fix: increase CLIP max_batch_size to 32

As we didn't configure any config.pbtxt file, max batch size was 4 by
default

models/clip/config.pbtxt

@@ -0,0 +1,26 @@
+backend: "onnxruntime"
+max_batch_size: 32
+input [
+  {
+      name: "pixel_values"
+      data_type: TYPE_FP32
+      dims: [ -1, -1, -1 ]
+  },
+  {
+      name: "attention_mask"
+      data_type: TYPE_INT64
+      dims: [ -1 ]
+  },
+  {
+      name: "input_ids"
+      data_type: TYPE_INT64
+      dims: [ -1 ]
+  }
+]
+output [
+  {
+      name: "image_embeds"
+      data_type: TYPE_FP32
+      dims: [ 512 ]
+  }
+]

robotoff/triton.py

@@ -13,6 +13,9 @@
 
 logger = get_logger(__name__)
 
+# Maximum batch size for CLIP model set in CLIP config.pbtxt
+CLIP_MAX_BATCH_SIZE = 32
+
 
 @cachetools.cached(cachetools.Cache(maxsize=1))
 def get_triton_inference_stub():
@@ -64,10 +67,8 @@ def generate_clip_embedding_request(images: list[Image.Image]):
 def generate_clip_embedding(images: list[Image.Image]) -> np.ndarray:
     embedding_batches = []
     stub = get_triton_inference_stub()
-    # max_batch_size is currently set to default value of 4 for CLIP
-    # TODO(raphael): Supply a custom model config file to increase this
-    # value
-    for image_batch in chunked(images, 4):
+
+    for image_batch in chunked(images, CLIP_MAX_BATCH_SIZE):
         request = generate_clip_embedding_request(image_batch)
         response = stub.ModelInfer(request)
         embedding_batch = np.frombuffer(