### What this PR does / why we need it?
Use nginx for package cache to speed up CI
- vLLM version: v0.14.0
- vLLM main:
d68209402d
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
Rectify the problem that the pcp and pd separation and kv pooling
scenario.
In the pooling scenario, multi_nodes_meta_mapping is empty. As a result,
an error is reported when the remote_host information is obtained
through the get_remote_port_send_num method.
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
### What this PR does / why we need it?
Due to the long-term lack of synchronization with the upstream code, a
problem that led to a decrease in the acceptance rate of the
Qwen3-30B-A3B-EAGLE3 draft model was introduced when fixing the
bug（#5967）. Now, synchronize with the upstream and fix this bug
### 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="Qwen/Qwen3-30B-A3B",
tensor_parallel_size=4,
gpu_memory_utilization=0.9,
enforce_eager=True,
speculative_config={
"method": "eagle3",
"model": "AngelSlim/Qwen3-a3B_eagle3"
"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:
d68209402d
Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
Co-authored-by: drslark <slarkblood@qq.com>
### What this PR does / why we need it?
The test case
`tests/e2e/singlecard/spec_decode/test_v1_spec_decode.py::test_llama_qwen_eagle_acceptance`
fails occasionally, such result seems not stable with method `eagle`,
for example:
[tests/e2e/singlecard/spec_decode/test_v1_spec_decode.py::test_llama_qwen_eagle_acceptance](https://github.com/vllm-project/vllm-ascend/actions/runs/21249578476/job/61147453980?pr=6151)
This PR skips the `eagle` tests to keep CI success
- vLLM version: v0.14.0
- vLLM main:
d68209402d
Signed-off-by: wjunLu <wjunlu217@gmail.com>
### What this PR does / why we need it?
Add nightly ci test for deepseek v3.1
- vLLM version: release/v0.13.0
- vLLM main:
bc0a5a0c08
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
### What this PR does / why we need it?
1.Incorporate the warm up of the EPLB into the profile run.
2.Reusing the same gather buffer
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
qwen3-235b aime baseline
| dataset | version | metric | mode | vllm-api-general-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 86.67 |
eplb The OOM issue does not occur.
| dataset | version | metric | mode | vllm-api-general-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 86.67 |
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
### What this PR does / why we need it?
Align max_num_batched_tokens with tp*pcp when using FLASHCOMM1 to avoid
assert error in `NPUModelRunner._dummy_run`.
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
---------
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
`vllm_ascend` already supports several speculative decoding strategies
such as MTP, EAGLE, N-gram, and suffix decoding. However, Medusa is not
yet supported. Medusa is an efficient speculative decoding framework
that leverages a lightweight draft model to propose multiple tokens in a
single step, which can significantly improve decoding throughput and
reduce latency.
To enable Medusa-based speculative decoding on Ascend hardware and
provide more decoding options for users, this PR adds Medusa support
into the `vllm_ascend` speculative decoding pipeline.
### Does this PR introduce _any_ user-facing change?
This PR introduces Medusa speculative decoding as an additional
speculative decoding method:
✔ Adds `MedusaProposer` and integrates it into the speculative decoding
registry
✔ Extends `SpecDcodeType` with a `MEDUSA` enum entry
✔ Updates `NPUModelRunner` to recognize and invoke Medusa during
decoding
✔ Adds Medusa-specific handling in the draft token generation logic
✔ Ensures backward compatibility — Medusa is only used when explicitly
enabled
Key code changes include:
* New file: `vllm_ascend/spec_decode/medusa_proposer.py`
* Register Medusa in `get_spec_decode_method`
* Extend proposer type hints to include `MedusaProposer`
* Add a Medusa-specific branch in `generate_draft_token_ids`
* Pass `sample_hidden_states` required by Medusa
### How was this patch tested?
Medusa is implemented as a new proposer class (`MedusaProposer`)
following the existing speculative decoding interface. The integration
works as follows:
1. Users enable Medusa via the speculative decoding configuration.
2. `get_spec_decode_method()` returns a `MedusaProposer` instance when
`method="medusa"`.
