### What this PR does / why we need it?
This PR fix the input constraints checks for the mlapo and bmm_transpose
operators.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
### Perf
64K/3K，1P1D，bs=32
before this pr:
TPOT 29ms, TTFT 47s，TPS 606 token/s
after this pr:
TPOT 29ms, TTFT 48s，TPS 636 token/s
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
1. Fix DeepSeek-V3.2-W8A8-Pruning mtp
2. Add DeepSeek-V3.2-W8A8-Pruning e2e test
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
11b6af5280
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
### What this PR does / why we need it?
Based on the RFC:https://github.com/vllm-project/vllm-ascend/issues/5604
This PR is a refactoring of vllm_ascend/distributed.
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
11b6af5280
Signed-off-by: lty <linhebiwen@gmail.com>
### What this PR does / why we need it?
Fix XliteModelRunner init failed when aclgraph is enabled. Ensure
function graph_capture of vllm.v1.worker.gpu_model_runner is replaced.
- vLLM version: v0.13.0
- vLLM main:
bde38c11df
Signed-off-by: changdawei1 <changdawei3@huawei.com>
### What this PR does / why we need it?
1. Rename num_iterations_eplb_update to expert_heat_collection_interval.
2. Rename num_wait_worker_iterations to algorithm_execution_interval.
3. Rename init_redundancy_expert to num_redundant_experts because the
variable with the same meaning in vLLM is named this way.
4. Delete gate_eplb because we don't need this feature.
5. Move eplb config into a dict in additional config.
6. Depend on pr5817
### Does this PR introduce _any_ user-facing change?
before this pr：
`--additional-config '{"dynamic_eplb":true,
"num_iterations_eplb_update": 4000, "num_wait_worker_iterations": 150,
"init_redundancy_expert": 16, "expert_map_path": "xxx.json"}'`
after this pr:
`--additional-config
'{"eplb_config":{"dynamic_eplb":true,"expert_heat_collection_interval":4000,
"algorithm_execution_interval":150,"num_redundant_experts": 16,
"expert_map_path": "xxx.json"}}'`
### How was this patch tested?
#### test qwen3-235b eplb num_redundant_experts=16
without pr5817
| dataset | version | metric | mode | vllm-api-general-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 83.33 |
with pr5817
| dataset | version | metric | mode | vllm-api-general-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 86.67 |
- vLLM version: v0.13.0
- vLLM main:
45c1ca1ca1
Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
### 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:
bde38c11df
---------
Signed-off-by: Zetong Li <slippersss@126.com>
### What this PR does / why we need it?
Fix acc bug when enbale dispatch_gmm_combine_decode and eplb.
After eplb, expert table may change, so mapping is needed, while
fused_mc2 miss the mapping.
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?
without this pr, qwen3-235b eplb with dispatch_gmm_combine_decode get
acc 3.33% on aime2024.
with this pr,
test qwen3-235b eplb on a single A3 node(ep16)
without dispatch_gmm_combine_decode
| dataset | version | metric | mode | vllm-api-stream-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 86.67 |
with dispatch_gmm_combine_decode
| dataset | version | metric | mode | vllm-api-stream-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 86.67 |
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: wangqiankun <wangqiankun13@huawei.com>
### What this PR does / why we need it?
we implement set_additional_forward_context in platform, it's necessary
to reuse code of gpu in model runner v2 by inheriting method in gpu
model runer v2. please see model runner v2's plan #5208
### 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: Ronald1995 <ronaldautomobile@163.com>
### What this PR does / why we need it?
Based on the RFC:https://github.com/vllm-project/vllm-ascend/issues/5604
This PR is a refactoring of vllm_ascend/distributed, moving all
kv_transfer realtaed codes into a dedicated folder, which has already
been done in vLLM
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: lty <linhebiwen@gmail.com>
### What this PR does / why we need it?
support dsv3.2 enable both mtp and full_decode_only
PR5626 To align with the community, the branch logic was modified.
Previously, dsv32 could not reach inside the branch, and now an
additional unpadded step is required, which causes transformations in
positions and num_input_tokens, leading to changes in the cos and sin
dimensions in sfa_v1.py. This, in turn, causes an illegal shape error
when passed to the operator.
