Hugging Face's logo

HectorHe
/
Deepseek-Coder-V2-Lite-13B-Instruct-Math10K-diff-info-Distill-forward-kl-new

Safetensors
deepseek_v2
custom_code
Model card Files
xet
Community
Deepseek-Coder-V2-Lite-13B-Instruct-Math10K-diff-info-Distill-forward-kl-new / training.log
HectorHe's picture
HectorHe
Training in progress, step 100
7a15969 verified
raw
history blame
44.1 kB
 2025-07-20 07:48:33 - INFO - __main__ - Model parameters ModelConfig(model_name_or_path='deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct', model_revision='main', torch_dtype='bfloat16', trust_remote_code=True, attn_implementation='flash_attention_2', use_peft=False, lora_r=16, lora_alpha=32, lora_dropout=0.05, lora_target_modules=None, lora_modules_to_save=None, lora_task_type='CAUSAL_LM', use_rslora=False, use_dora=False, load_in_8bit=False, load_in_4bit=False, bnb_4bit_quant_type='nf4', use_bnb_nested_quant=False)
2025-07-20 07:48:33 - INFO - __main__ - Script parameters ScriptArguments(dataset_name='lmms-lab/Math10K', dataset_config=None, dataset_train_split='train', dataset_test_split='test', gradient_checkpointing_use_reentrant=False, ignore_bias_buffers=False)
2025-07-20 07:48:33 - INFO - __main__ - Training parameters EfficientDistillationConfig(
_n_gpu=1,
accelerator_config={'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None, 'use_configured_state': False},
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
alpha=0.5,
auto_find_batch_size=False,
average_tokens_across_devices=False,
batch_eval_metrics=False,
benchmarks=[],
bf16=True,
bf16_full_eval=False,
callbacks=[],
ce_loss_scale=1.0,
chars_per_token=<CHARS_PER_TOKEN>,
chat_template=None,
completion_only_loss=None,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_persistent_workers=False,
dataloader_pin_memory=True,
dataloader_prefetch_factor=None,
dataset_batch_size=None,
dataset_kwargs=None,
dataset_num_proc=None,
dataset_text_field=text,
ddp_backend=None,
ddp_broadcast_buffers=None,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800000000,
debug=[],
deepspeed=None,
disable_dropout=True,
disable_tqdm=False,
do_eval=True,
do_predict=False,
do_train=False,
eos_token=<EOS_TOKEN>,
eval_accumulation_steps=None,
eval_delay=0,
eval_do_concat_batches=True,
eval_on_start=False,
eval_packing=None,
eval_steps=None,
eval_strategy=IntervalStrategy.NO,
eval_use_gather_object=False,
expert_num=6,
fp16=False,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False},
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
gradient_checkpointing_kwargs={'use_reentrant': False},
greater_is_better=None,
group_by_length=False,
half_precision_backend=auto,
hub_always_push=False,
hub_model_id=Deepseek-Coder-V2-Lite-13B-Instruct-Math10K-diff-info-Distill-forward-kl-new,
hub_model_revision=main,
hub_private_repo=None,
hub_strategy=HubStrategy.EVERY_SAVE,
hub_token=<HUB_TOKEN>,
ignore_data_skip=False,
include_for_metrics=[],
include_inputs_for_metrics=False,
include_num_input_tokens_seen=False,
include_tokens_per_second=False,
jit_mode_eval=False,
kl_loss_scale=1.0,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=1e-05,
length_column_name=length,
lmbda=0.0,
load_best_model_at_end=False,
local_rank=0,
log_level=info,
log_level_replica=warning,
log_on_each_node=True,
logging_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl/runs/Jul20_07-48-31_haozeh-dev-pod-1,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=1,
logging_strategy=IntervalStrategy.STEPS,
loss_type=token_specific,
lr_scheduler_kwargs={'min_lr_rate': 0.1},
lr_scheduler_type=SchedulerType.COSINE_WITH_MIN_LR,
max_grad_norm=1.0,
max_length=8192,
max_new_tokens=1024,
max_seq_length=None,
max_steps=-1,
metric_for_best_model=None,
model_init_kwargs=None,
mp_parameters=,
neftune_noise_alpha=None,
no_cuda=False,
num_of_sequences=None,
num_train_epochs=3,
optim=OptimizerNames.ADAMW_TORCH,
optim_args=None,
optim_target_modules=None,
output_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
overwrite_hub_revision=False,
overwrite_output_dir=True,
packing=False,
pad_to_multiple_of=None,
pad_token=<PAD_TOKEN>,
padding_free=False,
past_index=-1,
per_device_eval_batch_size=16,
per_device_train_batch_size=4,
prediction_loss_only=False,
