[vlm] Remove vision language config. (vllm-project#6089)

Signed-off-by: Xiaowei Jiang <[email protected]>
Co-authored-by: Roger Wang <[email protected]>

docs/source/dev/multimodal/multimodal_index.rst

@@ -10,8 +10,13 @@ vLLM provides experimental support for multi-modal models through the :mod:`vllm
 :class:`vllm.inputs.PromptStrictInputs` accepts an additional attribute ``multi_modal_data``
 which allows you to pass in multi-modal input alongside text and token prompts.
 
+.. note::
+   ``multi_modal_data`` can accept keys and values beyond the builtin ones, as long as a customized plugin is registered through 
+    :class:`vllm.multimodal.MULTIMODAL_REGISTRY`.
+
 By default, vLLM models do not support multi-modal inputs. To enable multi-modal support for a model, please follow :ref:`the guide for adding a new multimodal model. <adding_a_new_multimodal_model>`.
 
+
 # TODO: Add more instructions on how to do that once embeddings is in.
 
 Guides

docs/source/models/vlm.rst

@@ -8,18 +8,6 @@ vLLM provides experimental support for Vision Language Models (VLMs). This docum
 .. important::
     We are actively iterating on VLM support. Expect breaking changes to VLM usage and development in upcoming releases without prior deprecation.
 
-Engine Arguments
-----------------
-
-The following :ref:`engine arguments <engine_args>` are specific to VLMs:
-
-.. argparse::
-    :module: vllm.engine.arg_utils
-    :func: _vlm_engine_args_parser
-    :prog: -m vllm.entrypoints.openai.api_server
-    :nodefaultconst:
-
-.. important::
     Currently, the support for vision language models on vLLM has the following limitations:
 
     * Only single image input is supported per text prompt.
@@ -33,40 +21,33 @@ To initialize a VLM, the aforementioned arguments must be passed to the ``LLM``
 
 .. code-block:: python
 
-    llm = LLM(
-        model="llava-hf/llava-1.5-7b-hf",
-        image_token_id=32000,
-        image_input_shape="1,3,336,336",
-        image_feature_size=576,
-    )
+    llm = LLM(model="llava-hf/llava-1.5-7b-hf")
 
 .. important::
-    Currently, you have to specify ``image_feature_size`` to support memory profiling.
-    To avoid OOM during runtime, you should set this to the maximum value supported by the model.
-    The calculation of feature size is specific to the model. For more details, please refer to
-    the function :code:`get_<model_name>_image_feature_size` inside the corresponding model file.
+    We have removed all vision language related CLI args in the ``0.5.1`` release. **This is a breaking change**, so please update your code to follow
+    the above snippet. Specifically, ``image_feature_size`` is no longer required to be specified, and internally we will construct data structures for
+    every model to perform profiling with.
 
-    We will remove most of the vision-specific arguments in a future release as they can be inferred from the HuggingFace configuration.
+    This work is still ongoing. In the meantime, we internally hardcode ``image_feature_size = 3000`` through 
+    :meth:`MULTIMODAL_REGISTRY.get_num_input_tokens <vllm.multimodal.MultiModalRegistry.get_num_input_tokens>` 
+    for every model to be conservative in terms of GPU memory consumption. This hardcoded value will be replaced 
+    with a more accurate profiling strategy in the future.
 
 
 To pass an image to the model, note the following in :class:`vllm.inputs.PromptStrictInputs`:
 
 * ``prompt``: The prompt should follow the format that is documented on HuggingFace.
 * ``multi_modal_data``: This is a dictionary that follows the schema defined in :class:`vllm.multimodal.MultiModalDataDict`. 
 
