Download scripts/run_benchmark_pair_local.sh from yaro1214/daim: direct link, hf CLI and curl.
- Browser
- Download file 2.25 kB
-
https://huggingface.co/yaro1214/daim/resolve/main/scripts/run_benchmark_pair_local.sh
- Command line
-
hf download hf://yaro1214/daim/scripts/run_benchmark_pair_local.sh
-
curl -L -o run_benchmark_pair_local.sh https://huggingface.co/yaro1214/daim/resolve/main/scripts/run_benchmark_pair_local.sh
2.25 kB
| set -euo pipefail | |
| if [[ $# -ne 2 ]]; then echo "usage: $0 DATASET NUM_SAMPLES" >&2; exit 2; fi | |
| SCRIPT_DIR=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd) | |
| PROJECT_DIR=$(dirname "$SCRIPT_DIR") | |
| MODEL_ROOT=${MODEL_ROOT:-/mnt/bn/lsh-hl-fuse/zyj/model} | |
| TARGET_MODEL=${TARGET_MODEL:-$MODEL_ROOT/Qwen3-4B} | |
| RELAY_DRAFT=${RELAY_DRAFT:-$MODEL_ROOT/Qwen3-4B-target-hidden-relay} | |
| BASELINE_DRAFT=${BASELINE_DRAFT:-$MODEL_ROOT/Qwen3-4B-DFlash-b16} | |
| DATASET=$1; NUM_SAMPLES=$2 | |
| OUTPUT_ROOT=${OUTPUT_ROOT:-$PROJECT_DIR/outputs/local} | |
| MAX_NEW_TOKENS=${MAX_NEW_TOKENS:-2048} | |
| RUN_RELAY=${RUN_RELAY:-1} | |
| RUN_BASELINE=${RUN_BASELINE:-1} | |
| export HF_HOME=${HF_HOME:-$MODEL_ROOT/.cache/huggingface} | |
| export TRANSFORMERS_CACHE=${TRANSFORMERS_CACHE:-$HF_HOME} | |
| export TOKENIZERS_PARALLELISM=false | |
| export PYTHONPATH="$PROJECT_DIR:${PYTHONPATH:-}" | |
| export CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0} | |
| source "$PROJECT_DIR/.venv/bin/activate" | |
| cd "$PROJECT_DIR" | |
| DATASET_DIR="$OUTPUT_ROOT/$DATASET"; mkdir -p "$DATASET_DIR" | |
| if [[ ! -d "$TARGET_MODEL" ]]; then echo "TARGET_MODEL not found: $TARGET_MODEL" >&2; exit 2; fi | |
| if [[ "$RUN_RELAY" == 1 && ! -d "$RELAY_DRAFT" ]]; then echo "RELAY_DRAFT not found: $RELAY_DRAFT" >&2; exit 2; fi | |
| if [[ "$RUN_BASELINE" == 1 && ! -d "$BASELINE_DRAFT" ]]; then echo "BASELINE_DRAFT not found: $BASELINE_DRAFT" >&2; exit 2; fi | |
| if [[ "$RUN_RELAY" == 1 && ! -f "$DATASET_DIR/relay.DONE" ]]; then | |
| python -m torch.distributed.run --standalone --nproc_per_node=1 -m dflash.benchmark \ | |
| --backend transformers --model "$TARGET_MODEL" --draft-model "$RELAY_DRAFT" \ | |
| --dataset "$DATASET" --max-samples "$NUM_SAMPLES" --max-new-tokens "$MAX_NEW_TOKENS" \ | |
| --temperature 0 --block-size 16 --target-hidden-relay >"$DATASET_DIR/relay.log" 2>&1 | |
| touch "$DATASET_DIR/relay.DONE" | |
| fi | |
| if [[ "$RUN_BASELINE" == 1 && ! -f "$DATASET_DIR/dflash.DONE" ]]; then | |
| python -m torch.distributed.run --standalone --nproc_per_node=1 -m dflash.benchmark \ | |
| --backend transformers --model "$TARGET_MODEL" --draft-model "$BASELINE_DRAFT" \ | |
| --dataset "$DATASET" --max-samples "$NUM_SAMPLES" --max-new-tokens "$MAX_NEW_TOKENS" \ | |
| --temperature 0 --block-size 16 >"$DATASET_DIR/dflash.log" 2>&1 | |
| touch "$DATASET_DIR/dflash.DONE" | |
| fi | |