Claude (noether) dbf97a0fc1 FIX: panvk priv-heap exhaustion solved (shared TLS + WLS reuse); crash gone ub64->1024 ...

Device-shared TLS scratch + grid-independent WLS reuse collapse per-cmdbuf 304MB TLS and
per-dispatch WLS into ~1 BO each. Peak priv-VA 3985MB->908MB, no OOM at any ubatch (was crash
at >=72). Verified on panthor. Real upstreamable Mesa panvk bug fix (mesa-panvk-0010).

BUT prefill speedup NOT achieved: larger ubatch is 3x SLOWER (ub64=21, ub128=5.8, ub256=7.4 t/s)
because llama.cpp's L-variant matmul shader spills ~10x more registers (the 304MB stack) -> poor
Mali occupancy. ub64 @ 21 t/s stays the ceiling, ties CPU. Remaining lever = ggml-vulkan L-shader
Mali tuning (separate effort). See docs/RESULTS.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

2026-07-11 16:30:55 +02:00

..

benchmarks.md

panvk WLS fix v1 + confirmed mechanism (partial: reuse halves but doesn't eliminate)

2026-07-11 14:06:47 +02:00

fable-review.md

panvk WLS fix v1 + confirmed mechanism (partial: reuse halves but doesn't eliminate)

2026-07-11 14:06:47 +02:00

fix-plan.md

Add fix plan: cleanup_frequency=10 is the accumulation trigger; consumer-side bounding is the primary lever

2026-07-11 14:58:50 +02:00

RESULTS.md

FIX: panvk priv-heap exhaustion solved (shared TLS + WLS reuse); crash gone ub64->1024

2026-07-11 16:30:55 +02:00

root-cause-final.md

Phase 1: consumer-side levers falsified; accumulation is in-flight cmdbufs x 304MB TLS

2026-07-11 16:00:09 +02:00

root-cause.md

rocky-vulkan-llama: skeleton — deps, root-cause diagnosis, llama.cpp submit patch

2026-07-11 13:33:00 +02:00