Know what your fleet can handle before you find out the hard way
SlickToken simulates vLLM, SGLang, and TRT-LLM under realistic load — modeling KV cache pressure, queue dynamics, and SLA risk — so you can make hardware decisions with data instead of guesswork.
Seven planning tools. One question each.
Each tool answers a specific planning question. Run a simulation in under a minute — no live cluster required.
A100 vs H100 at my traffic level
Configure two fleets and compare throughput, KV utilization, SLA violation rate, and monthly cost side by side.
What's the probability we get overwhelmed?
Monte Carlo simulation across 50 independent traffic scenarios. Returns green / yellow / red with breach probabilities per metric.
What breaks if I upgrade to Llama 405B?
Compare any configuration change — model upgrade, GPU swap, engine switch, precision change, traffic spike — and see the exact delta before deploying.
When do I need more GPUs?
Project your fleet 3–24 months forward under exponential, linear, or S-curve growth. Flags the exact period where SLA violations begin.
How much does routing to a small model save?
Route different request classes to different models. See cost savings vs single-model serving, per-class latency, and bottleneck identification.
What's the true cost of my RAG pipeline?
Define multi-step pipelines with per-step failure rates. Reports wasted token spend from failed chains and cost per successful completion.
Running in production in under five minutes
SlickToken ships as a Docker image. No external services. No outbound connections. Runs fully offline after image pull.
Install your license key (This Section is not relevant for this release)
License comes preinstalled with Docker container. Expires July 2027.
You received a license key by email when you signed up. Install it before starting the container.
mkdir -p ~/.slicktoken
echo "YOUR_LICENSE_KEY" > ~/.slicktoken/license.key
chmod 600 ~/.slicktoken/license.key
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.slicktoken"
"YOUR_LICENSE_KEY" | Out-File "$env:USERPROFILE\.slicktoken\license.key" -Encoding ascii
export SLICKTOKEN_LICENSE_PATH=/etc/slicktoken/license.key
Start the container
Mount your license directory as read-only and expose port 8081.
docker run -p 8081:8081 slicktoken
Open the planning dashboard
The simulator starts immediately. No database setup, no additional configuration.
start http://localhost:8081/plan # Windows
xdg-open http://localhost:8081/plan # Linux
Free for one year, machine-independent
Your license is valid for 12 months from the issue date. It works on any number of machines — no hardware binding.
Full simulator access, all seven tools
Fleet comparison, risk assessment, deployment changes, capacity growth, model routing, workflow chains, and agentic presets — all included at no cost for the first year.
30-day warning, no surprise cutoff
The agent logs a warning 30 days before expiry. Submit an inquiry in our help section to renew. The expiry date cannot be extended without a new key.
Running a simulation
All seven tools follow the same pattern: configure a fleet, describe your workload, run. Results appear in 1–15 seconds depending on simulation type.
| Tool | Key inputs | Key outputs | Runtime |
|---|---|---|---|
| Fleet Comparison | Two GPU fleets, arrival rate, engine, SLA deadline | Throughput, KV utilization, SLA %, monthly cost, recommendations | ~1s |
| Risk Assessment | Fleet, arrival rate, trial count (20–100), thresholds | P(SLA breach), P(KV exhaustion), risk classification, distribution p10–p99 | ~5–15s |
| Deployment Changes | Baseline fleet, change type and parameters | Delta table per metric, auto-generated warnings, recommendation | ~2s |
| Capacity Growth | Fleet, growth rate, growth model, horizon (3–24 periods) | Per-period metrics table, breach period, GPUs needed at horizon, monthly cost trajectory | ~3–8s |
| Model Routing | Workload mix with per-class model assignments, fleet per model | Per-model throughput and latency, cost savings vs single-model, bottleneck identification | ~2s |
| Workflow Chains | Step definitions with failure rates and modes, arrival rate | Chain success rate, wasted token %, cost per successful chain, per-step failure breakdown | ~3s |
| Agentic Presets | Preset type, fleet assignment, arrival rate | E2E latency p50/p95/p99, bottleneck model, per-path SLA violations, cost per request | ~3s |
Saving and sharing results
Downloads the current fleet configuration as a JSON file. Reload it on any machine to restore your setup exactly.
Upload a previously saved fleet JSON. Validates all values before populating — malformed files are rejected safely.
Generates a self-contained HTML file of the current result. Viewable offline and shareable without the server running.
What the numbers mean
SlickToken models queue dynamics, KV memory pressure, and scheduling behavior — not GPU microarchitecture. Throughput figures are calibrated against published benchmarks.
Benchmark calibration
⁽¹⁾ Small model throughput is intentionally conservative — see Disclaimers. Ratio = simulated ÷ published real-world benchmark value.
Confidence by output type
Every result includes a calibration note for your specific GPU type.
Before you make decisions
SlickToken is a planning tool. Read these before using simulation results to justify hardware purchases or SLA commitments.
⚠ Not a replacement for load testing
- Simulation results predict behavioral trends, not exact production metrics. Validate with real load tests against your actual models and traffic patterns before provisioning hardware.
- The simulator does not model CUDA kernel launch latency, CPU scheduling overhead, network I/O between tensor parallel workers, or NUMA effects on multi-GPU systems.
- Results are most reliable for relative comparisons (Fleet A vs Fleet B) and least reliable for predicting absolute throughput at a specific request rate.
⚠ Throughput calibration scope
- Throughput figures are calibrated for 70B-class models (Llama 3.1 70B, Qwen 72B, DeepSeek V3) on H100 SXM5 hardware. Accuracy degrades for smaller models, older hardware, and non-standard configurations.
- Small models (7B–13B) may achieve 2–4× higher real throughput than the simulator shows because calibration multipliers do not account for smaller parameter counts enabling larger effective batch sizes.
- If you run the simulator against real hardware, sharing your calibration ratios in our feedback section improves accuracy for everyone in a future release.
⚠ SLA predictions are probabilistic
- The Risk Assessment tool reports probability distributions, not guarantees. A "green" classification means the simulated probability of SLA breach is below 5% — not that breach is impossible.
- Monte Carlo trials use a reproducible random seed sequence. Results are consistent but are not a substitute for observing actual traffic variance in production over time.
- Traffic burst events such as viral spikes, marketing launches, or incident replay may exceed arrival rates tested in simulation. Model your expected peak-to-average ratio explicitly using the Deployment Changes tool.
⚠ No data persistence between sessions
- The simulator stores no state between sessions. Fleet configurations, simulation results, and custom workflow definitions are lost when the browser tab is closed unless exported using the Export Report or Save Fleet buttons.
- Monitoring agent is not yet available. Disregard follow line.
- Monitoring agent incident data is stored at
~/.slicktoken/incidents.db. Back this directory up if it matters.
General
- SlickToken is provided as-is during the free tier period. It is not covered by an SLA and may change between versions.
- Cost estimates ($/hr, monthly cost) use configurable reference prices for illustrative purposes. Verify actual pricing with your cloud provider or hardware vendor.
- Model KV capacity figures are derived from published architecture specifications. Actual memory consumption may vary based on quantization implementation, framework version, and activation memory overhead not included in KV calculations.