Commit Graph
43 Commits
Author SHA1 Message Date
Mengqing Cao 986cd45397 [Version] Drop 0.16.0 support (#7153)
### What this PR does / why we need it?
Drop 0.16.0 support in main
- Fix eagle proposer break introduced by
https://github.com/vllm-project/vllm/pull/34552. Mainly change to use
the draft attention group to initialize the attention metadata builder.
- Fix the `ModelRunner` has no attribute `cudagraph_capture_sizes`
error, which is a bug in vLLM v0.17.0, and fixed by a later pr
https://github.com/vllm-project/vllm/pull/30515

- vLLM version: v0.16.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/4034c3d32e30d01639459edd3ab486f56993876d
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
2026-03-13 16:14:15 +08:00
kxand01267596 df1ee8070d [feat][spec decode]Unified draft parallel (#6766)
### What this PR does / why we need it?
Implement a unified parallelized speculative decoding in VLLM
Ascend,which can simultaneously support parallel speculative inference
schemes such as Pard, P-Eagle, etc. refer to
https://github.com/vllm-project/vllm-ascend/pull/6565 and
https://github.com/vllm-project/vllm-ascend/pull/4078

### How was this patch tested?

run with parallel drafting script:
export target=/model/Llama-3.1-8B-Instruct
export draft=/model/PARD-Llama-3.2-1B
export CUDA_VISIBLE_DEVICES=6
export ASCEND_RT_VISIBLE_DEVICES=6
vllm serve $target \
  --tensor-parallel-size 1 \
  --max-model-len 4096 \
  --no-enable-prefix-caching \
  --port 8811 \
--speculative-config '{"model": "/model/PARD-Llama-3.2-1B", "method":
"draft_model", "num_speculative_tokens": 8, "parallel_drafting": true}'

base script:
export target=/model/Llama-3.1-8B-Instruct
export draft=/model/PARD-Llama-3.2-1B
export CUDA_VISIBLE_DEVICES=6
export ASCEND_RT_VISIBLE_DEVICES=6
vllm serve $target \
  --tensor-parallel-size 1 \
  --max-model-len 4096 \
  --no-enable-prefix-caching \
  --port 8811

benchmark script:
MAX_CONCURRENCY=1
NUM_PROMPTS=80
vllm bench serve --port 8811 \
    --temperature 0 \
    --model /model/Llama-3.1-8B-Instruct \
    --backend openai-chat \
    --endpoint /v1/chat/completions \
    --dataset-name hf \
    --dataset-path philschmid/mt-bench \
    --num-prompts ${NUM_PROMPTS} \
    --max-concurrency ${MAX_CONCURRENCY} \
    --seed 1234

test results :
base(without spec decode): TTFT 79.46ms TPOT 26.99ms
output_tokens_throughput 36.75 tok/s
this pr(with parallel drafting): TTFT 72.24ms TPOT 13.45ms
output_tokens_throughput 72.98 tok/s
per-position acceptance(from position 0 to 7):
79.48%、56.93%、40%、27.90%、19.79%、14.25%、10.57%、7.61%.

----------------------------------------------------------------------
run on qwen3 model script :
export target=/model/Qwen3-1.7B
export draft=/model/PARD-Qwen3-0.6B
export CUDA_VISIBLE_DEVICES=1
export ASCEND_RT_VISIBLE_DEVICES=1

vllm serve $target \
  --tensor-parallel-size 1 \
  --max-model-len 4096 \
  --no-enable-prefix-caching \
  --port 8811 \
--speculative-config '{"model": "/model/PARD-Qwen3-0.6B", "method":
"draft_model", "num_speculative_tokens": 8, "parallel_drafting": true}'

cc  @NickJudyHvv
- vLLM version: v0.15.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/9562912cead1f11e8540fb91306c5cbda66f0007

---------

Signed-off-by: 01267596 <xiongkai123@cmbchina.com>
Signed-off-by: kx <1670186653@qq.com>
Signed-off-by: HF-001 <1670186653@qq.com>
Co-authored-by: 01267596 <xiongkai123@cmbchina.com>
2026-03-13 14:07:35 +08:00
Ronald c980e68d40 [Feature] support aclgraph for model runner v2 (#7110)
### What this PR does / why we need it?
This PR aims to support aclgraph for model runner v2, please see RFC
#5208. The PR contains these modifications:
- adapt to newest commit of vllm main branch.
- supply a unified interface of extra forward context for both model
runner v1 and model runner v2.
- implement graph mode for main model. 

### Does this PR introduce _any_ user-facing change?
no

### How was this patch tested?

- vLLM version: v0.16.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/4034c3d32e30d01639459edd3ab486f56993876d

---------

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
2026-03-13 09:11:46 +08:00
Zetong Li a2696006d1 [Refactor][EAGLE] 8/N delete mtp_proposer (re-pull) (#7033)
### What this PR does / why we need it?
**NOTE: This PR is re-pull of #7016 since ci mistakenly marked
unfinished pr as having passed.**

This PR aims to delete mtp_proposer. By fixing a bug in both dsv32 and
glm5, now it should be ok to remove mtp_proposer. The bug is actually
about unnecessary slicing of `slot_mapping`.

