2301_76761127/x-vla-widowx-npu
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

X-VLA-WidowX on Ascend NPU

This is the sole offline NPU inference entry for 2toINF/X-VLA-WidowX, fixed revision 8d7ea1aaa948665d44129a3ff488629b955fc0f9. Place that exact snapshot at ./model; model weights are not committed here.

The verified back-half run used Ascend 910B3 (npu:1) with Python 3.11, CANN 8.5.1, PyTorch 2.9.0, torch_npu 2.9.0.post1, transformers 4.57.6, and timm 1.0.20. The entry loads the complete local XVLA checkpoint and executes generate_actions, including the vision-language backbone and one action-diffusion denoising step, entirely on NPU. The recorded output shape is [1, 30, 20] with 600 values and exit status 0.

git clone https://gitcode.com/hf_mirrors/2toINF/X-VLA-WidowX.git model
git -C model checkout --detach 8d7ea1aaa948665d44129a3ff488629b955fc0f9
git -C model lfs pull
python3 -m pip install --no-deps --target ./pydeps -r requirements.txt
source /usr/local/Ascend/ascend-toolkit/set_env.sh
PYTHONPATH="$PWD/pydeps:$PYTHONPATH" python inference.py --model-path ./model --device npu:0 --output audit/npu_result.json

CPU/CUDA fallback is forbidden. The result records actual model/input/output devices, parameter count, output shape/value count and exit status. Precision and finiteness are recorded honestly but are non-blocking under the current environment-pass policy. Private logs remain outside this public repository.

The FastAPI, Uvicorn, json_numpy, and OpenCV modules imported by the published serving code are not used by this offline forward. The entry provides narrow import-only stubs for those modules instead of requiring unrelated server packages.

Verification evidence

The two DS-V4-Flash workflow views are independent plan-only conversations: an overall Ascend NPU migration design and a separate dependency, operator, full-NPU validation, and troubleshooting plan. They do not execute tools or modify repository files.

DS-V4-Flash overall Ascend NPU migration plan DS-V4-Flash dependency, operator, validation and troubleshooting plan NPU device call, first time point NPU device call, second time point Model result, full view Model result, concise view

本仓为 2301_76761127 队伍交付件(与备份账号合并提交)。