Follow the shared setup. Use a complete, revision-pinned Cosmos3-Nano-Transfer-Auto snapshot or your compatible adapted Transfer-Auto export. Keep assets outside the code checkout.
Set existing inputs and new output directories:
: "${TRANSFER_AUTO_HF:?set the complete local Transfer-Auto snapshot}"
: "${AUTO_BUNDLE:?set the validated local training bundle}"
: "${VIDEO_VAE:?set the matching Wan2.2_VAE.pth file}"
: "${SIM_AUTO_STAGE0_CHECKPOINT:?set a new DCP conversion directory}"
: "${IMAGINAIRE_OUTPUT_ROOT:?set a new shared training-output root}"
: "${SIM_AUTO_RUN_GROUP:?set a unique group for this training chain}"
export COSMOS_SIM_AUTO_LOCAL_BUNDLE="$AUTO_BUNDLE"
export RGB_VAE="$VIDEO_VAE"
export LIDAR_VAE="$TRANSFER_AUTO_HF/lidar_vae/diffusion_pytorch_model.safetensors"
export QWEN_TOKENIZER="$AUTO_BUNDLE/resources/tokenizer"
export SIM_AUTO_STAGE0_CHECKPOINT IMAGINAIRE_OUTPUT_ROOT SIM_AUTO_RUN_GROUP
export COSMOS_TRAINING=1 COSMOS_INTERNAL=0
python -m cosmos_framework.scripts.prepare_sim_auto_checkpoint \
--checkpoint-path "$TRANSFER_AUTO_HF" \
--output-path "$SIM_AUTO_STAGE0_CHECKPOINT"The converter checks the joint architecture, trained LiDAR heads and sensor
embeddings. It writes a weights-only iteration root containing model/.
It creates no optimizer, scheduler, trainer or EMA history. Do not convert the
LiDAR tokenizer into model DCP: its weights load separately.
A fresh TF launch skips only the absent net_ema. history and starts a new EMA
with rate 0.1 and shift 0. Clear that exception for full same-stage resume.
All nodes must see the same files. Check complete HF shards, tokenizer resources and DCP metadata before launch; strict native loading remains the compatibility check. File existence alone is insufficient.
Do not enable COSMOS_SMOKE: it can change architecture. Use explicit reductions
from the smoke guide.