feat: add tests for image types and remove alpha from png

integration-tests/models/__snapshots__/test_flash_gemma3/test_flash_gemma3_image_base64_rgb_jpg.json

@@ -0,0 +1,26 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null,
+      "message": {
+        "content": "Okay, let's analyze the image.\n\nThe image is a solid, bright white color. There is nothing else visible within it. \n\nIt's essentially a blank white canvas or a completely white square. \n\nIs there anything specific you'd like me to do with this image, such as describe it further or imagine what it might represent?",
+        "name": null,
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "usage": null
+    }
+  ],
+  "created": 1741965894,
+  "id": "",
+  "model": "google/gemma-3-4b-it",
+  "object": "chat.completion",
+  "system_fingerprint": "3.2.1-dev0-native",
+  "usage": {
+    "completion_tokens": 74,
+    "prompt_tokens": 277,
+    "total_tokens": 351
+  }
+}

integration-tests/models/__snapshots__/test_flash_gemma3/test_flash_gemma3_image_base64_rgb_png.json

@@ -0,0 +1,26 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null,
+      "message": {
+        "content": "Okay, let's analyze the image. \n\nThe image is entirely white, with a very subtle, faint outline of a stylized, cartoonish figure. It appears to be a simplified depiction of a person, likely a child, with a wide-eyed expression and a small, rounded body. \n\nIt's almost like a minimalist, iconic representation. \n\nDo you want me to try and describe it in more detail or perhaps speculate about the context of the image?",
+        "name": null,
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "usage": null
+    }
+  ],
+  "created": 1741965892,
+  "id": "",
+  "model": "google/gemma-3-4b-it",
+  "object": "chat.completion",
+  "system_fingerprint": "3.2.1-dev0-native",
+  "usage": {
+    "completion_tokens": 98,
+    "prompt_tokens": 277,
+    "total_tokens": 375
+  }
+}

integration-tests/models/__snapshots__/test_flash_gemma3/test_flash_gemma3_image_base64_rgba.json

@@ -0,0 +1,26 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null,
+      "message": {
+        "content": "Okay, let's analyze the image. \n\nThe transparent image reveals a stylized depiction of **a human head**. It's a minimalist, geometric representation, showing the basic shapes of the skull, eye sockets, and head outline. \n\nDo you want me to describe any specific element of the image in more detail?",
+        "name": null,
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "usage": null
+    }
+  ],
+  "created": 1741966313,
+  "id": "",
+  "model": "google/gemma-3-4b-it",
+  "object": "chat.completion",
+  "system_fingerprint": "3.2.1-dev0-native",
+  "usage": {
+    "completion_tokens": 67,
+    "prompt_tokens": 277,
+    "total_tokens": 344
+  }
+}

integration-tests/models/test_flash_gemma3.py

@@ -1,3 +1,7 @@
+import base64
+from io import BytesIO
+from PIL import Image
+
 import pytest
 
 
@@ -91,3 +95,76 @@ async def test_exceed_window(flash_gemma3, response_snapshot):
     )
     assert response.details.generated_tokens == 16
     assert response == response_snapshot
+
+
+# Helper function to convert a Pillow image to a base64 data URL
+def image_to_data_url(img: Image.Image, fmt: str) -> str:
+    buffer = BytesIO()
+    img.save(buffer, format=fmt)
+    img_data = buffer.getvalue()
+    b64_str = base64.b64encode(img_data).decode("utf-8")
+    mime_type = "image/png" if fmt.upper() == "PNG" else "image/jpeg"
+    return f"data:{mime_type};base64,{b64_str}"
+
+
+async def test_flash_gemma3_image_base64_rgba(flash_gemma3, response_snapshot):
+    # Create an empty 100x100 PNG image with alpha (transparent background)
+    img = Image.new("RGBA", (100, 100), (0, 0, 0, 0))
+    data_url = image_to_data_url(img, "PNG")
+    response = await flash_gemma3.chat(
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image_url", "image_url": {"url": data_url}},
+                    {
+                        "type": "text",
+                        "text": "What do you see in this transparent image?",
+                    },
+                ],
+            },
+        ],
+        max_tokens=100,
+    )
+    assert response == response_snapshot
+
+
+async def test_flash_gemma3_image_base64_rgb_png(flash_gemma3, response_snapshot):
+    # Create an empty 100x100 PNG image without alpha (white background)
+    img = Image.new("RGB", (100, 100), (255, 255, 255))
+    data_url = image_to_data_url(img, "PNG")
+    response = await flash_gemma3.chat(
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image_url", "image_url": {"url": data_url}},
+                    {"type": "text", "text": "What do you see in this plain image?"},
+                ],
+            },
+        ],
+        max_tokens=100,
+    )
+    assert response == response_snapshot
+
+
+async def test_flash_gemma3_image_base64_rgb_jpg(flash_gemma3, response_snapshot):
+    # Create an empty 100x100 JPEG image (white background)
+    img = Image.new("RGB", (100, 100), (255, 255, 255))
+    data_url = image_to_data_url(img, "JPEG")
+    response = await flash_gemma3.chat(
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image_url", "image_url": {"url": data_url}},
+                    {"type": "text", "text": "What do you see in this JPEG image?"},
+                ],
+            },
+        ],
+        max_tokens=100,
+    )
+    assert response == response_snapshot

server/text_generation_server/models/custom_modeling/gemma3/image_processing_gemma3.py

@@ -263,7 +263,7 @@ class Gemma3ImageProcessor(BaseImageProcessor):
         return_tensors: Optional[Union[str, TensorType]] = None,
         data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
         input_data_format: Optional[Union[str, ChannelDimension]] = None,
-        do_convert_rgb: bool = None,
+        do_convert_rgb: bool = True,
         do_pan_and_scan: bool = None,
         pan_and_scan_min_crop_size: int = None,
         pan_and_scan_max_num_crops: int = None,