push_to_hub=True,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_revision=False,
push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
ray_scope=last,
reduction=sum,
remove_unused_columns=True,
report_to=['wandb'],
restore_callback_states_from_checkpoint=False,
resume_from_checkpoint=None,
run_name=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
save_on_each_node=False,
save_only_model=False,
save_safetensors=True,
save_steps=100,
save_strategy=SaveStrategy.STEPS,
save_total_limit=1,
seed=1234,
skip_memory_metrics=True,
system_prompt=None,
teacher_model_init_kwargs=None,
teacher_model_name_or_path=None,
temperature=0.9,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torch_empty_cache_steps=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_cpu=False,
use_ipex=False,
use_legacy_prediction_loop=False,
use_liger=False,
use_liger_kernel=False,
use_mps_device=False,
wandb_entity=None,
wandb_project=None,
warmup_ratio=0.1,
warmup_steps=0,
weight_decay=0.0,
)
2025-07-20 07:48:42 - INFO - __main__ - *** Initializing model kwargs ***
2025-07-20 07:48:42 - INFO - __main__ - Model memory before loading model:Memory allocated: 0.0
Memory reserved: 0.0
2025-07-20 07:49:51 - INFO - __main__ - Model memory after loading model:Memory allocated: 4836.39697265625
Memory reserved: 7322.0
2025-07-20 07:49:52 - INFO - __main__ - MoE layers replaced with Dense MLP layers
2025-07-20 07:49:52 - INFO - __main__ - Model memory after replacing MoE with dense:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-07-20 07:49:52 - INFO - __main__ - Initializing EfficientDistillationTrainer...
2025-07-20 07:50:16 - INFO - __main__ - Model memory after trainer initialization:Memory allocated: 9670.91943359375
Memory reserved: 12800.0
2025-07-20 07:50:16 - INFO - __main__ - *** Starting training ***
2025-07-20 07:50:16 - INFO - __main__ - Model architecture: DeepseekV2ForCausalLM(
(model): DeepseekV2Model(
(embed_tokens): Embedding(102400, 2048)
(layers): ModuleList(
(0): DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=10944, bias=False)
(up_proj): Linear(in_features=2048, out_features=10944, bias=False)
(down_proj): Linear(in_features=10944, out_features=2048, bias=False)
(act_fn): SiLU()
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
(1-26): 26 x DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MoE(
(experts): ModuleList(
(0-63): 64 x DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=1408, bias=False)
(up_proj): Linear(in_features=2048, out_features=1408, bias=False)
(down_proj): Linear(in_features=1408, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(gate): MoEGate()
(shared_experts): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=2816, bias=False)
(up_proj): Linear(in_features=2048, out_features=2816, bias=False)
(down_proj): Linear(in_features=2816, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
)
(norm): DeepseekV2RMSNorm()
)
(lm_head): Linear(in_features=2048, out_features=102400, bias=False)
)
2025-07-21 06:19:44 - INFO - __main__ - Model parameters ModelConfig(model_name_or_path='deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct', model_revision='main', torch_dtype='bfloat16', trust_remote_code=True, attn_implementation='flash_attention_2', use_peft=False, lora_r=16, lora_alpha=32, lora_dropout=0.05, lora_target_modules=None, lora_modules_to_save=None, lora_task_type='CAUSAL_LM', use_rslora=False, use_dora=False, load_in_8bit=False, load_in_4bit=False, bnb_4bit_quant_type='nf4', use_bnb_nested_quant=False)
2025-07-21 06:19:44 - INFO - __main__ - Script parameters ScriptArguments(dataset_name='lmms-lab/Math10K', dataset_config=None, dataset_train_split='train', dataset_test_split='test', gradient_checkpointing_use_reentrant=False, ignore_bias_buffers=False)
2025-07-21 06:19:44 - INFO - __main__ - Training parameters EfficientDistillationConfig(
_n_gpu=1,
accelerator_config={'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None, 'use_configured_state': False},
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
alpha=0.5,
auto_find_batch_size=False,
average_tokens_across_devices=False,
batch_eval_metrics=False,
benchmarks=[],
bf16=True,
bf16_full_eval=False,
callbacks=[],
ce_loss_scale=1.0,