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
by ci

- vLLM version: v0.16.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/4034c3d32e30d01639459edd3ab486f56993876d

---------

Signed-off-by: Zetong Li <slippersss@126.com>
2026-03-06 17:11:22 +08:00
wangxiyuan 16c3b0b822 Revert "[Refactor][EAGLE] 8/N delete mtp_proposer" (#7030)
Reverts vllm-project/vllm-ascend#7016
It breaks E2E test
- vLLM version: v0.16.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/4034c3d32e30d01639459edd3ab486f56993876d
2026-03-06 11:24:05 +08:00
Zetong Li a60e179c7f [Refactor][EAGLE] 8/N delete mtp_proposer (#7016)
### What this PR does / why we need it?
This PR aims to delete mtp_proposer. By fixing a bug in both dsv32 and
glm5, now it should be ok to remove mtp_proposer. The bug is actually
about unnecessary slicing of `slot_mapping`.

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
by ci

- vLLM version: v0.16.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/15d76f74e2fdb12a95ea00f0ca283acf6219a2b7

---------

Signed-off-by: Zetong Li <slippersss@126.com>
2026-03-06 09:10:57 +08:00
wangxiyuan 13777bf3f0 [Spec Decode]clean up spec decode interface (#6947)
This pull request refactors the speculative decoding proposer interface
to align with upstream vLLM, removing the local `Proposer` interface and
renaming methods to `propose`.

This is the first step. In the future we should remove the class
register and just add few Ascend specified method once the arch in vLLM
is ready.

- vLLM version: v0.16.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/15d76f74e2fdb12a95ea00f0ca283acf6219a2b7

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
2026-03-05 14:30:10 +08:00
realliujiaxu 5def28dcd3 [Feat]support sequence parallelism by pass for VL models (#5632) 2026-02-27 08:27:41 +08:00
wangxiyuanandwxsIcey 2a826b5fad [Misc] upgrade to vllm main (#6646)
### What this PR does / why we need it?
This PR upgrades the core vLLM dependency to a newer version from the
main branch (`13397841ab469cecf1ed425c3f52a9ffc38139b5`). This is
necessary to keep our project up-to-date with the latest features and
fixes from upstream vLLM.

1.
https://github.com/vllm-project/vllm/commit/ac32e66cf95c40502ad7b00b2e80dfb0315bfee4
pass file is moved.

- vLLM version: v0.15.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/d7e17aaacd5ed1b4b4be6bcfef3a1b7cbc84fc9a

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: wxsIcey <1790571317@qq.com>
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
Co-authored-by: wxsIcey <1790571317@qq.com>
2026-02-10 14:08:59 +08:00
922e5c163b [main2main] upgrade vllm main 0202 (#6560)
### What this PR does / why we need it?
1. Fix `TypeError: FusedMoEParallelConfig.__init__() missing 1 required
positional argument: 'is_sequence_parallel'` due to
https://github.com/vllm-project/vllm/pull/32567
2. Fix ` TypeError: '>' not supported between instances of 'MagicMock'
and 'int'` due to https://github.com/vllm-project/vllm/pull/33035
3. Fix `TypeError: Can't instantiate abstract class AscendMLAImpl with
abstract methods forward_mha, forward_mqa` and AttributeError: 'bool'
object has no attribute 'process_weights_after_loading' due to
https://github.com/vllm-project/vllm/pull/33284
4. Fix `'AscendSharedFusedMoE' object has no attribute
'_routed_input_transform'`due to
https://github.com/vllm-project/vllm/pull/32790
5. Fix `NPUModelRunner._dummy_run() got an unexpected keyword argument
'num_active_loras'` due to
https://github.com/vllm-project/vllm/pull/32005
6. Fix the problem caused by` 'tuple' object has no attribute 'job_id'`
due to https://github.com/vllm-project/vllm/pull/27492
7. Fix the problem that all_moe_layers is not equal to vllm.moe_forward,
vllm.moe_forward_shared due to
https://github.com/vllm-project/vllm/pull/33184
8. Add patch to fix the problem "got multiple values for keyword
argument 'add_special_tokens'" due to
https://github.com/vllm-project/vllm/pull/32863
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.15.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.15.0

