Carry ggml-org#24423 (DiffusionGemma) onto b10775, and make the arch testable - #177

Closed

danielhanchen wants to merge 30 commits into

base/upstream-de8656bd9from

diffusiongemma-24423-b10775

Commits on Jun 10, 2026

diffusion-visual updates
danielhanchen
committed
c5fe75b
diffusion: fix Windows build, skip diffusion-gemma in test-llama-archs, drop debug hooks
danielhanchen
committed
c84e85a

Commits on Jun 11, 2026

diffusion-cli: note that --fit is not applied by the diffusion runner
danielhanchen
committed
7c200dd
diffusion-cli: honor -ot / --n-cpu-moe (copy tensor_buft_overrides into model params)
danielhanchen
committed
9b4beb7
diffusion: device-resident self-conditioning + cli throughput, /help, /clear
danielhanchen
committed
15ad8f4
diffusion: device-side sampling reductions (default on, ~1.4x/step)
danielhanchen
committed
d6cf0b2
ggml-cuda: fix diffusion sampler build on HIP/MUSA
danielhanchen
committed
53752ad
diffusion: add gemma visual server for streaming canvas frames
danielhanchen
committed
f53de16
diffusion: self-tokenize in gemma visual server
danielhanchen
committed
7ea238c
diffusion: stop per-step device-sample fallback spam
danielhanchen
committed
10a2613

Commits on Jun 12, 2026

Commits on Jun 13, 2026

diffusion-gemma-visual-server: emit channel markers + low-VRAM auto-size fallback
danielhanchen
committed
49fc372

Commits on Jun 14, 2026

diffusion-gemma-visual-server: size context by RAM when the model spills off the GPU
danielhanchen
committed
9b4dae8

Commits on Jun 17, 2026

diffusion-gemma: chunked causal prefill + cap encode outputs to the canvas
danielhanchen
committed
ef5e2dc

Commits on Jun 23, 2026

Commits on Jul 3, 2026

Fix merge conflicts
danielhanchen
committed
1d2faac

Commits on Jul 4, 2026

Add tool calling
danielhanchen
committed
c3fb972

Commits on Aug 10, 2026

Fix merge conflicts
danielhanchen
committed
1d87a1b
Fix merge conflicts
danielhanchen
committed
daca807

Commits on Aug 26, 2026

Merge b10630 into the DiffusionGemma pin
danielhanchen
committed
a15ca5b
diffusion-gemma : port the visual server to common_json
danielhanchen
committed
74acc40

Commits on Sep 3, 2026

Merge b10775 into the DiffusionGemma carry
danielhanchen
committed
ba18cd5
diffusion-gemma: give canvas_length an llm_kv id, and a test fixture
danielhanchen
committed
6851db7
diffusion-gemma: keep the fixture out of save_models
danielhanchen
committed
47073df
Merge b10786: n_ff_exp is per-layer now
danielhanchen
committed
ea0cdfd

Commits on Sep 4, 2026

diffusion-gemma: run the arch, the fixture was the blocker
danielhanchen
committed
4137a7d