chars_per_token=<CHARS_PER_TOKEN>,
chat_template=None,
completion_only_loss=None,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_persistent_workers=False,
dataloader_pin_memory=True,
dataloader_prefetch_factor=None,
dataset_batch_size=None,
dataset_kwargs=None,
dataset_num_proc=None,
dataset_text_field=text,
ddp_backend=None,
ddp_broadcast_buffers=None,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800000000,
debug=[],
deepspeed=None,
disable_dropout=True,
disable_tqdm=False,
do_eval=True,
do_predict=False,
do_train=False,
eos_token=<EOS_TOKEN>,
eval_accumulation_steps=None,
eval_delay=0,
eval_do_concat_batches=True,
eval_on_start=False,
eval_packing=None,
eval_steps=None,
eval_strategy=IntervalStrategy.NO,
eval_use_gather_object=False,
expert_num=6,
fp16=False,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False},
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
gradient_checkpointing_kwargs={'use_reentrant': False},
greater_is_better=None,
group_by_length=False,
half_precision_backend=auto,
hub_always_push=False,
hub_model_id=Deepseek-Coder-V2-Lite-13B-Instruct-Math10K-diff-info-Distill-forward-kl-new,
hub_model_revision=main,
hub_private_repo=None,
hub_strategy=HubStrategy.EVERY_SAVE,
hub_token=<HUB_TOKEN>,
ignore_data_skip=False,
include_for_metrics=[],
include_inputs_for_metrics=False,
include_num_input_tokens_seen=False,
include_tokens_per_second=False,
jit_mode_eval=False,
kl_loss_scale=1.0,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=1e-05,
length_column_name=length,
lmbda=0.0,
load_best_model_at_end=False,
local_rank=0,
log_level=info,
log_level_replica=warning,
log_on_each_node=True,
logging_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl/runs/Jul21_06-19-42_haozeh-dev-pod-1,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=1,
logging_strategy=IntervalStrategy.STEPS,
loss_type=token_specific,
lr_scheduler_kwargs={'min_lr_rate': 0.1},
lr_scheduler_type=SchedulerType.COSINE_WITH_MIN_LR,
max_grad_norm=1.0,
max_length=8192,
max_new_tokens=1024,
max_seq_length=None,
max_steps=-1,
metric_for_best_model=None,
model_init_kwargs=None,
mp_parameters=,
neftune_noise_alpha=None,
no_cuda=False,
num_of_sequences=None,
num_train_epochs=3,
optim=OptimizerNames.ADAMW_TORCH,
optim_args=None,
optim_target_modules=None,
output_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
overwrite_hub_revision=False,
overwrite_output_dir=True,
packing=False,
pad_to_multiple_of=None,
pad_token=<PAD_TOKEN>,
padding_free=False,
past_index=-1,
per_device_eval_batch_size=16,
per_device_train_batch_size=4,
prediction_loss_only=False,
push_to_hub=True,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_revision=False,
push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
ray_scope=last,
reduction=sum,
remove_unused_columns=True,
report_to=['wandb'],
restore_callback_states_from_checkpoint=False,
resume_from_checkpoint=None,
run_name=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
save_on_each_node=False,
save_only_model=False,
save_safetensors=True,
save_steps=100,
save_strategy=SaveStrategy.STEPS,
save_total_limit=1,
seed=1234,
skip_memory_metrics=True,
system_prompt=None,
teacher_model_init_kwargs=None,
teacher_model_name_or_path=None,
temperature=0.9,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torch_empty_cache_steps=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_cpu=False,
use_ipex=False,
use_legacy_prediction_loop=False,
use_liger=False,
use_liger_kernel=False,
use_mps_device=False,
wandb_entity=None,
wandb_project=None,
warmup_ratio=0.1,
warmup_steps=0,
weight_decay=0.0,
)
2025-07-21 06:20:08 - INFO - __main__ - *** Initializing model kwargs ***
2025-07-21 06:20:08 - INFO - __main__ - Model memory before loading model:Memory allocated: 0.0
Memory reserved: 0.0
2025-07-21 06:22:26 - INFO - __main__ - Model memory after loading model:Memory allocated: 4836.39697265625
Memory reserved: 7322.0
2025-07-21 06:22:27 - INFO - __main__ - MoE layers replaced with Dense MLP layers
2025-07-21 06:22:27 - INFO - __main__ - Model memory after replacing MoE with dense:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-07-21 06:22:27 - INFO - __main__ - Initializing EfficientDistillationTrainer...
