vllm-omni-humming 0.2.0

Companion plugin: load & serve pre-quantized humming Qwen-Image DiT pipelines on stock vLLM-Omni

vllm-omni-humming

Plugin that lets stock vLLM-Omni serve pre-quantized humming Qwen-Image checkpoints.

Install it into the vLLM-Omni environment and serve normally — the plugin registers itself through vLLM-Omni's standard plugin entry point, and quantized checkpoints are detected automatically from their transformer/config.json. No flags, no environment variables, no configuration.

Compatibility

vllm-omni-humming	vLLM-Omni	humming-kernels	vLLM	Python
0.2.0	0.24.x	>=0.1.11	0.24.0	>=3.10,<3.14

humming-kernels >= 0.1.11 is required (older versions produce corrupted images on some GPUs). Install it explicitly as shown below, or via this package's [kernels] extra.

What it patches

Five small, humming-only runtime patches against stock vLLM-Omni; other quantization methods are untouched. Each is a separate source module documenting the exact upstream symbol it targets:

• _preimport.py — import the real humming package before vLLM-Omni's stub.
• _factory.py — register humming as a quantization method for checkpoints with a quantization_config.
• _loader.py — load packed weights on CPU and repack per module on GPU (keeps the load peak low enough for 32 GiB cards), and accept humming's tensor-name suffixes.
• _modpin.py — route quantized modulation layers (img_mod.1/txt_mod.1), which stock vLLM-Omni keeps in bf16.

Install

Into the same venv that runs vllm-omni serve. Shown for CUDA 12.9 wheels; adapt the wheel URLs for your platform.

python -m venv venv && source venv/bin/activate
pip install -U pip

# 1) torch trio (cu129)
pip install torch==2.11.0 torchvision==0.26.0 torchaudio==2.11.0 \
  --index-url https://download.pytorch.org/whl/cu129

# 2) vLLM 0.24.0 cu129 wheel
printf 'torch==2.11.0+cu129\ntorchvision==0.26.0+cu129\ntorchaudio==2.11.0+cu129\n' > constraints.txt
PIP_CONSTRAINT=constraints.txt pip install \
  https://github.com/vllm-project/vllm/releases/download/v0.24.0/vllm-0.24.0+cu129-cp38-abi3-manylinux_2_28_x86_64.whl

# 3) vLLM-Omni 0.24.0
PIP_CONSTRAINT=constraints.txt pip install vllm-omni==0.24.0

# 4) this plugin
PIP_CONSTRAINT=constraints.txt pip install vllm-omni-humming==0.2.0

# 5) humming kernels (required)
pip install --no-deps "humming-kernels>=0.1.11"

Requirements: NVIDIA GPU (SM75+), driver ≥ 575. No CUDA toolkit needed — kernels are JIT-compiled on first serve (a few minutes once, cached afterwards).

Serve

vllm-omni serve /path/to/checkpoint \
  --omni --served-model-name Qwen/Qwen-Image-2512 \
  --enable-cpu-offload \
  --port 8124 --api-key "$KEY"

Keep --enable-cpu-offload on GPUs with ≤ 32 GiB; drop it on larger cards for full speed. Generate an image (OpenAI images API):

curl -s http://127.0.0.1:8124/v1/images/generations \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"model":"Qwen/Qwen-Image-2512","prompt":"A red panda on a mossy log at dawn","size":"1024x1024","num_inference_steps":20,"true_cfg_scale":2.5,"negative_prompt":"blurry, low quality, distorted, watermark","seed":1234}' \
  | python3 -c "import sys,json,base64; d=json.load(sys.stdin); open('out.png','wb').write(base64.b64decode(d['data'][0]['b64_json']))"

Always send a negative_prompt — vLLM-Omni disables classifier-free guidance without one.

Troubleshooting

• TypeError: HummingConfig.__init__() got an unexpected keyword argument ... — the plugin isn't installed in the serving venv (pip show vllm-omni-humming).
• Shape/dtype mismatch on img_mod.1 / txt_mod.1 at load — same cause.
• Corrupted images — humming-kernels too old; pip install --no-deps "humming-kernels>=0.1.11".

License

Apache-2.0. See LICENSE.