Datasets:

jaytonde05
/

kv-cache-quantization-commands

#!/bin/bash
# Exp 2 (accuracy) for one run: RULER (6 KV-sensitive tasks x 30 = 180 prompts per length) + DocVQA (1,000).
# Usage: GPU=0 PORT=8000 LENS="4k 32k 128k" ./exp2.sh RUN_NAME CONFIG [CONC_SCALE]
#   RUN_NAME   result folder name, e.g. BF16 or BF16_rep2
#   CONFIG     BF16 | BF16-FI | FP8 | FP4
#   CONC_SCALE 1 (default) or 0.5 for noise-floor reruns (halves every concurrency)
# Resumable: datasets whose result JSON already exists are skipped.
# Other models (defaults = Qwen3.8-27B): KVQ_MODEL KVQ_NAME KVQ_REASONING_PARSER KVQ_MAX_LEN KVQ_EXTRA_BODY
#   KVQ_DATA (data folder) KVQ_EXP_OUT (results folder) under ~/bench; MRCR=1 also runs MRCR (KVQ_MRCR_OUT).
#   KVQ_BACKEND overrides the attention backend, KVQ_SERVE_ARGS adds vllm serve args, KVQ_DOCVQA=0 skips DocVQA.
set -u
RUN=$1; CFG=$2; SCALE=${3:-1}
case $RUN in *_rep2) echo "== skip $RUN (noise-floor runs dropped)"; exit 0 ;; esac
GPU=${GPU:-0}; PORT=${PORT:-8000}; LENS=${LENS:-"4k 32k 128k"}
export CUDA_VISIBLE_DEVICES=$GPU
source ~/kvq-env/bin/activate
MODEL=${KVQ_MODEL:-~/models/Qwen3.8-27B}; NAME=${KVQ_NAME:-Qwen/Qwen3.8-27B}
RP=${KVQ_REASONING_PARSER:-qwen3}; MAXLEN=${KVQ_MAX_LEN:-163840}
EXTRA=${KVQ_EXTRA_BODY:-'{"reasoning_effort":"low"}'}
case $CFG in
  BF16)    KV=bfloat16; BE=FLASH_ATTN ;;
  BF16-FI) KV=bfloat16; BE=FLASHINFER ;;
  FP8)     KV=fp8_e4m3; BE=FLASH_ATTN ;;
  FP4)     KV=nvfp4;    BE=FLASHINFER ;;
  *) echo "unknown config $CFG"; exit 2 ;;
esac
BE=${KVQ_BACKEND:-$BE}
OUT=~/bench/${KVQ_EXP_OUT:-exp2}/$RUN; mkdir -p $OUT; D=~/bench/${KVQ_DATA:-data}
LOG=$OUT/server.log
echo "== RUN $RUN ($CFG): kv=$KV backend=$BE scale=$SCALE gpu=$GPU port=$PORT  $(date -u)"
vllm serve $MODEL --served-model-name $NAME --port $PORT --dtype bfloat16 \
  --max-model-len $MAXLEN --gpu-memory-utilization 0.90 --reasoning-parser $RP --seed 0 \
  --kv-cache-dtype $KV --attention-backend $BE ${KVQ_SERVE_ARGS:-} >> $LOG 2>&1 &
PID=$!
for i in $(seq 1 180); do
  curl -sf localhost:$PORT/health >/dev/null && { echo "server ready after $((i*10))s"; break; }
  kill -0 $PID 2>/dev/null || { echo "SERVER_EXITED"; tail -40 $LOG; echo "EXP2_FAIL $RUN"; exit 1; }
  sleep 10
done
curl -sf localhost:$PORT/health >/dev/null || { echo "SERVER_TIMEOUT"; kill $PID; echo "EXP2_FAIL $RUN"; exit 1; }
BENCH="vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:$PORT \
  --model $NAME --tokenizer $MODEL --skip-chat-template --disable-shuffle --no-oversample \
  --custom-output-len -1 --save-result --save-detailed --result-dir $OUT --num-warmups 0 \
  --temperature 0 --extra-body $EXTRA"
conc() { python3 -c "print(max(1,int($1*$SCALE)))"; }
run() {  # name file n concurrency [extra args]
  local name=$1 file=$2 n=$3 c; c=$(conc $4); shift 4
  if [ -s $OUT/$name.json ]; then echo "== skip $name (done)"; return; fi
  echo "== bench $name n=$n conc=$c  $(date -u +%T)"
  $BENCH --dataset-path $D/$file --num-prompts $n --max-concurrency $c --result-filename $name.json "$@" 2>&1 \
    | grep -E "Successful requests|Failed requests|Benchmark duration|Error 0|Traceback"
}
# Same concurrency for every config, sized to bf16 KV capacity (1.12M tokens).
declare -A CONC=([4k]=64 [32k]=24 [128k]=6)
# DocVQA first: it is small (~5-10 min).
[ "${KVQ_DOCVQA:-1}" = 1 ] && run docvqa docvqa_1000.jsonl 1000 32 --dataset-name custom_image \
    --custom-ensure-client-side-data --custom-output-len 16384
for L in $LENS; do
  run ruler_$L ruler_${L}_6t.jsonl 180 ${CONC[$L]} --dataset-name custom
done
if [ "${MRCR:-0}" = 1 ]; then
  echo "== mrcr  $(date -u +%T)"
  KVQ_TOK=$MODEL KVQ_EXTRA_BODY=$EXTRA python ~/bench/mrcr.py run $RUN $PORT 2>&1 | grep -E "^== |MRCR_DONE|Traceback"
fi
kill $PID; wait $PID 2>/dev/null
echo "EXP2_OK $RUN $(date -u)"