2025-07-21 06:23:02 - INFO - __main__ - Model memory after trainer initialization:Memory allocated: 9670.91943359375
Memory reserved: 12800.0
2025-07-21 06:23:02 - INFO - __main__ - *** Starting training ***
2025-07-21 06:23:02 - INFO - __main__ - Model architecture: DeepseekV2ForCausalLM(
(model): DeepseekV2Model(
(embed_tokens): Embedding(102400, 2048)
(layers): ModuleList(
(0): DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=10944, bias=False)
(up_proj): Linear(in_features=2048, out_features=10944, bias=False)
(down_proj): Linear(in_features=10944, out_features=2048, bias=False)
(act_fn): SiLU()
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
(1-26): 26 x DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MoE(
(experts): ModuleList(
(0-63): 64 x DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=1408, bias=False)
(up_proj): Linear(in_features=2048, out_features=1408, bias=False)
(down_proj): Linear(in_features=1408, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(gate): MoEGate()
(shared_experts): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=2816, bias=False)
(up_proj): Linear(in_features=2048, out_features=2816, bias=False)
(down_proj): Linear(in_features=2816, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
)
(norm): DeepseekV2RMSNorm()
)
(lm_head): Linear(in_features=2048, out_features=102400, bias=False)
)
2025-07-21 06:47:16 - INFO - __main__ - Model parameters ModelConfig(model_name_or_path='deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct', model_revision='main', torch_dtype='bfloat16', trust_remote_code=True, attn_implementation='flash_attention_2', use_peft=False, lora_r=16, lora_alpha=32, lora_dropout=0.05, lora_target_modules=None, lora_modules_to_save=None, lora_task_type='CAUSAL_LM', use_rslora=False, use_dora=False, load_in_8bit=False, load_in_4bit=False, bnb_4bit_quant_type='nf4', use_bnb_nested_quant=False)
2025-07-21 06:47:16 - INFO - __main__ - Script parameters ScriptArguments(dataset_name='lmms-lab/Math10K', dataset_config=None, dataset_train_split='train', dataset_test_split='test', gradient_checkpointing_use_reentrant=False, ignore_bias_buffers=False)
2025-07-21 06:47:16 - INFO - __main__ - Training parameters EfficientDistillationConfig(
_n_gpu=1,
accelerator_config={'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None, 'use_configured_state': False},
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
alpha=0.5,
auto_find_batch_size=False,
average_tokens_across_devices=False,
batch_eval_metrics=False,
benchmarks=[],
bf16=True,
bf16_full_eval=False,
callbacks=[],
ce_loss_scale=1.0,
chars_per_token=<CHARS_PER_TOKEN>,
chat_template=None,
completion_only_loss=None,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_persistent_workers=False,
dataloader_pin_memory=True,
dataloader_prefetch_factor=None,
dataset_batch_size=None,
dataset_kwargs=None,
dataset_num_proc=None,
dataset_text_field=text,
ddp_backend=None,
ddp_broadcast_buffers=None,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800000000,
debug=[],
deepspeed=None,
disable_dropout=True,
disable_tqdm=False,
do_eval=True,
do_predict=False,
do_train=False,
eos_token=<EOS_TOKEN>,
eval_accumulation_steps=None,
eval_delay=0,
eval_do_concat_batches=True,
eval_on_start=False,
eval_packing=None,
eval_steps=None,
eval_strategy=IntervalStrategy.NO,
eval_use_gather_object=False,
expert_num=6,
fp16=False,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False},
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
gradient_checkpointing_kwargs={'use_reentrant': False},
greater_is_better=None,
group_by_length=False,
half_precision_backend=auto,
hub_always_push=False,
hub_model_id=Deepseek-Coder-V2-Lite-13B-Instruct-Math10K-diff-info-Distill-forward-kl-new,
hub_model_revision=main,
hub_private_repo=None,
hub_strategy=HubStrategy.EVERY_SAVE,
hub_token=<HUB_TOKEN>,
ignore_data_skip=False,
include_for_metrics=[],
include_inputs_for_metrics=False,
include_num_input_tokens_seen=False,
include_tokens_per_second=False,
jit_mode_eval=False,
kl_loss_scale=1.0,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=1e-05,
length_column_name=length,
lmbda=0.0,
load_best_model_at_end=False,
local_rank=0,
log_level=info,
log_level_replica=warning,
log_on_each_node=True,
logging_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl/runs/Jul21_06-47-14_haozeh-dev-pod-1,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=1,
logging_strategy=IntervalStrategy.STEPS,
loss_type=token_specific,
lr_scheduler_kwargs={'min_lr_rate': 0.1},
