Fine-Tuning with Validation#

On this page, you’ll cover configuring and executing a fine-tuning run with some upstream validation. More specifically you’ll be fine-tuning a LLaMA3 8B model here as an example.

By the end, you should be comfortable kicking off your own fine-tuning run for the model of your choice.

Prerequisites#

You must have installed the Cerebras Model Zoo (click here if you haven’t).
You must be familiar with the Trainer and YAML format
Please ensure you have read Checkpointing
Please ensure you have read LLaMA3 8B pre-training

Configuring the Run#

This page will cover the two main flows you can employ to perform fine-tuning. One using a YAML configuration file and a training script that is packaged in the Cerebras ModelZoo. The other using pure Python to run on your own. They will be presented side-by-side so that you can compare the two flows as you progress through this tutorial.

Start with the Trainer configured in LLaMA3 8B pre-training.You will only need to make a few changes to this configuration to accomodate a fine-tuning run.

Fine-Tuning Using a Pre-trained Checkpoint#

To perform fine-tuning, a checkpoint from a previous training run is required. These checkpoints can be generated from previous runs or downloaded from online databases. For more information on porting a checkpoint from HuggingFace see Port a Hugging Face model to Cerebras Model Zoo. In this tutorial you will assume a checkpoint has already been generated after finishing Pretraining with Upstream Validation. For simplicty, let’s assume the checkpoint saved after the final step has the path:

./ckpts/checkpoint_10000.mdl

Configure Checkpoint State Loading#

To enable fine-tuning, you want to only load the model state from the checkpoint. Other checkpoint states such as the optimizer state or the training step should be reset.

To configure what states to load from the checkpoint, specify the following parameters to the callbacks key.

trainer:
  init:
    backend:  # CSX
      ...
    model:  # llama
      ...
    optimizer:  # AdamW
      ...
    schedulers:  # CosineDecayLR
      ...
    precision:  # DLS
      ...
    loop:
      ...
    checkpoint:
      ...
    callbacks:
      ...
      - LoadCheckpointStates:
          load_checkpoint_states: "model"
    ...

To configure what states to load from the checkpoint, construct a LoadCheckpointStates object as follows.

import cerebras.pytorch as cstorch
from cerebras.modelzoo import Trainer
from cerebras.modelzoo.trainer.callbacks import (
    CheckLoss,
    Checkpoint,
    ComputeNorm,
    LoadCheckpointStates,
    MixedPrecision,
    ModelEvalMetrics,
    TrainingLoop,
)
from cerebras.modelzoo.models.nlp.gpt2.model import Gpt2Model as LLaMA3


trainer = Trainer(
    backend=cstorch.backend("CSX", ...),
    model=lambda: LLaMA3(...),
    optimizer=lambda model: cstorch.optim.AdamW(...),
    schedulers=[...],
    precision=MixedPrecision(...),
    loop=TrainingLoop(...),
    checkpoint=Checkpoint(...),
    callbacks=[
        ...,
        LoadCheckpointStates("model"),
    ],
)

...

Load From a Checkpoint#

You now need to configure the trainer to load a checkpoint from a given path.

To configure the trainer to load from a given checkpoint, add the ckpt_path parameter to the fit key.

trainer:
  init:
    ...
  fit:
    train_dataloader:
      ...
    val_dataloader:
    - ...
    ckpt_path: ./ckpts/checkpoint_10000.mdl

To configure the trainer to load from a given checkpoint, specify ckpt_path in the Trainer’s fit method as follows.

import cerebras.pytorch as cstorch
from cerebras.modelzoo import Trainer

trainer = Trainer(...)
trainer.fit(
    train_dataloader=cstorch.utils.data.DataLoader(
        ...,
    ),
    val_dataloader=[
        ...,
    ],
    ckpt_path="./ckpts/checkpoint_10000.mdl",
)

Putting It All Together#

After the above adjustments, you should have a configuration that looks like this.

trainer:
  init:
    backend:
      backend_type: CSX
      cluster_config:
        num_csx: 16
    seed: 2024
    model:
      # Embedding
      vocab_size: 128256
      hidden_size: 4096
      position_embedding_type: "rotary"
      pos_scaling_factor: 1.0
      rope_theta: 500000.0
      rotary_dim: 128
      share_embedding_weights: false
      max_position_embeddings: 8192
      embedding_dropout_rate: 0.0
      embedding_layer_norm: false