1. The unpadded function is introduced to align with the community， and
in the community the function does not have the parameters
num_input_tokens and positions.
2. The positions are split and num_input_tokens=num_actual_tokens are
used to correspond to the function name unpad, so that the padded
positions and num_input_tokens are not output.
However, in fact, attention_v1 does not use the above two parameters.
This is done because we are concerned that some people might use these
parameters later and encounter shape mismatch issues if they are not
aware of this. Therefore, we have performed the cropping.
From the perspective of the source of acquisition, positions are not
cropped, so there is actually no need to add unpad in this case.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: cookieyyds <126683903+cookieyyds@users.noreply.github.com>
### What this PR does / why we need it?
When there is no kv cache in some devices, the `_compute_prefill_context
func` will return `None`, which is unexecpted. This PR replaces None
with full zeros/-inf tensors to avoid TypeError.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
```bash
pytest tests/e2e/multicard/4-cards/long_sequence/test_chunked_prefill.py -k test_models_chunked_prefill_with_empty_kvcache
```
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
This PR removes the unused `make_attention_mask` function from
`vllm_ascend/worker/v2/attn_utils.py`.
**Why it's dead code:**
- After PR #4870 (attention mask unification refactor), attention mask
generation has been centralized in the `AttentionMaskBuilder` singleton
class
- The mask is now generated directly by metadata builders when needed
(e.g., `AscendAttentionMetadataBuilder`, `AscendMLAMetadataBuilder`)
- The `make_attention_mask` function is no longer called anywhere in the
codebase
- The function's parameters (including `attn_mask` and `spec_attn_mask`)
were also removed from `build_attn_metadata` in the same refactor
**Changes:**
- Remove `make_attention_mask` function (24 lines) from
`vllm_ascend/worker/v2/attn_utils.py`
### Does this PR introduce _any_ user-facing change?
No. This is a code cleanup that removes dead code. No user-facing
behavior changes.
### How was this patch tested?
- Verified that `make_attention_mask` is not called anywhere in the
codebase (via `grep`)
- CI tests pass to ensure no regressions
- The function has been unused since PR #4870 was merged
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: lico67373 <918688502@qq.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
### What this PR does / why we need it?
this pr implement eagle spec decoding for model runner v2, please see
RFC https://github.com/vllm-project/vllm-ascend/issues/5208
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
vLLM version: v0.13.0
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
### What this PR does / why we need it?
While using the LLM Compressor quantization tool from the VLLM community
to generate quantized weights, the VLLM Ascend engine needs to be
adapted to support the compressed tensors quantization format.
1. Support Moe model W8A8 Int8 dynamic weight.
2. Specify W4A16 quantization configuration.
Co-authored-by: menogrey 1299267905@qq.com
Co-authored-by: kunpengW-code 1289706727@qq.com
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: LHXuuu <scut_xlh@163.com>
Signed-off-by: menogrey <1299267905@qq.com>
Signed-off-by: Wang Kunpeng <1289706727@qq.com>
Co-authored-by: menogrey <1299267905@qq.com>
Co-authored-by: Wang Kunpeng <1289706727@qq.com>
### 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:
2f4e6548ef
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
Fix the bug where the P-node's schedule dead after it force-frees a
request due to timeout and then receives the completed kv cache pulled
by the D-node again. By add list to recode all requests.
- vLLM version: release/v0.13.0
- vLLM main:
81786c8774
---------
Signed-off-by: liziyu <liziyu16@huawei.com>
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Co-authored-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
Fix layerwise connector for decoder tp size > num kv heads. In this case
prefiller should push kv cache to all decoder npu.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
- Use upstream util function (`_pre_process()` and `_post_process()`) to
reduce redundant codes. (Find more details at
https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/rotary_embedding/common.py#L184-L213)
- Merge Q/K split to simplify the logic of calling
`torch_npu.npu_rotary_mul()` for better performance (TPOT has been
reduced by **6.22%**).
### Does this PR introduce _any_ user-facing change?
no.