lr_scheduler_type=SchedulerType.COSINE_WITH_MIN_LR,
max_grad_norm=1.0,
max_length=8192,
max_new_tokens=1024,
max_seq_length=None,
max_steps=-1,
metric_for_best_model=None,
model_init_kwargs=None,
mp_parameters=,
neftune_noise_alpha=None,
no_cuda=False,
num_of_sequences=None,
num_train_epochs=3,
optim=OptimizerNames.ADAMW_TORCH,
optim_args=None,
optim_target_modules=None,
output_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
overwrite_hub_revision=False,
overwrite_output_dir=True,
packing=False,
pad_to_multiple_of=None,
pad_token=<PAD_TOKEN>,
padding_free=False,
past_index=-1,
per_device_eval_batch_size=16,
per_device_train_batch_size=4,
prediction_loss_only=False,
push_to_hub=True,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_revision=False,
push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
ray_scope=last,
reduction=sum,
remove_unused_columns=True,
report_to=['wandb'],
restore_callback_states_from_checkpoint=False,
resume_from_checkpoint=None,
run_name=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
save_on_each_node=False,
save_only_model=False,
save_safetensors=True,
save_steps=100,
save_strategy=SaveStrategy.STEPS,
save_total_limit=1,
seed=1234,
skip_memory_metrics=True,
system_prompt=None,
teacher_model_init_kwargs=None,
teacher_model_name_or_path=None,
temperature=0.9,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torch_empty_cache_steps=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_cpu=False,
use_ipex=False,
use_legacy_prediction_loop=False,
use_liger=False,
use_liger_kernel=False,
use_mps_device=False,
wandb_entity=None,
wandb_project=None,
warmup_ratio=0.1,
warmup_steps=0,
weight_decay=0.0,
)
2025-07-21 06:47:25 - INFO - __main__ - *** Initializing model kwargs ***
2025-07-21 06:47:25 - INFO - __main__ - Model memory before loading model:Memory allocated: 0.0
Memory reserved: 0.0
2025-07-21 06:48:39 - INFO - __main__ - Model memory after loading model:Memory allocated: 4836.39697265625
Memory reserved: 7322.0
2025-07-21 06:48:40 - INFO - __main__ - MoE layers replaced with Dense MLP layers
2025-07-21 06:48:40 - INFO - __main__ - Model memory after replacing MoE with dense:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-07-21 06:48:40 - INFO - __main__ - Initializing EfficientDistillationTrainer...
2025-07-21 06:49:02 - INFO - __main__ - Model memory after trainer initialization:Memory allocated: 9670.91943359375
Memory reserved: 12800.0
2025-07-21 06:49:02 - INFO - __main__ - *** Starting training ***
2025-07-21 06:49:02 - INFO - __main__ - Model architecture: DeepseekV2ForCausalLM(
(model): DeepseekV2Model(
(embed_tokens): Embedding(102400, 2048)
(layers): ModuleList(
(0): DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=10944, bias=False)
(up_proj): Linear(in_features=2048, out_features=10944, bias=False)
(down_proj): Linear(in_features=10944, out_features=2048, bias=False)
(act_fn): SiLU()
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
(1-26): 26 x DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MoE(
(experts): ModuleList(
(0-63): 64 x DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=1408, bias=False)
(up_proj): Linear(in_features=2048, out_features=1408, bias=False)
(down_proj): Linear(in_features=1408, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(gate): MoEGate()
(shared_experts): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=2816, bias=False)
(up_proj): Linear(in_features=2048, out_features=2816, bias=False)
(down_proj): Linear(in_features=2816, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
)
(norm): DeepseekV2RMSNorm()
)
(lm_head): Linear(in_features=2048, out_features=102400, bias=False)
)
2025-07-21 07:37:25 - INFO - __main__ - Model parameters ModelConfig(model_name_or_path='deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct', model_revision='main', torch_dtype='bfloat16', trust_remote_code=True, attn_implementation='flash_attention_2', use_peft=False, lora_r=16, lora_alpha=32, lora_dropout=0.05, lora_target_modules=None, lora_modules_to_save=None, lora_task_type='CAUSAL_LM', use_rslora=False, use_dora=False, load_in_8bit=False, load_in_4bit=False, bnb_4bit_quant_type='nf4', use_bnb_nested_quant=False)
2025-07-21 07:37:25 - INFO - __main__ - Script parameters ScriptArguments(dataset_name='lmms-lab/Math10K', dataset_config=None, dataset_train_split='train', dataset_test_split='test', gradient_checkpointing_use_reentrant=False, ignore_bias_buffers=False)
2025-07-21 07:37:25 - INFO - __main__ - Training parameters EfficientDistillationConfig(
_n_gpu=1,
