| # Fill the idle GPU 3: extend both curves at 65,512. | |
| set -uo pipefail; cd "$(dirname "$0")" | |
| export LD_LIBRARY_PATH="$(cd ../.. && pwd)/Tensor-Implementations/lib:$(cd ../.. && pwd)/Profiler/lib:${LD_LIBRARY_PATH:-}" | |
| export PATH=/usr/local/cuda/bin:$PATH; export PY=/root/Adhi/BluTrain/envv/bin/python | |
| L=eval_gpu3_logs; mkdir -p "$L" | |
| A680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1" | |
| say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$L/master.log" >&2; } | |
| say "=== GPU3: t64 + c500 + d100 @65,512 ===" | |
| A1B="$A680" HAY="data/mix128_b6_60/val_shard_00000.bin" GEN_DATA_ROOT="data/mix128_b6_60" \ | |
| CKPT_EXTRA="t64=t64.ckpt c500=t128c_s500.ckpt d100=t128d_s100.ckpt" CKPTS="t64 c500 d100" \ | |
| TASKS="niah_single" N=100 T=131072 YARN=32 YARN_ORIG=4096 PROMPT_LEN=65512 \ | |
| GPUS="3" NP=1 JOBS_PER_GPU=2 OUT="eg3_65512" ./run_longeval.sh > "$L/len65512.log" 2>&1 | |
| say " 65512: $(head -1 eg3_65512/summary.csv 2>/dev/null|tr -d '\r') -> $(grep -h '^niah_single,' eg3_65512/summary.csv 2>/dev/null|tr -d '\r')" | |
| say "EVAL_GPU3_DONE" | |