### What this PR does / why we need it?
- This PR proposes a P2P version of Disaggregated Prefill based on
llm_datadist which manages data transfer.
- This solution reconstructs previous offline single-node Disaggregated
Prefill solution, and supports multi-node and online serveing now.
- Currently this solution supports 1P1D situation of Deepseek hybrid
parallelism (P: TP+EP, D: DP+EP). Note that xPyD situation is considered
in the solution design, and will be supported soon within v1 engine.
---------
Signed-off-by: hw_whx <wanghexiang7@huawei.com>
Signed-off-by: ganyi <pleaplusone.gy@gmail.com>
Co-authored-by: hw_whx <wanghexiang7@huawei.com>
Co-authored-by: ganyi <pleaplusone.gy@gmail.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>
<!-- 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?
As custom deepseek modeling do some changes to support graph mode in
https://github.com/vllm-project/vllm-ascend/pull/585, so i follow it to
change custom deepseek_mtp modeling.
And some modifications for k>1 were not carried over by the
https://github.com/vllm-project/vllm-ascend/pull/429, now i add it.
In order to better take care of the MTP feature in the vllm-ascend
repository, I added cases related to graph mode(torchair), but i skip it
since torchair can not correctly clean up memory in vllmrunner.
Also i add some case for MTP quantization weights, but test weight is
not ready, so i skip it and i will open it when test quant weights is
ready.
https://github.com/vllm-project/vllm-ascend/pull/648 did not completely
fix the sample
change(https://github.com/vllm-project/vllm-ascend/issues/660) issue, I
added the relevant changes.
### Does this PR introduce _any_ user-facing change?
now, u can use following method to use mtp in deepseek v3/r1 float or
quant weights with eager mode.
```python
llm = LLM(
model="wemaster/deepseek_mtp_main_random_bf16",
tensor_parallel_size=2,
speculative_config={
"num_speculative_tokens": 1,
},
enforce_eager=True,
trust_remote_code=True,
disable_log_stats=False,
gpu_memory_utilization=0.8,
max_model_len=64,
)
```
or use mtp in deepseek v3/r1 float or quant weights with graph
mode（torchair）
```python
llm = LLM(
model="wemaster/deepseek_mtp_main_random_bf16",
tensor_parallel_size=2,
speculative_config={
"num_speculative_tokens": 1,
},
trust_remote_code=True,
additional_config={
'enable_graph_mode': True,
},
disable_log_stats=False,
gpu_memory_utilization=0.8,
max_model_len=64,
)
```
add notes:
1. now, we support k>1, so u can set num_speculative_tokens > 1 if there
is sufficient redundant computing power;
2. MTP is not supported in V1, we will support it when vLLM does it in
https://github.com/vllm-project/vllm/issues/13500.
3. if u run MTP failed by `segmentation fault`, u can follow v0.7.3
patch https://github.com/vllm-project/vllm-ascend/pull/236 file
`vllm_ascend/patch/patch_metrics.py` method
`__npu_async_metrics_collector_init__`
### How was this patch tested?
local tested passed and test by CI
Signed-off-by: mengwei805 <mengwei25@huawei.com>
### What this PR does / why we need it?
Enforce eager mode in the V1 engine ahead of the upcoming CANN and
torch_npu releases.
### Does this PR introduce _any_ user-facing change?
After this change, users will no longer need to manually set
enforce_eager=True.
### How was this patch tested?
Test it with regular offline inference examples.
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
Part of #499
Add qwen2.5-vl test on single npu, v1 engine is excluded because
qwen2.5-vl has some problems with v1 now, at the same time, this test
can also make #639 more credible
Signed-off-by: wangli <wangli858794774@gmail.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?
<!--
- 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 #
-->
This PR supports the access of vllm-acend to the piecewise_graph feature
provided by the v1 engine.
1. register unifiled_ascend_attention_with_output for piecewise_graph to
split graph.
2. support NPUGraph to accelerate kernel launch.
### Does this PR introduce _any_ user-facing change?
<!--
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.
-->
support npugraph to default， Users can disenable the npugraph feature by
configuring enforce_eager.
This has corresponding requirements for the versions of torch_npu and
CANN, and they need to support graph capture.
