metal : fix event synchronization

[ggerganov]

Overview

cont #20463
cont #18919

Fix the event synchronization logic when using virtual Metal devices.

Additional information

I think this was the reason why we didn't observe the regression from #17795 when running the pipeline parallel workflows with virtual devices.

GGML_METAL_DEVICES=4 CMAKE_OPTS="-DGGML_BLAS=OFF" scripts/compare-commits.sh master gg/metal-fix-events llama-bench -m ~/models/gpt-oss-20b/ggml-model-mxfp4.gguf -m ~/models/qwen3-0.6b-base/ggml-model-q8_0.gguf -m ~/models/qwen3.5-0.8b-base/ggml-model-q8_0.gguf -m ~/models/qwen3-30b-a3b/ggml-model-q8_0.gguf -fa 1 -ub 2048,512 -b 16384 -p 512,2048,4096,8192,16384,32768 -n 32,32,32 -r 1 -t 1

Model	Microbatch size	Test	t/s master	t/s gg/metal-fix-events	Speedup
gpt-oss 20B MXFP4 MoE	512	pp512	2350.44	2367.07	1.01
gpt-oss 20B MXFP4 MoE	512	pp2048	2376.94	2339.00	0.98
gpt-oss 20B MXFP4 MoE	512	pp4096	2303.54	2319.16	1.01
gpt-oss 20B MXFP4 MoE	512	pp8192	2154.49	2182.73	1.01
gpt-oss 20B MXFP4 MoE	512	pp16384	1887.93	1911.66	1.01
gpt-oss 20B MXFP4 MoE	512	pp32768	1506.42	1518.49	1.01
gpt-oss 20B MXFP4 MoE	512	tg32	102.71	116.25	1.13
gpt-oss 20B MXFP4 MoE	2048	pp512	2367.84	2382.75	1.01
gpt-oss 20B MXFP4 MoE	2048	pp2048	2703.35	2706.89	1.00
gpt-oss 20B MXFP4 MoE	2048	pp4096	2631.97	2607.63	0.99
gpt-oss 20B MXFP4 MoE	2048	pp8192	2453.51	2429.44	0.99
gpt-oss 20B MXFP4 MoE	2048	pp16384	2131.23	2123.23	1.00
gpt-oss 20B MXFP4 MoE	2048	pp32768	1670.28	1678.03	1.00
gpt-oss 20B MXFP4 MoE	2048	tg32	104.06	116.12	1.12
qwen3 0.6B Q8_0	512	pp512	13577.98	13815.02	1.02
qwen3 0.6B Q8_0	512	pp2048	11833.70	11821.11	1.00
qwen3 0.6B Q8_0	512	pp4096	9619.92	9847.19	1.02
qwen3 0.6B Q8_0	512	pp8192	7085.81	7305.02	1.03
qwen3 0.6B Q8_0	512	pp16384	4612.35	4687.70	1.02
qwen3 0.6B Q8_0	512	pp32768	2682.56	2703.95	1.01
qwen3 0.6B Q8_0	512	tg32	208.69	233.10	1.12
qwen3 0.6B Q8_0	2048	pp512	13482.82	13785.38	1.02
qwen3 0.6B Q8_0	2048	pp2048	13222.19	13268.26	1.00
qwen3 0.6B Q8_0	2048	pp4096	10966.75	10991.81	1.00
qwen3 0.6B Q8_0	2048	pp8192	8071.85	8064.77	1.00
qwen3 0.6B Q8_0	2048	pp16384	5171.93	5174.08	1.00
qwen3 0.6B Q8_0	2048	pp32768	2968.30	2961.29	1.00
qwen3 0.6B Q8_0	2048	tg32	212.48	234.86	1.11
qwen35 0.8B Q8_0	512	pp512	10575.46	10783.21	1.02
qwen35 0.8B Q8_0	512	pp2048	10868.20	10865.48	1.00
qwen35 0.8B Q8_0	512	pp4096	10485.53	10809.71	1.03
qwen35 0.8B Q8_0	512	pp8192	9635.95	10012.95	1.04
qwen35 0.8B Q8_0	512	pp16384	8207.64	8418.07	1.03
qwen35 0.8B Q8_0	512	pp32768	6277.44	6390.16	1.02
qwen35 0.8B Q8_0	512	tg32	154.54	177.99	1.15
qwen35 0.8B Q8_0	2048	pp512	10560.57	10777.40	1.02
qwen35 0.8B Q8_0	2048	pp2048	11856.34	11827.24	1.00
qwen35 0.8B Q8_0	2048	pp4096	11696.23	11783.99	1.01
qwen35 0.8B Q8_0	2048	pp8192	10955.16	10891.34	0.99
qwen35 0.8B Q8_0	2048	pp16384	9391.60	9455.61	1.01
qwen35 0.8B Q8_0	2048	pp32768	7202.31	7244.19	1.01
qwen35 0.8B Q8_0	2048	tg32	156.69	178.88	1.14

Extra CI: https://github.com/ggml-org/llama.cpp/actions/runs/24797481063

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO