thor-cosmos

The Idea

One command surface. Four Cosmos families. Full lifecycle.

Every NVIDIA Cosmos upstream repo already ships a justfile. thor-cosmos blends in: a single justfile with 42 recipes is the only command surface. A Strands agent calls the same recipes an operator would type — zero duplication, zero ambiguity.

Capabilities

Cosmos on the edge.

Flagship Pipeline

intbot_edge_vlm — real-time perception on Thor.

Deploy Cosmos-Reason2 end-to-end: download → quantize → export → ship → build → serve → infer.

# x86 GPU host
just prep-edge-model reason2-2b ./models/R2-fp8
scp -r ./models/R2-fp8/onnx cagatay@thor.local:~/R2-fp8-onnx

# Jetson AGX Thor
ssh cagatay@thor.local
cd ~/thor-cosmos
just build-engines ~/R2-fp8-onnx ~/R2-fp8-engines
just serve-start ~/R2-fp8-engines/llm ~/R2-fp8-engines/visual

# Real-time loop (RTP → VLM → NATS)
just perception-loop perception.vlm "describe the scene, count people"

Design philosophy

Thin Python. Thick justfile.

Every agent tool is a ~30-line wrapper that calls just_run(recipe, *args) and maps the output to a Strands ToolResult. All pipeline logic lives in the justfile. New capability? Add a recipe. Wrap it. Done.

@tool
def cosmos_quantize(model_dir, output_dir, dtype="fp16", quantization="fp8"):
    """Quantize a Cosmos VLM/LLM via `just quantize`."""
    proc = just_run("quantize", model_dir, output_dir, dtype, quantization,
                    timeout_s=60*60*3)
    return proc_result(proc,
        success_text=f"✅ quantized {model_dir} → {output_dir}",
        fail_text=f"quantization failed: {proc.get('stderr','')[:200]}")