The #364 GPU campaigns on the 2080 Ti exposed four distinct Vulkan-backend defects. Full provenance in the linked run logs; all reproduce under tests/bench_perf_sweep with GEIST_BENCH_BACKEND=vulkan.
- gemma4-e4b Q4_K_M:
GEIST_E_BACKEND at pp1024 (all repeats; pp128–512 fine) — run 34157240964. Smells like device-memory exhaustion/fragmentation (4.2 GB weights + KV on 11 GB).
- bitnet_b1_58-large TQ2_0:
GEIST_E_UNSUPPORTED at pp1024 (pp128–512 fine; same model+shape passes on cpu_neon at 198 tok/s) — run 34157633018.
- qwen3-0.6b Q8_0: hard hang in the first measured prefill — no output for 6 h until the job timeout killed it; run 34157843292. A plain-attention model, so this is not the DeltaNet gap below.
- qwen3.5 arch:
state_create fails at model load (decoder state_create failed) — expected in the sense that DeltaNet was only ever built for metal/cpu_neon, but the Vulkan backend should reject the arch with a clear capability error, not a generic load failure — runs 34178688335, 34178881180.
Until these are fixed the matrix carries the affected NVIDIA Vulkan cells as unmeasured/incompatible with this issue as the reference.
The #364 GPU campaigns on the 2080 Ti exposed four distinct Vulkan-backend defects. Full provenance in the linked run logs; all reproduce under
tests/bench_perf_sweepwithGEIST_BENCH_BACKEND=vulkan.GEIST_E_BACKENDat pp1024 (all repeats; pp128–512 fine) — run 34157240964. Smells like device-memory exhaustion/fragmentation (4.2 GB weights + KV on 11 GB).GEIST_E_UNSUPPORTEDat pp1024 (pp128–512 fine; same model+shape passes on cpu_neon at 198 tok/s) — run 34157633018.state_createfails at model load (decoder state_create failed) — expected in the sense that DeltaNet was only ever built for metal/cpu_neon, but the Vulkan backend should reject the arch with a clear capability error, not a generic load failure — runs 34178688335, 34178881180.Until these are fixed the matrix carries the affected NVIDIA Vulkan cells as unmeasured/incompatible with this issue as the reference.