ggml-meta: propagate buffer usage and call init on the new tensors

[max-krasnyansky]

Overview

Another try at fixing ggml-meta to propagate buffer usage and properly initialize new tensors.
Original PR was #26502

Tested with ggml-metal and ggml-hexagon backends.

GGML_METAL_DEVICES=2 ./build-macos/bin/llama-completion  -m ../gguf/gemma-4-E4B_q4_0-it.gguf  -p "just a test"  -n 32  -ngl 99 --split-mode tensor  --device MTL0,MTL1 -no-cnv

tensor_init moved after tensor->data is setup which should resolve issues with CUDA.

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

[cdanis]

As requested, tested with -sm tensor on a few different model archs on dual RTX A5000. Seems to work great, thanks @max-krasnyansky !

[max-krasnyansky]

@youyoulyz
Can you please check on your CUDA setup as well.

Looking good with ggml-metal on my M4 Pro

GGML_METAL_DEVICES=2 ./build-macos/bin/test-llama-archs
...

[max-krasnyansky]

double-checked on my CUDA setup.

./build-cuda/bin/llama-completion --device CUDA0,CUDA1 --split-mode tensor -m Qwen3.8-27B-Q4_0.gguf -f sample_prompt_4096.txt --ctx-size 8192 --jinja -st -v 
...
0.03.052.429 I llama_prepare_model_devices: creating a Meta device with 2 devices
0.03.052.432 I llama_prepare_model_devices: - device 0: CUDA0
0.03.052.432 I llama_prepare_model_devices: - device 1: CUDA1
...
0.14.309.941 I common_perf_print: prompt eval time =    3350.21 ms /  4018 tokens (    0.83 ms per token,  1199.33 tokens per second)
0.14.309.942 I common_perf_print:        eval time =    2984.28 ms /   186 runs   (   16.04 ms per token,    62.33 tokens per second)
0.14.309.943 I common_perf_print:       total time =    6434.54 ms /  4204 tokens
0.14.309.946 I common_perf_print: unaccounted time =      44.12 ms /   0.7 %      (total - sampling - prompt eval - eval) / (total)
0.14.309.947 I common_perf_print:    graphs reused =        184
0.14.310.083 I common_memory_breakdown_print: | memory breakdown [MiB] |  total     free    self   model   context   compute    unaccounted |
0.14.310.085 I common_memory_breakdown_print: |   - Meta() (Meta())    | 162305 = 145039 + (8028 =  7184 +     330 +     513) +        9237 |
0.14.310.086 I common_memory_breakdown_print: |   - Host               |                     710 =   682 +       0 +      28                |
0.14.310.190 D ~llama_context:     Meta() compute buffer size is 513.0177 MiB, matches expectation of 513.0177 MiB
0.14.310.194 D ~llama_context:  CUDA_Host compute buffer size is  28.0196 MiB, matches expectation of  28.0196 MiB

@ggerganov should be good to merge

[max-krasnyansky]

@ggerganov sorry for bugging you. It's blocking for #26501
I could probably workaround in there but seems like this is clean and simple, besides breaking CUDA on the first try :), but that's all tested now.

[ggerganov]

@ggerganov sorry for bugging you. It's blocking for #26501 I could probably workaround in there but seems like this is clean and simple, besides breaking CUDA on the first try :), but that's all tested now.

No problem at all. I was just planning to add more tests with -sm tensor to the CI, but can't get around to do that still.