transformers-backend-vllm-benchmark.sh

#!/usr/bin/env bash
# =============================================================================
#  Reproduce the benchmark behind vLLM PR #47187
#  "Make the Transformers modeling backend as fast as native vLLM"
# =============================================================================
#
#  What this measures
#  ------------------
#  For each model we compare THREE conditions on identical hardware, workload
#  and flags — the only thing that changes is the code path:
#
#    native  --model-impl vllm          vLLM's hand-written model  (the bar to match)
#    after   --model-impl transformers  generic Transformers backend, WITH  PR #47187
#    before  --model-impl transformers  generic Transformers backend, WITHOUT PR #47187
#
#  PR #47187 is 100% Python and lives entirely under
#  vllm/model_executor/models/transformers/ — no CUDA/C++/build changes. So the
#  honest before/after delta is obtained by `git checkout`-ing just those files
#  to the PR's parent commit and back. No recompile, nothing else moves.
#
#  Every number the graphic reports is a WITHIN-model, WITHIN-harness ratio
#  (% of native, and after/before), so the comparison is apples-to-apples even
#  though absolute tok/s differ across model sizes and harnesses.
#
#  Hardware used for the published figures: 8x NVIDIA H100 80GB.
#
#  Usage
#  -----
#    ./benchmark.sh                # run every condition for every model
#    ./benchmark.sh q3_235b        # run only conditions whose tag matches this
#  Results (JSON + logs) land in ./results/. Re-running skips completed configs.
# =============================================================================
set -uo pipefail

# --- configure these two paths for your machine ------------------------------
VLLM=${VLLM:-/fsx/harry/vllm}                 # editable vLLM checkout on PR #47187
VENV=${VENV:-/fsx/harry/.venv/bin}            # venv with that vLLM installed
# -----------------------------------------------------------------------------
PR=5bce653e09                                 # PR #47187 merge commit
PARENT="${PR}^"                               # its parent = "before"
TFDIR=vllm/model_executor/models/transformers/
OUT="$(cd "$(dirname "$0")" && pwd)/results"
PORT=8231
FILTER="${1:-}"                               # optional tag substring filter
mkdir -p "$OUT"

# Shared workload for every run: 1024-token prompts, 128-token generations.
INLEN=1024; OUTLEN=128; NPROMPTS=1000

# Model matrix that produced the graphic.
#   tag | model | harness | tp | gpus | extra-flags
# harness = "throughput" (offline) or "serve" (online; required for data parallel).
MODELS=(
  "q3_4b|Qwen/Qwen3-4B|throughput|1|0|"
  "q3_32b|Qwen/Qwen3-32B|throughput|2|0,1|"
  "q3_235b|Qwen/Qwen3-235B-A22B-FP8|serve|8|0,1,2,3,4,5,6,7|--data-parallel-size 8 --enable-expert-parallel --max-model-len 8192"
)

# Always leave the tree back on the PR files, whatever happens.
restore(){ git -C "$VLLM" checkout "$PR" -- "$TFDIR" 2>/dev/null; }
trap restore EXIT

# Node-local compile caches on tmpfs. On a networked filesystem the multi-rank
# inductor/triton autotune cache races and throws "CUDA driver error: file not
# found"; keeping caches on /dev/shm with single-threaded compile avoids it.
setcache(){
  local tag=$1 c=/dev/shm/benchcache/$1
  rm -rf "$c"; mkdir -p "$c"/{triton,inductor,vllm}
  export TRITON_CACHE_DIR="$c/triton" TORCHINDUCTOR_CACHE_DIR="$c/inductor"
  export VLLM_CACHE_ROOT="$c/vllm"    TORCHINDUCTOR_COMPILE_THREADS=1
}

free_gpus(){  # wait until the GPUs are actually released before the next run
  pkill -9 -f "EngineCore" 2>/dev/null; pkill -9 -f "vllm serve" 2>/dev/null
  pkill -9 -f "VLLM::" 2>/dev/null
  for _ in $(seq 1 90); do
    local u; u=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits \
                 | awk '{s+=$1} END{print s+0}')
    [ "$u" -lt 3000 ] && break; sleep 2
  done
}

