### What this PR does / why we need it?
bugfix for mtp fullgraph
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main:
83f478bb19
Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
### What this PR does / why we need it?
When using multi connector, the multi connector does not define
get_finished_count, which will cause the kv cache to be released
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main:
83f478bb19
---------
Signed-off-by: baxingpiaochong <771405853@qq.com>
### What this PR does / why we need it?
fix a typo in mooncake layerwise connector. There is only `requests`,
instead of `request` in `connector_metadata`. This pr fixes this typo
- vLLM version: v0.11.0rc3
- vLLM main:
83f478bb19
Signed-off-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
To adapt the torch_npu version to avoid the precision problem of
torchair deepseek. The torch_npu version may result in the different
branches in the ops register, the rms_norm ops has two branches
according to the verson_check, this pr unify the rms_norm in torchair by
patching quant_rms_norm to rms_norm to fix the accuracy issue in torchair scenario
- vLLM version: v0.11.0rc3
- vLLM main:
83f478bb19
Signed-off-by: hust17yixuan <303660421@qq.com>
### What this PR does / why we need it?
After refactoring vllm_ascend/models and FusedMoE, we are unable to pass
`gate` from deepseekv2.py to `AscendFusedMoE.forward`, which will result
in error when running deepseek v3/r1 with allgather.
Hence, this pr removes `gate` related computations from FusedMoE module
in eager/aclgraph mode.
### Does this PR introduce _any_ user-facing change?
`rm_router_logits` is deprecated in eager/aclgraph.
### How was this patch tested?
e2e & ut
- vLLM version: v0.11.0rc3
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.1
Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
### What this PR does / why we need it?
Part of https://github.com/vllm-project/vllm-ascend/pull/3106
Fix Hybrid kvcache sharing bug in same attention type
Change the `shared_by` logic so that the same attention spec could share
the same buffer instead of allocating more hbm.
After this pr, kvcache memory saved 50% in qwen3-next compared with
before (`self_attn:linear_attn=1:3` in an `attn_group`), and
`gpu_memory_utilization` could increase to `0.8` on Qwen3-Next when
running on A2 64G/card with tp4
<img width="2833" height="1540" alt="image"
src="https://github.com/user-attachments/assets/2a91fa99-fb0f-447c-9e8b-acd587890fbe"
/>
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
Test pass with the latest e2e test case on qwen3-next
- vLLM version: v0.11.0rc3
- vLLM main:
c9461e05a4
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
force with_prefill true after allreduce in kv producer
- vLLM version: v0.11.0rc3
- vLLM main:
c9461e05a4
---------
Signed-off-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
We have optimized the performance of long sequences：First,Modify the
input data format for attention calculation. Instead of using the
original BSND format, remove the logic for converting between TND and
BSND, and directly adopt the TND format.
The TND input format can be directly reused, which shortens the data
flow path. Converting to BSND is an unnecessary processing step.Second,
we switched the output update of the concatenated small operators to the
npu_attention_update fusion operator to improve performance.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main:
c9461e05a4
---------
Signed-off-by: pichangping <1337510399@qq.com>
### What this PR does / why we need it?
The current MatmulReduceScatter operator experiences performance
degradation in small-shape scenarios, so it determines whether to use
this operator by judging the size of the shape.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.1
---------
Signed-off-by: ZYang6263 <zy626375@gmail.com>
### What this PR does / why we need it?
dcp pcp support full aclgraph, including mla attention_v1
- vLLM version: v0.11.0rc3
- vLLM main:
c9461e05a4
Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
### What this PR does / why we need it?
The cache for MLA decode graph parameters was holding strong references
to tensors, preventing them from being garbage collected and leading to
increased memory usage.
This change wraps the cached tensors in weak references, allowing them
to be deallocated when no longer in use and reducing overall memory
pressure.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
None.
