### What this PR does / why we need it?
Running multimodal model with ascend scheduler may cause assert error
【assert (request.num_tokens - request.num_computed_tokens) == 1】
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
17b4c6685c
---------
Signed-off-by: fan2956 <zhoufan53@huawei.com>
### What this PR does / why we need it?
- Fixes the bug that Multiple calls (maybe >100) to eagle3-qwen3-8b often incurs "attn_mask index out of range" error
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
```
python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --served-model-name Eagle3 --port 8000 --model Qwen/Qwen3-8B --seed 42 -tp 1 --speculative_config '{"model": "Tengyunw/qwen3_8b_eagle3", "draft_tensor_parallel_size": 1, "num_speculative_tokens": 5, "method": "eagle3"}'
```
Co-authored-by: liuruijin17
[ricklrj@outlook.com](mailto:ricklrj@outlook.com)
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
Signed-off-by: Icey <1790571317@qq.com>
### What this PR does / why we need it?
1. Solved the issue where sizes capture failed for the Qwen3-32b-int8
model when aclgraph, dp1, and tp4 were enabled.
2. Added the exception thrown when sizes capture fails and provided a
solution
3. Add this common problem to the FAQ doc
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
ut
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.0
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
### What this PR does / why we need it?
Relying on #3044, this PR aims to further fix:
1. The forward error occured when `LogitsProcessorWithLoRA` calls
`AscendLogitsProcessor.forward`. Since `LogitsProcessorWithLoRA`
bypasses the MRO to call it, `super().forward(...)` in
`AscendLogitsProcessor.forward` will raise an error. This PR fixes it by
directly invoking `LogitsProcessor.forward(self, ...)`;
2. The shape mismatch in `add_lora_logits` in punica_npu.py. The
`lora_a_stacked` and `lora_b_stacked` are organized as [num_loras, 1,
lora_rank, hidden_size] and [num_loras, 1, vocab_size, lora_rank] shapes
respectively, but they are misunderstood in #1583---the last two
dimensions were assumed in reverse order, which causes errors in
`bgmv_shrink` and `bgmv_expand`. This PR fixes it by reverting it to the
previous version to align with the implementation in punica_cpu.py in
vllm.
### Dependencies
This PR depends on changes introduced by #3044 (LoRA support for
`AscendQKVParallelLinear` and `AscendMergedQKVParallelLinear` layers).
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
The LoRA-related tests, e.g., test_ilama_lora.py and
test_ilama_lora_tp2.py, use ilama-3.2-1B, and this model is regarded as
`TransformersForCausalLM`, where `embedding_modules` attribute lacks
`lm_head`. However, `LlamaForCausalLM` and most other models include
both `embed_tokens` and `lm_head` in `embedding_modules`. This attribute
contributes to `supported_lora_modules` when using LoRA in vllm.
Therefore, without `lm_head` in `embedding_modules`, current tests using
ilama-3.2-1B are unable to find the abve errors since
`LogitsProcessorWithLoRA` replacing `lm_head` is skipped. Simply using
Meta-Llama-3.1-8B-Instruct can reproduce the above errors and check
whether these fixes can work. What's more, it's necessary to add more
comprehensive tests for LoRA.
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
Signed-off-by: Zetong Li <slippersss@126.com>
### What this PR does / why we need it?
Fix quant_config input parameter bug in qwenvl series. Currently,
non-instantiated variables should be passed.
### Does this PR introduce _any_ user-facing change?
None
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.0
Signed-off-by: booker123456 <945658361@qq.com>
### What this PR does / why we need it?
Add vLLM 0.11.0 release hourly job to monitor release branch changes
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.0
Signed-off-by: Yikun Jiang <yikunkero@gmail.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?
- Pin vLLM commit to releases/v0.11.0 branch.
- Fix the break change by vLLM commit
d4d9899860
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
17b4c6685c
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
LoRA e2e test uses ilama-3.2-1B model. It uses transformers.py model
files. Its self-attention layer names end with "\*.attn", not
"\*.self_attn".
There are some other model attention layer names end with "*.attn", such
as baichuan.py, bert.py.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
pytest -sv tests/e2e/singlecard/test_ilama_lora.py
pytest -sv tests/e2e/multicard/test_ilama_lora_tp2.py
- vLLM version: v0.10.2
- vLLM main:
17b4c6685c
---------
Signed-off-by: paulyu12 <507435917@qq.com>
### What this PR does / why we need it?
Add e2e test related to weight updates in RL scenarios.
