### What this PR does / why we need it?
cherry-pick #1675 to main
This PR adds validation checking to torchair_graph_config for better
reliability.
Co-authored-by: whx-sjtu <2952154980@qq.com>
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.10.0
- vLLM main:
2836dd73f1
Signed-off-by: 22dimensions <waitingwind@foxmail.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>
### What this PR does / why we need it?
Cherry pick #1705 from v0.9.1-dev
Compared qwen2_5_vl.py, qwen2_5_vl_without_padding.py missing some
funtions. The purpose of this PR is to supplement these.
add:
- rot_pos_emb(self, grid_thw: torch.Tensor)
- get_window_index(self, grid_thw)
- _process_image_input(self, image_input)
- _process_video_input(self, video_input)
Co-authored-by: zheliuyu
[15750543867@163.com](mailto:15750543867@163.com)
Co-authored-by: wangli
[wangli858794774@gmail.com](mailto:wangli858794774@gmail.com)
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.0
- vLLM main:
207b750e19
Signed-off-by: wangli <wangli858794774@gmail.com>
bugfix cherry-pick from v0.9.1-dev
https://github.com/vllm-project/vllm-ascend/pull/2007
### What this PR does / why we need it?
Minimum reproducing code：
```python
# test.py
from vllm import LLM, SamplingParams
prompts = [
"Hello, my name is",
"The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
llm = LLM(model="Qwen2.5-VL-7B-Instruct", max_model_len=26240)
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}")
```
```bash
export USE_OPTIMIZED_MODEL=0
python test.py
```
exception as follow:
```
[rank0]: File "/home/xxx/vllm_ascend/models/qwen2_5_vl_without_padding.py", line 84, in forward
[rank0]: q = torch_npu.npu_rotary_mul(q, cos, sin)
[rank0]: File "/home/anaconda3/envs/xxx/lib/python3.10/site-packages/torch/_ops.py", line 1116, in __call__
[rank0]: return self._op(*args, **(kwargs or {}))
[rank0]: RuntimeError: Expected all tensors to be on the same device, but found at least two devices, npu:0 and cpu! (when checking argument for argument r1 in method wrapper__npu_rotary_mul)
```
In `AscendQwen2_5_VisionAttention_Without_Padding`,
`torch_npu.npu_rotary_mul(q, cos, sin)`， `cos`/`sin` on cpu, but `q` on
npu, so there will be an error.
`qwen2_5_vl_without_padding.py` need this bugfix, because
`AscendQwen2_5_VisionTransformer_Without_Padding.rot_pos_emb` in
wen2_5_vl_without_padding.py is from vllm and `inv_freq` will create on
cpu.
40d86ee412/vllm/model_executor/models/qwen2_5_vl.py (L482)
```python
inv_freq = 1.0 / (theta**(torch.arange(0, dim, 2, dtype=torch.float, device='cpu') / dim))
```
`qwen2_5_vl.py` do not need, because
`AscendQwen2_5_VisionRotaryEmbedding` in qwen2_5_vl.py rewrite
`AscendQwen2_5_VisionRotaryEmbedding` and `inv_freq` will create on
device.
```python
inv_freq = 1.0 / (theta**(torch.arange(0, dim, 2, dtype=torch.float) / dim))
```
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.10.0
- vLLM main:
18cc33dd60
Signed-off-by: pjgao <gaopengju3@huawei.com>
Co-authored-by: pjgao <gaopengju3@huawei.com>
This PR designs the shared expert multi-stream parallelism of
w8a8-dynamic-quantized MoE stage in more detail to achieve better
performance.
- vLLM version: v0.10.0
- vLLM main:
2cc571199b
Signed-off-by: whx-sjtu <2952154980@qq.com>
Support the inference of the Deepseekr1-w8a8-mtp model with
statically-quantized shared_head in MTP layers.
- vLLM version: v0.9.2
- vLLM main:
6eca337ce0
Signed-off-by: curryliu <120010041@link.cuhk.edu.cn>
### 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>
Signed-off-by: wuzhongjian <wuzhongjian_yewu@cmss.chinamobile.com>
### What this PR does / why we need it?
Fix duplicate 'torch.' prefix in qwen2-vl, qwen2.5-vl
- vLLM version: v0.9.2
- vLLM main:
dde295a934
Signed-off-by: wuzhongjian <wuzhongjian_yewu@cmss.chinamobile.com>
### What this PR does / why we need it?
Add prefix parameter to parent class initialization to avoid parameter
naming conflicts
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: v0.9.2
- vLLM main:
32142b3c62
### What this PR does / why we need it?
Optimizes the performance of the Qwen3 quantization model by registering
a custom model and adding the AddRmsNormQuant operation. Subsequent PRs
will focus on performance optimizations based on this custom model.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with existing test.
