seed: 0
output_dir: './output'
load_checkpoint: ''
src_strategy_path_or_dir: ''
auto_trans_ckpt: False
run_mode: 'predict'
trainer:
type: CausalLanguageModelingTrainer
model_name: 'telechat_35b'
use_parallel: True
parallel:
parallel_mode: "STAND_ALONE"
full_batch: False
strategy_ckpt_save_file: "./ckpt_strategy.ckpt"
parallel_config:
data_parallel: 1
model_parallel: 2
vocab_emb_dp: False
context:
mode: 0
max_device_memory: "58GB"
device_id: 0
save_graphs: False
save_graphs_path: "./graph"
mempool_block_size: "58GB"
affinity_cpu_list: "None"
model:
model_config:
type: TelechatConfig
batch_size: 1
net_name: "telechat3_36b"
seq_length: 8192
hidden_size: 6144
num_layers: 64
layers_group: 1
num_heads: 48
n_kv_heads: 8
embedding_size: 512
vocab_size: 131072
eod_reset: False
out_proj_has_bias: False
rms_norm_eps: 1.0e-5
bos_token_id: 1
eos_token_id: 2
pad_token_id: 3
ignore_token_id: -100
intermediate_size: 24576
res_dtype: "bfloat16"
compute_dtype: "bfloat16"
layernorm_compute_type: "float32"
softmax_compute_type: "float32"
rotary_dtype: "float32"
param_init_type: "bfloat16"
use_past: True
pretrain_seqlen: 8192
extend_method: "None"
use_flash_attention: True
block_size: 16
num_blocks: 512
is_dynamic: True
qkv_concat: False
use_past_shard: False
repetition_penalty: 1.03
max_decode_length: 512
top_k: 3
top_p: 1
do_sample: False
auto_map:
AutoModel: telechat.ParallelTelechatForCausalLM
AutoConfig: telechat_config.TelechatConfig
AutoTokenizer: [telechat_tokenizer.TelechatTokenizer, null]
arch:
type: ParallelTelechatForCausalLM
processor:
return_tensors: ms
tokenizer:
vocab_file: ""
unk_token: '<unk>'
bos_token: '<_start>'
eos_token: '<_end>'
pad_token: '<_pad>'
type: TelechatTokenizer