|
1 | 1 | # Release note |
2 | 2 |
|
| 3 | +## v0.10.2rc1 - 2025.09.16 |
| 4 | + |
| 5 | +This is the 1st release candidate of v0.10.2 for vLLM Ascend. Please follow the [official doc](https://vllm-ascend.readthedocs.io/en/) to get started. |
| 6 | + |
| 7 | +### Highlights |
| 8 | + |
| 9 | +- Add support for Qwen3 Next. Please note that expert parallel and MTP feature doesn't work with this release. We'll make it work enough soon. Follow the [official guide](https://vllm-ascend.readthedocs.io/en/latest/tutorials/multi_npu_qwen3_next.html) to get start [#2917](https://github.com/vllm-project/vllm-ascend/pull/2917) |
| 10 | +- Add quantization support for aclgraph [#2841](https://github.com/vllm-project/vllm-ascend/pull/2841) |
| 11 | + |
| 12 | +### Core |
| 13 | + |
| 14 | +- Aclgraph now works with Ray backend. [#2589](https://github.com/vllm-project/vllm-ascend/pull/2589) |
| 15 | +- MTP now works with the token > 1. [#2708](https://github.com/vllm-project/vllm-ascend/pull/2708) |
| 16 | +- Qwen2.5 VL now works with quantization. [#2778](https://github.com/vllm-project/vllm-ascend/pull/2778) |
| 17 | +- Improved the performance with async scheduler enabled. [#2783](https://github.com/vllm-project/vllm-ascend/pull/2783) |
| 18 | +- Fixed the performance regression with non MLA model when use default scheduler. [#2894](https://github.com/vllm-project/vllm-ascend/pull/2894) |
| 19 | + |
| 20 | +### Other |
| 21 | +- The performance of w8a8 quantization is improved. [#2275](https://github.com/vllm-project/vllm-ascend/pull/2275) |
| 22 | +- The performance of moe model is improved. [#2689](https://github.com/vllm-project/vllm-ascend/pull/2689) [#2842](https://github.com/vllm-project/vllm-ascend/pull/2842) |
| 23 | +- Fixed resources limit error when apply speculative decoding and aclgraph. [#2472](https://github.com/vllm-project/vllm-ascend/pull/2472) |
| 24 | +- Fixed the git config error in docker images. [#2746](https://github.com/vllm-project/vllm-ascend/pull/2746) |
| 25 | +- Fixed the sliding windows attention bug with prefill. [#2758](https://github.com/vllm-project/vllm-ascend/pull/2758) |
| 26 | +- The official doc for Prefill Decode Disaggregation with Qwen3 is added. [#2751](https://github.com/vllm-project/vllm-ascend/pull/2751) |
| 27 | +- `VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP` env works again. [#2740](https://github.com/vllm-project/vllm-ascend/pull/2740) |
| 28 | +- A new improvement for oproj in deepseek is added. Set `oproj_tensor_parallel_size` to enable this feature[#2167](https://github.com/vllm-project/vllm-ascend/pull/2167) |
| 29 | +- Fix a bug that deepseek with torchair doesn't work as expect when `graph_batch_sizes` is set. [#2760](https://github.com/vllm-project/vllm-ascend/pull/2760) |
| 30 | +- Avoid duplicate generation of sin_cos_cache in rope when kv_seqlen > 4k. [#2744](https://github.com/vllm-project/vllm-ascend/pull/2744) |
| 31 | +- The performance of Qwen3 dense model is improved with flashcomm_v1. Set `VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE=1` and `VLLM_ASCEND_ENABLE_FLASHCOMM=1` to enable it. [#2779](https://github.com/vllm-project/vllm-ascend/pull/2779) |
| 32 | +- The performance of Qwen3 dense model is improved with prefetch feature. Set `VLLM_ASCEND_ENABLE_PREFETCH_MLP=1` to enable it. [#2816](https://github.com/vllm-project/vllm-ascend/pull/2816) |
| 33 | +- The performance of Qwen3 MoE model is improved with rope ops update. [#2571](https://github.com/vllm-project/vllm-ascend/pull/2571) |
| 34 | +- Fix the weight load error for RLHF case. [#2756](https://github.com/vllm-project/vllm-ascend/pull/2756) |
| 35 | +- Add warm_up_atb step to speed up the inference. [#2823](https://github.com/vllm-project/vllm-ascend/pull/2823) |
| 36 | +- Fixed the aclgraph steam error for moe model. [#2827](https://github.com/vllm-project/vllm-ascend/pull/2827) |
| 37 | + |
| 38 | +### Known issue |
| 39 | +- The server will be hang when running Prefill Decode Disaggregation with different TP size for P and D. It's fixed by [vLLM commit](https://github.com/vllm-project/vllm/pull/23917) which is not included in v0.10.2. You can pick this commit to fix the issue. |
| 40 | +- The HBM usage of Qwen3 Next is higher than expected. It's a [known issue](https://github.com/vllm-project/vllm-ascend/issues/2884) and we're working on it. You can set `max_model_len` and `gpu_memory_utilization` to suitable value basing on your parallel config to avoid oom error. |
| 41 | +- We notice that lora doesn't work with this release due to the refactor of kv cache. We'll fix it soon. [2941](https://github.com/vllm-project/vllm-ascend/issues/2941) |
| 42 | +- Please do not enable chunked prefill with prefix cache when running with Ascend scheduler. The performance and accuracy is not good/correct. [#2943](https://github.com/vllm-project/vllm-ascend/issues/2943) |
| 43 | + |
3 | 44 | ## v0.10.1rc1 - 2025.09.04 |
4 | 45 |
|
5 | 46 | This is the 1st release candidate of v0.10.1 for vLLM Ascend. Please follow the [official doc](https://vllm-ascend.readthedocs.io/en/) to get started. |
|
0 commit comments