3. During decoding, `NPUModelRunner` detects that the active drafter is
a `MedusaProposer`.
4. Instead of the generic speculative decoding path, the Medusa-specific
`generate_token_ids()` method is invoked, which consumes:
* `valid_sampled_token_ids`
* `sampling_metadata`
* `spec_decode_metadata`
* `sample_hidden_states`
5. The proposed tokens are validated by the target model as usual.
When Medusa is not enabled, the decoding pipeline behaves exactly as
before, ensuring full backward compatibility.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: simplzyu <191163281@qq.com>
Signed-off-by: simplzyu <zhenyuguo@cmbchina.com>
### What this PR does / why we need it?
PCP/DCP splits the kv-cache onto different cards. After introducing the
parameter cp-kv-cache-interleave-size, the first size tokens will be
cached at Card 0, and so on.
However, if there are too few tokens, some cards will not store the
key-value pairs, resulting in values of 0, corrupted values, and
precision issues. Currently, additional operations are introduced to
avoid this precision problem.
After we integrate FIA operator in mla_cp._forward_decode and CANN
updates to 8.5.0, we now can remove these additional operations.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
passed all CI by CANN 8.5.0
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
Signed-off-by: dsxsteven <dsxsteven@sina.com>
Signed-off-by: dsxsteven <36877507+dsxsteven@users.noreply.github.com>
### What this PR does / why we need it?
Now `seq_lens` was not being reset correctly after each step due to
missing code that clears the sequence lengths. As a result, when
processing a smaller batch after a larger batch, the `seq_lens` from the
larger batch was still carried over. This caused the attention operator
to compute using an unnecessarily larger sequence length, leading to an
increased computation load and performance degradation.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: ZYang6263 <zy626375@gmail.com>
### What this PR does / why we need it?
When the P node accesses the proxy meteserver, add the SSL certificate
and the CA certificate path to improve security.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
By ci
- vLLM version: v0.13.0
- vLLM main:
bde38c11df
---------
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
Re-open `tests/e2e/singlecard/test_aclgraph_accuracy.py` and update its
golden results to match PTA 2.9.0
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: wjunLu <wjunlu217@gmail.com>
### What this PR does / why we need it?
update supported features
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
### What this PR does / why we need it?
Drop vLLM 0.13.0 support, upgrade to 0.14.0
- vLLM version: v0.13.0
- vLLM main:
d68209402d
---------
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
### What this PR does / why we need it?
dispatch_ffn_combine use hccl buffer as shared buffer, if hccl buffer
not enough，operator will error with "MTE out of range"
now add check for hccl buffer size, if not enough, will prompt "hccl
buffer is too small" and indicate what the expectation is.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
d68209402d
---------
Signed-off-by: lhchg <lhao_cheng@163.com>
### 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:
11b6af5280
Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>
### What this PR does / why we need it?
This PR aims to remove `use_aclgraph` and use `use_cuda_graph` just the
same as eagle_proposer in mtp_proposer. The reason of these changes are
described below.
There is a scenario that `use_aclgraph=True` while
`use_cuda_graph=False`, e.g. enabling `async_scheduling=True`. When
using deepseek v3.2, `common_attn_metadata.num_input_tokens` is
important and it should be the same as `num_input_tokens` entering into
model. In the above scenario, `use_aclgraph` accidentally pad
`num_tokens` to `num_input_tokens`, coinciding with
`common_attn_metadata.num_input_tokens`. But later eager mode is
triggered and actually we don't need padding. That means that the code
logic is incorrect but the running output looks fine.
