メインコンテンツまでスキップ

Runtime & Dispatch

VariableValueEffect
TINYLOOP_LOG_LEVEL0 | 1 | 20 silences output; 1 (default) prints model header + VRAM; 2 adds detail.
TINYLOOP_CUDA_PROFILE1Print per-phase CUDA event timings during benchmark.
TINYLOOP_DISABLE_KV_CACHE1Force uncached generation. ~10× slower; parity testing only.
TINYLOOP_DISABLE_WARMUP1Skip cache-warmup call at end of load_model.
TINYLOOP_DISABLE_MMAP1Disable mmap model loading. Fall back to chunked ifstream.
TINYLOOP_DISABLE_FLASH2_PREFILL1Disable tiled prefill attention.
TINYLOOP_DISABLE_CUBLAS_FP161Disable cuBLAS FP16 GEMM path.
TINYLOOP_USE_CUDNN_SDPA1Route prefill through cuDNN SDPA (FA-3 on Hopper, FA-2 on Ampere).
TINYLOOP_PREFILL_CHUNKNForce chunked prefill at N-token segments.
TINYLOOP_TOKEN_EXITNToken-stable early-exit in scoring.
TINYLOOP_TOKEN_EXIT_VERBOSE1Print early-exit diagnostics.
TINYLOOP_CUDA_POOL1Use CUDA memory pool (cudaMemPool) for allocations instead of bare cudaMalloc.
TINYLOOP_TOKENIZER_DIRdirectoryPath to GPT-2 BPE vocab.json + merges.txt.
TINYLOOP_FP16_RESIDUAL1Use FP16 residual stream instead of FP32. Halves buf.main bandwidth. Opt-in, verified on H100.