brunopires01/aimedlab-pulse-hf — Thin Endpoint Repo

This repository contains ONLY the runtime glue (handler.py + vendored llava/ source + requirements.txt) needed to serve PULSE-7B on a Hugging Face Inference Endpoint.

The model weights (14 GB safetensors) live in the upstream public repo ubden/aimedlab-pulse-hf and are downloaded automatically at cold-start via HF_MODEL_ID=ubden/aimedlab-pulse-hf.

Why a thin repo?

The stock HF Inference container (pytorch-gpu) ships with transformers ≥ 4.40, which is incompatible with the legacy llava package's pinned transformers==4.37.2. Vendoring the ~9 source files of LLaVA v1.2.0 directly under llava/ and pointing sys.path at the repo root avoids any pip install at runtime — the endpoint boots cleanly.

Layout

handler.py            # HF EndpointHandler — multimodal (image + text) -> text
requirements.txt      # only what the HF container does not already provide
llava/                # vendored from github.com/haotian-liu/LLaVA @ v1.2.0
  __init__.py
  constants.py
  conversation.py
  mm_utils.py
  utils.py
  model/
    __init__.py
    builder.py
    llava_arch.py
    utils.py
    language_model/
      __init__.py
      llava_llama.py
    multimodal_encoder/
      __init__.py
      builder.py
      clip_encoder.py
    multimodal_projector/
      __init__.py
      builder.py

Endpoint configuration

Env var	Default	Purpose
HF_MODEL_ID	ubden/aimedlab-pulse-hf	Hub repo to pull weights from
HF_MODEL_DIR	/repository	Local mount of THIS repo
HF_MODEL_BASE	(none)	Optional base model for delta
CONV_MODE	(auto)	Force a LLaVA conversation mode

Request payload

{
  "inputs": {
    "image": "https://.../ecg.jpg  |  data:image/jpeg;base64,...  |  <plain base64>",
    "query": "Describe the ECG.",
    "temperature": 0.0,
    "top_p": 0.9,
    "max_new_tokens": 512
  }
}

Response: {"generated_text": "...", "model": "...", "conv_mode": "llava_v1"}

Credits

• Model: PULSE-ECG/PULSE-7B — AIMedLab/PULSE
• Upstream deployment repo: ubden/aimedlab-pulse-hf — Ubden® Team
• LLaVA: haotian-liu/LLaVA v1.2.0 (Apache-2.0)