Installation¶
thor-cosmos runs in two places: your x86 GPU host (for model prep, training, generation) and your Jetson AGX Thor (for edge inference). The package is identical on both — the recipes know which steps only make sense where.
Prerequisites¶
- Python 3.10+
just— the command runnergit— for cloning Cosmos upstream repos- AWS / Anthropic / OpenAI / Ollama credentials — one is enough (the agent auto-detects)
On Jetson Thor additionally:¶
- TensorRT-Edge-LLM built from source (
llm_build,visual_build,trt_edgellm_server) - GStreamer with
nvv4l2decoder/nvjpegencfor HW-accelerated RTP capture - NATS (optional, for perception event publishing)
On x86 GPU host additionally:¶
tensorrt-edgellm-*CLI tools (forquantize,export-llm,export-visual)hf(huggingface_hub) — for model downloadscosmos-cli/cosmos-rl— if you plan to post-train Reason2
Install just¶
Install thor-cosmos¶
Configure .env¶
Copy the example:
Key variables:
| Variable | Default | Purpose |
|---|---|---|
THOR_COSMOS_PROVIDER |
bedrock |
bedrock / openai / ollama |
THOR_COSMOS_MODEL_ID |
global.anthropic.claude-opus-4-6-v1 |
Agent model |
COSMOS_PREDICT_REPO |
../cosmos-predict2.5 |
Upstream repo path |
COSMOS_TRANSFER_REPO |
../cosmos-transfer2.5 |
Upstream repo path |
COSMOS_REASON_REPO |
../cosmos-reason2 |
Upstream repo path |
COSMOS_XENNA_REPO |
../cosmos-xenna |
Upstream repo path |
COSMOS_COOKBOOK_REPO |
../cosmos-cookbook |
Upstream repo path |
TRT_ROOT |
/opt/tensorrt-edge-llm |
TensorRT-LLM build root (Thor only) |
VLM_HOST / VLM_PORT |
127.0.0.1 / 8080 |
Where the TRT-EdgeLLM server runs |
RTP_BIND / RTP_PORT |
0.0.0.0 / 5600 |
Camera RTP ingress |
NATS_URL |
nats://127.0.0.1:4222 |
Perception event bus |
Verify¶
just env # shows effective variables
just --list # shows all 42+ recipes
just smoke # env + sysinfo + serve-status sanity check
If just smoke prints no errors, you're ready. → Quickstart