### What this PR does / why we need it?
Some bug fixes, mainly including:
1. For A2, the number of experts each single card cannot be greater than
16 when using MC2. The PR fixed the error in the A2 moe communication
method selection, which would cause the selection of an incorrect
communication method when the number of model experts exceeds 256. For
example, when using an A2 16-cards model to load the PD-disaggregation D
node with Qwen3.5 series models, the incorrect MC2 method would be
chosen.
2. Fixed the issue where the layerwise connector sends the kv-cache of
the MTP layer multiple times when `num_spec_tokens` > 1. Now, the
kv-cache is sent only when the MTP layer is forward for the first time.
3. Fix the accuracy issue of qwen3.5 when using MTP for PD
disaggregation. The cause is that `num_decode_draft_tokens` does not
consider that `spec_tokens` are not existed during the first inference
when PD disaggregation (`spec_tokens` are generated during the first
inference). However, `spec_tokens_padding` is added by
`recomputed_scheduler`. As a result, `gdn_metadata` incorrectly
considers that the prefill with a length of 2 is performed.
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
Signed-off-by: zxr2333 <64738772+nwpu-zxr@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
### What this PR does / why we need it?
Qwen3.5 Moe supports enabling the dispatch_ffn_combine fusion operator.
Fix problem: In the w8a8 quantization scene, Qwen3.5 model's config.json
lacks the quantize field. The previous logic strictly relied on
quant_type == "w8a8_dynamic" to enable VLLM_ASCEND_ENABLE_FUSED_MC2.
This caused the dispatch_ffn_combine fusion operator to fail to activate
even when the environment variable was set.
Enable dispatch_ffn_combine fusion operator for BF16 scenarios.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: asunxiao <asunxiao@qq.com>
### What this PR does / why we need it?
This PR optimizes bias handling in `AscendRMSNorm` without changing the
intended
functional behavior.
In the current implementation, bias may be initialized for
`AscendRMSNorm` based
on configuration-level detection, even though some norm layers never
actually
load a bias weight. This can cause the inference path to enter the bias
branch
and execute an unnecessary `add_` operator.
To improve this, this PR introduces a loader-based flag to record
whether the
bias has actually been loaded. The bias addition is then executed only
when the
bias is truly present.
This optimization reduces redundant computation in inference and makes
the bias
application logic better aligned with the actual model weights.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
This PR fixes the bug for eagle3 and cp enable introduced by the
parallel speculative inference PR.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
tests and ut
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
Co-authored-by: kunpengW-code <1289706727@qq.com>
Co-authored-by: linsheng1 <1950916997@qq.com>
### What this PR does / why we need it?
Currently, chunked prefill is forcibly enabled. DeepSeek V3.1 W8A8C8
supports only the PD separation scenario. C8 refers to quantizing the KV
cache to int8, which aims to reduce the GPU memory usage of the KV cache
and improve the inference throughput.
Constraints:
1. Only the PD separation mode can be used and
MooncakeLayerwiseConnector can be used to run the model.
2. Currently, only the activation value supports dynamic quantization,
and the KV cache supports static quantization. C8 quantization with MTP
is not supported. You can use ModelSlim for quantization. The
quantization procedure is as follows:
pip install transformers==4.48.2
git clone https://gitcode.com/Ascend/msmodelslim.git
cd msmodelslim
bash install.sh
cd example/DeepSeek/
python3 quant_deepseek_w8a8.py --model_path <path/weight> --save_path
<path/quant_weight>
--anti_dataset../common/deepseek_anti_prompt_50_v3_1.json
--calib_dataset../common/deepseek_calib_prompt_50_v3_1.json --rot
--trust_remote_code True --fa_quant --dynamic --anti_method m6
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: pichangping <1337510399@qq.com>
Signed-off-by: Wang Kunpeng <1289706727@qq.com>
Co-authored-by: Wang Kunpeng <1289706727@qq.com>
### What this PR does / why we need it?
Two problems have been solved in this pr.
These problems occur in the `FULL_DECODE_ONLY` mode that `num_tokens`
should be padded to some value in `cudagraph_capture_sizes`.
