### 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?
This PR is aimed to fix the recomputing out of memory bug in decode
instance. When recomputing happens in decode, kv cache usage may exceed
the pre-allocated memory, and it will cause OOM.
So we propose a new scheduling strategy, when decode instance cannot
allocate new block for running requests, we will stop the request that
will be preempted. These stopped request will be recognied by proxy, and
they will be send to prefill instance again to calculate kvc and then
direct to decode instance.
This is a temporary plan to fix the bug. The long-term stratege is to
use CPU offload in decode instance.
### Does this PR introduce _any_ user-facing change?
An extra ascend configuration option **-- recompute_scheduler_enable =
True** is added to enable this strategy. The default value is False
### 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: CHEN <116010019@link.cuhk.edu.cn>
### 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?
Fix ZeroDivisionError when prefill_tp_size > num_kv_head, in this
situation, num_head_replica can be 0 and used to divide another value,
this PR restricts the minimum value of a to be 1. And this PR fix
tp_resharding README.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- 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: nwpu-zxr <zhouxuerong2@huawei.com>
Co-authored-by: liziyu <liziyu16@huawei.com>
### What this PR does / why we need it?
- Refacotr and integrate a unified `WeightPrefetchMethod`
- Integrate `gate_up_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": True,
"prefetch_ratio": {
"moe": {
"gate_up": 0.8
},
},
},
}
```
This feature is enabled by default, and can be disabled through this
configuration
### 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: yuzhup <15705211260@163.com>
This PR primarily focuses on two key changes:
1. Adjusts internal interface calls to optimize the interaction logic
between related modules.
2. Exposes an interface that allows users to select the EPLB algorithm,
enabling more flexible configuration based on specific usage scenarios.
These changes aim to enhance the usability of the system while ensuring
the stability of internal operations. Relevant unit tests have been
updated to cover the modified logic.
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: Che Ruan <cr623@ic.ac.uk>
Co-authored-by: Che Ruan <cr623@ic.ac.uk>
### What this PR does / why we need it?
Resolve the issue where, in the case of unequal TP (Tensor Parallelism),
the TP size is larger than the number of model attention kvcache heads,
causing the KV cache to generate duplicates, which leads to transmission
errors in the original code.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
By ci
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Co-authored-by: nwpu-zxr <zhouxuerong2@huawei.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>
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?
Add an option of enable frozen parameter
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
68dbde5dbb
Signed-off-by: 1Fire4 <wangdingyi2@huawei.com>
### Motivation
Currently dynamically experts balancing would stop-the-world.
Asynchronously expert load balancing would be better without flowing
problems:
Host-bound latency:
There are many cpu operations during EPLB such as
eplb-algorithm、creating p2p ops、and log2phy expert converting would
spend long cpu time, as ~1s.
Communication latency: The transfer time would cost much in the
situation without nvlink. As the weight of an expert maybe transfer to
multiple new positions, thus N times send/recv for one expert, with
result long latency. We had tested that batch_isend_irecv cost more
100ms for 16 experts weight transmission in A2 server of ascend.
SwiftBalancer would not stop-the-world anymore, in out test on NPU 1~2ms
cost for each layer while benefit 5ms-8ms decode latency with ep_size =
64.
The following updates have been made:
1、expert distribution recording with lower cost.
2、async cpu computing for eplb algo and other python operator.
3、new eplb algo with less expert rebalancing while almost the same
effect.
### Proposed Change
We will gradually migrate the EPLB logic to the VLLM community and
implement a generalized design. Relevant RFC:
https://github.com/vllm-project/vllm/issues/22246
The overall workflow involves:
<img width="801" height="302"
alt="474430541-23b06f58-23bc-44a3-a1be-00f268aeb15c"
src="https://github.com/user-attachments/assets/1d73a459-1b23-4b0a-812a-bf0a75debfed"
/>
1. Record experts distribution during forward. We using expert_token_num
after disptach instead of topk_ids, thus we got much smaller tensor
shape to reduce cost of hbm recording and add-operator.
2. Do all-gather for experts distribution. Using all-gather instead of
all-reduce as less traffic volume.
3. Wake up eplb worker process with experts distribution when
num_iterations comes. Run eplb algorithm in eplb worker.
4. Generate p2p send/recv ops and other operator such as log2phy would
cost long cpu time.
5. Lanch ibatch_send_recv in async_stream before forward.
6. After forward, wait for the ibatch_send_recv finish, then do uapte
expert map and expert weights.
