> ## Documentation Index
> Fetch the complete documentation index at: https://training-docs.cerebras.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Convert Legacy to Trainer YAML
> Learn how to convert legacy YAMLs into the new Trainer YAML configuration.
## Converter Tool
The YAML specification changed after version 2.2 to make use of the Trainer class. You can use the converter tool to convert any legacy YAML to work with later versions:
```bash ModelZoo CLI theme={null}
cszoo config convert_legacy /path/to/v1.yaml -o /path/to/v2.yaml
```
```python Python theme={null}
import yaml
from cerebras.modelzoo.trainer.utils import (
convert_legacy_params_to_trainer_params
)
with open("/path/to/legacy/params.yaml") as f:
legacy_params = yaml.load(f)
trainer_params = convert_legacy_params_to_trainer_params(
legacy_params
)
with open("/path/to/trainer/params.yaml", "w") as f:
yaml.dump(trainer_params, f)
```
## Parameter Index
If you have a Legacy YAML specification and want to find out how to specify a specific parameter in the Trainer YAML specification, please use the tabs below to find the Legacy and Trainer specifications for each parameter.
## `eval_input`
```yaml Legacy Spec theme={null}
eval_input:
...
```
```yaml Trainer Spec theme={null}
trainer:
fit:
val_dataloader:
...
trainer:
validate:
val_dataloader:
...
trainer:
validate_all:
val_dataloaders:
...
```
## `eval_input.micro_batch_size`
```yaml Legacy Spec theme={null}
eval_input:
micro_batch_size: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- ScopedValidateFlags:
csx.performance.micro_batch_size: ...
```
## `model`
```yaml Legacy Spec theme={null}
model:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
model:
...
```
## `model.compression`
```yaml Legacy Spec theme={null}
model:
compression:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- WeightCompression:
compressions:
...
```
## `model.fp16_type`
```yaml Legacy Spec theme={null}
model:
fp16_type: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
fp16_type: ...
```
## `model.lora_params`
```yaml Legacy Spec theme={null}
model:
lora_params: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
-Lora:
lora_params:
...
```
## `model.mixed_precision`
```yaml Legacy Spec theme={null}
model:
mixed_precision: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
enabled: ...
```
## `model.selective_grad`
```yaml Legacy Spec theme={null}
model:
selective_grad ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- SelectiveGrad:
selective_grads:
...
```
## `optimizer`
```yaml Legacy Spec theme={null}
optimizer:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
optimizer:
...
```
## `optimizer.grad_accum_steps`
```yaml Legacy Spec theme={null}
optimizer:
grad_accum_steps: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loop:
grad_accum_steps: ...
```
## `optimizer.initial_loss_scale`
```yaml Legacy Spec theme={null}
optimizer:
initial_loss_scale: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
initial_loss_scale: ...
```
## `optimizer.learning_rate`
```yaml Legacy Spec theme={null}
optimizer:
learning_rate:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
schedulers:
...
```
## `optimizer.log_summaries`
```yaml Legacy Spec theme={null}
optimizer:
log_summaries: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- ComputeNorm: ...
trainer:
init:
precision:
log_loss_scale: ...
```
## `optimizer.loss_scaling_factor`
```yaml Legacy Spec theme={null}
optimizer:
loss_scaling_factor: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
loss_scaling_factor: ...
```
## `optimizer.max_gradient_norm`
```yaml Legacy Spec theme={null}
optimizer:
max_gradient_norm: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
max_gradient_norm: ...
```
## `optimizer.max_gradient_value`
```yaml Legacy Spec theme={null}
optimizer:
max_gradient_value: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
max_gradient_value: ...
```
## `optimizer.max_loss_scale`
```yaml Legacy Spec theme={null}
optimizer:
max_loss_scale: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
max_loss_scale: ...
```
## `optimizer.min_loss_scale`
```yaml Legacy Spec theme={null}
optimizer:
min_loss_scale: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
min_loss_scale: ...
```
## `optimizer.steps_per_increase`
```yaml Legacy Spec theme={null}
optimizer:
steps_per_increase: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
steps_per_increase: ...
```
## `runconfig.act_memory_gi`
```yaml Legacy Spec theme={null}
runconfig:
act_memory_gi: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.act_memory_gi: ...
```
## `runconfig.autoload_last_checkpoint`
```yaml Legacy Spec theme={null}
runconfig:
autoload_last_checkpoint: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
checkpoint:
autoload_last_checkpoint: ...
```
## `runconfig.check_loss_values`
```yaml Legacy Spec theme={null}
runconfig:
check_loss_values: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- CheckLoss: ...
```
## `runconfig.checkpoint_path`
```yaml Legacy Spec theme={null}
runconfig:
checkpoint_path: ...
```
```yaml Trainer Spec theme={null}
trainer:
fit:
ckpt_path: ...
trainer:
validate:
ckpt_path: ...
trainer:
validate_all:
ckpt_paths: ...
```
## `runconfig.checkpoint_steps`
```yaml Legacy Spec theme={null}
runconfig:
checkpoint_steps: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
checkpoint:
steps: ...