-.. note::
-
-   ``multi_modal_data`` can accept keys and values beyond the builtin ones, as long as a customized plugin is registered through
-    :class:`vllm.multimodal.MULTIMODAL_REGISTRY`.
-
 .. code-block:: python
 
     # Refer to the HuggingFace repo for the correct format to use
     prompt = "USER: <image>\nWhat is the content of this image?\nASSISTANT:"
 
     # Load the image using PIL.Image
-    image = ...
-
+    image = PIL.Image.open(...)
+    
+    # Single prompt inference
     outputs = llm.generate({
         "prompt": prompt,
         "multi_modal_data": {"image": image},
@@ -75,6 +56,26 @@ To pass an image to the model, note the following in :class:`vllm.inputs.PromptS
     for o in outputs:
         generated_text = o.outputs[0].text
         print(generated_text)
+    
+    # Batch inference
+    image_1 = PIL.Image.open(...)
+    image_2 = PIL.Image.open(...)
+    outputs = llm.generate(
+        [
+            {
+                "prompt": "USER: <image>\nWhat is the content of this image?\nASSISTANT:",
+                "multi_modal_data": {"image": image_1},
+            },
+            {
+                "prompt": "USER: <image>\nWhat's the color of this image?\nASSISTANT:",
+                "multi_modal_data": {"image": image_2},
+            }
+        ]
+    )
+
+    for o in outputs:
+        generated_text = o.outputs[0].text
+        print(generated_text)
 
 A code example can be found in `examples/llava_example.py <https://github.com/vllm-project/vllm/blob/main/examples/llava_example.py>`_.
 
@@ -99,18 +100,17 @@ Below is an example on how to launch the same ``llava-hf/llava-1.5-7b-hf`` with
 
     python -m vllm.entrypoints.openai.api_server \
         --model llava-hf/llava-1.5-7b-hf \
-        --image-token-id 32000 \
-        --image-input-shape 1,3,336,336 \
-        --image-feature-size 576 \
         --chat-template template_llava.jinja
 
 .. important::
-    Currently, you have to specify ``image_feature_size`` to support memory profiling.
-    To avoid OOM during runtime, you should set this to the maximum value supported by the model.
-    The calculation of feature size is specific to the model. For more details, please refer to
-    the function :code:`get_<model_name>_image_feature_size` inside the corresponding model file.
-
-    We will remove most of the vision-specific arguments in a future release as they can be inferred from the HuggingFace configuration.
+    We have removed all vision language related CLI args in the ``0.5.1`` release. **This is a breaking change**, so please update your code to follow
+    the above snippet. Specifically, ``image_feature_size`` is no longer required to be specified, and internally we will construct data structures for
+    every model to perform profiling with.
+
+    This work is still ongoing. In the meantime, we internally hardcode ``image_feature_size = 3000`` through 
+    :meth:`MULTIMODAL_REGISTRY.get_num_input_tokens <vllm.multimodal.MultiModalRegistry.get_num_input_tokens>` 
+    for every model to be conservative in terms of GPU memory consumption. This hardcoded value will be replaced 
+    with a more accurate profiling strategy in the future.
 
 To consume the server, you can use the OpenAI client like in the example below:
 

examples/llava_example.py

@@ -10,12 +10,7 @@
 
 
 def run_llava():
-    llm = LLM(
-        model="llava-hf/llava-1.5-7b-hf",
-        image_token_id=32000,
-        image_input_shape="1,3,336,336",
-        image_feature_size=576,
-    )
+    llm = LLM(model="llava-hf/llava-1.5-7b-hf")
 
     prompt = "USER: <image>\nWhat is the content of this image?\nASSISTANT:"
 

examples/llava_next_example.py

@@ -7,13 +7,7 @@
 
 
 def run_llava_next():
-    llm = LLM(
-        model="llava-hf/llava-v1.6-mistral-7b-hf",
-        image_token_id=32000,
-        image_input_shape="1,3,336,336",
-        # Use the maximum possible value for memory profiling
-        image_feature_size=2928,
-    )
+    llm = LLM(model="llava-hf/llava-v1.6-mistral-7b-hf", max_model_len=4096)
 
     prompt = "[INST] <image>\nWhat is shown in this image? [/INST]"
     url = "https://h2o-release.s3.amazonaws.com/h2ogpt/bigben.jpg"

examples/openai_vision_api_client.py

@@ -3,9 +3,6 @@
 Launch the vLLM server with the following command:
 python -m vllm.entrypoints.openai.api_server \
     --model llava-hf/llava-1.5-7b-hf \
-    --image-token-id 32000 \
-    --image-input-shape 1,3,336,336 \
-    --image-feature-size 576 \
     --chat-template template_llava.jinja
 """
 import base64

examples/phi3v_example.py

@@ -14,15 +14,13 @@ def run_phi3v():
 
