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?
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?
Fix issues mentioned in
https://github.com/vllm-project/vllm-ascend/pull/2791 and some minor
refactoring.
1. Use Enum instead of string.
2. Avoid setting a new property to forward_context in
AscendFusedMoE.forward().
3. Enabling TokenDispatcherWithMoge.
4. Remove redundant code.
### 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
2. Aclgraph & eager
- vLLM version: v0.10.2
- vLLM main:
9607d5eb44
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?
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:
9607d5eb44
Signed-off-by: rjg-lyh <1318825571@qq.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:
ef7eefe17a
Signed-off-by: sdmyzlp <lrwei2@petalmail.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?
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:
dd39baf717
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
Co-authored-by: weijinqian0 <weijinqian@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?
This PR fused addrmsnorm op and w8a8 quant op to get better perf.
### 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:
0faf3cc3e8
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
This PR deletes ~2K lines of code about deepseek modeling. It falls back
CustomDeepseekV2 modules to original vllm implementations and adapts
some modifications in vllm about deepseek and moe.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E vllm serving with torchair graph mode and eager mode.
- vLLM version: v0.10.2
- vLLM main:
759ef49b15
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: yiz-liu <136800916+yiz-liu@users.noreply.github.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
1. Replace prepare/finalize operation in fused_moe.py by
moe_comm_method.prepare()/finalize()
2. Replace unified_fused_experts by moe_comm_method.fused_experts() in
fused_moe.py/w8a8_dynamic.py/w4a8_dynamic.py
3. Add calling _select_moe_comm_method in spec-decode proposers.
4. Currently, w4a8_dynamic does not support gatherep, use all2allv
instead.
5. Remove redundant code.
### Does this PR introduce _any_ user-facing change?
AllgatherEP switch is disabled in aclgraph/eager mode, just follow the
rules in modelrunner_v1._select_moe_comm_method()
### How was this patch tested?
e2e & ut
- vLLM version: v0.10.2
- vLLM main:
7f6f2c1182
Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
Co-authored-by: weijinqian0 <12153182+weijinqian0@users.noreply.github.com>
**Background:**
There are two principles about operator registration in PyTorch
- The same namespace can be only registered once by `TORCH_LIBRARY`
- The operator signatures can be only registered once by `def`
Considering that all custom operators defined in the current repo are
only used by Ascend, instead of defining a common operator schema by
vLLM, all accelerators then follow this operator schema and complete the
implementation based on their respective hardware, which is conducive to
functional abstraction.
Therefore, we can rename the operator registration namespace to an
Ascend-specific namespace(**_C_ascend**).
Related ISSUE: https://github.com/vllm-project/vllm-ascend/issues/2742
- vLLM version: main
- vLLM main:
f592b3174b
Signed-off-by: FFFrog <ljw1101.vip@gmail.com>
### What this PR does / why we need it?
This PR fixed the bug in register_custom_ops without forward_context. We
set try-except to consider this situation.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: main
- vLLM main:
7920de0a2a
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
`torch_npu.npu_apply_rotary_pos_emb` only support head_size and
rotary_dim equal 128. Error occurs when running GLM
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: main
- vLLM main:
404c85ca72
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
### What this PR does / why we need it?
modelslim will generate self.bias for rms norm in quantization, since
RMSNorm in vllm has no this parameter, so its nesscesary
to create a AscendQuantRmsNorm.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
tested by deepseek-v3.1-w8a8
<img width="2496" height="592" alt="image"
src="https://github.com/user-attachments/assets/004c6e76-3d7a-4a1f-b59f-a14304012663"
/>
- vLLM version: main
- vLLM main:
d6249d0699
Signed-off-by: 22dimensions <waitingwind@foxmail.com>
### What this PR does / why we need it?
1. Move ops/comm_utils to ops/moe/comm_utils
2. Move distributed/tensor_parallel/gather_from_sequence_parallel_region
to ops/moe/comm_utils
3. Delete distributed/tensor_parallel
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
e2e & ut
- vLLM version: main
- vLLM main:
a1213fae5f
---------
Signed-off-by: wuweiqiang24 <1005334931@qq.com>
Signed-off-by: wuweiqiang24 <wuweiqiang11@huawei.com>
### What this PR does / why we need it?
This PR prefetchs the weight of mlp layers in Qwen Dense Models to
optimize the performance in Decode phase mainly.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: main
- vLLM main:
a1213fae5f
Signed-off-by: rjg-lyh <1318825571@qq.com>
Co-authored-by: Shuming19 <313093131@qq.com>
### What this PR does / why we need it?
[Feat]support dynamic quantization in allgather
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: main
- vLLM main:
5931b7e5d9
Signed-off-by: withHades <244036962@qq.com>
Signed-off-by: WithHades <244036962@qq.com>
### What this PR does / why we need it?
fix ascend fused moe spelling error
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
0ae43dbf8c
- vLLM version: main
- vLLM main:
fcc0a3130a
Signed-off-by: zhaozixin <zhaozixin1@huawei.com>
Co-authored-by: zhaozixin <zhaozixin1@huawei.com>
### What this PR does / why we need it?
Currently, when executing to the Linear layer of the model in
vLLM-Ascend, the weights input format is ND in unquantized case and
skipped ascend case, which is slower than FRACTAL_NZ.
