test-conditioner / README.md
multimodalart's picture
multimodalart HF Staff
Tidy comments and docstrings
90d53e0 verified
|
Raw History Blame Contribute Delete
1.81 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade
metadata
title: Qwen3-VL conditioner
emoji: 🧠
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
pinned: false
short_description: Qwen3-VL 33B prompt conditioning as a service
suggested_hardware: zero-a10g

Qwen3-VL conditioner

Encodes prompts (optionally with keyframes or ordered media references) into conditioning tensors over the gradio API, running the 33B Qwen3-VL conditioner unquantized in bfloat16 on ZeroGPU. Built as the text encoding half of a split deployment for large video generation models.

Both endpoints take an optional trailing rewrite_prompt (default False). Turned on, the conditioner first writes the structured prompt format the video model was trained on — a deterministic task instruction line, then a shot-by-shot timeline and the two sound fields, written by the conditioner's own language-model head while it looks at the keyframes or references — and encodes that instead. The rewrite comes back under the plan's refined_prompt, and a rewrite that fails falls back to encoding the raw prompt rather than failing the call.

Where the code comes from

The weights are the public MiniMaxAI/MiniMax-H3 diffusers checkpoint — only text_encoder/, tokenizer/ and processor/ are ever fetched here.

diffusers is installed from the canonical MiniMax-H3 pull request, huggingface/diffusers#14371, pinned to the commit 665f5782 (refs/pull/14371/head) rather than to the moving minimax-h3-refactor branch. That PR is a WIP, so it needs re-pinning whenever it updates, and h3_split_blocks.py — which subclasses its block classes to cut the pipeline in two — has to be re-checked against the new head at the same time.