1. We found the length of `seq_lens_list` in drafter's `attn_metadata`
is 1 shorter than expected. It will raise a kernel exception to make
vllm crash.
e.g., `num_reqs` = 3, `cudagraph_capture_sizes` = [20],
`actual_seq_lengths_q` is padded well to [4, 8, 12, 20]. But
`seq_lens_list` = [5742, 4700, 7996], it is not padded.
3. Though the length of `seq_lens_list` in target's `attn_metadata` is
the same as expected in `FULL_DECODE_ONLY`, some data are corrupted at
the end of the list.
e.g., `num_reqs` = 3, `cudagraph_capture_sizes` = [20],
`actual_seq_lengths_q` is padded well to [4, 8, 12, 20]. But
`seq_lens_list` = [5742, 4700, 7996, 5738], it has corrupted at the end
of the list.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
This PR fixes a bug in Xlite
backend(https://atomgit.com/openeuler/GVirt/issues/3).
This PR adds support for mrope (Mixture-of-RoPE) and deepstack features
in the xlite backend. These features are necessary for running certain
multimodal models that utilize them.
The main changes include:
- Updating `_build_model_config` to parse mrope and deepstack
configurations from the model's `hf_config`.
- Modifying `XliteWrapper.__call__` to handle `deepstack_input_embeds`
and mrope positions during the model forward pass.
- Replacing `ModelAttnMeta` with the newer `AttnMeta` to accommodate the
new metadata fields required by these features.
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?
online server config:
```
python -m vllm.entrypoints.openai.api_server \
--model /mnt/nvme0n1/models/checkpoint-8200 \
--additional-config='{"xlite_graph_config": {"enabled": true}}' \
--tensor-parallel-size 4 \
--gpu-memory-utilization 0.9 \
--max-num-batched-tokens 8192 \
--max-num-seqs=20 \
--block-size 128 \
--max-model-len 8192 \
--trust-remote-code \
--served-model-name Qwen3-VL-8B \
--host localhost \
--generation-config vllm \
--port 6777
```
test_config:
```
vllm bench serve \
--max-concurrency ${maxconcurrency} \
--num-prompts ${num_prompts} \
--host ${HOST} \
--port ${PORT} \
--model ${MODEL_NAME} \
--dataset-name random \
--backend openai-chat \
--random-input-len 512 \
--random-output-len 512 \
--random-range-ratio 0.2 \
--temperature 0.6 \
--metric-percentiles "50,90,99" \
--tokenizer ${TOKENIZER_PATH} \
--endpoint /v1/chat/completions \
--ignore-eos
```
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
Signed-off-by: LVYANGGUO <lvyangguo@huawei.com>
Co-authored-by: LVYANGGUO <lvyangguo@huawei.com>
### What this PR does / why we need it?
Optimize the performance of the triton operator _topk_log_softmax_kernel
in model_runner_v2 to 1.04xH100，which is 7% of its original value.(issue
https://github.com/vllm-project/vllm-ascend/issues/5208)
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: wangx700 <wangxin700@huawei.com>
### What this PR does / why we need it?
This PR restores #7029, which adds W8A8C8 support for dsv3.2/glm5 using
the `lightning_indexer_quant` ops in the pd-mix stage.
The original PR was reverted by #7288 because the patch did not work
with the recompute scheduler.
This PR also fixes the patching issue so that it works correctly with
the recompute scheduler.
### Does this PR introduce _any_ user-facing change?
Yes. To enable LI C8, users need to set the `enable_sparse_c8` option to
`"true"` in `additional_config`.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
**Refactor: Replace npu_ring_mla with FIA in MLA prefill**
This PR refactors the MLA (Multi-Layer Attention) prefill implementation
by replacing `npu_ring_mla` with `npu_fused_infer_attention_score` (FIA)
operator, unifying the attention backend with the standard attention
implementation.
