Q4_K/Q5_K canonical ggml layout + FP32 MemSeg arena leak fix - #556

Merged

michalharakal merged 7 commits into

feature/ISSUE-555-q4k-q5k-arena-leak

Apr 28, 2026

Commits on Apr 28, 2026

fix: GC-reclaim FP32 MemSeg transpose/matmul output arenas

michalharakal
and
claude
committed
6863d53
feat: liveness-based freeing of intermediate tensors in ComputeGraphExecutor

michalharakal
and
claude
committed
2ec0991
fix: MemorySegmentTensorDataFactory uses Arena.ofAuto for per-op outputs

michalharakal
and
claude
committed
0ae4e32
perf: vectorize and parallelize CPU matmul kernels

michalharakal
and
claude
committed
b1a034d
fix(q4_k): apply canonical ggml layout in tensor data + SIMD matmul

michalharakal
and
claude
committed
e661eda
fix(q5_k): use canonical ggml `qh[l]` indexing in dequantQ5KFromBytes

michalharakal
and
claude
committed
e7a39c0
fix(test): drop parens from Q4_K test names for Kotlin/Native

michalharakal
and
claude
committed
063bdb3