      # Decoder
      num_hidden_layers: 32
      dropout_rate: 0.0
      layer_norm_epsilon: 1.0e-5
      norm_type: "rmsnorm"

      # Decoder - Attention
      num_heads: 32
      attention_type: "scaled_dot_product"
      attention_module: "multiquery_attention"
      attention_dropout_rate: 0.0
      use_projection_bias_in_attention: false
      use_ffn_bias_in_attention: false
      extra_attention_params:
          num_kv_groups: 8

      # Decoder - ffn
      filter_size: 14336
      nonlinearity: "swiglu"
      use_ffn_bias: false

      # Task-specific
      use_bias_in_output: false
      loss_scaling: "num_tokens"
      loss_weight: 1.0

      # Initializer
      initializer_range: 0.02

      # Cerebras parameters
      mixed_precision: True
      fp16_type: "cbfloat16"

    optimizer:
      AdamW:
        betas: [0.9, 0.95]
        correct_bias: True
        weight_decay: 0.1

    schedulers:
    - CosineDecayLR:
        initial_learning_rate: 3.0e-5
        end_learning_rate: 3.0e-6
        total_iters: 528

    precision:
      fp16_type: cbfloat16
      loss_scaling_factor: dynamic
      max_gradient_norm: 1.0

    loop:
      num_steps: 10000
      eval_frequency: 1000
      eval_steps: 1000

    checkpoint:
      steps: 1000

    callbacks:
    - ComputeNorm: {}
    - CheckLoss: {}
    - LoadCheckpointStates:
        load_checkpoint_states: "model"
    - ModelEvalMetrics: {}

    loggers:
    - ProgressLogger: {}
    - TensorBoardLogger: {}
  fit:
    train_dataloader:
      data_processor: GptHDF5MapDataProcessor
      data_dir: "/data/llama_v3_dataset_vocab128256/train"
      batch_size: 80
      micro_batch_size: 20
      shuffle: False
      shuffle_seed: 1337
      num_workers: 8
      prefetch_factor: 10
      persistent_workers: True # Important to avoid seeding at each epoch
    val_dataloader:
    - data_processor: GptHDF5MapDataProcessor
      data_dir: "/data/llama_v3_dataset_vocab128256/val"
      batch_size: 80
      micro_batch_size: 20
      shuffle: False
      shuffle_seed: 1337
      num_workers: 8
      prefetch_factor: 10
      persistent_workers: True # Important to avoid seeding at each epoch
    ckpt_path: ./ckpts/checkpoint_10000.mdl

import cerebras.pytorch as cstorch
from cerebras.modelzoo import Trainer
from cerebras.modelzoo.trainer.callbacks import (
    CheckLoss,
    Checkpoint,
    ComputeNorm,
    MixedPrecision,
    ModelEvalMetrics,
    TrainingLoop,
)
from cerebras.modelzoo.trainer.loggers import (
      ProgessLogger,
      TensorBoardLogger,
)
from cerebras.modelzoo.models.nlp.gpt2.model import Gpt2Model as LLaMA3


trainer = Trainer(
    backend=cstorch.backend(
        backend_type="CSX",
        cluster_config=cstorch.distributed.ClusterConfig(
            num_csx=16,
        ),
    ),
    seed=2024,
    model=lambda: LLaMA3(
        # Embedding
        vocab_size=128256,
        hidden_size=4096,
        position_embedding_type="rotary",
        pos_scaling_factor=1.0,
        rope_theta=500000.0,
        rotary_dim=128,
        share_embedding_weights=False,
        max_position_embeddings=8192,
        embedding_dropout_rate=0.0,
        embedding_layer_norm=False,

        # Decoder
        num_hidden_layers=32,
        dropout_rate=0.0,
        layer_norm_epsilon=1.0e-5,
        norm_type="rmsnorm",

        # Decoder - Attention
        num_heads=32,
        attention_type="scaled_dot_product",
        attention_module="multiquery_attention",
        attention_dropout_rate=0.0,
        use_projection_bias_in_attention=False,
        use_ffn_bias_in_attention=False,
        extra_attention_params=dict(
            num_kv_groups=8,
        ),

