[XXXXRT666]

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: Benchmarking and Testing

Summary

Track inflight Metal buffer use separately from array ownership. Reuse released storage on the sole stream still using it, with fences for earlier reads and writes and allocator checks for host writes and donation.

Benchmarks

Values are medians over 48 trials per build for small models, 6 for Qwen3 and Parakeet, and 8 for vision and Qwen3-ASR.

Extra active peak memory does not include model footprint.

Workload	Wall time	Total active peak (MB)	Extra active peak (MB)
Small Transformer encoder, 512 tokens	8.305 → 8.073	56.715 → 12.652	49.500 → 5.438 (−89.0%)
Small LM prefill, 384 tokens	4.617 → 4.576	34.954 → 13.470	28.291 → 6.808 (−75.9%)
Small LM decode, 16 steps	13.786 → 12.920	13.978 → 12.201	3.941 → 2.165 (−45.1%)
Small residual CNN, 160 × 160	1.021 → 1.016	3.245 → 2.635	2.272 → 1.662 (−26.8%)
Small U-Net, 128 × 128	1.215 → 1.178	8.251 → 5.501	7.875 → 5.125 (−34.9%)
Two-layer GRU, 128 steps	14.571 → 12.682	0.678 → 0.566	0.286 → 0.174 (−39.3%)
Qwen3-0.6B prefill, 2,048 tokens	820.45 → 810.14	1,868.88 → 1,404.88	732 → 268 (−63.4%)
Qwen3-0.6B prefill, 4,096 tokens	1,863.32 → 1,860.83	2,112.89 → 1,672.89	976 → 536 (−45.1%)
Qwen3-0.6B prefill, 6,144 tokens	3,338.36 → 3,353.98	2,348.91 → 1,940.91	1,212 → 804 (−33.7%)
Qwen3-0.6B decode, 8 tokens	149.95 → 154.42	1,630.71 → 1,614.20	269.83 → 253.33 (−6.1%)
Parakeet encoder, 33 s	221.03 → 222.37	1,656.07 → 1,341.83	419.58 → 105.06 (−75.0%)
ResNet-50, batch 1	7.903 → 8.028	72.89 → 57.31	23.71 → 8.12 (−65.7%)
ResNet-50, batch 32	175.34 → 169.00	768.41 → 322.53	710.34 → 264.45 (−62.8%)
ViT-L/16, batch 1	59.00 → 58.58	814.81 → 587.38	234.04 → 6.61 (−97.2%)
ViT-L/16, batch 16	745.47 → 766.32	1,222.04 → 671.70	636.97 → 86.64 (−86.4%)
Qwen3-ASR encoder, 33 s	169.03 → 168.82	2,483.54 → 1,841.06	982.18 → 339.70 (−65.4%)
Qwen3-ASR transcription, 33 s	1,700.34 → 1,627.52	2,483.67 → 1,841.07	985.94 → 343.35 (−65.2%)