### 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.
-->
it turn to default
---------
Signed-off-by: Bug Hunter Yan <yanpq@zju.edu.cn>
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
After extensive testing, we are happy to say that guided_decoding is
fully supported by npu, in this pr, we add guided_decoding integrated
with our test, mainly does the following things:
1. test v0 supported backends including ` "outlines",
"lm-format-enforcer","xgrammar"`
2. test v1 supported backends including ` "guidance", "xgrammar"`
---------
Signed-off-by: wangli <wangli858794774@gmail.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>
### What this PR does / why we need it?
Add a `VLLMAscendQuantizer` to support w8a8 static (W8A8) and dynamic on
linear and moe (W8A8_DYNAMIC), the quantizer will be enable if a model
has [quantize
filed](https://huggingface.co/vllm-ascend/Qwen2.5-0.5B-Instruct-w8a8/blob/main/config.json#L27).
If MindIE Turbo is installed, the MindIE Turbo Quantizer will apply,
otherwise will use VLLMAscendQuantizer directly.
- This patch fix installation docs to make installation work
- This patch enable norm quantization by patch `RMSNorm.__init__`,
`RMSNorm.forward_oot`, `NPUModelRunnerBase.load_model`
- Add `AscendW8A8LinearMethod` for W8A8
- Add `AscendW8A8DynamicLinearMethod` and
`AscendW8A8DynamicFusedMoEMethod` for W8A8_DYNAMIC
- Add a e2e test for `vllm-ascend/Qwen2.5-0.5B-Instruct-w8a8`
### Does this PR introduce _any_ user-facing change?
Yes, support w8a8 quantization. After this patch supported, users can
use below commands to run w8a8 models:
```
vllm serve /root/.cache/modelscope/hub/Qwen/Qwen2.5-7B-Instruct-w8a8 --served-model-name "qwen2.5-7B"
```
### How was this patch tested?
0. CI passed: add e2e test for `vllm-ascend/Qwen2.5-0.5B-Instruct-w8a8`
1. From @Yikun:
I test Qwen2.5-0.5B-Instruct-w8a8 for functional test all is well, pls
refer to
https://github.com/vllm-project/vllm-ascend/pull/580#issuecomment-2816747613
2. From @dingdingchaomian :
Use qwen2.5-72b-instruct model and deepseek-v2-lite-chat tested, both
models were quantized using Ascend's msmodelslim tool:
- Qwen2.5-72b-instruct were tested twice, one for w8a8 static and one
for w8a8 dynamic.
- Deepseek-v2-lite-chat were tested once because its quantization used
both static and dynamic w8a8.
Models were tested using both off line inference and online serving, and
both work well. The inference codes are exactly the same with the
examples in
https://vllm-ascend.readthedocs.io/en/latest/quick_start.html, with
model path and tensor parallel number changed.
---------
Signed-off-by: dingdingchaomian <wangce21@huawei.com>
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
Co-authored-by: dingdingchaomian <wangce21@huawei.com>
Co-authored-by: Angazenn <zengyanjia@huawei.com>
Co-authored-by: liujiaxu <liujiaxu4@huawei.com>
Co-authored-by: ApsarasX <apsarax@outlook.com>
Co-authored-by: ganyi1996ppo <pleaplusone.gy@gmail.com>
### What this PR does / why we need it?
This PR adds sleep mode feature for vllm-ascend, when sleeps, we do
mainly two things:
- offload model weights
- discard kv cache
RLHF tools(such as https://github.com/volcengine/verl and
https://github.com/OpenRLHF/OpenRLHF) have a strong need of sleep mode
to accelerate the training process.
This PR may solve #375 and #320 .
### Does this PR introduce _any_ user-facing change?
No existing user interfaces changed.
Users will have two new methods(`sleep()` and `wake_up()`) to use.
### How was this patch tested?
This PR is tested with Qwen/Qwen2.5-0.5B-Instruct.
At first, we have free NPU memory M1.
After `llm = LLM("Qwen/Qwen2.5-0.5B-Instruct", enable_sleep_mode=True)`
executed, we have free NPU memory M2. M2 < M1.