Due to CI issues, the newly added Python test files cannot locate the
correct path. As a temporary solution, use absolute paths to add test
cases.
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: Shangwei-Li <lishangwei2@huawei.com>
### What this PR does / why we need it?
Addresses a bug in DenseOptimRowParallelOp that occurs when tensor
parallelism is not used
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
### What this PR does / why we need it?
fix bugs when mtp>1, and reorder input batch when mtp is not accepted.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
by ci
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
---------
Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
### What this PR does / why we need it?
fix oom in aclgraph.
1. In the current token dispatch implementation, tensors are mounted on
class instances to facilitate parameter passing between different
methods. This approach prevents automatic recycling of these tensors. In
some cases, it may lead to out-of-memory error. To address this issue,
we manually set these tensors to None to release corresponding memory.
2. The `profile_run` method is designed to accurately estimate the
maximum NPU memory usage during vLLM inference. However, in certain
scenarios, MoE models perform inference via MC2, which includes
communication and consumes additional NPU memory. This leads to
inaccurate estimation by the profile run. We address this by actively
triggering the MC2 during profile run for initialization.```.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
Signed-off-by: WithHades <244036962@qq.com>
### What this PR does / why we need it?
PR #2894 make ascend_scheduler_config.enabled always be `True` for
non-mla models，when `ascend_scheduler_config.enabled=True `, it will
always initialize `AscendScheduler` which is a subclass of `Scheduler`,
but when we enbale async_scheduling,we need to initialize
`AsyncScheduler` in vllm, this will make async_scheduling can't be
enabled.
### Does this PR introduce _any_ user-facing change?
not-related
### How was this patch tested?
when user set `async_scheduling`, it means user don't want to use
`AscendScheduler`, so we shouldn't set `ascend_scheduler_config.enabled
= True`
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
What this PR does / why we need it?
The Qwen3 moe MC2 graph currently has two redundant computational
operator implementations. After npu_moe_distribute_dispatch_v2, the
cumsum and cast operations have been added. By using
expert_token_nums_type=0 and not converting weight_scale to float32,
these two operators can be eliminated, thereby improving inference
performance.
Does this PR introduce any user-facing change?
No
How was this patch tested?
No need
vLLM version: v0.10.2
vLLM main:
f225ea7dd9
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
---------
Signed-off-by: florenceCH <gaoxiang120@huawei.com>
Co-authored-by: florenceCH <gaoxiang120@huawei.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>
…to avoid unintentional copy ops blocking across different NPU streams,
improving disagg TTIT/TTFT (#2788)"
### What this PR does / why we need it?
This reverts commit 6995a7bc5b. We'll add
it back once the issue is fixed.
related issue: https://github.com/vllm-project/vllm-ascend/issues/3195
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
### What this PR does / why we need it?
This PR is for the adaptation and optimization of qwen3_vl and
qwen3_vl_moe on the Ascend platform.
### Does this PR introduce _any_ user-facing change?
None
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
b1068903fd
---------
Signed-off-by: booker123456 <945658361@qq.com>
### What this PR does / why we need it?
`ready` label now is used for trigger full e2e test now. If a PR is
ready and merge conflict then, no need to drop the ready label.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Just a github action change. No need for function test.
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Revise the EPLB feature guide content.Add eplb params to ascend config.
### Does this PR introduce any user-facing change?
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
Co-authored-by: offline0806 <3337230449@qq.com>
### What this PR does / why we need it?
It is a quick bugfix for the memory explosion issue that requires
further refactoring.
The dummy_run in eager mode may lead to OOM and the reason is that
`hidden_states` were not released in time.
The PR temporarily resolves the issue by manually clearing the cache,
and further refactoring will be conducted subsequently.
Before the modification, the dummy_run's memory showed an accumulation
issue.
<img width="1796" height="207" alt="image"
src="https://github.com/user-attachments/assets/05e2b04c-2f99-4085-9eda-c78b7d9a57b0"
/>
After modification, it can be observed that the memory is released
promptly.
And it was verified that the model responded normally after a single
data input.
- vLLM version: v0.10.2
- vLLM main:
b1068903fd
---------
Signed-off-by: chenmenglong <chenmenglong1@huawei.com>
### What this PR does / why we need it?
Correct the vllm interface e2e test Base container image name
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
Tests in vllm ci pipeline
- vLLM version: v0.10.2
- vLLM main:
52d0cb8458
Signed-off-by: leo-pony <nengjunma@outlook.com>
Upgrade vLLM to newest commit.
