Support for NemotronH Nano VLM with an optimized vision model (vLLM native)

[danielafrimi]

This PR builds on #23644

In addition to supporting the new VL model, it introduces the vision encoder implementation (C-RADIO
) using vLLM’s native layer.

To reduce code duplication, the implementation leverages InternVisionModel blocks.

[gemini-code-assist]

Code Review

This pull request adds support for the NemotronH Nano VLM model, including a native vLLM implementation of the C-RADIO vision encoder. The changes are well-structured, introducing new model files for nano_vlm and radio, along with corresponding tests. My review has identified a critical bug in the new test file that would prevent it from running correctly, a potential TypeError in the model initialization due to unsafe dictionary access, and the use of print statements for debugging which should be replaced with proper logging. Addressing these points will improve the robustness and maintainability of the new model support.

[gemini-code-assist]

The list comprehension for pixel_values seems to have a bug. The img_processor is called with a single PIL image, which returns a 3D tensor for pixel_values. However, a 4D slice [:, :, :, :640] is then applied to this 3D tensor, which will cause a runtime error.

To fix this, you can process each image as a list containing a single image to ensure the pixel_values tensor is 4D. Also, using a more descriptive loop variable name would improve readability.

Suggested change

	pixel_values = [
	img_processor(
	images,
	return_tensors='pt').pixel_values.to(torch_dtype)[:, :, :, :640]
	for images in images
	]
	pixel_values = [
	img_processor(
	[img],
	return_tensors='pt').pixel_values.to(torch_dtype)[:, :, :, :640]
	for img in images
	]

[gemini-code-assist]

The call to vit_dims.get(model_name) can return None if model_name is not found in the vit_dims dictionary. This would cause a TypeError when attempting to unpack the result into hidden_size, num_layers, num_heads, intermediate_size. It's safer to check if the model name exists in the dictionary before attempting to get the value.

Suggested change

	model_name = hf_config.args.get("model")
	hidden_size, num_layers, num_heads, intermediate_size = vit_dims.get(
	model_name)
	model_name = hf_config.args.get("model")
	if model_name not in vit_dims:
	raise ValueError(
	f"Unsupported ViT model type for Radio: {model_name}. "
	f"Supported types are: {list(vit_dims.keys())}")
	hidden_size, num_layers, num_heads, intermediate_size = vit_dims[
	model_name]

[gemini-code-assist]

There are print statements in the ClsToken initializer. In a library like vLLM, using print for logging can clutter the output and is generally discouraged. Please consider using the logger from vllm.logger for debugging information, or removing these statements if they are not necessary.