A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Docs] add the parallel sampling usage in LLMEngine and AsyncLLM (#24222)
William Song committed
Sep 18, 2025 at 20:37 UTC
c9ff9e6f0cf48615bad1525caeef3025c62b2720
1 file changed
+7
-1
vllm/sampling_params.py
+7
-1
@@ -81,7 +81,13 @@ class SamplingParams(
81
"""
82
83
n: int = 1
84
- """Number of output sequences to return for the given prompt."""
84
+ """Number of outputs to return for the given prompt request.
85
+
86
+ NOTE:
87
+ `AsyncLLM` streams outputs by default. When `n > 1`, all `n` outputs
88
+ are generated and streamed cumulatively per request. To see all `n`
89
+ outputs upon completion, use `output_kind=RequestOutputKind.FINAL_ONLY`
90
+ in `SamplingParams`."""
91
best_of: Optional[int] = None
92
"""Number of output sequences that are generated from the prompt. From
93
these `best_of` sequences, the top `n` sequences are returned. `best_of`