```
## `runconfig.cmd_memory_gi`
```yaml Legacy Spec theme={null}
runconfig:
cmd_memory_gi: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.cmd_memory_gi: ...
```
## `runconfig.compile_crd_memory_gi`
```yaml Legacy Spec theme={null}
runconfig:
compile_crd_memory_gi: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.compile_crd_memory_gi: ...
```
## `runconfig.compile_dir`
```yaml Legacy Spec theme={null}
runconfig:
compile_dir: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
compile_dir: ...
```
## `runconfig.compile_only`
```yaml Legacy Spec theme={null}
runconfig:
compile_only: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
compile_only: ...
```
## `runconfig.credentials_path`
```yaml Legacy Spec theme={null}
runconfig:
credentials_path: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
credentials_path: ...
```
## `runconfig.debug_args`
```yaml Legacy Spec theme={null}
runconfig:
debug_args:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.debug_args:
...
```
## `runconfig.debug_args_path`
```yaml Legacy Spec theme={null}
runconfig:
debug_args_path: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- DebugArgsPath:
debug_args_path: ...
```
## `runconfig.disable_strict_checkpoint_loading`
```yaml Legacy Spec theme={null}
runconfig:
disable_strict_checkpoint_loading: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
checkpoint:
disable_strict_checkpoint_loading: ...
```
## `runconfig.disable_version_check`
```yaml Legacy Spec theme={null}
runconfig:
disable_version_check: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
disable_version_check: ...
```
## `runconfig.dist_backend`
```yaml Legacy Spec theme={null}
runconfig:
dist_backend: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
dist_backend: ...
```
## `runconfig.drop_data`
```yaml Legacy Spec theme={null}
runconfig:
drop_data: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.drop_data: ...
```
## `runconfig.dump_activations`
```yaml Legacy Spec theme={null}
runconfig:
dump_activations: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- DumpActivations: ...
```
## `runconfig.enable_act_frequency`
```yaml Legacy Spec theme={null}
runconfig:
enable_act_frequency: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
logging:
enable_act_frequency: ...
```
## `runconfig.enable_distributed`
```yaml Legacy Spec theme={null}
runconfig:
enable_distributed: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
enable_distributed: ...
```
## `runconfig.eval_frequency`
```yaml Legacy Spec theme={null}
runconfig:
eval_frequency: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loop:
eval_frequency: ...
```
## `runconfig.eval_steps`
```yaml Legacy Spec theme={null}
runconfig:
eval_steps: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loop:
eval_steps: ...
```
## `runconfig.execute_crd_memory_gi`
```yaml Legacy Spec theme={null}
runconfig:
execute_crd_memory_gi: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.execute_crd_memory_gi: ...
```
## `runconfig.experimental.listeners`
```yaml Legacy Spec theme={null}
runconfig:
experimental:
listeners:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- Listener:
listeners:
...
```
## `runconfig.init_method`
```yaml Legacy Spec theme={null}
runconfig:
init_method: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
init_method: ...
```
## `runconfig.job_labels`
```yaml Legacy Spec theme={null}
runconfig:
job_labels:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
job_labels:
...
```
## `runconfig.job_priority`
```yaml Legacy Spec theme={null}
runconfig:
job_priority: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
job_priority: ...
```
## `runconfig.job_time_sec`
```yaml Legacy Spec theme={null}
runconfig:
job_time_sec: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
job_time_sec: ...
```
## `runconfig.lazy_initialization`
```yaml Legacy Spec theme={null}
runconfig:
lazy_initialization: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.lazy_initialization: ...
```
## `runconfig.load_checkpoint_states`
```yaml Legacy Spec theme={null}
runconfig:
load_checkpoint_states: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- LoadCheckpointStates:
load_checkpoint_states: ...
```
## `runconfig.log_initialization`
```yaml Legacy Spec theme={null}
runconfig:
log_initialization: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.log_initialization: ...
```
## `runconfig.log_input_summaries`
```yaml Legacy Spec theme={null}
runconfig:
log_input_summaries: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- LogInputSummaries: ...
```
## `runconfig.log_steps`
```yaml Legacy Spec theme={null}
runconfig:
log_steps: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
logging:
log_steps: ...
```
## `runconfig.logging`
```yaml Legacy Spec theme={null}
runconfig:
logging: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
logging:
log_level: ...
```
## `runconfig.main_process_id`
```yaml Legacy Spec theme={null}
runconfig:
main_process_id: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
main_process_id: ...
```
## `runconfig.max_checkpoints`
```yaml Legacy Spec theme={null}
runconfig:
max_checkpoints: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- KeepNCheckpoints:
n: ...
```
## `runconfig.max_steps`
```yaml Legacy Spec theme={null}
runconfig:
max_steps: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loop:
max_steps: ...
```
## `runconfig.mgmt_address`
```yaml Legacy Spec theme={null}
runconfig:
mgmt_address: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
mgmt_address: ...
```
## `runconfig.mgmt_namespace`
```yaml Legacy Spec theme={null}
runconfig:
mgmt_namespace: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
mgmt_namespace: ...
```
## `runconfig.model_dir`
```yaml Legacy Spec theme={null}
runconfig:
model_dir: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
model_dir: ...
