llm-torch-profiler-analysis

name	llm-torch-profiler-analysis
description	Unified LLM torch-profiler triage skill for `sglang`, `vllm`, `TensorRT-LLM`, and `TokenSpeed`. Use it to inspect an existing `trace.json(.gz)` or profile directory, or to drive live profiling against a running server when supported and return one three-table report with kernel, overlap-opportunity, and fuse-pattern tables.

Unified LLM Torch Profiler Analysis

Overview

Use this skill for torch.profiler analysis across:

• sglang
• vllm
• TensorRT-LLM
• TokenSpeed

There is only one public workflow:

• triage

Preferred unified entrypoint:

• scripts/analyze_llm_torch_profile.py

Backwards-compatibility shim (kept so older docker exec ... analyze_sglang_torch_profile.py ... calls keep working; it just forwards to the unified entrypoint):

• scripts/analyze_sglang_torch_profile.py

Markdown bundling helper:

• scripts/render_triage_markdown_bundle.py

triage always prints the same three tables:

• kernel table
• overlap-opportunity table
• fuse-pattern table

By default, all three tables only render rows at or above 1.0% cumulative GPU-time share. Rows below that are hidden by default unless the user asks for a lower cutoff.

Keep the fuse-pattern table source-backed and deterministic. Do not turn it into a fuzzy matcher.

If exact source-backed matching is weak but a kernel cluster is still close to a known family, add one short note after the tables with exactly one of:

• high
• medium
• low

Capability Matrix

Capability	SGLang	vLLM	TensorRT-LLM	TokenSpeed
Existing trace triage	yes	yes	yes	yes
Single-trace live capture	yes	yes, if torch profiler is enabled on server	requires profiler control endpoints	yes, if /start_profile and /stop_profile are exposed
Two-trace mapping+formal triage	yes	yes	yes	yes
Stage-separated live workload	yes	yes	yes, with a writable shared trace dir or per-stage host runner	yes, via workload-separated HTTP capture
--profile-by-stage capture	yes	no	no	no
--profile-prefix control	yes	usually ignored on HTTP profiler route	usually ignored on HTTP profiler route	yes, mapped to profile_id

For TensorRT-LLM, live capture only works when the server exposes /start_profile and /stop_profile, and when the deployment already provides a shared trace path plus the required env vars.

For TokenSpeed, this skill supports both existing trace triage and live capture against current servers that expose /start_profile and /stop_profile. The live helper sends output_dir, activities, with_stack, record_shapes, and profile_id in the start payload. TokenSpeed also has its own native profile_by_stage field for manual capture, but the unified helper uses workload-separated prefill/ and decode/ directories by default so the tables stay comparable across frameworks.

Real B200 Validation

The current B200 smoke run was captured on 2026-06-27 on GPUC5A6 (cirrascale-gpuc5a6) in container sglang_bbuf under:

• /data/bbuf/ai_infra_skills_pr72_20260627/profiler/qwen25_0_5b_live

Validated live profile:

Model	GPU	Workloads	Result
Qwen/Qwen2.5-0.5B-Instruct	1x B200	prefill 128->1, decode 32->8	generated separate prefill/*.trace.json.gz and decode/*.trace.json.gz; kernel, overlap, and fuse tables rendered with separate extend/prefill and decode sections

The same run verified Nsight Compute availability with a tiny CUDA matmul:

• /data/bbuf/ai_infra_skills_pr72_20260627/profiler/ncu_smoke_qwen_matmul.ncu-rep
• /data/bbuf/ai_infra_skills_pr72_20260627/profiler/ncu_smoke.log

After server and ncu cleanup, all eight B200 GPUs reported 0 MiB used.

The 2026-08-23 source refresh did not recapture a live prefill/decode trace or ncu --set basic smoke: the assigned B200 host closed SSH on port 22. Keep the 2026-06-27 artifacts as the last GPU-backed profiler evidence.