- vLLM version: v0.9.2
- vLLM main:
8d0a01a5f2
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
Remove ETP/EP maintained in branch main. We drop this as there is no
relevant scenarios to use ETP now, and we may subsequently advocate
implementing expert tensor parallelism in vLLM to support scenarios
where the expert is needed to be sliced
This is a part of #1422 backport.
Fixes https://github.com/vllm-project/vllm-ascend/issues/1396https://github.com/vllm-project/vllm-ascend/issues/1154
### Does this PR introduce _any_ user-facing change?
We'll not maintain etp/ep in vllm-ascend anymore, and use the tp/ep in
vllm instead.
### How was this patch tested?
CI passed with new added and existing test.
- vLLM version: v0.9.2
- vLLM main:
fe8a2c544a
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
Before patch, we can see
`vllm_ascend.models.deepseek_v2:CustomDeepseekV3ForCausalLM`, it seems
not friendly format.
```
WARNING 07-14 23:57:34 [registry.py:413] Model architecture DeepseekV2ForCausalLM is already registered, and will be overwritten by the new model class vllm_ascend.models.deepseek_v2:CustomDeepseekV2ForCausalLM.
WARNING 07-14 23:57:34 [registry.py:413] Model architecture DeepseekV3ForCausalLM is already registered, and will be overwritten by the new model class vllm_ascend.models.deepseek_v2:CustomDeepseekV3ForCausalLM.
WARNING 07-14 23:57:34 [registry.py:413] Model architecture Qwen3MoeForCausalLM is already registered, and will be overwritten by the new model class vllm_ascend.models.qwen3_moe:CustomQwen3MoeForCausalLM.
```
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Local Test.
- vLLM version: v0.9.2
- vLLM main:
bcdfb2a330
Signed-off-by: xleoken <xleoken@163.com>
### What this PR does / why we need it?
The previous code is
router_logits, _ = self.gate(hidden_states)
hidden_states = get_dp_group().all_gather(hidden_states, 0)
router_logits = get_dp_group().all_gather(router_logits, 0)
I want to change the two all_gathers to one, reduce one all_gather
communication, and make it
hidden_states = get_dp_group().all_gather(hidden_states, 0)
router_logits, _ = self.gate(hidden_states)
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
bash examples/run_dp_attention_etp16.sh
bash examples/run_dp_attention_etp16_benmark.sh
gsm8k accuracy verification
<img width="1809" alt="截屏2025-06-24 21 53 24"
src="https://github.com/user-attachments/assets/47eace3b-a86b-41b4-9de8-773f57fea33b"
/>
- vLLM version: v0.9.2
- vLLM main:
77f77a951e
---------
Signed-off-by: ttanzhiqiang <389825161@qq.com>
### What this PR does / why we need it?
Follow vllm-project/vllm lint way:
https://github.com/vllm-project/vllm/blob/main/.pre-commit-config.yaml
Enable pre-commit to avoid some low level error AMAP.
This pr is one step of #1241, The purpose is make linting system more
clear and convenient, on this step, Mainly did the following things:
yapf, actionlint, ruff, typos, isort, mypy, png-lint, signoff-commit,
enforce-import-regex-instead-of-re.
TODO:
- clang-format(check for csrc with google style)
need clean code, disable for now
- pymarkdown
need clean code, disable for now
- shellcheck
need clean code, disable for now
### Does this PR introduce _any_ user-facing change?
Only developer UX change:
https://vllm-ascend--1256.org.readthedocs.build/en/1256/developer_guide/contributing.html#run-lint-locally
```
pip install -r requirements-lint.txt && pre-commit install
bash format.sh
```
### How was this patch tested?
CI passed with new added/existing test.
Co-authored-by: Yikun [yikunkero@gmail.com](mailto:yikunkero@gmail.com)
Co-authored-by: wangli
[wangli858794774@gmail.com](mailto:wangli858794774@gmail.com)
- vLLM version: v0.9.1
- vLLM main:
5358cce5ff
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
When all_reduce_merge is in progress, shared_experts does not do
all_reduce in mlp, but waits until shared_experts+router_experts are
completed before doing all_reduce
In prefill and decode, as long as shared_experts+router_experts are
all_reduce, there will be benefits.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
bash examples/run_dp_attention_etp16.sh
bash examples/run_dp_attention_etp16_benmark.sh
- vLLM version: v0.9.1
- vLLM main:
977180c912
---------
Signed-off-by: ttanzhiqiang <389825161@qq.com>
vllm has released 0.9.2. This PR drop 0.9.1 support.
