Add support for siglip models

docs/source/exporters/onnx/overview.mdx

@@ -86,6 +86,7 @@ Supported architectures from [🤗 Transformers](https://huggingface.co/docs/tra
 - SEW
 - SEW-D
 - Speech2Text
+- SigLIP
 - SpeechT5
 - Splinter
 - SqueezeBert

optimum/exporters/onnx/model_configs.py

@@ -941,6 +941,31 @@ def generate_dummy_inputs(self, framework: str = "pt", **kwargs):
         return dummy_inputs
 
 
+class SiglipNormalizedConfig(CLIPNormalizedConfig):
+    pass
+
+
+class SiglipOnnxConfig(CLIPOnnxConfig):
+    NORMALIZED_CONFIG_CLASS = SiglipNormalizedConfig
+    DEFAULT_ONNX_OPSET = 13
+
+    @property
+    def inputs(self) -> Dict[str, Dict[int, str]]:
+        return {
+            "input_ids": {0: "text_batch_size", 1: "sequence_length"},
+            "pixel_values": {0: "image_batch_size", 1: "num_channels", 2: "height", 3: "width"},
+            # NOTE: No attention_mask
+        }
+
+
+class SiglipTextWithProjectionOnnxConfig(CLIPTextWithProjectionOnnxConfig):
+    pass
+
+
+class SiglipTextOnnxConfig(CLIPTextOnnxConfig):
+    pass
+
+
 class UNetOnnxConfig(VisionOnnxConfig):
     ATOL_FOR_VALIDATION = 1e-3
     # The ONNX export of a CLIPText architecture, an other Stable Diffusion component, needs the Trilu

optimum/exporters/tasks.py

@@ -900,6 +900,19 @@ class TasksManager:
             "audio-classification",
             onnx="SEWDOnnxConfig",
         ),
+        "siglip": supported_tasks_mapping(
+            "feature-extraction",
+            "zero-shot-image-classification",
+            onnx="SiglipOnnxConfig",
+        ),
+        "siglip-text-model": supported_tasks_mapping(
+            "feature-extraction",
+            onnx="SiglipTextOnnxConfig",
+        ),
+        "siglip-text-with-projection": supported_tasks_mapping(
+            "feature-extraction",
+            onnx="SiglipTextWithProjectionOnnxConfig",
+        ),
         "speech-to-text": supported_tasks_mapping(
             "feature-extraction",
             "feature-extraction-with-past",

optimum/utils/normalized_config.py

@@ -203,6 +203,7 @@ class NormalizedConfigManager:
         'perceiver',
         'roformer',
         'segformer',
+        'siglip',
         'squeezebert',
         'table-transformer',
     """

tests/exporters/exporters_utils.py

@@ -137,6 +137,7 @@
     "roformer": "hf-internal-testing/tiny-random-RoFormerModel",
     "sam": "fxmarty/sam-vit-tiny-random",
     "segformer": "hf-internal-testing/tiny-random-SegformerModel",
+    "siglip": "HuggingFaceM4/tiny-random-siglip",
     "splinter": "hf-internal-testing/tiny-random-SplinterModel",
     "squeezebert": "hf-internal-testing/tiny-random-SqueezeBertModel",
     "swin": "hf-internal-testing/tiny-random-SwinModel",