**Key changes:**
1. **Core prefill refactoring (`mla_v1.py`)**
- Replace `npu_ring_mla` with `npu_fused_infer_attention_score` in
`_forward_prefill` and `_compute_prefill_context`
- Use TND layout with `softmax_lse_flag=True` for prefill attention
- Use `npu_attention_update` to merge multiple chunk outputs with LSE
(Log-Sum-Exp)
- Change `attn_mask` from `get_final_mla_mask()` to
`get_splitfuse_attn_mask()` for FIA compatibility
2. **Data type handling**
- Add automatic float16 → bfloat16 conversion (FIA with TND layout only
supports bfloat16)
- Convert output back to original dtype after FIA computation
3. **Metadata optimization**
- Pre-calculate `actual_seq_lengths_q` in `AscendMLAPrefillMetadata`
- Pre-calculate `chunk_actual_seq_lengths_kv_list` in
`ChunkedContextMetadata`
- Move `torch.cumsum` operations from forward pass to metadata building
phase
4. **CP compatibility (`mla_cp.py`)**
- Add `_ring_mla_mask_builder` to get `npu_ring_mla`-compatible masks
for Context Parallel scenarios
- Add `chunk_actual_seq_lengths_kv_list` field to
`CPChunkedContextMetadata`
**Why we need it:**
- **Backend unification**: Aligns MLA prefill with standard attention
implementation (`attention_v1.py`)
- **Better chunked context support**: FIA + `npu_attention_update`
provides native LSE-based output merging
- **Future compatibility**: Prepares for eventual `npu_ring_mla` removal
across the codebase
### Does this PR introduce _any_ user-facing change?
**No.** This is a pure refactoring with no functional changes - same
behavior, unified backend.
---
- Related issue: #5463 (item 7)
- vLLM version: v0.14.1
Signed-off-by: lico67373 <918688502@qq.com>
### What this PR does / why we need it?
This reverts commit 7ed9e9de69, which
introduces an issue that the patch doesn't work with recompute scheduler
enabled.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
Fix the error that reports while initializing qwen3-reranker-0.6b model
with `--enable-lora`.
And add a testcase to verify the fix.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: paulyu12 <507435917@qq.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
### What this PR does / why we need it?
To support prefix cache for Qwen3.5/Next in vLLM-Ascend, this PR mainly
follows the design in
[#30877](https://github.com/vllm-project/vllm/pull/30877) and inherits
changes to functions which are overridden in vLLM-Ascend.
Note:
1. `--mamba-cache-mode align` && PD disaggregation is still not
supported yet in vLLM v0.17.0(see
https://github.com/vllm-project/vllm/blob/main/vllm/v1/core/sched/scheduler.py#L295).
2. The current implementation of hybrid kv cache might result in a very
large block_size when scheduling. For example, if we run Qwen3.5-35B-A3B
with `-tp 2`, the block_size is adjusted to 2048, which means that any
prefix shorter than 2048 will never be cached. Although this behavior is
consistent with vLLM, it still needs improvements in the future.
3. `--mamba-cache-mode align` requires to copy mamba states during
forward steps. vLLM uses a triton kernel to implement it. However, the
original version run into some bugs on Ascend hardwares. Thus we patch a
new triton kernel to avoid this bug.
### Does this PR introduce _any_ user-facing change?
To use mamba prefix cache, set `--enable-prefix-caching` and
`--mamba-cache-mode align`. Note that the mamba state copy function(see
[do_mamba_copy_block](https://github.com/vllm-project/vllm/blob/main/vllm/v1/worker/mamba_utils.py#L132))
does not provide a torch native version, thus it might have trouble if
users can't use triton.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Angazenn <supperccell@163.com>
### What this PR does / why we need it?
Conditional Import for Mooncake: The import of
mooncake.engine.TransferEngine was moved into a try-except block within
the GlobalTE class's constructor. This ensures that mooncake is only
imported when needed and provides a clear error message with
installation instructions if it's missing.
### Does this PR introduce _any_ user-facing change?
The error message "ascend_transport.so: cannot open shared object file:
No such file or directory" in the CI is fixed to ensure the normal
running of the CI.
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: 房建伟 <fangjianwei@fangjianweideMacBook-Air.local>
Co-authored-by: 房建伟 <fangjianwei@fangjianweideMacBook-Air.local>
### What this PR does / why we need it?
This pr fix the Unified draft parallel feature.
