tmplate

[floatlibai]

[floatlibai]

Wan-AI/Wan2.2-TI2V-5B-Diffusers

[floatlibai]

GPU基线
精度
修改 /vllm-omni/examples/offline_inference/text_to_video/text_to_video.py，使用 CPU 的随机数 generator，否则会出现 NPU/GPU 画面风格严重不一致的问题，对测试的人工校验增加困难：
image
在 /vllm-omni/examples/offline_inference/text_to_video/ 目录下创建一个 shell 脚本：
python text_to_video.py
--prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage."
--height 720
--width 1280
--model /data/zwc/weights/Wan-AI/Wan2___2-TI2V-5B-Diffusers/
--num-frames 81
--seed 42
--guidance-scale 4.0
--guidance-scale-high 3.0
--num-inference-steps 40
--fps 16
--output t2v_out.mp4
​
执行该 shell 脚本结束后，会在同目录下生成对应视频。

结果
image

NPU自测结果
精度
修改测试脚本 /vllm-omni/examples/offline_inference/text_to_video/text_to_video.py，使用 CPU 随机数 generator，否则会因NPU/GPU generator 逻辑不一致导致 NPU/GPU 画面风格严重不一致的问题：
image
在 /vllm-omni/examples/offline_inference/text_to_video/ 目录下创建一个 shell 脚本：

python text_to_video.py \
  --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
  --height 720 \
  --width 1280 \
  --model /data/zwc/weights/Wan-AI/Wan2___2-TI2V-5B-Diffusers/ \
  --num-frames 81 \
  --seed 42 \
  --guidance-scale 4.0 \
  --guidance-scale-high 3.0 \
  --num-inference-steps 40 \
  --fps 16 \
  --output t2v_out.mp4

[floatlibai]

[floatlibai]

--use-hsdp --hsdp-shard-size 2

[floatlibai]

VLLM_BATCH_INVARIANT=1 pip install -v -e . --no-build-isolation --no-deps

[floatlibai]

whl包安装
curl -k -O https://ascend-cann-open.obs.cn-north-4.myhuaweicloud.com/ops-batchinvariant/beta/20260603/batch_invariant-torch_ops_extension-1.0.0.zip
unzip batch_invariant-torch_ops_extension-1.0.0.zip
cd torch_ops_extension/batch_invariant_ops/
bash build_and_install.sh

[floatlibai]

vLLM安装

vllm

pip install -r requirements.txt --extra-index-url=https://triton-ascend.osinfra.cn/pypi/simple --trusted-host triton-ascend.osinfra.cn

export COMPILE_CUSTOM_KERNELS=1

pip install -e .  --extra-index-url=https://triton-ascend.osinfra.cn/pypi/simple --trusted-host triton-ascend.osinfra.cn

another

export VLLM_TARGET_DEVICE=empty
python3 use_existing_torch.py
pip install -r requirements/common.txt
pip install -e . --no-build-isolation

旧版本

export VLLM_TARGET_DEVICE=empty
pip install -r requirements/build.txt
pip install -r requirements/common.txt

# 安装vllm
pip install -e .

ascend

cd ../vllm-ascend
pip install -r requirements.txt
export COMPILE_CUSTOM_KERNELS=1
python setup.py develop

["'rollout/raw_reward':", 'train/train_rollout_logprob_abs_diff':", "'perf/rollout_time':", "'perf/train_time':", "'perf/actor_train_time': "]

["'rollout/raw_reward':", "'train/train_rollout_logprob_abs_diff':", "'perf/rollout_time':", "'perf/train_time':", "'perf/actor_train_time':", "'perf/step_time':"]