	#!/bin/bash
	# Exp 2 (accuracy) for one run: RULER (6 KV-sensitive tasks x 30 = 180 prompts per length) + DocVQA (1,000).
	# Usage: GPU=0 PORT=8000 LENS="4k 32k 128k" ./exp2.sh RUN_NAME CONFIG [CONC_SCALE]
	# RUN_NAME result folder name, e.g. BF16 or BF16_rep2
	# CONFIG BF16 \| BF16-FI \| FP8 \| FP4
	# CONC_SCALE 1 (default) or 0.5 for noise-floor reruns (halves every concurrency)
	# Resumable: datasets whose result JSON already exists are skipped.
	# Other models (defaults = Qwen3.8-27B): KVQ_MODEL KVQ_NAME KVQ_REASONING_PARSER KVQ_MAX_LEN KVQ_EXTRA_BODY
	# KVQ_DATA (data folder) KVQ_EXP_OUT (results folder) under ~/bench; MRCR=1 also runs MRCR (KVQ_MRCR_OUT).
	# KVQ_BACKEND overrides the attention backend, KVQ_SERVE_ARGS adds vllm serve args, KVQ_DOCVQA=0 skips DocVQA.
	set -u
	RUN=$1; CFG=$2; SCALE=${3:-1}
	case $RUN in *_rep2) echo "== skip $RUN (noise-floor runs dropped)"; exit 0 ;; esac
	GPU=${GPU:-0}; PORT=${PORT:-8000}; LENS=${LENS:-"4k 32k 128k"}
	export CUDA_VISIBLE_DEVICES=$GPU
	source ~/kvq-env/bin/activate
	MODEL=${KVQ_MODEL:-~/models/Qwen3.8-27B}; NAME=${KVQ_NAME:-Qwen/Qwen3.8-27B}
	RP=${KVQ_REASONING_PARSER:-qwen3}; MAXLEN=${KVQ_MAX_LEN:-163840}
	EXTRA=${KVQ_EXTRA_BODY:-'{"reasoning_effort":"low"}'}
	case $CFG in
	BF16) KV=bfloat16; BE=FLASH_ATTN ;;
	BF16-FI) KV=bfloat16; BE=FLASHINFER ;;
	FP8) KV=fp8_e4m3; BE=FLASH_ATTN ;;
	FP4) KV=nvfp4; BE=FLASHINFER ;;
	*) echo "unknown config $CFG"; exit 2 ;;
	esac
	BE=${KVQ_BACKEND:-$BE}
	OUT=~/bench/${KVQ_EXP_OUT:-exp2}/$RUN; mkdir -p $OUT; D=~/bench/${KVQ_DATA:-data}
	LOG=$OUT/server.log

	echo "== RUN $RUN ($CFG): kv=$KV backend=$BE scale=$SCALE gpu=$GPU port=$PORT $(date -u)"
	vllm serve $MODEL --served-model-name $NAME --port $PORT --dtype bfloat16 \
	--max-model-len $MAXLEN --gpu-memory-utilization 0.90 --reasoning-parser $RP --seed 0 \
	--kv-cache-dtype $KV --attention-backend $BE ${KVQ_SERVE_ARGS:-} >> $LOG 2>&1 &
	PID=$!
	for i in $(seq 1 180); do
	curl -sf localhost:$PORT/health >/dev/null && { echo "server ready after $((i*10))s"; break; }
	kill -0 $PID 2>/dev/null \|\| { echo "SERVER_EXITED"; tail -40 $LOG; echo "EXP2_FAIL $RUN"; exit 1; }
	sleep 10
	done
	curl -sf localhost:$PORT/health >/dev/null \|\| { echo "SERVER_TIMEOUT"; kill $PID; echo "EXP2_FAIL $RUN"; exit 1; }

	BENCH="vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:$PORT \
	--model $NAME --tokenizer $MODEL --skip-chat-template --disable-shuffle --no-oversample \
	--custom-output-len -1 --save-result --save-detailed --result-dir $OUT --num-warmups 0 \
	--temperature 0 --extra-body $EXTRA"

	conc() { python3 -c "print(max(1,int($1*$SCALE)))"; }

	run() { # name file n concurrency [extra args]
	local name=$1 file=$2 n=$3 c; c=$(conc $4); shift 4
	if [ -s $OUT/$name.json ]; then echo "== skip $name (done)"; return; fi
	echo "== bench $name n=$n conc=$c $(date -u +%T)"
	$BENCH --dataset-path $D/$file --num-prompts $n --max-concurrency $c --result-filename $name.json "$@" 2>&1 \
	\| grep -E "Successful requests\|Failed requests\|Benchmark duration\|Error 0\|Traceback"
	}
	# Same concurrency for every config, sized to bf16 KV capacity (1.12M tokens).
	declare -A CONC=([4k]=64 [32k]=24 [128k]=6)
	# DocVQA first: it is small (~5-10 min).
	[ "${KVQ_DOCVQA:-1}" = 1 ] && run docvqa docvqa_1000.jsonl 1000 32 --dataset-name custom_image \
	--custom-ensure-client-side-data --custom-output-len 16384
	for L in $LENS; do
	run ruler_$L ruler_${L}_6t.jsonl 180 ${CONC[$L]} --dataset-name custom
	done

	if [ "${MRCR:-0}" = 1 ]; then
	echo "== mrcr $(date -u +%T)"
	KVQ_TOK=$MODEL KVQ_EXTRA_BODY=$EXTRA python ~/bench/mrcr.py run $RUN $PORT 2>&1 \| grep -E "^== \|MRCR_DONE\|Traceback"
	fi

	kill $PID; wait $PID 2>/dev/null
	echo "EXP2_OK $RUN $(date -u)"