schema_version: "0.1"
name: "OpenRAL/rskill-qwen35_4b-any-general-nf4"
version: "0.1.0"
license: "apache-2.0"
role: "s2"
kind: "vlm"

# embodiment-agnostic: a perception rSkill (detector/vlm) needs only an RGB
# camera, not a specific robot. The rSkill<->robot gate exempts these kinds
# (loader._EMBODIMENT_AGNOSTIC_KINDS), so the tag list is empty —
# it matches any robot/scene.
embodiment_tags: ["any"]  # explicit embodiment-agnostic wildcard

sensors_required:
  - modality: "rgb"
    min_width: 336
    min_height: 336

actuators_required: []

runtime: "pytorch"
quantization:
  dtype: "int4"
  backend: "pytorch"
  extra:
    scheme: "nf4"
    quantizer: "bitsandbytes"
    compute_dtype: "bfloat16"
    min_params_to_quantize: 4000000
    loader: "transformers.AutoModelForImageTextToText"

min_vram_gb:
  fp32: 16.0
  bf16: 8.0
  int4: 2.5

# Deployable checkpoint: the pre-quantized NF4 weights (model.safetensors with
# an embedded bitsandbytes quantization_config). Loads directly as 4-bit
# (~3.3 GB) with no bf16 load spike — see tools/build_qwen_vlm_nf4_checkpoint.py.
weights_uri: "hf://OpenRAL/rskill-qwen35_4b-any-general-nf4"

chunk_size: 1
latency_budget:
  per_chunk_ms: 3000.0

# Provenance: the upstream Apache-2.0 model the NF4 checkpoint was quantized
# from, SHA-pinned for reproducibility (§8).
source_repo: "hf://Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"

description: >
  Qwen3.5-4B natively-multimodal video-language model packaged as an NF4
  bitsandbytes vlm rSkill. Accepts RGB image or video frames plus a
  natural-language query; returns a text answer. Embodiment-agnostic. No
  actuators. Apache-2.0.

actions:
  - "query"

objects:
  - "open-vocabulary object"
  - "text"
  - "scene region"
  - "spatial relation"

scenes:
  - "tabletop"
  - "kitchen"
  - "indoor"
  - "outdoor"
  - "warehouse"
  - "driving"