### What this PR does / why we need it?
Add cudagraph_capture_sizes for E2E CI test.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: menogrey <1299267905@qq.com>
1. refresh additional config doc
2. move kv config logic to platform.
3. improve `dump_config` init logic and rename it to `dump_config_path`
this change is user impacted. dump_config is changed from dict to
string.
4. correct `enable_async_exponential` type
5. remove useless `chunked_prefill_for_mla`
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
- Fix vllm break in the pr:
1.[Drop v0.14 deprecations
]https://github.com/vllm-project/vllm/pull/31285
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08
---------
Signed-off-by: ZT-AIA <1028681969@qq.com>
### What this PR does / why we need it?
Currently, our multi-node logs only show the master node's logs (via the
Kubernetes API), which is insufficient for effective problem
localization if other nodes experience issues. Therefore, this pull
request adds the ability to upload logs for other nodes.
Next plan: Output structured directory logs, including logs from each
node and the polog.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
The contiguous() operation temporarily increases memory usage, leading
to higher peak GPU memory, which necessitates reducing
gpu_memory_utilization. However, making tensors contiguous in
modelrunnerv1 significantly enhances operator performance, resulting in
greater end-to-end model benefits despite the memory overhead.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Currently, the vllm pr: https://github.com/vllm-project/vllm/pull/24252
is causing operator fusion to fail, which can be mitigated by patching
the backend. Once the problem is completely resolved, I will submit a
new pull request to remove the patch.
- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef
---------
Signed-off-by: wxsIcey <1790571317@qq.com>
Description:
This PR updates the implementation of the Triton operator for deployment
on NPU devices, focusing on optimizing grid size and memory handling
based on NPU limitations.
Design Plan:
Grid Calculation: The grid size is now dynamically calculated by batch
and dim to ensure that the number of programs executed does not exceed
the NPU's vector core capacity. This ensures optimal parallelism without
overloading the hardware.
Data Block Handling: Due to the limited on-chip memory (UB) on Ascend
NPUs, this implementation splits large data into smaller chunks of 32k
or less per block. The kernel performs a for-loop to process the data in
these smaller chunks, minimizing memory usage and avoiding potential
overflows.
Changes Compared to GPU Implementation:
Grid and Block Sizing:
For GPU, the grid and block size were determined based on available
thread counts and memory size. In contrast, the NPU version dynamically
adjusts these parameters using B_TILE and BLOCK_N to optimize for NPU’s
architecture.
Memory Chunking:
The original GPU implementation did not require chunking due to the
higher available memory and processing capacity. For the NPU, data is
divided into smaller chunks (32k or smaller) to comply with memory
constraints on the device. The kernel has been modified to handle this
chunking mechanism inside a loop.
Optimized Thread Usage:
The NPU implementation takes into account the hardware-specific thread
limit (24 threads per vector core), ensuring that the number of active
programs is aligned with the NPU's vector core count, avoiding
over-subscription that would lead to serial processing.
This PR ensures that the operator functions efficiently on Ascend NPU,
considering hardware limitations while maintaining the same
functionality and input parameters as the GPU implementation.
- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef
Signed-off-by: maoxx241 <maomaoyu870@gmail.com>
### What this PR does / why we need it?
Update the configuration for optimal performance of deepseek v3.2 in the usage tutorial.
- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08
---------
Signed-off-by: cookieyyds <126683903+cookieyyds@users.noreply.github.com>
Signed-off-by: Mengqing Cao <cmq0113@163.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
fix xlite decode-only e2e test, xlite decode-only mode utilizes
aclgraph's prefill and will be affected by aclgraph, so shortened test
length.
- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08
Signed-off-by: changdawei1 <changdawei3@huawei.com>
Co-authored-by: changdawei1 <changdawei3@huawei.com>
### What this PR does / why we need it?
[Bugfix] Fixing the issue where 128K context does not work in long
sequence scenarios.
This issue is caused by not splitting num_token according to pcp_size
during profile_run.
