[EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB #4216

845473182 · 2025-11-17T02:08:06Z

What this PR does / why we need it?

Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list into dynamic EPLB to support list-type parameters
This PR also modify the logic of loading model in dynamic-eplb scenario.
The operator is based on this pr: #3804

Does this PR introduce any user-facing change?

no

How was this patch tested?

vllm serve /home/weight/DeepSeek-V3.1_w8a8mix_mtp \
    --max_num_seqs 8 \
    --max-model-len 8192 \
    --max-num-batched-tokens 16384 \
    --tensor-parallel-size 8 \
    --data-parallel-size 2 \
    --enable-expert-parallel \
    --served-model-name ds_r1 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes \
    --no-enable-prefix-caching \
    --port 8999 \
    --quantization "ascend" \
    --gpu-memory-utilization 0.85 \
    --trust-remote-code \
    --compilation_config '{"cudagraph_capture_sizes":[1,2,4,8,16,32]}' \
    --additional-config='{"dynamic_eplb":true, "num_iterations_eplb_update":100, "num_wait_worker_iterations":100}'

input&output: 2k 2k
This PR:

Baseline:

vLLM version: v0.11.2
vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2

github-actions · 2025-11-17T02:08:14Z

👋 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.

github-actions · 2025-11-26T06:45:31Z

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

Signed-off-by: 白永斌 <[email protected]>

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

Signed-off-by: 欧派果奶我还要 <[email protected]>

wangxiyuan · 2025-11-28T09:21:57Z

vllm_ascend/ops/fused_moe/moe_comm_method.py

-            w1: torch.Tensor,
-            w2: torch.Tensor,
+            w1: torch.Tensor | list[torch.Tensor],
+            w2: torch.Tensor | list[torch.Tensor],


w1: list[torch.Tensor]

wangxiyuan · 2025-11-28T09:22:18Z

vllm_ascend/quantization/w8a8_dynamic.py

+            return moe_comm_method.fused_experts(
+                hidden_states=x,
+                pertoken_scale=pertoken_scale,
+                w1=layer.w13_weight,


[layer.w13_weight]

wangxiyuan · 2025-11-28T09:23:17Z

vllm_ascend/quantization/w8a8_dynamic.py

-            dynamic_scale_for_share=dynamic_scale_for_share,
-            dynamic_eplb=self.dynamic_eplb,
-            mc2_mask=kwargs.get("mc2_mask", None))
+        if self.dynamic_eplb:


if self.dynamic_eplb:
w1 = xx
else:
w1 = yyy

…ist operator into dynamic EPLB (vllm-project#4216) ### What this PR does / why we need it? Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list into dynamic EPLB to support list-type parameters This PR also modify the logic of loading model in dynamic-eplb scenario. The operator is based on this pr: vllm-project#3804 ### Does this PR introduce _any_ user-facing change? no ### How was this patch tested? ``` vllm serve /home/weight/DeepSeek-V3.1_w8a8mix_mtp \ --max_num_seqs 8 \ --max-model-len 8192 \ --max-num-batched-tokens 16384 \ --tensor-parallel-size 8 \ --data-parallel-size 2 \ --enable-expert-parallel \ --served-model-name ds_r1 \ --enable-auto-tool-choice \ --tool-call-parser hermes \ --no-enable-prefix-caching \ --port 8999 \ --quantization "ascend" \ --gpu-memory-utilization 0.85 \ --trust-remote-code \ --compilation_config '{"cudagraph_capture_sizes":[1,2,4,8,16,32]}' \ --additional-config='{"dynamic_eplb":true, "num_iterations_eplb_update":100, "num_wait_worker_iterations":100}' ``` input&output: 2k 2k This PR: <img width="1318" height="695" alt="fusion" src="https://github.com/user-attachments/assets/f8657813-0c02-42f4-8396-d99e730f48cd" /> Baseline: <img width="1323" height="690" alt="baseline" src="https://github.com/user-attachments/assets/e1323a78-af26-4523-820c-e20e5642a38e" /> - vLLM version: v0.11.2 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2 --------- Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]> Co-authored-by: 白永斌 <[email protected]>