1. Remove the useless func get_state_cls, it has been removed from vLLM
already.
e6750d0b18
2. Fix ut broken by
6160ba4151
- vLLM version: v0.10.2
- vLLM main:
b1068903fd
---------
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
vllm-ascend support [msMonitor
](https://gitcode.com/Ascend/mstt/tree/master/msmonitor)tool to collect
performance of vllm-ascend
### Does this PR introduce _any_ user-facing change?
1.add env MSMONITOR_USE_DAEMON；
2.user cann enable msMonitor tool by setting MSMONITOR_USE_DAEMON=1
before run vllm-ascend model；
3.MSMONITOR_USE_DAEMON and VLLM_TORCH_PROFILER_DIR cannot both set
### How was this patch tested?
1.run vllm-ascend model while not set MSMONITOR_USE_DAEMON=1 or set
MSMONITOR_USE_DAEMON=0, model will run successfully;
2.run vllm-ascend model while set MSMONITOR_USE_DAEMON=1, run msMonitor
tool to collect profile data;
3.run vllm-ascend model while set MSMONITOR_USE_DAEMON=1 and
VLLM_TORCH_PROFILER_DIR, will raise error
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
Signed-off-by: mei-feiyao <1332490378@qq.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>
When MTP and oprojTP are enabled, it triggers the recompilation of the
torchair graph, leading to a decrease in performance, and this PR fixes
this issue.
- vLLM version: v0.10.2
- vLLM main:
486c5599e3
---------
Signed-off-by: zzhx1 <zzh_201018@outlook.com>
### What this PR does / why we need it?
Add OOT platform E2E test case to be run in the vllm buildkite pipeline.
Note: added test case is not run in vllm-ascend CI.
### Does this PR introduce _any_ user-facing change?
NA
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
Signed-off-by: leo-pony <nengjunma@outlook.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?
When processing a mix of large and small requests, the TTFT of responses
is significantly reduc\ed. Please refer to
https://github.com/vllm-project/vllm/pull/10235, which achieves the same
effect by simply limiting the number of prompt fills for long requests.
This solution can be applied to both AscendScheduler (V0) and vLLM
Scheduler (V1). Tests show that TTFT can be significantly improved when
handling such mixed requests. However, This capability is currently
missing when Ascend Scheduler is enabled.
This benchmark used the Qwen3-8B model, with a context length of 128K,
running on a single card.
Regarding dataset selection, the sharegpt_clean dataset is used, with
its content concatenated and cropped. Small requests with token=50 and
medium requests with token=10240 were constructed (there were also large
requests with token=102400, but these were ignored because when using
the Prefill First scheduling strategy, max_num_batched_tokens will not
be set to such a large value). When loading vLLM, set
max_num_batched_tokens=22000. This length can accommodate two
medium-sized requests and some short requests, reflecting an extreme
scenario where the budget is almost entirely occupied by longer
requests.
Next, we mix 990 small requests and 100 medium requests into one type of
load scenario (hereinafter referred to as 10%), and similarly generate
load scenarios with 5% medium requests and 1% load scenarios.
Performance tests were conducted separately for enabling vLLMScheduler,
AscendScheduler, and AscendScheduler (long prompt concurrency set to 1).
- vLLM version: v0.10.2
- vLLM main:
1dfea5f4a9
---------
Signed-off-by: Csrayz <jover@cmbchina.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?
In the P node timeout release mechanism during PD separation, the req_id
that requires timeout release is transmitted from the scheduler to the
worker. If the KV cache between PDs is transferred too quickly, the P
node's req_id may be released twice. The first release is when the D
node notifies the P node that the KV cache has been pulled, and the
second release is when the scheduler transmits the timeout release to
the worker.
To address this bug, an intermediate component is introduced to manage
the release of req_ids.
Pull kv and forward2 may occur one after the other in timing. The
previous timeout defaulted to forward2 being before pull_kv.
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
---------
Signed-off-by: baxingpiaochong <771405853@qq.com>
### What this PR does / why we need it?
When we copy the sampled valid token ids from device to host, avoid
using tolist which would trigger a CUDA wise stream sync if the source
is on device. We change it to use non-blocking copy followed by an
explicit CUDA event sync.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
Bring up vLLM server
```bash
VLLM_USE_V1=1 vllm serve Qwen/Qwen2.5-14B-Instruct --disable-l
og-requests -tp 8 --max-num-seqs 64 --no-enable-prefix-caching --max_num_batched_tokens=8000
```
## Before：
## After
As shown in the figure, the TTFT decreased
- vLLM version: v0.10.2
- vLLM main:
9607d5eb44
---------
Signed-off-by: jesse <szxfml@gmail.com>
### What this PR does / why we need it?
This PR aims to address the incompatibility of the `.masked_scatter_`
operation in the current `_merge_multimodal_embeddings` function on
Ascend. For now, it reverts to the previous version of the CPU
operation, which can be executed asynchronously on the device side to
enhance performance.