- vLLM version: v0.11.0rc3
- vLLM main:
c9461e05a4
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
Fix Qwen3NextGatedDeltaNet, caused by
https://github.com/vllm-project/vllm/pull/26437
### How was this patch tested?
```
def main():
prompts = [
"窗前明月光，",
"The president of the United States is Mr.",
"The capital of France is",
"The future of AI is",
"感时花溅泪，",
"家书抵万金啥意思？",
"plz tell me a story: ",
]
# Create a sampling params object.
sampling_params = SamplingParams(max_tokens=100, temperature=0.6, top_k=40, top_p=0.95)
# Create an LLM.
llm = LLM(
model="/root/.cache/modelscope/hub/models/Qwen/Qwen3-Next-80B-A3B-Instruct",
tensor_parallel_size=4,
enforce_eager=True,
trust_remote_code=True,
max_model_len=256,
gpu_memory_utilization=0.7,
block_size=64
)
# Generate texts from the prompts.
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}")
```
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: Icey <1790571317@qq.com>
### What this PR does / why we need it?
Remove codes of dbo.
Currently, vLLM has supported dbo with pr:
https://github.com/vllm-project/vllm/pull/23693.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main:
17c540a993
Signed-off-by: zzzzwwjj <1183291235@qq.com>
### What this PR does / why we need it?
Caps the calculated maximum number of tokens at 512.
This prevents allocating an excessively large buffer when a cudagraph
capture size is not specified, mitigating the risk of out-of-memory
errors.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
None.
- vLLM version: v0.11.0rc3
- vLLM main:
17c540a993
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
1. Rename common_fused_moe.py to fused_moe.py.
2. Rename fused_moe_prepare_and_finalize.py / FusedMoEPrepareAndFinalize
to prepare_finalize.py / PrepareAndFinalize.
3. Rename vllm_ascend/ops/moe to vllm_ascend/ops/fused_moe.
4. Move vllm_ascend/ops/fused_moe.py to
vllm_ascend/ops/fused_moe/fused_moe.py
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
e2e & ut
- vLLM version: v0.11.0rc3
- vLLM main:
17c540a993
Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
### What this PR does / why we need it?
We optimized the _prepare_input method in eagle_proposer and no longer
use the _prepare_eagle_input_sequential method, improving the
performance of eagle-3.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
```
python3 -m vllm.entrypoints.openai.api_server
--host 0.0.0.0
--port 13963
--dtype bfloat16
--model meta-llama/Llama-3.1-8B-Instruct
--served-model-name Llama-3.1-8B-Instruct
--tensor-parallel-size 1
--gpu-memory-utilization 0.85
--max-model-len 32768
--trust-remote-code
--seed 42
--no-enable-prefix-caching
--speculative_config '{"method":"eagle3","model":"yuhuili/EAGLE3-LLaMA3.1-Instruct-8B","num_speculative_tokens":2,"draft_tensor_parallel_size":1}'
```
Co-authored-by: QilaiZhang (245706640@qq.com )
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: lio <1983142975@qq.com>
### What this PR does / why we need it?
Since Attention and LinearAttention share the same ```slot_mapping```,
and the ```slot_mapping``` for LinearAttention is all zeros, the
```slot_mapping``` for Attention gets overwritten, resulting in the
computed output being all zeros.
This PR removes the uniformly managed ```self.slot_mapping``` and
directly passes the ```slot_mapping``` from ```input_batch.blocktable```
to ```attn_metadata```, along with modifying the relevant references.
Due to hardware, the data type of ```block_table.slot_mapping``` needs
to be set to int32.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed with existing test.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: QilaiZhang <245706640@qq.com>
This PR fix the bug related with running multi-modal models with
AscendScheduler. This bug was introduced by PR #2372 by using the same
parameter names as vLLM with different default values.
Currently I fix this bug by changing the default values of these two
parameters to align with vLLM.
- vLLM version: v0.11.0rc3
- vLLM main:
17c540a993
Signed-off-by: hw_whx <wanghexiang7@huawei.com>
Co-authored-by: hw_whx <wanghexiang7@huawei.com>
### What this PR does / why we need it?
Modify the recalculation logic to prevent waiting requests from filling
up the D node KVCache
- vLLM version: v0.11.0rc3
- vLLM main:
17c540a993
Signed-off-by: underfituu <hzhucong@163.com>
### What this PR does / why we need it?
This PR refactors the Ascend attention implementation to align with
vLLM's core interfaces, simplifying the code and improving
maintainability.
