### What this PR does / why we need it?
There is a zero-like operator before the attention operation in each
decoding stage. After analysis, this operator can be eliminated. The
purpose of this PR is to remove this operator and improve performance.
---------
Signed-off-by: ZYang6263 <zy626375@gmail.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?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
Signed-off-by: hust17yixuan <303660421@qq.com>
### What this PR does / why we need it?
Fix mtp tochair bug cuased by #2719
Since FIA need extra space for padding, we need to enforce
`self.max_num_seqs > self.scheduler_config.max_num_seqs` in KV consumer
+ MTP
This means that, `self.max_num_seqs` **>** the actual maximum requests
(`self.scheduler_config.max_num_seqs`)
### 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: xuyexiong <xuyexiong@huawei.com>
### What this PR does / why we need it?
we notice that `patch_main` is never used. Usually the patch is for all
version. And if it's for specified version, we can use `vllm_version_is`
instead. So let's remove the useless sub folder in patch module to make
it clear.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
**Problem Description:**
The existing implementation for the w4a8-dynamic linear method only
supports the old quantization format from msmodelslim. When attempting
to load models quantized with the new version, vLLM encounters errors
due to mismatched tensor shapes and unprocessed quantization parameters.
Relavant issues:
- https://github.com/vllm-project/vllm-ascend/issues/3192
- https://github.com/vllm-project/vllm-ascend/issues/3152
**Proposed Changes:**
1. Add support for w4a8 dynamic(new format) in
AscendW4A8DynamicLinearMethod and TorchairAscendW4A8DynamicLinearMethod
2. Add unit tests and e2e tests for w4a8 dynamic new and old format
models
<details>
<summary><b>details</b></summary>
1. **Support for new w4a8-dynamic format:**
* Detects quantization format by reading the "version" field in
quant_description to ensure backward compatibility.
* Handles the new pre-packed weight format (`2x int4` in an `int8`),
which has a halved dimension. It tells the vLLM loader how to unpack it
using `_packed_dim` and `_packed_factor`.
* Supports the new `scale_bias` parameter, setting its shape based on
the layer type, as required by msmodelslim. For api consistency and
future use, the `layer_type` parameter was also added to other
quantization methods.
* Updates the weight processing logic: new format weights are handled
with `.view(torch.int32)` since they're pre-packed, while old ones are
processed with `npu_convert_weight_to_int4pack`.
2. **New unit and E2E tests:**
* Added unit tests that verify the logic for both the old and new
formats.
* Split the distributed E2E test to confirm that both old and new format
models work correctly.
</details>
Theoretically, these changes will provide support for all common new
version w4a8(dynamic) models from msmodelslim.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
I implement relevant unit tests and e2e tests and test the changes with
following commands:
```bash
# unit tests
python -m pytest tests/ut/quantization/test_w4a8_dynamic.py tests/ut/torchair/quantization/test_torchair_w4a8_dynamic.py -v
# e2e tests
pytest tests/e2e/singlecard/test_quantization.py -v -s
pytest tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen3_W4A8DYNAMIC_new_version -v -s
pytest tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen3_W4A8DYNAMIC_old_version -v -s
pytest tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_W4A8DYNAMIC -v -s
```
I also tested Hunyuan-1.8B-Instruct quantized with the new w4a8-dynamic
format:
```
vllm serve ./models/Hunyuan-1.8B-Instruct-quantized --gpu-memory-utilization 0.96 --quantization ascend --max-model-len 9600 --seed 0 --max-num-batched-tokens 16384
```
All tests mentioned passed locally.
**NOTE: I use quantization model from my own repo in
test_offline_inference_distributed.py**. Here is the description:
[Anionex/Qwen3-1.7B-W4A8-V1](https://modelscope.cn/models/Anionex/Qwen3-1.7B-W4A8-V1/summary)
(including quantization steps).This should be replaced by a model in
vllm-ascend ci modelscope repo.
Thanks for reading!
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: Anionex <1005128408@qq.com>
### What this PR does / why we need it?
The `force_attention` parameter is designed for flash infer kernel
warmup, we don't actually need it on Ascend device (at least for
now).And it tends to make things more complicated. So we replace the
`force_attention` parameter with `aclgraph_runtime_mode` in the
attention metadata creation logic.
This change makes the control flow more explicit by directly using the
graph runtime mode to determine how to build attention metadata, rather
than relying on an intermediate boolean flag. This simplification
removes redundant logic and clarifies the conditions for building
attention metadata for full decode graph mode.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
DP + `FULL_DECODE_ONLY` + online serving.