        # Decoder - ffn
        filter_size=14336,
        nonlinearity="swiglu",
        use_ffn_bias=False,

        # Task-specific
        use_bias_in_output=False,
        loss_scaling="num_tokens",
        loss_weight=1.0,

        # Initializer
        initializer_range=0.02,

        # Cerebras parameters
        mixed_precision=True,
        fp16_type="cbfloat16",
    ),
    optimizer=lambda model: cstorch.optim.AdamW(
        model.parameters(),
        lr=0.01,  # This is a placeholder
        betas=[0.9, 0.95],
        correct_bias=True,
        weight_decay: 0.1,
    ),
    schedulers=[
        lambda optimizer: cstorch.optim.lr_scheduler.CosineDecayLR(
            initial_learning_rate=3.0e-5,
            end_learning_rate=3.0e-6,
            total_iters=528,
        )
    ],
    precision=MixedPrecision(
        fp16_type="cbfloat16",
        loss_scaling_factor="dynamic",
        max_gradient_norm=1.0,
    ),
    loop=TrainingLoop(
      num_steps=10000,
      eval_frequency=1000,
      eval_steps=1000,
    ),
    checkpoint=Checkpoint(steps=1000),
    callbacks=[
        ComputeNorm(),
        CheckLoss(),
        LoadCheckpointStates("model"),
        ModelEvalMetrics(),
    ],
    loggers=[
        ProgessLogger(),
        TensorBoardLogger(),
    ],
)
trainer.fit(
    train_dataloader=cstorch.utils.data.DataLoader(
        registry.get_data_processor("GptHDF5MapDataProcessor"),
        data_dir="/data/llama_v3_dataset_vocab128256/train",
        batch_size=80,
        micro_batch_size=20,
        shuffle=False,
        shuffle_seed=1337,
        num_workers=8,
        prefetch_factor=10,
        persistent_workers=True, # Important to avoid seeding at each epoch
    ),
    val_dataloader=[
        cstorch.utils.data.DataLoader(
            registry.get_data_processor("GptHDF5MapDataProcessor"),
            data_dir="/data/llama_v3_dataset_vocab128256/val",
            batch_size=80,
            micro_batch_size=20,
            shuffle=False,
            shuffle_seed=1337,
            num_workers=8,
            prefetch_factor=10,
            persistent_workers=True, # Important to avoid seeding at each epoch
        ),
    ],
    ckpt_path="./ckpts/checkpoint_10000.mdl",
)

Start Fine-Tuning#

Now that you have a fully configured Trainer, all there is to do now is to kick off the run and start fine-tuning.

Let’s assume that the YAML configuration that you put together above is written to a file called ./finetune_llama_8b.yaml.

Then, to run fine-tuning using the training script that comes packaged as part of ModelZoo, you can run the following on the command line

python modelzoo/models/nlp/llama/run.py CSX \
    --mode train_and_eval \
    --params ./finetune_llama_8b.yaml \

Let’s assume that the python code that we put together above is written to a file called ./finetune_llama_8b.py.

Then, to run fine-tuning, all there is to do is to execute that python script.

python ./finetune_llama_8b.py

Monitoring the Run#

Once compilation finishes and the Wafer-Scale Cluster is programmed for execution, you should start seeing progress logs that look like