…ist operator into dynamic EPLB (vllm-project#4216) Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list into dynamic EPLB to support list-type parameters This PR also modify the logic of loading model in dynamic-eplb scenario. The operator is based on this pr: vllm-project#3804 no ``` vllm serve /home/weight/DeepSeek-V3.1_w8a8mix_mtp \ --max_num_seqs 8 \ --max-model-len 8192 \ --max-num-batched-tokens 16384 \ --tensor-parallel-size 8 \ --data-parallel-size 2 \ --enable-expert-parallel \ --served-model-name ds_r1 \ --enable-auto-tool-choice \ --tool-call-parser hermes \ --no-enable-prefix-caching \ --port 8999 \ --quantization "ascend" \ --gpu-memory-utilization 0.85 \ --trust-remote-code \ --compilation_config '{"cudagraph_capture_sizes":[1,2,4,8,16,32]}' \ --additional-config='{"dynamic_eplb":true, "num_iterations_eplb_update":100, "num_wait_worker_iterations":100}' ``` input&output: 2k 2k This PR: <img width="1318" height="695" alt="fusion" src="https://github.com/user-attachments/assets/f8657813-0c02-42f4-8396-d99e730f48cd" /> Baseline: <img width="1323" height="690" alt="baseline" src="https://github.com/user-attachments/assets/e1323a78-af26-4523-820c-e20e5642a38e" /> - vLLM version: v0.11.2 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2 --------- Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]> Co-authored-by: 白永斌 <[email protected]>

…ist operator into dynamic EPLB (vllm-project#4216) Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list into dynamic EPLB to support list-type parameters This PR also modify the logic of loading model in dynamic-eplb scenario. The operator is based on this pr: vllm-project#3804 no ``` vllm serve /home/weight/DeepSeek-V3.1_w8a8mix_mtp \ --max_num_seqs 8 \ --max-model-len 8192 \ --max-num-batched-tokens 16384 \ --tensor-parallel-size 8 \ --data-parallel-size 2 \ --enable-expert-parallel \ --served-model-name ds_r1 \ --enable-auto-tool-choice \ --tool-call-parser hermes \ --no-enable-prefix-caching \ --port 8999 \ --quantization "ascend" \ --gpu-memory-utilization 0.85 \ --trust-remote-code \ --compilation_config '{"cudagraph_capture_sizes":[1,2,4,8,16,32]}' \ --additional-config='{"dynamic_eplb":true, "num_iterations_eplb_update":100, "num_wait_worker_iterations":100}' ``` input&output: 2k 2k This PR: <img width="1318" height="695" alt="fusion" src="https://github.com/user-attachments/assets/f8657813-0c02-42f4-8396-d99e730f48cd" /> Baseline: <img width="1323" height="690" alt="baseline" src="https://github.com/user-attachments/assets/e1323a78-af26-4523-820c-e20e5642a38e" /> - vLLM version: v0.11.2 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2 --------- Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]> Co-authored-by: 白永斌 <[email protected]> Signed-off-by: Che Ruan <[email protected]>

…ist operator into dynamic EPLB (vllm-project#4216) ### What this PR does / why we need it? Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list into dynamic EPLB to support list-type parameters This PR also modify the logic of loading model in dynamic-eplb scenario. The operator is based on this pr: vllm-project#3804 ### Does this PR introduce _any_ user-facing change? no ### How was this patch tested? ``` vllm serve /home/weight/DeepSeek-V3.1_w8a8mix_mtp \ --max_num_seqs 8 \ --max-model-len 8192 \ --max-num-batched-tokens 16384 \ --tensor-parallel-size 8 \ --data-parallel-size 2 \ --enable-expert-parallel \ --served-model-name ds_r1 \ --enable-auto-tool-choice \ --tool-call-parser hermes \ --no-enable-prefix-caching \ --port 8999 \ --quantization "ascend" \ --gpu-memory-utilization 0.85 \ --trust-remote-code \ --compilation_config '{"cudagraph_capture_sizes":[1,2,4,8,16,32]}' \ --additional-config='{"dynamic_eplb":true, "num_iterations_eplb_update":100, "num_wait_worker_iterations":100}' ``` input&output: 2k 2k This PR: <img width="1318" height="695" alt="fusion" src="https://github.com/user-attachments/assets/f8657813-0c02-42f4-8396-d99e730f48cd" /> Baseline: <img width="1323" height="690" alt="baseline" src="https://github.com/user-attachments/assets/e1323a78-af26-4523-820c-e20e5642a38e" /> - vLLM version: v0.11.2 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2 --------- Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]> Co-authored-by: 白永斌 <[email protected]>

github-actions bot added module:ops module:quantization labels Nov 17, 2025

github-actions bot added the merge-conflicts label Nov 26, 2025

github-actions bot removed the merge-conflicts label Nov 27, 2025

845473182 changed the title ~~Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list into dynamic EPLB~~ [EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB Nov 28, 2025