Then we call `llm.sleep(level=1)`, we have free NPU memory M3.
We have M3 > M2, M3 is very close to M1.
Plus, we have the same output tokens before sleep and after wake up,
with the config of `SamplingParams(temperature=0, max_tokens=10)` and
with the same input tokens of course.
This PR is utilizing the CMake procedure of #371 , thanks a lot.
Signed-off-by: Shuqiao Li <celestialli@outlook.com>
### What this PR does / why we need it?
Adopt custom kernel rotary embedding in actual model inference,
customized rotary_embedding will generate contiguous query and key in
the cpp side to reduce the overhead of two contiguous and index_select
compared with rotary_embedding in torch_npu. For now, rotary_embedding
can only support the scenario of `is_neox = true`, non-neox version rope
will be updated soon in the future.
---------
Signed-off-by: ganyi <pleaplusone.gy@gmail.com>
This PR Fixes scheduler problems in last PR:
1. change position of DT test to validate it.
2. fix format of copyright.
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
Backport: https://github.com/vllm-project/vllm-ascend/pull/252
This support speculative decoding in Ascend, including speculating with
a draft model、by matching n-grams in the prompt、using MLP speculators
and using EAGLE based draft models.
Backport: https://github.com/vllm-project/vllm-ascend/pull/423
spec decode MultiStepWorker support TP1DraftModelRunner fully, support
run the draft_model_runner with multi-step prepare on the NPU directly
and support draft_model_runner use MLA.
1. before this pr, `MultiStepWorker` would not step into the branch
using NPU prepare, but only into the branch using CPU prepare (`line 52`
of `vllm_ascend/patch/patch_multi_step_worker.py`). Although this has
`no effect` on the `correct operation` of speculative decoding and the
performance of the two branches is basically the same as of the current
version, I support entering this branch in this PR. In general, there
are two main changes in `patch_multi_step_worker.py`: first, the
`is_cuda_like()` check is removed and the `TP1DraftModelRunner`
rewritten in vllm_ascend is used; second, the
`supports_gpu_multi_step()` function is made to return true on NPU
devices when outer Multi_step_worker could work correct.
3. before this pr, `TP1DraftModelRunner` only supports Attention on NPU,
but not MLA. The relevant adaptation is in
`vllm_ascend/worker/draft_model_runner.py`. Although I don’t know why
the `input_positions` of `model_input.attn_metadata` in vllm-ascend
needs to be added in `execute_model`, it is done in `model_runner.py`,
so I also made corresponding changes. Otherwise, when atten_backend is
MLA, it will prompt that input_positions cannot be found.
4. I commented out two lines in `draft_model_runner.py` in `line118` to
support the scenario of K>1.
```
# lora_mapping=model_input.lora_mapping,
# lora_requests=model_input.lora_requests,
```
I added comments. In the future, when vllm-ascend supports lora feature,
the changes here can be restored.
TODO：
- [ ] revert the patch when the related issues are addressed in vllm
### How was this patch tested?
CI passed with new added test.
- e2e test for medusa proposer:
tests/singlecard/spec_decode/e2e/test_medusa_correctness.py
- e2e test for mlp proposer:
tests/singlecard/spec_decode/e2e/test_mlp_correctness.py
- e2e test for n-gram proposer:
tests/singlecard/spec_decode/e2e/test_ngram_correctness.py
Tests for patched files:
- tests/singlecard/spec_decode/test_dynamic_spec_decode.py
- tests/singlecard/spec_decode/test_multi_step_worker.py
- tests/singlecard/spec_decode/test_ngram_worker.py
- tests/singlecard/spec_decode/test_spec_decode_worker.py
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
Co-authored-by: mengwei805 <mengwei25@huawei.com>
This PR adds AscendScheduler to vllm v1 engine.
This scheduler currently supports v0-style prefill-first scheduling
strategy.
In the future more schedule methods will be supported by this scheduler.
---------
Signed-off-by: hw_whx <wanghexiang7@huawei.com>
Co-authored-by: hw_whx <wanghexiang7@huawei.com>
### What this PR does / why we need it?
This PR enable custom ops build by default.
### Does this PR introduce _any_ user-facing change?
Yes, users now install vllm-ascend from source will trigger custom ops
build step.
