The upstream vLLM Qwen2_5_VisionAttention.forward() now requires a
`sequence_lengths` argument, but Qwen3_VisionBlock.forward() does not
pass it, causing a TypeError at init time.
Instead of patching the vLLM package directly, introduce a
Qwen3Omni_VisionTransformer subclass in vllm-omni that overrides
forward() to properly compute and propagate sequence_lengths via
MMEncoderAttention (following qwen3_vl.py's pattern). Both thinker
and talker now import the fixed subclass.
Also bump the vLLM wheel to e196028dbaa41b.
Made-with: Cursor
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Updated the vLLM wheel commit hash in Dockerfile.ci to a60985b07eaf0fefa4c353ad2309b745707591ed.
- Modified the mock shared memory handle name in test_basic_connectors.py for clarity.
- Increased GPU memory utilization to 0.9 and adjusted max_num_batched_tokens to 4096 in qwen2_5_omni_ci.yaml to enable testing on H800.
- Refactored GPU model runner classes to remove deprecated activate_lora parameter and ensure compatibility with num_active_loras.
- Enhanced request handling in the OmniGPUModelRunner class to improve streaming request updates.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Updated the vLLM wheel to commit 24650715105a26b41c88f4164777b5e1a0f3200b in Dockerfile.ci.
- Modified the `calculate_metrics` function to handle `tokenizer` as optional (can be None) in metrics.py.
- Adjusted the `benchmark` function to accept `tokenizer` as optional in patch.py.
- Enhanced condition checks in OmniARScheduler to ensure valid token generation in omni_ar_scheduler.py.
- Updated input processing to include supported tasks in input_processor.py.
- Cleaned up unnecessary code in output_processor.py.
- Added support for SSL ciphers in the API server configuration in api_server.py.
- Enabled speculative decoding in GPU model runners for improved performance in gpu_generation_model_runner.py and gpu_model_runner.py.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Improved the test_overridden_method_signatures_compatible function to better handle missing parameters in method signatures.
- Added logic to account for VAR_KEYWORD parameters, ensuring more accurate detection of signature mismatches.
Signed-off-by: [Your Name] <your.email@example.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Introduced tests to ensure method signatures in OmniLLM match those of the base LLM class, preventing runtime TypeErrors due to parameter mismatches.
- Updated the _run_engine method in OmniLLM to include an optional output_type parameter.
Signed-off-by: [Your Name] <your.email@example.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Updated Dockerfile to use a new vllm wheel version.
- Modified test_put_get_inline to clarify that it always uses SHM and updated metadata assertions.
- Added handling for speculative configuration in GPU model runners to delay clearing connector metadata.
- Refactored payload decoding in OmniGPUModelRunner for better clarity and maintainability.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Update Dockerfile for vllm wheel version and enhance input processing in Omni models. Added decoding for byte messages in async requests, included new metrics in benchmarks, and improved pause state handling in the scheduler. Refined input and output processing methods for better multimodal support and error handling.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Resolve arg_utils merge conflict and align CI base image tag to v0.16.0 for the vLLM rebase track.
Co-authored-by: Cursor <cursoragent@cursor.com>
Signed-off-by: Taichang Zhou <tzhouam@connect.ust.hk>
- Introduced OmniMRotaryEmbedding to extend MRotaryEmbedding with methods for computing input positions across various multimodal scenarios, including image, video, and audio inputs.
- Updated patch.py to replace the original MRotaryEmbedding with OmniMRotaryEmbedding in the module patching logic.
- Created new files for the OmniMRotaryEmbedding implementation and its associated methods.
Signed-off-by: [Your Name] <[Your Email]>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Changed the multimodal_output field to a private attribute _multimodal_output for better encapsulation.
- Added a property method to access multimodal_output, allowing for a clearer interface and improved handling of request outputs.
Signed-off-by: [Your Name] <[Your Email]>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Updated the audio tensor extraction logic to access the correct output structure in Qwen2.5 Omni, Qwen3 Omni, and Qwen3 TTS implementations.
- Ensured consistency in handling audio outputs for improved functionality.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Eliminated the _process_embeds method and related imports from preprocess.py, streamlining the input preprocessing workflow.