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
---------
Signed-off-by: booker123456 <945658361@qq.com>
### What this PR does / why we need it?
modify the version compatibility between vllm and vllm-ascend, the main
branch of vllm-ascend corresponds to the v0.10.2 tag of vllm.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9
Signed-off-by: weiguihua2 <weiguihua2@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>
### What this PR does / why we need it?
Bump vLLM commit hash to
f225ea7dd9
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
5aeb925452
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
Updates the `cudagraph_support` attribute to `aclgraph_support` to use
terminology appropriate for the Ascend platform (ACL graphs instead of
CUDA graphs).
This change also explicitly disables graph support for the MLA attention
backend.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
None needed.
- vLLM version: v0.10.2
- vLLM main:
5aeb925452
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
This PR implements the renaming of the environment variable
VLLM_LLMDD_RPC_PORT to VLLM_ASCEND_LLMDD_RPC_PORT, as proposed and
tracked in
[#2450](https://github.com/vllm-project/vllm-ascend/pull/2450). The
renaming is intended to align the variable naming convention with other
Ascend-specific environment variables in the vllm-ascend codebase,
enhancing consistency and clarity for developers and users working with
Ascend-based deployments.
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
CI passed with existing test.
- vLLM version: v0.10.2
- vLLM main:
9607d5eb44
Signed-off-by: wyu0-0 <woshilynn@163.com>
## Purpose
This Pull Request enhances the EPLB (Expert Parallelism Load Balancing)
system by introducing a novel balancing algorithm: FlashLB.
## Motivation
1. The default algorithm adopts a two-stage greedy strategy:
a. Replica allotment: Determine the number of expert replicas by
minimizing the maximum load per replica (Min Max Replica, MMR).
b. Replica placement: Distribute replicas across devices by repeatedly
assigning the heaviest replica to the least loaded device (Longest
Processing Time First, LPT).
However, this sequential process lacks inter-stage collaborative
optimization, often leading to suboptimal load balancing. For example,
in the simple case shown in the figure below: given 8 logical experts
with hotness values of 600, 560, 120, 120, 20, 10, 10, 10, and 2
replicas allocated per device across 8 devices, the EPLB algorithm
yields a maximum per-device hotness of 232, while our proposed FlashLB
algorithm can reduce this value to 205.
2. The default algorithm relies on the averaged expert hotness over a
fixed time window for optimization. While this provides a coarse
approximation of the hotness distribution, it fails to capture
oscillatory deviations and temporal correlations of expert hotness
observed across iterations in real-world scenarios, limiting
optimization quality.
3. The default algorithm periodically regenerates the expert placement
table. However, it generates the table for each individual layer, and
the new table does not account for correlations with the previous one;
these two factors collectively lead to nearly full-scale expert
reassignment.
## FlashLB Algorithm Principle
1. Joint Optimization
FlashLB achieves joint optimization of replica allotment and placement
through group-based decision-making. Each group gradually determines the
replica count and placement for a subset of experts, ensuring that the
expected inter-device load balance (considering both deployed and
pending expert replicas) is holistically optimized. To attain superior
load balancing, FlashLB employs tree search to expand the solution space
while integrating pruning and precompilation techniques for
acceleration, thereby delivering load balancing that is both
high-quality and practically efficient.
2. Multi-Shot Enhancement
FlashLB partitions each profiling interval (e.g., 1024 iterations) into
consecutive smaller sub-intervals (e.g., 16 iterations), each capturing
independent hotness measurements. It then performs multi-shot
optimization to co-optimize these sub-intervals simultaneously—enabling
adaptation to time-variant expert hotness while enhancing robustness.
3. Incremental Adjustment
To reduce the overhead of frequent expert re-deployment, FlashLB
introduces an incremental adjustment scheme operating at both
inter-layer and intra-layer levels:
a. Inter-Layer: Hotness variations are tracked at the layer level. Only
layers with fluctuations exceeding a predefined threshold trigger
re-computation of expert placement, avoiding unnecessary redeployment
for stable layers；
b. Intra-Layer (Optional): A lightweight incremental LPT algorithm
(LPT-Incremental) is applied. Instead of recomputing full placement for
all experts in a layer, it selectively adjusts only the hottest experts
or those with replica count changes, further reducing migration
overhead.