     # Note: The default setting of max_num_seqs (256) and
     # max_model_len (128k) for this model may cause OOM.
+    # You may lower either to run this example on lower-end GPUs.
+
     # In this example, we override max_num_seqs to 5 while
     # keeping the original context length of 128k.
     llm = LLM(
         model=model_path,
         trust_remote_code=True,
-        image_token_id=32044,
-        image_input_shape="1,3,1008,1344",
-        # Use the maximum possible value for memory profiling
-        image_feature_size=2653,
         max_num_seqs=5,
     )
 

tests/distributed/test_multimodal_broadcast.py

@@ -20,9 +20,9 @@
 model = os.environ["TEST_DIST_MODEL"]
 
 if model.startswith("llava-hf/llava"):
-    from ..models.test_llava import model_and_vl_config, run_test
+    from ..models.test_llava import models, run_test
 elif model.startswith("microsoft/Phi-3-vision"):
-    from ..models.test_phi3v import model_and_vl_config, run_test
+    from ..models.test_phi3v import models, run_test
 else:
     raise NotImplementedError(f"Unsupported model: {model}")
 
@@ -44,7 +44,7 @@ def test_models(hf_runner, vllm_runner, image_assets,
         hf_runner,
         vllm_runner,
         image_assets,
-        model_and_config=model_and_vl_config[0],
+        model=models[0],
         size_factors=[1.0],
         dtype=dtype,
         max_tokens=max_tokens,

tests/entrypoints/openai/test_vision.py

@@ -39,12 +39,6 @@ def server(ray_ctx):
         "--max-model-len",
         "4096",
         "--enforce-eager",
-        "--image-token-id",
-        "32000",
-        "--image-input-shape",
-        "1,3,336,336",
-        "--image-feature-size",
-        "576",
         "--chat-template",
         str(LLAVA_CHAT_TEMPLATE),
     ])

tests/models/test_llava.py

@@ -3,7 +3,6 @@
 import pytest
 from transformers import AutoTokenizer
 
-from vllm.config import VisionLanguageConfig
 from vllm.multimodal.utils import rescale_image_size
 from vllm.sequence import SampleLogprobs
 
@@ -21,49 +20,27 @@
     "USER: <image>\nWhat's in this image?\nASSISTANT:",
 })
 
+IMAGE_TOKEN_ID = 32000
 
-def iter_llava_configs(model_name: str):
-    image_hw_to_feature_size = {
-        (336, 336): 576,
-    }
-
-    for (h, w), f in image_hw_to_feature_size.items():
-        input_shape = (1, 3, h, w)
-        yield (model_name,
-               VisionLanguageConfig(image_feature_size=f,
-                                    image_token_id=32000,
-                                    image_input_shape=input_shape))
-
-
-model_and_vl_config = [
-    *iter_llava_configs("llava-hf/llava-1.5-7b-hf"),
-]
+models = ["llava-hf/llava-1.5-7b-hf"]
 
 
 def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
                                          Optional[SampleLogprobs]],
-                      vlm_config: VisionLanguageConfig, model_id: str):
-    """Sanitize vllm output to be comparable with hf output.
-    The function reduces `input_ids` from 1, 32000, 32000, ..., 32000,
-    x1, x2, x3 ... to 1, 32000, x1, x2, x3 ...
-    It also reduces `output_str` from "<image><image>bla" to "bla".
-    """
+                      model: str):
+    """Sanitize vllm output to be comparable with hf output."""
     output_ids, output_str, out_logprobs = vllm_output
-    image_token_id = vlm_config.image_token_id
 
