### What this PR does / why we need it?
Qwen3 MoE supports SP. In scenarios like AlltoAll, AlltoAllv, and MC2,
replacing AllReduce with Reduce-Scatter and AllGather achieves
computational benefits in norm operations while saving one AllGather
communication. This feature is enabled during the P-phase and delivers
notable gains in long-sequence scenarios (e.g., 16k–25k), with
performance improvements reaching 5%–10%.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
```
compilation_config={
"pass_config":{
"enable_sequence_parallelism": True
}
},
enable_expert_parallel=True,
```
- vLLM version: v0.10.0
- vLLM main:
9edd1db02b
---------
Signed-off-by: libaokui <libaokui@huawei.com>
Co-authored-by: libaokui <libaokui@huawei.com>
### What this PR does / why we need it?
Support MTP with：
- [x] V0 Scheduler
- [x] TorchAir
- [x] Single DP
- [x] Multi DP
- [x] Disaggregate PD
Known issues：
- [ ] Not support V1 Scheduler (chunked prefill), will be supported in a
few weeks
- [ ] vllm v0.10.0 does not support metrics with `DP > 1` right now,
need to comment out the line 171-175 in file
`vllm/vllm/v1/metrics/loggers.py`
```
if (len(self.engine_indexes) > 1
and vllm_config.speculative_config is not None):
raise NotImplementedError("Prometheus metrics with Spec Decoding "
"with >1 EngineCore per AsyncLLM is not "
"supported yet.")
```
To start an online server with torchair enabled, here is an example:
```
python -m vllm.entrypoints.openai.api_server \
--model="/weights/DeepSeek-R1_w8a8/" \
--trust-remote-code \
--max-model-len 40000 \
--tensor-parallel-size 4 \
--data_parallel_size 4 \
--max-num-seqs 16 \
--no-enable-prefix-caching \
--enable_expert_parallel \
--served-model-name deepseekr1 \
--speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
--quantization ascend \
--host 0.0.0.0 \
--port 1234 \
--additional-config '{"ascend_scheduler_config":{"enabled":true,"enable_chunked_prefill":false},"torchair_graph_config":{"enabled":true,"graph_batch_sizes":[16]},"enable_weight_nz_layout":true}' \
--gpu_memory_utilization 0.9
```
offline example with torchair enabled
```
from vllm import LLM, SamplingParams
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(max_tokens=16, temperature=0)
# Create an LLM.
llm = LLM(
model="/home/data/DeepSeek-R1_w8a8/",
tensor_parallel_size=16,
max_num_seqs=16,
gpu_memory_utilization=0.9,
distributed_executor_backend="mp",
enable_expert_parallel=True,
speculative_config={
"method": "deepseek_mtp",
"num_speculative_tokens": 1,
},
trust_remote_code=True,
enforce_eager=False,
max_model_len=2000,
additional_config = {
'torchair_graph_config': {
'enabled': True,
"graph_batch_sizes": [16],
'enable_multistream_shared_expert': False,
},
"ascend_scheduler_config": {
"enabled": True
},
# 'expert_tensor_parallel_size': 16,
}
)
# Generate texts from the prompts.
# llm.start_profile()
outputs = llm.generate(prompts, sampling_params)
# llm.stop_profile()
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
```
- vLLM version: v0.10.0
- vLLM main:
302962e806
---------
Signed-off-by: xuyexiong <xuyexiong@huawei.com>
### What this PR does / why we need it?
Cherry pick #1291 from v0.9.1-dev, This pr implement the synchronization
of whether `dbo` is enabled across all dp ranks. specifically, it
performed allreduce op across multiple DP ranks, only when all the dp
rank is `enable_dbo`, it is enabled
Co-authored-by: shikang-hangzhou <459956190@qq.com>
Co-authored-by: wangli <wangli858794774@gmail.com>
- vLLM version: v0.10.0
- vLLM main:
2836dd73f1
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
This PR fixes the bug `local variable 'decode_hs_or_q_c' referenced
before assignment` when running chunked-prefill with torchair. We should
calculate `decode_hs_or_q_c` whether or not torchair graphics mode is
enabled.
backport of #1378
fix https://github.com/vllm-project/vllm-ascend/issues/1369
- vLLM version: v0.10.0
- vLLM main:
0e36abf993
---------
Signed-off-by: whx-sjtu <2952154980@qq.com>
Signed-off-by: MengqingCao <cmq0113@163.com>
Co-authored-by: whx-sjtu <2952154980@qq.com>
This PR removes the restriction that TP cannot be greater than 4 in
torchair scenario, because current newest version of CANN has fixed this
bug.
