LlamaForCausalLM(
  (model): LlamaModel(
    (embed_tokens): Embedding(49152, 576)
    (layers): ModuleList(
      (0-14): 15 x LlamaDecoderLayer(
        (self_attn): LlamaSdpaAttention(
          (q_proj): Linear(in_features=576, out_features=576, bias=False)
          (k_proj): Linear(in_features=576, out_features=192, bias=False)
          (v_proj): Linear(in_features=576, out_features=192, bias=False)
          (o_proj): Linear(in_features=576, out_features=576, bias=False)
          (rotary_emb): LlamaRotaryEmbedding()
        )
        (mlp): LigerSwiGLUMLP(
          (gate_proj): Linear(in_features=576, out_features=1536, bias=False)
          (up_proj): Linear(in_features=576, out_features=1536, bias=False)
          (down_proj): Linear(in_features=1536, out_features=576, bias=False)
        )
        (input_layernorm): LigerRMSNorm((576,), eps=1e-05, offset=0.0)
        (post_attention_layernorm): LigerRMSNorm((576,), eps=1e-05, offset=0.0)
      )
    )
    (norm): LigerRMSNorm((576,), eps=1e-05, offset=0.0)
    (rotary_emb): LlamaRotaryEmbedding()
  )
  (lm_head): Linear(in_features=576, out_features=49152, bias=False)
)

Resource Usage

Max Train VRAM Use: 13.4793 GB
Available VRAM: 23.6497 GB
GPUs:
- 1x NVIDIA GeForce RTX 4090
CPUs: 48
CPU Memory: 251.5386 GB
CPU Memory Bandwidth: 1200 GB/s

Distillation (Teacher -> Student) Architecture Difference:

Architecture: LlamaForCausalLM -> LlamaForCausalLM
Total Parameters: 134,515,008 -> 81,413,568
Data Type (dtype): torch.float32 -> torch.float32
Model Size: 0.25 GB -> 0.30 GB

Module Diff Details

diff

--- teacher model modules
+++ student model modules
@@ -2,7 +2,7 @@
   (model): LlamaModel(
     (embed_tokens): Embedding(49152, 576)
     (layers): ModuleList(
-      (0-29): 30 x LlamaDecoderLayer(
+      (0-14): 15 x LlamaDecoderLayer(
         (self_attn): LlamaSdpaAttention(
           (q_proj): Linear(in_features=576, out_features=576, bias=False)
           (k_proj): Linear(in_features=576, out_features=192, bias=False)
@@ -10,17 +10,16 @@
           (o_proj): Linear(in_features=576, out_features=576, bias=False)
           (rotary_emb): LlamaRotaryEmbedding()
         )
-        (mlp): LlamaMLP(
+        (mlp): LigerSwiGLUMLP(
           (gate_proj): Linear(in_features=576, out_features=1536, bias=False)
           (up_proj): Linear(in_features=576, out_features=1536, bias=False)
           (down_proj): Linear(in_features=1536, out_features=576, bias=False)
-          (act_fn): SiLU()
         )
-        (input_layernorm): LlamaRMSNorm((576,), eps=1e-05)
-        (post_attention_layernorm): LlamaRMSNorm((576,), eps=1e-05)
+        (input_layernorm): LigerRMSNorm((576,), eps=1e-05, offset=0.0)
+        (post_attention_layernorm): LigerRMSNorm((576,), eps=1e-05, offset=0.0)
       )
     )
-    (norm): LlamaRMSNorm((576,), eps=1e-05)
+    (norm): LigerRMSNorm((576,), eps=1e-05, offset=0.0)
     (rotary_emb): LlamaRotaryEmbedding()
   )
   (lm_head): Linear(in_features=576, out_features=49152, bias=False)

Train Dataset

Trained on 706,573,563 tokens from the wikimedia/wikipedia dataset.

Num Samples: 998,000
Subset: 20231101.en
Split: train

Training Objective

DistillationObjective(
    logits_loss_component=LossComponent(
        weight=1,
        loss_fn='kl'
    ),
    hs_loss_component=LossComponent(
        weight=0
    ),
    attn_loss_component=LossComponent(
        weight=0
    )
)

Hyperparameters

The following hyperparameters were used during training:

Expand

learning_rate: 0.0002
train_batch_size: 4
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 8
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: polynomial
num_epochs: 1.0
distillation_objective: DistillationObjective( logits_loss_component=LossComponent( weight=1, loss_fn='kl' ), hs_loss_component=LossComponent( weight=0 ), attn_loss_component=LossComponent( weight=0 ) )
lr_scheduler: <torch.optim.lr_scheduler.LambdaLR object at 0x718c02862f80>
student_model_name_or_path: None
student_config_name_or_path: None
student_model_config: {'num_hidden_layers': 15}
reinitialize_weights: None
copy_teacher_modules: [('lm_head', False)]
student_model_as_bitnet: False
student_use_liger_kernel: True
teacher_model_name_or_path: HuggingFaceTB/SmolLM-135M
teacher_load_in_8bit: False
teacher_load_in_4bit: False
dataset_uri: wikimedia/wikipedia
dataset_subset: 20231101.en
dataset_split: train
dataset_column_name: text
dataset_sample_size: 1000000
dataset_test_size: 0.002
dataset_shuffle: False
dataset_shuffle_seed: 42
dataset_trust_remote_code: False
weight_decay: 0.0
max_grad_norm: 1.0
warmup_ratio: 0.0
warmup_steps: 0
gradient_checkpointing: True

Framework Versions

Distily 0.5.0
Transformers 4.44.2
Pytorch 2.4.1+cu121
Datasets 3.0.0

distily/distily_distsmollm_long

Summary

Model Architecture:

Resource Usage

Distillation (Teacher -> Student) Architecture Difference:

Train Dataset

Training Objective

Hyperparameters

Framework Versions

distily/
distily_distsmollm_long