### Co-author
Co-authored-by: raindaywhu raindaywhu@raindaywhu@ 163.con
Co-authored-by: njuyuan yuanjl19@smail.nju.edu.cn
Co-authored-by: qmkakaxi wjh1594260677@qq.com
Co-authored-by: Skywalker-EP 173723846@qq.com
- vLLM version: v0.10.2
- vLLM main:
567939953b
---------
Signed-off-by: offline0806 <z00858301@china.huawei.com>
Co-authored-by: offline0806 <z00858301@china.huawei.com>
### What this PR does / why we need it?
In memory of #677 , a long overdue milestone. Now DeepSeek V3/R1 should
be OK with ACL Graph.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
Working on it.
- vLLM version: v0.10.2
- vLLM main:
68dbde5dbb
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
This PR introduces Oproj matrix tensor model parallel to achieve
decreasing of memory consumption. It only support graph mode in pure DP
scenario.
In deepseek r1 w8a8 PD disagregated Decode instance, using pure DP, with
oproj_tensor_parallel_size = 8, we have 1 ms TPOT increasing, saved 5.8
GB NPU memory per RANK. We got best performance when
oproj_tensor_parallel_size=4 without TPOT increasing.
performance data:
<img width="1442" height="442" alt="image"
src="https://github.com/user-attachments/assets/83270fc5-868a-4387-b0a9-fac29b4a376d"
/>
### Does this PR introduce _any_ user-facing change?
This PR introduces one new config in `additional_config`.
| Name | Effect | Required | Type | Constraints |
| :---------------------------- |
:--------------------------------------- | :------- | :--- |
:----------------- |
| oproj_tensor_parallel_size | Split the o_proj matrix along the row
dimension (head num * head dim) into oproj_tensor_parallel_size pieces.
| No | int | default value is None, once this value is set, the feature
will be enabled, head num * head dim must be divisible by this value. |
example
`--additional_config={"oproj_tensor_parallel_size": 8}`
### How was this patch tested?
- vLLM version: v0.10.1.1
- vLLM main:
eddaafc1c7
---------
Signed-off-by: zzhx1 <zzh_201018@outlook.com>
Co-authored-by: zzh <zzh_201018@outlook.com>
### What this PR does / why we need it?
This PR ports #2312#2506#2531 to main branch.
Original implementation of torchair caching forces users to make
everything prepared, fix all the configuration and enable
`use_cached_npu_graph`, and it might cause some problems confusing to
understand and tackle for users. It is better to compile the graph twice
instead of reusing the old kvcaches and cached torchair graph. And the
extra duration time is acceptable. Additionally, this pr fixes a
recompilation problem of torchair graph mode caused by
`running_in_graph` variable in `AscendMLATorchairImpl`.
### Does this PR introduce _any_ user-facing change?
If users want to enabling torchair.cache_compile with high compilation
speed, it is recommended to enable both `use_cached_kv_cache_bytes` and
`use_cached_graph` in `torchair_graph_config`. Without
`use_cached_kv_cache_bytes`, we'll compile torchair computation graph
twice to avoid runtime error caused by configuration mismtaches (the
second compilation will be much faster). Additionally, we've made a
change to how the TORCHAIR_CACHE_HOME enviroment variable is utilized to
enhance safety and prevent accidental file deletion by adding a suffix
directory.
### How was this patch tested?
CI and e2e vllm serving pass.
- vLLM version: v0.10.1.1
- vLLM main:
70549c1245
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
support torchair mode
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.10.1.1
- vLLM main:
5438967fbc
Signed-off-by: zhangdepeng <zhangdepeng2@huawei.com>
Signed-off-by: p00465316 <panchao13@huawei.com>
Co-authored-by: zhangdepeng <zhangdepeng2@huawei.com>
### What this PR does / why we need it?
This PR introduces LMhead tensor model parallel to achieve decreasing of
memory consumption, and TPOT performance improvement. It support both
eager mode and graph mode.
In deepseek r1 w8a8 PD disagregated Decode instance, using pure DP, with
lmhead_tensor_parallel_size = 8, we have 1 ms TPOT optimization, saved
1.48 GB NPU memory per RANK.
performance data:
<img width="1444" height="438" alt="image"
src="https://github.com/user-attachments/assets/3c5ef0d3-a7c7-46fd-9797-4de728eb0cb0"
/>
### Does this PR introduce _any_ user-facing change?
This PR introduces one new config in `additional_config`.
| Name | Effect | Required | Type | Constraints |
| :---------------------------- |
:--------------------------------------- | :------- | :--- |
:----------------- |
| lmhead_tensor_parallel_size | Split the lm_head matrix along the
column dimension (vocab_size) into lmhead_tensor_parallel_size pieces |
No | int | default value is None, once this value is set, the feature
will be enabled, vocab_size must be divisible by this value. |
example
`--additional_config={"lmhead_tensor_parallel_size": 8}`
### How was this patch tested?
- vLLM version: v0.10.1.1
- vLLM main:
de533ab2a1
---------
Signed-off-by: zzhx1 <zzh_201018@outlook.com>
Co-authored-by: zhangzihang <zzh_201018@outlook.com>
### What this PR does / why we need it?
enable_shared_pert_dp is currently on by default. This optimization is
currently only valid for deepseek series models. The default opening
affects the accuracy of the qwen series models.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
use parameter --additional_config='{"enable_shared_expert_dp": true}'
- vLLM version: v0.10.0
- vLLM main:
d983769c41
Signed-off-by: Wang Kunpeng <1289706727@qq.com>
### What this PR does / why we need it?
1.Shared Expert Sharding Strategy Update: Switched from TP-aligned to
pure DP for shared experts, enabling more efficient execution.