- vLLM version: v0.10.0
- vLLM main:
04ff4be310
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
A refactoring of forward_context and model_runner_v1, add some context
which is necessary in model inference into forward_context, and refactor
dummy_run logic, make it more reasonable.
Some details for this PR:
Add `ascend_forward_context`;
Update mc2_v2 op, and support `active_mask` param;
Update scripts in examples dir;
refactor `dummy_run` logic;
Add soc_version for A2 and A3;
### Does this PR introduce _any_ user-facing change?
No change at user-facing.
### How was this patch tested?
- vLLM version: v0.10.0
- vLLM main:
57c22e57f9
Signed-off-by: zzzzwwjj <1183291235@qq.com>
### What this PR does / why we need it?
- Upgrade to v0.10.0
- Drop v0.9.2 version compatibility
- Add patch for
`vllm_ascend/patch/worker/patch_common/patch_sampler_gather_logprobs.py`
as workaround of
f3a683b7c9
for v0.10.0 and also add e2e test `test_models_prompt_logprobs`
- Pin transformers<4.54.0 as workaround of
https://github.com/vllm-project/vllm-ascend/issues/2034
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- Test locally:
`VLLM_USE_MODELSCOPE=true pytest -sv
tests/e2e/singlecard/test_offline_inference.py::test_models_prompt_logprobs`
- CI passed
- vLLM version: v0.9.2
- vLLM main:
7728dd77bb
---------
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
Before do attention module refactor, we can do some code cleanup to make
the next step easier.
What this PR does:
1. remove uesless `common_prefix_len` for attention builder
2. remove uesless `is_only_prefill` and `num_input_tokens` in attention
metadata.
3. remove `CommonAttentionMetadata` and ues `query_start_loc` instead,
`CommonAttentionMetadata` is over designed and uesless
4. update the attention backend input parameters to keep the same as
vLLM.
5. Rename attention name to the same style with `ASCEND` prefix
- vLLM version: v0.9.2
- vLLM main:
107111a859
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
There is a lot torchair specified logic in common code. It results hard
code maintenance. We will create a new torchair module to launch
torchair related logic there. I plan to add 4 PR.
1. Refactor worker
2. Refactor utils (this PR)
- simple change that move all torchair related util function to torchair
module
3. Refactor model_runner
4. Refactor attention
- vLLM version: v0.9.2
- vLLM main:
8188196a1c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
vLLM commit
752c6ade2e
removed `blocksparse_params` for attention backend. This PR does the
same change to make CI happy.
- vLLM version: v0.9.2
- vLLM main:
9499e26e2a
---------
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
Co-authored-by: Yikun Jiang <yikunkero@gmail.com>
### What this PR does / why we need it?
mla attention still using the gpu_input_batch's attr:`swap_states`, which will lead to
an error `AttributeError: 'InputBatch' object has no attribute 'swap_states'`
This PR fixed the mla input patch error
### How was this patch tested?
will be tested by #1136
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
When use AscendScheduler with prefix-cache enabled and chunk-prefill
disabled, there will be accuray problem because there is no branch in
mla_v1 to process this scenario. This PR fixes it.
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
After #1094, decode might be executed with non-compiled mode, despite of
`torchair_graph_config.enabled`, causing multistream mla to fail, which
assumes torchair compiled mode for decode when
`torchair_graph_config.enabled == True`.
Augment that assumption to fix this.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Tested both offline, and by graph mode mla e2e testcase.
---------
Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
### What this PR does / why we need it?
Fix the issue of insufficient cached cosine and sine length in MLA's
TorchAir graph mode, which causes accuracy deviation during
long-sequence inference.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
We tested the accuracy of this patch with DeepSeek R1 e2e becnhmark
serving, and get 83.33 sore for AIME2024 dataset with DP4TP4EP16
setting.
Signed-off-by: linfeng-yuan <1102311262@qq.com>
<!-- 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?
This PR is used for resolved [issue
1147](https://github.com/vllm-project/vllm-ascend/issues/1147)
1. Move fused_moe code into one file `fused_moe.py`.
2. Integrate branch conditions into function `get_fused_moe_state`.
<!--
- 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?
1. This PR has removed the env `VLLM_ENABLE_MC2`, because I think this
env is useless, we can make judgments based on the current scenario
without this env, it will only increase complexity.
