2301_76761127/cryofm-v2-npu
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

CryoFM v2 on Ascend NPU

This repository provides a minimal, auditable Ascend NPU inference entry point for the official ByteDance-Seed/cryofm-v2 pretraining checkpoint. It constructs the upstream CryoFM 3D UNet, loads the complete safetensors checkpoint strictly, and runs one full 3D forward pass on NPU.

Run

Install the listed Python dependencies without replacing the platform-provided torch and torch_npu, then run:

git clone https://github.com/ByteDance-Seed/cryofm.git cryofm-source
git -C cryofm-source checkout --detach 64486814723010bd200525cda2fa463a25ba1339
git clone https://gitcode.com/hf_mirrors/ByteDance-Seed/cryofm-v2.git model
git -C model checkout --detach 4e308f7f028af46ca2c7ee5af81e29775bc370dd
git -C model lfs pull
python3 -m pip install --target ./pydeps -r requirements.txt
source /usr/local/Ascend/ascend-toolkit/set_env.sh
PYTHONPATH="$PWD/pydeps" python inference.py \
  --model-dir "$PWD/model" \
  --source-dir "$PWD/cryofm-source" \
  --device npu:0 \
  --output result.json

The public source checkout is fixed to commit 64486814723010bd200525cda2fa463a25ba1339. The hosted rerun used CANN 8.5.1 with platform PyTorch 2.9.0 and torch_npu 2.9.0; requirements intentionally leave that pair untouched.

The JSON result records the requested, model, input, and output devices together with tensor shapes, value counts, finite/non-finite counts, and wall-clock forward time. The entry point intentionally rejects non-NPU devices and contains no CPU or CUDA inference fallback.

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

The device views show the same Python PID at two distinct times. The result views show complete and concise views of the same successful real-NPU run.

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)。