hotpath 0.2.3

Profile vLLM inference under RL-style rollout workloads.

hotpath

Profiler for LLM inference. Kernel timing, request lifecycle tracing, and disaggregation analysis for vLLM and SGLang.

What it does

Profile a live vLLM or SGLang endpoint with real traffic: capture CUDA kernel timing, Prometheus server metrics, and per-request latency breakdowns.

Analyze the results: prefill vs decode phase breakdown, KV cache efficiency, prefix sharing patterns, queue depth over time, TTFT and decode-per-token distributions.

Advise on disaggregation: an analytical M/G/1 queueing model estimates whether splitting prefill and decode onto separate GPU pools improves throughput. If recommended, hotpath generates ready-to-use deployment configs for vLLM, llm-d, and Dynamo.

Install

pip install hotpath

Quick start

Profile a live vLLM server:

hotpath serve-profile \
  --endpoint http://localhost:8000 \
  --traffic prompts.jsonl \
  --concurrency 4 \
  --duration 300 \
  --output .hotpath/run

View results:

hotpath serve-report .hotpath/run/serve_profile.db

Generate disaggregation deployment configs:

hotpath disagg-config .hotpath/run/serve_profile.db --format all

For full server-side timing (queue wait, prefill, decode phases), start vLLM with debug logging and pass the log file:

VLLM_LOGGING_LEVEL=DEBUG vllm serve <model> 2>vllm.log &

hotpath serve-profile \
  --endpoint http://localhost:8000 \
  --traffic prompts.jsonl \
  --server-log vllm.log \
  --concurrency 4 \
  --duration 300

For kernel-level GPU phase breakdown, add --nsys:

hotpath serve-profile --endpoint http://localhost:8000 --traffic prompts.jsonl --nsys

Traffic file format

JSONL, one request per line:

{"prompt": "Explain KV cache eviction policy.", "max_tokens": 256}
{"prompt": "Write a Python retry decorator with exponential backoff.", "max_tokens": 400}

ShareGPT format is also accepted.

Commands

Command	Description
serve-profile	Profile a live vLLM/SGLang server with traffic replay
serve-report	Print a serving analysis report
disagg-config	Generate deployment configs for disaggregated serving
profile	GPU kernel profiling under RL-style rollout workloads
report	View a saved kernel profile
diff	Compare two kernel profiles
bench	Benchmark individual GPU kernel implementations
export	Export profile data to JSON, CSV, or OTLP
doctor	Check local profiling environment
lock-clocks	Lock GPU clocks for reproducible measurements

System requirements

• Linux
• NVIDIA GPU with CUDA driver
• nsys (for kernel profiling; not required for serving analysis)
• vLLM or SGLang (for serving analysis)

Build from source

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel
ctest --test-dir build --output-on-failure

Install from source:

python3 -m venv .venv && . .venv/bin/activate
pip install .

Requirements: CMake 3.28+, C++20 compiler, SQLite3.

How it works

hotpath is a single C++ binary with no runtime dependencies beyond SQLite3.

Data is collected from three sources:

1. Kernel traces -- nsys captures GPU kernel execution. hotpath parses the output, categorizes kernels (GEMM, attention, MoE, etc.), and classifies them as prefill or decode phase by timing correlation with server events.

2. Server metrics -- Prometheus metrics from vLLM or SGLang /metrics endpoints are polled at 1 Hz. Batch size, queue depth, KV cache utilization, and preemption counts are tracked over the profiling window.

3. Request lifecycle -- vLLM debug logs are parsed to extract per-request timestamps: arrival, queue wait, prefill start, decode start, completion. These are stored as structured traces and can be exported as OpenTelemetry spans.

The disaggregation advisor uses a simplified M/G/1 queueing model to estimate whether splitting prefill and decode onto separate GPU pools would improve throughput. It searches over P:D ratios and accounts for KV transfer overhead to produce a concrete recommendation with estimated throughput improvement.

All data is stored in SQLite databases for offline analysis and comparison across runs.

Release notes

See CHANGELOG.md.