application/octet-stream support in content type deserialization

no reason not to accept it

Signed-off-by: Raphael Glon <[email protected]>

Author: oOraph

src/huggingface_inference_toolkit/serialization/base.py

@@ -1,3 +1,4 @@
+from huggingface_inference_toolkit.const import HF_TASK
 from huggingface_inference_toolkit.serialization.audio_utils import Audioer
 from huggingface_inference_toolkit.serialization.image_utils import Imager
 from huggingface_inference_toolkit.serialization.json_utils import Jsoner
@@ -38,7 +39,13 @@
 
 class ContentType:
     @staticmethod
-    def get_deserializer(content_type: str):
+    def get_deserializer(content_type: str, task: str):
+        if content_type.lower().startswith("application/octet-stream"):
+            if "audio" in task or "speech" in task:
+                return Audioer
+            elif "image" in task:
+                return Imager
+
         # Extract media type from content type
         extracted = content_type.split(";")[0]
         if extracted in content_type_mapping:

src/huggingface_inference_toolkit/webservice_starlette.py

@@ -89,10 +89,11 @@ async def metrics(request):
 async def predict(request):
     global INFERENCE_HANDLERS
     try:
+        task = request.path_params.get("task", HF_TASK)
         # extracts content from request
         content_type = request.headers.get("content-Type", os.environ.get("DEFAULT_CONTENT_TYPE")).lower()
         # try to deserialize payload
-        deserialized_body = ContentType.get_deserializer(content_type).deserialize(
+        deserialized_body = ContentType.get_deserializer(content_type, task).deserialize(
             await request.body()
         )
         # checks if input schema is correct
@@ -108,7 +109,7 @@ async def predict(request):
             )
 
         # We lazily load pipelines for alt tasks
-        task = request.path_params.get("task", HF_TASK)
+
         if task == "feature-extraction" and HF_TASK in [
             "sentence-similarity",
             "sentence-embeddings",