vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-06 18:39:04 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add explaination for why the output layer weights should be kept in metal shared memory

Browse source
This commit is contained in:
Lizonghang 2025-01-25 23:51:16 +04:00
parent 1ca9a43bd1
commit ac5d63b09e
1 changed files with 3 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
common/arg.cpp
View file

@ -738,6 +738,9 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
} }
).set_env("LLAMA_ARG_CUDA_MEM")); ).set_env("LLAMA_ARG_CUDA_MEM"));
// #ifdef GGML_USE_METAL // #ifdef GGML_USE_METAL
// // warn: if the output layer weights are not kept in metal shared memory, its mmap-ed weight data
// // could be released by the OS and reloaded repeatedly, which causes additional disk I/O latency.
// // so we recommend to keep the output layer weights in metal shared memory.
// add_opt(llama_arg( // add_opt(llama_arg(
// {"--keep-out-in-metal"}, // {"--keep-out-in-metal"},
// format("whether to keep output weights in metal memory (default: %s)", params.keep_out_in_metal ? "true" : "false"), // format("whether to keep output weights in metal memory (default: %s)", params.keep_out_in_metal ? "true" : "false"),