fix: derive checkpoint_format from resolved local path (LoRA base weights silently not loaded)

[Rouhi-Amirreza]

What

checkpoint_format is computed from the configured checkpoint.load_path string, but get_checkpoint_path() may resolve a DCP-style URI to a consolidated .pt file. The LoRA key-mapping branch is gated on checkpoint_format == "pt", so PEFT-wrapped models (use_lora=True) load zero parameters under the non-strict load and train from random init — with no error raised. This derives the format from the resolved file instead.

Fixes #160. Also matches the long-standing symptom reported in nvidia-cosmos/cosmos-predict2#176 (LoRA model produces identical/degenerate results vs. base; checkpoint-loading warnings).

Change

One conditional in cosmos_predict2/_src/predict2/utils/model_loader.py, immediately after the get_checkpoint_path(...) resolution:

    if str(local_s3_ckpt_fp).endswith(".pt"):
        checkpoint_format = "pt"

Verification

Cosmos-Predict2.5-2B LoRA post-training (rank 32, single H100), per the issue's reproduction:

	Before	After
Startup log	_IncompatibleKeys(missing_keys=[<all base params>]), no LoRA-mapping line	Mapped 689 LoRA keys from checkpoint to model
First-iter loss	~3.04 (random base)	~0.03 (loaded base)
Generation	noise	coherent video

No behavior change for runs whose load_path already ends in .pt (the condition only promotes dcp→pt when the resolved file is a .pt).