Skip to content

Commit

Permalink
Add llama3 ft example yamls (#1686)
Browse files Browse the repository at this point in the history
Co-authored-by: Chuck Tang <[email protected]>
Co-authored-by: Saaketh Narayan <[email protected]>
  • Loading branch information
3 people authored Dec 5, 2024
1 parent ff3d901 commit 05563e1
Show file tree
Hide file tree
Showing 2 changed files with 169 additions and 8 deletions.
19 changes: 11 additions & 8 deletions mcli/mcli-llama2-finetune.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -21,9 +21,12 @@ compute:

# The below is injected as a YAML file: /mnt/config/parameters.yaml
parameters:
tokenizer_name: meta-llama/Llama-2-7b-hf
max_seq_len: 4096
global_seed: 17
variables:
tokenizer_name: meta-llama/Llama-2-7b-hf
global_seed: 17
max_seq_len: 4096

max_seq_len: ${variables.max_seq_len}

# Run Name
run_name: # If left blank, will be read from env var $RUN_NAME
Expand All @@ -42,17 +45,17 @@ parameters:

# Tokenizer
tokenizer:
name: ${tokenizer_name}
name: ${variables.tokenizer_name}
kwargs:
model_max_length: ${max_seq_len}
model_max_length: ${variables.max_seq_len}

# Dataloaders
train_loader:
name: finetuning
dataset:
hf_name: mosaicml/dolly_hhrlhf
split: train
max_seq_len: ${max_seq_len}
max_seq_len: ${variables.max_seq_len}
allow_pad_trimming: false
decoder_only_format: true
shuffle: true
Expand All @@ -75,7 +78,7 @@ parameters:
dataset:
hf_name: mosaicml/dolly_hhrlhf
split: test
max_seq_len: ${max_seq_len}
max_seq_len: ${variables.max_seq_len}
allow_pad_trimming: false
decoder_only_format: true
# packing_ratio:
Expand Down Expand Up @@ -114,7 +117,7 @@ parameters:
global_train_batch_size: 64

# System
seed: ${global_seed}
seed: ${variables.global_seed}
device_eval_batch_size: 8
device_train_microbatch_size: auto
precision: amp_bf16
Expand Down
158 changes: 158 additions & 0 deletions mcli/mcli-llama3-70b-instruct-finetune.yaml
Original file line number Diff line number Diff line change
@@ -0,0 +1,158 @@
integrations:
- integration_type: git_repo
git_repo: mosaicml/llm-foundry
git_branch: v0.15.0
# git_commit: # OR use your commit hash
pip_install: .[gpu]
ssh_clone: false # Should be true if using a private repo

command: |
cd llm-foundry/scripts
composer train/train.py /mnt/config/parameters.yaml
image: mosaicml/llm-foundry:2.5.1_cu124-latest
name: llama3.1-70b-finetune

compute:
# Note: Finetuning the 70b model requires at least 16x80GB GPUs
gpus: 16 # Number of GPUs to use
## These configurations are optional
# cluster: TODO # Name of the cluster to use for this run
# gpu_type: h100_80gb # Type of GPU to use. We use h100_80gb in our experiments

# The below is injected as a YAML file: /mnt/config/parameters.yaml
parameters:
variables:
tokenizer_name: meta-llama/Llama-3.1-70B-Instruct
global_seed: 17
max_seq_len: 4096

max_seq_len: ${variables.max_seq_len}
# Run Name
run_name: # If left blank, will be read from env var $RUN_NAME

max_split_size_mb: 512

# Model
model:
name: hf_causal_lm
init_device: mixed
pretrained_model_name_or_path: meta-llama/Llama-3.1-70B-Instruct
pretrained: true
# Note: you must have set the HF_TOKEN environment variable and have access to the llama3 models
use_auth_token: true
use_flash_attention_2: true

# Tokenizer
tokenizer:
name: ${variables.tokenizer_name}
kwargs:
model_max_length: ${variables.max_seq_len}
# Dataloaders
train_loader:
name: finetuning
dataset:
hf_name: mosaicml/dolly_hhrlhf
split: train
max_seq_len: ${variables.max_seq_len}
allow_pad_trimming: false
decoder_only_format: true
shuffle: true
# # Use packing_ratio: 'auto' to automatically profile and select the highest observed packing ratio with
# # zero waste. In practice, this may result in > 0 waste because profiling is done on only a portion
# # of the dataset.
# # Or use `python llmfoundry/scripts/misc/profile_packing.py --yaml-path /path/to/this/yaml/ ...`
# # to profile this run's optimal packing_ratio as it depends on GPU count,
# # batch size, sequence length
# packing_ratio: auto
drop_last: true
num_workers: 8
pin_memory: false
prefetch_factor: 2
persistent_workers: true
timeout: 0

eval_loader:
name: finetuning
dataset:
hf_name: mosaicml/dolly_hhrlhf
split: test
max_seq_len: ${variables.max_seq_len}
allow_pad_trimming: false
decoder_only_format: true
# packing_ratio:
shuffle: false
drop_last: true
num_workers: 8
pin_memory: false
prefetch_factor: 2
persistent_workers: true
timeout: 0

# Optimization
scheduler:
name: cosine_with_warmup
t_warmup: 100ba
alpha_f: 0.1

# Note: You may want to change learning rate, betas, weight decay
optimizer:
name: decoupled_lionw
lr: 5.0e-7
betas:
- 0.9
- 0.95
weight_decay: 0.0

algorithms:
gradient_clipping:
clipping_type: norm
clipping_threshold: 1.0

max_duration: 1ep
eval_first: false
eval_interval: 1ep
eval_subset_num_batches: -1
global_train_batch_size: 16

# System
seed: ${variables.global_seed}
device_eval_batch_size: 1
device_train_microbatch_size: 1
precision: amp_bf16

# FSDP
fsdp_config:
state_dict_type: sharded # Note: we enable sharded checkpointing to avoid GPU OOM
sharding_strategy: FULL_SHARD
mixed_precision: PURE
activation_checkpointing: true
activation_checkpointing_reentrant: false
activation_cpu_offload: false
limit_all_gathers: true

# Logging
progress_bar: false
log_to_console: true
console_log_interval: 1ba

callbacks:
speed_monitor:
window_size: 10
lr_monitor: {}
memory_monitor: {}
runtime_estimator: {}

load_weights_only: true # Only load the weights, not the optimizer state, LR schedule, etc

# loggers:
# wandb: {}

# Checkpoint to local filesystem or remote object store
# save_interval: 2000ba
# save_num_checkpoints_to_keep: 1 # Important, this cleans up checkpoints saved to DISK
# save_folder: ./{run_name}/checkpoints
# save_folder: s3://my-bucket/my-folder/{run_name}/checkpoints

# Load from local filesystem or remote object store
# load_path: ./gpt-1b/checkpoints/latest-rank{rank}.pt
# load_path: s3://my-bucket/my-folder/gpt-1b/checkpoints/latest-rank{rank}.pt

0 comments on commit 05563e1

Please sign in to comment.