### How was this patch tested?
#### ✅ Functional test
Launch the server:
```bash
export VLLM_USE_MODELSCOPE=True
vllm serve /root/.cache/modelscope/hub/models/Qwen/Qwen2.5-VL-7B-Instruct \
--dtype bfloat16 \
--limit-mm-per-prompt '{"image": 1}' \
--max-model-len 16384 \
--max-num-batched-tokens 16384
```
Query the server:
```bash
curl -X POST http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-VL-7B-Instruct",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": [
{"type": "image_url", "image_url": {"url": "https://modelscope.oss-cn-beijing.aliyuncs.com/resource/qwen.png"}},
{"type": "text", "text": "What is the text in the illustrate? How does it look?"}
]}
],
"max_tokens": 100
}'
```
Output:
```
{"id":"chatcmpl-b2911ab6989ef098","object":"chat.completion","created":1768202780,"model":"/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-VL-7B-Instruct","choices":[{"index":0,"message":{"role":"assistant","content":"The text in the illustration is \"TONGYI Qwen.\" The word \"TONGYI\" is written in blue, and \"Qwen\" is written in gray. The text appears to be part of a logo or branding design, with \"TONGYI\" being more prominent and \"Qwen\" being slightly smaller and positioned below it. The font style is modern and clean, with \"TONGYI\" having a slightly bolder appearance compared to \"Qwen.\"","refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":null,"reasoning_content":null},"logprobs":null,"finish_reason":"length","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":78,"total_tokens":178,"completion_tokens":100,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null}
```
#### ✅ Benchmark
Run:
```bash
export VLLM_USE_MODELSCOPE=False
export HF_ENDPOINT="https://hf-mirror.com"
vllm bench serve \
--model /root/.cache/modelscope/hub/models/Qwen/Qwen2.5-VL-7B-Instruct \
--backend openai-chat \
--endpoint /v1/chat/completions \
--dataset-name hf \
--hf-split train \
--dataset-path lmarena-ai/vision-arena-bench-v0.1 \
--num-prompts 10 \
--no-stream
```
Before this PR:
```
============ Serving Benchmark Result ============
Successful requests: 10
Failed requests: 0
Benchmark duration (s): 5.96
Total input tokens: 7191
Total generated tokens: 996
Request throughput (req/s): 1.68
Output token throughput (tok/s): 167.05
Peak output token throughput (tok/s): 261.00
Peak concurrent requests: 10.00
Total token throughput (tok/s): 1373.16
---------------Time to First Token----------------
Mean TTFT (ms): 964.43
Median TTFT (ms): 858.48
P99 TTFT (ms): 1691.45
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 63.08
Median TPOT (ms): 40.86
P99 TPOT (ms): 241.30
---------------Inter-token Latency----------------
Mean ITL (ms): 40.16
Median ITL (ms): 33.61
P99 ITL (ms): 250.30
==================================================
```
After this PR:
```
============ Serving Benchmark Result ============
Successful requests: 10
Failed requests: 0
Benchmark duration (s): 5.71
Total input tokens: 7191
Total generated tokens: 996
Request throughput (req/s): 1.75
Output token throughput (tok/s): 174.45
Peak output token throughput (tok/s): 279.00
Peak concurrent requests: 10.00
Total token throughput (tok/s): 1433.95
---------------Time to First Token----------------
Mean TTFT (ms): 992.14
Median TTFT (ms): 938.30
P99 TTFT (ms): 1728.71
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 59.16
Median TPOT (ms): 37.65
P99 TPOT (ms): 234.89
---------------Inter-token Latency----------------
Mean ITL (ms): 36.55
Median ITL (ms): 30.73
P99 ITL (ms): 170.72
==================================================
```
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: shen-shanshan <467638484@qq.com>
