Vinh H. Pham

Open-source AI engineer in Ho Chi Minh City, working on diffusion/video generation, 4-bit diffusion inference, LLM/VLM quantization, model serving, and practical ML infrastructure.

GitHub | Hugging Face | Website

Diffusion · Video generation · SVDQuant / 4-bit inference · LLM/VLM quantization · Model serving · Open-source ML

Open-source impact

Project	✓ Merged PRs	Representative merged work
huggingface/diffusers ★ 34,623 / forks 7,356	24	Nunchaku Lite quantization backend, LTX2 distilled checkpoint support, framewise LTX Video VAE encoding/decoding
huggingface/transformers ★ 166,749 / forks 34,699	8	PoolFormer fast image processor, BridgeTower fast image processor
huggingface/blog	1	Bringing Nunchaku 4-bit Diffusion Inference to Diffusers, co-authored with the Diffusers team
sgl-project/sglang ★ 36,508 / forks 9,189	1	Z-Image text encoder config fix
d2l-ai/d2l-vi ★ 663 / forks 254	108	Vietnamese ML education translation/revision
mlbvn/ml-yearning-vi ★ 1,234 / forks 395	23	Vietnamese ML education translation/revision
d2l-ai/d2l-en ★ 29,724 / forks 5,140	7	Documentation fixes

Selected upstream work includes the Nunchaku Lite 4-bit quantization backend in Diffusers, LTX2 distilled checkpoint support, LTX Video VAE framewise encoding/decoding, Diffusers pipeline/test improvements, a Z-Image fix in SGLang, and Vietnamese ML education work across Dive into Deep Learning and Machine Learning Yearning.

Projects

Project	What it does
nunchaku-lite	Lean runtime for SVDQuant W4A4 diffusion models in standard Diffusers pipelines, with native CUDA kernels published as nunchaku-lite-kernels on the Hub and torch.compile support
diffuse-compressor	Model-agnostic SVDQuant toolkit that calibrates and quantizes diffusion transformers to INT4 / NVFP4 and exports Nunchaku-compatible checkpoints
lite-infer	Hub organization with 30 prequantized Nunchaku Lite checkpoints: FLUX.1 / FLUX.2 Klein, Qwen-Image and Qwen-Image-Edit, Z-Image-Turbo, ERNIE-Image-Turbo, Krea-2-Turbo, and LTX-2.3 Distilled
diffuser_layerdiffuse	Transparent image generation with Diffusers (LayerDiffuse)

Adopted work

Project	Where it is used	Adopting project scale
rootonchair/nunchaku-lite	Diffusers: official quantization backend with a dedicated docs page SD.Next: built-in Nunchaku-Lite inference engine that installs the package and ships 11 lite-infer checkpoints in its model reference	★ 34,623 / forks 7,356 ★ 7,346 / forks 585
rootonchair/LTX-2-19b-distilled	Listed in vLLM Omni's supported-model table for LTX2DistilledOneStagePipeline and LTX2DistilledTwoStagePipeline	★ 7,101 / forks 1,802
rootonchair/diffuser_layerdiffuse	SD.Next includes a LayerDiffuse: Transparent Image extension and links back to this project in the extension UI	★ 7,346 / forks 585

Selected AI systems work

• Video: Diffusers integrations, LTX video support, LTX-2.3 Distilled Diffusers conversions, transparent image generation, and model execution workflows.
• 4-bit diffusion: SVDQuant INT4 / NVFP4 pipelines for image and video transformers, including MiniMax-H3, Ideogram v4, and ERNIE-Image-Turbo.
• VLM quantization: GGUF and AWQ releases for image-text models such as Lens-Turbo, Vintern-3B, Vintern-1B, EraX-VL-7B, and InternVL2.5-4B.
• Serving: quantized/GGUF model artifacts, runtime tooling, llama.cpp-based OCR serving, and diffusion model compression experiments.
• Tooling: practical patches across Hugging Face Diffusers/Transformers and related open-source ML projects.

Current focus

I am mostly interested in making generative models easier to run, adapt, compress, and serve in real systems.