1. In Draft model proposer, there are exceed 1 attention layers in
target model, thus removing the assertion on layer number.
2. we should get block size through `draft_attn_groups` instead of
`attn_metadata_builder` after 0.17.0.
3. `attn_update_stack_num_spec_norm` shouldn't be done when unified
draft parallel is enabled
### How was this patch tested?
Test pass with
`tests/e2e/singlecard/spec_decode/test_v1_spec_decode.py::test_parallel_drafting_acceptance`,
which is already included in CI
- vLLM version: v0.17.0
- vLLM main:
4034c3d32e
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
Reapply the auto-detect quantization format feature (originally in
#6645, reverted in #6873) and extend it to support remote model
identifiers (e.g., `org/model-name`).
Changes:
- Reapply auto-detection of quantization method from model files
(`quant_model_description.json` for ModelSlim, `config.json` for
compressed-tensors)
- Add `get_model_file()` utility to handle file retrieval from both
local paths and remote repos (HuggingFace Hub / ModelScope)
- Update `detect_quantization_method()` to accept remote repo IDs with
optional `revision` parameter
- Update `maybe_update_config()` to work with remote model identifiers
- Add platform-level `auto_detect_quantization` support
- Add unit tests and e2e tests for both local and remote model ID
scenarios
Closes#6836
### Does this PR introduce _any_ user-facing change?
Yes. When `--quantization` is not explicitly specified, vllm-ascend will
now automatically detect the quantization format from the model files
for both local directories and remote model IDs.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
### What this PR does / why we need it?
LMCache-Ascend is LMCache's solution on the Ascend platform and one of
the KVCache pooling solutions for Ascend. We hope to integrate
LMCache-Ascend into the vLLM-Ascend community as one of the official
KVCache pooling solutions for vLLM-Ascend.
We added a new LMCacheAscendConnector in vLLM-Ascend and registered it.
### Does this PR introduce _any_ user-facing change?
Users can specify the kvconnector using `--kv-transfer-config`, allowing
them to freely choose which kvconnector to use, without any user-facing
change.
### How was this patch tested?
Test by specifying `--kv-transfer-config
'{"kv_connector":"LMCacheAscendConnector","kv_role":"kv_both"}'`
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: chloroethylene <jjysama@gmail.com>
### What this PR does / why we need it?
Drop 0.16.0 support in main
- Fix eagle proposer break introduced by
https://github.com/vllm-project/vllm/pull/34552. Mainly change to use
the draft attention group to initialize the attention metadata builder.
- Fix the `ModelRunner` has no attribute `cudagraph_capture_sizes`
error, which is a bug in vLLM v0.17.0, and fixed by a later pr
https://github.com/vllm-project/vllm/pull/30515
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: MengqingCao <cmq0113@163.com>
### What this PR does / why we need it?
This PR supports W8A8C8 in dsv3.2/glm5 with lightning_indexer_quant ops
in pd-mix stage mainly.
Because the code for the current PD-disaggregated scenario is still
under refactoring and cleanup, this PR prioritizes ensuring the C8
functionality in the pd-mix scenario.
The next steps are planned in two parts:
① Once the optimized scatter operator is updated, we will replace the
original operator to improve the performance of storing k_scale.
② Once the code logic for the PD-disaggregated scenario becomes stable,
we will carry out more comprehensive validation and make appropriate
adaptations.
③ Because enabling C8 currently introduces several new operators whose
performance still needs improvement, performance may regress in some
scenarios. Therefore, only after all the operators are fully ready can
we ensure that this feature does not cause any performance degradation.
At that point, we will enable this feature by default and remove the
switch in `additional_config`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
CI passed with new added/existing test.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: rjg-lyh <1318825571@qq.com>
### What this PR does / why we need it?
This PR aims to support aclgraph for model runner v2, please see RFC
#5208. The PR contains these modifications:
- adapt to newest commit of vllm main branch.
- supply a unified interface of extra forward context for both model
runner v1 and model runner v2.
