implementation detail · filed under inference & serving
Standardized sampling configuration
Applies a shared sampling configuration across use cases; the cited instance specifies temperature, top-p, and top-k values.
- sources
- 3
- model
- 1
- lab adopt it
- 1
- strongest
- default
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1default 2
Documented in
Evidence
3 spans quoted from the sources, strongest treatment first.
Use the following standardized sampling configuration across all use cases: - temperature=1.0 - top_p=0.95 - top_k=64
defaultinference servingin Gemma 4Google DeepMind
Use the following standardized sampling configuration across all use cases: - temperature=1.0 - top_p=0.95 - top_k=64
defaultinference servingin Gemma 4Google DeepMind
Use the following standardized sampling configuration across all use cases:
usedinference servingin Gemma 4Google DeepMind
Filed alongside
Other methods under inference & serving :: decoding strategy.
Speculative decodingMulti-Token PredictionDSparkEAGLEDFlashPresence PenaltyBest-of-N scaffoldingNEXTN speculative decodingMulti-layer EAGLESpeculative samplingMulti-stage candidate filteringRecursive shared MTP-head draftingTask-specific sampling parametersChat Prefix CompletionConcurrency-aware draft-length tuningDistribution-matched draft-model fine-tuningEAGLE-3-style draft-model fine-tuningFused recurrent replay kernelGrammar-constrained decodingKV-cache sharing between drafter and targetLongest-trace selectionMTP-1 speculative decodingSame-checkpoint target and draft weightsThroughput-aware dynamic verification-length scheduling