accelerator_config={'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None, 'use_configured_state': False},
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
alpha=0.5,
auto_find_batch_size=False,
average_tokens_across_devices=False,
batch_eval_metrics=False,
benchmarks=[],
bf16=True,
bf16_full_eval=False,
callbacks=[],
ce_loss_scale=1.0,
chars_per_token=<CHARS_PER_TOKEN>,
chat_template=None,
completion_only_loss=None,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_persistent_workers=False,
dataloader_pin_memory=True,
dataloader_prefetch_factor=None,
dataset_batch_size=None,
dataset_kwargs=None,
dataset_num_proc=None,
dataset_text_field=text,
ddp_backend=None,
ddp_broadcast_buffers=None,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800000000,
debug=[],
deepspeed=None,
disable_dropout=True,
disable_tqdm=False,
do_eval=True,
do_predict=False,
do_train=False,
eos_token=<EOS_TOKEN>,
eval_accumulation_steps=None,
eval_delay=0,
eval_do_concat_batches=True,
eval_on_start=False,
eval_packing=None,
eval_steps=None,
eval_strategy=IntervalStrategy.NO,
eval_use_gather_object=False,
expert_num=6,
fp16=False,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False},
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
gradient_checkpointing_kwargs={'use_reentrant': False},
greater_is_better=None,
group_by_length=False,
half_precision_backend=auto,
hub_always_push=False,
hub_model_id=Deepseek-Coder-V2-Lite-13B-Instruct-Math10K-diff-info-Distill-forward-kl-new,
hub_model_revision=main,
hub_private_repo=None,
hub_strategy=HubStrategy.EVERY_SAVE,
hub_token=<HUB_TOKEN>,
ignore_data_skip=False,
include_for_metrics=[],
include_inputs_for_metrics=False,
include_num_input_tokens_seen=False,
include_tokens_per_second=False,
jit_mode_eval=False,
kl_loss_scale=1.0,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=1e-05,
length_column_name=length,
lmbda=0.0,
load_best_model_at_end=False,
local_rank=0,
log_level=info,
log_level_replica=warning,
log_on_each_node=True,
logging_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl/runs/Jul21_07-37-23_haozeh-dev-pod-1,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=1,
logging_strategy=IntervalStrategy.STEPS,
loss_type=token_specific,
lr_scheduler_kwargs={'min_lr_rate': 0.1},
lr_scheduler_type=SchedulerType.COSINE_WITH_MIN_LR,
max_grad_norm=1.0,
max_length=8192,
max_new_tokens=1024,
max_seq_length=None,
max_steps=-1,
metric_for_best_model=None,
model_init_kwargs=None,
mp_parameters=,
neftune_noise_alpha=None,
no_cuda=False,
num_of_sequences=None,
num_train_epochs=3,
optim=OptimizerNames.ADAMW_TORCH,
optim_args=None,
optim_target_modules=None,
output_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
overwrite_hub_revision=False,
overwrite_output_dir=True,
packing=False,
pad_to_multiple_of=None,
pad_token=<PAD_TOKEN>,
padding_free=False,
past_index=-1,
per_device_eval_batch_size=16,
per_device_train_batch_size=4,
prediction_loss_only=False,
push_to_hub=True,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_revision=False,
push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
ray_scope=last,
reduction=sum,
remove_unused_columns=True,
report_to=['wandb'],
restore_callback_states_from_checkpoint=False,
resume_from_checkpoint=None,
run_name=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
save_on_each_node=False,
save_only_model=False,
save_safetensors=True,
save_steps=100,
save_strategy=SaveStrategy.STEPS,
save_total_limit=1,
seed=1234,
skip_memory_metrics=True,
system_prompt=None,
teacher_model_init_kwargs=None,
teacher_model_name_or_path=None,
temperature=0.9,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torch_empty_cache_steps=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_cpu=False,
use_ipex=False,
use_legacy_prediction_loop=False,
use_liger=False,
use_liger_kernel=False,
use_mps_device=False,
wandb_entity=None,
wandb_project=None,
warmup_ratio=0.1,
warmup_steps=0,
weight_decay=0.0,
)
2025-07-21 07:37:36 - INFO - __main__ - *** Initializing model kwargs ***
2025-07-21 07:37:36 - INFO - __main__ - Model memory before loading model:Memory allocated: 0.0
Memory reserved: 0.0
2025-07-21 07:38:50 - INFO - __main__ - Model memory after loading model:Memory allocated: 4836.39697265625
Memory reserved: 7322.0
2025-07-21 07:38:50 - INFO - __main__ - MoE layers replaced with Dense MLP layers
2025-07-21 07:38:50 - INFO - __main__ - Model memory after replacing MoE with dense:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-07-21 07:38:50 - INFO - __main__ - Initializing EfficientDistillationTrainer...
