#### What this PR does / why we need it?
This PR adapt DispatchGmmCombineDecode operator to eplb tensor list and
expert token numbers.
This operator support gmm1, gmm2, gmm1Scale and gmm2Scale in format of
list.
This operator support couting how many token each local expert recieves
by expertTokensNum .
- vLLM version: v0.13.0
- vLLM main:
7157596103
More info about this operator, please refer to RFC: issue
https://github.com/vllm-project/vllm-ascend/issues/5476
# What this PR does / why we need it?
Add DeepSeek-V3.2-W8A8 dual-node nightly CI test and update A3 nightly
test configuration:
1. Add DeepSeek-V3.2-W8A8 dual-node test:
tests/e2e/nightly/multi_node/config/DeepSeek-V3_2-W8A8-A3-dual-nodes.yaml
- 2 nodes, 16 NPUs per node (32 NPUs total)
- Configuration: 2P+1D (data-parallel-size=4, tensor-parallel-size=8,
data-parallel-size-local=2)
- Includes performance and accuracy benchmarks with GSM8K dataset
2. Update A3 nightly workflow: .github/workflows/nightly_test_a3.yaml
- Added DeepSeek-V3.2-W8A8 dual-node test to the A3 nightly test matrix
- Test name: multi-node-dpsk3.2-2node
3. Improve test scripts: Updated
.github/workflows/_e2e_nightly_multi_node.yaml and related scripts for
better multi-node testing support
test on A3 instances
- Performance baseline: 1 (threshold: 0.97)
- Accuracy baseline: 95% (threshold: 5%)
- Test dataset: GSM8K with 512 prompts for performance, gsm8k-lite for
accuracy
---------
Signed-off-by: guozr <guozr1997@hotmail.com>
Co-authored-by: guozr <guozr1997@hotmail.com>
### What this PR does / why we need it?
Purpose: initialize padded slot mapping buffer to prevent garbage
values.
In PCP mode, the `pcp_padded_slot_mapping` buffer is reused across
invocations. Without explicit initialization, this buffer retain stale
values from previous runs, which can lead to incorrect results.
This change ensures the buffer is filled with -1.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: F.Liu <liufeng248@huawei.com>
Co-authored-by: F.Liu <liufeng248@huawei.com>
### What this PR does / why we need it?
1. add `COMPILATION_PASS_KEY` constant
2. clean up useless platform interface `empty_cache`, `synchronize`,
`mem_get_info`, `clear_npu_memory`
3. rename `CUSTOM_OP_REGISTERED` to `_CUSTOM_OP_REGISTERED`
4. remove uesless env `VLLM_ENABLE_CUDAGRAPH_GC`
NPUPlatform is the interface called by vLLM. Do not call it inner
vllm-ascend.
### Does this PR introduce _any_ user-facing change?
This PR is just a cleanup. All CI should pass.
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
This PR builds upon PR
https://github.com/vllm-project/vllm-ascend/pull/5011 and aims to
further enhance the npu_graph_ex_passes module. Based on prior work, we
have added graph optimization support for the add_rms_quant fused
operator in scenarios where a bias term is present—ensuring the fusion
pattern is correctly registered and matched into the computation graph.
For validation, we switched to the Qwen3-235B-A22B-W8A8 model for
SPPatternWithBias and Qwen3-32B model for SPPattern. Benchmark results
show that, compared to the unfused baseline, enabling this fusion pass
significantly improves inference throughput for W8A8 quantized models.
For more details can refer to the
RFC:https://github.com/vllm-project/vllm-ascend/issues/4715
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
```
llm = LLM(
model=model,
tensor_parallel_size=GPUs_per_dp_rank,
enforce_eager=False,
enable_expert_parallel=enable_expert_parallel,
trust_remote_code=trust_remote_code,
gpu_memory_utilization=0.98,
max_num_batched_tokens=512,
# load_format="dummy",
max_model_len=2048,
max_num_seqs=16,
quantization="ascend",
additional_config={
"refresh": True,
"enable_npugraph_ex": True
},
compilation_config={
"cudagraph_capture_sizes": [8, 16],
"cudagraph_mode": "FULL_DECODE_ONLY",
},
)
if profile_dir:
llm.start_profile()
outputs = llm.generate(prompts, sampling_params)
if profile_dir:
llm.stop_profile()
for i, output in enumerate(outputs):
if i >= 5:
break
prompt = output.prompt
generated_text = output.outputs[0].text
print(
f"DP rank {global_dp_rank}, Prompt: {prompt!r}, "
f"Generated text: {generated_text!r}"
)
```
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: cjian <2318164299@qq.com>
### What this PR does / why we need it?
[Bugfix] fix dcp_only bug and add e2e accuracy test for dcp only and pcp
only
this pr fix the bug of accuracy test when decode_parallel_size>1 and
prefill_context_parallel_size=1.
