File size: 5,767 Bytes
881f9f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
#!/bin/bash
# PyTorch Profiler wrapper for vLLM-backed benchmark runs.
#
# vLLM ships with built-in torch.profiler support that activates when the
# server is launched with --profiler-config (vLLM >= 0.19.0; the older
# VLLM_TORCH_PROFILER_DIR env-var path was removed). When the flag is set,
# vLLM exposes /start_profile and /stop_profile HTTP endpoints on the serve
# process and writes Chrome-trace-compatible JSON files into the configured
# torch_profiler_dir.
#
# This wrapper:
#   1. Confirms the already-running vLLM server is reachable
#   2. Hits /start_profile, runs the target command, hits /stop_profile
#   3. Writes profile_meta.json alongside the trace
#   4. Returns the exit code of the target command
#
# It does NOT launch vLLM. The canonical capture route is the benchmark
# wrapper (TORCH_PROFILE=1 bash scripts/run_experiment.sh <config>), which
# constructs the --profiler-config flag automatically. Use this script only
# for ad-hoc / debugging captures against a manually-launched vLLM serve.
#
# Usage:
#   bash profiling/scripts/run_vllm_torch_profile.sh <output_dir> \
#       -- <target command that drives inference>
#
# Example: profile the smoke chat prompt end-to-end:
#   bash profiling/scripts/run_vllm_torch_profile.sh profiling/traces/smoke \
#       -- curl -s http://127.0.0.1:8000/v1/chat/completions \
#            -H 'Content-Type: application/json' \
#            -d @/tmp/chat.json
#
# IMPORTANT — ordering:
#
#   This wrapper assumes a vLLM server is ALREADY running in a separate
#   Slurm allocation (or srun --pty shell) and was launched with
#   --profiler-config pointing at a writable absolute path. If vLLM wasn't
#   started with the flag, /start_profile returns an error and this script
#   aborts.
#
#   To start vLLM with profiling enabled (manual recipe):
#     TRACE_DIR="$PWD/profiling/traces/<runid>_torch"
#     mkdir -p "$TRACE_DIR"
#     python -u -m vllm.entrypoints.openai.api_server \
#         --model models/Llama-3.1-8B-Instruct \
#         --served-model-name Llama-3.1-8B-Instruct \
#         --port 8000 --max-model-len 32768 --dtype float16 \
#         --enable-auto-tool-choice --tool-call-parser llama3_json \
#         --profiler-config "{\"profiler\":\"torch\",\"torch_profiler_dir\":\"$TRACE_DIR\"}"
#
#   Or, more typically, use the benchmark wrapper:
#     TORCH_PROFILE=1 bash scripts/run_experiment.sh <config>
#   which handles the flag construction in run_experiment.sh:783-785.
#
# Output:
#   $OUTPUT_DIR/*.pt.trace.json.gz (Chrome trace; vLLM 0.19 emits gzipped form.
#                                   Open via https://ui.perfetto.dev directly,
#                                   or `gunzip -k <file>.pt.trace.json.gz` and
#                                   load the resulting .pt.trace.json in
#                                   chrome://tracing)
#   $OUTPUT_DIR/profile_meta.json  (runid, timestamps, target command)

set -euo pipefail

OUTPUT_DIR="${1:?Usage: $0 <output_dir> -- <command> [args...]}"
shift

if [ "${1:-}" != "--" ]; then
    echo "ERROR: expected -- between output dir and command." >&2
    echo "Usage: $0 <output_dir> -- <command> [args...]" >&2
    exit 1
fi
shift

if [ "$#" -lt 1 ]; then
    echo "ERROR: no command given after --." >&2
    exit 1
fi

VLLM_HOST="${VLLM_HOST:-127.0.0.1}"
VLLM_PORT="${VLLM_PORT:-8000}"
BASE_URL="http://$VLLM_HOST:$VLLM_PORT"

mkdir -p "$OUTPUT_DIR"

# Confirm vLLM is reachable and has profiling endpoints
if ! curl -s "$BASE_URL/health" > /dev/null 2>&1; then
    echo "ERROR: vLLM not reachable at $BASE_URL/health" >&2
    echo "       Start vLLM with --profiler-config '{\"profiler\":\"torch\",\"torch_profiler_dir\":\"<abs-path>\"}' before running this wrapper." >&2
    exit 1
fi

echo "run_vllm_torch_profile: starting profiler via $BASE_URL/start_profile" >&2
START_RESP="$(curl -s -o /tmp/start_profile.out -w '%{http_code}' -X POST "$BASE_URL/start_profile" || true)"
if [ "$START_RESP" != "200" ]; then
    echo "ERROR: /start_profile returned HTTP $START_RESP" >&2
    echo "Response body:" >&2
    cat /tmp/start_profile.out >&2 || true
    echo "" >&2
    echo "Most common cause: vLLM was not started with --profiler-config (vLLM >= 0.19.0)." >&2
    exit 1
fi

START_TS="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
echo "run_vllm_torch_profile: running target command" >&2
echo "  cmd: $*" >&2

set +e
"$@"
CMD_RC=$?
set -e

STOP_TS="$(date -u +%Y-%m-%dT%H:%M:%SZ)"

echo "run_vllm_torch_profile: stopping profiler" >&2
STOP_RESP="$(curl -s -o /tmp/stop_profile.out -w '%{http_code}' -X POST "$BASE_URL/stop_profile" || true)"
if [ "$STOP_RESP" != "200" ]; then
    echo "WARNING: /stop_profile returned HTTP $STOP_RESP (target rc=$CMD_RC)" >&2
    cat /tmp/stop_profile.out >&2 || true
fi

# Write meta alongside the trace so downstream analysis has context
META_FILE="$OUTPUT_DIR/profile_meta.json"
python3 - "$META_FILE" "$START_TS" "$STOP_TS" "$CMD_RC" "$@" <<'PY'
import json
import sys

meta_path = sys.argv[1]
start_ts = sys.argv[2]
stop_ts = sys.argv[3]
cmd_rc = int(sys.argv[4])
cmd = sys.argv[5:]

meta = {
    "start_ts": start_ts,
    "stop_ts": stop_ts,
    "target_command": cmd,
    "target_exit_code": cmd_rc,
    "notes": "Chrome trace written by vLLM into the directory configured via "
             "--profiler-config (vLLM >= 0.19.0). Emitted as gzipped "
             "*.pt.trace.json.gz. Upload the .gz directly to "
             "https://ui.perfetto.dev (handles gzip transparently), or "
             "`gunzip -k <file>.pt.trace.json.gz` first and load the resulting "
             ".pt.trace.json in chrome://tracing.",
}
with open(meta_path, "w") as f:
    json.dump(meta, f, indent=2)
PY

echo "run_vllm_torch_profile: done. Target rc=$CMD_RC. Meta: $META_FILE" >&2
exit "$CMD_RC"