However, `common_attn_metadata.num_input_tokens` should mean
`num_input_tokens` entering into model. So we should update
`common_attn_metadata.num_input_tokens = num_input_tokens` after
padding. Therefore, we can safely use normal `use_cuda_graph` instead of
problematic `use_acl_graph`.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
by ci
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
Signed-off-by: Zetong Li <slippersss@126.com>
### What this PR does / why we need it?
Fix Qwen3VL dense quant model load weights Error.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
The Qwen3VL quantized model service initialized successfully. Inference
requests are processed correctly, and valid responses are returned.
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: 李少鹏 <lishaopeng21@huawei.com>
### What this PR does / why we need it?
Replace the npu_multi_head_latent_attention with FIA operator in
mla_cp.py _forward_decode.
Adjust mla_attn_dpc_pcp in acl_graph.py
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: 白永斌 <baiyongbin3@h-partners.com>
Signed-off-by: Bai Yongbin <845473182@qq.com>
Signed-off-by: tongyuzhou <t00886357@china.huawei.com>
Co-authored-by: 白永斌 <baiyongbin3@h-partners.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: tongyuzhou <t00886357@china.huawei.com>
### What this PR does / why we need it?
Upgrade wheel building's CANN to 8.5.0 and update the Docs
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: wjunLu <wjunlu217@gmail.com>
### What this PR does / why we need it?
Install clang in dokerfile for triton ascend
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
### What this PR does / why we need it?
Upgrade PTA to 2.9.0
- vLLM version: v0.13.0
- vLLM main:
d68209402d
---------
Signed-off-by: wjunLu <wjunlu217@gmail.com>
### What this PR does / why we need it?
This PR builds upon PR
https://github.com/vllm-project/vllm-ascend/pull/5011 and aims to
further enhance the npu_graph_ex_passes module. Based on prior work, we
have added graph optimization support for the add_rms_quant fused
operator in scenarios where a bias term is present—ensuring the fusion
pattern is correctly registered and matched into the computation graph.
For validation, we switched to the Qwen3-235B-A22B-W8A8 model for
QKVNormRopeWithBias and Qwen3-32B model for QKVNormRope . Benchmark
results show that, compared to the unfused baseline, enabling this
fusion pass significantly improves inference throughput for W8A8
quantized models.
For more details can refer to the
RFC:https://github.com/vllm-project/vllm-ascend/issues/4715
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
```
llm = LLM(
model=model,
tensor_parallel_size=GPUs_per_dp_rank,
enforce_eager=False,
enable_expert_parallel=enable_expert_parallel,
trust_remote_code=trust_remote_code,
gpu_memory_utilization=0.98,
max_num_batched_tokens=512,
# load_format="dummy",
max_model_len=2048,
max_num_seqs=16,
quantization="ascend",
additional_config={
"refresh": True,
"enable_npugraph_ex": True
},
compilation_config={
"cudagraph_capture_sizes": [8, 16],
"cudagraph_mode": "FULL_DECODE_ONLY",
},
)
if profile_dir:
llm.start_profile()
outputs = llm.generate(prompts, sampling_params)
if profile_dir:
llm.stop_profile()
for i, output in enumerate(outputs):
if i >= 5:
break
prompt = output.prompt
generated_text = output.outputs[0].text
print(
f"DP rank {global_dp_rank}, Prompt: {prompt!r}, "
f"Generated text: {generated_text!r}"
)
```
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: cjian <2318164299@qq.com>
### What this PR does / why we need it?
The issue of the D node mistakenly sending the pull-end signal twice,
leading to the P node printing logger errors abnormally, has been
resolved.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
By ci
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
This helps to fix a bug in for pa get_workspace. In earlier
implementation, we use `_npu_paged_attention_get_workspace` in
`_update_pa_attn_params`. However, this might cause some potential
memory problems as it dynamically allocate new memory for workspace when
calling this api. Therefor, we move this back to capturing, and use a
fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
This patch purpose to optimize the lint check term. The main idea is to
reduce unnecessary installation time.
