본문으로 건너뛰기

Quantisation Environment Variables

VariableValueEffect
TINYLOOP_CROSS_ITER_ROUND1Stochastic cross-iteration rounding. O(√L) error. Needs v5 file with sign_bit.
TINYLOOP_CANCEL_ROUND1Deterministic error cancellation. O(1) error. Architecture-exclusive. Needs v5 sign_bit. Takes priority over stochastic.
TINYLOOP_USE_MARLIN1Marlin INT4 tensor-core dispatch. Ampere/Ada only. 21× over naive INT4. 0.67% PPL drift.
TINYLOOP_DUAL_BIT1Dual-bit weight swap: clone loop block to INT2 at load time. Early iterations use INT2, late use INT4. 3.2% faster. Architecture-exclusive.
TINYLOOP_DUAL_BIT_SWITCHintIteration index to switch from INT2 to INT4 (default 3 = first 4 iters INT2).
TINYLOOP_TRIPLE_BIT1Triple-bit: adds INT8 copy for iter 0 on top of dual-bit. Requires TINYLOOP_DUAL_BIT=1. Highest quality for first iteration.
TINYLOOP_ENABLE_INT4_GEMM_SMALL_BT1Opt-in small-BT INT4 fast path (negative result, kept gated).