What I load in ComfyUI, and which models do true instruction edits vs. needing img2img. Includes the encoders, VAEs and ControlNets you need.
🔄 In a Training Loop
Alex Jones
axjns
AI & ML interests
I build AI/ML infrastructure
Recent Activity
liked a dataset about 4 hours ago
axjns/llmfit-benchmarks updated a dataset about 4 hours ago
axjns/llmfit-benchmarks published a dataset about 4 hours ago
axjns/llmfit-benchmarksOrganizations
Inference cost decides your architecture. These papers, read top to bottom, build the mental model: where the FLOPs and the bytes actually go.
-
Efficiently Scaling Transformer Inference
Paper • 2211.05102 • Published • 4 -
Scaling Laws for Neural Language Models
Paper • 2001.08361 • Published • 11 -
Training Compute-Optimal Large Language Models
Paper • 2203.15556 • Published • 13 -
Fast Transformer Decoding: One Write-Head is All You Need
Paper • 1911.02150 • Published • 9
What I load in ComfyUI, and which models do true instruction edits vs. needing img2img. Includes the encoders, VAEs and ControlNets you need.
Everything I run on a 64GiB unified-memory Strix Halo box, with my own measured numbers. The finding: MoE decodes ~6x faster than dense here.
Inference cost decides your architecture. These papers, read top to bottom, build the mental model: where the FLOPs and the bytes actually go.
-
Efficiently Scaling Transformer Inference
Paper • 2211.05102 • Published • 4 -
Scaling Laws for Neural Language Models
Paper • 2001.08361 • Published • 11 -
Training Compute-Optimal Large Language Models
Paper • 2203.15556 • Published • 13 -
Fast Transformer Decoding: One Write-Head is All You Need
Paper • 1911.02150 • Published • 9