845473182 marked this pull request as ready for review November 28, 2025 07:48

github-actions bot added the module:tests label Nov 29, 2025

weijinqian0 added ready read for review ready-for-test start test by label for PR labels Nov 29, 2025

白永斌 added 19 commits November 30, 2025 00:13

fix init adaptor bug and moe_mlp dow_proj type error

1de03a9

Signed-off-by: 白永斌 <[email protected]>

fix pre-commit

a7799c5

Signed-off-by: 白永斌 <[email protected]>

fix pre-commit

53d930d

Signed-off-by: 白永斌 <[email protected]>

fix parameter naming error

0c47052

Signed-off-by: 白永斌 <[email protected]>

fix pre-commit

4897ab4

Signed-off-by: 白永斌 <[email protected]>

unify the data types of weight and weight scale in moe_mlp

6d94c2e

Signed-off-by: 白永斌 <[email protected]>

standarlized using enable_custom_op()

aaa0bd9

Signed-off-by: 白永斌 <[email protected]>

fix pre-commit

40126a6

Signed-off-by: 白永斌 <[email protected]>

fix pre-commit

1dc4e22

Signed-off-by: 白永斌 <[email protected]>

fix ut

ac6dfc9

Signed-off-by: 白永斌 <[email protected]>

fix ut

cc2c9b8

Signed-off-by: 白永斌 <[email protected]>

fix e2e-light

5811819

Signed-off-by: 白永斌 <[email protected]>

fix unquantized ut

d1cb0e3

Signed-off-by: 白永斌 <[email protected]>

fix init adaptor bug and moe_mlp dow_proj type error

96de387

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix pre-commit

c3aa478

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix pre-commit

5990199

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix parameter naming error

225e4f2

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix pre-commit

f35fa16

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

unify the data types of weight and weight scale in moe_mlp

48eb2ba

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

白永斌 added 7 commits November 30, 2025 14:02

standarlized using enable_custom_op()

de43903

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix pre-commit

55a1eb5

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix pre-commit

18b13f5

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix ut

765a3c3

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix ut

016ff63

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix e2e-light

2627e98

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

fix unquantized ut

1777304

Signed-off-by: 白永斌 <[email protected]> Signed-off-by: 欧派果奶我还要 <[email protected]>

845473182 force-pushed the gmm_swiglu_quant_tensor_list branch from fef26ce to 1777304 Compare November 30, 2025 06:05

fix precommit

b929db2

Signed-off-by: 欧派果奶我还要 <[email protected]>

wangxiyuan approved these changes Nov 30, 2025

View reviewed changes

wangxiyuan merged commit bc67696 into vllm-project:main Nov 30, 2025
22 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB #4216

[EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB #4216

Uh oh!

845473182 commented Nov 17, 2025 •

edited by github-actions bot

Loading

Uh oh!

github-actions bot commented Nov 17, 2025

Uh oh!

github-actions bot commented Nov 26, 2025

Uh oh!

wangxiyuan Nov 28, 2025

Uh oh!

wangxiyuan Nov 28, 2025

Uh oh!

wangxiyuan Nov 28, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

[EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB #4216

[EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB #4216

Uh oh!

Conversation

845473182 commented Nov 17, 2025 • 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!

github-actions bot commented Nov 17, 2025

Uh oh!

github-actions bot commented Nov 26, 2025

Uh oh!

wangxiyuan Nov 28, 2025

Choose a reason for hiding this comment

Uh oh!

wangxiyuan Nov 28, 2025

Choose a reason for hiding this comment

Uh oh!

wangxiyuan Nov 28, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

845473182 commented Nov 17, 2025 •

edited by github-actions bot

Loading