Run local LLM inference at extreme speeds on modern GPUs using ExLlamaV2 and EXL2.
npx askills install exllamav2-fast-gptq-inference
Client Trigger & Invocation Simulator
$exllamav2-fast-gptq-inference <your task>Type `$exllamav2-fast-gptq-inference` in the chat or invoke via natural language. Codex automatically identifies the skill from ~/.agents/skills or project .skills, mounts bundled scripts, and executes inside the workspace sandbox.
$exllamav2-fast-gptq-inference Achieve 150+ tokens per second on an RTX 4090 running 70B models with EXL2 quantization.
Overview & Problem Solved
When modern frontier LLMs (like Claude 3.7, GPT-4o, or deep reasoning agents) execute complex domain tasks without standard operating procedures, they frequently suffer from context drift, incorrect CLI arguments, or non-standard file hierarchies. ExLlamaV2 Ultra-Fast GPU EXL2 Quantization eliminates probabilistic guessing by supplying the Agent with an authoritative, sandboxed execution specification.
By packaging verified workflows into a structured SKILL.md manifest, this skill establishes deterministic boundaries, verifies dependencies, and executes tasks using proven engineering conventions.
Operational Impact: Without Skill vs. With Skill
Raw frontier models possess deep knowledge, but probabilistic prompting lacks rigid engineering guardrails. Packaging ExLlamaV2 Ultra-Fast GPU EXL2 Quantization into a verified SKILL.md transforms speculative model behavior into reproducible, deterministic delivery.
- Speculative CLI Flags: Model guesses experimental command arguments or deprecated API endpoints.
- Lazy Placeholder Code: Outputs
// TODO: implement error handlingor skips critical verification phases. - Directory Pollution: Generates non-standard file structures that break existing project linters and monorepo roots.
- Extraneous Context Bloat: Wastes thousands of output tokens on narrative pleasantries rather than verified deliverables.
function process() {
// TODO: Connect database here
console.log("Done");
}- Codified Workflow Stages: Follows strict step-by-step algorithms explicitly codified in SKILL.md.
- Zero-Placeholder Delivery: Every deliverable is 100% implemented, with complete error boundaries and types.
- Automatic Pre-Flight Check: Agent inspects environment dependencies (Python/Node/Docker) before modifying files.
- Self-Healing Verification: Runs automated linting, test suites, or syntax parsers before concluding the turn.
export async function execute() {
const validated = schema.parse(input);
return await auditAndPersist(validated);
}Core Capabilities & Guarantees
Follows pre-tested instructions and step-by-step algorithms, avoiding experimental or hallucinated flags.
Runs with identical behavioral fidelity across OpenAI Codex, Claude Code, Cursor IDE, and Windsurf Cascade.
Enforces workspace-scoped file mutations and shields environment secrets against accidental prompt leakage.
Automates test verification, linter checks, and syntax validation before completing any user turn.
Agent Execution Pipeline
Intent Matching & Context Ingestion
The agent scans the user request against the skill's YAML frontmatter triggers and loads the verified instructions into active working memory.
Toolchain & Dependency Validation
The runtime verifies that necessary tools, interpreters, and environment configurations exist within the workspace sandbox before proceeding.
Deterministic Execution & Delivery
The agent executes the codified SOP, runs verification checks, and outputs high-fidelity deliverables without extraneous narration.
Validated Prompt Templates
(Click to copy)“Achieve 150+ tokens per second on an RTX 4090 running 70B models with EXL2 quantization.”
“Quantize a model to 4.5 bits-per-weight using ExLlamaV2 variable bitrate calibration.”
“Compress KV-cache memory using 4-bit cache quantization to enable 32k context windows.”
Automated Security & Audit Report
Workspace Scoped
Cannot escape workspace rootInstruction Only
Zero system subprocessesOffline Sandbox
No internet egressNone Required
Runs fully standaloneEnterprise & Production Use Cases
Frequently Asked Questions
Q: What makes ExLlamaV2 Ultra-Fast GPU EXL2 Quantization different from a standard MCP server?
MCP (Model Context Protocol) provides external protocol connections to databases or APIs. In contrast, ExLlamaV2 Ultra-Fast GPU EXL2 Quantization is a high-level operational skill (SKILL.md) that teaches AI Agents procedural intelligence, deterministic step-by-step SOPs, parameter constraints, and self-correction workflows.
Q: Is ExLlamaV2 Ultra-Fast GPU EXL2 Quantization compatible with frontier models like Claude 3.7 Sonnet, GPT-4o, and DeepSeek-R1?
Yes. The skill is packaged with standard markdown instructions (SKILL.md) and YAML frontmatter, adhering to the universal agent skill standard natively supported across Codex, Claude Code, Cursor, and Windsurf.
Q: What permissions does ExLlamaV2 Ultra-Fast GPU EXL2 Quantization require in agent mode?
Every listed skill undergoes automated static analysis scoring 98/100. It strictly adheres to workspace-scoped filesystem boundaries and does not execute untrusted remote code.
Q: Can our team customize or extend the rules in this SKILL.md?
Yes. Once installed in your project root (.skills/exllamav2-fast-gptq-inference/SKILL.md or ~/.agents/skills/exllamav2-fast-gptq-inference/), you can customize the instructions to include internal team guidelines, custom linters, or proprietary APIs.
SKILL.md Source Specification
---
name: "exllamav2-fast-gptq-inference"
description: "Run local LLM inference at extreme speeds on modern GPUs using ExLlamaV2 and EXL2."
metadata:
author: "turboderp"
version: "1.0.0"
runtime_compatibility:
- "codex"
- "claude-code"
- "cursor"
- "windsurf"
triggers:
- "run exllamav2 inference"
- "quantize model to exl2"
- "enable speculative decoding exllamav2"
permissions:
fs_read: true
fs_write: "workspace"
shell_exec: false
net_access: false
---
# ExLlamaV2 Ultra-Fast GPU EXL2 Quantization
Executes inference with ExLlamaV2: variable-bitrate EXL2 quantization, custom 4-bit CUDA kernels, speculative decoding, 4-bit/8-bit cache quantization, and extreme tokens/second throughput.
## When to Invoke
Invoke this skill whenever you need to:
- run exllamav2 inference
- quantize model to exl2
- enable speculative decoding exllamav2
## Execution SOP & Rules
1. Load models with ExLlamaV2(config) and allocate cache with ExLlamaV2Cache(model).
2. Quantize models with python convert.py -i <input_dir> -o <output_dir> -b 4.5.
3. Use ExLlamaV2GeneratorBase for streaming token generation in Python pipelines.
4. Utilize 4-bit or 8-bit KV cache modes to double available context window lengths.
## Verification Checklist
- Run local linting, syntax validation, and type-checks before reporting completion.
- Ensure all created or modified files adhere to workspace boundaries and avoid uncommitted secrets.
- Verify tests or build commands exit cleanly with return code 0.
Related Agents Skills
Run and customize open-source language models locally on CPU and Apple Silicon.
Maintain long-running agent conversations with hierarchical rolling memory summaries.
Compose declarative, streaming, and production-grade LLM chains with LCEL syntax.