OpenAI API: Changes to enable multi-modal for 3.2 11B

torchchat/usages/openai_api.py

@@ -19,7 +19,8 @@
 
 from PIL import Image
 
-from torchtune.data import Message, padded_collate
+from torchtune.data import Message, padded_collate_tiled_images_and_mask
+
 from torchtune.models.llama3_2_vision._model_builders import llama3_2_vision_transform
 
 from torchchat.cli.download import is_model_downloaded, load_model_configs
@@ -288,13 +289,50 @@ def __init__(self, *args, **kwargs):
                 else self.model.text_transformer_args.max_seq_length
             )
         except:
-            # can not find max_seq_length in model config, use default value
-            self.max_seq_length = 128
+            self.max_seq_length = 2048
+            print(f"can not find max_seq_length in model config, use default value: {self.max_seq_length}")
         # The System fingerprint is a unique identifier for the model and its configuration.
         self.system_fingerprint = (
             f"{self.builder_args.device}_{self.builder_args.precision}"
         )
 
+    def _openai_messages_to_torchtune_messages( 
+        self, messages: List[_AbstractMessage]
+    ) -> List[Message]:
+        """Convert a list of OpenAI API messages to a list of TorchTune messages.
+
+        Args:
+            messages: A list of OpenAI API messages.
+
+        Returns:
+            A list of Torchtune Messages.
+        """
+        torchtune_messages = []
+        for message in messages:
+            torchtune_contents = []
+            if isinstance(message["content"], list):
+                for content_dict in message["content"]:
+                    converted_content = []
+                    if content_dict["type"] == "text":
+                        converted_content.append(
+                            {"type": "text", "content": content_dict["text"]}
+                        )
+                    elif content_dict["type"] == "image_url":
+                        base64_decoded = base64.b64decode(
+                                    content_dict["image_url"].split(";base64,")[1]
+                                )
+                        image = Image.open(BytesIO(base64_decoded))
+                        converted_content.append(
+                            {
+                                "type": "image",
+                                "content": image,
+                            }
+                        )
+                    torchtune_messages.append( 
+                        Message(role=message["role"], content=converted_content, eot=False)
+                    )
+        return torchtune_messages
+
     def _openai_messages_to_torchtune(
         self, messages: List[_AbstractMessage]
     ) -> List[Message]:
@@ -376,15 +414,32 @@ def chunked_completion(self, completion_request: CompletionRequest):
             transform = llama3_2_vision_transform(
                 str(self.tokenizer_args.tokenizer_path)
             )
-            torchtune_messages = self._openai_messages_to_torchtune(
+            torchtune_messages = self._openai_messages_to_torchtune_messages(
                 completion_request.messages
             )
             data = transform(
                 {"images": images, "messages": torchtune_messages}, inference=True
             )
-            batch = padded_collate([data], self.builder_args.device)
-            batch.pop("mask")
-            encoded = batch["tokens"]
+            seq_len = len(data["tokens"])
+            total_response_length = seq_len + completion_request.max_tokens
+            causal_mask = torch.tril(
+                torch.ones(
+                    size=(total_response_length, total_response_length),
+                    dtype=torch.bool,
+                )
+            )
+            input_pos = torch.arange(total_response_length)
+
+            with torch.no_grad():
+                with torch.device(self.builder_args.device):
+                    batch = padded_collate_tiled_images_and_mask([data], pad_direction="left", pad_max_images=1)
+                    batch["encoder_input"]["images"] = batch["encoder_input"]["images"].to(self.builder_args.precision)
+                    batch["causal_mask"] = causal_mask
+                    batch["input_pos"] = input_pos[None, :seq_len]
+                    batch["encoder_mask"] = batch["encoder_mask"][:, :seq_len]
+
+                    #batch = padded_collate([data], self.builder_args.device)
+            encoded = batch["tokens"].view(-1)
         else:
             tokens = self.chat_formatter.encode_dialog_prompt(
                 dialog=[