```
## `runconfig.mount_dirs`
```yaml Legacy Spec theme={null}
runconfig:
mount_dirs:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
mount_dirs:
...
```
## `runconfig.num_act_servers`
```yaml Legacy Spec theme={null}
runconfig:
num_act_servers: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
max_act_per_csx: ...
```
## `runconfig.num_csx`
```yaml Legacy Spec theme={null}
runconfig:
num_csx: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
num_csx: ...
```
## `runconfig.num_epochs`
```yaml Legacy Spec theme={null}
runconfig:
num_epochs: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loop:
num_epochs: ...
```
## `runconfig.num_steps`
```yaml Legacy Spec theme={null}
runconfig:
num_steps: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loop:
num_steps: ...
```
## `runconfig.num_wgt_servers`
```yaml Legacy Spec theme={null}
runconfig:
num_wgt_servers: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
max_wgt_servers: ...
```
## `runconfig.num_workers_per_csx`
```yaml Legacy Spec theme={null}
runconfig:
num_workers_per_csx: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
num_workers_per_csx: ...
```
## `runconfig.op_profiler_config`
```yaml Legacy Spec theme={null}
runconfig:
op_profiler_config:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- OpProfiler:
...
```
## `runconfig.precision_opt_level`
```yaml Legacy Spec theme={null}
runconfig:
precision_opt_level: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
precision:
precision_opt_level: ...
```
## `runconfig.python_paths`
```yaml Legacy Spec theme={null}
runconfig:
python_paths:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
cluster_config:
python_paths:
...
```
## `runconfig.retrace_every_iteration`
```yaml Legacy Spec theme={null}
runconfig:
retrace_every_iteration: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.retrace_every_iteration: ...
```
## `runconfig.save_initial_checkpoint`
```yaml Legacy Spec theme={null}
runconfig:
save_initial_checkpoint: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
checkpoint:
save_initial_checkpoint: ...
```
## `runconfig.seed`
```yaml Legacy Spec theme={null}
runconfig:
seed: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
seed: ...
```
## `runconfig.steps_per_epoch`
```yaml Legacy Spec theme={null}
runconfig:
steps_per_epoch: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loop:
steps_per_epoch: ...
```
## `runconfig.sync_batchnorm`
```yaml Legacy Spec theme={null}
runconfig:
sync_batchnorm: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
sync_batchnorm: ...
```
## `runconfig.target_device`
```yaml Legacy Spec theme={null}
runconfig:
target_device: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
backend_type: ...
```
## `runconfig.transfer_processes`
```yaml Legacy Spec theme={null}
runconfig:
transfer_processes: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.performance.transfer_processes: ...
```
## `runconfig.validate_only`
```yaml Legacy Spec theme={null}
runconfig:
validate_only: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
backend:
validate_only: ...
```
## `runconfig.wgt_memory_gi`
```yaml Legacy Spec theme={null}
runconfig:
wgt_memory_gi: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.wgt_memory_gi: ...
```
## `runconfig.wrk_memory_gi`
```yaml Legacy Spec theme={null}
runconfig:
wrk_memory_gi: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- GlobalFlags:
csx.debug.wrk_memory_gi: ...
```
## `runconfig.wsc_log_level`
```yaml Legacy Spec theme={null}
runconfig:
wsc_log_level:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
logging:
wsc_log_level:
...
```
## `sparsity`
```yaml Legacy Spec theme={null}
sparsity:
...
```
```yaml Trainer Spec theme={null}
trainer:
init:
sparsity:
...
```
## `sparsity.add_summaries`
```yaml Legacy Spec theme={null}
sparsity:
add_summaries: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- LogSparsity: ...
```
## `train_input`
```yaml Legacy Spec theme={null}
train_input:
...
```
```yaml Trainer Spec theme={null}
trainer:
fit:
train_dataloader:
...
```
## `train_input.micro_batch_size`
```yaml Legacy Spec theme={null}
train_input:
micro_batch_size: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
callbacks:
- ScopedTrainFlags:
csx.performance.micro_batch_size: ...
```
## `wandb.group`
```yaml Legacy Spec theme={null}
wandb:
group: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loggers:
- WandbLogger:
group: ...
```
## `wandb.job_type`
```yaml Legacy Spec theme={null}
wandb:
job_type: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loggers:
- WandbLogger:
job_type: ...
```
## `wandb.project`
```yaml Legacy Spec theme={null}
wandb:
project: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loggers:
- WandbLogger:
project: ...
```
## `wandb.resume`
```yaml Legacy Spec theme={null}
wandb:
resume: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loggers:
- WandbLogger:
resume: ...
```
## `wandb.run_id`
```yaml Legacy Spec theme={null}
wandb:
run_id: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loggers:
- WandbLogger:
run_id: ...
```
## `wandb.run_name`
```yaml Legacy Spec theme={null}
wandb:
run_name: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loggers:
- WandbLogger:
run_name: ...
```
## `wandb.tags`
```yaml Legacy Spec theme={null}
wandb:
tags: ...
```
```yaml Trainer Spec theme={null}
trainer:
init:
loggers:
- WandbLogger:
tags: ...
```