2.O_Proj AllReduce → ReduceScatter: Reduced communication overhead by
using ReduceScatter, made possible by pure DP sharding.
3.AllGather Postponed: Delayed to after QKV down projection to reduce
synchronization impact during prefill.
### How was this patch tested?
Adding ut case in `tests/ut/attention/test_mla_v1.py`
#### How to run
use parameter `--additional_config='{"enable_shared_expert_dp": true}'`
##### a.How to run eager mode
eg:
python -m vllm.entrypoints.openai.api_server --model=/model_path
--trust-remote-code -tp 8 -dp 2 --enable_expert_parallel --port 8002
--max-model-len 5120 --max-num-batched-tokens 16384 --enforce-eager
--disable-log-requests
--additional_config='{"ascend_scheduler_config":{"enabled":true},"enable_shared_expert_dp":
true,"chunked_prefill_for_mla":true}'
##### b.How to run graph mode
eg:
python -m vllm.entrypoints.openai.api_server --model=/model_path
--trust-remote-code -tp 8 -dp 2 --enable_expert_parallel --port 8002
--max-model-len 5120 --max-num-batched-tokens 16384
--disable-log-requests
--additional_config='{"ascend_scheduler_config":{"enabled":true},"enable_shared_expert_dp":
true,"chunked_prefill_for_mla":true,"torchair_graph_config":{"enabled":true}}'
- vLLM version: v0.10.0
- vLLM main:
9edd1db02b
---------
Signed-off-by: Wang Kunpeng <1289706727@qq.com>
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
Co-authored-by: SlightwindSec <slightwindsec@gmail.com>
### What this PR does / why we need it?
In fact, the kimi-k2 model is similar to the deepseek model, and we only
need to make a few changes to support it. what does this pr do:
1. Add kimi-k2-w8a8 deployment doc
2. Update quantization doc
3. Upgrade torchair support list
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.0
- vLLM main:
9edd1db02b
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
cherry-pick #1675 to main
This PR adds validation checking to torchair_graph_config for better
reliability.
Co-authored-by: whx-sjtu <2952154980@qq.com>
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.10.0
- vLLM main:
2836dd73f1
Signed-off-by: 22dimensions <waitingwind@foxmail.com>
### What this PR does / why we need it?
Remove ETP/EP maintained in branch main. We drop this as there is no
relevant scenarios to use ETP now, and we may subsequently advocate
implementing expert tensor parallelism in vLLM to support scenarios
where the expert is needed to be sliced
This is a part of #1422 backport.
Fixes https://github.com/vllm-project/vllm-ascend/issues/1396https://github.com/vllm-project/vllm-ascend/issues/1154
### Does this PR introduce _any_ user-facing change?
We'll not maintain etp/ep in vllm-ascend anymore, and use the tp/ep in
vllm instead.
### How was this patch tested?
CI passed with new added and existing test.
- vLLM version: v0.9.2
- vLLM main:
fe8a2c544a
Signed-off-by: MengqingCao <cmq0113@163.com>
This commit
78fe77534b
from vllm reverted the change for FusedMoEParallelConfig
This PR do the same to fix the CI error
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
This PR supports torchair graph mode with non-mla backend on both 800IA2
and 300I Duo platforms. The main change is to add
`attention_v1_torchair.py` to support specific attention related
operations that are required by torchair.
### Does this PR introduce _any_ user-facing change?
Before this PR, vLLM-Ascend only allows deepseek to use torchair. Now we
can also use it with pangu. Besides, we add a support model list to
control which type of models that can use torchair.
### How was this patch tested?
We have test it with PanguProMoE on both 800IA2 and 300I Duo platforms,
and model generates answer normally.
---------
Signed-off-by: angazenn <zengyanjia@huawei.com>
Signed-off-by: tianyitang <tangtianyi4@huawei.com>
Co-authored-by: angazenn <zengyanjia@huawei.com>
Co-authored-by: tianyitang <tangtianyi4@huawei.com>
This PR added the unit test framework to enable ut for vLLM Ascend. Unit
test runs on CPU machines. It'll be ran once lint check is passed the
same as e2e test.
