File size: 5,082 Bytes
1e59964 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 | #!/usr/bin/env bash
###############################################################################
# StepProbe — 14B scale-up experiment.
#
# Adds a single (model × benchmark × method) cell at 14B scale without
# rerunning the whole matrix. Answers the Q1 reviewer's inevitable "does
# this generalise beyond 8B?" question by producing one diagnosed cell with
# full metrics/CIs. FP16 reference is not available at 14B on a 24 GB card,
# so we diagnose against the FP16 reference from the closest sibling model
# (r1-qwen-7b) as a partial proxy — noted explicitly in the paper's
# limitations if used for the main table.
#
# Expected runtime: ~60-90 min on 3090 Ti (14B GPTQ inference on 500 problems).
###############################################################################
set -euo pipefail
PROJECT_DIR="$(cd "$(dirname "$0")" && pwd)"
cd "$PROJECT_DIR"
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-1}"
export HF_HUB_DOWNLOAD_TIMEOUT=300
MODEL_HF="${MODEL_HF:-deepseek-ai/DeepSeek-R1-Distill-Qwen-14B}"
MODEL_TAG="${MODEL_TAG:-r1-qwen-14b}"
QUANT_METHOD="${QUANT_METHOD:-gptq}"
BITS="${BITS:-4}"
QUANT_TAG="${QUANT_METHOD}_w${BITS}"
BENCHMARK="${BENCHMARK:-math500}"
PY="${PY:-python}"
GPU_MEM="${GPU_MEM:-0.55}"
QUANT_ENV_PY="${QUANT_ENV_PY:-/home/aiteam1/anaconda3/envs/sonthh-stepprobe-quant/bin/python}"
QUANT_MODEL_DIR="${PROJECT_DIR}/results/quantized_models/${MODEL_TAG}_${QUANT_TAG}"
INF_DIR="${PROJECT_DIR}/results/inference/${QUANT_TAG}/${MODEL_TAG}"
SEG_DIR="${PROJECT_DIR}/results/segmented/${QUANT_TAG}/${MODEL_TAG}"
DIAG_DIR="${PROJECT_DIR}/results/diagnosis/${QUANT_TAG}/${MODEL_TAG}"
METRICS_DIR="${PROJECT_DIR}/results/metrics"
# FP16 reference at 14B doesn't fit on a 24 GB card. Fall back to the 7B
# sibling's reference traces, which the 14B should follow closely in
# high-level reasoning structure.
REF_DIR="${PROJECT_DIR}/results/segmented/fp16/r1-qwen-7b"
LOG_DIR="${PROJECT_DIR}/logs"
TIMESTAMP=$(date +%Y%m%d_%H%M%S)
LOG_FILE="${LOG_DIR}/scale_up_${TIMESTAMP}.log"
mkdir -p "$METRICS_DIR" "$LOG_DIR"
log() { echo "[$(date '+%H:%M:%S')] $1" | tee -a "$LOG_FILE"; }
log "=============================================="
log "14B scale-up: $MODEL_TAG / $QUANT_TAG / $BENCHMARK"
log "=============================================="
# 1. Offline quantization (only needed for awq/gptq).
if [[ "$QUANT_METHOD" == "awq" || "$QUANT_METHOD" == "gptq" ]]; then
if [[ ! -f "${QUANT_MODEL_DIR}/config.json" ]]; then
log "RUN: offline quantization ($QUANT_METHOD w${BITS}) — ~30 min"
"$QUANT_ENV_PY" ${PROJECT_DIR}/scripts/quantize_models.py \
--model "$MODEL_HF" \
--method "$QUANT_METHOD" \
--bits "$BITS" \
--group-size 128 \
--output "$QUANT_MODEL_DIR" 2>&1 | tee -a "$LOG_FILE"
else
log "SKIP: quantized model exists at $QUANT_MODEL_DIR"
fi
VLLM_MODEL="$QUANT_MODEL_DIR"
else
VLLM_MODEL="$MODEL_HF"
fi
# 2. Inference.
OUT_FILE="${INF_DIR}/${BENCHMARK}_run0.jsonl"
if [[ -f "$OUT_FILE" ]]; then
log "SKIP: inference — $OUT_FILE exists"
else
log "RUN: vLLM inference on $BENCHMARK"
mkdir -p "$INF_DIR"
$PY ${PROJECT_DIR}/scripts/run_inference.py \
--model "$VLLM_MODEL" \
--quant "$QUANT_METHOD" --bits "$BITS" \
--benchmark "$BENCHMARK" \
--output "$INF_DIR" \
--max-tokens 4096 --num-runs 1 \
--gpu-memory-utilization "$GPU_MEM" 2>&1 | tee -a "$LOG_FILE"
fi
# 3. Segment + diagnose + metrics.
if [[ ! -f "${SEG_DIR}/${BENCHMARK}_run0.jsonl" ]]; then
mkdir -p "$SEG_DIR"
log "RUN: segment"
$PY -m stepprobe.segment --input "$INF_DIR" --output "$SEG_DIR" \
--quant "$QUANT_TAG" 2>&1 | tee -a "$LOG_FILE"
fi
if [[ ! -f "${DIAG_DIR}/${BENCHMARK}_run0.jsonl" ]]; then
if [[ ! -d "$REF_DIR" ]]; then
log "ERROR: FP16 reference dir $REF_DIR missing — scale-up needs a"
log " reference; at 14B the own-model FP16 doesn't fit on a 24GB"
log " card. Either run 'bash run_all.sh --phase 4 --models small' to"
log " populate r1-qwen-7b references, or set REF_DIR=... via env."
exit 1
fi
mkdir -p "$DIAG_DIR"
log "RUN: diagnose (ref = r1-qwen-7b FP16)"
$PY -m stepprobe.diagnose \
--ref "$REF_DIR" --hyp "$SEG_DIR" --output "$DIAG_DIR" \
--alignment dtw 2>&1 | tee -a "$LOG_FILE"
fi
log "RUN: metrics + bootstrap CIs"
$PY -m stepprobe.metrics \
--diagnosis "$DIAG_DIR" --output "$METRICS_DIR" \
--model "$MODEL_TAG" --quant "$QUANT_TAG" 2>&1 | tee -a "$LOG_FILE"
$PY ${PROJECT_DIR}/scripts/compute_ci.py \
--diagnosis "${PROJECT_DIR}/results/diagnosis" \
--output "$METRICS_DIR" --n-boot 5000 2>&1 | tee -a "$LOG_FILE"
log ""
log "DONE — 14B scale-up cell"
log " Metrics: ${METRICS_DIR}/${MODEL_TAG}_${QUANT_TAG}_${BENCHMARK}_run0_metrics.json"
log " Log: $LOG_FILE"
log ""
log "To include in the paper figures: rerun scripts/make_paper_figures.py"
log " — the 14B cell will appear in fig_paper_3_forest.pdf and Table 1."
|