### 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:
https://github.com/vllm-project/vllm/commit/5aeb9254521023f97aca292b3478aa7ff485ffb2
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:
https://github.com/vllm-project/vllm/commit/9607d5eb449711b349d4c2bee0a9c94afcc7ed14
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:
https://github.com/vllm-project/vllm/commit/9607d5eb449711b349d4c2bee0a9c94afcc7ed14
---------
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:
https://github.com/vllm-project/vllm/commit/9607d5eb449711b349d4c2bee0a9c94afcc7ed14
---------
Signed-off-by: underfituu <hzhucong@163.com>
### What this PR does / why we need it?
This is the follow-up PR of #2128 .
Moves graph parameter management components, including `GraphParams`,
`get_graph_params`, and `set_graph_params`, from the generic `utils.py`
to the more specific `compilation/acl_graph.py`.
Additionally, extracts the `update_attn_params` logic from the
`NPUModelRunner` class into a standalone function within the `acl_graph`
module.
This refactoring improves code organization by centralizing ACL
graph-related logic into its own dedicated module, enhancing modularity
and clarity.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
None needed.
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
A new kv_role "kv_both" is added to run mixed deployment scenarios. The
mixed deployment will involve a decode phase, where with_prefill should
be false.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/c60e6137f0bf2034853919b3a9d705d7e06b93cf
Signed-off-by: fems14 <1804143737@qq.com>
### What this PR does / why we need it?
This PR removed the redundant log prints in register_custom_ops.py, in
order to make output clear.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/9607d5eb449711b349d4c2bee0a9c94afcc7ed14
Signed-off-by: rjg-lyh <1318825571@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:
https://github.com/vllm-project/vllm/commit/9607d5eb449711b349d4c2bee0a9c94afcc7ed14
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
Follow up `UniformTypeKVCacheSpecs` changes introduced by
https://github.com/vllm-project/vllm/pull/25101, which support different
hidden size in uniform type kvcache specs
This also fix the CI issue about `TypeError: AttentionGroup.__init__()
missing 1 required positional argument: 'kv_cache_spec'`
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Tests passed with exsiting e2e tests.
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/c60e6137f0bf2034853919b3a9d705d7e06b93cf
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
Add new npu_fused_infer_attention_score op to improve perfomance in
splitfuse cases and resolve long-seq mask problems .
1. The original op's performance is suboptimal in certain scenarios,
necessitating optimization through the _new op_
(npu_fused_infer_attention_score)。
2. For ultra-long sequences (128k), the original operator will allocate
a large attn_mask, which consumes excessive CPU memory. In contrast, the
_new op_ supports a fixed-size compressed mask, effectively resolving
this issue.
NOTE1: The current PR retains the original logic and uses a version
check of the CANN package to determine whether the _new op_ can be
enabled. This ensures no impact on existing users. In future versions,
this version check and the original logic will be deprecated, and the
_new op_ scheduling will be uniformly adopted.
NOTE2: This pr relies on future CANN version, which is not available
now.
NOTE3: To enable the new op in chunked prefill, the parameter
additional_config should be set like `--additional-config
'{"ascend_scheduler_config":
{"enabled":true,"enable_chunked_prefill":true}}' \` at least.
### 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/6c5f82e5aa87cd73ce03ce10fc44138f75ee1aea
---------
Signed-off-by: tangtianyi <tangtianyi4@huawei.com>
Signed-off-by: Angazenn <supperccell@163.com>
Co-authored-by: Angazenn <supperccell@163.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:
https://github.com/vllm-project/vllm/commit/6d8246aaffff3ebec84767e373212a7b8da328e2
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
1. Refactor ci to reuse base workflow and enable main 2 hours trigger
job:
- Extract e2e test in to _e2e_test.yaml
- Reuse _e2e_test in light / full job
- Enable main 2 hours trigger job
2. Rename e2e test to ascend test to make sure action display label
3. Re-enable ut coverage which was failed since
https://github.com/vllm-project/vllm-ascend/commit/5bcb4c1528311b8f20c91a9239ac38172f002fcb
and disable on
https://github.com/vllm-project/vllm-ascend/commit/6d8bc38c7b0faba0b666585a09b7523c7714b9cc
### Does this PR introduce _any_ user-facing change?
Only developer behavior changes:
- Every job trigger full test with vllm release and hash
- Run full job per 2 hours with vllm main
- e2e light test (30 mins): `lint` (6mins) ---> ut (10mins) --->
`v0.10.2 + main / 4 jobs` (15mins)
- e2e full test (1.5h): `ready label` ---> `v0.10.2 + main / 4 jobs`,
about 1.5h
- schedule test: 2hours ---> `v0.10.2 + main / 4 jobs`, about 1.5h
### How was this patch tested?
CI passed
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/c60e6137f0bf2034853919b3a9d705d7e06b93cf
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
### What this PR does / why we need it?
Fix shape not match when test LLM-Research/Phi-4-mini-instruct accuarcy
### Does this PR introduce _any_ user-facing change?
Users can't set dynamic batch_size or use lm_eval test accuracy when
using models(sliding_window)
### How was this patch tested?
accuarcy of LLM-Research/Phi-4-mini-instruct is ok :
```
vllm (pretrained=LLM-Research/Phi-4-mini-instruct,max_model_len=4096,dtype=auto,tensor_parallel_size=1), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version| Filter |n-shot| Metric | |Value | |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k| 3|flexible-extract| 5|exact_match|↑ |0.8105|± |0.0108|
| | |strict-match | 5|exact_match|↑ |0.8097|± |0.0108|
```
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/3c96e7b8a1ffac31b947c753f4fe1d0cd14abf87
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
### What this PR does / why we need it?
Some custom models in vllm-ascend define packed_modules_mapping, which
prevent keeping same model class with vllm community. So move these
custom packed_modules_mapping to quant utils.py. After this pr, some
custom models can be removed.
