Add diffusion-gemma block-diffusion support - #1

Open

lnigam wants to merge 27 commits into

diffusion-gemma-review

Commits on Jun 9, 2026

convert: add diffusion_gemma4 (block-diffusion Gemma 4) conversion support
lnigam
committed
5bddf43
model: register diffusion-gemma4 arch (reuses gemma4 graph)
lnigam
committed
e814e44
model: diffusion-gemma4 bidirectional graph + self-conditioning input
lnigam
committed
aac0349
model: wire self-conditioning MLP into the diffusion-gemma4 graph
lnigam
committed
08e6dfa
examples: block-diffusion generation CLI for diffusion-gemma4
lnigam
committed
27a4fc5
diffusion-gemma4: self-conditioning input channel (feedback)
lnigam
committed
0804f74
diffusion-gemma4: prompt conditioning via prefix attention
lnigam
committed
2c40f4a
diffusion-gemma4: apply the chat template to the prompt
lnigam
committed
d7dc3ea
diffusion-gemma4: fix prompt conditioning (keep prompt tokens active)
lnigam
committed
724ba77
diffusion-gemma4: greedy read-out of the final canvas
lnigam
committed
212a698
diffusion-gemma4: extract the clean final answer
lnigam
committed
0daa231
diffusion-gemma4: offload to GPU by default
lnigam
committed
27c35eb
diffusion_gemma4: KV-cache reuse for block-diffusion generation
lnigam
committed
03b9c2b
diffusion-gemma (v7): rename arch + multimodal (gemma4 vision) support
lnigam
committed
868ef51
diffusion-gemma cli: -n controls block count, generation timing, inline-argmax output
lnigam
committed
9c53a9d
diffusion-gemma: place precomputed transposed embedding on the offloaded backend
lnigam
committed
5ffc5f4
diffusion-gemma cli: report encoder-phase prefill timing
lnigam
committed
ea2e26a
diffusion-gemma: top-k host sampling (top-k softmax/entropy/self-cond) with k-annealing
lnigam
committed
459e35e
diffusion-gemma: sparse top-k self-conditioning via on-device F16 gather
lnigam
committed
7e33c60
diffusion-gemma: OpenAI-compatible HTTP server (llama-server analogue)
lnigam
committed
5cc7006
diffusion-gemma: add CUDA graph GPU sampling server path
lnigam
committed
ff726a2
diffusion-gemma: keep self-cond sampling on device
lnigam
committed
378fbfd
diffusion-gemma: keep denoising loop on device
lnigam
committed
f7cea9c
diffusion-gemma: use output flag for persistent inputs
lnigam
committed
7e65661
diffusion-gemma: checkpoint device-loop early stop
lnigam
committed
2b7758f
diffusion-gemma: make device early stop every step
lnigam
committed
3b51143
examples: update diffusion gemma mtmd bitmap loading
lnigam
committed
e1fc535