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: hfadzxy <starmoon_zhang@163.com>
2026-02-05 19:31:17 +08:00
meihancandwjunLu fea197ad50 [Main2Main] Upgrade vllm commit to 0123 (#6169)
### What this PR does / why we need it?
1. ✅ Upgrade vllm commit to: 0115
(8471b27df97c3eb79f891802fc0e858f8f7ac6a0)
Modify import paths due to the refactors:
https://github.com/vllm-project/vllm/pull/32245
https://github.com/vllm-project/vllm/pull/32060
Test result:
https://github.com/vllm-project/vllm-ascend/actions/runs/21034239336/job/60490156965?pr=5913
2. ✅Upgrade vllm commit to: 0119
(9a1f16da1e423ede2c2f52a9850cbfbb39cefe96)
Fix `WorkerProc.__init__() missing 1 required positional argument:
'is_driver_worker'` due to
https://github.com/vllm-project/vllm/pull/28506
Test result:
https://github.com/vllm-project/vllm-ascend/actions/runs/21156263050/job/60841668755?5569
3. ✅Upgrade vllm commit to:
0120(148117ea2e689cd43df4be6892671a17cdae5833)
1. Add `skip_compiled` param in `set_forward_context` due to
https://github.com/vllm-project/vllm/pull/30385
2. Modify `tests/ut/spec_decode/test_eagle_proposer.py` due to
https://github.com/vllm-project/vllm/pull/24322
change `self.max_num_tokens =
vllm_config.scheduler_config.max_num_batched_tokens + max_batch_size`
3. Modify UT import paths due to the
refactors:https://github.com/vllm-project/vllm/pull/32060
Test result:
https://github.com/vllm-project/vllm-ascend/actions/runs/21204851770/job/60999046946
4. ✅Upgrade vllm commit to:
0121(f23fb5a7c1b61350c5c40ca1115d3bf8cf2b8cc9)
1. vLLM switched `uses_mrope` from target to draft model config, making
`positions`/`mrope_positions` mutually exclusive, breaking vllm-ascend's
direct self.positions access and tests missing
`draft_model_config.uses_mrope`.
https://github.com/vllm-project/vllm/pull/32048
2. Moved bs_to_padded_graph_size from CompilationConfig to
CudagraphDispatcher due to the refactor
https://github.com/vllm-project/vllm/pull/30143
3. Remove unused `maybe_setup_kv_connector` due to
https://github.com/vllm-project/vllm/pull/32077
Test result:
https://github.com/vllm-project/vllm-ascend/actions/runs/21217728738/job/61043738834
6. ✅Upgrade vllm commit to:
0122(8ebf271bb6d1e7e9b1a55be73d755ef1a57dbbe5)
Updating FusedMoEParallelConfig (added enable_eplb) and FusedMoEConfig
due to https://github.com/vllm-project/vllm/pull/32414
Test result:
https://github.com/vllm-project/vllm-ascend/actions/runs/21249922546/job/61148613054
8. ✅Upgrade vllm commit to:
0123(dc917cceb877dfd13f98c538c4c96158047d98bd)
Setting temperature=0.0 due to the removal of the default temperature
value in https://github.com/vllm-project/vllm/pull/32723
Test result:
https://github.com/vllm-project/vllm-ascend/actions/runs/21280796875
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.14.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60

---------

Signed-off-by: wjunLu <wjunlu217@gmail.com>
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
Co-authored-by: wjunLu <wjunlu217@gmail.com>
2026-01-27 08:44:36 +08:00
wangxiyuan 4e3919e965 Reapply "[Refactor] Unify full-graph parameter update logic (#6041)" (#6227) (#6231)
This reverts commit 95649344aa.

The CI failure doesn't related to this change. Let's reapply it.

- vLLM version: v0.14.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60
2026-01-26 09:04:54 +08:00
wangxiyuan 95649344aa Revert "[Refactor] Unify full-graph parameter update logic (#6041)" (#6227)
This reverts commit 8966a99710.

It breaks the test
`tests/e2e/singlecard/spec_decode/test_mtp_eagle_correctness.py::test_deepseek_mtp_correctness[True-FULL_DECODE_ONLY-2-wemaster/deepseek_mtp_main_random_bf16]`

- vLLM version: v0.14.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60
2026-01-25 15:25:38 +08:00
8966a99710 [Refactor] Unify full-graph parameter update logic (#6041)
### What this PR does / why we need it?

**Refactor: Unify full-graph parameter update logic**

This PR consolidates the scattered full-graph parameter update logic
into a unified approach, improving code architecture and eliminating
duplication.

**Key improvements:**

1. **Unified interface**
- Create `update_full_graph_params` as the single entry point for all
full-graph updates
   - Replace multiple scattered update calls with one unified function
- Remove ~50 lines of duplicated if-else logic across
`model_runner_v1.py` and `eagle_proposer.py`

2. **Better architecture**
- Move update logic to respective Backend classes
(`AscendAttentionBackend`, `AscendMLABackend`)
   - Each Backend manages its own parameter update logic internally
   - Simplify caller code to just dispatch to the appropriate Backend

3. **Cleaner parameter handling**
   - Remove unnecessary `pcp_size` and `dcp_size` parameter passing
   - Get parallel configuration directly from distributed groups
   - Consistent with how other parts of the codebase obtain these values

**Why we need it:**
- **Maintainability**: Future changes only need to be made in one place
per Backend
- **Code quality**: Follows DRY principle and Single Responsibility
Principle
- **Readability**: Cleaner, more intuitive code structure

### Does this PR introduce _any_ user-facing change?

**No.** This is a pure refactoring with no functional changes - same
behavior, cleaner code.

### How was this patch tested?

- All existing unit tests pass with updated mocks
- No new tests needed (pure refactoring, no behavior changes)
- CI validates correctness

