Matt Mastracci

@mmastrac

Also @mmastrac@hachyderm.io

following
followers

Jev-like DiffusionGemma patch reportedly hits 22 structured generations per second at concurrency 32

A user reports testing the patch on a DGX Spark with almost no tuning. With one request at a time, the reported median (p50) latency was about 0.21 seconds.

5 Sources, 31d ago, first seen 31d ago

DiffusionGemma-as-Jev reportedly runs near-real-time vision detection from a phone

A user says the system, also called djev, uses its native vision tower—the component that processes visual input.

3 Sources, 33d ago, first seen 33d ago

Jev vs. DiffusionGemma in live evaluations

A user tentatively favors DiffusionGemma after testing it against Jev, reporting a rough tie on intelligence and no speed advantage for Jev.

1 Source, 34d ago, first seen 34d ago

DiffusionGemma reportedly gets a 3–10x speedup by fixing structured response tokens in place

A developer says their diffgemma implementation reduces the work needed to generate answers and suggests the technique could probably also be used in vLLM.

1 Source, 35d ago, first seen 35d ago

Featured in

Jev-like DiffusionGemma patch reportedly hits 22 structured generations per second at concurrency 32

DiffusionGemma-as-Jev reportedly runs near-real-time vision detection from a phone

Jev vs. DiffusionGemma in live evaluations

DiffusionGemma reportedly gets a 3–10x speedup by fixing structured response tokens in place

Featured in

Jev-like DiffusionGemma patch reportedly hits 22 structured generations per second at concurrency 32

DiffusionGemma-as-Jev reportedly runs near-real-time vision detection from a phone

Jev vs. DiffusionGemma in live evaluations

DiffusionGemma reportedly gets a 3–10x speedup by fixing structured response tokens in place