batch-invariance

In bfloat16 a trainer gives the same token a different log probability depending on its batch shape; measured on 8 models, with the controls that survived

Batch-invariant inference nodes for guaranteed reproducibility in ComfyUI. ThinkingMachines + ECHO 2.0 + Nemotron patterns.

Bitwise batch-invariant matmul and attention kernels for MLX on Apple Silicon: identical logits at any batch size, with the harness that proves it.

Batch-invariance verifier for llama.cpp continuous batching. Per-cell diffing, a five-verdict output contract, and signed GREEN or RED certificates. Mock-sourced passes are non-promotable by construction. 138 tests, CI green.

量測 LLM 推論的決定性：同一 prompt 重複執行時，輸出從第幾個 token 開始分歧。

Measures whether an LLM inference engine returns the same tokens when requests share engine steps. Real results for OpenVINO GenAI on an Intel CPU and Arc iGPU.