Profiling Capture

XProf Profiling Capture

Produce XProf profiling files from sgl-jax for later analysis. sgl-jax is JAX/TPU, so the profiler is jax.profiler and it emits XProf/TensorBoard artifacts (*.xplane.pb + trace.json.gz under plugins/profile/), not torch.profiler. Scope is capture only — producing the trace, not reading it. Analysis belongs to a separate skill.

Workflow

1. Pick a capture path (A/B/C) from the table below by your situation.
2. Warm up before tracing so JIT/autotune is excluded — never trace a cold run. Path A uses --warmup-requests; Path B auto-warms.
3. Default to one non-stage trace; prefill/decode are separated at analysis time (out of scope here). Add --profile-by-stage only when you need pre-split dirs.
4. For prefill, force cache-missing prompts (--dataset-name random / --disable-radix-cache) plus --flush-cache — unless the change under test is cache-related.
5. Run the capture. The trace dir is server-side — confirm it is writable on the host.
6. Wait for flush: confirm *.xplane.pb exists and profile_status == idle. Whether the needed device compute actually landed is confirmed at analysis time, not here.
7. Prefer a single-rank trace when TP emits one per rank.
8. Hand off (see Handoff) — do not analyze; analysis is a separate skill.

Pick a capture path

A — live server

bench_serving --profile is the recommended default: it warms up, arms the profiler, drives a real dataset workload, then stops — one trace over the whole run. A single window reliably captures device ops for both stages; separating prefill from decode is an analysis-time step (out of scope here).

Keep startup precompilation enabled on the live server (the default): do not pass --disable-precompile when launching sgl_jax.launch_server.

Default command:

python -m sgl_jax.bench_serving \
  --backend sgl-jax --base-url http://127.0.0.1:30000 \
  --dataset-name random --num-prompts 8 \
  --random-input-len 4096 --random-output-len 128 --warmup-requests 10 \
  --flush-cache --profile --profile-num-steps 5

Key flags:

Need pre-split prefill/ + decode/ dirs to compare the stages directly? Add --profile-by-stage, optionally with --profile-stages prefill decode — pass one stage to profile only that stage, omit for both.

Manual path: POST /start_profile (body = ProfileReqInput, may include output_dir) → send requests → POST /stop_profile → GET /profile_status until status == "idle" (else "in_progress"):

# 1. arm the profiler (body = ProfileReqInput; all fields optional).
#    Omit num_steps so the window runs until you call stop_profile below.
curl -s -X POST http://127.0.0.1:30000/start_profile \
  -H 'Content-Type: application/json' \
  -d '{"output_dir": "/tmp/myprofile"}'

# 2. send your requests here (bench_serving, curl /generate, real traffic, …)

# 3. stop, then poll status until it flushes to idle
curl -s -X POST http://127.0.0.1:30000/stop_profile
until [ "$(curl -s http://127.0.0.1:30000/profile_status | jq -r .status)" = idle ]; do sleep 1; done

start_profile / stop_profile also accept GET; start_profile with no body uses server defaults. Key ProfileReqInput fields: output_dir, num_steps, start_step, profile_by_stage, profile_stages (e.g. ["prefill","decode"]).

Traces are written server-side, into SGLANG_JAX_PROFILER_DIR (default /tmp), read from the server process's environment. To steer bench_serving output, export it at server launch — a running process cannot be changed. The manual /start_profile path can pass output_dir per request instead. Either way the dir must be writable on the host.

B — offline batch

bench_one_batch takes the same server args as launch_server, so fill the model and parallelism from your own deployment context, then add the batch shape and --profile:

Replace ... with your launch_server config; writes to *.tb/.

python -m sgl_jax.bench_one_batch \
  --model-path <model> --tp-size <n> --device tpu ... \
  --batch-size 1 --input-len 4096 --output-len 128 --profile

• Match your server config. Pass the same parallelism/model flags you launch with (--tp-size, --dp-size, --trust-remote-code, --dtype, …); mismatched config profiles a different program than you run in production.
• Auto-warms up first (a full untraced run), so JIT/autotune is already excluded — do not add warmup here.
• The trace covers prefill plus all decode steps mixed in one *.tb/ dir, with no stage separation.
• Only tp_rank 0 writes the trace.
• Dir name pattern: <prefix>_batch<bs>_input<il>_output<ol>.tb (--profile-filename-prefix).

C — kernel microbench

Use jax.profiler.trace(dir) around a warmed-up run_kernel() as the low-overhead default. Add LLO utilization, compiler dumps, or Ironwood/v7x periodic counters only when the profiling question needs that evidence. The modes are composable but add compilation, profile-size, or runtime overhead. Full selection guide, flags, verification, and delivery details are in references/pallas-kernel-profiling.md.

Capture quality

Details in references/capture-recipes.md. Musts:

• Warm up (--warmup-requests 5) so JIT/autotune is excluded (a step = one forward over the batch; decode step = one token, prefill step ≈ one prompt batch).
• Capture both stages in one non-stage trace; splitting prefill from decode is done at analysis time. Reach for --profile-by-stage only when you need pre-split dirs, knowing its prefill trace is unreliable.
• Avoid the prefix-cache shortcut for prefill. --dataset-name random gives unique (cache-missing) prompts, or run with --disable-radix-cache. Add --flush-cache too — warmup populates the radix cache, so random alone still traces hits where run prompts overlap warmup; flush clears that once before the run. Decode is unaffected. Exception: when profiling a KV/radix/prefix/HiCache change, do not flush — you want controlled cache state, and flushing turns a reuse trace into miss-then-hit.
• Wait for flush, then confirm the capture. *.xplane.pb exists and profile_status == idle are necessary but not sufficient — a truncated window still writes a valid-but-empty file. Whether the device compute you need actually landed is confirmed when the trace is opened for analysis (the device/TensorCore track is non-empty over the stage of interest), not from the file itself. A stage file orders of magnitude smaller than its sibling is under-capture, not success.
• Prefer a single-rank trace when TP emits one per rank.

Output layout

$SGLANG_JAX_PROFILER_DIR/            # server-side; default /tmp
    plugins/profile/<run>/<host>.xplane.pb           # default: single tree (+ .trace.json.gz)
# with --profile-by-stage: split into prefill/ and decode/ subtrees instead
    prefill/plugins/profile/<run>/<host>.xplane.pb
    decode/plugins/profile/<run>/<host>.xplane.pb

Handoff

Do not analyze here — this skill is capture-only. Hand the trace to a companion analysis skill if one is installed in your environment, else open it manually:

• Prefer standalone XProf. Pass the complete logdir that contains plugins/profile (not plugins/profile or the .xplane.pb file):

  pip install -U xprof
  xprof --logdir=/tmp/myprofile --port=8791   # open http://localhost:8791
  

• TensorBoard remains a fallback; install xprof, which now provides its Profile tab and supersedes the formerly recommended tensorboard-plugin-profile package:

  pip install -U tensorboard xprof
  tensorboard --logdir=/tmp/myprofile   # open http://localhost:6006 → Profile tab
  

Both commands consume the full logdir containing plugins/profile. Always report:

• Trace dir(s) and the stage(s) captured.
• Workload shape — dataset, input/output len, warmup / --profile-num-steps.
• Server config — GET /get_server_info.
• Verification — *.xplane.pb present and profile_status == idle.