---

- vLLM version: v0.13.0

Signed-off-by: lico67373 <918688502@qq.com>
Co-authored-by: drslark <slarksblood@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
2026-01-24 20:12:57 +08:00
Angazenn 019a2fe6e6 [Eagle3]enhance skipping dp allreduce and add it into eagle proposer (#6192)
### What this PR does / why we need it?
This PR:
1. Enhances the logic of `_skip_all_reduce_across_dp_group` to skip all
cpu dp allreduce for dense models. This is also for purpose 2.
2. Adds `_skip_all_reduce_across_dp_group` into eagle_proposer. Now
models like Qwen3-235b supports eagle3 spec decode. A typical setting
for these moe models on pd disaggregation often introduce `dp_size > 1`.
This requires `set_forward_context` to call a cpu dp allreduce to
retrieve `num_tokens_across_dp` on all cases. Skipping this allreduce
greatly improves performance.

- vLLM version: v0.14.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60

---------

Signed-off-by: Angazenn <supperccell@163.com>
2026-01-24 11:29:42 +08:00
anon189Ty 7725314b26 [Feat] Merge the multi eagle graphs to one graph (#5940)
### What this PR does / why we need it?
This PR merge all steps of draft model in fullgraph mode, to avoid the
synchronize between each graph, reduce the bubble time.

#### Key ideas:
- The "model forward" of the step 0 (first step) and remaining steps are
captured together as a "Callable", rather than capturing each model
individually.
- "update_attn_params" is moved outside the entire graph, meaning that
all "attn_metadata" required by all steps are constructed before
"replay", and the "attn_params" of all steps are updated at once.
- Remove synchronization between the main model graph and draft model
graph.

#### Key params/functions:
- params: draft_attn_metadatas, attn_metadata_multi_steps,
slot_mapping_group
- functions: _run_merged_draft, attn_update_stack_num_spec_norm,
update_attn_params, _propose, dummy_run

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9

Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>
2026-01-23 08:37:02 +08:00
zhaomingyu13anddrslark 34fb628248 [BugFix] Support setting tp=1 for the Eagle draft model to take effect (#6097)
According to the official documentation, the parameter
"draft_tensor_parallel_size": 1 is supposed to be applied to the Eagle3
model. However, based on actual debugging, it was found that the number
of tensor parallelisms (tp) of the Eagle model is consistent with that
of the target model. The setting of tp for the draft model did not take
effect as expected.

**Note:** This feature has not been superimposed and tested with `sp`
and `dp`. It will be adapted later
No
```python
from vllm import LLM, SamplingParams

def main():
    prompts = [
        "The future of AI is",
    ]
    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
    llm = LLM(
            model="meta-llama/Llama-3.1-8B-Instruct",
            tensor_parallel_size=4,
            gpu_memory_utilization=0.9,
            enforce_eager=True,
            speculative_config={
                "method": "eagle3",
                "model": "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B"
                "draft_tensor_parallel_size": 1,
                "num_speculative_tokens": 3,
            },
        )
    outputs = llm.generate(prompts, sampling_params)
    print(f"Outputs: {outputs}")
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
```

Fixes vllm-project/vllm#31345

### What this PR does / why we need it?

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60

Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
Co-authored-by: drslark <slarksblood@qq.com>
2026-01-22 11:36:23 +08:00
drslark b2475099a0 [main][Bugfix] Fixed an problem related to embeddings sharing (#5967)
### What this PR does / why we need it?

Cancel the embeddings sharing when the embeddings of main model and the
embeddings of eagle model are different.

### Does this PR introduce _any_ user-facing change?

N/A

### How was this patch tested?

Cause i don't have `Meta-Llama-3.1-8B-Instruc`t locally, i commented it
and run:

```shell
pytest -s tests/e2e/singlecard/spec_decode/test_v1_spec_decode.py::test_llama_qwen_eagle_acceptance
```

The output is fine:

```text
.

======================================================================================================================== warnings summary =========================================================================================================================
<frozen importlib._bootstrap>:241
  <frozen importlib._bootstrap>:241: DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute

<frozen importlib._bootstrap>:241
  <frozen importlib._bootstrap>:241: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
====================================================================================================== 3 passed, 1 skipped, 2 warnings in 196.19s (0:03:16) =======================================================================================================

```

- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/2c24bc6996cb165fce92f780b388a5e39b3f4060

Signed-off-by: drslark <slarksblood@qq.com>
2026-01-20 21:34:28 +08:00
Song Zhixin 2b6dc100b5 Eagle3 mm support, enablement on qwen3vl (#4848)
### What this PR does / why we need it?
follow pr
[https://github.com/vllm-project/vllm/pull/20788](https://github.com/vllm-project/vllm/pull/20788)
, Eagle3 mm support, enablement on qwen3vl
target model
[Qwen/Qwen3-VL-8B-Instruct]([https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct])
eagle3
[MNN/Qwen3-VL-8B-Instruct-Eagle3](https://www.modelscope.cn/models/MNN/Qwen3-VL-8B-Instruct-Eagle3)
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?