During `profile_run`, a warm-up is performed based on
`self.max_num_tokens`. When PCP is enabled, each PCP group will only
schedule up to `self.max_num_tokens / pcp_size`. After `profile_run` is
completed, the original scheduling size needs to be restored.
This is a temporary workaround; once
https://github.com/vllm-project/vllm/pull/28988/files is implemented,
this part can be removed.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
### What this PR does / why we need it?
Fixed the error in the CI process for
vllm-ascend/tests/e2e/nightly/ops/triton/test_rejection_sampler.py
Error: test_rejection_sampler_block_verify_triton_kernel: duplicate
parametrization of 'vocab_size'.
- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08
Signed-off-by: chenaoxuan <cax1165@163.com>
`VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE` is only used together with
`VLLM_ASCEND_ENABLE_PREFETCH_MLP` which is useless totally. This PR
remove it.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
When matmul_and_reduce is enabled, the prefix attribute is required.
However, in some models, the prefix is not passed correctly, causing
errors when starting the service.
The issue of incorrect prefix passing will be fixed in vLLM in the
future.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: Wang Kunpeng <1289706727@qq.com>
### What this PR does / why we need it?
The variable `self.num_pcp_pads` was incorrectly truncated during
assignment, causing errors in certain scenarios such as PD
disaggregated. This issue has now been resolved.
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
Co-author by: QiuChunshuo <qiuchunshuo@huawei.com>
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: daishixun <dsxsteven@sina.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
add xlite e2e test
- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef
Signed-off-by: DaweiChang <405739598@qq.com>
### What this PR does / why we need it?
1. MagicMTP (paper: "Block Verification Accelerates Speculative
Decoding") was introduced to consider the influence among multiple draft
tokens, improving the acceptance rate without compromising accuracy.
2. The rejection sampling logic in rejection_sampler.py was restructured
using Triton-Ascend, enabling it to operate under high concurrency, thus
resolving CPU and NPU operator bottlenecks and enhancing throughput.
### Does this PR introduce _any_ user-facing change?
MagicMTP will automatically take effect when the parameter
"num_speculative_tokens" >= 3.
- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2
Signed-off-by: chenaoxuan <cax1165@163.com>
### What this PR does / why we need it?
This pull request introduces an L2 normalization kernel implemented in
Triton, specifically optimized for Ascend NPUs.
### Does this PR introduce _any_ user-facing change?
No, this PR does not introduce any user-facing changes.
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
bc0a5a0c08
---------
Signed-off-by: Ascendyh <hw7osiris@outlook.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
Revert [KV-Sharing] Support KV-Sharing feature in CLA models (#4138) as
it causes deepseek v3.2 hang error
- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
`VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP` is not used anywhere, let's
remove it.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
1. Use optimized apply_top_k_top_p for NPU platfrom in rejection
sampler; (avoid scatter elements which can reduce ~26ms TPOT with bs=24
per DP)
2. <del>Avoid D2H Synchronization before calling npu_top_k_top_p
introduced by parameter validation which improves inference speed with
`async_scheduling` enabled;</del> In order to elminate the D2H
synchronization introduced by parameter validation before calling
`npu_top_k_top_p`, we directly drop this fused operator since the
performance improvement is not significant compared to async_scheduling
and may bring potential accuracy problem.
3. Refactor the implementation of AscendTopKTopPSampler to align that of
vLLM.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E serving test with combinations of `k=500` and `p=0.95` with
async_scheduling in single node and wide-EP scenarios.
- vLLM version: v0.11.0
- vLLM main:
83f478bb19
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Co-authored-by: realliujiaxu <realliujiaxu@163.com>
### What this PR does / why we need it?
Skip some failed ops tests
- vLLM version: release/v0.13.0
- vLLM main:
5fbfa8d9ef
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
Add pa_shape_list description to qwen dense tutorial.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: ZYang6263 <zy626375@gmail.com>
Co-authored-by: zzzzwwjj <34335947+zzzzwwjj@users.noreply.github.com>
### What this PR does / why we need it?
- This PR removes the Expert Parallel (EP) HCCL buffer allocation that
was previously introduced by the fused-op `dispatch_ffn_combine` (#3532
), since the fused-op has switch to MC2 HCCL buffer (#5156 ).
