1k77 comments

Results 8 comments of


                                            1k77

vllm version error about Ascend NPU

Please provide detailed environment information.

Shape mismatch on vllm+fsdp on ascend npus (A3 node) with qwen-30B model

You can refer to the following issue: https://github.com/volcengine/verl/issues/3862 . If the problem cannot be resolved, please provide detailed environment information.

bash recipe/dapo/run_dapo_qwen3_8b_base_npu.sh报错

Please provide detailed environment information.

FP16 training has an anomaly on the NPU.

@glowwormX Hi! I think the issue might be related to the versions of vllm or vllm-ascend, so I pulled the latest branch to see if the problem persists. > vllm:...

FP16 training has an anomaly on the NPU.

> [@1k77](https://github.com/1k77) Okay, I'll give it a try. Did you manage to reproduce it on v0.11.0? Yes, I did reproduce it on v0.11.0 with DeepSeek-R1-Distill-Qwen-1.5B.

What situations could cause the 'Memory usage increased after sleeping' problem?

Hi, please provide your script.

Repeated output when training Qwen2.5-VL

Please provide your script.

RuntimeError: Tensor size mismatch on apply_multimodal_rotary_pos_emb with ulysses_sequence_parallel_size=2

Please provide your detailed environment informaiton.