2025-07-21 07:39:14 - INFO - __main__ - Model memory after trainer initialization:Memory allocated: 9670.91943359375
Memory reserved: 12800.0
2025-07-21 07:39:14 - INFO - __main__ - *** Starting training ***
2025-07-21 07:39:14 - INFO - __main__ - Model architecture: DeepseekV2ForCausalLM(
(model): DeepseekV2Model(
(embed_tokens): Embedding(102400, 2048)
(layers): ModuleList(
(0): DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=10944, bias=False)
(up_proj): Linear(in_features=2048, out_features=10944, bias=False)
(down_proj): Linear(in_features=10944, out_features=2048, bias=False)
(act_fn): SiLU()
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
(1-26): 26 x DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MoE(
(experts): ModuleList(
(0-63): 64 x DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=1408, bias=False)
(up_proj): Linear(in_features=2048, out_features=1408, bias=False)
(down_proj): Linear(in_features=1408, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(gate): MoEGate()
(shared_experts): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=2816, bias=False)
(up_proj): Linear(in_features=2048, out_features=2816, bias=False)
(down_proj): Linear(in_features=2816, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
)
(norm): DeepseekV2RMSNorm()
)
(lm_head): Linear(in_features=2048, out_features=102400, bias=False)
)
2025-07-21 17:36:15 - INFO - __main__ - Model parameters ModelConfig(model_name_or_path='deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct', model_revision='main', torch_dtype='bfloat16', trust_remote_code=True, attn_implementation='flash_attention_2', use_peft=False, lora_r=16, lora_alpha=32, lora_dropout=0.05, lora_target_modules=None, lora_modules_to_save=None, lora_task_type='CAUSAL_LM', use_rslora=False, use_dora=False, load_in_8bit=False, load_in_4bit=False, bnb_4bit_quant_type='nf4', use_bnb_nested_quant=False)
2025-07-21 17:36:15 - INFO - __main__ - Script parameters ScriptArguments(dataset_name='lmms-lab/Math10K', dataset_config=None, dataset_train_split='train', dataset_test_split='test', gradient_checkpointing_use_reentrant=False, ignore_bias_buffers=False)
2025-07-21 17:36:15 - INFO - __main__ - Training parameters EfficientDistillationConfig(
_n_gpu=1,
accelerator_config={'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None, 'use_configured_state': False},
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
alpha=0.5,
auto_find_batch_size=False,
average_tokens_across_devices=False,
batch_eval_metrics=False,
benchmarks=[],
bf16=True,
bf16_full_eval=False,
callbacks=[],
ce_loss_scale=1.0,
chars_per_token=<CHARS_PER_TOKEN>,
chat_template=None,
completion_only_loss=None,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_persistent_workers=False,
dataloader_pin_memory=True,
dataloader_prefetch_factor=None,
dataset_batch_size=None,
dataset_kwargs=None,
dataset_num_proc=None,
dataset_text_field=text,
ddp_backend=None,
ddp_broadcast_buffers=None,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800000000,
debug=[],
deepspeed=None,
disable_dropout=True,
disable_tqdm=False,
do_eval=True,
do_predict=False,
do_train=False,
eos_token=<EOS_TOKEN>,
eval_accumulation_steps=None,
eval_delay=0,
eval_do_concat_batches=True,
eval_on_start=False,
eval_packing=None,
eval_steps=None,
eval_strategy=IntervalStrategy.NO,
eval_use_gather_object=False,
expert_num=6,
fp16=False,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False},
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
gradient_checkpointing_kwargs={'use_reentrant': False},
greater_is_better=None,
group_by_length=False,
half_precision_backend=auto,
hub_always_push=False,
hub_model_id=Deepseek-Coder-V2-Lite-13B-Instruct-Math10K-diff-info-Distill-forward-kl-new,
hub_model_revision=main,
hub_private_repo=None,
hub_strategy=HubStrategy.EVERY_SAVE,
hub_token=<HUB_TOKEN>,
ignore_data_skip=False,
include_for_metrics=[],
include_inputs_for_metrics=False,
include_num_input_tokens_seen=False,
include_tokens_per_second=False,
jit_mode_eval=False,
kl_loss_scale=1.0,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=1e-05,
length_column_name=length,
lmbda=0.0,
load_best_model_at_end=False,
local_rank=0,
log_level=info,
log_level_replica=warning,
log_on_each_node=True,
logging_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl/runs/Jul21_17-36-13_haozeh-dev-pod-1,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=1,
