Groq inference provider

[VladOS95-cyber]

Feature request

Since several inference provider APIs were added to HF. It would be beneficial to add Groq inference provider as well to expand a list of supported providers.

Motivation

Groq is a cutting-edge AI inference provider that has recently gained significant attention for its exceptional performance and speed. The company's Language Processing Unit (LPU) Inference Engine offers several key benefits and advantages:

1. Groq's unique approach to AI inference is based on several innovative features like SRAM onlz architecture, co-located compute and memory, kernel-less compiler.
2. Over 500,000 developers are currently using Groq's API keys.
3. It achieved up to 18x faster output tokens throughput compared to other cloud-based inference providers in the Anyscale LLMPerf Leaderboard.
4. Models support: Llama 3 series (8B, 70B, vision models), Mixtral 8x7B, Gemma 2 9B, Whisper Large V3 and so on.

Your contribution

I could try to dive deep into this topic and provide implementation myself.