pytest ./tests/e2e/singlecard/test_completion_with_prompt_embeds.py -vv

vLLM with eagle3 :
```bash
vllm serve /model/Qwen3-VL-8B-Instruct   --enforce-eager   --port 9100    --max-model-len 32768   --max-num-seqs 32   --tensor-parallel-size 2   --allowed-local-media-path /model/gx/images  --speculative-config '{
    "method": "eagle3",
    "model": "/model/hf/Qwen3-VL-8B-Instruct-Eagle3",
    "num_speculative_tokens": 3
  }'
```
vLLM without eagle3 :
```bash
vllm serve /model/Qwen3-VL-8B-Instruct   --enforce-eager   --port 9100    --max-model-len 32768   --max-num-seqs 32   --tensor-parallel-size 2   --allowed-local-media-path /model/gx/images 
```

bench:
```
vllm bench serve   --backend openai-chat   --base-url http://127.0.0.1:9100   --tokenizer /model/Qwen3-VL-8B-Instruct   --endpoint /v1/chat/completions   --model /model/Qwen3-VL-8B-Instruct   --dataset-name random  --num-prompts 50   --max-concurrency 5   --temperature 0   --top-p 1.0   --seed 123
```

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: jesse <szxfml@gmail.com>
2026-01-19 08:58:07 +08:00
Zetong Li ea01aeaab7 [Refactor][EAGLE] 4/N extract common methods from eagle and mtp (#5870)
### What this PR does / why we need it?
This PR aims to extract common methods from eagle_proposer and
mtp_proposer. This is a small step towards merging eagle and mtp.

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
by ci

- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/bde38c11df0ea066a740efe9b77fff5418be45df

---------

Signed-off-by: Zetong Li <slippersss@126.com>
2026-01-15 10:24:35 +08:00
zhaomingyu13 01805fbd7d Revert "[BugFix] Support setting tp=1 for the Eagle draft model to take effect (#5519)"(#5902)
This reverts commit d886b81971. it breaks pd function

- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/bde38c11df0ea066a740efe9b77fff5418be45df

Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
2026-01-14 20:55:10 +08:00
drslark 48ec97821a [Bugfix] Fixed an accuracy problem of sp with eagle3 (#5816)
### What this PR does / why we need it?
Fixed an accuracy problem when using eagle3 with sp.

The problem is described in
https://github.com/vllm-project/vllm-ascend/issues/5825.

It also adds a much more precise way to determine whether drafter should
use `sp` or not.

Also, it changes the `eager` of drafter to be a real `eager` in frontend
to avoid a `fx-graph` problem.

### Does this PR introduce _any_ user-facing change?

N/A

### How was this patch tested?

For simpilicity, we test it as in
https://github.com/vllm-project/vllm-ascend/issues/5825.

And we get the same result of `eagle3` with `sp` disabled.

```text
--------------------------------------------------
total_num_output_tokens: 1000
num_drafts: 437
num_draft_tokens: 1311
num_accepted_tokens: 564
mean acceptance length: 2.29
--------------------------------------------------
acceptance at token 0: 0.62
acceptance at token 1: 0.40
acceptance at token 2: 0.27
acceptance at token 3: 0.00
acceptance at token 4: 0.00
acceptance at token 5: 0.00
```

* vLLM version: v0.13.0
* vLLM main:
https://github.com/vllm-project/vllm/commit/2f4e6548efec402b913ffddc8726230d9311948d

Signed-off-by: drslark <slarksblood@qq.com>
2026-01-14 09:00:37 +08:00
zhaomingyu13anddrslark d886b81971 [BugFix] Support setting tp=1 for the Eagle draft model to take effect (#5519)
### What this PR does / why we need it?
According to the official documentation, the parameter
"draft_tensor_parallel_size": 1 is supposed to be applied to the Eagle3
model. However, based on actual debugging, it was found that the number
of tensor parallelisms (tp) of the Eagle model is consistent with that
of the target model. The setting of tp for the draft model did not take
effect as expected.

**Note:** This feature has not been superimposed and tested with `sp`
and `dp`. It will be adapted later
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
```python
from vllm import LLM, SamplingParams

def main():
    prompts = [
        "The future of AI is",
    ]

    # Create a sampling params object.
    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
    # Create an LLM.
    llm = LLM(
            model="meta-llama/Llama-3.1-8B-Instruct",
            tensor_parallel_size=4,
            gpu_memory_utilization=0.9,
            enforce_eager=True,
            speculative_config={
                "method": "eagle3",
                "model": "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B"
                "draft_tensor_parallel_size": 1,
                "num_speculative_tokens": 3,
            },
        )

    # Generate texts from the prompts.
    outputs = llm.generate(prompts, sampling_params)
    print(f"Outputs: {outputs}")
    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: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/45c1ca1ca1ee8fa06df263c8715e8a412ff408d4

Fixes vllm-project/vllm#31345

Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
Co-authored-by: drslark <slarksblood@qq.com>
2026-01-13 09:14:30 +08:00
zhenwenqi2024 97f6be8108 [feature]dcp&pcp support mlapo (#5672)
### What this PR does / why we need it?
mlapo in deepseek is a huge performance improvement in decode, this pr
support pcp & dcp with mlapo

### Does this PR introduce _any_ user-facing change?
NO

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/2f4e6548efec402b913ffddc8726230d9311948d

---------

Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
2026-01-08 23:49:23 +08:00
drslark ccbc5e2ba1 [Feat][Bugfix][main] Adapted SP to eagle3 (#5562)
### What this PR does / why we need it?

Adapted sp to eagle3.

There may still be some problems, e.g., accuracy in some scenes,
`sp`+`dp`...

We will fix them later.

### Does this PR introduce _any_ user-facing change?

N/A

### How was this patch tested?

We tested it mainly in a new `e2e`.