Real H100 Validation

The current reference run is the 4x H100 matrix captured on 2026-04-23 on h100_sglang under:

• /data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix_v3

Rendered markdown bundle:

• /data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix_v3/h100_large_model_matrix_v3_bundle.md

Validated model directories:

• mixtral_8x7b_instruct
• qwen2_5_32b_instruct
• qwen3_32b

Each model directory contains:

• analysis_sglang.txt
• analysis_vllm.txt
• analysis_trtllm.txt
• framework-specific trace roots and probe artifacts

Validated matrix:

Model	SGLang	vLLM	TensorRT-LLM	Result
mistralai/Mixtral-8x7B-Instruct-v0.1	4x H100	4x H100	4x H100	three tables rendered correctly on all three frameworks; benchmark probes returned direct, non-empty text
Qwen/Qwen2.5-32B-Instruct	4x H100	4x H100	4x H100	three tables rendered correctly on all three frameworks; benchmark probes returned direct, non-empty text
Qwen/Qwen3-32B	4x H100	4x H100	4x H100	three tables rendered correctly on all three frameworks; vLLM and TensorRT-LLM chat probes often emitted <think> prefixes

Use this run as the main H100 reference. The older 2026-04-22 single-card Qwen3 matrix is still useful for bring-up, but it is not the default reference anymore. TokenSpeed support was added later and is covered by existing-trace triage and HTTP profiler-control support, but it is not part of this older H100 validation matrix yet.

Stage-separated workload validation captured on 2026-05-01 on h100_sglang:

• /data/bbuf/validate/unified_llm_profiler_skill/runs/20260501_stage_split_validation
• /data/bbuf/validate/unified_llm_profiler_skill/runs/20260501_stage_split_validation_large

Validated models:

Model	GPU	Workloads	Result
Qwen/Qwen2.5-0.5B-Instruct	1x H100	prefill 4090->1, decode 1->2048	generated separate prefill/*.trace.json.gz and decode/*.trace.json.gz; kernel, overlap, and fuse tables rendered with separate extend/prefill and decode sections
Qwen/Qwen2.5-1.5B-Instruct	1x H100	prefill 4090->1, decode 1->2048	generated separate prefill/*.trace.json.gz and decode/*.trace.json.gz; kernel, overlap, and fuse tables rendered with separate extend/prefill and decode sections
Qwen/Qwen2.5-7B-Instruct	1x H100	prefill 4090->1, decode 1->2048	generated separate traces; prefill kernel table captured 28-layer GEMM/FA3/RMSNorm work, decode captured 5-step graph launches, and fuse rows were split by stage
Qwen/Qwen2.5-14B-Instruct	1x H100	prefill 4090->1, decode 1->2048	generated separate traces; prefill kernel table captured 48-layer GEMM/FA3/RMSNorm work, decode captured 5-step graph launches, and fuse rows were split by stage
Qwen/Qwen3-8B	2x H100, TP=2	prefill 4090->1, decode 1->2048, warmup 10/capture 5	generated separate prefill/decode traces and all three tables; unique probe prompts avoided prefix-cache pollution in the prefill table
mistralai/Mistral-7B-Instruct-v0.3	2x H100, TP=2	prefill 4090->1, decode 1->2048, warmup 10/capture 5	generated separate prefill/decode traces and all three tables; server logs showed no repeated-prompt prefix-cache shortcut during the active prefill window

This validation also covers the compatibility fix for older SGLang profiler state machines: workload-separated live capture labels stages by output directory and avoids nesting SGLang's internal profile_by_stage state machine inside each workload. The helper adds one internal scheduler guard step because SGLang increments forward_ct before checking whether the profiler should stop; without that guard, a num_steps=1 prefill capture can stop just before the actual prefill forward. The 2026-05-01 two-card validation artifacts for the additional models are:

• /data/bbuf/validate/core_skill_validation_20260501/qwen3_8b/profiler
• /data/bbuf/validate/core_skill_validation_20260501/mistral_7b_instruct_v03/profiler

To render a validated run into one markdown document:

python3 scripts/render_triage_markdown_bundle.py \
  --analysis-root /data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix_v3 \
  --output /data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix_v3/h100_large_model_matrix_v3_bundle.md

The bundle groups by model and keeps the three tables for each framework.

