dasLLAMA vulkan arc: decode-tail fusions, K-quant GEMV, memory-priority shield - #3581

Merged

borisbat merged 35 commits into

bbatkin/dasllama-cm2-arc

Jul 27, 2026

Commits on Jul 26, 2026

dasLLAMA: extract disk->compute layout transforms into dasllama_layout

borisbat
and
claude
committed
dasLLAMA: daspkg release manifest for lcpp_bench (dasllama-bench bundle)

borisbat
and
claude
committed
dasLLAMA vulkan tier: size the stream-slot reserve from the model, not a constant

borisbat
and
claude
committed
dasLLAMA tuner: gate failures record the reference body, and stop deleting a locked array

borisbat
and
claude
committed
dasLLAMA tuner: q51 probe resolves per-variant mr off the layout companion registry

borisbat
and
claude
committed
dasLLAMA: consolidate plane-stride truth into dasllama_gemm_schema

borisbat
and
claude
committed
dasLLAMA vulkan tier: the fastest GEMM path is the default

borisbat
and
claude
committed
dasLLAMA bench: -m accepts a comma-separated model list

borisbat
and
claude
committed
llvm -exe: route das main through a catch boundary — panics print and exit nonzero

borisbat
and
claude
committed
dasLLAMA: q8n groundwork — the cm2 twin's format, device mode, gather, and lens rails

borisbat
and
claude
committed
dasLLAMA: the cm2 twin - decode-in-load prefill GEMM + q8n GEMV on the stack path

borisbat
and
claude
committed
dasLLAMA: the cm2 arena - the resident driver rides q8n planes and the f16-fed twin prefill

borisbat
and
claude
committed
dasLLAMA vulkan tier: memory-priority residency shield (phase 4 wiring)

borisbat
and
claude
committed
dasLLAMA vulkan tier: streamed copies ride a dedicated transfer queue (phase 5)

borisbat
and
claude
committed
dasLLAMA vulkan tier: heat cache serves mixed expert triples, skip logs name the reason

borisbat
and
claude
committed

Commits on Jul 27, 2026