- 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?
Port #1916 and #2157 to master branch to fuse operators in deepseek moe
layers, which can reduce scheduling overhead on devices. Note that this
feature is valid only when `tp_size = 1` and
`multistream_overlap_shared_expert` is enabled with torchair graph mode.
### Does this PR introduce _any_ user-facing change?
Users can enable this feature with `--additional-config
'{"torchair_graph_config":{"enabled":true, "enable_super_kernel":true},
"multistream_overlap_shared_expert":true}'`.
### How was this patch tested?
E2E deepseek serving with 2P1D disaggregated prefill scenarios.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
Using non-blocking operations for device-to-host transfers can lead to
data corruption in later steps. The CPU tensor is accessed right after
the transfer is triggered, but the transfer might not be complete yet.
As a result, the data could be wrong. This problem was seen in the A3
environment during `profile_run`.
### How was this patch tested?
CI pass.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>
This PR deletes model codes of deepseek_v2 and deepseek_v3 to reuse the
model file from vLLM.
vLLM Ascend now uses custom ops register way instead of model file
hard-coding.
- 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?
when using dynamic eplb, patch v1 executor to avoid create child process
failed.
### How was this patch tested?
deepseek in v3.
- 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?
Optimize torchair kv_consumer padding logic. Only pad when it is spec
decoding
### 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: xuyexiong <xuyexiong@huawei.com>
This PR adds support for redundant experts in the EPLB.
Key points:
- Use global_num_experts = num_experts + num_redundant_experts
consistently.
- Backward compatible when num_redundant_experts=0.
Tested
On a 16-rank setup (W8A8) with static EPLB and expert_map_path,
verifying router logits shape and successful requests.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: yechao237 <yechao20180411@gmail.com>
### What this PR does / why we need it?
In memory of https://github.com/vllm-project/vllm-ascend/pull/2610 and
#3449 Fix Mtp torchair pd bug.
In the pd Disaggregation scenario, the first token of the inference
after the d node receives the kv follows the eager mode.
Fixes:
Running with MTP torchair graph mode with Prefilling Decoding
Disaggregation , if all requests processed by the D node are requests
just transmitted from the P node, it will break the torchair graph.
Reason: During PD Disaggregation , the P node only transmits the KV
cache and prompt to the D node, not the actual tokens inferred (neither
the main model tokens nor the MTP tokens are transmitted). Therefore,
the D node will treat this request as one without MTP tokens for
inference (seq_len=1).
The community does not have graph mode issues because the community's
attention has a seq_len=1 for each batch during the decode phase.
We have issues because the graph mode pads according to processing 2
tokens per request. When there are some seq_len=1 and some seq_len=2,
padding is done at the end. If all requests received by the D node are
seq_len=1, padding cannot be performed normally according to the
attention's fia operator constraints.
Solution:
The kv consumer uses extra torchair graph padding to avoid breaking FIA
graph constrains (The one this PR implemented).
### 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: xuyexiong <xuyexiong@huawei.com>
This reverts commit b0ae203e72.
### What this PR does / why we need it?
The fix is not ready yet, conflict with #3411 need to revert first. Will
fix this issue later
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
Signed-off-by: xuyexiong <xuyexiong@huawei.com>
### What this PR does / why we need it?
In memory of https://github.com/vllm-project/vllm-ascend/pull/2610
In the pd Disaggregation scenario, the first token of the inference
after the d node receives the kv follows the eager mode.
Fixes:
Running with MTP torchair graph mode with Prefilling Decoding
Disaggregation , if all requests processed by the D node are requests
just transmitted from the P node, it will break the torchair graph.
Reason: During PD Disaggregation , the P node only transmits the KV
cache and prompt to the D node, not the actual tokens inferred (neither
the main model tokens nor the MTP tokens are transmitted). Therefore,
the D node will treat this request as one without MTP tokens for
inference (seq_len=1).
The community does not have graph mode issues because the community's
attention has a seq_len=1 for each batch during the decode phase.
We have issues because the graph mode pads according to processing 2
tokens per request. When there are some seq_len=1 and some seq_len=2,
padding is done at the end. If all requests received by the D node are
seq_len=1, padding cannot be performed normally according to the
attention's fia operator constraints.
Solution:
The kv consumer uses extra torchair graph padding to avoid breaking FIA
graph constrains (The one this PR implemented).