1. The installation of vllm is not must, only append the path of vllm
src to the `PATHONPATH` is effective
2. This installation of `requirements-dev.txt` is not must, we have a
pre-built image `quay.io/ascend-ci/vllm-ascend:lint` with all the
requirements installed in advance.
**NOTE**: the conditions for triggering image builds are: 1).Daily
scheduled build; 2) Build when requirements are modified; 3) Manual
build. This ensures that the dependencies in our image are up-to-date to
the greatest extent possible.
3. The `mypy` was separated from the `pre-commit` hook for performance
reasons; we found that integrating `mypy` into the `pre-commit` hook
resulted in poor performance.
4. Reduce the CPU core consumption from 16 -> 8
### Does this PR introduce _any_ user-facing change?
The end-to-end lint time was optimized from 20min/per PR to 8min/per PR
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
When running the Qwen2.5-Omni-7B model on Ascend NPU, the engine fails
during the profiling/warmup stage with the following error:
`AclNN_Runtime_Error(EZ9903): rtKernelLaunchWithHandleV2 failed: 507035.
The vector core execution is abnormal.`
error log:
https://github.com/vllm-project/vllm-ascend/actions/runs/21144534911/job/60806765393#step:17:6412
This error is specifically triggered by the `triton_mrope` kernel when
handling the unique `mrope_section` configurations of the Omni model.
Other multimodal models with standard sections (e.g., [16, 24, 24]) or
standard LLMs work correctly with Triton.
Modified vllm_ascend/ops/rotary_embedding.py to add a conditional check
before calling forward_triton.
1. For standard LLMs (mrope_interleaved = True ), it continues to use
Triton for acceleration.
2. For complex configurations (like Qwen2.5-Omni mrope_interleaved =
False ), it now falls back to the native super().forward_oot() path,
which uses the stable torch_npu or PyTorch implementation.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
### What this PR does / why we need it?
When using the full_decode_only mode, the vllm framework will still use
the torch.fx.passes.split_module.split_module API to process the
corresponding GraphModule of the model.
However, the output of this API may cause the output of the fx graph to
no longer be a tuple, and torch.compile enforces strict checks on this.
Previously, we manually modified the fx graph, which introduced an
abnormality in the model output type.
In this PR, we switched to using PyTorch's native API to modify the FX
graph, and removed the code that was previously added to handle output
type anomalies.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
---------
Signed-off-by: chencangtao <chencangtao@huawei.com>
Co-authored-by: chencangtao <chencangtao@huawei.com>
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}")
```
Fixesvllm-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:
d68209402d
Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
Co-authored-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
Operator `DispatchGmmCombineDecode` does not support non-W8A8 scenarios
and cannot share the same communication domain with Operator
`Dispatch`/`Combine`.
> for instance, when the draft model uses a non-W8A8 MOE architecture
while the main model employs a W8A8 MOE architecture.
Therefore days ago, I implemented an interception that unconditionally
disables Operator `DispatchGmmCombineDecode` whenever the speculative
mode is `EAGLE` or `EAGLE-3`. [PR:
5293](https://github.com/vllm-project/vllm-ascend/pull/5293)
However, this approach was not precise enough.
This PR further refines the logic by specifically identifying the draft
model's configuration: Operator `DispatchGmmCombineDecode` will now be
disabled only when the draft model uses an MOE architecture and is
non-W8A8.