2. This PR has removed the env `USING_LCCL_COM`, because this env has
already expired.
3. `additional_config.expert_tensor_parallel_size` has already expired,
and now we also use parameter `enable_expert_parallel`, consistent with
the vLLM.
<!--
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?
<!--
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: zzzzwwjj <1183291235@qq.com>
### What this PR does / why we need it?
W_UV/W_UK_T cannot be converted to nz, because this position will be
fused into transposebatchmatmul, which does not support nz. The weights
are actually converted back to nd in each run.
### Does this PR introduce _any_ user-facing change?
Use #1098 as the baseline, p90 TPOT 90.79ms->88.58ms, improve TPOP 2ms
### How was this patch tested?
use #1101
---------
Signed-off-by: ttanzhiqiang <389825161@qq.com>
### What this PR does / why we need it?
Move all vector operations to a secondary stream, with the expected
overlaping being:
```
| q_rmsnorm | | kv_norm_rope_cache | | q_rope |
| matmul W_DQ | matmul W_DKV | index | index | matmul W_UQ | split | matmul W_KV_T |
```
Currently, the `IndexByTensor` operators introduced by computation of
`cos` and `sin` can't be offloaded to the secondary stream due to a
known bug of graph fusion optimization pass. So we instead keep it in
the main stream, only requires it be computed before `matmul W_UQ` to
avoid hindering later overlapping. The problem may be solved by later
optimization (#993), which hoists the computation of `cos` and `sin` up
to the first layer.
### Does this PR introduce _any_ user-facing change?
Controlled by `torchair_graph_config.enable_multistream_mla`, defaulted
to False.
### How was this patch tested?
Tested on 1x16 910 node, with tailored 2 layer DSKv2.
Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
### What this PR does / why we need it?
Improve assertion on Graph mode with MLA.
When running deepseek with graph mode, the fused MLA op only support
`numHeads / numKvHeads ∈ {32, 64, 128}`, thus we improve the assertion
info here to avoid users confused with this.
### Does this PR introduce _any_ user-facing change?
Adjusting tp size is required when running deepseek-v3/r1 with graph
mode. deepseek-v2-lite is not supported in graph mode.
### How was this patch tested?
Test locally as the CI machine could not run V3 due to the HBM limits.
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
The former PR https://github.com/vllm-project/vllm-ascend/pull/736
select the valid token inside the `input_ids` and `position_ids` breaks
the necessary padding required by torchair. In this PR, we pending the
pad logic after the multimodal part.
Signed-off-by: ganyi <pleaplusone.gy@gmail.com>
### What this PR does / why we need it?
Based on the design of dual-batch overlap proposed by Deepseek team and
also the implementation of fused moe in VLLM project, we implement the
multi-stream(also known as dual-batch) overlap for deepseek+mla on
Ascend NPU. We split the input batch of model into two microbatches and
then overlap the comp/comm ops in attention and moe layers using two
streams to improve the performance. Our approach can be easily extended
when adding dispatch/combine communications for moe layer.
Compared with the previously proposed
[draft](https://github.com/vllm-project/vllm-ascend/pull/842), we use
one stream for computation ops and the other for communication ops,
separately. In out opinions, it is beneficial for arranging the order of
executing different ops and thus avoiding the contention of
computation/communication resources.
ref: [overlap for
llama](https://github.com/vllm-project/vllm/pull/15787/files)
ref: [dbo in
sglang](https://github.com/sgl-project/sglang/pull/4068/files#diff-b4937569fc71f6ad215181b633b2f89c7183a2b4ac39e41fc22635599a9be7de)
### Does this PR introduce _any_ user-facing change?
Adding an env variable "VLLM_ENABLE_DBO". Users can enable dbo by
setting "VLLM_ASCEND_ENABLE_DBO=1"
See /examples/offline_dualbatch_overlap_npu.py for more info.
### How was this patch tested?
This patch can be tested with vllm-0.9.0 using its online service with
benchmark tests. We have decoupled the func of dbo from vllm and it
should be able to run without any modification to the code of vllm(some
modifications is better to implement in vllm though).
Any advice/discussion is welcome.
### Performance Benchmark
We have ran the benchmark_serving script of vllm to test the performance
after using dual-batch overlap.