This PR supplements the execution logic for Linear layer. When
VLLM_ASCEND_ENABLE_MLP_OPTIMIZE=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.
- vLLM version: main
- vLLM main:
267c80d31f
Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>
### What this PR does / why we need it?
Remove compatibility maintenance for vllm v0.10.1 and v0.10.1.1
### Does this PR introduce _any_ user-facing change?
branch main of vllm-ascend will not be compatible with vllm v0.10.1 and
v0.10.1.1
### How was this patch tested?
CI passed with existing test.
- vLLM version: v0.10.1.1
- vLLM main:
6fb2788163
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
In reinforcement learning scenarios, weight updates are required, but
the current inference applies a transpose operation to the weights,
altering their shape. This causes a shape mismatch with the training
weights, triggering an error during weight updates.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.10.1.1
- vLLM main:
6fb2788163
Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
### What this PR does / why we need it?
Really strange that `register_oot` doesn't work with `SharedFusedMoE`,
so we have to add this patch, for now.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
This PR won't have any effect in DeepSeek since we currently still stick
with the old `CustomDeepseekV2`.
- vLLM version: v0.10.1.1
- vLLM main:
0cdd213641
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
Optimize rope by caching sin and cos at the first layer in Qwen Models.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.10.1.1
- vLLM main:
562663a044
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
Signed-off-by: ZYang6263 <zy626375@gmail.com>
Signed-off-by: rjg-lyh <1318825571@qq.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
Co-authored-by: ZYang6263 <51255902183@stu.ecnu.edu.cn>
Co-authored-by: ZYang6263 <zy626375@gmail.com>
### What this PR does / why we need it?
Flashcomm_v1 optim in Qwen Dense Models.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.10.1.1
- vLLM main:
5e537f45b4
Co-authored-by: 1024daniel <xxltju324@gmail.com>
### What this PR does / why we need it?
The current implementation will result in duplicate generation of
`sin_cos_cache` in rope when `kv_seqlen` > 4k, because the
initialization length of the `sin_cos_cache` is only 4k.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
After this PR merged, sin_cos_cache will not increase in forward func,
so `test_native_rope_deepseek_forward_cache_handling` is not necessary.
- vLLM version: v0.10.1.1
- vLLM main:
60f0843ef8
Signed-off-by: zzzzwwjj <1183291235@qq.com>
### What this PR does / why we need it?
1. Move prepare/finalize operation from moe_comm_method to
/ops/moe/fused_moe_prepare_and_finalize
2. Adapt to token_dispatcher in moe_comm_method
3. Move
moe_comm_method/experts_selector/token_dispatcher/fused_moe_prepare_and_finalize
to /ops/moe
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
e2e & ut
- vLLM version: v0.10.1.1
- vLLM main:
f4962a6d55
Signed-off-by: weichen <calvin_zhu0210@outlook.com>
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?
line 408 already declared mc2_mask , remove duplicated unused code
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
CI passed with existing test.
- vLLM version: v0.10.1.1
- vLLM main:
60f0843ef8
Signed-off-by: machenglong <machenglong_yewu@cmss.chinamobile.com>
### What this PR does / why we need it?
quantization patch is unused code
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
tested by CI
- vLLM version: v0.10.1.1
- vLLM main:
f4962a6d55
Signed-off-by: 22dimensions <waitingwind@foxmail.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?
Delete redundant codes related to communication
### Does this PR introduce _any_ user-facing change?
not involve
### How was this patch tested?
not involve
- vLLM version: v0.10.1.1
- vLLM main:
6c7af8110a
---------
Signed-off-by: 刘哲续 <liuzhexu1@huawei.com>
Co-authored-by: 刘哲续 <liuzhexu1@huawei.com>
### What this PR does / why we need it?
Refactors the Mixture-of-Experts (MoE) communication method selection
logic. The choice between all-gather, all-to-all, and mc2 is now
determined by expert parallel configuration, SoC version (A2/A3), and
token count for better performance.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
Added.
- vLLM version: v0.10.1.1
- vLLM main:
eafa8dcde6
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
### What this PR does / why we need it?
1. Similar to #2384 , this PR add a torchair-specific modeling for
pangu.
