Repository navigation
[Cohere] Enable Cohere Transcribe - #38120
Conversation
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
|
Documentation preview: https://vllm--38120.org.readthedocs.build/en/38120/ |
There was a problem hiding this comment.
Code Review
This pull request integrates the Cohere-Transcribe model by updating model paths, enabling its online availability, adding it to the supported models documentation, and incorporating it into transcription API correctness tests. The tests now utilize a standardized English text normalizer from Whisper for evaluation. A review comment suggests refactoring the model registry entry for CohereASRForConditionalGeneration to avoid duplication and improve maintainability.
|
Hi @ekagra-ranjan, the pre-commit checks have failed. Please run: uv pip install pre-commit>=4.5.1
pre-commit install
pre-commit run --all-filesThen, commit the changes and push to your branch. For future commits, Tip Is
|
|
Some CI is failing due to the model name issue. I'll fix it |
…neration Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com> Signed-off-by: Nithin Chalapathi <nithin.ch10@gmail.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com> Signed-off-by: Rishi Puri <riship@nvidia.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Followup to #35809
Turns on
Cohere Transcribemodel (official name) refered asCohereAsrForConditionalGenerationwithing vLLM code. The official HF repo will beCohereLabs/cohere-transcribe-03-2026.The PR also adds the model to test_transcription_api_correctness.py but doesnt enable it yet. Its adds additional value to this test because this model touches all the recent changes that was done to the variable length encoder inputs. Earlier, vLLM only had padded length encoder inputs which was suited for whisper but not for Cohere-Transcribe.
That test calls a normalizer from the tokenizer but not all the models have a normalizer so we use the standard normalizer in the test for all the models in that test.
Cmd to run the model:
Install vLLM nightly and dependencies
Start server
Send request