FLOTorch FloTest

LLM inference load testing and benchmarking tool. Measure TTFT, TPS, latency percentiles, and throughput of any OpenAI-compatible or SageMaker endpoint under sustained concurrent load.

Features

• Single-command benchmarking with auto-generated synthetic prompts
• Accurate concurrency control with ramp-up/ramp-down phases
• Streaming and non-streaming support
• Comprehensive metrics: TTFT, TTFNT, TPS, RPM, ITL, E2E latency, percentiles (p25–p99)
• JSON and CSV report exports with per-request logs
• Cache-hit simulation for testing endpoint caching behavior
• OpenAI-compatible API and AWS SageMaker backends

Installation

Requires Node.js 22.19+.

# npm
npm install -g @flotorch/flotest

# pnpm
pnpm add -g @flotorch/flotest

# yarn
yarn global add @flotorch/flotest

After installation, the flotest command is available globally.

Quick Start

1. Generate a config file

flotest init

This launches an interactive wizard that asks for:

• Provider adapter — openai or sagemaker (default: openai)
• Model name — the model identifier your endpoint expects
• Base URL — API endpoint (default: https://api.openai.com/v1)
• Concurrency — number of parallel requests (default: 10)
• Input tokens mean — average input token count per request (default: 512)
• Output tokens mean — average output token count per request (default: 256)
• Max requests — total number of requests to send (default: 100)
• Streaming — whether to stream responses (default: y)

Writes config.json to the current directory. You can specify a custom path:

flotest init my-test.json

2. Set credentials

OpenAI / OpenAI-compatible:

export OPENAI_API_KEY="sk-..."

AWS SageMaker:

The SageMaker backend reads standard AWS environment variables. At minimum, set AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY. Provide AWS_SESSION_TOKEN when using temporary credentials (e.g., aws sts assume-role).

export AWS_REGION="us-east-1"
export AWS_ACCESS_KEY_ID="AKIA..."
export AWS_SECRET_ACCESS_KEY="wJalr..."
# export AWS_SESSION_TOKEN="FwoGZX..."  # only for temporary/session credentials

Alternatively, configure credentials via ~/.aws/credentials and set AWS_REGION (or AWS_DEFAULT_REGION).

3. Run the load test

flotest run -c config.json

This runs the full pipeline: generate prompts → run benchmark → generate report.

Results are saved to ./results/<run-id>/ containing:

| File | Description | | ----------------------- | ------------------------------------------------------------------ | | summary.json | Aggregated metrics (latency, throughput, error rates, percentiles) | | run_log.jsonl | Per-request metrics streamed during the run | | prompts.jsonl | All generated prompts | | individual_responses/ | Full response data for each request | | config.resolved.json | Final merged configuration used |

Commands

| Command | Description | | ------------- | ------------------------------------------------------ | | run | Full pipeline: generate → bench → report (default) | | generate | Generate and save prompts only | | bench | Run benchmark using pre-generated prompts | | report | Generate report from existing benchmark results | | init [path] | Interactively create a config file |

flotest run -c config.json        # full pipeline
flotest generate -c config.json   # prompts only
flotest bench -c config.json      # benchmark only
flotest report -c config.json     # report only

CLI Options

Any config value can be overridden from the command line:

| Flag | Short | Description | | --------------------- | ----- | ------------------------------------------ | | --config <path> | -c | Path to config JSON (required) | | --run-id <id> | | Custom run ID (default: ISO timestamp) | | --model <name> | -m | Override provider.model | | --concurrency <n> | -n | Override benchmark.concurrency | | --max-requests <n> | | Override benchmark.maxRequests | | --max-duration <n> | | Override benchmark.maxDuration (seconds) | | --output-dir <path> | -o | Override benchmark.outputDir | | --base-url <url> | | Override provider.baseURL | | --streaming | | Enable streaming | | --no-streaming | | Disable streaming |

Example — override concurrency and model on the fly:

flotest run -c config.json -n 50 -m gpt-4o

Configuration Reference

The config file is JSON with five sections:

{
  "provider": {
    "adapter": "openai", // "openai" | "sagemaker"
    "model": "gpt-4o", // model identifier (required)
    "baseURL": "https://api.openai.com/v1", // API endpoint
    "systemPrompt": "You are a helpful assistant.", // optional system message
    "config": {}, // backend-specific options
  },
  "benchmark": {
    "concurrency": 10, // parallel requests (required)
    "inputTokens": { "mean": 512, "stddev": 51 }, // input token distribution
    "outputTokens": { "mean": 256, "stddev": 26 }, // output token distribution
    "maxRequests": 100, // total requests (required if no maxDuration)
    "maxDuration": 60, // duration in seconds (required if no maxRequests)
    "timeout": 600, // per-request timeout in seconds (default: 600)
    "streaming": true, // stream responses (default: true)
    "cachePercentage": 0, // % of requests reusing previous prompts (0–100)
    "outputDir": "./results", // results directory (default: ./results)
    "inputFile": "prompts.jsonl", // pre-generated prompts (for bench command)
    "rampUp": {
      // optional: gradually increase concurrency
      "duration": 30, //   over N seconds, or
      "requests": 50, //   over N requests
    },
    "rampDown": {
      // optional: gradually decrease concurrency
      "duration": 15,
    },
  },
  "generator": {
    "enabled": false, // use synthetic prompt generator
    "prompt": "Custom instruction...", // optional custom prompt template
    "corpus": "./my-corpus.txt", // optional custom corpus file
  },
  "http": {
    // optional: keep-alive connection pool shared by all requests
    "maxConnections": 10, // pool size per host (default: benchmark.concurrency)
    "keepAliveTimeout": 60, // seconds an idle connection stays open (default: 60)
    "connectTimeout": 10, // seconds to wait for a TCP connection (default: 10)
  },
  "reporter": {
    "adapters": ["json", "csv"], // export formats (default: ["json"])
  },
}

At least one of maxRequests or maxDuration is required. If stddev is omitted, it defaults to 10% of the mean.

