A technical tool for calculating Time To First Token (TTFT), token generation speed, and VRAM footprints for Large Language Models on specific hardware.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install llm-perf-estimator
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install llm-perf-estimator using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is LLM Performance Estimator?
The LLM Performance Estimator is a sophisticated utility designed to provide developers and AI engineers with precise infrastructure requirements and performance benchmarks. By analyzing model architectures—including dense, MoE, and hybrid models—it calculates critical metrics like weight memory, KV cache size, and activation peaks. This allows users of Openclaw Skills to plan deployments effectively without needing to perform manual, error-prone calculations.
This skill bridges the gap between raw model weights and real-world hardware constraints. It accounts for various quantization formats (from FP16 to INT4) and incorporates hardware-specific variables like HBM bandwidth and TFLOPS. Whether you are sizing a consumer RTX 4090 or a data-center H100, this tool ensures you know exactly how your model will behave before you hit run.
LLM Performance Estimator Use Cases
Primary Scenarios
- Determining if a specific LLM will fit in available GPU VRAM at different quantization levels.
- Benchmarking expected latency (TTFT) for user-facing applications.
- Comparing the inference efficiency of MoE versus dense model architectures on specific hardware.
- Infrastructure planning and GPU procurement based on target tokens-per-second requirements.
- Debugging performance bottlenecks by comparing theoretical maximums with actual observed hardware speed.
How LLM Performance Estimator Works
- Resolve Model Architecture: The skill identifies model parameters via presets, a provided HuggingFace config.json, or manual user input.
- GPU Specification Mapping: Hardware capabilities (VRAM, Bandwidth, TFLOPS) are retrieved from a built-in database of RTX, A-series, H-series, and Apple Silicon chips.
- Quantization Analysis: The system selects the appropriate bytes-per-parameter and compute data type based on the requested format (e.g., INT4, FP8).
- VRAM Calculation: Weight memory, KV cache, and peak activation memory are calculated, including a 15% framework overhead buffer.
- Compute Latency Estimation: Prefill latency (TTFT) is estimated using Model FLOP Utilization (MFU) factors.
- Bandwidth-Bound Decoding: Token generation speed is calculated based on memory bandwidth utilization and bytes read per decode step.
- Reporting: A detailed Markdown report is generated, summarizing the feasibility and expected performance metrics.
LLM Performance Estimator Setup
To use this feature within your environment, you can invoke it directly via the CLI or through interactive prompts. This is a core part of the Openclaw Skills library.
# Example: Estimating Qwen 7B performance on an RTX 4090
/llm-perf-estimator Qwen2.5-7B RTX4090 2048 512 fp16
# Example: Interactive mode to define custom hardware
/llm-perf-estimator
LLM Performance Estimator Data Schema & Taxonomy
The skill organizes its analysis into a structured hierarchy to ensure accuracy across various LLM configurations:
| Data Layer | Details Included |
|---|---|
| Model Metadata | Total params, activated params, layers, hidden size, attention heads, and vocab size. |
| GPU Profile | VRAM capacity, BF16/FP8 TFLOPS, INT8 TOPS, and HBM Bandwidth (GB/s). |
| Quantization Map | Bytes/param mapping (e.g., 2.0 for FP16, 0.5 for INT4) and compute data types. |
| VRAM Breakdown | Sectioned data for Weights, KV Cache, Activations, and Framework Overhead. |
LLM Performance Estimator Advanced Features
Power-User Capabilities
- Hybrid Attention Support: Specifically handles models with mixed linear and full attention layers, adjusting KV cache calculations accordingly.
- MoE Specialization: Distinguishes between total and activated parameters for Mixture-of-Experts architectures to provide realistic memory vs. compute speed trade-offs.
- Custom Hardware Definitions: Allows users to input custom VRAM and bandwidth specs for unlisted or future hardware.
- Config Parsing: Automatically extracts technical specifications from HuggingFace or ModelScope config.json files.
- Proactive Optimization: Automatically suggests the minimum quantization level required if the current configuration exceeds available VRAM.
SKILL.md
Loading
Related Openclaw Skills
OpenClaw Model Switcher
A specialized WebUI and CLI utility for rapid switching and management of AI model providers within the OpenClaw ecosystem.
ProProfs Project Integration
A powerful AI agent skill for managing ProProfs Project workflows, including tasks, resources, and time tracking through the Membrane framework.
Rakuten Shop Analysis
A specialized AI agent skill for extracting and analyzing structured data from Rakuten storefronts using shop links or shop codes.
Test Upload Skill
A specialized diagnostic tool for validating the upload and publishing lifecycle of AI coding agent skills on the clawhub.ai platform.
Agent Briefing — Not For Humans
A zero-configuration skill for monitoring daily AI agent briefings and extracting structured product compatibility scores from the Not For Humans channel.
Seedance 2.0 Prompt Engineering to Video
A specialized tool for designing production-grade English prompts and generating high-control AI videos using the Seedance 2.0 model on WeryAI.
Featured*
