Skip to content

goniz/mlx-vulkan

Repository files navigation

mlx-vulkan

Home for the Development of MLX Vulkan backend

Benchmark Results

CI benchmark history from AMD Radeon 8060S (Strix Halo). Detailed data is in benchmarks/results.csv.

Qwen3-0.6B Prompt Throughput

Qwen3-0.6B prompt TPS

Qwen3-0.6B Generation Throughput

Qwen3-0.6B generation TPS

Qwen3.6-35B-A3B Prompt Throughput

Qwen3.6-35B-A3B prompt TPS

Qwen3.6-35B-A3B Generation Throughput

Qwen3.6-35B-A3B generation TPS

Latest Results

Model Bits Prompt TPS Generation TPS Peak memory (GB) mlx-vulkan mlx Run
mlx-community/Qwen3-0.6B-8bit 8bit 1360.581 87.609 2.056 c0c3da7 5d618c8 run
mlx-community/Qwen3-0.6B-bf16 bf16 2410.127 65.748 2.614 c0c3da7 5d618c8 run
mlx-community/Qwen3.6-35B-A3B-8bit 8bit 121.011 21.135 40.350 c0c3da7 5d618c8 run

Model Generation Report

Serial generation smoke tests validate that each model produces coherent output on Vulkan.

Model Output Coherent Peak memory (GB) Sample Error
mlx-community/Qwen3-0.6B-bf16 pass pass 1.148 Okay, the user wants a concise sentence about why Vulkan acceleration is useful. Let...
mlx-community/Qwen3-0.6B-8bit pass pass 1.032 Okay, the user wants a concise sentence about why Vulkan acceleration is useful. Let...
LiquidAI/LFM2.5-1.2B-Instruct-MLX-8bit pass pass 1.396 Vulkan acceleration enhances performance by enabling efficient parallel processing and reduci...
mlx-community/Qwen3.5-2B-bf16 pass pass 4.529 Thinking Process: 1. Analyze the Request: * Task: Write one concise sentence. * Topic: Wh...
mlx-community/gemma-4-e2b-it-bf16 pass pass 10.005 <|channel>thought 1. Analyze the Request: The user wants a concise sentence explaining...
mlx-community/gemma-4-e4b-it-4bit pass pass 5.19 <|channel>thought 1. Analyze the request: The user wants one concise sentence explainin...
mlx-community/gemma-4-26b-a4b-it-4bit pass pass 14.092 <|channel>thought * Topic: Why Vulkan acceleration is useful. * Constraint: One concise sente...
mlx-community/Qwen3.6-35B-A3B-8bit pass pass 35.819 Here's a thinking process: 1. Analyze User Input: - Topic: Vulkan acceleration - **Re...
mlx-community/gpt-oss-20b-MXFP4-Q8 pass pass 13.689 <|channel|>analysis<|message|>We need to write one concise sentence about why Vulkan accelera...
mlx-community/Qwen3.6-27B-8bit pass pass 28.461 Here's a thinking process: 1. Analyze User Input: - Topic: Vulkan acceleration - **Re...

About

Home for the Development of MLX Vulkan backend

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Sponsor this project

 

Packages

 
 
 

Contributors