H100 notes:

• all three frameworks now render kernel, overlap, and fuse tables with separate extend/prefill and decode sections when the trace contains a clean stage split
• SGLang live capture is validated and calls the server profiler API directly instead of shelling out to sglang.profiler
• SGLang trace flush can lag well beyond a few seconds, so the runner waits longer for artifacts than the earlier implementation
• SGLang kernel-site reconstruction keeps sampling disabled in the mapping path so the optimized parser does not perturb SGLang table output; equality rechecks matched for Mixtral-8x7B-Instruct-v0.1, Qwen3-32B, and nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8
• vLLM live capture requires --output-dir to match the server torch_profiler_dir; the validated H100 flow uses --profiler-config {"profiler":"torch","torch_profiler_dir":"..."} and then drives /start_profile and /stop_profile
• TensorRT-LLM validation stays on --backend pytorch; the H100 flow writes the trace with TLLM_TORCH_PROFILE_TRACE and then analyzes the saved trace
• TensorRT-LLM current mainline was rechecked at da38c1d2e0dffd073b7dfb6d69e15ee7b45d84a9 on 2026-08-23; PyTorch profiling still uses record_shapes=True and with_modules=True, but not with_stack=True; keep the override path for table-quality Python locations unless the target image proves otherwise
• SGLang v0.5.18 turns the DSV4 fused MHC post+pre path on by default, including SM12x. A decode trace that still shows split MHC post and pre kernels is now more likely a disabled/regressed path than a novel fuse idea
• TokenSpeed live capture is still HTTP profiler-control plus existing-trace triage; it is not part of the older H100 validation matrix
• TokenSpeed trace analysis has first-class registry rows for native TokenSpeed CuTe DSL MLA, MLA KV pack + FP8 quantize, fused top-k/top-p sampling, persistent lm_head GEMM, and NVFP4 GEMM + SwiGLU + quant; live capture still requires an existing torch-profiler trace until the target TokenSpeed image exposes a supported profiler API
• on this host, keep all trace roots under /data/..., not /home/...

When To Use It

• inspect a torch.profiler trace or profile directory from sglang, vllm, TensorRT-LLM, or TokenSpeed
• profile a live serving endpoint and analyze the result
• summarize which kernel families dominate prefill or decode
• map kernels back to Python code paths
• judge whether a code path still leaves overlap opportunity
• check whether an already-known fusion or overlap path should have applied

Diffusion Backend Gate

For diffusion benchmark or profiling work, only analyze traces produced by the native SGLang diffusion backend.

If the run that generated the trace logs any of:

• Falling back to diffusers backend
• Using diffusers backend
• Loaded diffusers pipeline

stop the workflow instead of analyzing the trace. Handle it as a backend-selection issue, not as native-kernel profiler evidence.

Main Flows

Stage-Separated Live Capture Contract

Live capture must not use one mixed prompt as the default. By default, analyze_llm_torch_profile.py --url ... captures two labeled workloads and then renders the same three tables with separate stage sections:

• prefill: synthetic input length 4090, output length 1
• decode: synthetic input length 1, output length 2048

Every live profiler path warms up 10 steps before arming the profiler and then captures 5 active steps by default. Keep this warmup/active split aligned across SGLang, vLLM, and TensorRT-LLM before comparing kernel tables.

Use these options to override the contract when the benchmark workload is known:

--profile-workload both \
--warmup-steps 10 --num-steps 5 \
--prefill-input-len 4090 --prefill-output-len 1 \
--decode-input-len 1 --decode-output-len 2048

Allowed --profile-workload values:

• both: default; capture prefill and decode separately
• prefill: capture only the long-input / one-token workload
• decode: capture only the one-input / long-output workload
• legacy: keep the old --probe-prompt / --probe-max-new-tokens behavior

For sglang-sota-humanize-loop, do not use the defaults if the slow SGLang benchmark scenario has a known input/output distribution. Set the profiler lengths from that slow scenario instead: prefill uses the slow input length with output 1, and decode uses input 1 with the slow output length. For a mixed dataset, profile the slowest representative bucket such as the p50 or p95 input/output pair used in the benchmark report, and record the bucket in the artifact notes.