-    tokenizer = AutoTokenizer.from_pretrained(model_id)
-    image_token_str = tokenizer.decode(image_token_id)
+    tokenizer = AutoTokenizer.from_pretrained(model)
     eos_token_id = tokenizer.eos_token_id
 
     hf_output_ids = [
         token_id for idx, token_id in enumerate(output_ids)
-        if token_id != image_token_id or output_ids[idx - 1] != image_token_id
+        if token_id != IMAGE_TOKEN_ID or output_ids[idx - 1] != IMAGE_TOKEN_ID
     ]
 
-    hf_output_str = output_str \
-        .replace(image_token_str * vlm_config.image_feature_size, "")
-    assert hf_output_str[0] == " "
-    hf_output_str = hf_output_str[1:]
+    assert output_str[0] == " "
+    hf_output_str = output_str[1:]
     if hf_output_ids[-1] == eos_token_id:
         hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
 
@@ -74,7 +51,7 @@ def run_test(
     hf_runner: Type[HfRunner],
     vllm_runner: Type[VllmRunner],
     image_assets: _ImageAssets,
-    model_and_config: Tuple[str, VisionLanguageConfig],
+    model: str,
     *,
     size_factors: List[float],
     dtype: str,
@@ -92,7 +69,6 @@ def run_test(
     Note, the text input is also adjusted to abide by vllm contract.
     The text output is sanitized to be able to compare with hf.
     """
-    model_id, vlm_config = model_and_config
     images = [asset.pil_image for asset in image_assets]
 
     inputs_per_image = [(
@@ -106,12 +82,11 @@ def run_test(
     # will hurt multiprocessing backend with fork method (the default method).
 
     # max_model_len should be greater than image_feature_size
-    with vllm_runner(model_id,
+    with vllm_runner(model,
                      dtype=dtype,
                      tensor_parallel_size=tensor_parallel_size,
                      distributed_executor_backend=distributed_executor_backend,
-                     enforce_eager=True,
-                     **vlm_config.as_cli_args_dict()) as vllm_model:
+                     enforce_eager=True) as vllm_model:
         vllm_outputs_per_image = [
             vllm_model.generate_greedy_logprobs(prompts,
                                                 max_tokens,
@@ -120,7 +95,7 @@ def run_test(
             for prompts, images in inputs_per_image
         ]
 
-    with hf_runner(model_id, dtype=dtype, is_vision_model=True) as hf_model:
+    with hf_runner(model, dtype=dtype, is_vision_model=True) as hf_model:
         hf_outputs_per_image = [
             hf_model.generate_greedy_logprobs_limit(prompts,
                                                     max_tokens,
@@ -136,15 +111,15 @@ def run_test(
         check_logprobs_close(
             outputs_0_lst=hf_outputs,
             outputs_1_lst=[
-                vllm_to_hf_output(vllm_output, vlm_config, model_id)
+                vllm_to_hf_output(vllm_output, model)
                 for vllm_output in vllm_outputs
             ],
             name_0="hf",
             name_1="vllm",
         )
 
 
-@pytest.mark.parametrize("model_and_config", model_and_vl_config)
+@pytest.mark.parametrize("model", models)
 @pytest.mark.parametrize(
     "size_factors",
     [
@@ -161,14 +136,13 @@ def run_test(
 @pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [128])
 @pytest.mark.parametrize("num_logprobs", [5])
-def test_models(hf_runner, vllm_runner, image_assets, model_and_config,
-                size_factors, dtype: str, max_tokens: int,
-                num_logprobs: int) -> None:
+def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
+                dtype: str, max_tokens: int, num_logprobs: int) -> None:
     run_test(
         hf_runner,
         vllm_runner,
         image_assets,
-        model_and_config,
+        model,
         size_factors=size_factors,
         dtype=dtype,
         max_tokens=max_tokens,