Skip to content
KernelIndex
Search⌘K
Sign in

Top p sampling from probs v129280

0 eligible runs
sampling

Top-p (nucleus) sampling from probabilities with vocab_size=129280. Filters probabilities using cumulative probability threshold, then samples from the filtered distribution. Captured from DeepSeek V3/R1.

Current records

No published measurement for the selected workload.

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
claude-opus-4-1-20250805 / cudaea19e3claude-opus-4-1-20250805
cuda
—
—
Apache-2.0 · source
claude-opus-4-1-20250805 / triton7a27f9claude-opus-4-1-20250805
triton
—
—
Apache-2.0 · source
gemini-2.5-pro / cuda1e905dgemini-2.5-pro
cuda
—
—
Apache-2.0 · source
gpt-5-2025-08-07_cuda_371839gpt-5-2025-08-07
cuda
—
—
Apache-2.0 · source
gpt-5-2025-08-07 / triton8dfa99gpt-5-2025-08-07
triton
—
—
Apache-2.0 · source
gpt-o3 / cuda14adb8gpt-o3
cuda
—
—
Apache-2.0 · source
gpt-o3 / triton579f5dgpt-o3
triton
—
—
Apache-2.0 · source

Semantics

Inputs and outputs
probsfp32 [batch_size, vocab_size]
top_pfp32 [batch_size]
samplesint64 [batch_size]
Axes and behavior
batch_sizevariable
vocab_sizeconstant = 129280
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliastop_p_sampling_from_probs_v129280modeldeepseek-v3modeldeepseek-r1sha256f1b91d34068a…
No source imports for this operation yet.How records are decidedJSON