More info about this operator, please refer to RFC: issue
https://github.com/vllm-project/vllm-ascend/issues/5476
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Acc
test qwen3-235b eplb on a single A3 node(ep16),
with dispatch_gmm_combine_decode
```shell
nic_name="xxxx"
local_ip="xxx.xxx.xxx.xxx"
export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export VLLM_ASCEND_ENABLE_FUSED_MC2=2
echo "VLLM_ASCEND_ENABLE_FUSED_MC2=${VLLM_ASCEND_ENABLE_FUSED_MC2}"
export HCCL_OP_EXPANSION_MODE="AIV"
export HCCL_BUFFSIZE=512
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD
vllm serve /dataset/Qwen3-235B-A22B-Instruct-2507-w8a8-QuaRot/ \
--served-model-name "qwen" \
--host 0.0.0.0 \
--port 8004 \
--async-scheduling \
--tensor-parallel-size 4 \
--data-parallel-size 4 \
--max-num-seqs 64 \
--max-model-len 40960 \
--max-num-batched-tokens 16384 \
--gpu-memory-utilization 0.9 \
--enable-expert-parallel \
--no-enable-prefix-caching \
--quantization "ascend" \
--trust-remote-code \
--speculative_config \
'{
"method": "eagle3",
"model": "/dataset/Qwen3-235B-A22B-Instruct-2507-speculator-eagle3/",
"num_speculative_tokens": 2
}' \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
2>&1 | tee qwen3_235b_eagle3.log
```
| dataset | version | metric | mode | vllm-api-stream-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 80.00 |
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: wangqiankun <wangqiankun13@huawei.com>
This PR addresses a request ID mismatch issue in the PD
(Prefill-Decoding) separation deployment scenario for vllm-ascend.
Upstream vLLM recently mitigated request ID collisions by appending a
random suffix to each request_id (e.g., req-123 → req-123-abc), refer to
[PR-27987](https://github.com/vllm-project/vllm/pull/27987 ) &
[PR-29665](https://github.com/vllm-project/vllm/pull/29665). While this
works in single-node deployments, it breaks compatibility in
PD-separated setups: the Producer (Prefill node) and Consumer (Decoding
node) end up with different request_id values, preventing the Consumer
from correctly retrieving the KV cache generated by the Producer.
To resolve this, this PR introduces a new field remote_request_id in the
metadata passed via mooncake_connector. The Producer preserves and
forwards the original (unmodified) request_id as remote_request_id. The
Consumer then uses this remote_request_id—instead of its locally
generated suffixed ID—to fetch the correct KV cache from the Prefill
node.
This ensures consistent request identification across PD nodes while
maintaining compatibility with upstream vLLM’s request ID deduplication
mechanism.
<img width="1279" height="781" alt="image"
src="https://github.com/user-attachments/assets/274238c1-dab6-4d3a-9ee4-6e578679b762"
/>
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: ghphotoframe <854746559@qq.com>
Co-authored-by: jiangweixiang <jwx02384838@antgroup.com>
Signed-off-by: zzhx1 <zzh_201018@outlook.com>
### What this PR does / why we need it?
> Extracted from PR #5513
Based on the Sharded-CP feature PR:#4702;
RFC:https://github.com/vllm-project/vllm/issues/30055
### Support FULL_DECODE_ONLY Mode under PD-Mixed Scenario:
Extends DSA-CP to handle the FULL_DECODE_ONLY execution mode when
running in a prefill-decode mixed (PD-mixed) serving environment,
improving throughput and resource utilization for decode-intensive
workloads.
**In pure prefill nodes:**
- Both q_proj and o_proj are sharded across world ranks, using
**broadcast** for weights distribution.
**In PD-mixed nodes (supporting both prefill and decode):**
- q_proj is fully replicated (not sharded) to avoid communication
overhead during decoding.
- o_proj Using the original TP `RowParallelLinear` method to store
weights
**During prefill execution:**
- o_proj forwards through all_gather to collect weights, reconstructing
the complete o_proj weights on each card.
**During decode (graph replay phase):**
- Additional all_to_all (before o_proj) and reduce_scatter (after
o_proj) are introduced to enable sequence-parallel output aggregation
while maintaining correctness under SFA CP.