```shell
pytest -s tests/e2e/singlecard/spec_decode/test_v1_spec_decode.py::test_llama_qwen_eagle_acceptance
```

```text
.

=============================== warnings summary ===============================
<frozen importlib._bootstrap>:241
  <frozen importlib._bootstrap>:241: DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute

<frozen importlib._bootstrap>:241
  <frozen importlib._bootstrap>:241: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
============= 3 passed, 1 skipped, 2 warnings in 142.05s (0:02:22) =============
```

It passed.

- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/7157596103666ee7ccb7008acee8bff8a8ff1731

Signed-off-by: drslark <slarksblood@qq.com>
2026-01-08 15:33:52 +08:00
LICO67373andweijinqian0 380f089fbf [Refactor] Fix AttentionMaskBuilder singleton and remove redundant pcp_prefill_mask (#4870)
## What this PR does / why we need it?

This PR fixes the `AttentionMaskBuilder` singleton initialization issue
introduced in PR #4779 and removes the unused `pcp_prefill_mask` field.

### Background

After PR #4779 made `AttentionMaskBuilder` a singleton with `@singleton`
decorator, the class constructor now requires a `device` parameter.
However, two initialization sites were still using the old parameterless
constructor, causing failures.

### Changes

1. **Fix singleton initialization**
- Fixed `AttentionMaskBuilder()` → `AttentionMaskBuilder(self.device)`
in `AscendMLAMetadataBuilder.__init__()`
- Fixed `AttentionMaskBuilder()` → `AttentionMaskBuilder(self.device)`
in `AscendAttentionMetadataBuilder.__init__()`

2. **Remove unused field**
- Removed `pcp_prefill_mask` field from
`AscendPrefillContextParallelMetadata` (never used in codebase)
   - Updated related test assertions

### Related

- Issue #5463
- PR #4779 (Unify all mask generation methods)
- PR #5389 (Make AttentionMaskBuilder singleton)

## Does this PR introduce _any_ user-facing change?

No. This is an internal refactoring.

## How was this patch tested?

- ✅ Local testing: No linter errors
- ✅ Unit tests for attention modules verified
- ⏳ CI pipeline

Signed-off-by: lico67373 <918688502@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
2026-01-07 17:09:52 +08:00
Zetong Li fe3f2c7702 [Refactor][EAGLE] 3/N delete redundant methods in mtp_proposer (#5420)
### What this PR does / why we need it?
This PR aims to delete redundant methods in mtp_proposer. All the
deleted methods now can be found in eagle_proposer. We also remove some
methods in eagle_proposer since they are identical to those in
vllm-eagle.

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
by ci

- vLLM version: release/v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/81786c87748b0177111dfdc07af5351d8389baa1

---------

Signed-off-by: Zetong Li <slippersss@126.com>
2026-01-06 16:47:39 +08:00
Shanshan Shen b94d589769 [MM][Bugfix] Update hf_config to hf_text_config (#5319)
### What this PR does / why we need it?

Following https://github.com/vllm-project/vllm-ascend/pull/5205, update
`hf_config` to `hf_text_config`.

Find more details at
https://github.com/vllm-project/vllm-ascend/pull/5205#issuecomment-3675417534
and
https://github.com/vllm-project/vllm-ascend/pull/5205#issuecomment-3677920872.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: release/v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/5fbfa8d9ef15948599631baeb91e8220b2ee9bcc

Signed-off-by: shen-shanshan <467638484@qq.com>
2026-01-06 16:41:39 +08:00
wjunLu 3cf059a72b [Main2Main] Upgrade vllm commit to 0105 (#5595)
### What this PR does / why we need it?

Upgrade vllm commit to 0105 (8be6432bdaf6275664d857b1e5e9bf8ed1ce299e)

1. Remove `maybe_padded_num_tokens` arg in `model_runner_v1.py` since
https://github.com/vllm-project/vllm/pull/31517 deleted unused arg

2. Remove dense `Qwen/Qwen3-0.6B` in
`tests/e2e/multicard/test_aclgraph_capture_replay.py` and
`tests/e2e/multicard/test_data_parallel.py` due to
https://github.com/vllm-project/vllm/pull/30739
where offline data parallel mode will not be supported/useful for dense
models

3. Adapt `vllm_ascend/worker/worker.py` due to
https://github.com/vllm-project/vllm/pull/31584

4. Adapt `self.block_size` calling due to
https://github.com/vllm-project/vllm/pull/31540

5. Modify `test_mla_v1.py` due to
https://github.com/vllm-project/vllm/pull/28454 , which refactorred
`get_head_size()`

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/7157596103666ee7ccb7008acee8bff8a8ff1731

Signed-off-by: wjunLu <wjunlu217@gmail.com>
2026-01-06 08:44:29 +08:00
