vla.cpp
The open-source project I'm P.I.C of at VinRobotics - bringing Vision-Language-Action models to local hardware, in the spirit of llama.cpp for robotics.
View on GitHub ↗Khanh Nguyen
- Build efficient inference engines for neural nets across various devices.
- Optimize VLAs, VLMs, and LLMs for edge deployment.
- Interested in searching and studying efficient neural net architectures.
AI Software Engineer
Making capable AI models run fast on real hardware.
I work on inference engineering, model optimization, and edge AI for robotics - building lean C/C++ inference engines that bring VLMs, VLAs, and LLMs to CPUs, accelerators, and robots. Inspired by llama.cpp: a few lines of simple code, big results on your own hardware. The two projects I'm proudest of are vla.cpp and vla.simd - compare their measured latency across devices in the explorer. Two companion pages go deeper: VLA Hub benchmarks every model they serve on every device, and VLA Arch illustrates each model's architecture as block diagrams.
Latency
ms / call
Actions / s
Memory
- Image
- Views
- Action chunk
- Setup
Same model on other devices
Pinned projects
A few things I'm proud of - small in code, big in scope.
vla.simd
Efficient CPU inference for language-conditioned manipulation - shared SIMD micro-kernels (AVX2 / NEON) plus IMPACT, a language-conditioned policy that supplies 30+ actions per second on a Raspberry Pi 5.
View on GitHub ↗Model Optimization 101
Hands-on notebooks for learning model optimization for edge deployment - quantization, distillation, pruning, NAS and GPU kernels, each measuring size, latency and accuracy before vs. after.
View on GitHub ↗REX
Neural network Representation EXchange - a hands-on study of how different frameworks represent and serialize neural networks.
View on GitHub ↗FaceGen
Multiple Appropriate Facial Reaction Generation - synthesizing diverse listener reactions from a speaker's audio-visual cues. 3rd place at REACT 2024.
View on GitHub ↗Say hello
Always happy to talk edge AI, optimization, and robotics.
Games
Z next · X close
←↑↓→ or WASD to walk · Z to interact · X to close
Read the computer and the book in the lobby. Press Z at an arcade cabinet to play; the red button on the game window leads back here.
Left player
Score
0
Best
0
Lines
0
Level
1
Next
Keys
- ←→
- Move
- ↑X
- Rotate ↻
- Z
- Rotate ↺
- ↓
- Soft drop
- Space
- Hard drop
- P
- Pause
Keys
- AD
- Move
- W
- Rotate
- S
- Soft drop
- F
- Hard drop
- Space
- Pause
Keys
- ←→
- Move
- ↑
- Rotate
- ↓
- Soft drop
- Enter
- Hard drop
- Space
- Pause
Score
0
target 300
Hand
-
Pick up to 5 cards
Hands
4
Discards
3
Deck
44
Best
-
How to play
Pick up to five cards and play them as a poker hand. Its chips × mult is added to your score.
Reach the target within four hands. Discards swap cards without scoring.
After each blind, take a joker. Only the cards that make the hand score, and jokers act on those.
Time
0:00
Best
-
Done
0 / 0
How to play
Join all islands into one group. Each number is how many bridges touch that island.
Bridges run straight, never cross, and at most two join a pair.
Tap between two islands, or drag from one to the other, to add a bridge. Tap again for a second one, a third time to remove them.
You : CPU
0 : 0
Rally
0
Best rally
0
Keys
- WS
- Move
- ↑↓
- Move
- Space
- Pause
Keys
- WS
- Left player
- ↑↓
- Right player
- Space
- Pause
Score
0
Best
-
Pairs
0
Level
1 / 13
Still
Time
4:00
How to play
Tap two matching aliens to clear them. The line between them bends at most twice and only crosses empty space, the gap around the board included.
Clear every pair before the time runs out. Quick matches in a row score a combo.
Each level sets a gravity: after a match the other tiles stay put, fall, rise, slide, or split two ways.
Arrows and Z work too; H hints, P pauses.
Play the full game here