### How was this patch tested?
By image build and e2e CI
---------
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
- Add a new runner to the continuous integration system and keep the
original CI runner until the new runner runs stably
- Add distributed test cases
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
Enable Expert-Parallel for ascend devices.
### Does this PR introduce _any_ user-facing change?
Enable EP
add `enable_expert_parallel=True` in your offline inference scripts,
like this:
```python
llm = LLM(
model="/path/to/model",
trust_remote_code=True,
tensor_parallel_size=4,
max_model_len=4096,
enforce_eager=True,
distributed_executor_backend="mp",
enable_expert_parallel=True,
)
```
### How was this patch tested?
Please use the `main` branch of vLLM.
---------
Signed-off-by: Yizhou Liu <liuyizhou5@h-partners.com>
Co-authored-by: Yizhou Liu <liuyizhou5@h-partners.com>
### What this PR does / why we need it?
Add dispatch job to leverage jobs to dynamic devices include 2 stage as
below:
The dispatch job will spend extra about `10s * parallel number + 30s`
time to wait other job launch container and release lock.
- **Stage 1: Acquire lock**
add a dispatch job, this job use lockfile to acquire locks and then get
device number dynamically
- **Stage 2.1: Launch container with dynamic device**
pass the device number via output and start the container job with
dynamic device
- **Stage 2.2: Release lock**
once the job started, release the lock.
In the backend, we use multiple path to setup multiple self host runners
as load balancer:
```
$ pwd
/home/action
$ ll | grep actions
drwx------ 6 action action 4096 Mar 7 08:55 actions-runner-01
drwx------ 6 action action 4096 Mar 7 08:55 actions-runner-02
drwx------ 6 action action 4096 Mar 7 08:55 actions-runner-03
drwx------ 6 action action 4096 Mar 7 08:56 actions-runner-04
drwx------ 4 action action 4096 Jan 24 22:08 actions-runner-05
drwx------ 4 action action 4096 Jan 24 22:08 actions-runner-06
```
```
adduser -G docker action
su action
pip3 install docker prettytable
sudo yum install procmail
```
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
- CI passed
- E2E test manully, triggered 3 jobs in parallel:
- [1st
job](https://github.com/vllm-project/vllm-ascend/actions/runs/13711345757/job/38348309297)
dispatch to /dev/davinci2.
- [2nd
job](https://github.com/vllm-project/vllm-ascend/actions/runs/13711348739/job/38348316250)
dispatch to /dev/davinci3
- [3rd
job](https://github.com/vllm-project/vllm-ascend/actions/runs/13711351493/job/38348324551)
dispatch to /dev/davinci4
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
### What this PR does / why we need it?
vLLM Ascend plugin (vllm-ascend) is a backend plugin for running vLLM on
the Ascend NPU.
This plugin is the recommended approach for supporting the Ascend
backend within the vLLM community. It adheres to the principles outlined
in the [RFC]: Hardware pluggable, providing a hardware-pluggable
interface that decouples the integration of the Ascend NPU with vLLM.
This patch also include changes to make CI work and use cache speed up
e2e test, including:
1. Change push (post merge ci) and pull_request (pr ci) trigger branch
to main
2. Make mypy work by ignore base_communicator and clear unused deps
3. Several improvements for vllm_ascend_test:
- use cache (pip, ms, hf) speed up e2e test (25mins --> 5mins)
- switch `git clone` command to `action/checkout` to speedup checkout
and
- Enable sv for pytest for better info dump
- Remove network host to resole `docker: conflicting ontions: cannot
attach both user-defined and non-user-definednetwork-modes`, which is a
problem on docker 1.45 but not on 1.39.
4. Adapt MLA decode optimizations:
cabaf4eff3
### Does this PR introduce _any_ user-facing change?
Yes, init the PR.
### How was this patch tested?
- This is the first PR to make ascend NPU work on vLLM. All code is
tested on ascend with vLLM V0 Engine.
- CI passed
---------
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: MengqingCao <cmq0113@163.com>
Co-authored-by: wangshuai09 <391746016@qq.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>
Co-authored-by: wangli <wangli858794774@gmail.com>
