# Historical settings for the superseded 2026-06-24 Gemma 4 26B A4B Q8 B70
# 95 tok/s record. The current Gemma result packet and
# repro/gemma4-26b-a4b-q8-b70-125tps-20260701 document the newer
# 124.977 tok/s strict cold-suite Q8-target record.
#
# Source this file before running the helper scripts if you need to override
# local paths. The scripts also source it automatically with these defaults.

export LLAMA_CPP_COMMIT="${LLAMA_CPP_COMMIT:-c926ad09857517978575d6a74d225b463f7417a0}"
export LLAMA_CPP_DIR="${LLAMA_CPP_DIR:-/home/steve/src/llama.cpp-gemma-record-stack}"
export LLAMA_CPP_BUILD_DIR="${LLAMA_CPP_BUILD_DIR:-${LLAMA_CPP_DIR}/build-sycl-b70-aot-bmg-g31}"
export LLAMA_SYCL_DEVICE_ARCH="${LLAMA_SYCL_DEVICE_ARCH:-intel_gpu_bmg_g31}"

export GEMMA_MODEL_DIR="${GEMMA_MODEL_DIR:-/mnt/fast-ai/llm-models/gemma4-26b-a4b-it-q8-gguf}"
export GEMMA_Q8_MODEL="${GEMMA_Q8_MODEL:-${GEMMA_MODEL_DIR}/gemma-4-26B-A4B-it-UD-Q8_K_XL.gguf}"
export GEMMA_Q8_EXPECTED_BYTES="${GEMMA_Q8_EXPECTED_BYTES:-27636230944}"

export GEMMA_MTP_F16_MODEL="${GEMMA_MTP_F16_MODEL:-${GEMMA_MODEL_DIR}/mtp-gemma-4-26B-A4B-it.gguf}"
export GEMMA_MTP_F16_EXPECTED_BYTES="${GEMMA_MTP_F16_EXPECTED_BYTES:-461766816}"
export GEMMA_MTP_Q40_MODEL="${GEMMA_MTP_Q40_MODEL:-${GEMMA_MODEL_DIR}/MTP/gemma-4-26B-A4B-it-Q4_0-MTP.gguf}"
export GEMMA_MTP_Q40_EXPECTED_BYTES="${GEMMA_MTP_Q40_EXPECTED_BYTES:-321126560}"

export GPU_INDEX="${GPU_INDEX:-0}"
export PORT="${PORT:-18300}"
export CTX_SIZE="${CTX_SIZE:-8192}"
export BATCH_SIZE="${BATCH_SIZE:-512}"
export UBATCH_SIZE="${UBATCH_SIZE:-512}"
export THREADS="${THREADS:-16}"
export POLL="${POLL:-100}"
export CANARY_REPEATS="${CANARY_REPEATS:-96}"
export BENCH_REPEATS="${BENCH_REPEATS:-8}"
export PROMPT_TOKENS="${PROMPT_TOKENS:-512}"
export MAX_TOKENS="${MAX_TOKENS:-512}"
export BENCH_PROMPT_MODE="${BENCH_PROMPT_MODE:-filled-long}"

export MTP_N_MAX="${MTP_N_MAX:-7}"
export MTP_N_MIN="${MTP_N_MIN:-2}"
export MTP_P_MIN="${MTP_P_MIN:-0.12}"
export MTP_DRAFT_DEVICE="${MTP_DRAFT_DEVICE:-SYCL0}"
export MTP_DRAFT_THREADS="${MTP_DRAFT_THREADS:-32}"
export MTP_DRAFT_THREADS_BATCH="${MTP_DRAFT_THREADS_BATCH:-32}"
export MTP_DRAFT_FAST_ARGMAX="${MTP_DRAFT_FAST_ARGMAX:-1}"
export MTP_DRAFT_PROFILE="${MTP_DRAFT_PROFILE:-1}"
export MTP_BACKEND_SAMPLING="${MTP_BACKEND_SAMPLING:-0}"
export MTP_EXTRA_ARGS="${MTP_EXTRA_ARGS:---ctx-checkpoints 0}"

export ONEAPI_DEVICE_SELECTOR="${ONEAPI_DEVICE_SELECTOR:-level_zero:${GPU_INDEX}}"
export GGML_SYCL_ENABLE_VMM="${GGML_SYCL_ENABLE_VMM:-0}"
export GGML_SYCL_DISABLE_OPT="${GGML_SYCL_DISABLE_OPT:-0}"
export FLASH_ATTN="${FLASH_ATTN:-off}"