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: Chen Chen <0109chenchen@gmail.com>
### What this PR does / why we need it?
[E2E] Optimize e2e test.
- Remove the test_basic_camem testcase.
- Change Qwen2.5-0.5B-Instruct-W8A8 to Qwen3-0.6B-W8A8
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: menogrey <1299267905@qq.com>
### What this PR does / why we need it?
Some E2E testcases are not in our CI workflow, this PR add them back.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: menogrey <1299267905@qq.com>
RFC: https://github.com/vllm-project/vllm-ascend/issues/4629
Reason：
The functions related to Cp differ significantly from those of normal
MLA-Attention, but the coupling is quite severe.
Steps：
1)Extract common code AscendMLAMetadataBuilder.build to 4 functions:
build_prefill_metadata, build_decode_metadata,build_cp_metadata,
build_chunked_metadata
todo：
1)refactor function _compute_prefill_context;
2)refactor function _mla_preprocess,_mla_decode_preprocess
3）Extract public data and processing functions from the attention_cp.py
and mla_cp.py files to the common_cp file.
vLLM version: 0.13.0rc3
vLLM main:
ad32e3e19c
- vLLM version: 0.13.0rc3
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wujinyuan1 <wjy9595@qq.com>
Signed-off-by: wujinyuan1 <wujinyuan1@huawei.com>
Co-authored-by: wujinyuan1 <wjy9595@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
In certain scenarios (such as smoke testing), the source code is used to
update the vllm-ascend version for running updated models (such as
Qwen3-VL). However, vllm and vllm-ascend themselves have no restrictions
on the transformer version, and the transformer will not be updated,
resulting in errors when launching the model.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: 李少鹏 <lishaopeng21@huawei.com>
### What this PR does / why we need it?
### Does this PR introduce _any_ user-facing change?
Fix vllm break:
1. [Enable cuda graph for deepepHT, 5.3% throughput improvement, 4.4%
TTFT improvement] (https://github.com/vllm-project/vllm/pull/29558)
Fix Solution: Add the now-necessary `all2all_backend` parameter. The
impact of this parameter on the original `set_splitting_ops_for_v1`
implementation is only that graph mode is disabled in `vllm` if
`deepep_high_throughput` is enabled; it has no effect on the
`vllm-ascend` logic.
2.[Migrate legacy ViT MultiHeadAttention to new MMEncoderAttention
interface ] (https://github.com/vllm-project/vllm/pull/30684)
Fix Solution: The reason why the GPU does not need to convert qkv to 3D
is that the GPU's flash_attention operator is compatible with 3D and 4D
(b s h d and s b ( h d)), but the NPU's flash_attention_unpad operator
only supports 3D (s b ( h d)). Therefore, we need to introduce the
reshape_qkv_to_3d operation.
4.Skip Tencent-Hunyuan/HunyuanOCR test case, as it has following issue
in upgrade vllm code:
https://github.com/vllm-project/vllm-ascend/issues/5297
### How was this patch tested?
Co-authored-by: zxwang <1476209578@qq.com>
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: leo-pony <nengjunma@outlook.com>