For unit test, this PR created a new folder called `ut` under `tests`
module. All the test file in `ut` should keep the same with the code in
`vllm-ascend`. The file name should be start with `test_` prefix. For
example, in this PR. the `test_ascend_config.py` is added for
`ascend_config.py` test.
A new fille `worker/test_worker_v1.py` is also added as the placeholder.
This file should be the unit test for `vllm-ascend/worker/worker_v1.py`.
Additional, a new `fake_weight` folder is added, it contains the
config.json from `facebook/opt-125m`, so that the test will not always
visit huggingface.
TODO:
We should add all the unit test file one by one in the future.
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Move all vector operations to a secondary stream, with the expected
overlaping being:
```
| q_rmsnorm | | kv_norm_rope_cache | | q_rope |
| matmul W_DQ | matmul W_DKV | index | index | matmul W_UQ | split | matmul W_KV_T |
```
Currently, the `IndexByTensor` operators introduced by computation of
`cos` and `sin` can't be offloaded to the secondary stream due to a
known bug of graph fusion optimization pass. So we instead keep it in
the main stream, only requires it be computed before `matmul W_UQ` to
avoid hindering later overlapping. The problem may be solved by later
optimization (#993), which hoists the computation of `cos` and `sin` up
to the first layer.
### Does this PR introduce _any_ user-facing change?
Controlled by `torchair_graph_config.enable_multistream_mla`, defaulted
to False.
### How was this patch tested?
Tested on 1x16 910 node, with tailored 2 layer DSKv2.
Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
Contains on #1111 for completeness.
<!-- Thanks for sending a pull request!
BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html
-->
### What this PR does / why we need it?
Implement multi-stream parallelism for MoE layers with shared experts,
where computation of shared experts will be overlapped with expert token
dispatch and combine. Also, when multi-stream is enabled, weights of
shared experts will be force to replicate across all cards, regardless
of any tensor parallelism configurations, to avoid AllReduce operations.
With the expected overlaping being:
```
| shared gate_up | shared act | | shared down |
| dispatch | routed gate_up, act, down | combine |
```
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.
- Please clarify why the changes are needed. For instance, the use case
and bug description.
- Fixes #
-->
### Does this PR introduce _any_ user-facing change?
No.
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
### How was this patch tested?
Tested on 1x16 910 node, with tailored 2 layer DSKv2.
<!--
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: sdmyzlp <lrwei2@petalmail.com>
### What this PR does / why we need it?
Add EPLB expert map import capabilities
### Does this PR introduce _any_ user-facing change?
When importing the EPLB expert map you need import expert map file by
vllm args additional_config
### How was this patch tested?
1.You need to collect expert hotness and generate an expert placement
file based on the hotness and the EPLB algorithm, or you can directly
use an existing expert placement table.
2.When launching vLLM, enable EC2 and pass the configuration via the
command-line argument:
--additional-config '{"expert_map_path": "/xxx/xxx/xx.json"}
Co-authored-by: songshanhu07 <1763685535@qq.com>
---------
Signed-off-by: songshanhu07 <1763685535@qq.com>
Signed-off-by: Yuxiao-Xu <664988918@qq.com>
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: songshanhu07 <1763685535@qq.com>
Co-authored-by: Xu Yuxiao <xuyuxiao2@huawei.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
View optimization in torchair (defaulted to on for Transpose with any of
its axis being 1) prevents the weight Transpose to be fused with later
GroupedMatmul, which decrease the performance of MoE layer when expert
parallelism equals the total number of experts (e.g. EP256 for DSKv3).
Add an option to solve this problem by disabling the optimization.
### Does this PR introduce _any_ user-facing change?
Controlled by
`additional_config.torchair_graph_config.enable_view_optimize`,
defaulted to `True`.
### How was this patch tested?
Tested on 1x16 910 node, with tailored 2 layer DSKv2.
Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
Fix the ascend config check logic:
1. refactor check_ascend_config to make it clear:
1. torchair graph should not work with enforce_eager=True
2. aclgraph should not work with torchair graph
3. add refresh config for rlhf case
4. fix a typo in model runner
5. change expert_tensor_parallel_size default to 0 to keep the same as
before
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Support MOE inner Multi-stream for Deepseek.
This feature requires graph mode with mc2 enabled.
---------
Signed-off-by: David9857 <985700846@qq.com>
More and more config options are added to additional_config. This PR
provide a new AscendConfig to manage these config options by an easier
way to make code cleaner and readable.
This PR also added the `additional_config` doc for users.
Added the test_ascend_config.py to make sure the new AscendConfig works
as expect.
TODO: Add e2e test with torchair and deepseek once the CI resource is
available.
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