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
---------
Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
### What this PR does / why we need it?
Revert PR 5253 to fix the smoking problem
### Does this PR introduce _any_ user-facing change?
Does not.
### How was this patch tested?
It was tested in the failure case.
Signed-off-by: Rifa <865071616@qq.com>
### What this PR does / why we need it?
[P/D] Performance enhancement of Layerwise connector in TP asymmetric
scenarios
1. Session fusion: For transmission tasks at each layer, aggregate
transmission tasks with the same destination and merge them into a
single task for assignment.
2. Alltoall aggregation: For TP asymmetric scenarios, perform all
alltoall operations at once according to the block granularity for all
requests.
[RFC]: CDCP Scheduling for Disaggregated Prefilling with KV Cache
Layerwise Push Support
https://github.com/vllm-project/vllm-ascend/issues/4842
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
45c1ca1ca1
---------
Signed-off-by: liziyu <liziyu16@huawei.com>
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Co-authored-by: nwpu-zxr <zhouxuerong2@huawei.com>
Co-authored-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
There was an accuracy issue with **Qwen3-Next-80B-A3B-Instruct-W8A8**
model in the old version of **Triton-Ascend**, so, we are now adding one
nightly test to maintain it.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: IncSec <1790766300@qq.com>
### What this PR does / why we need it?
When using the swa parameter in fia, headDim does not currently support
256, and when gemma3's headDim is equal to 256, an error will occur.
Therefore, code rollback is required, and it will be incorporated after
cann supports it.
### Does this PR introduce _any_ user-facing change?
Remove swa parameter of fia.
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
---------
Signed-off-by: nsdie <yeyifan@huawei.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
#### Overview
This PR fixes a shape mismatch bug between `expert_placement_map` and
`log2phy_expert_map` when **redundant experts** are enabled in the
vLLM-Ascend platform. The issue occurred during the initialization of
expert maps and their updates via EPLB (Expert Load Balancer)
adjustment, leading to potential tensor shape errors and incorrect
expert routing in distributed MoE deployments.
#### Key Changes
1. **Unify expert map shape calculation logic**
- Ensure the shape of `expert_placement_map` and `log2phy_expert_map`
strictly aligns with the total number of experts (including redundant
experts) during initialization.
- Update the shape adjustment logic in EPLB dynamic update process to
match the initial expert map dimensions.
2. **Add shape consistency checks**
- Add assertion statements to verify the shape consistency of the two
maps after initialization and EPLB adjustment, preventing silent shape
mismatches in subsequent operations.
#### Impact
- Resolves tensor shape errors when using redundant experts with EPLB on
Ascend platform.
- Ensures correct expert routing and load balancing for MoE models with
redundant expert configurations.
- No breaking changes to existing functionality; compatible with
non-redundant expert deployments.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: Che Ruan <cr623@ic.ac.uk>
Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
Co-authored-by: Che Ruan <cr623@ic.ac.uk>
Co-authored-by: shenchuxiaofugui <1311027364@qq.com>
### What this PR does / why we need it?
This PR aims to delete redundant methods in mtp_proposer. All the
deleted methods now can be found in eagle_proposer. We also remove some
methods in eagle_proposer since they are identical to those in
vllm-eagle.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
by ci
- vLLM version: release/v0.13.0
- vLLM main:
81786c8774
---------
Signed-off-by: Zetong Li <slippersss@126.com>
### What this PR does / why we need it?
Pin the xlite version to avoid CI failures during its upgrade.
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: changdawei1 <changdawei3@huawei.com>
### What this PR does / why we need it?
Upgrade vllm commit to 0106
- vLLM version: v0.13.0
- vLLM main:
8be6432bda
Signed-off-by: wjunLu <wjunlu217@gmail.com>
### What this PR does / why we need it?
update bisheng version in 20260105
- vLLM version: v0.13.0
- vLLM main:
8be6432bda
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
## Purpose
This PR add unit test for `compute_slot_mapping` function in
`block_table.py` with various `pcp_size` & `dcp_size` &
`cp_kv_cache_interleave_size`.