2. Fixes a bug introduced by routed_scaling_factor in #2675 .
3. remove eager test case for pangu since there has already been a
torchair test case.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.10.1.1
- vLLM main:
6997a25ac6
---------
Signed-off-by: zengyanjia <z00883269@china.huawei.com>
Signed-off-by: Angazenn <supperccell@163.com>
Co-authored-by: zengyanjia <z00883269@china.huawei.com>
### What this PR does / why we need it?
Fix long context seq accuracy problem for `GLM4.5`.
When `max_tokens=1000`, there is cyclic output problem like:
```bash
00 00 00 00 00 00 00 00 00 00 00 00 00 00
```
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
```python
import os
os.environ["VLLM_USE_MODELSCOPE"] = "True"
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
from vllm import LLM, SamplingParams
def main():
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(max_tokens=1000, temperature=0.0)
# Create an LLM.
llm = LLM(model="/root/.cache/modelscope/hub/models/ZhipuAI/GLM-4___5",
tensor_parallel_size=8,
enforce_eager=True,
trust_remote_code=True,
max_model_len=1024)
# Generate texts from the prompts.
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
if __name__ == "__main__":
main()
```
- vLLM version: v0.10.1.1
- vLLM main:
0235103cbb
---------
Signed-off-by: Shanshan Shen <87969357+shen-shanshan@users.noreply.github.com>
Signed-off-by: shen-shanshan <467638484@qq.com>
### What this PR does / why we need it?
This PR enables `npu_moe_gating_top_k_softmax` when running quantized
MoE (such as W8A8). This op in fact makes no distinction between
quantized and non-quantized scenarios. Introducing this op reduces 3~4ms
for TPOT.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- vLLM version: v0.10.1.1
- vLLM main:
ce30dca5c4
Signed-off-by: Angazenn <supperccell@163.com>
Clean up useless code which is only used for torchair in rotary_embedding
- vLLM version: v0.10.1.1
- vLLM main:
a344a5aa0a
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Fix CI Break: upstream adds routed_scaling_factor in forward_oot
interface, vllm-ascend needs to adapt
### Does this PR introduce _any_ user-facing change?
NA
### How was this patch tested?
E2E and UT
- vLLM version: v0.10.1.1
- vLLM main:
3e330fcb21
Signed-off-by: leo-pony <nengjunma@outlook.com>
### What this PR does / why we need it?
After moved torchair related rope ops into torchair_ops, split the
torchair from the origin rope ops to make the code clean.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
vLLM version: main
vLLM main:
ab9f2cfd19
- vLLM version: v0.10.1.1
- vLLM main:
81eea3d348
Signed-off-by: hust17yixuan <303660421@qq.com>
### What this PR does / why we need it?
There are a lot of redundant codes related to moe here, and the
structure is not very clear.
We did the following things：
we have placed the relatively independent code related to apply_mlp into
a separate file;
removed the environment variables of alltoall_buffer and alltoall_seq.
Remove the code related to alltoall_buffer and alltoall_seq, and retain
the sole TokenDispatcher inheritance class.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
e2e&ut
- vLLM version: v0.10.1.1
- vLLM main:
4071c76cf3
---------
Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian0 <12153182+weijinqian0@users.noreply.github.com>
### What this PR does / why we need it?
* **Unify execution paths:** Consolidates the quantized and
non-quantized execution paths into a single `fused_experts` function,
removing duplicated logic and making the control flow clearer and easier
to maintain.
* **W8A8 dynamic quantization:** Adds support for W8A8 dynamic
quantization inside the unified MoE kernel. Communication routines are
updated to correctly handle dynamic quantization scales for activations.
* **Weight pre-processing:** Prae-transpose the `w13` and `w2` weight
matrices (as implemented in PR #2025) so that quantized and
non-quantized models follow the same code path for the MoE gating,
up-projection, and down-projection operations.
* **All-to-all communication:** Adds an `all-to-all` collective
communication pattern. For large token counts on modern hardware,
`all-to-all` is more efficient than the previous `all-gather` strategy.
However, `all-to-all` is not really captured and replayed due to
multiple D2H operations which will trigger synchronization, and thus
raise error when capture graphs. We only use `all-to-all` when fallback
to `compiled_graph_for_general_shape`.
* **Dynamic communication selection:** The model runner now selects the
optimal MoE communication method (`mc2`, `allgather`, or `alltoall`) at
runtime based on token count and the Ascend SoC version.
* **Limitation:** `all-gather` is not yet supported for quantized
models, which means there is still something left to do on A2.
### Does this PR introduce _any_ user-facing change?
None.
### How was this patch tested?
No further test cases needed.
- vLLM version: v0.10.1.1
- vLLM main:
d660c98c1b
---------
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