1. Single-trace triage from an existing profile dir or trace

python3 scripts/analyze_llm_torch_profile.py \
  --input /path/to/profile_dir_or_trace.json.gz

Use this when one trace is enough. The overlap table stays conservative in single-trace mode and will tell you when a mapping/formal pair is needed.

2. Single-trace live capture from SGLang

python3 scripts/analyze_llm_torch_profile.py \
  --framework sglang \
  --url http://127.0.0.1:30000 \
  --output-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example/sglang_profile_live \
  --num-steps 5 \
  --warmup-steps 10 \
  --profile-by-stage \
  --profile-workload both

The script sends POST /start_profile to the SGLang server directly. Keep --output-dir under /data/... so later analysis and docs can see the trace. The script writes server_args.json, warms up with the same workload shape, sends the active probe requests after profiling is armed, captures separate prefill/ and decode/ profile roots by default, and waits longer for trace flush than the earlier implementation. For the default workload-separated capture, the directory name labels the stage and the SGLang internal profile_by_stage mode is not used inside each workload. This avoids mixing a one-token prefill probe with a separate decode profile. The helper still adds one internal guard step because older SGLang profilers check the target counter before running the next forward.

3. Single-trace live capture from vLLM

Launch vLLM with torch profiler enabled, for example:

vllm serve meta-llama/Llama-3.1-8B-Instruct \
  --profiler-config '{"profiler":"torch","torch_profiler_dir":"/data/bbuf/validate/unified_llm_profiler_skill/runs/example/vllm_profile"}'

Then run:

python3 scripts/analyze_llm_torch_profile.py \
  --framework vllm \
  --url http://127.0.0.1:8000 \
  --output-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example/vllm_profile \
  --num-steps 5 \
  --warmup-steps 10 \
  --no-profile-by-stage \
  --profile-workload both

For vLLM, --output-dir must point to the same torch_profiler_dir the server uses. The current vLLM profiler config already defaults torch_profiler_with_stack=true, so the runner only needs to set torch_profiler_dir. On h100_sglang, external vLLM containers should mount both:

• /data/.cache/huggingface:/root/.cache/huggingface
• /data/bbuf/validate/unified_llm_profiler_skill:/data/bbuf/validate/unified_llm_profiler_skill

4. Single-trace live capture from TensorRT-LLM

Use this only when the server exposes POST /start_profile and POST /stop_profile, and the trace path is shared with the current machine.

Typical env expectations are:

• TLLM_PROFILE_START_STOP=<start>-<stop> such as 10-20
• TLLM_TORCH_PROFILE_TRACE=/shared/path/trace.json or .json.gz

Then run:

python3 scripts/analyze_llm_torch_profile.py \
  --framework trtllm \
  --url http://127.0.0.1:8000 \
  --output-dir /shared/path \
  --num-steps 5 \
  --no-profile-by-stage \
  --profile-workload both

If the deployment does not expose the profiler control endpoints, fall back to analyzing an existing trace instead of trying live capture. If the TensorRT-LLM trace output is configured as one fixed file path, use scripts/run_trtllm_pytorch_profile_host.sh --stage prefill and --stage decode instead of direct --profile-workload both, so each stage gets its own trace file.

On the current TensorRT-LLM mainline path, py_executor.py creates the torch profiler with record_shapes=True and with_modules=True but not with_stack=True. For table-quality validation, use the override generator:

python3 scripts/make_trtllm_py_executor_override.py \
  --source /path/to/original/py_executor.py \
  --output /data/bbuf/validate/unified_llm_profiler_skill/overrides/trtllm/py_executor_with_stack.py

The matrix runner does this automatically on H100 before TensorRT-LLM capture starts.

