-
Notifications
You must be signed in to change notification settings - Fork 8.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
ggml llama: align structs for memory optimization on 64-bit platforms #7267
base: master
Are you sure you want to change the base?
Conversation
GermanAizek
commented
May 13, 2024
- ggml_type_traits_t (80 -> 72 bytes)
- llama_batch (72 -> 64 bytes)
- llama_model_params (56 -> 48 bytes)
- hash_node (32 -> 24 bytes)
- ggml_compute_state (32 -> 24 bytes)
- gguf_tensor_info (88 -> 80 bytes)
- ggml_type_traits_t (80 -> 72 bytes) - llama_batch (72 -> 64 bytes) - llama_model_params (56 -> 48 bytes) - hash_node (32 -> 24 bytes) - ggml_compute_state (32 -> 24 bytes) - gguf_tensor_info (88 -> 80 bytes)
The changes to the llama.h public structs are effectively a API breaking change for no real benefit. The other structs are less sensitive since they are internal to ggml, but I don't see how this is worth the risk of introducing bugs. |
What I find interesting "sample time" has regressed about 0.02 ms for me. Not like that has much meaning but still interesting.
|
Which compiler did you compile on for which platform? |
main: built with cc (Debian 12.2.0-14) 12.2.0 for x86_64-linux-gnu `./llama-bench -m /mnt/36TB/AI/llama-3-8B-Instruct-More-abliterated/ggml-model-f16.gguf -t 24 -r 20 -pg 512,128 Master:
PR:
|
@USBhost, I have:
master:
PR:
|