## Test Plan
```
pytest tests/ut/worker/test_block_table.py
```
## Test Result
```
==== 3 passed, 2 warnings in 0.20s ====
```
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
mv ops to correct path
:`tests/e2e/nightly/single_node/ops/singlecard_ops/triton`
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
Fix Smoke Testing Bug for DSR1 longseq
We need to make this change because the daily smoke test case is
throwing an error: "max_tokens or max_completion_tokens is too large:
32768.This model's maximum context length is 32768 tokens and your
request has 128 input tokens". We encounter this error due to
max-out-len equals to max-model-len. We can fix this error by increasing
max-model-len argument in the script.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: daishixun <dsxsteven@sina.com>
This reverts commit fb9fdcdbe4.
### What this PR does / why we need it?
this pr breaks the smoke test because of that leads the error of
aclnnNeScalar:Kernel Run failed. opType: 25, NotEqual
launch failed for NotEqual, errno:361001
<img width="1149" height="166"
alt="A6C9453D-4F0B-4256-DD80-A9C181DAB2D9"
src="https://github.com/user-attachments/assets/cab9c4b8-3fd1-4c6b-b424-474b46042726"
/>
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: zxwang <1476209578@qq.com>
### What this PR does / why we need it?
Add nightly test for triton split_rmsnorm_rope
### 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: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
- Problem: In MLA+MLAPO, KV-consumer deployments keep
fused_qkv_a_proj/q_proj weights and quant params even though MLAPO uses
the prepacked buffers, increasing memory footprint on decode nodes.
- Fix: Conditionally drop those tensors only when
`kv_transfer_config.is_kv_consumer` to reclaim memory (consistent with
the SFA behavior #4774 ).
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: Chen Chen <0109chenchen@gmail.com>
### What this PR does / why we need it?
Remove the incorrectly depicted DCP all_gather operation in the prefill
stage PCP for GQA diagram.
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
update triton-ascend version to 20260105
- vLLM version: v0.13.0
- vLLM main:
7157596103
---------
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
### What this PR does / why we need it?
In mooncake kvpool, `local_hostname` is not used. Instead, the local IP
is obtained directly via `get_ip()`. Therefore, remove this parameter to
avoid confusion.
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: LCAIZJ <leichao139636@163.com>
### What this PR does / why we need it?
In scenarios where models like
[Moonlight](https://modelscope.cn/models/moonshotai/Moonlight-16B-A3B-Instruct)
(using MLA but without `rope_scaling` in config.json) invoke
`AscendRotaryEmbedding`. `_cos_cache` and `_sin_cache` are not recorded
correctly.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
45c1ca1ca1
Signed-off-by: Debonex <719893090@qq.com>
### What this PR does / why we need it?
This fixes a bug that occurred when running `test_camem.py` in the
triton-ascend environment `NPU function error:
aclrtGetMemInfo(ACL_HBM_MEM, &device_free, &device_total)`
- vLLM version: v0.13.0
- vLLM main:
5326c89803
---------
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
### What this PR does / why we need it?
[UT]add triton ops ut : test_fused_qkvzba_split_reshape_cat
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
pytest -sv tests/ut/ops/test_fused_qkvzba_split_reshape_cat.py
- vLLM version: v0.13.0
- vLLM main:
5326c89803
---------
Signed-off-by: ZT-AIA <1028681969@qq.com>
### What this PR does / why we need it?
Add a new workflow to allow developers submit a pull_request downloading
new models for CI cache
- vLLM version: release/v0.13.0
- vLLM main:
254f6b9867
Signed-off-by: wangli <wangli858794774@gmail.com>
Fixes#3386
- Update Qwen3-30B-A3B.md to use A3-specific image tag
- Update Qwen3-Dense.md to provide both A2 and A3 image options
- Update Qwen3-Next.md to use A3-specific image for Atlas A3
environments
Previously, documentation only mentioned A2 images (vllm-ascend:version)
but Atlas A3 machines require A3-specific images
(vllm-ascend:version-a3). This change ensures users select the correct
image for their hardware.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
Signed-off-by: hu-qi <huqi1024@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Fixes#2727
- Add NNAL to the software requirements table with version information
- Add note explaining that prebuilt Docker images include NNAL
- Add warning message for manual installation when encountering
libatb.so errors
- Improve visibility of NNAL installation instructions to prevent
runtime errors
This addresses the issue where users encounter 'libatb.so not found'
errors due to missing NNAL installation in their environment.