lilinsiman 52863c4165 [Refactor][EAGLE] 2/N: load model and generate token (#5437)
### What this PR does / why we need it?
1. Refactor eagle and mtp function: load_model and generate_token_ids
2. Remove redundant code in mtp and eagle file
3. Refactor the UT of file

2/N of Refactor and merge mtp and eagle
Relational RFC: https://github.com/vllm-project/vllm-ascend/issues/5467

### Does this PR introduce _any_ user-facing change?
no

### How was this patch tested?
ut and tests

- vLLM version: release/v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/81786c87748b0177111dfdc07af5351d8389baa1

---------

Signed-off-by: lilinsiman <lilinsiman@gmail.com>
2026-01-05 14:07:54 +08:00
zhenwenqi2024andzzzzwwjj 5d9fde9819 [Feature] Refactor PCP &DCP related code (#5214)
### What this PR does / why we need it?
Refactor pcp& dcp related code. we use pcp_manager class to Unifiy
Manage pcp & dcp . as we do this , many code can be deleted from
model_runner, and can avoid break pcp & dcp by other developments.
RFC:https://github.com/vllm-project/vllm-ascend/issues/5449
### Does this PR introduce _any_ user-facing change?
NO

### How was this patch tested?

- vLLM version: release/v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
Co-authored-by: zzzzwwjj <34335947+zzzzwwjj@users.noreply.github.com>
2025-12-31 09:29:57 +08:00
Zetong Li 92353c0643 [Refactor][EAGLE] 1/N delete __init__ in mtp_proposer (#5176)
### What this PR does / why we need it?
This PR aims to refactor eagle-related modules in vllm-ascend.

This is the starting PR of eagle refactoring. Provided with vllm-eagle,
ascend-eagle and ascend-mtp, we first let ascend-mtp inherit from
ascend-eagle and let ascend-eagle inherit from vllm-eagle. As a
initialization, we just delete `__init__` in mtp_proposer and simplify
the corresponding logic in eagle_proposer.

Based on "vllm-eagle <----- ascend-eagle <----- ascend-mtp", our target
is to gradually delete ascend-mtp and enable ascend-eagle to converge to
vllm-eagle. So the main workspace is eagle_proposer. In this way, we
hope that contributors can concurrently refactor eagle.

Incoming changes:
1. delete common methods in vllm-eagle & ascend-eagle & ascend-mtp
2. delete `load_model` in mtp_proposer
3. delete `dummy_run` and `propose` in mtp_proposer
4. ......

RFC: #5467

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
by ci

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: Zetong Li <slippersss@126.com>
2025-12-29 16:25:52 +08:00
3e67e8276c [Feature] Support to use fullgraph with eagle (#5118)
### What this PR does / why we need it?
    
We support to use full graph with eagle. 

Change list:
1. Distinguish between processing graph_params and draft_graph_params in
attention_v1.
    2. Adapt the full-graph mode in eagle_proposer, include:
        1). If use full graph, make Fullgraph Wrapper when load model.
2). Build a new meatadata, set running mode in FULL and mark attention
update in dummy_run when in Fullgraph mode.
3). Fixed and fill any attn_metadata, such as
attn_metadata.slot_mapping.
        4). Add a descriptor.
        5). Set running mode and triggered update metadata.
3. Trans is_mtp_model to is_draft_model, and add the update of
workspace.

NOTE:
When set async_scheduling=True, the draft model will enforce execution
in eager mode.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 09:54:51 +08:00
zhangsicheng5 78aa7f2693 [feature] support pcp + mtp in full graph (#4572)
1. support pcp + mtp in full graph
2. pcp/dcp related mtp bugfix
3. support pcp + mtpx

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

Signed-off-by: zhangsicheng5 <zhangsicheng5@huawei.com>
2025-12-22 16:13:39 +08:00
Qiu ea6206bb18 [bugfix][ACLGraph][MTP] deletes cudagraph_batch_sizes in MtpProposer (#5183)
### What this PR does / why we need it?
This PR deletes `cudagraph_batch_sizes` in `MtpProposer` and reuses the
one in `NPUModelRunner`.

During our deployment of DeepSeek-V3.2 with MTP across machines 2P2D and
conducting AISBench stress testing, an error occurred (see below). After
investigation, we found that
`compilation_config.cudagraph_capture_sizes` is modified by
`adjust_cudagraph_sizes_for_spec_decode` in `NPUModelRunner`. This
modification only updates `cudagraph_batch_sizes` in `NPUModelRunner`
but is not synchronized to `MtpProposer`. After discussion (CC @yiz-liu)
, we believe it is unnecessary to maintain `cudagraph_batch_sizes` in
`MtpProposer`; it should directly use the variable from
`NPUModelRunner`.

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
2025-12-22 14:08:27 +08:00