### What this PR does / why we need it?
To support dispatch_ffn_combine ep32 enabled
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Single operator tested
---------
Signed-off-by: lhchg <lhao_cheng@163.com>
come from: https://github.com/vllm-project/vllm-ascend/issues/5463
Reason:
During the iteration process of the hardware version, there may be a
large number of iterations for the operators, which can lead to
short-term compatibility differences. Therefore, an intermediate
adaptation layer is provided to accommodate the short-term differences
in operators.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Signed-off-by: weijinqian0 <1184188277@qq.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
[Refactor] Modify the binding logic to allocate CPU cores for each NPU
card
### What this PR does / why we need it?
Modify the binding logic to allocate CPU cores for each NPU card based
on NUMA affinity, while isolating acl_thread/release_thread and other
processes to prevent mutual interference.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
c85cc045f8
Signed-off-by: rowzwel_dx <1392851715@qq.com>
- vLLM version: v0.13.0
- vLLM main:
7157596103
Signed-off-by: Rozwel-dx <1392851715@qq.com>
### 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:
45c1ca1ca1Fixesvllm-project/vllm#31345
Signed-off-by: zhaomingyu <zhaomingyu13@h-partners.com>
Co-authored-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
this pr support use triton mrope like cuda_forward, which performance is
equal to ascendc ops
this triton ops should use cann 8.5.0
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
test in qwen3-vl-235b acc textvqa
native 81.82
npu triton 81.58
cuda triton 81.52
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: shiyuan680 <917935075@qq.com>
### What this PR does / why we need it?
When running memcache in the A2 environment, the logic for registering
memory needs to be added. Additionally, there is a link establishment
conflict between memcache and HCCS during initialization in A2, so the
link should be established in advance.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103
---------
Signed-off-by: fangjianwei <f30058701@china.huawei.com>
Co-authored-by: fangjianwei <f30058701@china.huawei.com>
### What this PR does / why we need it?
Add docstrings for Metadata and MetadataBuilder classes in the attention
module to improve code readability.
Related to #5463 (Item 11: Add some comments for CommonMetadata and
others)
**Modified files:**
- `vllm_ascend/attention/context_parallel/common_cp.py`: Added comments
for `AscendPCPMetadata`, `CPChunkedContextMetadata`,
`AscendMetadataForPrefill`, `AscendMetadataForDecode`
- `vllm_ascend/attention/utils.py`: Added comments for
`AscendPrefillContextParallelMetadata`
- `vllm_ascend/attention/mla_v1.py`: Added comments for
`ChunkedContextMetadata`, `AscendMLADecodeMetadata`
- `vllm_ascend/attention/attention_v1.py`: Added comments for
`AscendMetadata`, `AscendAttentionMetadataBuilder`
- `vllm_ascend/attention/context_parallel/attention_cp.py`: Added
comments for `AscendAttentionCPMetadataBuilder`
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Documentation only, no functional changes.
Signed-off-by: lico67373 <918688502@qq.com>
### What this PR does / why we need it?
This PR addresses an issue in piecewise graph mode when Multi-Threading
Parallelism (MTP) is enabled. Specifically, the original dummy run
sequence performs the following steps in order:
1. Sync DP (input length = 1 + k)
2. Dispatch (input length = 1 + k, with padding==graph size)
However, in the model execution phase, the sequence differs, resulting
in:
1. Padding (input length = 1, with padding)
2. Sync DP (input length = 1 + k)
3. Dispatch (input length 1 + k != graph size 1 + k, with padding)
This discrepancy leads to a mismatch between the input sizes used in the
model execution and those expected by the dispatch graph, causing an
inconsistency in graph size.
This PR ensures that the dispatch graph size aligns correctly by
modifying the sequence of operations during model execution to match the
dummy run sequence, resolving the mismatch issue.
### 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: LiuYi-UP <1150854440@qq.com>
### What this PR does / why we need it?
In the chunked prefill scenario, CP needs to align the
`max_context_chunk` to the `cp_virtual_block_size`, but the current
implementation only aligns it to the `block_size`. For
PD-disaggregation, `cp_kv_cache_interleave_size` is typically set equal
to `block_size`, in which case `cp_virtual_block_size=block_size *
dcp_size * pcp_size`. Under specific conditions, this can lead to
misalignment of certain chunks, subsequently triggering assertion check
errors.
