### What this PR does / why we need it?
This patch adds support for the xlite graph wrapper to vllm_ascend.
Xlite provides operator implementations of the transformer network on
Ascend hardware. For details about xlite, please refer to the following
link: https://gitee.com/openeuler/GVirt/blob/master/xlite/README.md
The latest performance comparison data between xlite and the default
aclgraph mode is as follows:
## Qwen3 32B TPS 910B3(A2) Online Inference Performance Comparison
- aclgraph: main(c4a71fc6)
- xlite-full: main(c4a71fc6) + xlite-full
- xlite-decode-only: main(c4a71fc6) + xlite-decode-only
- diff1: Performance comparison between xlite-full and aclgraph
- diff2: Performance comparison between xlite-decode-only and aclgraph
### Does this PR introduce _any_ user-facing change?
Enable the xlite graph mode by setting xlite_graph_config:
--additional-config='{"xlite_graph_config": {"enabled": true}}' #
Enabled for decode only
--additional-config='{"xlite_graph_config": {"enabled": true,
"full_mode": true}}' # Enabled for prefill and decode
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: lulina <lina.lulina@huawei.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
It's safe to drop ascend scheduler now. The related test and doc has
been removed already
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
The main goal of this PR to alleviate the high maintenance burden from
model duplication when we are going to do the model optimization. Some
of our optimized models diverges a little from the vllm's modeling, but
needs to rewrite several part of original one, brings negligible
maintenance bruden to the vllm-ascend.In order to solve that, we propose
to leverage `torch.compile` and `inductor pattern matcher`,
automatically fuse the pattern we want to merge. For more details can
refer to the RFC https://github.com/vllm-project/vllm-ascend/issues/4239
This pr integrates `AddRMSNorm` and the `Quant` operator, which can
improve the inference speed of models using `w8a8 `quantization.
### Does this PR introduce _any_ user-facing change?
Yes, add new additional_config
### How was this patch tested?
```python
def main():
prompts = [
"The president of the United States is Mr.",
]
# Create a sampling params object.
sampling_params = SamplingParams(max_tokens=100, temperature=0.6, top_k=40, top_p=0.95)
# Create an LLM.
llm = LLM(
model="/root/.cache/modelscope/hub/models/vllm-ascend/Qwen3-8B-W8A8",
# enforce_eager=True,
tensor_parallel_size=1,
trust_remote_code=True,
gpu_memory_utilization=0.7,
quantization="ascend",
)
# Generate texts from the prompts.
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
```
```text
Prompt: 'The president of the United States is Mr.', Generated text: ' Trump. The president of the United States is Mr. Biden. Which of the following statements is correct? \n\nA. Mr. Trump is Mr. Biden. \nB. Mr. Trump is not Mr. Biden. \nC. The president of the United States is not Mr. Trump. \nD. The president of the United States is not Mr. Biden.\n\nThe question presents a contradiction: it states that "The president of the United States is Mr. Trump" and "The president of'
```
- vLLM version: 86e178f7c4d8c3b0eaf3c8e3f810a83f63b90e24
- vLLM main:
86e178f7c4
---------
Signed-off-by: Icey <1790571317@qq.com>
Signed-off-by: wxsIcey <1790571317@qq.com>
Move Custom ops register to correct place to make CI happy
- vLLM version: 86e178f7c4d8c3b0eaf3c8e3f810a83f63b90e24
- vLLM main:
86e178f7c4
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
1. Fixes the environment path used to locate custom op shared libraries.
2. Uses empty tensor initialization for op outputs instead of
zero-initialization for better efficiency.
- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2
---------
Signed-off-by: QianChenxi <chenxi.qian.cq@outlook.com>
### What this PR does / why we need it?
Support shared expert DP for deepseek_mtp feature.
`shared_expert_dp` requires `SP==True`, with corresponding parameter
restrictions.
Previously, due to the coupling between `shared_expert_dp` and torchair,
and the removal of `deepseek_mtp` in vllm_ascend, shared expert dp of
deepseek_mtp was temporarily removed.
Currently, by performing the `reduce_scatter` on the input of
deepssek_mtp in `mtp_proposer.py`, we ensure that it matches the
dimensions of `input_embedding`, and then perform the `all_gather` on
the output of mtp.