This is the validated TensorRT-LLM flow on h100_sglang:

1. launch trtllm-serve with TLLM_PROFILE_START_STOP=<start>-<stop> and TLLM_TORCH_PROFILE_TRACE=/data/.../trace.json
2. run a few benchmark requests
3. analyze the emitted trace with --input /data/.../trace.json

5. Single-trace live capture or triage from TokenSpeed

For a running TokenSpeed server that exposes the profiler routes, the unified helper can drive live capture:

python3 scripts/analyze_llm_torch_profile.py \
  --framework tokenspeed \
  --url http://127.0.0.1:8000 \
  --output-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example/tokenspeed_profile \
  --num-steps 5 \
  --warmup-steps 10 \
  --no-profile-by-stage \
  --profile-workload both \
  --profile-prefix ts-triage

The helper sends POST /start_profile with:

• output_dir: the --output-dir path
• activities: ["CPU", "GPU"]
• with_stack: true
• record_shapes: false
• profile_id: --profile-prefix, with -prefill or -decode appended during workload-separated capture

It then sends OpenAI-compatible probe requests and calls POST /stop_profile. TokenSpeed writes files such as ts-triage-prefill-TP-0.trace.json.gz under the output directory. If the server was launched with multiple TP ranks, expect one trace per rank.

Existing TokenSpeed torch-profiler traces can still be analyzed directly:

python3 scripts/analyze_llm_torch_profile.py \
  --framework tokenspeed \
  --input /path/to/tokenspeed_profile_dir_or_trace.json.gz

TokenSpeed's own manual profiler control surface can also be used:

curl -X POST http://127.0.0.1:8000/start_profile \
  -H 'Content-Type: application/json' \
  -d '{"output_dir":"/data/bbuf/profiles/tokenspeed","activities":["CPU","GPU"],"with_stack":true,"record_shapes":false,"profile_id":"ts-manual"}'

# send representative workload here

curl -X POST http://127.0.0.1:8000/stop_profile

For server-side automatic stop, pass num_steps. For TokenSpeed-native EXTEND/DECODE split, pass profile_by_stage: true; this produces files with stage suffixes such as -EXTEND and -DECODE.

TokenSpeed's benchmark driver can capture traces too:

tokenspeed bench serve \
  --base-url http://127.0.0.1:8000 \
  --model <model> \
  --dataset-name random \
  --random-input-len 4090 \
  --random-output-len 1 \
  --num-prompts 64 \
  --profile \
  --profile-num-steps 5 \
  --extra-body '{"output_dir":"/data/bbuf/profiles/tokenspeed","activities":["CPU","GPU"],"with_stack":true,"profile_id":"ts-bench"}'

If output_dir is omitted, TokenSpeed falls back to TOKENSPEED_PROFILER_DIR and then /tmp.

Use scripts/probe_llm_server.py with --framework tokenspeed for a small OpenAI-compatible endpoint probe before or after trace collection:

python3 scripts/probe_llm_server.py \
  --framework tokenspeed \
  --url http://127.0.0.1:8000 \
  --requests 6 \
  --max-tokens 48

For sglang-sota-humanize-loop, keep TokenSpeed profiler evidence aligned to the same slow scenario bucket as the benchmark result. Prefer the unified workload-separated live capture when possible; if only a mixed agentic trace is available, label that limitation in analysis/root-cause.md before comparing it to SGLang prefill/decode traces.

6. Two-trace triage from existing profile dirs or traces

python3 scripts/analyze_llm_torch_profile.py \
  --mapping-input /path/to/graph_off_profile_dir \
  --formal-input /path/to/graph_on_profile_dir

Use this when you need stronger overlap attribution and kernel-to-source mapping.

7. Two-trace triage from running servers

python3 scripts/analyze_llm_torch_profile.py \
  --framework sglang \
  --mapping-url http://127.0.0.1:31025 \
  --formal-url http://127.0.0.1:31026 \
  --num-steps 5 \
  --profile-by-stage

For vllm or TensorRT-LLM, use the same shape but pass:

• --framework vllm or --framework trtllm
• --mapping-output-dir ...
• --formal-output-dir ...
• --no-profile-by-stage

For TokenSpeed, either use --mapping-url and --formal-url against servers that expose /start_profile and /stop_profile, or pass two existing trace directories with --mapping-input and --formal-input.

--framework auto uses trace, path, server-argument, and endpoint signatures. If none is conclusive, the analyzer now stops and asks for an explicit sglang, vllm, trtllm, or tokenspeed value; never interpret an unknown trace as SGLang by default.

profile_by_stage

--profile-by-stage is only meaningful on the SGLang live-capture path.