### Key Changes:
* **Align with vLLM's Attention Interface**: The `forward` method
signature in `AscendAttentionBackendImpl` now matches the base
`AttentionImpl` in vLLM, removing the custom `trace_flag`.
* **Enable Opaque Attention Operator**: By adding `opaque_attention_op`
to `AscendPlatform`, we allow vLLM to wrap our attention kernel in its
standard `vllm.unified_attention_with_output` operator. This avoids the
need for a custom call path.
* **Remove Obsolete Code**:
* The custom op `vllm.unified_ascend_attention_with_output` has been
deleted as it is now redundant.
* The `trace_flag` and its associated logic were removed, reducing code
complexity.
* An outdated quantization branch within the attention implementation
was cleaned up.
* **Improve Readability**: Renamed output variables (`output` vs.
`intermediate_output`) and added comments to clarify the in-place nature
of the attention output.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
No extra tests needed.
- vLLM version: v0.11.0rc3
- vLLM main:
17c540a993
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
This PR boosts performance by introducing a fused kernel for the matrix
matmul and reduce scatter operations. It supports both unquantized
(e.g., BFloat16) and W8A8 quantized models.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: ZYang6263 <zy626375@gmail.com>
### What this PR does / why we need it?
In multi-Tensor Parallel (TP) scenarios, the KV pool only queries the
first GPU card. When keys on other cards are released, the query result
still returns as successful, introducing accuracy issues. This PR
modifies the KV pool's query logic to check all cards, resolving this
problem.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: fems14 <1804143737@qq.com>
### What this PR does / why we need it?
In certain scenarios, the performance of synchronously loading data from
the pool is better than that of asynchronously loading data. Therefore,
a control logic (or switch) for asynchronous loading from the pool has
been added.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: fems14 <1804143737@qq.com>
### What this PR does / why we need it?
fix bug: In the mooncake pooling scenario, when the client closes the
request, the server fails to locate the request, leading to the server
hanging.oling scenario, when the client closes the request, the server
fails to locate the request, leading to the server hanging.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
Pull up the PD separated pooling service, send requests using aisbench,
press CTRL+C twice, and check if the vllm_ascend service exit.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: linhebiwen <linhebiwen@gmail.com>
### What this PR does / why we need it?
Check all expert maps when using muilty instance.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
Qwen 235B in double A3.
case1：master has expert map, slave has not expert map.
case2: master has expert map, slave has error expert map.
case3: master has expert map,slave has correct expert map.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: offline0806 <3337230449@qq.com>
Co-authored-by: offline0806 <3337230449@qq.com>
### What this PR does / why we need it?
`vanilla_chunked_prefill_mla` and `vanilla_decode_mla` is unused, so
remove it.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: zzzzwwjj <1183291235@qq.com>
This PR moves the communication operation of shared experts out of extra
stream because I found that this might cause rtMemcpy related errors
when running shared experts multistream with aclgraph.
Furthermore, I utilize a global variable as extra stream object to avoid
allocating streams for each layer in full-graph mode.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
On Arm systems, os.sched_yield() does not take effect, causing the GIL
(Global Interpreter Lock) to remain unrelinquished and resulting in CPU
bound issues. This PR applies a patch to sched_yield in vLLM, making the
process execute time.sleep(0) instead to release the GIL.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: fems14 <1804143737@qq.com>
### What this PR does / why we need it?
The #3624 PR fix the precision of deepseek torchair, but don't consider
the limitation of torch compile which results in the recompile, This PR
fixs this problem
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: hust17yixuan <303660421@qq.com>
Resolves a `TypeError: got an unexpected keyword argument 'layer_type'`.
A recent change (PR #3311) started passing the `layer_type` argument
when calling `get_pergroup_param()`. This specific implementation does
not use this parameter, causing the error.
This patch adds `layer_type=None` to the method signature to maintain
API compatibility and ignore the unused argument.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
### What this PR does / why we need it?
Fix mooncake connector. In scenarios where TP is not equal, when the
prefill TP size is less than the number of key-value heads,
_get_remote_tp_ranks_for_req will return a list of np.arrays. Performing
an operation like int in list of np.arrays will cause an error.