- vLLM version: v0.9.1
- vLLM main:
b942c094e3
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
This PR supports torchair graph mode with non-mla backend on both 800IA2
and 300I Duo platforms. The main change is to add
`attention_v1_torchair.py` to support specific attention related
operations that are required by torchair.
### Does this PR introduce _any_ user-facing change?
Before this PR, vLLM-Ascend only allows deepseek to use torchair. Now we
can also use it with pangu. Besides, we add a support model list to
control which type of models that can use torchair.
### How was this patch tested?
We have test it with PanguProMoE on both 800IA2 and 300I Duo platforms,
and model generates answer normally.
---------
Signed-off-by: angazenn <zengyanjia@huawei.com>
Signed-off-by: tianyitang <tangtianyi4@huawei.com>
Co-authored-by: angazenn <zengyanjia@huawei.com>
Co-authored-by: tianyitang <tangtianyi4@huawei.com>
### What this PR does / why we need it?
support pangu moe w8a8c8
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed with new added test.
Signed-off-by: zhuyilin <809721801@qq.com>
### What this PR does / why we need it?
This PR introduces an expert rearrange algorithm for PanguProMoE model.
Different from the original grouped topk, it filters out the top experts
that are allocated more tokens. Therefore, we can load less experts when
calculating gmm.
We have test this algorithm for PanguProMoE-72B on 300I Duo platform and
800I A2 platform. On 300I Duo platform, we find that `num_voted_experts`
set to 5 achieves both good performance and accuracy. While on 800I A2,
we still set it to 8 to use original pangu grouped topk.
### 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?
In this PR, we support H2P communication optimization when running
PanguProMoE with dp_size > 1. H2P use `reduce_scatter` and `all_gather`
to replace `all_reduce` to improve performance:
original layer:
input_layernorm --> attn --> tp all_reduce --> post_attention_layernorm
--> dp all_gather --> moe/mlp --> dp reduce_scatter --> tp all_reduce
now:
input_layernorm --> tp all_gather --> attn --> tp reduce_scatter -->
post_attention_layernorm --> all_rank all_gather --> moe/mlp -->
all_rank reduce_scatter
Besides, because `reduce_scatter` requires num_tokens that can be
divided by group size, we need pad the seqs based on
`max_tokens_across_dp`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
This PR has been tested with both offline and online inference using
PanguProMoE-72B.
---------
Signed-off-by: angazenn <zengyanjia@huawei.com>
Co-authored-by: angazenn <zengyanjia@huawei.com>
### What this PR does / why we need it?
- Fix vLLM EPLB break
e9fd658a73
by recovering load_weights back to [v0.9.1
version](07b8fae219)
temporarily.
- Fix transformers>=4.53.0 image processor break
Related: https://github.com/vllm-project/vllm-ascend/issues/1470
- Mirror torch_npu requirements to pyproject.toml
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed
---------
Signed-off-by: MengqingCao <cmq0113@163.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?
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?
Support Pangu Pro MoE model (https://arxiv.org/abs/2505.21411)
### Does this PR introduce _any_ user-facing change?
Yes, new model supported
### How was this patch tested?
Test locally
---------
Signed-off-by: angazenn <zengyanjia@huawei.com>
Co-authored-by: angazenn <zengyanjia@huawei.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?
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>
Contains on #1111 for completeness.
<!-- 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?
Implement multi-stream parallelism for MoE layers with shared experts,
where computation of shared experts will be overlapped with expert token
dispatch and combine. Also, when multi-stream is enabled, weights of
shared experts will be force to replicate across all cards, regardless
of any tensor parallelism configurations, to avoid AllReduce operations.
With the expected overlaping being:
```
| shared gate_up | shared act | | shared down |
| dispatch | routed gate_up, act, down | combine |
```
<!--
- 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?
No.
<!--
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?
Tested on 1x16 910 node, with tailored 2 layer DSKv2.
<!--
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: sdmyzlp <lrwei2@petalmail.com>
Add unpadded Qwen2.5-VL for verl scenario.
When using vllm-ascend for verl scenario, set `USE_OPTIMIZED_QWEN2_5_VL`
(default `1`) to `0` to use unpadded Qwen2.5-VL to avoid errors.