### What this PR does / why we need it?
### 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>
Signed-off-by: hu-qi <huqi1024@gmail.com>
Co-authored-by: zhangyiming <34808445+menogrey@users.noreply.github.com>
### Does this PR introduce _any_ user-facing change?
suffix spec decode method rely on `arctic-inference` library. This PR
add it into requirements to make sure the function works by default
### How was this patch tested?
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: frankie-ys <yongshengwang@cmbchina.com>
Signed-off-by: frankie <wangyongsheng686@gmail.com>
Currently, the vllm pull request
(https://github.com/vllm-project/vllm/pull/24252) is causing operator
fusion to fail. This issue was previously fixed by patching the backend.
The root cause has been identified, and the problem can be resolved with
this pull request.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wxsIcey <1790571317@qq.com>
### What this PR does / why we need it?
Replace multiple PyTorch operations with a fused Triton kernel to
determine token indices for sampling during speculative decoding. This
reduces kernel launch overhead and memory traffic, improving overall
performance on Ascend hardware.
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
This documentation provides a comprehensive technical guide for
deploying **vLLM-Ascend** using a **Prefill-Decode (PD) colocated
architecture** integrated with **Mooncake**, a high-performance
distributed KV Cache transfer engine. As Large Language Model (LLM)
serving scales, managing KV Cache efficiently across distributed nodes
is essential for reducing latency and optimizing hardware utilization.
The tutorial focuses on a multi-instance setup using Huawei **Atlas 800T
A2** nodes. By leveraging Mooncake’s distributed memory pooling, vLLM
instances can achieve seamless **cross-node KV Cache reuse**. This
capability allows an instance to retrieve precomputed cache from a
remote node's DRAM via high-speed **RoCE** networks, effectively
bypassing redundant prefill computations.
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: release/v0.13.0
- vLLM main:
0bfd7484fd
---------
Signed-off-by: zhangmuzhibangde <1037640609@qq.com>
Signed-off-by: zhangmuzhi_yuwan <1037640609@qq.com>
Co-authored-by: Yikun Jiang <yikunkero@gmail.com>
### What this PR does / why we need it?
Fix the bug in the PCP overlay feature
1、Fix the bug related to PCP and EPLB overlap by including PCP size in
the word_size calculation.
2、In the PCP pooling scenario, a prompt has been added for setting the
cp_kv_cache_interleave_size.
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
### What this PR does / why we need it?
1. Refactor eagle and mtp function: load_model and generate_token_ids
2. Remove redundant code in mtp and eagle file
3. Refactor the UT of file
2/N of Refactor and merge mtp and eagle
Relational RFC: https://github.com/vllm-project/vllm-ascend/issues/5467
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
ut and tests
- vLLM version: release/v0.13.0
- vLLM main:
81786c8774
---------
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
### What this PR does / why we need it?
Since the _npu_ring_mla operator deteriorates in long-sequencescenarios,
the long sequence is split into shorter sequences for input to improve
performance.
- vLLM version: v0.13.0
- vLLM main:
5326c89803
---------
Signed-off-by: pichangping <1337510399@qq.com>
### What this PR does / why we need it?
skip xlite-decode-only e2e test, since it's unstable
- vLLM version: release/v0.13.0
- vLLM main:
81786c8774
Signed-off-by: changdawei1 <changdawei3@huawei.com>
### What this PR does / why we need it?
In the training-inference switching scenario, there is no need to resume
the model weights during KV cache resumption, as this would lead to
format mismatch.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: p00465316 <panchao13@huawei.com>
Co-authored-by: p00465316 <panchao13@huawei.com>
### What this PR does / why we need it?
Fix chunk prefill bug for long_sequence feature
When there are two requests with chunk prefill enabled in the
long-sequence scenario, if one request has only 1 token during
scheduling, it will be identified as a decode request and trigger an
error. This PR fixes the issue.
Closes: https://github.com/vllm-project/vllm-ascend/issues/5445
- vLLM version: release/v0.13.0
- vLLM main:
81786c8774
---------
Signed-off-by: LookAround <lixushi@huawei.com>
