Upload scripts/eval_basex32.sh with huggingface_hub

7efe038 verified about 2 months ago

1.84 kB

#!/usr/bin/env bash
# base680 at the SAME config as t128c: T=131072, YaRN x32, ORIG=4096.
# Identical scale AND identical real lengths -> the only difference is the
# extension training. This is the "what did 128k training actually buy" test,
# and it removes the real-length confound in my earlier virtual-position claim.
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:$PATH; export PY=/root/Adhi/BluTrain/envv/bin/python
A680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
L=eval_bx32_logs; mkdir -p "$L"; M="$L/master.log"
say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
one(){
  local PL="$1"; local G="$2"
  local O="bx32_${PL}"
  A1B="$A680" HAY="data/mix128_b6_60/val_shard_00000.bin" GEN_DATA_ROOT="data/mix128_b6_60" \
  CKPT_EXTRA="base=base680.ckpt" CKPTS="base" TASKS="niah_single" N=100 \
  T=131072 YARN=32 YARN_ORIG=4096 PROMPT_LEN="$PL" GPUS="$G" NP=1 JOBS_PER_GPU=1 OUT="$O" \
    ./run_longeval.sh > "$L/len_${PL}.log" 2>&1
  say "  base680 x32 len $PL (virtual $((PL/32))) -> $(grep -h '^niah_single,' $O/summary.csv 2>/dev/null|cut -d, -f2|tr -d '\r')"
}
say "=== base680 @ x32, T=131072 -- matched to t128c ==="
one  4096 0 & one  8192 1 & one 16360 2 & one 32744 3 &
wait
one 65512 0 & one 131000 1 &
wait
say "=== base680 x32 vs t128c x32 (identical config) ==="
printf "  %-8s %-8s %-8s %s\n" len virtual base680 t128c | tee -a "$M"
for c in "4096 99" "8192 97" "16360 99" "32744 92" "65512 45" "131000 28"; do
  set -- $c; b=$(grep -h '^niah_single,' bx32_$1/summary.csv 2>/dev/null|cut -d, -f2|tr -d '\r')
  printf "  %-8s %-8s %-8s %s\n" "$1" "$(($1/32))" "${b:-.}" "$2" | tee -a "$M"
done
say "BASEX32 DONE"

	#!/usr/bin/env bash
	# base680 at the SAME config as t128c: T=131072, YaRN x32, ORIG=4096.
	# Identical scale AND identical real lengths -> the only difference is the
	# extension training. This is the "what did 128k training actually buy" test,
	# and it removes the real-length confound in my earlier virtual-position claim.
	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:$PATH; export PY=/root/Adhi/BluTrain/envv/bin/python
	A680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
	L=eval_bx32_logs; mkdir -p "$L"; M="$L/master.log"
	say(){ echo "[$(date -u +%H:%M:%S)] $*" \| tee -a "$M" >&2; }
	one(){
	local PL="$1"; local G="$2"
	local O="bx32_${PL}"
	A1B="$A680" HAY="data/mix128_b6_60/val_shard_00000.bin" GEN_DATA_ROOT="data/mix128_b6_60" \
	CKPT_EXTRA="base=base680.ckpt" CKPTS="base" TASKS="niah_single" N=100 \
	T=131072 YARN=32 YARN_ORIG=4096 PROMPT_LEN="$PL" GPUS="$G" NP=1 JOBS_PER_GPU=1 OUT="$O" \
	./run_longeval.sh > "$L/len_${PL}.log" 2>&1
	say " base680 x32 len $PL (virtual $((PL/32))) -> $(grep -h '^niah_single,' $O/summary.csv 2>/dev/null\|cut -d, -f2\|tr -d '\r')"
	}
	say "=== base680 @ x32, T=131072 -- matched to t128c ==="
	one 4096 0 & one 8192 1 & one 16360 2 & one 32744 3 &
	wait
	one 65512 0 & one 131000 1 &
	wait
	say "=== base680 x32 vs t128c x32 (identical config) ==="
	printf " %-8s %-8s %-8s %s\n" len virtual base680 t128c \| tee -a "$M"
	for c in "4096 99" "8192 97" "16360 99" "32744 92" "65512 45" "131000 28"; do
	set -- $c; b=$(grep -h '^niah_single,' bx32_$1/summary.csv 2>/dev/null\|cut -d, -f2\|tr -d '\r')
	printf " %-8s %-8s %-8s %s\n" "$1" "$(($1/32))" "${b:-.}" "$2" \| tee -a "$M"
	done
	say "BASEX32 DONE"