gemma : allow offloading the output tensor

[slaren]

Apply the same solution as with falcon to allow offloading the output tensor.

Device 0: NVIDIA GeForce RTX 3090 Ti, compute capability 8.6, VMM: yes

model	size	params	backend	ngl	test	t/s master	t/s PR	speedup
gemma 7B Q4_K - Medium	4.77 GiB	8.54 B	CUDA	99	pp 512	1996.45 ± 206.39	3959.67 ± 110.36	1.98
gemma 7B Q4_K - Medium	4.77 GiB	8.54 B	CUDA	99	tg 128	60.60 ± 0.26	110.51 ± 0.30	1.82

@JohannesGaessler for some reason, compare-llama-bench.py does not work in this case. Do you know what may be reason?

Nvm, I think this is because this change caused the reported model size/params count to change.

Traceback (most recent call last):
  File "/home/diego/code/llama.cpp/scripts/compare-llama-bench.py", line 305, in <module>
    gpu_blas = bool(rows_full[0][KEY_PROPERTIES.index("gpu_blas")])
IndexError: list index out of range

This query seems to return 0 rows:

SELECT tb.cpu_info, tb.gpu_info, tb.n_gpu_layers, tb.main_gpu, tb.cuda, tb.opencl, tb.metal, tb.gpu_blas, tb.blas, tb.model_filename, tb.model_type, tb.model_size, tb.model_n_params, tb.n_batch, tb.n_threads, tb.type_k, tb.type_v, tb.no_kv_offload, tb.mul_mat_q, tb.tensor_split, tb.n_prompt, tb.n_gen, tb.n_prompt, tb.n_gen, AVG(tb.avg_ts), AVG(tc.avg_ts) FROM test tb JOIN test tc ON tb.cpu_info = tc.cpu_info AND tb.gpu_info = tc.gpu_info AND tb.n_gpu_layers = tc.n_gpu_layers AND tb.main_gpu = tc.main_gpu AND tb.cuda = tc.cuda AND tb.opencl = tc.opencl AND tb.metal = tc.metal AND tb.gpu_blas = tc.gpu_blas AND tb.blas = tc.blas AND tb.model_filename = tc.model_filename AND tb.model_type = tc.model_type AND tb.model_size = tc.model_size AND tb.model_n_params = tc.model_n_params AND tb.n_batch = tc.n_batch AND tb.n_threads = tc.n_threads AND tb.type_k = tc.type_k AND tb.type_v = tc.type_v AND tb.no_kv_offload = tc.no_kv_offload AND tb.mul_mat_q = tc.mul_mat_q AND tb.tensor_split = tc.tensor_split AND tb.n_prompt = tc.n_prompt AND tb.n_gen = tc.n_gen AND tb.build_commit = '89febfed' AND tc.build_commit = '22ca4ddb' GROUP BY tb.cpu_info, tb.gpu_info, tb.n_gpu_layers, tb.main_gpu, tb.cuda, tb.opencl, tb.metal, tb.gpu_blas, tb.blas, tb.model_filename, tb.model_type, tb.model_size, tb.model_n_params, tb.n_batch, tb.n_threads, tb.type_k, tb.type_v, tb.no_kv_offload, tb.mul_mat_q, tb.tensor_split, tb.n_prompt, tb.n_gen, tb.n_gen, tb.n_prompt ORDER BY tb.cpu_info, tb.gpu_info, tb.n_gpu_layers, tb.main_gpu, tb.cuda, tb.opencl, tb.metal, tb.gpu_blas, tb.blas, tb.model_filename, tb.model_type, tb.model_size, tb.model_n_params, tb.n_batch, tb.n_threads, tb.type_k, tb.type_v, tb.no_kv_offload, tb.mul_mat_q, tb.tensor_split, tb.n_prompt, tb.n_gen, tb.n_gen, tb.n_prompt;

This is the data in the tables:

sqlite> select * from test;
89febfed|2230|1|0|0|0|0|0|1|1|13th Gen Intel(R) Core(TM) i9-13900K|NVIDIA GeForce RTX 3090 Ti|models/gemma-7b-it-Q4_K_M.gguf|gemma 7B Q4_K - Medium|5121183744|8538074112|512|16|f16|f16|99|layer|0|0|1|0.00|1|512|0|2024-02-21T20:40:28Z|268801109|48015341|1944.929686|281.264199
89febfed|2230|1|0|0|0|0|0|1|1|13th Gen Intel(R) Core(TM) i9-13900K|NVIDIA GeForce RTX 3090 Ti|models/gemma-7b-it-Q4_K_M.gguf|gemma 7B Q4_K - Medium|5121183744|8538074112|512|16|f16|f16|99|layer|0|0|1|0.00|1|0|128|2024-02-21T20:40:30Z|2103439045|12619252|60.854476|0.364524
22ca4ddb|2231|1|0|0|0|0|0|1|1|13th Gen Intel(R) Core(TM) i9-13900K|NVIDIA GeForce RTX 3090 Ti|models/gemma-7b-it-Q4_K_M.gguf|gemma 7B Q4_K - Medium|5563772928|9324899328|512|16|f16|f16|99|layer|0|0|1|0.00|1|512|0|2024-02-21T20:40:42Z|128998178|2139241|3969.902909|64.418657
22ca4ddb|2231|1|0|0|0|0|0|1|1|13th Gen Intel(R) Core(TM) i9-13900K|NVIDIA GeForce RTX 3090 Ti|models/gemma-7b-it-Q4_K_M.gguf|gemma 7B Q4_K - Medium|5563772928|9324899328|512|16|f16|f16|99|layer|0|0|1|0.00|1|0|128|2024-02-21T20:40:43Z|1162848942|3929889|110.075492|0.371931

[cebtenzzre]

This change breaks CPU inference of the Q4_0 quant, with or without #5650:

llm_load_tensors: ggml ctx size =    0.10 MiB
ggml_new_object: not enough space in the context's memory pool (needed 101968, available 101600)
[1]    60495 segmentation fault (core dumped)

[slaren]

Should be fixed in #5651