# Poetry Video Workspace

This workspace is pinned to `/apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang`.
All generated outputs, local models, ComfyUI, and Bernini deployments should stay inside this directory.

## Mount notes

- `/apdcephfs_sgxy/share_304169095` is a shared Ceph mount.
- `/apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang` is writable.
- `/apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/Models` is the shared read-only model mount.
- `/apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/.git` is mounted read-only, so normal git writes fail there.
- This workspace uses `.repo_git` as the writable git directory:

```bash
git --git-dir=.repo_git --work-tree=. status
git --git-dir=.repo_git --work-tree=. log --oneline --decorate -n 5
```

## Path resolution

Python and shell entrypoints resolve model paths in this order:

1. Explicit env vars such as `QWEN_IMAGE_MODEL`, `WAN_VACE_MODEL`, `QWEN35_27B_MODEL`.
2. Local model copies under `ozhang/models`.
3. Shared fallback under `/apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/Models`.

## Start the local stack

Start the local Qwen 3.5 planning service:

```bash
bash /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/VGen/scripts/deploy/start_qwen35_27b.sh
```

Run the main agent-loop pipeline with 2-minute-oriented defaults:

```bash
RUN_NAME=poetry_full_local_$(date +%Y%m%d_%H%M%S)
bash /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/VGen/scripts/run_test.sh   --run-name "$RUN_NAME"
```

Direct invocation is also supported:

```bash
bash /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/VGen/scripts/run_poetry_teaching_pipeline-agentloop.sh   --poem "枫桥夜泊
张继
月落乌啼霜满天，江枫渔火对愁眠。
姑苏城外寒山寺，夜半钟声到客船。"   --run-name poetry_full_local_manual   --segments 8   --clip-seconds 14   --target-total-seconds 120   --min-segment-seconds 12   --max-segment-seconds 16   --image-device cuda:0   --video-usp-devices 1,2,3,4,5   --overwrite
```

## Bernini local deployment

The uploaded archive has been unpacked to `Bernini-main/`.
Use an isolated local env at `ozhang/venv/bernini`; the wrappers under `scripts/bernini/` now auto-prefer it when present.

```bash
bash /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/scripts/bernini/install_venv.sh
```

`install_venv.sh` installs the pinned CUDA 12.4 PyTorch stack, VeOmni, Bernini editable package, and `flash-attn==2.8.3` by default. If you only want the Python stack first, run it with `INSTALL_FLASH_ATTN=0`.

Recommended local weight locations:

```bash
hf download ByteDance/Bernini-Diffusers --local-dir /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/models/Bernini-Diffusers
hf download ByteDance/Bernini-R-Diffusers --local-dir /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/models/Bernini-R-Diffusers
```

Local wrappers are available under `scripts/bernini/`. By default they keep the current proxy settings; if your shell injects the broken `star-proxy.oa.com:3128` proxy inside a sandboxed runtime, run them with `BERNINI_CLEAR_PROXY=1`:

```bash
bash /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/scripts/bernini/download_models.sh all
bash /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/scripts/bernini/run_full_t2v.sh
PORT=9500 SHARE=1 bash /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/scripts/bernini/run_full_gradio.sh
PORT=7860 SHARE=1 bash /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/scripts/bernini/run_renderer_gradio.sh
```

Direct launches are still available:

```bash
cd /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/Bernini-main
BERNINI_CONFIG=/apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/models/Bernini-Diffusers   bash scripts/bernini/run_t2v.sh

BERNINI_R_CONFIG=/apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/models/Bernini-R-Diffusers   bash scripts/bernini_r/run_gradio.sh
```

## ComfyUI local deployment

The uploaded archive has been unpacked to `ComfyUI-master/`.
A tracked source config now lives at `configs/comfyui/extra_model_paths.yaml`, and `ComfyUI-master/extra_model_paths.yaml` points to it.
Search order is local `ozhang/models` first, shared Ceph mount second.

```bash
cd /apdcephfs_sgxy/share_304169095/hunyuan/juliayxzhu/dev/ozhang/ComfyUI-master
python main.py --listen 127.0.0.1 --port 8188
```