[vLLM] Add Omni token counting and text replacement offsets for the Transformers backend (#49435)
* init: Add all files
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* refactor: Split out the modeling-file changes
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* fix: Count Qwen2.5-Omni audio tokens from padded feature frames
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* refactor: Build Gemma3n processor inputs through the ProcessorMixin hooks like Gemma 3
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* refactor: Move the Gemma3n text type check into validate_inputs
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* refactor: Move the Qwen-Omni processors onto the OmniModalProcessorMixin hooks
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* fix: Process audio as usual when load_audio_from_video gets no videos
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* refactor: Address the round 3 review on the Qwen-Omni processors
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* refactor: Address the round 4 review on the Qwen-Omni processors
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* refactor: Address the round 5 review on the Omni processors
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* test: Move the audio token count test into the Mixin
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* refactor: Address the round 7 review on the Omni processors
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* test: Seed the Cohere ASR double-shift test before the model init
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* fix: Count only standalone audio against the placeholders in the Omni mixin
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
* fix: Return a list of prompts from the Gemma3n inputs layout for audio without text
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>
---------
Signed-off-by: Harshal Janjani <harshaljanjani@gmail.com>