---
license: apache-2.0
tags:
  - conda-env
  - reproducibility
  - avsqwen
  - qwen2.5-omni
pretty_name: AVSQwen Conda Environments
---

# AVSQwen Packed Conda Environments

Pre-packed conda environments for reproducing [AVSQwen](https://github.com/<your-org>/AVSQwen)
training & inference. Extract and activate; **no `pip install` needed**.

| File | Size | Purpose |
|---|---|---|
| `avsqwen_swift.tar.gz` | 4.11 GB | Main env: ms-swift + vllm + GRPO training + inference |
| `cod.zip` | 8.42 GB | Auxiliary env: COD / dataset preprocessing |

## Usage

```bash
# Download (or: `huggingface-cli download ...`)
huggingface-cli download Vegetabot/AVSQwen-Envs --repo-type dataset --local-dir envs/

# 1. avsqwen_swift (main env: ms-swift + vllm + GRPO training + inference)
mkdir -p ~/envs/avsqwen_swift
tar -xzf envs/avsqwen_swift.tar.gz -C ~/envs/avsqwen_swift
source ~/envs/avsqwen_swift/bin/activate
conda-unpack || true        # fix shebangs if the archive was built via conda-pack

# 2. cod (auxiliary env for COD / dataset preprocessing)
mkdir -p ~/envs/cod
unzip envs/cod.zip -d ~/envs/cod
source ~/envs/cod/bin/activate
conda-unpack || true
```

> These archives target **linux-x86_64 + CUDA 12.x**. If your system differs,
> rebuild from the `pip` / `conda` specs in the AVSQwen repo instead.

## Integrity

SHA-256 checksums (also stored in `SHA256SUMS.txt`):

```
34fc367009ff8291f99cceb79181ca3cd52b2fae9f15947153efb6669edfe320  avsqwen_swift.tar.gz
ae62fe9801981716404d5e310b98b17a802fe4e29ff5f8373b810d50a6d662a8  cod.zip
```