`python -m vllm.entrypoints.openai.api_server \
--model=DeepSeek-R1-W8A8 \
--trust-remote-code \
--distributed-executor-backend=mp \
-tp=16 \
--port 8006 \
--max-num-seqs 390 \
--max-model-len 32768 \
--max-num-batched-tokens 65536 \
--block-size 128 \
--compilation_config 0 \
--gpu-memory-utilization 0.90 \
--disable-log-requests \
--additional-config
'{"expert_tensor_parallel_size":1,"enable_inter_dp_scheduling":true,"init_torchair_graph_batch_sizes":true,"trace_recompiles":true,"ascend_scheduler_config":{},"enable_graph_mode":false}'`
and run benchmark with the parameters of :
`--dataset-name random --random-input-len 4096 --random-output-len 1
--num-prompts 200 --max-concurrency 8 --request-rate 5
--metric-percentiles 90`
1. test with the version using allgather+allreduce in Ascend 910B (tp16
ep16 + deepseek r1 w8a8)
2. test with the version using alltoall:
prefill qps: 0.90 -> 1.01
Mean TTFT：8226->7432ms
The overlap approach when using alltoall communication can be further
optimized by overlapping micro-batch1's moe comp with micro-batch2's
dispatch a2a comm
---------
Signed-off-by: zhuohuan <zxdu1997@gmail.com>
More and more config options are added to additional_config. This PR
provide a new AscendConfig to manage these config options by an easier
way to make code cleaner and readable.
This PR also added the `additional_config` doc for users.
Added the test_ascend_config.py to make sure the new AscendConfig works
as expect.
TODO: Add e2e test with torchair and deepseek once the CI resource is
available.
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Fix deepseek with v1, this error is introdeced by
https://github.com/vllm-project/vllm-ascend/pull/945. and this pr fix
the block table of mla
### How was this patch tested?
CI passed with new addedtest.
Signed-off-by: Mengqing Cao <cmq0113@163.com>
<!-- 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?
Currently, the implementation for MLA V1 pads q, k, v to `head_dim` 256
to conform to early MLA kernel. But the new MLA kernel supports
`head_dim` that can't be devided by 128. Therefore we can remove those
unnecessary paddings to boost the performance
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
<!--
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: angazenn <zengyanjia@huawei.com>
Co-authored-by: angazenn <zengyanjia@huawei.com>
### What this PR does / why we need it?
Update attention nz and mla nz modules to improve TPOP 6ms performance
Convert W_UV and W_UK_T to NPU format in mla_v1.py
Convert layer.weight to NPU format in w8a8.py
Signed-off-by: ttanzhiqiang <389825161@qq.com>
### What this PR does / why we need it?
this PR fix CI failure broken by vllm.
1. add moe_config for fused_moe
2. adjust the change for kv cache group from vllm. currently vllm-ascend
doesn't support this feature. this is just a quick fix for backward
compatibility
fix: #872
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
Support the features of prefix cache and chunked prefill in v0/v1.
---------
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
Deepseek v3 now adopt vanilla chunked prefill on MLA part which is
ineffcient for computing but necessary for chunked prefill. Since PR
https://github.com/vllm-project/vllm-ascend/pull/543 bring v0 scheduler
into vllm-ascend, we can now adopt torch_npu._npu_flash_attention inside
the mla backend for more performance boost. Also there are some
redundant computation inside the rope, which is also removed. This PR
should bring some performance gain for deepseek eager mode inference.
---------
Signed-off-by: ganyi <pleaplusone.gy@gmail.com>
Replace torch function with reshape_and_cache fused op for better
performance. The `reshape_and_cache` function wasn't working because it
expected torch.int32 tensor, but a torch.int64 tensor was provided.
Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>
### What this PR does / why we need it?
The pr will fix some bug about spec decode / MTP
The pr add a mtp e2e UT `test_mtp_correctness.py`
**vllm_ascend/attention/attention.py**
1. add support `self.attn_mask_cache` only has 1 element to cover scene
in which both spec docode and chunked prefill are enabled.
**vllm_ascend/distributed/parallel_state.py**
1. remove 2 assert because spec decode worker would use init_worker
twice
**vllm_ascend/models/deepseek_mtp.py**
1. remove unused params;
2. add support w8a8 in `CustomDeepSeekMTP`
**vllm_ascend/quantization/quant_config.py**
1. use `AscendUnquantizedFusedMoEMethod` instead of
`UnquantizedFusedMoEMethod`
**other**
1. replace `from vllm.logger import init_logger` to `from vllm.logger
import logger` all of the vllm-ascend project
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
Signed-off-by: mengwei805 <mengwei25@huawei.com>