• With --profile-workload both / prefill / decode, workload directories are the stage labels; the live-capture helper disables SGLang's internal stage profiler per workload, warms up first, and captures the requested active step count for the selected workload.
• On legacy or hand-captured SGLang serving, internal profile_by_stage is still useful because prefill and decode usually have very different bottlenecks.
• On the current profile-v2 path inside SGLang, stage-based profiling is effectively the normal path.
• PD-disaggregated serving adds one extra rule: prefill workers and decode workers must be profiled separately. That is stricter than ordinary profile_by_stage.
• For vllm, TensorRT-LLM, and TokenSpeed, disable it with --no-profile-by-stage.

How To Choose The Triage Shape

Single-trace triage

Use when you want the lowest-friction report:

• one trace is already available
• you mainly want kernel share and fusion clues
• you are comparing two runs side by side by running triage once per trace

Prefer this by default.

Two-trace triage

Use when you need:

• a stronger overlap answer
• graph-off source mapping plus graph-on final behavior
• more trustworthy overlap recommendations in the middle table

1. mapping trace with graph disabled or with the lower-fusion / more-readable config
2. formal trace with the real serving optimizations enabled

Do not call the mapping pass a "fast profile". It exists to recover kernel -> cpu_op -> python scope.

Workflow

Single-trace workflow

1. If the user only wants a diagnosis, one trace is enough.
2. Prefer one-rank traces over merged traces whenever the profiler emitted both.
3. For a live server, let the script drive the profiler only when the framework-specific prerequisites are already met.
4. Prefer --profile-workload both; use legacy only when reproducing an old trace contract.
5. Prefer workload-separated SGLang capture; use internal --profile-by-stage mainly for legacy or manually collected traces.
6. When on h100_sglang, create or clean the target trace directory through docker exec sglang_bbuf ... so the path is definitely writable under /data.

Two-trace workflow

1. Produce a mapping trace first with graph disabled or the lower-fusion configuration.
2. Produce a formal trace second with the real serving optimizations enabled.
3. Run triage for the three-table report.
4. Read the results in this order:
   • kernel table
   • overlap-opportunity table
   • fuse-pattern table
5. Before calling something a "new" optimization idea, compare the top rows against both references/fuse-overlap-catalog.md and references/overlap-catalog.md. Check mainline rows first, then the status-audited PR evidence. Prefer reporting:
   • an existing fused or overlap path that should already apply here
   • an existing path that appears disabled, unsupported, or regressed in this trace
   • an upstream pattern that is mainline elsewhere but missing locally, or still open upstream
   • a truly new opportunity only when no catalog entry fits
6. If no exact pattern fully matches but the trace is still close to a known family, add one flat similarity note after the tables. Use high, medium, or low only. Base that note on the full pattern shape, not on one kernel name alone. Prefer semantic cues such as producer-consumer chain, source locations, CPU op names, TP context, and model-specific structure. Do not rewrite the script table itself to include these heuristic judgments.

References

Load these only when needed:

• references/source-map.md
  • upstream SGLang profiler entrypoints and trace-writing paths; still most useful for SGLang-specific source follow-up
• references/heuristics.md
  • overlap labels, dependency-risk interpretation, and limits
• references/fuse-overlap-catalog.md
  • mixed source-backed catalog of existing fuse and overlap patterns, including mainline rows plus explicitly labeled open, merged, and closed-unmerged PR evidence
• references/vllm-torch-compile-fusions.md
  • current vLLM torch.compile fusion passes and the source patterns they target
• references/overlap-catalog.md
  • overlap-only lookup table across LLM, VLM, diffusion, disaggregation, HiSparse, and speculative scheduling

Output Contract

Return:

• trace path or generated profile path
• framework
• model/server args when available
• kernel table
• overlap-opportunity table
• fuse-pattern table
• optional similarity note with high / medium / low when exact matching is inconclusive
• one short summary of what dominates the run
• whether the overlap read came from single-trace triage or mapping/formal two-trace triage