Loading Cosmos3-Nano from a local snapshot path makes Reasoner text-only

[haifujing]

Bug Description

Loading the following Cosmos3-Nano snapshot through an absolute local path causes Reasoner image/video input to be treated as text-only:

nvidia/Cosmos3-Nano revision 411f42a

The snapshot contains vision_encoder/model.safetensors and multimodal processor files. However:

1. Its config omits include_visual, which defaults to False.
2. Its tokenizer configuration creates LLMTokenizerProcessor instead of Qwen3VLProcessor.

Using --checkpoint-path Cosmos3-Nano works because the registered YAML explicitly enables the visual tower and uses the multimodal processor.

Reproduction

Input JSON:

{
  "name": "video_reasoner",
  "model_mode": "reasoner",
  "prompt": "Describe what is happening in this video in one sentence.",
  "vision_path": "/absolute/path/to/video.mp4",
  "video_fps": 1.0,
  "max_new_tokens": 64,
  "do_sample": false
}

Command:

LD_LIBRARY_PATH= CUDA_VISIBLE_DEVICES=0 \
python -m cosmos_framework.scripts.inference \
  --parallelism-preset=latency \
  -i /tmp/reasoner-video.json \
  -o /tmp/cosmos3-output \
  --checkpoint-path /path/to/snapshots/411f42a8fdfb8c5b2583cb8786e0938f49796eaa \
  --seed=0 \
  --no-guardrails

Expected Behavior

The bundled vision tower and Qwen3VLProcessor should be loaded, allowing Reasoner image/video input.

Actual Behavior

The visual tower is initially disabled:

checkpoint's reasoner LM has no visual tower
(model config include_visual is false)

After manually enabling it, the text-only processor rejects video input:

LLMTokenizerProcessor does not implement apply_chat_template

Suggested Fix

When include_visual is omitted, infer it as True only if the checkpoint index references an existing vision_encoder shard. An explicit False should remain
authoritative.

When the checkpoint bundles the required processor and tokenizer files, load a local Qwen3VLProcessor instead of the legacy text-only processor.

System Information

• Python 3.13.14
• PyTorch 2.10.0+cu130
• CUDA 13.0
• NVIDIA RTX 5880 Ada Generation