The kv producer provides the correct tokens to the kv consumer, so that
our graph mode constraints are not broken, and all logic is the same as
the PD mixed deployment . Since we are using the community scheduler,
the modification requires patching the vllm scheduler, but
theoretically, performance should be better. (Maybe later )
### 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: xuyexiong <xuyexiong@huawei.com>
### What this PR does / why we need it?
Adapt deepseek-v3.2 to vllm 0.11.0, removing the useless patch.
The final goal is to remove all the patches and align the code arch to
vllm, thus we need to do the following work in next prs.
TODO:
- [x] remove patch on attention spec
- [ ] refactor the kvcache creation logic
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
1. CI passed with existing test.
2. Test pass with deepseek-v3.2-exp
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: MengqingCao <cmq0113@163.com>
What this PR does / why we need it?
1.Record expert map without dynamic eplb.
2.Add export PYTHONOPTIMIZE=1 when using dynamic eplb.
3.change eplb doc
Does this PR introduce any user-facing change?
How was this patch tested?
Qwen3_moe in A3.
- vLLM version: 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?
Avoid printing some warning msg as below :
UserWarning: To copy construct from a tensor, it is recommended to use
sourceTensor.clone().detach ...
### 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: SunnyLee219 <3294305115@qq.com>
### What this PR does / why we need it?
when infer deepseek mtp layer with multistream_moe, we should pass a
boolean to evaluate this feature and fix bugs when we are in mtp layer
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
### What this PR does / why we need it?
Fix the issue of missing NZ conversion for quantized weights in GMM
after moe_dispatch operator in torchair scenario, which does not involve
aclgraph & single scenarios.
### How was this patch tested?
vllm serving passed with lower latency (~5ms TPOT with bs_per_rank=28 &
ep_size=32)
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
Currently, when executing to the Linear layer of models in vLLM-Ascend,
the weights format is ND in unquantized case and skipped ascend case.
This PR supplements the execution logic for Linear layer. We use a new
global variable: VLLM_ASCEND_ENABLE_NZ. When VLLM_ASCEND_ENABLE_NZ=1 and
CANN version is 8.3, the weights of the Linear layer will be converted
to FRACTAL_NZ, in both unquantized case and skipped ascend case. We also
use VLLM_ASCEND_ENABLE_NZ to control the existing NZ conversion, such as
w8a8-quantized case.
### Does this PR introduce _any_ user-facing change?
Add a new global variable VLLM_ASCEND_ENABLE_NZ. If you want to use NZ
format, you should set VLLM_ASCEND_ENABLE_NZ=1.
### 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: anon189Ty <Stari_Falcon@outlook.com>
This PR will accomplish the following tasks:
**optimize SP**
In the old version implementation, the first layer was all_reduce, which
used rms to split chunks. We changed it to perform reduce_scatter on the
embedding side, replace one all_reduce operation and one chunk with one
reduce_scatter operation.
**Support qwen3 next**
Since Qwen3 Next includes a linear attention module, the prefix name of
this module cannot take effect directly.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
### What this PR does / why we need it?
When using dynamic eplb, moe load is not imported. We fix this problem
by modifying the return value of hidden states in torchair.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
DeepseekV3 in A3.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: daishixun <dsxsteven@sina.com>
### What this PR does / why we need it?
When using dynamic eplb,it will be blocking by nz tensor.We fix these
prolems by clone src tensor and recv tensor.
### Does this PR introduce any user-facing change?
### How was this patch tested?
Qwen3_moe in A3.
- 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?
when ops torchair_fused_experts_with_mc2 is called, we need pass a tp
group, but now it only pass when quantized scenario, we need also pass
it when unquantized.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
### What this PR does / why we need it?
Resolved the issue of EPLB failure caused by changes in the log2phy map
due to device type modifications when using MTP rotation position
encoding.
### Does this PR introduce any user-facing change?
### How was this patch tested?
https://github.com/vllm-project/vllm/commit/releases/v0.11.0
- vLLM version: 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?
