dflash: enable Qwen3.6-27B, Qwen3.6-35B-A3B, Qwen3.5-122B-A10B, Qwen3-Coder-Next, Gemma 4 31B on Vulkan#79

Closed

gboddaer wants to merge 16 commits into

Anbeeld:mainfrom

gboddaer:enable-dflash-qwen3-coder-next-on-vulkan

Commits on Jun 21, 2026

dflash: enable Qwen3-Coder-Next on Vulkan
gboddaer
committed

Commits on Jun 23, 2026

dflash: sync Vulkan DFlash code from vulkan-dflash-cross-ring (ceb4f697d)
Gert Boddaert
committed
dflash: keep n_rs_seq for QWEN35MOE (fix 122B GPU cross-ring corruption)
Gert Boddaert
committed
dflash: rotate DeltaNet RS snapshots for n_seq_tokens<K (fix stochastic-sampling corruption)
Gert Boddaert
committed
docs: rename + expand DFlash-on-Vulkan doc to cover various models
Gert Boddaert
committed
docs: refresh VULKAN_BUILD_VERIFICATION for current PR head (efb07137c)
Gert Boddaert
committed
tests: add DFlash regression guards; extract RS write-back slot helper
Gert Boddaert
committed

Commits on Jun 25, 2026

dflash: slide drafter KV without the GPU cross-ring (fix CPU-path decode fails)
gboddaer
committed

Commits on Jun 29, 2026

server: extend loop guard to visible output region
gboddaer
committed

Commits on Jun 30, 2026

docs: document Gemma 4 Vulkan DFlash FA corruption root cause and workaround
Gert Boddaert
committed
docs: update enable-and-fix-dflash-on-vulkan-various-models with latest status
Gert Boddaert
committed