Tag

#diffusion-models

Posts tagged with diffusion-models

2 posts

DiffusionGemma builds entire responses through parallel denoising instead of one token at a time, making full-response latency the metric that matters for serving economics.

Engineering • Nov 06, 2026 •11 min read

DiffusionGemma Makes Full Response Latency the Metric

DiffusionGemma teardown: parallel decoding trades first-token latency for full-response latency, drops the KV cache, and breaks streaming and tool calls.

#diffusiongemma #diffusion-models +3

A single-GPU workstation running a from-scratch diffusion model training job, the core scenario behind this cost and throughput breakdown.

Engineering • Sep 11, 2026 •11 min read

Training a Diffusion Model From Scratch in 3.5 Days

Training a diffusion model from scratch can cost hundreds, not millions. The full math behind a 210M DiT built in 3.5 days on one RTX PRO 6000.

#diffusion-models #training-economics +3