Upload scripts/e_short.sh with huggingface_hub

064effe verified about 2 months ago

2.11 kB

#!/usr/bin/env bash
# Does the wider cache COST anything at short context?
# retrain_corrected_cache.sh measured a prior cache switch going 75 -> 14-33 at
# ~4k while improving 1k and 16k. My 32k-cache result only covered 8k/16k/32k,
# so the regime where the known regression lives is UNMEASURED. This runs
# 1,024 and 4,096 under BOTH caches for both models.
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=eshort_logs; mkdir -p "$L"; M="$L/master.log"
say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
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"
HAY=Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
run(){ # T SCALE PL tag
  local T=$1 S=$2 PL=$3 TAG=$4
  local O="eshort_${TAG}_len${PL}"
  [ $((S*4096)) -eq $T ] || { say "YaRN invariant broken: $S*4096 != $T"; return 1; }
  [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ] && {
    say "  $TAG len $PL cached: $(grep '^niah_single,' $O/summary.csv)"; return 0; }
  A1B="$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=$S YARN_ORIG=4096 PROMPT_LEN=$PL GPUS=0,1 NP=1 OUT="$O" \
    ./run_longeval.sh >> "$L/${TAG}_${PL}.log" 2>&1
  if [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ]; then
    say "  $TAG len $PL -> $(head -1 $O/summary.csv) | $(grep '^niah_single,' $O/summary.csv)"
  else say "  $TAG len $PL FAILED"; grep -ahA2 'what()' "$O"/gen/*.log 2>/dev/null|head -3|tee -a "$M" >&2; fi
}
say "=== short-context cost of the wider cache ==="
for PL in 1024 4096; do
  run 16384 4 $PL cache16k
  run 32768 8 $PL cache32k
done
say "=== SHORT-CONTEXT RESULTS (cols: base680,p1final) ==="
for PL in 1024 4096; do for TAG in cache16k cache32k; do
  f="eshort_${TAG}_len${PL}/summary.csv"
  [ -f "$f" ] && say "  len $PL $TAG  $(head -1 $f) -> $(grep '^niah_single,' $f)"
done; done
say "DONE"

	#!/usr/bin/env bash
	# Does the wider cache COST anything at short context?
	# retrain_corrected_cache.sh measured a prior cache switch going 75 -> 14-33 at
	# ~4k while improving 1k and 16k. My 32k-cache result only covered 8k/16k/32k,
	# so the regime where the known regression lives is UNMEASURED. This runs
	# 1,024 and 4,096 under BOTH caches for both models.
	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=eshort_logs; mkdir -p "$L"; M="$L/master.log"
	say(){ echo "[$(date -u +%H:%M:%S)] $*" \| tee -a "$M" >&2; }
	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"
	HAY=Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
	run(){ # T SCALE PL tag
	local T=$1 S=$2 PL=$3 TAG=$4
	local O="eshort_${TAG}_len${PL}"
	[ $((S4096)) -eq $T ] \|\| { say "YaRN invariant broken: $S4096 != $T"; return 1; }
	[ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ] && {
	say " $TAG len $PL cached: $(grep '^niah_single,' $O/summary.csv)"; return 0; }
	A1B="$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=$S YARN_ORIG=4096 PROMPT_LEN=$PL GPUS=0,1 NP=1 OUT="$O" \
	./run_longeval.sh >> "$L/${TAG}_${PL}.log" 2>&1
	if [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ]; then
	say " $TAG len $PL -> $(head -1 $O/summary.csv) \| $(grep '^niah_single,' $O/summary.csv)"
	else say " $TAG len $PL FAILED"; grep -ahA2 'what()' "$O"/gen/*.log 2>/dev/null\|head -3\|tee -a "$M" >&2; fi
	}
	say "=== short-context cost of the wider cache ==="
	for PL in 1024 4096; do
	run 16384 4 $PL cache16k
	run 32768 8 $PL cache32k
	done
	say "=== SHORT-CONTEXT RESULTS (cols: base680,p1final) ==="
	for PL in 1024 4096; do for TAG in cache16k cache32k; do
	f="eshort_${TAG}_len${PL}/summary.csv"
	[ -f "$f" ] && say " len $PL $TAG $(head -1 $f) -> $(grep '^niah_single,' $f)"
	done; done
	say "DONE"