# ---- offline throughput harness (dense models) ------------------------------
run_throughput(){ # tag model impl tp gpus extra...
  local tag=$1 model=$2 impl=$3 tp=$4 gpus=$5; shift 5
  local json="$OUT/$tag.json" log="$OUT/$tag.log"
  grep -q tokens_per_second "$json" 2>/dev/null && { echo "SKIP $tag (done)"; return; }
  setcache "$tag"
  echo ">>> $tag : throughput impl=$impl tp=$tp gpus=$gpus"
  CUDA_VISIBLE_DEVICES=$gpus VLLM_LOGGING_LEVEL=INFO timeout 5400 \
    "$VENV/vllm" bench throughput \
      --model "$model" --model-impl "$impl" --tensor-parallel-size "$tp" \
      --dataset-name random --input-len $INLEN --output-len $OUTLEN \
      --num-prompts $NPROMPTS --output-json "$json" "$@" > "$log" 2>&1
  echo -n "    "; grep -o '"tokens_per_second":[^,]*' "$json" 2>/dev/null || echo "NO RESULT"
  echo    "    fused-op log lines: $(grep -cE '^Fused:' "$log")"
  free_gpus
}

# ---- online serving harness (data-parallel MoE) -----------------------------
run_serve(){ # tag model impl tp gpus extra...
  local tag=$1 model=$2 impl=$3 tp=$4 gpus=$5; shift 5
  local json="$OUT/$tag.json" slog="$OUT/${tag}_server.log" clog="$OUT/${tag}_client.log"
  grep -q total_token_throughput "$json" 2>/dev/null && { echo "SKIP $tag (done)"; return; }
  setcache "$tag"; rm -f "$json"
  echo ">>> $tag : serve impl=$impl gpus=$gpus"
  CUDA_VISIBLE_DEVICES=$gpus VLLM_LOGGING_LEVEL=INFO \
    "$VENV/vllm" serve "$model" --model-impl "$impl" --port $PORT "$@" > "$slog" 2>&1 &
  local spid=$! ready=0
  for _ in $(seq 1 400); do
    [ "$(curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:$PORT/health)" = 200 ] \
      && { ready=1; break; }
    kill -0 $spid 2>/dev/null || { echo "    server died (see $slog)"; break; }
    sleep 3
  done
  if [ "$ready" = 1 ]; then
    echo "    server ready -> client"
    "$VENV/vllm" bench serve --model "$model" --host 127.0.0.1 --port $PORT \
      --dataset-name random --random-input-len $INLEN --random-output-len $OUTLEN \
      --num-prompts $NPROMPTS --ignore-eos --save-result --result-filename "$json" > "$clog" 2>&1
    echo -n "    "; grep -oE '"total_token_throughput":[^,}]*' "$json" 2>/dev/null || echo "NO RESULT"
    echo    "    fused-op log lines: $(grep -cE '^Fused:' "$slog")"
  fi
  kill -9 $spid 2>/dev/null; free_gpus
}

dispatch(){ # spec impl condtag
  IFS='|' read -r tag model harness tp gpus extra <<< "$1"
  [ -n "$FILTER" ] && [[ "$tag" != *"$FILTER"* ]] && return
  # shellcheck disable=SC2086
  run_"$harness" "${tag}_$3" "$model" "$2" "$tp" "$gpus" $extra
}

# === Phase A: native + after — run on the PR files ===========================
echo "######## PHASE A: native + after (PR #47187 files) ########"
git -C "$VLLM" checkout "$PR" -- "$TFDIR"
for spec in "${MODELS[@]}"; do
  dispatch "$spec" vllm         native
  dispatch "$spec" transformers after
done

# === Phase B: before — swap ONLY the transformers/ dir to the parent commit ==
echo "######## PHASE B: before (parent commit ${PARENT}) ########"
git -C "$VLLM" checkout "$PARENT" -- "$TFDIR"
git -C "$VLLM" status --porcelain "$TFDIR"
for spec in "${MODELS[@]}"; do
  dispatch "$spec" transformers before
done

restore
echo "######## DONE — results in $OUT ; tree restored to PR files ########"