vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

bench: update doc for batched bench

Browse source
This commit is contained in:
Ting Sun 2024-04-05 09:11:41 +07:00
parent fa4802f983
commit 0b1ef49fbf
1 changed files with 4 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

8
examples/batched-bench/README.md
View file

@ -10,16 +10,16 @@ There are 2 modes of operation:
- `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`) - `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`)
```bash ```bash
./batched-bench MODEL_PATH [N_KV_MAX] [N_UBATCH] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL> ./batched-bench MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>
# LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared # LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared
./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 512 0 99 ./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 2048 512 0 99
# LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared # LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared
./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 512 1 99 ./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 2048 512 1 99
# custom set of batches # custom set of batches
./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 2048 512 0 999 0 128,256,512 128,256 1,2,4,8,16,32 ./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 2048 512 512 0 999 0 128,256,512 128,256 1,2,4,8,16,32
``` ```
## Sample results ## Sample results