NeMo
Nemotron-4-340B-Instruct / model_config.yaml
okuchaiev's picture
Add files using large-upload tool
365b81f verified
raw
history blame
5.97 kB
mcore_gpt: true
micro_batch_size: 1
global_batch_size: 256
tensor_model_parallel_size: 8
pipeline_model_parallel_size: 4
virtual_pipeline_model_parallel_size: null
encoder_seq_length: 4096
max_position_embeddings: 4096
num_layers: 96
hidden_size: 18432
ffn_hidden_size: 73728
num_attention_heads: 96
init_method_std: 0.0063
use_scaled_init_method: true
hidden_dropout: 0.0
attention_dropout: 0.0
ffn_dropout: 0.0
kv_channels: null
apply_query_key_layer_scaling: true
normalization: layernorm1p
layernorm_epsilon: 1.0e-05
do_layer_norm_weight_decay: false
make_vocab_size_divisible_by: 128
pre_process: true
post_process: true
persist_layer_norm: true
bias: false
activation: squared-relu
headscale: false
transformer_block_type: pre_ln
openai_gelu: false
normalize_attention_scores: true
position_embedding_type: rope
rotary_percentage: 0.5
attention_type: multihead
share_embeddings_and_output_weights: false
num_query_groups: 8
tokenizer:
library: sentencepiece
type: null
model: nemo:8223bf8eaa194eb8920af568bb52e2d0_megatron_2.model
vocab_file: null
merge_file: null
tokenizer_model: nemo:eb5528fdec5c4083affa2c97958eeef7_megatron_2.model
sentencepiece_legacy: false
native_amp_init_scale: 4294967296
native_amp_growth_interval: 1000
hysteresis: 2
fp32_residual_connection: false
fp16_lm_cross_entropy: false
megatron_amp_O2: true
grad_allreduce_chunk_size_mb: 125
grad_div_ar_fusion: true
gradient_accumulation_fusion: false
bias_activation_fusion: false
bias_dropout_add_fusion: false
masked_softmax_fusion: true
seed: 1234
resume_from_checkpoint: null
use_cpu_initialization: false
onnx_safe: false
apex_transformer_log_level: 30
gradient_as_bucket_view: false
sync_batch_comm: false
activations_checkpoint_granularity: null
activations_checkpoint_method: null
activations_checkpoint_num_layers: 1
num_micro_batches_with_partial_activation_checkpoints: null
activations_checkpoint_layers_per_pipeline: null
sequence_parallel: false
transformer_engine: false
fp8: false
fp8_e4m3: false
fp8_hybrid: false
fp8_margin: 0
fp8_interval: 1
fp8_amax_history_len: 1
fp8_amax_compute_algo: most_recent
reduce_amax: true
use_emha: false
optim:
name: distributed_fused_adam
lr: 3.002e-07
weight_decay: 0.1
betas:
- 0.9
- 0.98
sched:
name: CosineAnnealing
warmup_steps: 10
constant_steps: 400
min_lr: 3.0e-07
bucket_cap_mb: 200
overlap_grad_sync: false
contiguous_grad_buffer: true
precision: bf16-mixed
data:
chat: true
chat_prompt_tokens:
system_turn_start: <extra_id_0>
turn_start: <extra_id_1>
label_start: <extra_id_2>
end_of_turn: '
'
end_of_name: '
'
sample: true
num_workers: 2
dataloader_type: single
train_ds:
file_path: /dataset/daring-anteater_commercial.shuf.removelong.jsonl
global_batch_size: 128
micro_batch_size: 1
shuffle: true
memmap_workers: null
max_seq_length: 4096
min_seq_length: 1
drop_last: true
concat_sampling_probabilities: null
label_key: output
add_eos: false
add_sep: false
add_bos: false
truncation_field: input
index_mapping_dir: /indexmap_dir
prompt_template: '<extra_id_0>System
{system message}
<extra_id_1>User
{turn 1 user message}
<extra_id_1>Assistant
<extra_id_2>{turn 1 assistant label}
{turn 1 assistant message}
<extra_id_1>User
{turn 2 user message}
<extra_id_1>Assistant
<extra_id_2>{turn 2 assistant label}
{turn 2 assistant message}
<extra_id_1>'
hf_dataset: true
truncation_method: right
validation_ds:
file_path: /dataset/daring-anteater_commercial.shuf.removelong.jsonl
names: null
global_batch_size: 128
micro_batch_size: 1
shuffle: false
memmap_workers: null
max_seq_length: 4096
min_seq_length: 1
drop_last: false
label_key: output
add_eos: false
add_sep: false
add_bos: false
write_predictions_to_file: false
output_file_path_prefix: null
truncation_field: input
index_mapping_dir: /indexmap_dir
prompt_template: '<extra_id_0>System
{system message}
<extra_id_1>User
{turn 1 user message}
<extra_id_1>Assistant
<extra_id_2>{turn 1 assistant label}
{turn 1 assistant message}
<extra_id_1>User
{turn 2 user message}
<extra_id_1>Assistant
<extra_id_2>{turn 2 assistant label}
{turn 2 assistant message}
<extra_id_1>'
tokens_to_generate: 32
hf_dataset: true
truncation_method: right
metric:
name: loss
average: null
num_classes: null
test_ds:
prompt_template: '<extra_id_0>System
{system message}
<extra_id_1>User
{turn 1 user message}
<extra_id_1>Assistant
<extra_id_2>{turn 1 assistant label}
{turn 1 assistant message}
<extra_id_1>User
{turn 2 user message}
<extra_id_1>Assistant
<extra_id_2>{turn 2 assistant label}
{turn 2 assistant message}
<extra_id_1>'
data_impl: jsonl
splits_string: null
seq_length: 4096
skip_warmup: true
reset_position_ids: false
reset_attention_mask: false
eod_mask_loss: false
index_mapping_dir: /indexmap_dir
data_prefix:
train:
- /datasets/v30_benign-walrus_clip153600.jsonl
validation:
- /datasets/v30_benign-walrus_clip153600.jsonl
test:
- /datasets/v30_benign-walrus_clip153600.jsonl
answer_only_loss: true
restore_from_path: /models/340B_100p_CT_100B
save_nemo_on_validation_end: true
use_flash_attention: null
pipeline_model_parallel_split_rank: 0
dpo:
log_prob_forward_micro_batch_size: 2
ref_policy_kl_penalty: 0.3
average_log_probs: false
sft_loss_coeff: 1.0e-05
optimize_ref_policy_kl_penalty: false
preference_loss: reward_rev_dpo
gt_reward_scale: 1.0
apply_rope_fusion: false
target: nemo.collections.nlp.models.language_modeling.megatron_gpt_model.MegatronGPTModel
nemo_version: 1.22.0