UPSTREAM PR #18919: metal : support virtual devices

[loci-dev]

Mirrored from ggml-org/llama.cpp#18919

Support virtual Metal devices. Allows simulating multi-GPU environments on Mac using the new GGML_METAL_DEVICES environment variable.

GGML_METAL_DEVICES=4 ./bin/llama-completion -m [model.gguf]

...

0.02.020.033 I llama_memory_breakdown_print: | memory breakdown [MiB]    |  total     free    self   model   context   compute    unaccounted |
0.02.020.034 I llama_memory_breakdown_print: |   - MTL0 (Apple M2 Ultra) | 165150 = 158091 + (1916 =   780 +    1024 +     112) +        5143 |
0.02.020.034 I llama_memory_breakdown_print: |   - MTL1 (Apple M2 Ultra) | 165150 = 158091 + (1738 =   780 +     896 +      62) +        5320 |
0.02.020.036 I llama_memory_breakdown_print: |   - MTL2 (Apple M2 Ultra) | 165150 = 158091 + (1198 =   240 +     896 +      62) +        5861 |
0.02.020.037 I llama_memory_breakdown_print: |   - MTL3 (Apple M2 Ultra) | 165150 = 158091 + (2205 =  1137 +     768 +     300) +        4853 |
0.02.020.037 I llama_memory_breakdown_print: |   - Host                  |                     364 =   296 +       0 +      68                |

[loci-review]

Explore the complete analysis inside the Version Insights

[loci-review]

Explore the complete analysis inside the Version Insights

Based on the analysis, no functions were identified with measurable performance changes between the base and target versions. This indicates no meaningful performance impact from the code changes.