- implement graph mode for main model.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
### What this PR does / why we need it?
This PR fixes a bug in `reshape_kvcache_tensors` when reshaping the
Mamba cache for models like Qwen3.5. The previous implementation did not
correctly handle cases where the KV cache tensors have different data
types. This change ensures that slicing is performed based on byte
offsets before reshaping the tensors, which correctly handles
heterogeneous dtypes.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
### What this PR does / why we need it?
When GLM5 target model uses rotary quant, the final hidden states passes
to MTP need to do an extra rotary.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Wangbingjie <wangbj1207@126.com>
Signed-off-by: wangbj127 <256472688+wangbj127@users.noreply.github.com>
### What this PR does / why we need it?
The merged graph of draft in `FULL` mode is broken now.
This pr solves it.
Also, `actual_seq_lengths_q` in `model_runner` is found redundant, so,
it is removed.
It depends on https://github.com/vllm-project/vllm-ascend/pull/7144 and
https://github.com/vllm-project/vllm-ascend/pull/7148.
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Test code is shown as below:
```python
prompts = [
"1.Who are you?",
"2. Who are you?",
]
sampling_params = SamplingParams(temperature=0.0, top_p=0.95, top_k=40, max_tokens=200)
llm = LLM(
model="/home/some-model/Meta-Llama-3.1-8B-Instruct",
tensor_parallel_size=1,
max_num_seqs=32,
# enforce_eager=True,
disable_log_stats=False,
distributed_executor_backend="mp",
gpu_memory_utilization=0.7,
async_scheduling=True,
speculative_config={
"enforce_eager": True,
"model": "/home/some-model/EAGLE3-LLaMA3.1-Instruct-8B",
"disable_padded_drafter_batch": False,
"method": "eagle3",
"num_speculative_tokens": 3,
},
compilation_config={
"cudagraph_mode": "FULL",
"cudagraph_num_of_warmups": 1,
},
max_model_len=4096,
enable_prefix_caching=False,
)
outputs = llm.generate(prompts, sampling_params)
```
The result before:
```text
File "/vllm-workspace/vllm-ascend/vllm_ascend/attention/attention_v1.py", line 575, in full_graph_fia
graph_params.events[num_tokens].append(event)
~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^
KeyError: 132
```
The result after:
```text
--------------------------------------------------
total_num_output_tokens: 400
num_drafts: 242
num_draft_tokens: 726
num_accepted_tokens: 156
mean acceptance length: 1.64
--------------------------------------------------
acceptance at token 0: 0.42
acceptance at token 1: 0.16
acceptance at token 2: 0.07
```
We also test `FULL_DECODE_ONLY` mode.
The result is:
```text
--------------------------------------------------
total_num_output_tokens: 400
num_drafts: 244
num_draft_tokens: 732
num_accepted_tokens: 155
mean acceptance length: 1.64
--------------------------------------------------
acceptance at token 0: 0.42
acceptance at token 1: 0.16
acceptance at token 2: 0.06
```
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
Fix lint failed due to the merging of a previous PR.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: wangli <wangli858794774@gmail.com>
### What this PR does / why we need it?
On the 310P device, when running ACLGraph together with the n-gram
speculative decoding algorithm, both graph capture and graph replay
require `uniform_decode_query_len` and do not depend on
`attention_state`. This leads to a rather interesting and unexpected
issue on 310P: during decode-only, execution does **not** enter the
graph, while in the split-fuse state (that is, the chunked prefill
state), it instead enters graph execution directly.
The issue can be resolved by forcibly setting `uniform_decode_query_len`
to `1`, so that 310P captures only the decode-only graph, and replay is
then controlled through `attention_state`.
### Does this PR introduce _any_ user-facing change?
NO
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Tflowers-0129 <2906339855@qq.com>
### What this PR does / why we need it?
Fix the issue where no exception is thrown when graph capture fails.
- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef
Signed-off-by: WithHades <244036962@qq.com>
### What this PR does / why we need it?
Currently when tp>1, we have extra processes on tp rank0 device which
consumes extra HBM memory. This is caused by `import
torch_npu._inductor` before set_device which introduces extra
initialization of device.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
All ci passed.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: whx-sjtu <2952154980@qq.com>
# Feature: FlashLB algorithm
## Purpose
This Pull Request enhances the EPLB (Expert Parallelism Load Balancing)
system by introducing a novel load balancing algorithm: FlashLB.
1. The default algorithm adopts two separate sub-procedures to optimize
expert replication and placement independently:
a. **Expert Replica Allotment Sub-procedure** : Determines the number of
replicas for all experts. At each step, it greedily adds one more
replica to the expert with the highest per-replica load, aiming to
minimize load skew at the expert replica granularity (Min Max Replica,
MMR).
b. **Expert Replica Placement Sub-procedure** : Distributes all replicas
across devices. First, it sorts the generated replicas in descending
order of hotness, then iteratively places the currently hottest replica
onto the device with the lowest cumulative load and available slots.
However, this simplistic combination of two separate procedures lacks
synergy and often leads to sub-optimal load balancing. For example, in
the simple scenario illustrated below: Given 8 logical experts with
hotness values [600, 560, 120, 120, 20, 10, 10, 10], and 2 replicas
allocated per device across 8 devices, the default EPLB algorithm
results in a maximum per-device hotness of 232 (peak-average load ratio
1.28), while our proposed FlashLB algorithm reduces this value to 205
(peak-average load ratio 1.13).
<figure><img
src="https://github.com/user-attachments/assets/b9b10fab-651e-4524-9942-adbca8d044a4"
width="90%"</figure>
2. The default algorithm simply aggregates hotness measurements across
the entire profiling window. While this provides a coarse approximation
of the hotness distribution, it fails to capture the time-phased
variations and temporal correlations in expert hotness (both within and
between experts) across iterations—phenomena that have been observed in
real-world scenarios. Such single-point hotness estimation degrades the
solution quality of the load balancing algorithm.
3. The default algorithm regularly recalculates updated expert placement
results for all layers without discrimination. Considering that
excessive expert updates can impact Service Level Objectives (SLOs),
such full-scale redeployment leads to excessively high adjustment
overhead, which negatively affects end-to-end performance.
## FlashLB Algorithm Principle
### 1. Joint Optimization of Replica Allotment and Placement
FlashLB achieves joint optimization of replica allotment and placement
through a novel tree search approach, combined with carefully designed e
Fl fficient pruning and lightweight look-ahead estimation. We partition
all experts into several subsets, and for each subset, hierarchically
determine the optimal replica count and placement. Leveraging efficient
pruning and lightweight look-ahead estimation, the process consistently
aims to optimize the globally expected inter-device load balance degree
(considering both deployed and unexplored experts) while ensuring
sufficient computational efficiency. Additionally, precompilation
techniques are employed for acceleration, delivering load balancing that
is both high-quality and practically efficient.
### 2. Multi-Episode Enhancement
Instead of performing full-duration averaging like the default
algorithm, FlashLB partitions each profiling interval (e.g., 1024
iterations) into multiple consecutive smaller episodes (e.g., 16
iterations). This preserves hotness fluctuation and correlation
information. It then constructs a multi-objective optimization problem
to co-optimize these episodes simultaneously, enabling adaptability to
interleaved hotness patterns and improving statistical robustness.
### 3. Layer-wise Cherry-Picking Redeployment
To reduce the overhead of frequent expert redeployment, FlashLB
introduces a cherry-picking redeployment scheme. During each algorithmic
decision cycle, it real-time tracks load balance degree of all layers
and triggers expert placement updates only for those layers whose
peak-average ratio exceeds a predefined threshold. This avoids
unnecessary redeployment for stable layers, significantly reducing
adjustment overhead and thereby improving end-to-end performance gains.
## Co-author:
Co-authored-by: Skywalker-EP 173723846@qq.com
This PR mainly introduces two key optimizations for load balancing
scheduling:
1. **Add per-step heat collection function**:
Support real-time collection of per-step heat information during model
inference. This enables more fine-grained load balancing decisions by
taking per-step heat as the optimization target, improving scheduling
accuracy for dynamic and fluctuating workloads.
2. **Update FlashLB algorithm**:
Upgrade the FlashLB scheduling logic to better adapt to multi-stage heat
distribution scenarios. The improved algorithm can comprehensively
perceive and utilize multi-stage heat characteristics, achieving more
stable and efficient load balancing under complex expert deployment and
dynamic traffic patterns.
---------
Signed-off-by: Mercykid-bash <ruanche0218@gmail.com>
Signed-off-by: xuzewei28 <xuzewei2@h-partners.com>
Co-authored-by: xuzewei28 <xuzewei2@h-partners.com>
…the quantization layer name
### What this PR does / why we need it?
This PR modifies the loading logic for layer name prefixes in quantized
models. The goal is to reduce or eliminate the need for point-to-point
(hardcoded) modifications by leveraging the built-in mapper mechanism
already provided in vLLM's model code. For models that do not yet have a
corresponding mapper, the original point-to-point modification approach
has been retained to ensure backward compatibility.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
The changes were validated using an offline deployment script to launch
and verify multiple multimodal models. Testing confirmed that the
updated loading logic correctly handles layer name prefixes across
different model architectures, with no regression in model
initialization or inference behavior.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Matrix_K <zhangke144@huawei.com>
Signed-off-by: Feng-xiaosuo <tengchang1@huawei.com>
Co-authored-by: Matrix_K <zhangke144@huawei.com>
### What this PR does / why we need it?
Fixed the error of speculative decoding in FULL mode when `num_spec + 1`
not in `cudagraph_capture_sizes`.
Now, we can run speculative decoding in FULL mode, but with drafter as
eager.
It depends on https://github.com/vllm-project/vllm-ascend/pull/7144 .
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Test code is shown as below:
```python
prompts = [
"1.Who are you?",
"2. Who are you?",
]
sampling_params = SamplingParams(temperature=0.0, top_p=0.95, top_k=40, max_tokens=200)
llm = LLM(
model="/home/some-model/Meta-Llama-3.1-8B-Instruct",
tensor_parallel_size=1,
max_num_seqs=32,
# enforce_eager=True,
disable_log_stats=False,
distributed_executor_backend="mp",
gpu_memory_utilization=0.7,
async_scheduling=True,
speculative_config={
"enforce_eager": True,
"model": "/home/some-model/EAGLE3-LLaMA3.1-Instruct-8B",
"disable_padded_drafter_batch": False,
"method": "eagle3",
"num_speculative_tokens": 2,
},
compilation_config={
"cudagraph_mode": "FULL",
"cudagraph_num_of_warmups": 1,
},
max_model_len=4096,
enable_prefix_caching=False,
)
outputs = llm.generate(prompts, sampling_params)
```
The result before:
```text
File "/vllm-workspace/vllm/vllm/v1/cudagraph_dispatcher.py", line 140, in _create_padded_batch_descriptor
assert num_tokens_padded % uniform_decode_query_len == 0
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError
```
The result after:
```text
--------------------------------------------------
total_num_output_tokens: 400
num_drafts: 249
num_draft_tokens: 498
num_accepted_tokens: 149
mean acceptance length: 1.60
--------------------------------------------------
acceptance at token 0: 0.43
acceptance at token 1: 0.17
```
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
### What this PR does / why we need it?
related to vllm PR #34043 this pr delete func
‘relax_for_mixed_batch_cudagraphs’, num_reqs no longer equals the actual
number of requests, due to fia operator requires that
query_start_loc[-1] equals the total number of computed tokens, so this
func delete cause the ifa error.
In full graph mode, set num_reqs_paded = num_reqs to fix the error
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
### What this PR does / why we need it?
We should only display this log message when layer_sharding is enabled.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
### What this PR does / why we need it?
This PR adds support for the Ascend950 chip. This includes:
- Updating build scripts (`CMakeLists.txt` and `setup.py`) to recognize
the Ascend950 chip and set appropriate compilation flags.
- Disabling a set of custom operators that are not yet supported on the
Ascend950 hardware target.
