Add support for Prithvi in Online serving mode (vllm-project#21518)

Signed-off-by: Michele Gazzetti <[email protected]>
Co-authored-by: Cyrus Leung <[email protected]>
Signed-off-by: Jinzhen Lin <[email protected]>

Author: 2 people

tests/entrypoints/openai/test_skip_tokenizer.py

@@ -0,0 +1,93 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import base64
+import io
+
+import numpy as np
+import pytest
+import requests
+import torch
+
+from ...utils import RemoteOpenAIServer
+
+MODEL_NAME = "christian-pinto/Prithvi-EO-2.0-300M-TL-VLLM"
+DTYPE = "float16"
+
+
+@pytest.fixture(autouse=True)
+def v1(run_with_both_engines):
+    # Simple autouse wrapper to run both engines for each test
+    # This can be promoted up to conftest.py to run for every
+    # test in a package
+    pass
+
+
+@pytest.fixture(scope="module")
+def server():
+    args = [
+        "--task",
+        "embed",
+        # use half precision for speed and memory savings in CI environment
+        "--dtype",
+        DTYPE,
+        "--enforce-eager",
+        "--trust-remote-code",
+        "--skip-tokenizer-init",
+        "--max-num-seqs",
+        "32"
+    ]
+
+    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
+        yield remote_server
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_single_request(server: RemoteOpenAIServer, model_name: str):
+
+    pixel_values = torch.full((6, 512, 512), 1.0, dtype=torch.float16)
+    location_coords = torch.full((1, 2), 1.0, dtype=torch.float16)
+
+    buffer_tiff = io.BytesIO()
+    torch.save(pixel_values, buffer_tiff)
+    buffer_tiff.seek(0)
+    binary_data = buffer_tiff.read()
+    base64_tensor_embedding = base64.b64encode(binary_data).decode('utf-8')
+
+    buffer_coord = io.BytesIO()
+    torch.save(location_coords, buffer_coord)
+    buffer_coord.seek(0)
+    binary_data = buffer_coord.read()
+    base64_coord_embedding = base64.b64encode(binary_data).decode('utf-8')
+
+    prompt = {
+        "model":
+        model_name,
+        "additional_data": {
+            "prompt_token_ids": [1]
+        },
+        "encoding_format":
+        "base64",
+        "messages": [{
+            "role":
+            "user",
+            "content": [{
+                "type": "image_embeds",
+                "image_embeds": {
+                    "pixel_values": base64_tensor_embedding,
+                    "location_coords": base64_coord_embedding,
+                },
+            }],
+        }]
+    }
+
+    # test single pooling
+    response = requests.post(server.url_for("pooling"), json=prompt)
+    response.raise_for_status()
+
+    output = response.json()["data"][0]['data']
+
+    np_response = np.frombuffer(base64.b64decode(output), dtype=np.float32)
+
+    assert len(np_response) == 524288

vllm/engine/multiprocessing/client.py

@@ -97,11 +97,16 @@ def __init__(self, ipc_path: str, engine_config: VllmConfig,
         self.model_config = engine_config.model_config
         self.decoding_config = engine_config.decoding_config
 
-        # Create the tokenizer group.
-        self.tokenizer = init_tokenizer_from_configs(
-            model_config=self.model_config,
-            scheduler_config=engine_config.scheduler_config,
-            lora_config=engine_config.lora_config)
+        if self.vllm_config.model_config.skip_tokenizer_init:
+            self.tokenizer = None
+
+        else:
+            # Create the tokenizer group.
+            self.tokenizer = init_tokenizer_from_configs(
+                model_config=self.model_config,
+                scheduler_config=engine_config.scheduler_config,
+                lora_config=engine_config.lora_config)
+
         self.input_preprocessor = InputPreprocessor(self.model_config,
                                                     self.tokenizer)
 
@@ -375,7 +380,10 @@ async def get_input_preprocessor(self) -> InputPreprocessor:
         return self.input_preprocessor
 
     async def get_tokenizer(self, lora_request: Optional[LoRARequest] = None):
-        return await self.tokenizer.get_lora_tokenizer_async(lora_request)
+        if self.tokenizer is None:
+            return None
+        else:
+            return await self.tokenizer.get_lora_tokenizer_async(lora_request)
 
     async def get_vllm_config(self) -> VllmConfig:
         return self.vllm_config

vllm/entrypoints/openai/serving_engine.py

@@ -880,7 +880,10 @@ async def _preprocess_chat(
         _chat_template_kwargs.update(chat_template_kwargs or {})
 
         request_prompt: Union[str, list[int]]
-        if isinstance(tokenizer, MistralTokenizer):
+
+        if tokenizer is None:
+            request_prompt = "placeholder"
+        elif isinstance(tokenizer, MistralTokenizer):
             request_prompt = apply_mistral_chat_template(
                 tokenizer,
                 messages=messages,
@@ -910,7 +913,14 @@ async def _preprocess_chat(
             request = tool_parser(tokenizer).adjust_request(  # type: ignore
                 request=request)
 
-        if isinstance(request_prompt, str):
+        if tokenizer is None:
+            assert isinstance(request_prompt, str), (
+                "Prompt has to be a string", \
+                "when the tokenizer is not initialised"
+            )
+            prompt_inputs = TextTokensPrompt(prompt=request_prompt,
+                                             prompt_token_ids=[1])
+        elif isinstance(request_prompt, str):
             prompt_inputs = await self._tokenize_prompt_input_async(
                 request,
                 tokenizer,

vllm/entrypoints/openai/serving_pooling.py

@@ -96,7 +96,11 @@ async def create_pooling(
                 self.max_model_len, truncate_prompt_tokens)
             lora_request = self._maybe_get_adapters(request)
 
-            tokenizer = await self.engine_client.get_tokenizer(lora_request)
+            if self.model_config.skip_tokenizer_init:
+                tokenizer = None
+            else:
+                tokenizer = await self.engine_client.get_tokenizer(lora_request
+                                                                   )
 
             if isinstance(request, PoolingChatRequest):
                 (

vllm/model_executor/models/prithvi_geospatial_mae.py

@@ -103,7 +103,10 @@ def apply(
         mm_kwargs = {}
 
         for k, v in mm_data.items():
-            mm_kwargs[k] = v
+            if isinstance(v, dict) and k == "image":
+                mm_kwargs.update(v)
+            else:
+                mm_kwargs[k] = v
         mm_placeholders = {"image": [PlaceholderRange(offset=0, length=0)]}
 
         # This model receives in input a multi-dimensional tensor representing