- Refacotr and integrate a unified `WeightPrefetchMethod`
- Integrate `qkv_proj.weight` and `o_proj.weight` in quantized Attention
modules
- Prefetching these weights ahead of matmul-like operators imporves
performance by reducing L2 cache transfer latency
### Does this PR introduce _any_ user-facing change?
Add a new config in `--additional-config` for configuration:
```json
{
"weight_prefetch_config": {
"enabled": false,
"prefetch_ratio": {
"attn": {
"qkv": 1.0,
"o": 1.0,
},
},
},
}
```
This feature is enabled by default, and can be disabled through this
configuration
### How was this patch tested?
- vLLM version: v0.11.0
---------
Signed-off-by: yuzhup <15705211260@163.com>
Signed-off-by: zhoux77899 <zhouxiang100@huawei.com>
Co-authored-by: yuzhup <15705211260@163.com>
### What this PR does / why we need it?
the multistream moe in tochari only validate in decode, but can't be
applied to chunked prefill, So add some judgments to isolate the
scenario
### 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?
1. Move additional functionalities from fused_moe.py to
common_fused_moe.py and remove fused_moe.py
2. Remove unnecessary custom classes from qwen3_moe.py, and it will be
completely removed after we release vllm-ascend v0.11.0
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Qwen3-30B-A3B/Qwen3-30B-A3B-W8A8/DeepSeek-V3-W4A8-Pruing/deepseek-mtp/pangu-pro-moe-pruing:
1. Enable/Disable EP
3. Aclgraph & eager
4. SP
- vLLM version: v0.11.0
---------
Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
Co-authored-by: weijinqian0 <12153182+weijinqian0@users.noreply.github.com>
### What this PR does / why we need it?
1.Support deepseek w4a8 per-channel quantization
2.The eager mode supports converting weights to the NZ format
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
#### How to get weights using Modelslim
##### Installation steps
git clone https://gitcode.com/Ascend/msit.git
cd msit/msmodelslim
bash install.sh
##### Generate w4a8 per-channel weights
cd /example/DeepSeek
Command reference: msmodelslim/example/DeepSeek/README.md
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
---------
Signed-off-by: Wang Kunpeng <1289706727@qq.com>
### What this PR does / why we need it?
Upgrade vLLM to newest commit
- Fix the aclgraph doesn't work problem, caused by
24fab45d96
- Fix PoolerOutput import error, caused by
755ed7b05b
- Fix the aclgraph weight load error to keep the same with torchair fix.
4492e3a554
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
All test should pass
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
---------
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Remove useless PD check in deepseek
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Co-authored-by: wangxiaoteng <wangxiaoteng@huawei.com>
### What this PR does / why we need it?
To cut down the memory usage of large weight matrices, we often rely on
various linear operations:
- `ReplicatedLinear`: Stores the entire matrix, consuming excessive
memory.
- `RowParallelLinear`: Requires an `all_reduce` to merge answer,
introducing additional communication overhead and potential accuracy
loss. Each token is handled across multiple devices rather than a single
device, which is undesirable in SP scenario.
- ...
Furthermore, in multi-way Data Parallelism (DP) configurations, layers
typically store redundant weight copies.
This PR introduces a shared-weight plugin for layers inheriting from
`LinearBase`. It offers the following advantages:
- It evenly distributes a set of layers with identical structures across
devices. Each layer retains its complete weights, eliminating redundant
memory usage.
- It supports asynchronous broadcasting to prefetch weights for upcoming
layers.
- It preserves the custom `process_weights_after_loading()` method to
make keeping NZ format possible.
- It is compatible with any linear class that inherits from
`LinearBase`, thereby preserving all the features of the original linear
implementation.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
vLLM main:
f4a948f33f
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
---------
Signed-off-by: clrs97 <524936896@qq.com>
Co-authored-by: CalvinXKY <kyxiezju@163.com>
### What this PR does / why we need it?
Fixes#3096
1. Fix kv cache initialization error of attention layer. There are some
models with layer name like `attn.attn`, instead of `self_attn`, but the
initialization of kv cache tensors only check for `self_attn` and
`attn.attn`, which leding to the error `AssertionError: Some layers are
not correctly initialized`
2. Set the default value of input arg `sampling_metadata` in
`compute_logits` for the modeling files in vllm-ascend. Thus fixing the
error `Qwen3NextForCausalLM.compute_logits() missing 1 required
positional argument: 'sampling_metadata'`
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
test locally with internlm
- vLLM version: v0.10.2
- vLLM main:
5aeb925452
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
What this PR does / why we need it?
there are two sets of sp implementations for moe and dense models. One
is called sequence_parallelism, and the other is flashcomm_v1.
We did the following things：
Merge two sets of code with the same implementation into one.
Remove the implementation of sequence_parallelism, as this solution
cannot support aclgraph.
Does this PR introduce any user-facing change?