Signed-off-by: zxwang <1476209578@qq.com>
Co-authored-by: zxwang <1476209578@qq.com>
### What this PR does / why we need it?
Move all_reduce logic to AscendFusedMoE.forward, reuse vLLM's logic.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
e2e & ut
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: weichen <calvin_zhu0210@outlook.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
efect e2e ci test：
1. tests/e2e/singlecard/pooling/test_embedding.py: remove the eager
parameter and rename test case
2. tests/e2e/singlecard/pooling/test_scoring.py: Rename test cases
3. tests/e2e/singlecard/pooling/test_classification.py: Rename test case
4. tests/e2e/singlecard/test_quantization.py: remove the eager parameter
and chage model to vllm-ascend/Qwen2.5-0.6B-W8A8 and Rename test case
5. tests/e2e/multicard/test_shared_expert_dp.py: Rename test cases
6. tests/e2e/singlecard/test_sampler.py: Rename test cases
7. tests/e2e/singlecard/test_aclgraph_accuracy.py: Rename test cases
8. tests/e2e/multicard/test_offline_inference_distributed.py: Rename
test cases and remove the eager parameter
9. tests/e2e/multicard/long_sequence/test_accuracy.py: Rename test cases
and remove the eager parameter
10. tests/e2e/multicard/long_sequence/test_basic.py: Rename test cases
and remove the eager parameter
11.tests/e2e/multicard/test_expert_parallel.py:remove the eager
parameter
12.tests/e2e/multicard/test_full_graph_mode.py:remove the eager
parameter
13.tests/e2e/multicard/test_ilama_lora_tp2.py:remove the eager parameter
14.tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py:remove
the eager parameter
15.tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py:remove the
eager parameter
16.tests/e2e/singlecard/test_aclgraph_accuracy.py:remove the eager
parameter
17.tests/e2e/singlecard/test_camem.py:remove the eager parameter
18.tests/e2e/singlecard/test_ilama_lora.py:remove the eager parameter
19.tests/e2e/singlecard/test_multistream_overlap_shared_expert.py:remove
the eager parameter
20.tests/e2e/singlecard/test_vlm.py:remove the eager parameter
21.tests/e2e/singlecard/test_xli:remove the eager parameter
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
### What this PR does / why we need it?
Using `spawn` in continuous testing scenarios
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
[Kthena](https://github.com/volcano-sh/kthena) is a Kubernetes-native
LLM inference platform that transforms how organizations deploy and
manage Large Language Models in production. Built with declarative model
lifecycle management and intelligent request routing, it provides high
performance and enterprise-grade scalability for LLM inference
workloads.
The platform extends Kubernetes with purpose-built Custom Resource
Definitions (CRDs) for managing LLM workloads, supporting multiple
inference engines (vLLM, SGLang, Triton) and advanced serving patterns
like prefill-decode disaggregation.
This pr added a example on deloying llm on Ascend Kubernetes clusters.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: Zhonghu Xu <xuzhonghu@huawei.com>
### What this PR does / why we need it?
- Standardize test case naming in `vllm-ascend/tests/e2e/multicard/` to
follow the `<model>_<feature>_<distributed>` convention.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: MrZ20 <2609716663@qq.com>
Signed-off-by: root <root@LAPTOP-VQKDDVMG.localdomain>
Co-authored-by: root <root@LAPTOP-VQKDDVMG.localdomain>
### What this PR does / why we need it?
Add dynamic EPLB CI for qwen3-moe-30B-W8A8
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
### What this PR does / why we need it?
Support KV-Sharing feature in CLA (cross layer attention) models, which
sharing kv cache in some layers.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
This patch add handling of `XDRotaryEmbedding` in modelrunner to support
for `hunyuan-vl`
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
CI passed with added/exist tests
Closes: https://github.com/vllm-project/vllm-ascend/issues/4992
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
Following https://github.com/vllm-project/vllm/pull/29873, register
`AscendApplyRotaryEmb` CustomOp and remove related patch.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
#### ✅ Test Qwen2.5-VL
Run:
```bash
vllm serve /root/.cache/modelscope/hub/models/Qwen/Qwen2.5-VL-7B-Instruct \
--max_model_len 16384
```
Output:
```
{"id":"chatcmpl-b02c1ff3415d2462","object":"chat.completion","created":1766129265,"model":"/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-VL-7B-In struct","choices":[{"index":0,"message":{"role":"assistant","content":"The text in the illustration is \"TONGYI Qwen.\" The word \"TONGYI\" is writ ten in blue, and \"Qwen\" is written in gray. The text appears to be part of a logo or branding design.","refusal":null,"annotations":null,"audio": null,"function_call":null,"tool_calls":[],"reasoning":null,"reasoning_content":null},"logprobs":null,"finish_reason":"stop","stop_reason":null,"tok en_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":78,"total_tokens":129,"completion_tokens":51,"prompt_tokens_d
```
#### ✅ Test Qwen3-VL
Run:
```bash
vllm serve /root/.cache/modelscope/hub/models/Qwen/Qwen3-VL-8B-Instruct \
--max_model_len 16384
```
Output:
```
{"id":"chatcmpl-a3a7de5a900a9321","object":"chat.completion","created":1766129586,"model":"/root/.cache/modelscope/hub/models/Qwen/Qwen3-VL-8B-Instruct","choices":[{"index":0,"message":{"role":"assistant","content":"The text in the illustration is **“TONGYI Qwen”**.\n\n### How it looks:\n- **“TONGYI”** is written in **uppercase letters** in a **bold, modern sans-serif font**, colored **blue**.\n- **“Qwen”** is written in **lowercase letters** in a **slightly thinner, elegant sans-serif font**, colored **dark gray**.\n- The two lines of text are stacked vertically, with “TONG","refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":null,"reasoning_content":null},"logprobs":null,"finish_reason":"length","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":112,"total_tokens":212,"completion_tokens":100,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null}
```
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: shen-shanshan <467638484@qq.com>