### Does this PR introduce _any_ user-facing change?
tested by CI
### How was this patch tested?
tested by CI
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/5089fd749cbe4233a29f29ce706d56c47464c117
Signed-off-by: 22dimensions <waitingwind@foxmail.com>
Add missing barrier when no implicit synchonize by `repeat_interleave`
is available. Otherwise, the `non_blocking=True` copy of `output_splits`
and `input_splits` from NPU may failed to complete before later
`async_all_to_all` uses them.
### What this PR does / why we need it?
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/ef7eefe17a7dc212ddb8a8aabd7760218a10e25e
Signed-off-by: sdmyzlp <lrwei2@petalmail.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:
https://github.com/vllm-project/vllm/commit/6d8246aaffff3ebec84767e373212a7b8da328e2
Signed-off-by: xuyexiong <xuyexiong@huawei.com>
### What this PR does / why we need it?
This pr fixes a few issues on prefill disaggregation:
1. Fix prefill disaggregation kvcache addr alignment issue, llmdatadist
needs the addr of tensors to be aligned with 2M
2. Fix prefill disaggregation kvcache shape error, llmdatadist requires
k/v tensors with shape [num_blocks, ...], however the implentment before
this pr is [2, num_blocks, ...], which will break prefill disaggregation
3. Use hybrid kv cache only when running qwen3_next to fix accuracy
issue on prefill disaggregation.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Tested locally by @liziyu179
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/4f02b77de4e794a0d417ed98a26884208f75e043
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
For sleep mode level 2, we discarded model both weights and kv_cache,
but the problems is: When we discard weights, we also discard some
tensors representing the model state which we called
`model.named_buffers()`, such as: `running_mean / running_var` in
BatchNorm、rope cos-sin cache ... when we update weights, but forgot to
update buffers as well, this will lead to some unknown issue
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/5963b98b465007e3cfb0d39447e4459a8afa96dc
---------
Signed-off-by: wangli <wangli858794774@gmail.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:
https://github.com/vllm-project/vllm/commit/ef7eefe17a7dc212ddb8a8aabd7760218a10e25e
---------
Signed-off-by: SunnyLee219 <3294305115@qq.com>
This pr fixes two problems while `multistream_moe` enabled in torchair
graph mode:
1. check `TorchairAscendW8A8DynamicFusedMoEMethod` instead of incorrect
`AscendW8A8DynamicFusedMoEMethod`
2. mc2_mask should be chunked no matter `replace_allreduce` is True or
False in forward function of `TorchairAscendFusedMoE`
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/0fb2551c238c7ccbcf6f25ef4646ce6c92f684d1
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
The current linear.py has the following issues:
- There is redundant conditional logic in the `comm_group` and `forward`
selection for classes such as `AscendMergedColumnParallelLinear`.
- Inconsistent comm_group selection logic exists among
`AscendMergedColumnParallelLinear`, `AscendColumnParallelLinear`, and
`AscendQKVParallelLinear`.
To address these two issues, this PR encapsulates `comm_group` and
`forward` into classes and extracts the classes selection logic into
common functions. For future additions of custom communication groups or
forward methods, it will only be necessary to extend
`CustomColumnParallelOp` or `CustomRowParallelOp` and add new selection
logic.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/dd39baf7175c5e79faef071c67bb318eadb7752f
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
Co-authored-by: weijinqian0 <weijinqian@huawei.com>
### What this PR does / why we need it?
https://github.com/vllm-project/vllm-ascend/pull/2849 moves the
implementation of `shared_expert_dp` to torchair deepseek_modeling.
However, the calling of `set_forward_context` with `enforce_eager` and
`shared_expert_dp` falls back to the implementation of
model_runner_v1.py and set the global attn_metadata as a dictionary. It
leads to a RuntimerError when attn_metadata is got from the forward
context and used in torchair_deepseek_v2.py. This PR fixes this problem
by introducing the transformation of attn_metadata in this file.
Note that current E2E testing lacks the case of deepseek with
`shared_expert_dp`. We need to add an ST with `shared_expert_dp` in
testing workflow.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
e2e vllm serving with `enable_shared_expert_dp: true` passed.
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/de3e53a75ba9f31f446926911b7c44561af3b2ee
Signed-off-by: linfeng-yuan <1102311262@qq.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:
https://github.com/vllm-project/vllm/commit/567939953b7a9cb0ded6bf0bb21a76917b8fed97
---------
Signed-off-by: offline0806 <z00858301@china.huawei.com>
Co-authored-by: offline0806 <z00858301@china.huawei.com>
### What this PR does / why we need it?
1. In memory of #2509, Fix mtp torchair in pd Disaggregation scenario
2. fix mla bug in SpecDecoding Scenario, since num_decodes !=
num_decode_tokens
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.2
- vLLM main:
https://github.com/vllm-project/vllm/commit/5206ab20ba7477e5457c2e64469590d548fa15e6
Signed-off-by: xuyexiong <xuyexiong@huawei.com>