| Train Device=CSX, Step=1, Loss=1.41992, Rate=16.30 samples/sec, GlobalRate=16.30 samples/sec
| Train Device=CSX, Step=2, Loss=1.41016, Rate=20.40 samples/sec, GlobalRate=19.13 samples/sec
| Train Device=CSX, Step=3, Loss=1.39062, Rate=21.93 samples/sec, GlobalRate=20.25 samples/sec
| Train Device=CSX, Step=4, Loss=1.38281, Rate=22.45 samples/sec, GlobalRate=20.84 samples/sec
| Train Device=CSX, Step=5, Loss=1.36719, Rate=22.57 samples/sec, GlobalRate=21.17 samples/sec
| Train Device=CSX, Step=6, Loss=1.42188, Rate=22.54 samples/sec, GlobalRate=21.39 samples/sec
| Train Device=CSX, Step=7, Loss=1.39258, Rate=22.52 samples/sec, GlobalRate=21.54 samples/sec
| Train Device=CSX, Step=8, Loss=1.36914, Rate=22.44 samples/sec, GlobalRate=21.64 samples/sec
| Train Device=CSX, Step=9, Loss=1.37695, Rate=22.33 samples/sec, GlobalRate=21.71 samples/sec
| Train Device=CSX, Step=10, Loss=1.35938, Rate=22.38 samples/sec, GlobalRate=21.78 samples/sec
| Train Device=CSX, Step=11, Loss=1.36719, Rate=21.32 samples/sec, GlobalRate=21.67 samples/sec
| Train Device=CSX, Step=12, Loss=1.39844, Rate=21.72 samples/sec, GlobalRate=21.69 samples/sec
| Train Device=CSX, Step=13, Loss=1.38672, Rate=21.88 samples/sec, GlobalRate=21.71 samples/sec
| Train Device=CSX, Step=14, Loss=1.34961, Rate=21.91 samples/sec, GlobalRate=21.73 samples/sec
| Train Device=CSX, Step=15, Loss=1.33203, Rate=21.88 samples/sec, GlobalRate=21.74 samples/sec
| Train Device=CSX, Step=16, Loss=1.33008, Rate=21.91 samples/sec, GlobalRate=21.75 samples/sec
| Train Device=CSX, Step=17, Loss=1.33984, Rate=21.88 samples/sec, GlobalRate=21.76 samples/sec
| Train Device=CSX, Step=18, Loss=1.31250, Rate=21.88 samples/sec, GlobalRate=21.76 samples/sec
| Train Device=CSX, Step=19, Loss=1.36133, Rate=21.91 samples/sec, GlobalRate=21.77 samples/sec
| Train Device=CSX, Step=20, Loss=1.30664, Rate=23.15 samples/sec, GlobalRate=21.87 samples/sec
| Train Device=CSX, Step=21, Loss=1.30078, Rate=22.52 samples/sec, GlobalRate=21.88 samples/sec
| Train Device=CSX, Step=22, Loss=1.31250, Rate=22.23 samples/sec, GlobalRate=21.89 samples/sec
| Train Device=CSX, Step=23, Loss=1.30664, Rate=21.10 samples/sec, GlobalRate=21.82 samples/sec
| Train Device=CSX, Step=24, Loss=1.30469, Rate=22.73 samples/sec, GlobalRate=21.90 samples/sec
| Train Device=CSX, Step=25, Loss=1.28906, Rate=21.42 samples/sec, GlobalRate=21.84 samples/sec
...
| Eval Device=CSX, GlobalStep=1000, Batch=1, Loss=1.21875, Rate=21.47 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=2, Loss=1.24219, Rate=22.65 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=3, Loss=1.26562, Rate=22.06 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=4, Loss=1.25195, Rate=21.90 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=5, Loss=1.27539, Rate=21.80 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=6, Loss=1.23047, Rate=21.79 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=7, Loss=1.22852, Rate=20.72 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=8, Loss=1.27734, Rate=21.24 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=9, Loss=1.23633, Rate=22.57 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=10, Loss=1.27930, Rate=22.10 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=11, Loss=1.23438, Rate=20.86 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=12, Loss=1.24609, Rate=21.31 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=13, Loss=1.23633, Rate=21.47 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=14, Loss=1.23633, Rate=21.48 samples/sec, GlobalRates=21.66 samples/sec
| Eval Device=CSX, GlobalStep=1000, Batch=15, Loss=1.21680, Rate=22.66 samples/sec, GlobalRates=21.66 samples/sec
...

Note

The performance numbers that you get will vary depending on how many Cerebras systems you are using and which generation systems you are using.

If you open up the TensorBoard you can more closely monitor the run be observing the trends in the graphs of the various logged metrics.

tensorboard --bind_all --logdir="./model_dir"

As can be seen above, the screenshots were taken at around step 8000. At this point you can observe that so far, the run seems to progressing well. The losses appear to be trending downwards and the model wise gradient norms don’t appear overly abnormal.

Porting the Model to Hugging Face#

Once the fine-tuning run has finished, you can port the model and checkpoint to Hugging Face.

To learn more about how to do this, see Port a trained and fine-tuned model to Hugging Face.

Conclusion#

With that, you have completed your first fine-tuning run with validation on the Cerebras Wafer-Scale Cluster using the ModelZoo Trainer!

By now, you should understand how to write your own Trainer configuration and how to kick off a training job from a checkpoint on the Cerebras Wafer-Scale Cluster. You can now take this knowledge and fine-tune your very own model.

Pretraining with Downstream Validation

Multi-Phase Training