SageMaker provider.config options

| Key | Type | Default | Description | | --------------- | --------------------------- | ---------- | ---------------------------------------------------------------------------------------------------------------------------------------- | | requestFormat | "openai" | "sagemaker" | "openai" | Controls request body format. "openai" sends messages array (modern LMI/vLLM). "sagemaker" sends raw inputs string (legacy TGI). |

When adapter is "sagemaker":

• provider.model is the SageMaker endpoint name (not a model ID).
• provider.baseURL overrides the default SageMaker runtime URL (https://runtime.sagemaker.<region>.amazonaws.com). Useful for VPC endpoints or custom domains.
• Streaming uses the /endpoints/<model>/invocations-response-stream path; non-streaming uses /endpoints/<model>/invocations.
• Requests are signed with AWS Signature V4 using the configured credentials.

Metrics Collected

Per-request

• TTFT — Time to first token (ms)
• TTFNT — Time to first non-thinking token (for reasoning models)
• E2E Latency — End-to-end request latency (ms)
• Inter-token latencies — Time between successive tokens (streaming)
• Output throughput — Tokens per second
• Input/output token counts
• Phase — ramp-up, steady, or ramp-down
• Cache hit — Whether the request was a cache hit
• Error details — Error message and code if failed

Summary

• Success/failure counts and error rate
• RPM (requests per minute) and overall TPS (tokens per second)
• Percentiles (p25, p50, p75, p90, p95, p99) for all latency and throughput metrics
• Error code frequency breakdown
• Phase-level breakdown (requests and error rates per phase)
• Cache hit rate

Examples

Load test an OpenAI-compatible endpoint

{
  "provider": {
    "adapter": "openai",
    "model": "gpt-4o",
    "baseURL": "https://api.openai.com/v1"
  },
  "benchmark": {
    "concurrency": 20,
    "inputTokens": { "mean": 256 },
    "outputTokens": { "mean": 128 },
    "maxRequests": 500,
    "streaming": true
  }
}

Load test a self-hosted model (vLLM, Ollama, etc.)

{
  "provider": {
    "adapter": "openai",
    "model": "meta-llama/Llama-3-8B",
    "baseURL": "http://localhost:8000/v1"
  },
  "benchmark": {
    "concurrency": 50,
    "inputTokens": { "mean": 512 },
    "outputTokens": { "mean": 256 },
    "maxDuration": 120,
    "streaming": true,
    "rampUp": { "duration": 30 },
    "rampDown": { "duration": 15 }
  }
}

Load test an AWS SageMaker endpoint (LMI / vLLM container)

Modern SageMaker LMI and vLLM containers accept the Chat Completions messages format automatically — no extra configuration needed. Set adapter to "sagemaker" and use your SageMaker endpoint name as model.

{
  "provider": {
    "adapter": "sagemaker",
    "model": "my-llama3-endpoint",
    "systemPrompt": "You are a helpful assistant."
  },
  "benchmark": {
    "concurrency": 20,
    "inputTokens": { "mean": 512 },
    "outputTokens": { "mean": 256 },
    "maxRequests": 200,
    "streaming": true
  }
}

# Set AWS credentials
export AWS_REGION="us-east-1"
export AWS_ACCESS_KEY_ID="AKIA..."
export AWS_SECRET_ACCESS_KEY="..."
# export AWS_SESSION_TOKEN="..."   # only needed for temporary credentials

flotest run -c config.json

The tool calls https://runtime.sagemaker.<region>.amazonaws.com/endpoints/<model>/invocations-response-stream for streaming or .../invocations for non-streaming, signing each request with AWS Signature V4.

SageMaker with legacy request format (TGI / HuggingFace containers)

Older containers that don't support the messages field expect a raw text string via the inputs field. Set requestFormat to "sagemaker" in provider.config:

{
  "provider": {
    "adapter": "sagemaker",
    "model": "my-tgi-endpoint",
    "config": {
      "requestFormat": "sagemaker"
    }
  },
  "benchmark": {
    "concurrency": 10,
    "inputTokens": { "mean": 256 },
    "outputTokens": { "mean": 128 },
    "maxRequests": 100,
    "streaming": false
  }
}

With the legacy format, the prompt is sent as { "inputs": "<prompt>", "parameters": { "max_new_tokens": N } }. No chat template is applied — format your prompts accordingly.

SageMaker with custom endpoint URL and ramp-up

If your SageMaker endpoint uses a custom domain or VPC endpoint, override baseURL:

{
  "provider": {
    "adapter": "sagemaker",
    "model": "my-vllm-endpoint",
    "baseURL": "https://vpce-0123456789abcdef-ab12cd34.runtime.sagemaker.us-west-2.vpce.amazonaws.com"
  },
  "benchmark": {
    "concurrency": 50,
    "inputTokens": { "mean": 1024 },
    "outputTokens": { "mean": 512 },
    "maxRequests": 1000,
    "streaming": true,
    "rampUp": { "duration": 60 },
    "rampDown": { "duration": 30 }
  },
  "reporter": {
    "adapters": ["json", "csv"]
  }
}

Time-bounded test with CSV output

{
  "provider": {
    "adapter": "openai",
    "model": "gpt-4o-mini",
    "baseURL": "https://api.openai.com/v1"
  },
  "benchmark": {
    "concurrency": 10,
    "inputTokens": { "mean": 128 },
    "outputTokens": { "mean": 64 },
    "maxDuration": 300,
    "streaming": true
  },
  "reporter": {
    "adapters": ["json", "csv"]
  }
}

License

MIT