### benchmark:
- TTFT increased by **527%**
- TPOT increased by **180%**
<img width="1550" height="938" alt="image"
src="https://github.com/user-attachments/assets/9b7a03d8-a3db-4a99-8923-6e5bfcfecf72"
/>
### Does this PR introduce _any_ user-facing change?
None
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: zzhx1 <zzh_201018@outlook.com>
Signed-off-by: zzhxx <zhangzihang23@mails.ucas.ac.cn>
Co-authored-by: clrs97 <524936896@qq.com>
### What this PR does / why we need it?
1) Default enable MLAPO for deepseek MLA Attention W8A8 models on PD
disagregation D Instance, for example: DeepSeekV3-W8A8,
DeepSeek-R1-W8A8.
2) Default enable MLAPO for DeepSeek SFA Attention W8A8 models,
currently is DeepSeek-V3.2-W8A8.
### Does this PR introduce _any_ user-facing change?
Don't need use manully to VLLM_ASCEND_ENABLE_MLAPO=1, to enable MLAPO
feature for deepseek w8a8 model
The effect of enabling MLAPO SFA model deployed on a single A3 Node:
Test
with:tests/e2e/nightly/single_node/models/test_deepseek_v3_2_exp_w8a8.py
dataset: gsm8k-lite，without set MTP, FULL GRAPH, has 19% promote：
未默认开启 MLAPO 时：
├─────────────────────────┤
│ TTFT │ 14055.8836 ms │
├─────────────────────────┤
│ ITL │ 66.8171 ms. │
├─────────────────────────┤
│ Output Token Throughput │ 104.9105 token/s │
├─────────────────────────┤
默认开启 MLAPO 时：
├─────────────────────────┤
│ TTFT │ 3753.1547 ms │
├─────────────────────────┤
│ ITL. │ 61.4236 ms. │
├─────────────────────────┤
│ Output Token Throughput │ 125.2075 token/s│
├─────────────────────────┤
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
---------
Signed-off-by: leo-pony <nengjunma@outlook.com>
### What this PR does / why we need it?
Supplement PD separation parameters of DeepSeek V3.1
The recommended parameter configuration for DeepSeek V3.1 in the EP32
scenario after PD separation has been adjusted, and the core parameters
have been described in detail.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: chenmenglong <chenmenglong1@huawei.com>
### What this PR does / why we need it?
1. Implement a **high-performance Triton custom kernel** for the rotary
position embedding (RoPE) operator on **Ascend NPU** platform
2. Fix critical bugs in the Triton RoPE kernel registration and
invocation process: including incorrect fake impl function name
matching, wrong torch ops namespace for kernel call, missing self
parameter in cos/sin slice fetching, and syntax errors in function type
annotations.
3. Achieve **extreme performance optimization** for the core RoPE
operator: the single inference latency is reduced from **57.1 μs** to
**9 μs**, with **6.34x performance improvement** and **84.24% latency
reduction**.
4. The RoPE operator is a **hot path** that is executed in every
transformer layer during LLM inference, the optimization will directly
reduce the overall inference latency and improve the throughput of LLM
serving on Ascend NPU.
5. Keep full backward compatibility: the Triton kernel is enabled only
when `HAS_TRITON=True`, and automatically fall back to the original
Ascend NPU native implementation if Triton is not available, no
functional regression.
### Does this PR introduce _any_ user-facing change?
**NO**
- No changes to any public APIs, interfaces or inference behaviors of
vLLM.
- No impact on the text generation quality and correctness of the large
model.
- The optimization is transparent to end users, only the inference speed
(latency/throughput) is improved without any functional change.
### How was this patch tested?
1. **Environment Validation**: Tested on Ascend NPU platform with
vLLM-Ascend framework, Triton library installed and enabled
(`HAS_TRITON=True`).
2. **Kernel Registration Test**: Verified the Triton RoPE kernel
(`rope_forward_triton`) is successfully registered to
`torch.ops._C_ascend` namespace without any
`ValueError/NameError/SyntaxError`.
