### What this PR does / why we need it?
Supports generalized FlashComm2 optimization, which reduces
communication overhead, decreases RmsNorm computation, and saves one
AllGather step by replacing Allreduce operations in the Attention module
with pre-AlltoAll and post-AllGather operations (used in combination
with FlashComm1). This feature is enabled during the Prefill phase and
is recommended to be used together with FlashComm1, delivering broad
performance improvements, especially in long sequence scenarios with
large tensor parallelism (TP) configurations. Benchmark tests show that
under TP16DP1 configuration, it can improve the prefill performance of
the DeepSeek model by 8% on top of FlashComm1.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
---------
Signed-off-by: zzhxx <2783294813@qq.com>
Signed-off-by: Levi-JQ <yujinqi2@huawei.com>
Co-authored-by: Levi-JQ <yujinqi2@huawei.com>
Co-authored-by: zzhxx <2783294813@qq.com>
### What this PR does / why we need it?
The code bug caused an empty bubble. When the npu_paged_cache_load
operator was called, it forcibly transferred seq_len2 to the device,
which triggered synchronization and interrupted the CPU operator's
launch stream.
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
---------
Signed-off-by: underfituu <hzhucong@163.com>
### What this PR does / why we need it?
This PR fixes a bug where the workspace of `_npu_paged_attention` in
setup is smaller than execution. For current implementation of
FULL_DECODE_ONLY with `_npu_paged_attention`, we use
`_npu_paged_attention_get_workspace` when capturing with `max_model_len`
as `seq_lens`. This assumes that PA with larger `seq_lens` inputs should
have larger workspace than smaller `seq_lens`. However, there are rare
cases where PA with smaller `seq_lens` incurs larger space. So I add
`get_workspace` directly into `update_attn_params`.
This change might introduce small(≈1%) performance degradation for low
num_tokens(such as 1) in decode phase, and there is no other known
memory issues. So I think this change is acceptable. We can remove this
if new attention op (such as `npu_fused_infer_attention_score`) does not
have such problems.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
Fixes a compatible bug with `torch_npu.npu_fused_infer_attention_score`
which is discribed in
https://github.com/vllm-project/vllm-ascend/issues/4020.
@momo609 tells us this solution.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
The environment is same with this issue,
https://github.com/vllm-project/vllm-ascend/issues/4020.
We modify the code according to
https://github.com/vllm-project/vllm-ascend/pull/3918.
And run below codes:
```python
# run with Qwen3-next-mtp
prompts = [
"Who are you?",
]
sampling_params = SamplingParams(temperature=0.0, top_p=0.95, top_k=40, max_tokens=128)
llm = LLM(model="/home/model/Qwen3-Next-80B-A3B-Instruct",
tensor_parallel_size=4,
enforce_eager=True,
distributed_executor_backend="mp",
gpu_memory_utilization=0.7,
speculative_config={
"method": "qwen3_next_mtp",
"num_speculative_tokens": 1,
},
max_model_len=4096)
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}")
```
Outputs:
```text
Prompt: 'Who are you?', Generated text: ' I am Qwen, a large-scale language model independently developed by the Tongyi Lab under Alibaba Group. I am designed to answer questions, create text such as stories, official documents, emails, scripts, and more, as well as perform logical reasoning, programming, and other tasks. If you have any questions or need assistance, feel free to let me know anytime!'
```
Now, `torch_npu.npu_fused_infer_attention_score` is compatible with
Qwen3-Next.
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
Add Add constraints for sequence parallelism for unsupported scenarios:
1. tp_size > 1
2. enable_expert_parallel must be True for MoE model
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
### What this PR does / why we need it?
- global_segment_size and local_buffer_size use constants for unified
management.
- Newly added support for input formats ending with GB, MB, KB, and B,
while being compatible with existing input methods.
### Does this PR introduce _any_ user-facing change?
- Users can use new input methods
- The documentation has also been modified
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
---------
Signed-off-by: 李子琦 <liziqi_ing@163.com>
### What this PR does / why we need it?
Refactor kv cache tensor initialization logic.
1. Unify the kvcache tensor initialization logic of deepseek and normal
models
2. spilt `initialize_kv_cache_tensors` into `_allocate_kv_cache_tensors`
and `_reshape_kv_cache_tensors`, following gpu modelrunner in vllm
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
CI passed with existing test.
1. prefill disaggregation scenario
4. deepseek + aclgraph/eager mode
5. qwen3 next
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
move quant before allgather in Allgather EP, rely on
https://github.com/vllm-project/vllm-ascend/pull/3334
Deepseek R1 W8A8 performance on A2 with
`HCCL_ALGO="level0:NA;level1:pipeline"`:
| Seq length | Mean TTFT (ms) main | Mean TTFT (ms) this PR |
|----------|----------|----------|
| 4k | 375.21 | 364.99 |
| 16k | 1465.23 | 1421.75 |
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
### What this PR does / why we need it?
Set adxl engine as the default Mooncake backend, because Ascend
Transport is no longer maintained.
Update README to include instructions for installing the adxl backend
Mooncake.
### Does this PR introduce _any_ user-facing change?
Users need to compile and install the mooncake backend for adxl
according to the revised README instructions.
### How was this patch tested?
By CI.
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
### What this PR does / why we need it?
Fix#3891.
The empty of `moe_comm_method` in the above issue is due to the wrong
check for MoE models. To be specific, the method `is_moe_model` only
checks whether a text-only model is a MoE model, without considering
multi-modal models, e.g., `VL` and `Omni`.
Check the config dict recursively to find if it has a key contains
"expert", without checking the model architecture.
It is worth noting that, we can't verify a model by if it contains
`FusedMoE` module because `is_moe_model` is called somewhere before the model loading, e.g., it's called when updating the ACLGraph config in
platform initialization.
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
---------
Signed-off-by: shen-shanshan <467638484@qq.com>
### What this PR does / why we need it?
Protect the scene where the first problem occurs. The execution should
be interrupted when the video memory application fails, rather than
waiting until an illegal address is accessed.
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
NA
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
Signed-off-by: leo-pony <nengjunma@outlook.com>
### What this PR does / why we need it?
In the `update_*attn_params` functions, the
`torch.npu.stream(update_stream)` context manager was previously located
inside the for-loop that updates parameters for each layer. This
resulted in redundant stream initiations for every layer, adding
unnecessary overhead.
This commit refactors the code by moving the stream context manager to
wrap the entire for-loop. This ensures that the update stream is
initiated only once per function call, rather than for each layer. This
change reduces 90us in each decode model.
update stream in every layer:
<img width="1720" height="383" alt="image"
src="https://github.com/user-attachments/assets/70e4cb69-5bc1-4180-a67d-c99132134be6"
/>
remove update stream in every layer:
<img width="1269" height="175" alt="image"
src="https://github.com/user-attachments/assets/0e290edb-b0ce-48fe-b032-1b924ade6ae5"
/>
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac
Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>