Upload scripts/e128k.sh with huggingface_hub

f827ec2 verified about 2 months ago

2.43 kB

#!/usr/bin/env bash
# T=131,072 / YaRN 32 x 4096. A PRE-REGISTERED test of the virtual-position model.
#
# Hypothesis: score is governed by where the last token lands inside the 4,096
# range the model was pretrained on (real position / scale), not by token count.
#
# PREDICTIONS, written before the run:
#    4,096 tok -> virtual   128  : 92-99   (was 92 at x16)
#   32,744 tok -> virtual 1,023  : 85-92   (was 54 at x16)  <- the big one
#   65,512 tok -> virtual 2,047  : 45-55   (was 32 at x16)
# If 32,744 does NOT jump from 54 into the 80s, the virtual-position model is
# wrong and the gains so far were something else.
set -uo pipefail
cd "$(dirname "$0")"
export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
source /root/Adhi/BluTrain/envv/bin/activate 2>/dev/null
L=e128k_logs; mkdir -p "$L"; M="$L/master.log"
say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
T=131072; SCALE=32; ORIG=4096
[ $((SCALE*ORIG)) -eq $T ] || { say "FATAL YaRN invariant"; exit 1; }
HAY=Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
[ "$(stat -c%s "$HAY")" = 199754694 ] || { say "FATAL haystack differs"; exit 1; }
A1B_680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
say "=== T=$T YaRN ${SCALE}x${ORIG} -- predictions: 4096->92-99, 32744->85-92, 65512->45-55 ==="
for PL in 4096 32744 65512; do
  O="e128k_len${PL}"
  [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ] && {
    say "  len $PL cached: $(grep '^niah_single,' $O/summary.csv)"; continue; }
  say "eval prompt_len=$PL  (virtual $((PL/SCALE)))"
  A1B="$A1B_680" HAY="$HAY" GEN_DATA_ROOT="Data_Loader/Data/evalsrc" \
  CKPT_EXTRA="p1final=blumodelcp_run12_step_1200.ckpt base680=varybatch.ckpt" \
  CKPTS="p1final base680" TASKS=niah_single N=100 \
  T=$T YARN=$SCALE YARN_ORIG=$ORIG PROMPT_LEN=$PL GPUS=0,1 NP=1 OUT="$O" \
    ./run_longeval.sh >> "$L/eval_${PL}.log" 2>&1
  if [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ]; then
    say "  len $PL (virtual $((PL/SCALE))) -> $(head -1 $O/summary.csv) | $(grep '^niah_single,' $O/summary.csv)"
  else
    say "  len $PL FAILED"; grep -ahA2 'what()\|out of memory' "$O"/gen/*.log 2>/dev/null|head -4|tee -a "$M" >&2
  fi
done
say "=== RESULTS T=$T (cols base680,p1final) ==="
for PL in 4096 32744 65512; do
  f="e128k_len${PL}/summary.csv"
  [ -f "$f" ] && say "  len $PL virtual $((PL/SCALE))  $(grep '^niah_single,' $f)"
done
say "DONE"

	#!/usr/bin/env bash
	# T=131,072 / YaRN 32 x 4096. A PRE-REGISTERED test of the virtual-position model.
	#
	# Hypothesis: score is governed by where the last token lands inside the 4,096
	# range the model was pretrained on (real position / scale), not by token count.
	#
	# PREDICTIONS, written before the run:
	# 4,096 tok -> virtual 128 : 92-99 (was 92 at x16)
	# 32,744 tok -> virtual 1,023 : 85-92 (was 54 at x16) <- the big one
	# 65,512 tok -> virtual 2,047 : 45-55 (was 32 at x16)
	# If 32,744 does NOT jump from 54 into the 80s, the virtual-position model is
	# wrong and the gains so far were something else.
	set -uo pipefail
	cd "$(dirname "$0")"
	export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
	source /root/Adhi/BluTrain/envv/bin/activate 2>/dev/null
	L=e128k_logs; mkdir -p "$L"; M="$L/master.log"
	say(){ echo "[$(date -u +%H:%M:%S)] $*" \| tee -a "$M" >&2; }
	T=131072; SCALE=32; ORIG=4096
	[ $((SCALE*ORIG)) -eq $T ] \|\| { say "FATAL YaRN invariant"; exit 1; }
	HAY=Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
	[ "$(stat -c%s "$HAY")" = 199754694 ] \|\| { say "FATAL haystack differs"; exit 1; }
	A1B_680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
	say "=== T=$T YaRN ${SCALE}x${ORIG} -- predictions: 4096->92-99, 32744->85-92, 65512->45-55 ==="
	for PL in 4096 32744 65512; do
	O="e128k_len${PL}"
	[ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ] && {
	say " len $PL cached: $(grep '^niah_single,' $O/summary.csv)"; continue; }
	say "eval prompt_len=$PL (virtual $((PL/SCALE)))"
	A1B="$A1B_680" HAY="$HAY" GEN_DATA_ROOT="Data_Loader/Data/evalsrc" \
	CKPT_EXTRA="p1final=blumodelcp_run12_step_1200.ckpt base680=varybatch.ckpt" \
	CKPTS="p1final base680" TASKS=niah_single N=100 \
	T=$T YARN=$SCALE YARN_ORIG=$ORIG PROMPT_LEN=$PL GPUS=0,1 NP=1 OUT="$O" \
	./run_longeval.sh >> "$L/eval_${PL}.log" 2>&1
	if [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ]; then
	say " len $PL (virtual $((PL/SCALE))) -> $(head -1 $O/summary.csv) \| $(grep '^niah_single,' $O/summary.csv)"
	else
	say " len $PL FAILED"; grep -ahA2 'what()\\|out of memory' "$O"/gen/*.log 2>/dev/null\|head -4\|tee -a "$M" >&2
	fi
	done
	say "=== RESULTS T=$T (cols base680,p1final) ==="
	for PL in 4096 32744 65512; do
	f="e128k_len${PL}/summary.csv"
	[ -f "$f" ] && say " len $PL virtual $((PL/SCALE)) $(grep '^niah_single,' $f)"
	done
	say "DONE"