### What this PR does / why we need it?
[Doc] Add new contributors and relative scripts.
Usage of scripts:
- `export GITHUB_TOKEN=<your github token>`
- `bash tools/collect_user_first_contribution.sh
vllm-project/vllm-ascend <base_sha> <head_sha>` and save the result to
one temporary file such as `contributors.txt`
- `python tools/format_contributors.py contributors.txt --start <start
index now>`
- Use the output to update the `contributors.md`
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: menogrey <1299267905@qq.com>
### Motivation.
**Limitations of the current vLLM v1 scheduling strategy**
vLLM v1 scheduling currently enables chunkedprefill by default, which
processes prefill and decode requests simultaneously in a single
scheduling session. This can impact the overall system throughput and
performance in some scenarios.
Balance scheduling addresses this issue by synchronizing the number of
running queues across all schedulers to delay the scheduling of new
requests, thereby improving the overall system's steady-state decoding
time. This achieves:
✅Adding `balance_gather` to the scheduler synchronizes the number of
requests in the running queues between DPs.
✅Balance scheduling improves the decode steady-state time, thereby
increasing the overall output throughput of the inference system.
### Proposed Change.
**1.Feature Overview**
In the vLLM scheduler, running requests (i.e., requests that are already
undergoing pre-filled computation) have the highest priority, followed
by waiting requests (i.e., requests that have not yet been computed).
As shown in the diagram above, when the entire inference system exits
from a steady state, the scheduler will schedule a batch of new requests
for prefill operations and then synchronize them among the dynamic
programming (DP) models. This can cause some DP models that are entirely
decoded to synchronize with the number of prefilled tokens. Frequent
prefill scheduling by certain DP models can lead to a deterioration in
the overall system output throughput.
Balance scheduling synchronizes the number of running queue requests
across different DPs, and only schedules new requests for prefilling
when at least every scheduler has fewer than max_nun_requst.
**2.Implementation Design**
**3.Experiment Results**
- Fixed-length input scenario: In the performance test scenario with
3.5K fixed-length input and 1.5K fixed-length output, the throughput
performance was improved by approximately **18%** after adding balance
scheduling.
| Method | Model | Input Len | Request Count | Output Len | BatchSize |
Average TTFT | Average TPOT | e2e duration | Input Token Throughput |
Output Token Throughput | Request Throughput
| ---- | ---- | ---- | ---- | ---- | ---- | ---- | ---- | ---- | ---- |
---- | ---- |
| Baseline | DeepSeekV3.1 | 3500 | 512 | 1500 | 128 | 6600 | 86.85 |
591.9s | 3030.5 | 1297.3 | 0.86 |
| Balance scheduling | DeepSeekV3.1 | 3500 | 512 | 1500 | 128 | 7012 |
70.63 | 501.7s | 3575.7 | 1530.7 | 1.02 |
**4.Demo PR**
[#29721 ](https://github.com/vllm-project/vllm/pull/29721)
---------
Signed-off-by: GDzhu01 <809721801@qq.com>
### What this PR does / why we need it?
Update the weight download URL. Because the model was renamed.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: menogrey <1299267905@qq.com>