- Performing a codebase-wide refactoring of `pipe_barrier()` calls to
the namespaced `AscendC::PipeBarrier<>()` for improved code consistency
and adherence to the latest API standards.
Ascend950DT e2e passed (Qwen3-32B-MXFP8) and CI passed
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
### What this PR does / why we need it?
fix penality ops for new version, and achieved a 10% performance
improvement
### How was this patch tested?
pytest
tests/e2e/nightly/single_node/ops/singlecard_ops/triton/test_penality.py
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
Signed-off-by: shiyuan680 <917935075@qq.com>
### What this PR does / why we need it?
Fix the issue #6143 .
### Does this PR introduce _any_ user-facing change?
Allow to start the server with "--enable-lora && --fully-sharded-loras
&& --tensor_parallel_size 2".
### How was this patch tested?
pytest -sv tests/e2e/multicard/2-cards/test_llama32_lora_tp2.py
- vLLM version: v0.15.0
- vLLM main:
d7e17aaacd
---------
Signed-off-by: paulyu12 <507435917@qq.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
### What this PR does / why we need it?
Currently, we call lookup_client for looking up token hit in KV Pool,
however, when token length < block size, the key will be empty and there
is no point to lookup in KV Pool backend since there will never be a
hit.
Hence, add early return in `get_num_new_matched_tokens` when `token_len`
< `block_size`
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Pz1116 <zpbzpb123123@gmail.com>
Co-authored-by: fems14 <1804143737@qq.com>
### What this PR does / why we need it?
Fix the wrong usage of `model_type`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
By CI.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
### What this PR does / why we need it?
Initial version to support minimax-m2.5 on vllm-ascend.
This commit coverting original fp8 weight to a quantilized bf16 to
support Minimax-m2.5 on NPU.
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
### Test Report
Self tested precision summary, where the official precision score of
AIME2025 is 86.3
<img width="426" height="84" alt="image"
src="https://github.com/user-attachments/assets/a3ce2452-92fa-4713-962e-862248e0b61a"
/>
---------
Signed-off-by: limuyuan <limuyuan3@huawei.com>
Signed-off-by: SparrowMu <52023119+SparrowMu@users.noreply.github.com>
Co-authored-by: limuyuan <limuyuan3@huawei.com>
### What this PR does / why we need it?
Mooncake Layerwise Connector supports hybrid attention manager with
multiple kvcache groups.
### Does this PR introduce _any_ user-facing change?
Yes.
### How was this patch tested?
By CI.
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>
### What this PR does / why we need it?
The ops `torch_npu.npu_recurrent_gated_delta_rule` currently does not
support `ssm_state` inputs in float32 format,
we temporarily retain the _forward_core implementation with triton for
Qwen3_5
---------
Signed-off-by: pppeng <zepengliu912@qq.com>
Signed-off-by: pppeng <60355449+ppppeng@users.noreply.github.com>
### What this PR does / why we need it?
This is a bug fix to resolve the issue where the MOE model fails to load
quantized weights in w4a8 format when EP is not enabled.The parameters
["weight_scale_second", "weight_offset_second", "scale_bias"] shall be
parsed in per-group mode, regardless of other conditions.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: 李少鹏 <lishaopeng21@huawei.com>
### What this PR does / why we need it?
Fix acceptance and high-concurrency bug in eagle3 and cp enabled
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
tests and ut
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: lilinsiman <lilinsiman@gmail.com>
### What this PR does / why we need it?
New Quantization Method: Introduced support for the W8A8SC static linear
quantization scheme specifically for 310P hardware, enabling more
efficient model compression.
Refactored the save_sharded_state_310.py to avoid multi-process issue.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
W8A8SC quant E2E test.
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: pu-zhe <zpuaa@outlook.com>
### What this PR does / why we need it?
Fix the LoRA e2e test accuracy issue that introduced by the upstream PR
https://github.com/vllm-project/vllm/pull/32005
### How was this patch tested?
pytest -sv tests/e2e/singlecard/test_llama32_lora.py
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
---------
Signed-off-by: paulyu12 <507435917@qq.com>
Signed-off-by: yupeng <507435917@qq.com>
