llama_config.ini

[ft_instance_hyperparameter]
data_type=fp16
enable_custom_all_reduce=0

tensor_para_size=1
pipeline_para_size=1

model_name=llama_7b
model_dir=/data/models/llama-7b-infer/1-gpu

; model_name=llama_13b
; model_dir=/data/models/llama-13b-hf-infer/2-gpu

; model_name=llama_33b
; model_dir=/data/models/llama-33b-hf-infer/4-gpu

; model_name=llama_65b
; model_dir=/data/models/llama-65b-hf-infer/8-gpu

[request]
beam_width=1 # beam width for beam search
top_k=1 ; k value for top k sampling
top_p=0.0 ; p value for top p sampling
temperature=0.0 ; Use for sampling
repetition_penalty=1.0 ; Use for sampling
presence_penalty=0.0  ; Only one of repetition_penalty and presence_penalty are allowed.
len_penalty=0.0
beam_search_diversity_rate=0.0
request_batch_size=1 # determine by the request
request_output_len=256 # determine by the request

[llama_7b]
head_num = 32
size_per_head = 128
inter_size = 11008
num_layer = 32
rotary_embedding = 128
layernorm_eps = 1e-06
vocab_size = 32000
start_id = 0
end_id = 1
weight_data_type = fp16

[llama_13b]
head_num = 40
size_per_head = 128
inter_size = 13824
num_layer = 40
rotary_embedding = 128
layernorm_eps = 1e-06
vocab_size = 32000
start_id = 0
end_id = 1
weight_data_type = fp16

[llama_33b]
head_num = 52
size_per_head = 128
inter_size = 17920
num_layer = 60
rotary_embedding = 128
layernorm_eps = 1e-06
vocab_size = 32000
start_id = 0
end_id = 1
weight_data_type = fp16

[llama_65b]
head_num = 64
size_per_head = 128
inter_size = 22016
num_layer = 80
rotary_embedding = 128
layernorm_eps = 1e-05
vocab_size = 32000
start_id = 0
end_id = 1
weight_data_type = fp16