logging_strategy=IntervalStrategy.STEPS,
loss_type=token_specific,
lr_scheduler_kwargs={'min_lr_rate': 0.1},
lr_scheduler_type=SchedulerType.COSINE_WITH_MIN_LR,
max_grad_norm=1.0,
max_length=8192,
max_new_tokens=1024,
max_seq_length=None,
max_steps=-1,
metric_for_best_model=None,
model_init_kwargs=None,
mp_parameters=,
neftune_noise_alpha=None,
no_cuda=False,
num_of_sequences=None,
num_train_epochs=3,
optim=OptimizerNames.ADAMW_TORCH,
optim_args=None,
optim_target_modules=None,
output_dir=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
overwrite_hub_revision=False,
overwrite_output_dir=True,
packing=False,
pad_to_multiple_of=None,
pad_token=<PAD_TOKEN>,
padding_free=False,
past_index=-1,
per_device_eval_batch_size=16,
per_device_train_batch_size=4,
prediction_loss_only=False,
push_to_hub=True,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_revision=False,
push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
ray_scope=last,
reduction=sum,
remove_unused_columns=True,
report_to=['wandb'],
restore_callback_states_from_checkpoint=False,
resume_from_checkpoint=None,
run_name=data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/new/diff_info_distill_forward_kl,
save_on_each_node=False,
save_only_model=False,
save_safetensors=True,
save_steps=100,
save_strategy=SaveStrategy.STEPS,
save_total_limit=1,
seed=1234,
skip_memory_metrics=True,
system_prompt=None,
teacher_model_init_kwargs=None,
teacher_model_name_or_path=None,
temperature=0.9,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torch_empty_cache_steps=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_cpu=False,
use_ipex=False,
use_legacy_prediction_loop=False,
use_liger=False,
use_liger_kernel=False,
use_mps_device=False,
wandb_entity=None,
wandb_project=None,
warmup_ratio=0.1,
warmup_steps=0,
weight_decay=0.0,
)
2025-07-21 17:36:34 - INFO - __main__ - *** Initializing model kwargs ***
2025-07-21 17:36:34 - INFO - __main__ - Model memory before loading model:Memory allocated: 0.0
Memory reserved: 0.0
2025-07-21 17:37:34 - INFO - __main__ - Model memory after loading model:Memory allocated: 0.0
Memory reserved: 0.0
2025-07-21 17:37:34 - INFO - __main__ - MoE layers replaced with Dense MLP layers
2025-07-21 17:37:34 - INFO - __main__ - Model memory after replacing MoE with dense:Memory allocated: 0.0
Memory reserved: 0.0
2025-07-21 17:37:34 - INFO - __main__ - Initializing EfficientDistillationTrainer...
2025-07-21 17:38:14 - INFO - __main__ - Model memory after trainer initialization:Memory allocated: 31126.0048828125
Memory reserved: 36896.0
2025-07-21 17:38:14 - INFO - __main__ - *** Starting training ***
2025-07-21 17:38:14 - INFO - __main__ - Model architecture: DeepseekV2ForCausalLM(
(model): DeepseekV2Model(
(embed_tokens): Embedding(102400, 2048)
(layers): ModuleList(
(0): DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=10944, bias=False)
(up_proj): Linear(in_features=2048, out_features=10944, bias=False)
(down_proj): Linear(in_features=10944, out_features=2048, bias=False)
(act_fn): SiLU()
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
(1-26): 26 x DeepseekV2DecoderLayer(
(self_attn): DeepseekV2FlashAttention2(
(q_proj): Linear(in_features=2048, out_features=3072, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=2048, out_features=576, bias=False)
(kv_a_layernorm): DeepseekV2RMSNorm()
(kv_b_proj): Linear(in_features=512, out_features=4096, bias=False)
(o_proj): Linear(in_features=2048, out_features=2048, bias=False)
(rotary_emb): DeepseekV2YarnRotaryEmbedding()
)
(mlp): DeepseekV2MoE(
(experts): ModuleList(
(0-63): 64 x DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=1408, bias=False)
(up_proj): Linear(in_features=2048, out_features=1408, bias=False)
(down_proj): Linear(in_features=1408, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(gate): MoEGate()
(shared_experts): DeepseekV2MLP(
(gate_proj): Linear(in_features=2048, out_features=2816, bias=False)
(up_proj): Linear(in_features=2048, out_features=2816, bias=False)
(down_proj): Linear(in_features=2816, out_features=2048, bias=False)
(act_fn): SiLU()
)
)
(input_layernorm): DeepseekV2RMSNorm()
(post_attention_layernorm): DeepseekV2RMSNorm()
)
)
(norm): DeepseekV2RMSNorm()
)
(lm_head): Linear(in_features=2048, out_features=102400, bias=False)
)