Converting the list of np.arrays into a single np.array resolves this
issue.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
qwen235B
P tp16, D tp1
P tp8, D tp1
P tp4, D tp1
P tp8, D tp2
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: liziyu <liziyu16@huawei.com>
Signed-off-by: underfituu <hzhucong@163.com>
Co-authored-by: underfituu <hzhucong@163.com>
### What this PR does / why we need it?
This PR introduces a new model loader called Netloader, which leverages
high-bandwidth P2P direct transfer between NPU cards to achieve weight
loading. Netloader is implemented as a plugin through the newly added
'register_model_loader' function in vLLM 0.10. It facilitates the
process of weight loading by sending weights from a pre-loaded model
(server) to an empty model of a newly started instance (client). The
server operates concurrently with normal inference tasks through
sub-threads and the 'stateless_init_torch_distributed_process_group' in
vLLM. The client initiates a transfer request after verifying that the
model and partitioning method are the same as the server's, and uses
HCCL's collective communication (send/recv) to load the weights in the
order they are stored in the model.
Application Scenarios:
1. Significantly Reduces Inference Instance Startup Time By reusing the
weights of already loaded instances and performing high-speed transfers
directly between computing cards, this method reduces model loading
latency compared to traditional remote/local pull methods.
2. Reduces Network and Storage Pressure Avoids the need to repeatedly
download weight files from remote repositories, reducing the impact on
centralized storage and network traffic, thereby enhancing overall
system stability and service quality.
3. Improves Resource Utilization and Reduces Costs Accelerating the
loading process reduces reliance on redundant computing pools, allowing
computing resources to be elastically scaled and reclaimed as needed.
4. Enhances Business Continuity and High Availability In fault recovery
scenarios, new instances can quickly take over existing services,
avoiding prolonged business interruptions and improving the system's
high availability and user experience.
### Does this PR introduce _any_ user-facing change?
Netloader utilizes the existing --load-format=netloader and
--model-loader-extra-config to be activated. The
model-loader-extra-config needs to be input as a JSON string (as it is
now)
Afterwards, you can check whether the outputs for the same sentence are
consistent when the temperature is set to 0.
Signed-off-by: destinysky <kangrui10@126.com>
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: destinysky <kangrui10@126.com>
### What this PR does / why we need it?
The precision of deepseek torchair is broken by #3465 , which due to the
origin patch or rmsnorm in torchair. This PR fixes the precision of
deepseek torchair
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: hust17yixuan <303660421@qq.com>
### What this PR does / why we need it?
This PR refactors SequenceRowParallelOp forward. In order to further
expand the operator inclusion scope in dynamic judgment scenarios, this
PR customizes the entire matmul computation and communication as a
custom operator masking. With this refactor, it will support directly
writing code such as common operation fusion into the
`SequenceRowParallelOp` class's member function `matmul_and_reduce`,
without the need to register more redundant custom masking operators.
### How was this patch tested?
CI passed with existing test.
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
- `qkv_proj.weight` prefetching has been implemented with `Quant` op,
when `AddRmsNormQuant` is enabled (#3465) `qkv_proj.weight` prefetching
won't work
- Implement `qkv_proj.weight` prefetching with `AddRmsNormQuant`
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
Tested on `Qwen3-235B-A22B-W8A8`
<img width="1868" height="109" alt="image"
src="https://github.com/user-attachments/assets/0bc28082-0287-4d5c-b8f6-f907c3134d36"
/>
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: zhoux77899 <zhouxiang100@huawei.com>
### What this PR does / why we need it?
Move the creation of dummy attention metadata to occur after the ACL
graph runtime mode is determined. This ensures the metadata is
initialized with the correct configuration during a profile run.
Additionally, remove the `attn_metadata` existence check before updating
MLA attention parameters. This change prevents the update from being
skipped when metadata is not yet available, ensuring parameters are set
correctly.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
None.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
Corrects the attribute access for retrieving the device from `q_a_proj`
to `q_proj`. This prevents an `AttributeError` as `q_a_proj` does not
exist on the class instance.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
Need MLAPO tests.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
