fix: chat_template loading and spatial_shapes unbound variable in Eagle VL

[cagataycali]

Summary

Two bug fixes for GR00T N1.5/N1.6 inference with transformers>=4.51:

Bug 1: chat_template is None → ValueError: Cannot render chat template

File: gr00t/model/gr00t_n1d6/processing_gr00t_n1d6.py

Root cause: AutoProcessor.from_pretrained() loads tokenizer_config.json which has a simpler template without <image> token support. The proper multimodal template lives in chat_template.json but is not loaded by the processor.

Fix: After AutoProcessor.from_pretrained(), check if chat_template is None. If so, load it from chat_template.json (or fall back to tokenizer_config.json) in the Eagle model directory.

Bug 2: spatial_shapes unbound → UnboundLocalError

File: gr00t/model/modules/nvidia/Eagle-Block2A-2B-v2/modeling_eagle3_vl.py

Root cause: In extract_feature(), when select_layer != -1 (GR00T uses -4), the else branch only grabs hidden_states[select_layer] but never captures spatial_shapes from vision_model_output. Line 366 then uses spatial_shapes which was never assigned.

Fix: Initialize spatial_shapes = None before the branch, and extract it from vision_model_output in both if and else branches.

Reproduction

from gr00t.model.gr00t_n1d6 import Gr00tN1D6Policy

policy = Gr00tN1D6Policy(
    model_path="nvidia/GR00T-N1.5-3B",
    embodiment_tag="new_embodiment",
)
# Bug 1: ValueError on policy.get_action() due to None chat_template
# Bug 2: UnboundLocalError on spatial_shapes in Eagle VL forward pass

Testing

Verified with 12 integration tests on an NVIDIA L40S GPU (Python 3.10, transformers==4.51.3, gr00t==0.1.0). All pass after these fixes.

Changes

File	Lines	Description
processing_gr00t_n1d6.py	+18/-1	Load chat_template.json fallback when AutoProcessor misses it
modeling_eagle3_vl.py	+6/-2	Initialize spatial_shapes = None; extract in both branches

Additional note

chat_template.json exists in the source tree at gr00t/model/modules/nvidia/Eagle-Block2A-2B-v2/chat_template.json but is not included in the pip-installed wheel (gr00t==0.1.0). Users installing via pip need to manually copy this file. Consider adding it to the wheel packaging.