vllm-project/vllm

[Bug]: Add tests for all parallel sampling parameter combinations

Open

#21.948 aperta il 30 lug 2025

Vedi su GitHub
 (7 commenti) (0 reazioni) (1 assegnatario)Python (16.816 fork)batch import
buggood first issuekeep-openstale

Metriche repository

Star
 (80.034 star)
Metriche merge PR
 (Merge medio 3g 17h) (993 PR mergiate in 30 g)

Descrizione

Verify that we have tests that explicitly exercise parallel sampling (n>1 request sampling parameter) for all of the following combinations:

  • Via AsyncLLM.generate, via LLMEngine add_request() / step()
  • For output_kind equal to each of CUMULATIVE, DELTA, FINAL_ONLY

Ideally a test for each of AsyncLLM and LLMEngine in the appropriate files, with output_kind parameterized.

Additionally we can still test LLM.generate() but this enforces FINAL_ONLY. There is already a test for this one here, but I'm not sure about the other cases.

@sethkimmel3 has reported that the LLMEngine + CUMULATIVE combination is not working properly, this should be exposed and fixed if so.

Guida contributor