This is cherry-picked from 0.7.3-dev
Signed-off-by: shen-shanshan <467638484@qq.com>
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Shanshan Shen <467638484@qq.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>
### What this PR does / why we need it?
Support MOE inner Multi-stream for Deepseek.
This feature requires graph mode with mc2 enabled.
---------
Signed-off-by: David9857 <985700846@qq.com>
### What this PR does / why we need it?
Optimize the performance of calculation logic in sampler and deepseekv2.
### Does this PR introduce _any_ user-facing change?
Added VLLM_ENABLE_TOPK_OPTIMZE config in sampler
### How was this patch tested?
pytest test_sampler.py
Signed-off-by: wangxiaoxin (A) <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin (A) <wangxiaoxin7@huawei.com>
Co-authored-by: ZhengWG <zwg0606@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?
the interface of qwen2.5vl changes from column linear to qkv linear,
this makes our weight pad func become abnormal, thus we optimize
split_qkv func to fix this bug.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
with CI
Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
### What this PR does / why we need it?
1. In previous PRs https://github.com/vllm-project/vllm-ascend/pull/580https://github.com/vllm-project/vllm-ascend/pull/784, I saved GPU memory
by promptly deleting unnecessary tensors. For tensors passed from
upper-layer functions, I used a list container to transfer the parameter
and then popped the tensor from the list within the inner function to
achieve deletion. Recently, I discovered a better implementation in
sglang—the `dispose_tensor` function and I recommend adopting this
approach.
2. Dispose `hidden_states` and `residual` from the previous layer once
they're no longer used.
3. Avoid to generate `self.inputs_embeds` in `ModelRunnerV1` in
non-multimodal scenarios.
With the aforementioned optimizations, using the DeepSeek-R1-W8A8 model
under the conditions of `TP=16` and `max-model-len=32768`, we can save
1.3GB of npu memory.
**Reference**: https://github.com/sgl-project/sglang/pull/6147
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
---------
Signed-off-by: ApsarasX <apsarax@outlook.com>
Tweak packed_modules_mapping to support W8A8 weights.
<!-- 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 #
-->
### 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.
-->
### 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: Yizhou Liu <liu_yizhou@outlook.com>
The [vllm
commit](67da5720d4)
changed the input and rotary position embedding for qwen 2.5 vl which
break CI. This PR fix the CI failure for qwen2.5 vl in quick
Signed-off-by: wangxiyuan <wangxiyuan1007@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?
1. This PR introduces native `all_to_all` communication operator to fix
`allgather` bugs when dp_size > 1. Besides, it adds a naive
implementation of force-load-balance when doing profile runs.
2. The operator `npu_dequant_swiglu_quant` only supports input
hidden_states with dtype `torch.int32`. This tensor occupies space of
`global_bs * seq_len * topk * hidden_size`, which might be very large as
`ep_size` grows. Therefore we need to disable this operator and use
original `swiglu` && `quantize`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By performing offline inference:
---------
Signed-off-by: angazenn <zengyanjia@huawei.com>
Co-authored-by: angazenn <zengyanjia@huawei.com>
### What this PR does / why we need it?
Fix the method of importing environment variables in DeepSeek model to
support successful compilation via aclgraph.
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
In the w8a8 quantization code of `fused_experts`, the output of almost
every operator is assigned a new variable name. If we want to save NPU
memory, we manually `del` these variables to end their lifecycle, which
fills the code with `del` statements and looks inelegant.
Therefore, I plan to names the output of most operators as
`hidden_states`, thereby ending the lifecycle of the previous
`hidden_states`.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Signed-off-by: ApsarasX <apsarax@outlook.com>
-->
### 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 #
-->
1. Improve inference speed and usability for deepsek models with NPU
graph mode.
2. Modify some codes to adapt to CANN 8.1.RC1.beta1.
3. Add a switch for NPU graph mode and its cache.
### 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.
-->
This PR provides an experimental configuration to enable NPU graph mode
for Deepseek models. User can set
additional_config={'enable_graph_mode': True} to try this feature. Note
that this feature currently only supports for V0 engine.
### 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.
-->
This patch was tested with the newest torch_npu 2.5.1
(https://pypi.org/project/torch-npu/#files) and CANN 8.1.RC1.beta1
toolkit&nnal&kernels
(https://www.hiascend.com/developer/download/community/result?module=cann)
released in 25/30 April.
Signed-off-by: linfeng-yuan <1102311262@qq.com>
