Add EPQ pager source and reproducible scripts

2b6ab7c verified 24 days ago

2.53 kB

#!/usr/bin/env bash
# I/O honesty audit: prove the expert weights are read from the block device and
# are NOT served from (or hidden in) the page cache.
#
#   ./audit_io.sh <model-first-shard.gguf> [sdq-chat] [ram-budget-mb] [n-predict]
#
# Two runs with identical settings:
#   A) default                              pager uses O_DIRECT (bypasses the cache)
#   B) SDQ_NO_DIRECT_IO=1 + --no-fadvise    control: buffered reads kept in cache
#
# If this measurement were blind, A and B would look the same. They do not:
# B leaves gigabytes of the GGUF resident in the page cache, A does not.
set -u
MODEL="$(readlink -f "$1")"
BIN="$(readlink -f "${2:-./llama.cpp/build/bin/sdq-chat}")"
BUDGET="${3:-8192}"
NPRED="${4:-32}"
PROMPT="Explain in detail how paging mixture-of-experts weights from SSD works, \
and why the working set is much smaller than the weights."

cd "$(dirname "$0")/.."
SECOND="$(echo "$MODEL" | sed -E 's/-00001-of-/-00002-of-/')"
IFS=',' read -r -a SHARDS <<< "$MODEL,$SECOND"

if [ ! -x /tmp/pagecache_resident ]; then
    echo "building the mincore helper ..."
    g++ -O2 -o /tmp/pagecache_resident sdq-ling/pagecache_resident.cpp
fi

run() {
    local label="$1"; shift
    echo
    echo "================================================================"
    echo "[audit] $label"
    echo "================================================================"
    WATCH_INTERVAL=0.5 "$@" python3 sdq-ling/watch_io.py "$MODEL,$SECOND" \
        "$BIN" -m "$MODEL" -p "$PROMPT" -n "$NPRED" -c 2048 -t 16 -b 128 \
        --temp 0 --no-chat --no-interactive --ram-budget-mb "$BUDGET" \
        2>/dev/null | grep -E "^\[watch\]"
    grep -E "全程 SSD 實際讀取|O_DIRECT 讀取模型檔|O_DIRECT 不可用" /tmp/watch_stderr.log | tail -3
    grep -E "^assistant:" /tmp/watch_stdout.log | head -1
}

echo "[audit] model shards:"
/tmp/pagecache_resident "${SHARDS[@]}"

run "A) default (O_DIRECT)" env SDQ_RAM_BUDGET_MB="$BUDGET"
echo
echo "================================================================"
echo "[audit] B) control: buffered reads + page cache kept"
echo "================================================================"
WATCH_INTERVAL=0.5 SDQ_NO_DIRECT_IO=1 python3 sdq-ling/watch_io.py "$MODEL,$SECOND" \
    "$BIN" -m "$MODEL" -p "$PROMPT" -n "$NPRED" -c 2048 -t 16 -b 128 \
    --temp 0 --no-chat --no-interactive --no-fadvise --ram-budget-mb "$BUDGET" \
    2>/dev/null | grep -E "^\[watch\]"
grep -E "全程 SSD 實際讀取|O_DIRECT 不可用" /tmp/watch_stderr.log | tail -2

	#!/usr/bin/env bash
	# I/O honesty audit: prove the expert weights are read from the block device and
	# are NOT served from (or hidden in) the page cache.
	#
	# ./audit_io.sh <model-first-shard.gguf> [sdq-chat] [ram-budget-mb] [n-predict]
	#
	# Two runs with identical settings:
	# A) default pager uses O_DIRECT (bypasses the cache)
	# B) SDQ_NO_DIRECT_IO=1 + --no-fadvise control: buffered reads kept in cache
	#
	# If this measurement were blind, A and B would look the same. They do not:
	# B leaves gigabytes of the GGUF resident in the page cache, A does not.
	set -u
	MODEL="$(readlink -f "$1")"
	BIN="$(readlink -f "${2:-./llama.cpp/build/bin/sdq-chat}")"
	BUDGET="${3:-8192}"
	NPRED="${4:-32}"
	PROMPT="Explain in detail how paging mixture-of-experts weights from SSD works, \
	and why the working set is much smaller than the weights."

	cd "$(dirname "$0")/.."
	SECOND="$(echo "$MODEL" \| sed -E 's/-00001-of-/-00002-of-/')"
	IFS=',' read -r -a SHARDS <<< "$MODEL,$SECOND"

	if [ ! -x /tmp/pagecache_resident ]; then
	echo "building the mincore helper ..."
	g++ -O2 -o /tmp/pagecache_resident sdq-ling/pagecache_resident.cpp
	fi

	run() {
	local label="$1"; shift
	echo
	echo "================================================================"
	echo "[audit] $label"
	echo "================================================================"
	WATCH_INTERVAL=0.5 "$@" python3 sdq-ling/watch_io.py "$MODEL,$SECOND" \
	"$BIN" -m "$MODEL" -p "$PROMPT" -n "$NPRED" -c 2048 -t 16 -b 128 \
	--temp 0 --no-chat --no-interactive --ram-budget-mb "$BUDGET" \
	2>/dev/null \| grep -E "^\[watch\]"
	grep -E "全程 SSD 實際讀取\|O_DIRECT 讀取模型檔\|O_DIRECT 不可用" /tmp/watch_stderr.log \| tail -3
	grep -E "^assistant:" /tmp/watch_stdout.log \| head -1
	}

	echo "[audit] model shards:"
	/tmp/pagecache_resident "${SHARDS[@]}"

	run "A) default (O_DIRECT)" env SDQ_RAM_BUDGET_MB="$BUDGET"
	echo
	echo "================================================================"
	echo "[audit] B) control: buffered reads + page cache kept"
	echo "================================================================"
	WATCH_INTERVAL=0.5 SDQ_NO_DIRECT_IO=1 python3 sdq-ling/watch_io.py "$MODEL,$SECOND" \
	"$BIN" -m "$MODEL" -p "$PROMPT" -n "$NPRED" -c 2048 -t 16 -b 128 \
	--temp 0 --no-chat --no-interactive --no-fadvise --ram-budget-mb "$BUDGET" \
	2>/dev/null \| grep -E "^\[watch\]"
	grep -E "全程 SSD 實際讀取\|O_DIRECT 不可用" /tmp/watch_stderr.log \| tail -2