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
when graph mode is picewise，replay by synchronize will be effect
performance, sync almost cost 250us
### Does this PR introduce _any_ user-facing change?
only sync when graph mode contain full mode
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: wangyongjun <wangyongjun7@huawei.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
### All-gather KV Cache for Communication Overlap:
- This PR adjusts the calculation order in the SFA.
- split `index_select` into `indexer_select_pre_process` and
`indexer_select_post_process`.
- Combine `nope`, `rope` and `index-k` into a tensor to perform
asynchronous all-gather.
### benchmark:
input=40k && num_batch_token=20k
- before:
```
Mean TTFT (ms): 2614.52
Median TTFT (ms): 3148.03
P50 TTFT (ms): 3148.03
P90 TTFT (ms): 3163.48
P99 TTFT (ms): 3170.20
```
- after:
```
Mean TTFT (ms): 2529.92
Median TTFT (ms): 3051.69
P50 TTFT (ms): 3051.69
P90 TTFT (ms): 3067.31
P99 TTFT (ms): 3072.15
```
### 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>
### What this PR does / why we need it?
The condition for determining padding in the fullgraph overlay with MTP
and PCP has been modified to accommodate corner cases where the shape
capture size is manually specified.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
ut and tests
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
### What this PR does / why we need it?
Add new function to mooncake layerwise connector, including:
1. supports sparse attention, for DeepSeek-V3.2
2. Distribute transfer tasks to redundant kv_head cards
This PR is related to [[RFC]: CDCP Scheduling for Disaggregated
Prefilling with KV Cache Layerwise Push
Support](https://github.com/vllm-project/vllm-ascend/issues/4842)
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
Signed-off-by: liziyu <liziyu16@huawei.com>
Co-authored-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
This PR fixes Minimax model loading in vLLM Ascend backend by:
Adding model type check for "minimax" and "minimax_m2" to replace "mlp"
prefix with "block_sparse_moe"
Implementing special handling for Minimax expert layer naming
conventions
Adding Minimax configuration to packed_modules_model_mapping for proper
qkv_proj and experts module handling
Without these changes, Minimax models fail to load on Ascend devices due
to incompatible layer naming and module packing.
### Does this PR introduce _any_ user-facing change?
Yes. Users can now successfully load and run Minimax models on Ascend
hardware with vLLM. This enables inference capabilities for this model
family on Ascend devices.
### How was this patch tested?
Local Testing:
Verified model loading for minimax-xxx and minimax_m2-xxx model variants
on Atlas 800I A2 hardware
Tested inference with sample prompts using vLLM's OpenAI-compatible API
server
Benchmark Validation:
Compared throughput and latency metrics against GPU baseline
Verified memory usage stays within expected limits for different batch
sizes
Tested multi-card inference scenarios with tensor parallelism
- vLLM version: v0.13.0
- vLLM main:
8be6432bda
---------
Signed-off-by: Feng-xiaosuo <tengchang1@huawei.com>
### What this PR does / why we need it?
Fix the PCP port mapping error issue.In a multi-node PD separation
scenario, when the PCP feature is enabled, there is an issue with the
ZMQ transmission port. Specifically, the IP and port received by Side D
do not match. The cause of this issue is an error in the port mapping
update strategy logic.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
By ci
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
1.Fixed memory retention on certain GPUs caused by missing PUT
operations.
2.Fixed performance degradation resulting from architectural
incompatibilities in the underlying refactor.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: fems14 <1804143737@qq.com>
### What this PR does / why we need it?
support mxfp8 quantization (qwen liner layer)
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: wangyao <iwangyao@outlook.com>
### What this PR does / why we need it?
This PR fixes the precision issue from improper Tensor maintenance in
`vllm_ascend/ops/linear_op.py` under the Verl reinforcement learning
(RL) scenario. issue:
https://github.com/vllm-project/vllm-ascend/issues/5747
Key changes:
1. Remove the custom class member `self.weight_t` in
`vllm_ascend/ops/linear_op.py`;
2. Adjust the input logic of the `npu_mm_all_reduce_base` operator to
directly fetch weight parameters from the model's `nn.Parameters`,
instead of using pre-created Tensors.