3. **Functional Correctness Test**: Run large model (GLM4/MoE) inference
on the Ascend NPU platform, the generated text content is **completely
correct** (no garbled text, no logical errors), consistent with the
original implementation.
4. **Performance Benchmark Test**: Measure the single execution latency
of the RoPE operator before/after optimization, confirm the latency is
stably reduced from 57.1 μs to 9 μs, the performance gain is valid and
stable.
5. **Fallback Mechanism Test**: Manually disable Triton
(`HAS_TRITON=False`), verify the code correctly falls back to the
original Ascend NPU native RoPE implementation, no service crash and
normal inference.
6. **Compatibility Test**: Test with different tensor shapes/sizes of
query/key, all cases work correctly with the Triton kernel, no shape
mismatch error.
- operator supply by Hexiang Wang
- vLLM version: v0.13.0
- vLLM main:
11b6af5280
---------
Signed-off-by: ZCG12345 <2097562023@qq.com>
### What this PR does / why we need it?
Update causal_conv1d_update ops for better perf.
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
---------
Signed-off-by: SunnyLee219 <3294305115@qq.com>
### What this PR does / why we need it?
update triton ascend version in 3.2.0
- vLLM version: v0.13.0
- vLLM main:
d68209402d
Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com>
### What this PR does / why we need it?
In long-sequence scenarios, the chunked-prefill component may encounter
dimension misalignment issues, which previously occurred during
precision testing on the code_generate_lite dataset. This PR removes
redundant computations and instead derives the value using existing
results and straightforward calculations.
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
This PR makes `AscendMLAMetadataBuilder` and `AscendSFAMetadataBuilder`
properly inherit from the base class `MLACommonMetadataBuilder` in vllm
by adding `super().__init__()` calls.
**Changes:**
- Add `super().__init__()` call in `AscendMLAMetadataBuilder.__init__()`
- Add `super().__init__()` call in `AscendSFAMetadataBuilder.__init__()`
- Extract `ascend_chunked_prefill_workspace_size()` to
`vllm_ascend/attention/utils.py` to avoid code duplication
- Override `determine_chunked_prefill_workspace_size()` to support
Ascend-specific 128k tokens workspace size (vs 64k in parent class)
- Update unit tests to mock parent class `__init__` for proper isolation
**Why we need it:**
- Follow proper Python inheritance patterns by calling
`super().__init__()`
- Reduce code duplication by reusing parent class initialization logic
- Better maintainability as parent class changes will be automatically
inherited
Part of issue #5463 item 10
### Does this PR introduce _any_ user-facing change?
No, this is an internal refactoring that does not change any user-facing
behavior.
Signed-off-by: lico67373 <918688502@qq.com>
### What this PR does / why we need it?
add dispath_ffn_combine_bf16
- vLLM version: v0.13.0
- vLLM main:
bde38c11df
---------
Signed-off-by: guanguan0308 <1546542263@qq.com>
### What this PR does / why we need it?
In [PR 5040](https://github.com/vllm-project/vllm-ascend/pull/5040), the
`dispatch_gmm_combine_decode` operator was configured with an incorrect
global_bs parameter. This PR is to fix the bug.
The global_bs provided as input should have the same meaning as in the
`moe_distributed_dispatch` operator, specifically: (the maximum batch
size across all cards) * (expert parallel world size).
However, the implementation incorrectly used the variable
max_num_tokens, which does not account for tensor parallelism. This
error likely resulted in an unnecessarily large (overestimated) value.
More info about this operator, please refer to RFC: issue
https://github.com/vllm-project/vllm-ascend/issues/5476
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Acc
test qwen3-235b eplb on a single A3 node(ep16),
with dispatch_gmm_combine_decode
| dataset | version | metric | mode | vllm-api-stream-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 80.00 |
- vLLM version: v0.13.0
- vLLM main:
11b6af5280
Signed-off-by: wangqiankun <wangqiankun13@huawei.com>