zhenwenqi2024andMengqing Cao 950570f8d1 [Bugfix]delele profile_run in model_runner (#5122)
### What this PR does / why we need it?
delete sekf.in_profile_run in model_runner to make EPLB works as expect
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
Signed-off-by: Mengqing Cao <cmq0113@163.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
2025-12-17 23:48:34 +08:00
anon189Tyanddrslark 5b1da4e914 [Feat] Support async_scheduler and disable_padded_drafter_batch in eagle (#4893)
### What this PR does / why we need it?
We refactored the eagle_proposer.py to adapt the framework of eagle.py
in vllm-v0.12.0, to support the logit of padded drafter batch and
async-scheduler.

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>
Co-authored-by: drslark <slarksblood@qq.com>
2025-12-16 22:06:40 +08:00
zhenwenqi2024 4ed2951400 【Feature】refactor npu_modelrunner for profile_run (#4993)
### What this PR does / why we need it?
(1)refactor npu_model_runner for profile_run
(2) move _select_moe_comm_method to ascend_forward_context
(3) delete _init_model_kwargs in npu_model_runner

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
Na
- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
Signed-off-by: zhenwenqi2024 <155598497+zhenwenqi2024@users.noreply.github.com>
2025-12-16 17:44:04 +08:00
Jade Zheng c064d11fd7 [Cleanup] Remove unused attn_metadata parameter from Proposer classes (#4862)
The `attn_metadata` is not used by any draft proposer, so we can remove
it.


- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>
2025-12-15 21:21:38 +08:00
Li Wang 8d2998d0e4 [Misc] Upgrade vllm hash to 12_14 (#5000)
### What this PR does / why we need it?

### Does this PR introduce _any_ user-facing change?
1. fix https://github.com/vllm-project/vllm/pull/27938
2. fix https://github.com/vllm-project/vllm/pull/27145
pooling models now supports chunked prefill and prefix caching,
3. fix https://github.com/vllm-project/vllm/pull/30181
define the CPU fields in the field config where they really belong.
4. fix https://github.com/vllm-project/vllm/pull/28168
define the CPU fields in the field config where they really belong.
5. fix https://github.com/vllm-project/vllm/pull/30201
some moudle rename
6. fix https://github.com/vllm-project/vllm/pull/29067
fusedmoe moudle refactor
7. fix https://github.com/vllm-project/vllm/pull/29066
fusedmoe moudle refactor
8. fix https://github.com/vllm-project/vllm/pull/29624
### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: wangli <wangli858794774@gmail.com>
2025-12-15 19:54:23 +08:00
MengLong Chen fa367e3b1a [CI] Add mtp_proposer ut (#4397)
### What this PR does / why we need it?
Add mtp_proposer ut

- vLLM version: v0.12.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9

---------

Signed-off-by: chenmenglong <chenmenglong1@huawei.com>
2025-12-12 20:41:31 +08:00
7f2673ea2d upgrade vLLM to main (#4608)
1. fix https://github.com/vllm-project/vllm/pull/28542
The model structure modifications we involved in are:
     - Qwen2.5-VL(still exist some patch)
     - Qwen2-VL
     - Qwen2
     - DeepSeek series
     - Qwen-moe series
2. fix https://github.com/vllm-project/vllm/pull/29121
   the output token now  type changed from np to `list[list[int]]`

3. fix https://github.com/vllm-project/vllm/pull/29262
    `xformers` backend for multimodal now has been deprecated
4. fix https://github.com/vllm-project/vllm/pull/29342

5. fix https://github.com/vllm-project/vllm/pull/28579
6. fix https://github.com/vllm-project/vllm/pull/28718
7. fix https://github.com/vllm-project/vllm/issues/28665
8. fix https://github.com/vllm-project/vllm/pull/26847
vllm introduced the `optimization-level`, some default config has been
changed, and the param `--enforce-eager` has been deprecated
9. fix http://github.com/vllm-project/vllm/pull/29223 it retuns tuple
for sampler.
10. fix https://github.com/vllm-project/vllm/pull/29471 we'll remove the
related patch to avoid this kind of error.

Co-authored-by: hfadzxy <starmoon_zhang@163.com>
Co-authored-by: wangli <wangli858794774@gmail.com>


- vLLM version: v0.11.2

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: wangli <wangli858794774@gmail.com>
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
Co-authored-by: wangli <wangli858794774@gmail.com>
Co-authored-by: hfadzxy <starmoon_zhang@163.com>
2025-12-02 22:10:52 +08:00
Zhu Yi Lin 755b635844 [TEST] Add eagle proposer ut (#4447)
### What this PR does / why we need it?
Add eagle proposer ut

- vLLM version: v0.11.2

Signed-off-by: GDzhu01 <809721801@qq.com>
2025-11-27 21:59:31 +08:00