> In the vllm model, it is recommended to avoid creating additional
parameter copies (such as self.weight_t) for computation; if already
created, they must be synchronized with the model's original parameters.
This is because parameter synchronization between training and inference
in the Verl reinforcement learning (RL) scenario may cause memory
address changes to nn.Parameters, and unsynchronized extra Tensors will
reference old memory without updating with the parameters—ultimately
leading to precision issues.
### Does this PR introduce _any_ user-facing change?
No.
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
Signed-off-by: icerain-alt <450125138@qq.com>
Co-authored-by: Shangwei-Li <lishangwei@mail.ustc.edu.cn>
### What this PR does / why we need it?
The customized ascend operator sgmv_expand and sgmv_shrink applies only
to the scenario where rank is 8,16,32,64. When rank >= 128, the operator
is out of range, causing the model to report an error.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
Depends on this commit https://github.com/vllm-project/vllm/pull/31408
- vLLM version: release/v0.13.0
- vLLM main:
254f6b9867
---------
Signed-off-by: ZT-AIA <1028681969@qq.com>
Signed-off-by: ZT-AIA <63220130+ZT-AIA@users.noreply.github.com>
### What this PR does / why we need it?
To support tensorList for dispatch_ffn_combine, to adjust eplb
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Single Operator Testing
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: lhchg <lhao_cheng@163.com>
Co-authored-by: lihaocheng <lihaosheng1@h-partners.com>
### What this PR does / why we need it?
This PR fixes a bug in Xlite
backend(https://atomgit.com/openeuler/GVirt/issues/1), The direct cause
of the problem is that the XModel::PrepareAttn function obtained an
illegal number of tokens to be inferred, -540. This illegal value is due
to the padding feature of inference in graph mode and the residual state
across steps. This issue is triggered when a prefill request is newly
added in a step and a decode ends simultaneously. It is first fixed
using num_decode_tokens instead of attn_metadata.num_decodes.
1. In graph mode, vllm_ascend has padding characteristics. In the
_prepare_inputs function, if the number of tokens to be inferred is less
than the set threshold (8 in this case), the attn_metadata.num_decode
array will be expanded to 8.
2. Meanwhile, vllm_ascend uses the class variable self.query_start_loc
of NPUModelRunner to record the tokens to be inferred. Due to poor
coordination with the graph mode padding mechanism when crossing steps,
in some cases (such as when a decode request is completed in a certain
step and a new prefill request is added at the same time), negative
values may be calculated for attn_metadata.query_lens.
3. After type conversion, the negative values in query_lens cause an
overflow. Xlite detects that the number of tokens to be inferred for the
decode request is too large and triggers a "decode len too long" alert.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Same with https://atomgit.com/openeuler/GVirt/issues/1
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: wwwumr <1127858301@qq.com>
**What this PR does / why we need it?**
This PR fixes a bug in NetLoader
[PR#2888](https://github.com/vllm-project/vllm-ascend/pull/2888). The
bug was caused by
[PR#3612](https://github.com/vllm-project/vllm-ascend/pull/3612)
([1/N][Refactor] Refactor code to adapt with vllm main), which removed
the `stateless_init_device_torch_dist_pg` function from platform.py,
leading to a failure in the call. This PR adds a way to create a
stateless process group that does not depend on external code.
**Does this PR introduce any user-facing change?**
No
**How was this patch tested?**
Same with
[PR#2888](https://github.com/vllm-project/vllm-ascend/pull/2888)
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
---------
Signed-off-by: destinysky <kangrui10@126.com>
### What this PR does / why we need it?
The rotary algorithm in deepseek indexer should be neox-style instead of
gptj style. PR #4413 fix this accuracy bug with new triton kernel. This
PR fixes original pytorch version.
### Does this PR introduce _any_ user-facing change?
None
### How was this patch tested?
CI passed with existing test.
- vLLM version: 86e178f7c4d8c3b0eaf3c8e3f810a83f63b90e24
- vLLM main:
86e178f7c4
Signed-off-by: whx-sjtu <2952154980@qq.com>
