Vulkan: a model runs end to end, token-exact — 16 native kernels, variant pipeline, two CI gates - #80

Merged

localai-bot merged 18 commits into

mainfrom

row/BACKEND-VULKAN-A1

Aug 7, 2026

Commits on Aug 6, 2026

plan(vulkan): VK-A1 shader-variant pipeline implementation plan
mudler
committed
77e7e13
build(vulkan): pin glslang 16.5.0 and prove the committed SPIR-V reproduces
mudler
committed
0b9f9c7
ci(vulkan): gate the committed SPIR-V against staleness
mudler
committed
0fececc
fix(vulkan): repair the stale op-surface gate and give Vulkan a CI leg
mudler
committed
36896ff
feat(vulkan): record specialization-constant IDs, and rule out the workgroup size
mudler
committed
9579f94
feat(vulkan): specialize pipelines, and make vt_cast's dtype pair the first axis
mudler
committed
ba5ea0c
refactor(vulkan): emit SPIR-V words into a .cpp, leaving declarations in the header
mudler
committed
196ea46
record(vulkan): VK-A1 landed, and the campaign's baseline was wrong
mudler
committed
feea8c7
feat(vulkan): native GEMM, both orientations (kMatmul + kMatmulBT)
mudler
committed
e32c5ed
feat(vulkan): native embedding gather and greedy argmax
mudler
committed
3bfa1f1
record(vulkan): 12 native kernels, 75 still on the reference tier
mudler
committed
03f4d3c
feat(vulkan): native block-paged attention (kPagedAttention)
mudler
committed
34a3efe
record(vulkan): 13 native kernels, and AMD testing answered NO on this box
mudler
committed
b1b68c1
feat(vulkan): native KV-cache write (kReshapeAndCache)
mudler
committed
2c86f79
record(vulkan): 14 native kernels, and the RoPE decision that comes before more work
mudler
committed
d9ca9d9
feat(vulkan): a model runs END TO END on Vulkan, token-exact vs the vLLM oracle
mudler
committed
f4738bb
record(vulkan): a model runs e2e on Vulkan, token-exact -- 16 native kernels
mudler
committed
0be5d45

Commits on Aug 7, 2026

Merge origin/main into row/BACKEND-VULKAN-A1
mudler
committed
461ce6b

Navigation Menu

Vulkan: a model runs end to end, token-exact — 16 native kernels, variant pipeline, two CI gates - #80

Vulkan: a model runs end to end, token-exact — 16 native kernels, variant pipeline, two CI gates#80
localai-bot merged 18 commits into
mainfrom
row/BACKEND-VULKAN-A1

Commits on Aug 6, 2026

plan(vulkan): VK-A1 shader-variant pipeline implementation plan

build(vulkan): pin glslang 16.5.0 and prove the committed SPIR-V reproduces

ci(vulkan): gate the committed SPIR-V against staleness

fix(vulkan): repair the stale op-surface gate and give Vulkan a CI leg

feat(vulkan): record specialization-constant IDs, and rule out the workgroup size

feat(vulkan): specialize pipelines, and make vt_cast's dtype pair the first axis

refactor(vulkan): emit SPIR-V words into a .cpp, leaving declarations in the header

record(vulkan): VK-A1 landed, and the campaign's baseline was wrong

feat(vulkan): native GEMM, both orientations (kMatmul + kMatmulBT)

feat(vulkan): native embedding gather and greedy argmax

record(vulkan): 12 native kernels, 75 still on the reference tier

feat(vulkan): native block-paged attention (kPagedAttention)

record(vulkan): 13 native kernels, and AMD testing answered NO on this box

feat(vulkan): native KV-cache write (kReshapeAndCache)

record(vulkan): 14 native kernels, and the RoPE decision that comes before more work

feat(vulkan): a model runs END TO END on Vulkan, token-exact vs the vLLM oracle

record(vulkan): a model runs e2e on Vulkan, token-exact -- 16 native kernels

Commits on Aug 7, 2026

Merge origin/main into row/BACKEND-VULKAN-A1