expert_selection.log

2025-06-30 08:15:30 - 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)
2025-06-30 08:15:30 - 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)
2025-06-30 08:15:30 - 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=3,
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-Distill-6-experts-test-token-specific-3-epoch-scaled-3times,
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/6_experts_test_token_specific_3_epoch_scaled_3times/runs/Jun30_08-15-29_ip-172-31-78-111,
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=600,
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/6_experts_test_token_specific_3_epoch_scaled_3times,
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/6_experts_test_token_specific_3_epoch_scaled_3times,
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,
)
2025-06-30 08:15:31 - INFO - __main__ - *** Initializing model kwargs ***
2025-06-30 08:15:31 - INFO - __main__ - Model memory in step 1, before model initialization (0):Memory allocated: 0.0
Memory reserved: 0.0
2025-06-30 08:15:54 - INFO - __main__ - Model memory in step 1, after model initialization:Memory allocated: 4836.39697265625
Memory reserved: 7322.0
2025-06-30 08:15:54 - INFO - __main__ - Model memory in step 2, before data collator initialization:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-06-30 08:15:54 - INFO - __main__ - Model memory in step 2, after data collator initialization:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-06-30 08:15:54 - INFO - __main__ - Model memory in step 4, before trainer initialization:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-06-30 08:17:44 - 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)
2025-06-30 08:17:44 - 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)
2025-06-30 08:17:44 - 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=3,
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-Distill-6-experts-token-specific-3-scaled,
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/6_experts_test_token_specific_3_epoch_scaled_3times/runs/Jun30_08-17-44_ip-172-31-78-111,
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=600,
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/6_experts_test_token_specific_3_epoch_scaled_3times,
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/6_experts_test_token_specific_3_epoch_scaled_3times,
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,
)
2025-06-30 08:17:45 - INFO - __main__ - *** Initializing model kwargs ***
2025-06-30 08:17:45 - INFO - __main__ - Model memory in step 1, before model initialization (0):Memory allocated: 0.0
Memory reserved: 0.0
2025-06-30 08:18:10 - INFO - __main__ - Model memory in step 1, after model initialization:Memory allocated: 4836.39697265625
Memory reserved: 7322.0
2025-06-30 08:18:10 - INFO - __main__ - Model memory in step 2, before data collator initialization:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-06-30 08:18:10 - INFO - __main__ - Model memory in step 2, after data collator initialization:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-06-30 08:18:10 - INFO - __main__ - Model memory in step 4, before trainer initialization:Memory allocated: 4836.39697265625
Memory reserved: 6442.0
2025-06-30 08:18:11 - INFO - __main__ - Model memory in step 4, after trainer initialization:Memory allocated: 4836.3974609375
Memory reserved: 6442.0
2025-06-30 08:18:11 - INFO - __main__ - Model memory in step 5, before prediction:Memory allocated: 4836.3974609375
Memory reserved: 6442.0
2025-06-30 08:18:11 - INFO - __main__ - Running prediction on test subset to record expert activations...
2025-06-30 08:18:21 - INFO - __main__ - Model memory in step 5, after prediction:Memory allocated: 3789.373046875
Memory reserved: 16034.0
2025-06-30 08:18:21 - INFO - __main__ - Top k experts selected: {'model.layers.1.mlp': [51, 61, 44, 45, 14, 22], 'model.layers.2.mlp': [27, 25, 18, 13, 3, 23], 'model.layers.3.mlp': [54, 25, 41, 23, 28, 57], 'model.layers.4.mlp': [37, 21, 33, 49, 11, 14], 'model.layers.5.mlp': [54, 47, 35, 20, 52, 9], 'model.layers.6.mlp': [22, 1, 13, 45, 42, 47], 'model.layers.7.mlp': [58, 43, 24, 18, 44, 62], 'model.layers.8.mlp': [47, 39, 56, 30, 54, 58], 'model.layers.9.mlp': [31, 13, 22, 24, 12, 32], 'model.layers.10.mlp': [47, 19, 42, 2, 13, 22], 'model.layers.11.mlp': [29, 11, 17, 10, 59, 22], 'model.layers.12.mlp': [5, 56, 3, 59, 4, 26], 'model.layers.13.mlp': [10, 42, 58, 14, 47, 17], 'model.layers.14.mlp': [51, 7, 27, 18, 31, 61], 'model.layers.15.mlp': [24, 55, 5, 17, 14, 41], 'model.layers.16.mlp': [61, 33, 63, 49, 19, 9], 'model.layers.17.mlp': [0, 26, 43, 32, 27, 29], 'model.layers.18.mlp': [5, 56, 42, 36, 2, 1], 'model.layers.19.mlp': [2, 23, 24, 36, 40, 0], 'model.layers.20.mlp': [1, 56, 38, 20, 48, 58], 'model.layers.21.mlp': [5, 13, 15, 28, 19, 10], 'model.layers.22.mlp': [58, 32, 31, 3, 45, 14], 'model.layers.23.mlp': [20, 0, 58, 45, 33, 42], 'model.layers.24.mlp': [62, 7, 42, 47, 10, 63], 'model.layers.25.mlp': [45, 48, 39, 11, 46, 38], 'model.layers.26.mlp': [46, 49, 6, 13, 11, 57]}
2025-06-30 08:18:21 - INFO - __main__ - Top k experts saved to: data/DeepSeek-Coder-V2-Lite-Instruct/distill/math10K/6_experts_test_token_specific_3_epoch_scaled_3times/top_6_experts_lmms-lab_Math10K.json
2025-06-30 08:18:21 - INFO - __main__ - Model memory before cleanup:Memory allocated: 3789.373046875
Memory reserved: 16034.0
2025-06-30 08:18:24 - INFO - __main__ - Model memory after cleanup:Memory allocated: 3788.38427734375
Memory reserved: 5794.0
2025-06-30 08:18:24 - INFO - __main__ - Expert selection completed successfully. Run part 2 for training.