fix(multimodal): handle tensor images in exact-token preprocessing (#30368)

Signed-off-by: Zhuangcheng(Jesse) Gu <zcgu@connect.hku.hk>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Mick <mickjagger19@icloud.com>

python/sglang/srt/multimodal/processors/base_processor.py

@@ -1732,7 +1732,14 @@ class BaseMultimodalProcessor(ABC):
 
         """
         assert images is not None
-        image_sizes = [(image.height, image.width) for image in images]
+        image_sizes = [
+            (
+                tuple(image.shape[-2:])
+                if isinstance(image, torch.Tensor)
+                else (image.height, image.width)
+            )
+            for image in images
+        ]
         num_image_tokens = self._processor._get_num_multimodal_tokens(
             image_sizes=image_sizes
         ).num_image_tokens

test/registered/vlm/test_token_id_retokenize_e2e.py

@@ -9,6 +9,9 @@ With SGLANG_MM_AVOID_RETOKENIZE ON (default), the server keeps the user's
 original tokens verbatim and only expands the image placeholder, so prompt_tokens
 stays faithful to what the client sent.
 
+The test uses JPEG so CUDA decoding returns a CHW tensor, covering the same
+exact-token path as PIL-backed images.
+
 For each model we launch a real server twice with the same predefined,
 non-canonical prompt ("Describe" split into "D"+"escribe") plus one image:
 
@@ -42,8 +45,8 @@ register_cpu_ci(est_time=123, suite="stage-b-test-cpu-intel")
 def _data_uri():
     img = Image.new("RGB", (64, 64), (128, 128, 128))
     buf = io.BytesIO()
-    img.save(buf, format="PNG")
-    return "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode()
+    img.save(buf, format="JPEG")
+    return "data:image/jpeg;base64," + base64.b64encode(buf.getvalue()).decode()
 
 
 def _build_drift_prompt(model, image_token):