### How was this patch tested?
baseline:
<img width="1184" height="692" alt="image"
src="https://github.com/user-attachments/assets/9680d53a-7b1d-481a-accc-b8f3dae2b9e3"
/>
enable shared_expert_dp and multistream_overlap_shared_expert:
<img width="1167" height="687" alt="image"
src="https://github.com/user-attachments/assets/2531d06b-dfda-4e24-8628-6f4b0f677ddc"
/>
TPOT: 48ms -> 45.4ms
Average TPS per rank: 117.6 -> 126.1
- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2
---------
Signed-off-by: chenmenglong <chenmenglong1@huawei.com>
Signed-off-by: zengran <zengran2@huawei.com>
Co-authored-by: zengran <zengran2@huawei.com>
Ascend scheduler was added for non chunk prefill case before, since that
the npu ops didn't work well with chunked prefill.
Now the ops with chunked prefill work better, it's time to remove the
ascend scheduler to use vLLM default scheduler.
- vLLM version: v0.11.2
---------
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
This PR introduces support for adding custom CANN `aclnn` ops to
`vllm-ascend`, allowing users to define and use their own custom
operators.
Key changes include:
- Building and installing custom ops into the `vllm-ascend`-specified
directory
- Binding the `aclnn` op interface to the `torch.ops._C_ascend` module
- Enabling invocation of these ops within `vllm-ascend`
This PR includes a sample custom op:
`aclnnGroupedMatmulSwigluQuantWeightNzTensorList`, which is adapted from
the CANN operator
[`aclnnGroupedMatmulSwigluQuantWeightNZ`](https://www.hiascend.com/document/detail/zh/canncommercial/83RC1/API/aolapi/context/aclnnGroupedMatmulSwigluQuantWeightNZ.md).
Its input parameters `weight` and `weight_scale` now accept
`list[torch.Tensor]` (i.e., `at::TensorList`).
### Does this PR introduce _any_ user-facing change?
No.
- vLLM version: v0.11.2
---------
Signed-off-by: QianChenxi <chenxi.qian.cq@outlook.com>
### What this PR does / why we need it?
While using the LLM Compressor quantization tool from the VLLM community
to generate quantized weights, the VLLM Ascend engine needs to be
adapted to support the compressed tensors quantization format.
1. Add AscendCompressedTensorsConfig to replace CompressedTensorsConfig
in vllm.
2. Support CompressedTensorsW8A8 static weight.
- weight: per-channel, int8, symmetric; activation: per-tensor, int8,
symmetric.
4. Support CompressedTensorsW8A8Dynamic weight.
- weight: per-channel, int8, symmetric; activation: per-token, int8,
symmetric, dynamic.
5. Modify the override_quantization_method in AscendQuantConfig.
Co-authored-by: taoqun110 taoqun@huawei.com
Co-authored-by: chenxi-hh chen464822955@163.com
- vLLM version: v0.11.2
---------
Signed-off-by: LHXuuu <scut_xlh@163.com>
Signed-off-by: chenxi-hh <chen464822955@163.com>
Signed-off-by: chenxi-hh <32731611+chenxi-hh@users.noreply.github.com>
Co-authored-by: chenxi-hh <chen464822955@163.com>
Co-authored-by: chenxi-hh <32731611+chenxi-hh@users.noreply.github.com>
### What this PR does / why we need it?
Currently, there are two paths to judge the chip type in code,
`get_ascend_soc_version` use `get_soc_version` api in torch_npu, and
`is_310p` `use _build_info.__soc_version__`, which generate when
install. We need to unify the two paths.
We need to unify these codes based on the following points:
1. We need to ensure consistency in chip type judgment between compiling
and running states;
2. In compiling state, we need chip type to complete op's compilation,
but in running state, we only need device
type(910B/910_93/310P/910_95/etc) to make code branch judgement;
3. In compiling state, torch_npu may not have been installed yet, so we
can't use torch_npu's api.
Based on the above points, we have made the following changes:
1. When user set env `SOC_VERSION`, use it; when not set, query
soc_version by `npu-smi`;
2. generate device_type based on soc_version when compiling, and write
`__device_type__` instead of `__soc_version__` in `_build_info.py`;
3. In running state, use `__device_type__` to judge code branch.
### Does this PR introduce _any_ user-facing change?
When not set env `SOC_VERSION`, it will not be `ASCEND910B1` by default,
we will query soc_version by `npu-smi`. And env `SOC_VERSION` must be in
the list `soc_to_device` in `setup.py`.
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
Signed-off-by: zzzzwwjj <1183291235@qq.com>
There is a lot hack code for v0.11.0, which makes the code hard to
upgrade to newer vLLM version. Since v0.11.0 will release soon. Let's
drop v0.11.0 support first. Then we'll upgrade to v0.11.2 soon.
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Add error log for VL models when enabling
`VLLM_ASCEND_ENABLE_FLASHCOMM1=1` or `VLLM_ASCEND_ENABLE_FLASHCOMM=1`
(for backward compatibility).