No
How was this patch tested?
e2e&ut
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
---------
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
### What this PR does / why we need it?
This miscellaneous contains several small fixes:
1) fix initialization and forward bugs of DeepseekMTPLayer with
`shared_expert_dp` enabled.
2) fix a tensor shape mismatches after o_proj caused by a work-aroud
change in NPUModelRunner.
3) avoid unnecessary decline of kv_cache memory (default: 64MB) with
`use_cached_kv_cache_bytes` disabled.
4) fall back `fused_moe_state` from `MC2` to `All2All` since the padding
logic of `mc2_mask` is incompatible with input hidden_states when
`shared_expert_dp` enabled.
Once this PR is merged, users can launch disaggregated_prefill
deployments (large_ep) with `deepseek_mtp` and `shared_expert_dp` as
`v0.9.1-dev` branch. The remaining problem of kv_cache tokens decline
compared to `v0.9.1-dev` will be resolved by
https://github.com/vllm-project/vllm-ascend/pull/3073.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E vllm serving about deepseek_mtp with torchair graph mode and
`enable_shared_expert_dp` with eager mode. Large ep deployments are also
tested with this PR.
- vLLM version: v0.10.2
- vLLM main:
5aeb925452
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Note: This depends on [vLLM
#25161](https://github.com/vllm-project/vllm/pull/25161) and the
torch\_npu release from September 30.
### What this PR does / why we need it?
This pull request adds `FULL_DECODE_ONLY` mode for GQA/MHA models (MLA
models like DeepSeek V3/R1 are not included). Key improvements include:
* **Reduced dispatch latency:** By replaying the entire model execution
graph at once, we cut overhead compared with multiple smaller replays.
* **Stabilized multi-device performance:** Captureing the whole model as
one static graph also mitigates the dispatch fluctuations across
devices.
* **Stream/resource savings:** Consolidating graph captures frees up
streams, allowing more graphs to be captured.
**Known issues:**
1. `_npu_paged_attention` currently manages its own workspace in
`torch_npu`, which can deadlock when synchronizing during graph replay —
we’re working on a fix.
There may be other corner cases. This PR is the first in a planned
series; we’ll continue to iterate and address remaining issues in
follow-ups.
This is essentially a port of #1503 and #1677, but includes two major
changes:
1. Let `graph_dispatcher` decide the graph mode instead of hard-coding
it in the backend, which decouples Full Graph and Piecewise Graph and
could make it possible to remove dynamo.
2. Adapt to the new `attn_group` logic, but leave a small hack in
`update_graph_params`; multi-attention models may or may not be fully
supported yet.
### Does this PR introduce _any_ user-facing change?
```python
compilation_config={
"cudagraph_mode": "FULL_DECODE_ONLY",
},
```
### How was this patch tested?
Tests included.
- vLLM version: v0.10.2
- vLLM main:
9607d5eb44
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
While running quantized deepseek models with unquantized MTP layer, free
NPU memory abnormally decreases for `2*HCCL_BUFFSIZE` bytes. This
results from the wasted VRAM buffer allocation casued by calling
`dist.all_to_all_single` without correct device process group argument.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
We run vllm online serving with quantized deepseek-r1 and unquantized
MTP layer, and observed that free_memory increased without redundat VRAM
buffer for HCCL communication op (all_to_all_single).
- vLLM version: v0.10.2
- vLLM main:
6d8246aaff
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
This PR prepares for deleting this enviroment variable,
`VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE`, as vllm requires `fullgraph=True`
to run
- Fixes https://github.com/vllm-project/vllm/issues/21834
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
See CI
- vLLM version: v0.10.2
- vLLM main:
99cc41ad50
---------
Signed-off-by: Lucas Kabela <lucaskabela@meta.com>
### What this PR does / why we need it?
The speculative decode phase of chunkedprefill has taken an incorrect
path, should always use TND layout for speculative decoding.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
6d8246aaff
Signed-off-by: xuyexiong <xuyexiong@huawei.com>
This PR puts the calculation of shared experts into a separate stream,
overlaping with routing experts.
- vLLM version: v0.10.2
- vLLM main:
fbd6523ac0
---------
Signed-off-by: whx-sjtu <2952154980@qq.com>
### What this PR does / why we need it?
Remove chunked prefill for mla branch in mla , and change dtype of
prefill_mask to avoid accuracy problem
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
ef7eefe17a
---------
Signed-off-by: SunnyLee219 <3294305115@qq.com>
