dasLLAMA: native f16 KV mirrors on reified kernel templates - the vulkan resident driver goes codec-driven - #3681

Merged

borisbat merged 20 commits into

masterfrom

bbatkin/cm2-prefill-gemm

Aug 10, 2026

Commits on Aug 10, 2026

dasSpirv: l-tile wg shapes + [spirv_decode] class-method form

borisbat
and
claude
committed
2ac9bb8
dasLLAMA: the cm2 l-tile kernel on native fmt-0 planes

borisbat
and
claude
committed
e33c332
dasLLAMA: mode 4 serves native fmt-0 through the l-tile

borisbat
and
claude
committed
1d4fcf7
dasLLAMA: the l-tile fast/edge path split - interior tiles run unclamped

borisbat
and
claude
committed
f489f09
dasLLAMA: cm2 m-tile + split-k selection for underfilled prefill GEMMs

borisbat
and
claude
committed
3df8e45
dasLLAMA: the cm2 no-split arms keep literal loop bounds - the split

borisbat
and
claude
committed
e600210
dasLLAMA: delete the fmt-6 q8n side stack; INVENTORY.md retires to

borisbat
and
claude
committed
903edc1
dasLLAMA: DASLLAMA_VK_SPV_OVERRIDE - serve <kernel>.spv files instead

borisbat
and
claude
committed
8aef7e8
dasSpirv: the cm2 flash-attention rail - coopmatReduceNV, per-element

borisbat
and
claude
committed
8be1c9a
dasLLAMA: cm2 flash attention - FaCm2H64/H128 tiles serve the prefill

borisbat
and
claude
committed
1df5808
dasLLAMA: gate cm2 flash attention to hs=64 - the h128 tile set spills

borisbat
and
claude
committed
30b925c
dasLLAMA: the h128 fa Bc=32 cut - oracle-exact, measured no-change, gate

borisbat
and
claude
committed
5dbc848
dasLLAMA: the h128 fa collapse was VRAM oversubscription, not the kernel

borisbat
and
claude
committed
5383f05
dasLLAMA: regen ENVIRONMENT.md after the master rebase - the EnvConfig world already carries HF_HOME/HOME

borisbat
and
claude
committed
7c700f8
dasLLAMA: native f16 KV mirrors in the vulkan resident driver - reified kernel templates, shadows deleted

borisbat
and
claude
committed
aaaf58e
dasLLAMA: hot-path contracts on the vulkan driver - the lint sweep the rail never had

borisbat
and
claude
committed
7df329f
dasLLVM: always-on jit codegen timing + tune permutation-stamp log; CLAUDE.md: build all of your own modules

borisbat
and
claude
committed
561f62c
dasLLAMA: CODEREVIEW.md learns the KV-codec world + changed-set lint to zero

borisbat
and
claude
committed
7db93ce
review round 1: the four-agent findings - codec/cap gates, fa pre-flight, split-k range cap, VHZ_SK bit, orphan tile delete, display names, stale-f32 text

borisbat
and
claude
committed
e2ff151
dasLLAMA: gpu_tier contract test catches up to the kv16/byte-row seam (Copilot round 2)

borisbat
and
claude
committed
554eff9

Navigation Menu

Uh oh!

dasLLAMA: native f16 KV mirrors on reified kernel templates - the vulkan resident driver goes codec-driven - #3681

dasLLAMA: native f16 KV mirrors on reified kernel templates - the vulkan resident driver goes codec-driven#3681
borisbat merged 20 commits into
masterfrom
bbatkin/cm2-prefill-gemm

Commits on Aug 10, 2026

dasSpirv: l-tile wg shapes + [spirv_decode] class-method form

dasLLAMA: the cm2 l-tile kernel on native fmt-0 planes

dasLLAMA: mode 4 serves native fmt-0 through the l-tile

dasLLAMA: the l-tile fast/edge path split - interior tiles run unclamped

dasLLAMA: cm2 m-tile + split-k selection for underfilled prefill GEMMs

dasLLAMA: the cm2 no-split arms keep literal loop bounds - the split

dasLLAMA: delete the fmt-6 q8n side stack; INVENTORY.md retires to

dasLLAMA: DASLLAMA_VK_SPV_OVERRIDE - serve <kernel>.spv files instead

dasSpirv: the cm2 flash-attention rail - coopmatReduceNV, per-element

dasLLAMA: cm2 flash attention - FaCm2H64/H128 tiles serve the prefill

dasLLAMA: gate cm2 flash attention to hs=64 - the h128 tile set spills

dasLLAMA: the h128 fa Bc=32 cut - oracle-exact, measured no-change, gate

dasLLAMA: the h128 fa collapse was VRAM oversubscription, not the kernel

dasLLAMA: regen ENVIRONMENT.md after the master rebase - the EnvConfig world already carries HF_HOME/HOME

dasLLAMA: native f16 KV mirrors in the vulkan resident driver - reified kernel templates, shadows deleted

dasLLAMA: hot-path contracts on the vulkan driver - the lint sweep the rail never had

dasLLVM: always-on jit codegen timing + tune permutation-stamp log; CLAUDE.md: build all of your own modules

dasLLAMA: CODEREVIEW.md learns the KV-codec world + changed-set lint to zero

review round 1: the four-agent findings - codec/cap gates, fa pre-flight, split-k range cap, VHZ_SK bit, orphan tile delete, display names, stale-f32 text

dasLLAMA: gpu_tier contract test catches up to the kv16/byte-row seam (Copilot round 2)