This is a temporary fix for
https://github.com/vllm-project/vllm-ascend/issues/4132.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
Signed-off-by: shen-shanshan <467638484@qq.com>
### What this PR does / why we need it?
Currently, the default `cudagraph_capture_size` in vLLM is `[1, 2, 4 ,8
,16 ,24 ,... , max_capture_size]`. However, this is not always the best
choice on different situations. This PR aims to change the default
setting when running Qwen3-MoE on full dp (`dp_size > 1` && `tp_size ==
1`) setting, which is usually applied in Large-Scale EP.
old :
`[1, 2, 4 ,8 ,16 ,24 ,... , max_capture_size]`
new:
`[1, 2, 5 ,10 ,15, 16 ,24 ,... , max_capture_size]`
This is mainly because the performance of `_npu_paged_attention` op
degrades dramatically on old settings. We hope to provide better
performance if users do not set specific `cudagraph_capture_size`.
### Does this PR introduce _any_ user-facing change?
The default `cudagraph_capture_size` is modified in above cases.
However, if `cudagraph_capture_size` has already set by users, this PR
won't have any influence on this.
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
Fix moe error when sp chunked the hidden_states by disabling sp by a hacky way
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
### What this PR does / why we need it?
The current library only supports the FullDecodeOnly graph mode, which
enables full graph execution during the decode. This PR extends support
to allow full graph execution in both the prefill and decode, referred
to as FULL graph mode.
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
### What this PR does / why we need it?
Add import_kernels interface to avoid import useless vLLM C library
Closes#3488. Reopen#3498 for CI.
### How was this patch tested?
CI tested.
- vLLM version: v0.11.0
- vLLM main:
2918c1b49c
---------
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
Drop VLLM_USE_V1 usage. This env has been removed from vLLM already.
- vLLM version: v0.11.0
- vLLM main:
83f478bb19
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
This PR reverts the changes introduced in PR #2894 Initially, due to
performance issues with the older version of the chunked prefill ops,
the default behavior was to use the Ascend scheduler to disable the
chunked prefill feature. However, with the improvements in the
performance of the new chunked prefill ops, this interception strategy
has been removed. This change also aligns with the community's default
configuration behavior.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.11.0
- vLLM main:
83f478bb19
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
1. Revert [bugfix for mtp in
fullgraph](0948483642)
and support it when vllm supports
2. raise error when cudagraph_capture_sizes can't be an integer multiple
of uniform_decode_query_len
3. bugfix when max_num_seqs=14 in mtp=2 scenario
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
83f478bb19
---------
Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
### What this PR does / why we need it?
bugfix for mtp fullgraph
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main:
83f478bb19
Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
### What this PR does / why we need it?
This PR refactors the Ascend attention implementation to align with
vLLM's core interfaces, simplifying the code and improving
maintainability.
### Key Changes:
* **Align with vLLM's Attention Interface**: The `forward` method
signature in `AscendAttentionBackendImpl` now matches the base
`AttentionImpl` in vLLM, removing the custom `trace_flag`.
* **Enable Opaque Attention Operator**: By adding `opaque_attention_op`
to `AscendPlatform`, we allow vLLM to wrap our attention kernel in its
standard `vllm.unified_attention_with_output` operator. This avoids the
need for a custom call path.
* **Remove Obsolete Code**:
* The custom op `vllm.unified_ascend_attention_with_output` has been
deleted as it is now redundant.
* The `trace_flag` and its associated logic were removed, reducing code
complexity.
* An outdated quantization branch within the attention implementation
was cleaned up.
* **Improve Readability**: Renamed output variables (`output` vs.
`intermediate_output`) and added comments to clarify the in-place nature
of the attention output.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
No extra tests needed.
- vLLM version: v0.11.0rc3
- vLLM main:
17c540a993
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
This reverts commit
bf87606932.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E vllm serving with `enable_shared_expert_dp: true` in eager mode as
before.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
Adds a validation check to prevent running with an incompatible
configuration.
The `ASCEND_LAUNCH_BLOCKING=1` environment variable, used for debugging,
enforces synchronous execution which is incompatible with ACL Graph.
This change raises an explicit error to inform the user about the
conflict and how to resolve it, preventing a more obscure failure later.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
None needed.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
This PR is aimed to fix the recomputing out of memory bug in decode
instance. When recomputing happens in decode, kv cache usage may exceed
the pre-allocated memory, and it will cause OOM.
So we propose a new scheduling strategy, when decode instance cannot
allocate new block for running requests, we will stop the request that
will be preempted. These stopped request will be recognied by proxy, and
they will be send to prefill instance again to calculate kvc and then
direct to decode instance.
This is a temporary plan to fix the bug. The long-term stratege is to
use CPU offload in decode instance.
