Skip to main content
Process
tools•January 15, 2025

JetBlock Optimizer

Deterministic inference for Nemotron 3 in ComfyUI — same seed, same output, every time.

Most LLM inference is nondeterministic — run the same prompt twice, get different results. JetBlock fixes that for Nemotron 3's hybrid architecture in ComfyUI.

**The Problem** Batch-size variance (not temperature) causes nondeterminism. When server load changes, batch sizes change, and outputs drift.

What JetBlock Does

  • Forces batch-invariant operations across Mamba-2, MoE, and attention layers
  • Three modes: strict (1.6x slower), standard (1.2x slower), relaxed (full speed)
  • Cascade control for reasoning budget between /think and /no_think modes

**Why It Matters** Reproducible outputs enable real on-policy reinforcement learning and consistent creative iteration.

ComfyUIDeterminismNemotronThinkingMachines