- Updated the main processing logic to focus solely on token inputs, enhancing clarity and maintainability.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Reintroduced the load_chat_template import in api_server.py for better template handling.
- Updated the logic in serving_chat.py to streamline the resolution of chat template content formats, enhancing clarity and functionality.
- Adjusted the application of chat templates to ensure proper handling of tools and chat templates.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Changed the import path for multimodal processing to improve module organization.
- Removed the unused multimodal configuration parameter from the Qwen3OmniMoeThinker initialization, streamlining the class definition.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Removed the unused import of supports_mm_encoder_only from interfaces.
- Updated the dummy run check to utilize the multimodal configuration directly, enhancing clarity and future extensibility.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Updated import statements in chat_utils.py to replace the internal content format type with the public one for better clarity.
- Refactored api_server.py to streamline the chat template loading process by replacing the previous method with a direct function call.
- Enhanced serving_chat.py by reintroducing safe_apply_chat_template functions for better template handling.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Modified the pip install command for vllm to use the correct URL format for direct installation from the specified commit hash, ensuring a more reliable installation process.
- Removed the fallback option for pre-release installation to streamline the installation command.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Updated the import path for AnthropicServingMessages in api_server.py to enhance clarity and maintain consistency with the new module structure.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Changed the pip install command for vllm to download the wheel file directly instead of using an extra index URL, ensuring a more reliable installation process.
- Added a fallback option to install the pre-release version of vllm if the direct installation fails.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Refactored import statements in api_server.py to enhance module organization, including changes to the routing and server setup.
- Updated the method for removing routes to ensure compatibility with the new app structure.
- Adjusted the handling of chat completion requests to improve error responses and streamline processing.
- Modified serving modules to align with the new import paths and enhance code clarity.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Modified the pip install command for vllm to include --upgrade, --force-reinstall flags, and an extra index URL, ensuring the latest version is installed from the specified source during testing.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Modified the pip install command for vllm to include --upgrade and --force-reinstall flags, ensuring the latest version is installed during the CI process.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Added a warning in GPUARWorker and GPUGenerationWorker to indicate that the v2 model runner does not include omni hooks, forcing the use of the v1 model runner.
- Updated import statements in gpu_ar_model_runner.py and gpu_generation_model_runner.py to include necessary functions from parallel_state.
- Removed unused imports from both model runner files to streamline the code.
Signed-off-by: [Your Name] <your.email@example.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Deleted the MRotaryEmbedding class and its associated methods from mrope.py to streamline the codebase.
- Removed the get_llm_pos_ids_for_vision function from vision.py as it is no longer needed.
- Updated import statements in qwen2_5_omni.py to reflect the removal of the vision module.
Signed-off-by: [Your Name] <your.email@example.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Updated OmniInputProcessor to streamline input processing, including enhanced handling of prompt embeddings and additional information.
- Introduced OmniTextPrompt and OmniEmbedsInputs to support new input structures with optional metadata.
- Modified input preprocessing methods to accommodate new prompt types and ensure compatibility with multimodal inputs.
- Improved error handling and type checking for data parallel configurations.
Signed-off-by: [Your Name] <your.email@example.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Changed type hint for stage_connector_spec from 'any' to 'Any' for improved type checking and consistency.
- Ensured compatibility with typing standards in the arg_utils module.
Signed-off-by: [Your Name] <your.email@example.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Updated OmniARScheduler to wrap output in OmniSchedulerOutput, including transfer metadata for finished requests needing KV cache transfer.
- Enhanced OmniGenerationScheduler to handle routed experts and finish reasons more effectively, ensuring proper output structure.
- Introduced performance statistics handling in both schedulers for improved monitoring.
Signed-off-by: [Your Name] <your.email@example.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Introduced mm_encoder_only parameter to OmniModelConfig.
- Added a deprecation warning for the `--convert mm_encoder_only` option, advising users to switch to `--mm-encoder-only`.
Signed-off-by: [Your Name] <your.email@example.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
- Modified the constructor to accept torch.Tensor as an option for prompt_embeds.
- Introduced a new static method _maybe_decode_prompt_embeds to handle the conversion from PromptEmbedsPayload to torch.Tensor.
- Preserved the original prompt_embeds payload for reference.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Zhou Taichang