# Set $CHECKPOINT_DIR to the directory holding your model weights
# (Pi0 / Pi0.5 / GROOT format; the library auto-detects config).
export CHECKPOINT_DIR=/path/to/your/pi05_checkpoint
python examples/quickstart.py --checkpoint "$CHECKPOINT_DIR"Auto-detects GPU (SM110 → Thor backend, SM120 → Blackwell backend) and runs one inference pass with precision check.
Run inside Docker container (<your_container>):
# Precision check: cosine similarity vs PyTorch reference
python examples/thor/eval_precision.py \
--checkpoint_dir $CHECKPOINT_DIR
# LIBERO benchmark (full evaluation)
python examples/thor/eval_libero.py \
--checkpoint_dir $CHECKPOINT_DIR \
--task_suite libero_spatial| Metric | Value |
|---|---|
| E2E Latency | 44.3 ms |
| Cosine vs PyTorch | 0.9998 |
| LIBERO Spatial | 98.2% (491/500) |
# Precision check
python examples/blackwell/eval_precision.py \
--checkpoint /path/to/checkpoint
# LIBERO benchmark
python examples/blackwell/eval_libero.py \
--checkpoint /path/to/checkpoint \
--task_suite libero_spatial| Metric | Value |
|---|---|
| Latency (2-view) | 17.88 ms |
| Cosine vs PyTorch | 0.999638 |
| LIBERO Spatial | 100% (50/50) |
- Create
examples/<hardware>/directory - Copy an existing example and adjust engine/FMHA paths
- The library auto-detects SM version; you may also force backend:
model = flash_rt.load_model("pi05", ckpt, backend="thor")