fix: Handle Gemma 3 required pixel_values parameter in continuous batching

[lubauss]

Summary

• Fixes Gemma 3 continuous batching mode which failed with Model.__call__() missing 1 required positional argument: 'pixel_values'
• Gemma 3's model requires pixel_values as a positional argument, unlike Qwen2-VL which makes it optional
• The MLLMModelWrapper now injects pixel_values=None for text-only requests

Technical Details

The issue was in how different MLLM models handle the pixel_values parameter:

Qwen2-VL:

def __call__(self, input_ids, pixel_values: Optional = None, ...)

Gemma 3:

def __call__(self, input_ids, pixel_values, ...)  # Required!

When BatchGenerator calls the model with model(input_ids, cache=cache), Gemma 3 fails because pixel_values is missing.

Test plan

• Server starts successfully with Gemma 3 27B 4-bit
• Text generation works
• Streaming works
• Prefix caching works (2x speedup observed)
• Concurrent requests work

🤖 Generated with Claude Code