This incremental strategy significantly reduces adjustment costs while
maintaining balanced performance across layers and devices.
## Co-author:
Co-authored-by: Skywalker-EP 173723846@qq.com
- vLLM version: v0.10.2
- vLLM main:
9607d5eb44
---------
Signed-off-by: sdmyzlp <lrwei2@petalmail.com>
Signed-off-by: Che Ruan <cr623@ic.ac.uk>
Signed-off-by: Shanshan Shen <87969357+shen-shanshan@users.noreply.github.com>
Signed-off-by: shen-shanshan <467638484@qq.com>
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
Signed-off-by: 22dimensions <waitingwind@foxmail.com>
Signed-off-by: zhanghaiwen <zhanghaiwen@cmss.chinamobile.com>
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
Signed-off-by: Lucas Kabela <lucaskabela@meta.com>
Signed-off-by: wangli <wangli858794774@gmail.com>
Signed-off-by: MengqingCao <cmq0113@163.com>
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: Icey <1790571317@qq.com>
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Signed-off-by: dependabot[bot] <support@github.com>
Signed-off-by: tangtianyi <tangtianyi4@huawei.com>
Signed-off-by: Angazenn <supperccell@163.com>
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Signed-off-by: rjg-lyh <1318825571@qq.com>
Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
Signed-off-by: fems14 <1804143737@qq.com>
Co-authored-by: sdmyzlp <117554856+sdmyzlp@users.noreply.github.com>
Co-authored-by: Che Ruan <cr623@ic.ac.uk>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>
Co-authored-by: Yikun Jiang <yikunkero@gmail.com>
Co-authored-by: 22dimensions <waitingwind@foxmail.com>
Co-authored-by: zhanghw0354 <zhanghaiwencmss@139.com>
Co-authored-by: zhanghaiwen <zhanghaiwen@cmss.chinamobile.com>
Co-authored-by: zhangxinyuehfad <59153331+zhangxinyuehfad@users.noreply.github.com>
Co-authored-by: Lucas Kabela <lucasakabela@gmail.com>
Co-authored-by: Li Wang <wangli858794774@gmail.com>
Co-authored-by: MengqingCao <cmq0113@163.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Icey <1790571317@qq.com>
Co-authored-by: linfeng-yuan <1102311262@qq.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: tianyitang <tangtianyi4@huawei.com>
Co-authored-by: Angazenn <supperccell@163.com>
Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
Co-authored-by: rjg-lyh <83491835+rjg-lyh@users.noreply.github.com>
Co-authored-by: weichen <132029610+Pr0Wh1teGivee@users.noreply.github.com>
Co-authored-by: weijinqian0 <12153182+weijinqian0@users.noreply.github.com>
Co-authored-by: fems14 <74094523+fems14@users.noreply.github.com>
### What this PR does / why we need it?
This PR addresses a critical issue where Node D (Device) failures cause
Node P (Processor) to hang due to inability to release KV cache.
**Trigger Scenarios:**
1. Node D fails mid-inference (e.g., network disconnection)
2. Node D rejects requests at a certain stage (e.g., via API server)
3. Load-test script termination causes Node P or D to abort queued
requests
**Root Cause Analysis:**
1. Currently, Node D sends a "KV cache pull complete, release approved"
message to Node P
2. This message is transmitted via the worker connector. If PD
connection breaks or requests are rejected upstream, Node D cannot send
the message
3. Node P will never release KV cache without receiving this message
**Solution:**
Following VLLM community's approach (NIXL connector timeout mechanism),
we're implementing:
- A timeout mechanism with comprehensive warnings
- Updated README documentation
- Reference: VLLM's optimization PR
[#20139](https://github.com/vllm-project/vllm/pull/20139)
### Does this PR introduce _any_ user-facing change?
None
### How was this patch tested?
None
- vLLM version: v0.10.2
- vLLM main:
9607d5eb44
---------
Signed-off-by: underfituu <hzhucong@163.com>
### What this PR does / why we need it?
Fix the impact to LoRA that
https://github.com/vllm-project/vllm/pull/25249 brought.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
pytest -sv tests/e2e/singlecard/test_ilama_lora.py
pytest -sv tests/e2e/multicard/test_ilama_lora_tp2.py
- vLLM version: v0.10.2
- vLLM main:
9607d5eb44
---------
Signed-off-by: paulyu12 <507435917@qq.com>
