[Bug]: Prefix caching doesn't work for LlavaOneVision

[sleepwalker2017]

Your current environment

The generated dummy input is video, but the preprocessor tries to get image from the dict, and then it crashes.

After I walk around this, the code still fails to run.

It complains this:

  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/engine/core.py", line 264, in run_engine_core
    engine_core.run_busy_loop()
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/engine/core.py", line 302, in run_busy_loop
    outputs = self.step()
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/engine/core.py", line 125, in step
    output = self.model_executor.execute_model(scheduler_output)
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/executor/uniproc_executor.py", line 72, in execute_model
    output = self.worker.execute_model(scheduler_output)
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/worker/gpu_worker.py", line 203, in execute_model                                                                                                              output = self.model_runner.execute_model(scheduler_output)
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/worker/gpu_model_runner.py", line 472, in execute_model
    encoder_outputs = self._gather_encoder_outputs(scheduler_output)
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/worker/gpu_model_runner.py", line 456, in _gather_encoder_outputs
    assert req_id in self.encoder_cache

Model Input Dumps

model=/data/models/llava-onevision-qwen2-7b-ov-hf/
VLLM_USE_V1=1 VLLM_ENABLE_V1_MULTIPROCESSING=1 python3 mmmu_bench.py --model $model --num-prompts 500  --image-hit-rate 0.3

The mmmu_bench.py comes from here:
#11187

🐛 Describe the bug

  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/engine/core.py", line 264, in run_engine_core
    engine_core.run_busy_loop()
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/engine/core.py", line 302, in run_busy_loop
    outputs = self.step()
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/engine/core.py", line 125, in step
    output = self.model_executor.execute_model(scheduler_output)
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/executor/uniproc_executor.py", line 72, in execute_model
    output = self.worker.execute_model(scheduler_output)
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/worker/gpu_worker.py", line 203, in execute_model                                                                                                              output = self.model_runner.execute_model(scheduler_output)
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/worker/gpu_model_runner.py", line 472, in execute_model
    encoder_outputs = self._gather_encoder_outputs(scheduler_output)
  File "/data/test/prefix-caching-vlm/vllm/vllm/v1/worker/gpu_model_runner.py", line 456, in _gather_encoder_outputs
    assert req_id in self.encoder_cache

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[DarkLight1337]

Most multi-modal models don't support V1 yet. You can check the Supported Models page for more details (there is a V1 column for multi-modal models).

[sleepwalker2017]

Most multi-modal models don't support V1 yet. You can check the Supported Models page for more details (there is a V1 column for multi-modal models).

Thank you! Is there a link for that? I didn't find documents about the V1 Engine.

[DarkLight1337]

Most multi-modal models don't support V1 yet. You can check the Supported Models page for more details (there is a V1 column for multi-modal models).

Thank you! Is there a link for that? I didn't find documents about the V1 Engine.

It is pinned in the list of issues .

[DarkLight1337]

cc @ywang96 perhaps we should add a link to the V1 column header?

[sleepwalker2017]

Most multi-modal models don't support V1 yet. You can check the Supported Models page for more details (there is a V1 column for multi-modal models).

Thank you! Is there a link for that? I didn't find documents about the V1 Engine.

It is pinned in the list of issues .

I see this one #8779, but I didn't find any examples about its usage.

It seems the V1 engine is not totally the same to use as the old one.

[DarkLight1337]

Most multi-modal models don't support V1 yet. You can check the Supported Models page for more details (there is a V1 column for multi-modal models).

Thank you! Is there a link for that? I didn't find documents about the V1 Engine.

It is pinned in the list of issues .

I see this one #8779, but I didn't find any examples about its usage.

It seems the V1 engine is not totally the same to use as the old one.

It is still in development which is why we don't have user-facing docs about it yet. For now, you can enable it by setting environment variable VLLM_USE_V1=1.

[ywang96]

@sleepwalker2017 V1 is only available for experimental use and not all multimodal models have yet been supported on V1. You can check our latest documentation here https://docs.vllm.ai/en/latest/models/supported_models.html#id3 (the V1 column) to see which models are supported.

[sleepwalker2017]

https://docs.vllm.ai/en/latest/models/supported_models.html#id3

Thank you for the clear explanation!

[sleepwalker2017]

Most multi-modal models don't support V1 yet. You can check the Supported Models page for more details (there is a V1 column for multi-modal models).

Hi, I have another question, what is needed if I want to add support for a multi-modal model in V1 engine? On the condition that it's already supported by the old engine.

[ywang96]

Most multi-modal models don't support V1 yet. You can check the Supported Models page for more details (there is a V1 column for multi-modal models).

Hi, I have another question, what is needed if I want to add support for a multi-modal model in V1 engine? On the condition that it's already supported by the old engine.

@sleepwalker2017 There are a few key changes you'll need to do:

1. the model's input processor needs to return the PlaceholderRange that tracks where exactly the placeholder tokens for each image starts and its length.
2. The output of get_multimodal_embeddings needs to be
   • Either a tuple of flattened image embeddings (2D tensor of shape [feature_size, hidden_size]), with each corresponding to an image, or
   • A batched 3D tensor in case feature_size is constant across images.

Feel free to take a look at #10699 to see the changes needed. Also for now we only support image modality on V1.

[sleepwalker2017]

changes

Thank you! I'll check that!

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!