### Does this PR introduce _any_ user-facing change?
An extra ascend configuration option **-- recompute_scheduler_enable =
True** is added to enable this strategy. The default value is False
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: CHEN <116010019@link.cuhk.edu.cn>
### What this PR does / why we need it?
shared expert dp for deepseek and deepseek_mtp, could be combined with
sp to improve performance.
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
Co-authored-by: realliujiaxu <realliujiaxu@163.com>
### What this PR does / why we need it?
Adapt deepseek-v3.2 to vllm 0.11.0, removing the useless patch.
The final goal is to remove all the patches and align the code arch to
vllm, thus we need to do the following work in next prs.
TODO:
- [x] remove patch on attention spec
- [ ] refactor the kvcache creation logic
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
1. CI passed with existing test.
2. Test pass with deepseek-v3.2-exp
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
[Feat] Supports Aclgraph for bge-m3
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
```
pytest -s tests/e2e/singlecard/test_embedding.py
pytest -s tests/e2e/singlecard/test_embedding_aclgraph.py
```
to start an online server with bs 10, each batch's seq length=8192, we
set --max-num-batched-tokens=8192*10 to ensure encoder is not chunked:
```
vllm serve /home/data/bge-m3 --max_model_len 1024 --served-model-name "bge-m3" --task embed --host 0.0.0.0 --port 9095 --max-num-batched-tokens 81920 --compilation-config '{"cudagraph_capture_sizes":[8192, 10240, 20480, 40960, 81920]}'
```
For bs10, each batch's seq length=8192, QPS is improved from 85 to 104,
which is a 22% improvement, lots of host bound is reduced.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: xuyexiong <xuyexiong@huawei.com>
Co-authored-by: wangyongjun <1104133197@qq.com>
### What this PR does / why we need it?
Adds support for capturing the Multi-Layer Attention (MLA) decode
operation into an ACL graph. This improves performance by compiling the
attention kernel for single-token decoding.
Key changes include:
- Implementing the graph capture logic for the MLA kernel, including
workspace management and parameter updates.
- Modifying the rotary embedding (RoPE) handling to use pre-allocated
tensors, which is a requirement for graph capture.
- Adding a `build_for_graph_capture` method to the MLA metadata builder
to create dummy metadata during the graph compilation phase.
Known issues:
- Currently, MTP is not supported in FULL_DECEDE_ONLY mode -- we're
working on a fix
- We are preparing to remove update_mla_attn_params with
auto_dispatch_capture
### Does this PR introduce _any_ user-facing change?
compilation_config={
"cudagraph_mode": "FULL_DECODE_ONLY",
},
### How was this patch tested?
- vLLM version: v0.11.0
---------
Signed-off-by: panchao-hub <315134829@qq.com>
Signed-off-by: p00465316 <panchao13@huawei.com>
Co-authored-by: p00465316 <panchao13@huawei.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
Fix dp+ep+tp inplace copy error when sp chunked the `hidden_states`.
### How was this patch tested?
test locally with the following scripts
```bash
python examples/offline_data_parallel.py \
--model="Qwen/Qwen3-30B-A3B" \
--dp-size=2 \
--tp-size=2 \
--enable-expert-parallel
```
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
PR #2894 make ascend_scheduler_config.enabled always be `True` for
non-mla models，when `ascend_scheduler_config.enabled=True `, it will
always initialize `AscendScheduler` which is a subclass of `Scheduler`,
but when we enbale async_scheduling,we need to initialize
`AsyncScheduler` in vllm, this will make async_scheduling can't be
enabled.
### Does this PR introduce _any_ user-facing change?
not-related
### How was this patch tested?
when user set `async_scheduling`, it means user don't want to use
`AscendScheduler`, so we shouldn't set `ascend_scheduler_config.enabled
= True`
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
### What this PR does / why we need it?
Upgrade vLLM to newest commit
- Fix the aclgraph doesn't work problem, caused by
24fab45d96
- Fix PoolerOutput import error, caused by
755ed7b05b
- Fix the aclgraph weight load error to keep the same with torchair fix.
4492e3a554
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
All test should pass
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
---------
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
What this PR does / why we need it?
there are two sets of sp implementations for moe and dense models. One
is called sequence_parallelism, and the other is flashcomm_v1.
We did the following things：
Merge two sets of code with the same implementation into one.
Remove the implementation of sequence_parallelism, as this solution
cannot support aclgraph.
Does this PR introduce any user-facing change?
No
How was this patch tested?
e2e&ut
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
---------
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
Note: This depends on [vLLM
#25161](https://github.com/vllm-project/vllm/pull/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:
9607d5eb44
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
