Mistral Nemo LoRA training has super high grad_norm

[Nero10578]

Please check that this issue hasn't been reported before.

• I searched previous Bug Reports didn't find any similar reports.

Expected Behavior

Before the gradient accumulation fixes and changes with transformers recently, the grad_norm when training Mistral Nemo 12B was below 1.0 like normal. Could also be because of changes to using chat_templates?

This was using the same config with previous versions of axolotl and transformers:

Current behaviour

Gradient Normalization is now around 5 when training:

Steps to reproduce

Train Mistral Nemo 12B Instruct with LoRA. I used the same config as I did back then when this works fine.

Only difference is now I am using chat_templates, where I replace the chat template in the Mistral tokenizer_config.json with the chat template shown here so that it can accept repeating same roles.

{%- for message in messages %}
      {%- if message['role'] == 'system' -%}
          {{- message['content'] -}}
      {%- else -%}
          {%- if message['role'] == 'user' -%}
              {{-'[INST] ' + message['content'].rstrip() + ' [/INST]'-}}
          {%- else -%}
              {{-'' + message['content'] + '</s>' -}}
          {%- endif -%}
      {%- endif -%}
  {%- endfor -%}
  {%- if add_generation_prompt -%}
      {{-''-}}
  {%- endif -%}

I did this by changing the chat templates in the Mistral Nemo tokenizer config to this:

  "chat_template": "{%- for message in messages %}{%- if message['role'] == 'system' -%}{{- message['content'] -}}{%- else -%}{%- if message['role'] == 'user' -%}{{-'[INST] ' + message['content'].rstrip() + ' [/INST]'-}}{%- else -%}{{-'' + message['content'] + '</s>' -}}{%- endif -%}{%- endif -%}{%- endfor -%}{%- if add_generation_prompt -%}{{-''-}}{%- endif -%}",

If this is the wrong way to do it, that might be causing the high grad_norm? But that seems unlikely since the dataset seems to be tokenized properly when I use preprocess --debug.

Config yaml

base_model: /home/user/models/Mistral-Nemo-Instruct-2407
model_type: AutoModelForCausalLM

train_on_inputs: false
group_by_length: false
load_in_8bit:
load_in_4bit: false
strict: false
sequence_len: 8192
bf16: auto
flash_attention: true

shuffle_merged_datasets: true

# Data
datasets:
  - path: /home/user/datasets/conversations-escaped.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value

warmup_steps: 10
dataset_prepared_path: ./lora_last_run_prepared

# Iterations
num_epochs: 1
saves_per_epoch: 8
saves_total_limit: 8

# Evaluation
val_set_size: 0.0025
eval_max_new_tokens: 128
eval_sample_packing: false
evals_per_epoch: 8
eval_table_size:

# LoRA
output_dir: ./lora_out1
adapter: lora
lora_model_dir:
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true
loraplus_lr_ratio: 16
save_safetensors: true

# Sampling
sample_packing: true
pad_to_sequence_len: true

# Batching
gradient_accumulation_steps: 16
micro_batch_size: 1
gradient_checkpointing: unsloth

# wandb
wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
wandb_project: mistral-nemo
wandb_entity: # A wandb Team name if using a Team
wandb_watch:
wandb_name: nemo-v1.3-8192
wandb_run_id: # Set the ID of your wandb run
wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training

# Optimizer
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00001

# Misc
auto_resume_from_checkpoints: true
logging_steps: 1
weight_decay: 0.0

special_tokens:
  pad_token: <pad>

# Multi-GPU
deepspeed:
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: MistralDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD

Possible solution

No response

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

3.11

axolotl branch-commit

main/db51a9e4

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this bug has not been reported yet.
• I am using the latest version of axolotl.
• I have provided enough information for the maintainers to reproduce and diagnose the issue.

[NanoCode012]

Hey, thanks for the report.

Could also be because of changes to using chat_templates?

Are you able to test one with chat_template before the GA patch (making sure you install the old requirements)?

[winglian]

@Nero10578 btw, I say you're using FSDP, how many GPUs? thanks

[winglian]

It could also be that the grad norm is higher because it doesn't expect the repeating of roles

[Nero10578]

Hey, thanks for the report.

Could also be because of changes to using chat_templates?

Are you able to test one with chat_template before the GA patch (making sure you install the old requirements)?

Will have to test, but I suspect this is the source of the issue.

@Nero10578 btw, I say you're using FSDP, how many GPUs? thanks

I am using 2x3090Ti with FSDP with CPU offloading.

It could also be that the grad norm is higher because it doesn't expect the repeating of roles

I don't think so because the dataset didn't change and it was fine before.

[Nero10578]

So tested it out with the commit before the GA patch 718cfb2 and with the latest commit 724b660. These are the differences in grad_norm:

Before GA patch:

After GA Patch:

The thing is that now that I have let both run, the GA patch does make the loss better, can be seen in how the first eval shows lower loss after the GA patch.

Before GA patch:

After GA Patch:

I am using a different config with Liger kernels enabled:

base_model: /home/user/models/Mistral-Nemo-Instruct-2407
model_type: AutoModelForCausalLM

train_on_inputs: false
group_by_length: false
load_in_8bit:
load_in_4bit: false
strict: false
sequence_len: 8192
bf16: auto
flash_attention: true

shuffle_merged_datasets: true

# Data
datasets:
  - path: /home/user/datasets/conversations-escaped.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value

warmup_steps: 10
dataset_prepared_path: ./lora_last_run_prepared

# Iterations
num_epochs: 1
saves_per_epoch: 8
saves_total_limit: 8

# Evaluation
val_set_size: 0.0025
eval_max_new_tokens: 128
eval_sample_packing: false
evals_per_epoch: 8
eval_table_size:

# LoRA
output_dir: ./lora_out3
adapter: lora
lora_model_dir:
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true
peft_use_rslora: false
loraplus_lr_ratio: 16
save_safetensors: true

# Sampling
sample_packing: true
pad_to_sequence_len: true

# Batching
gradient_accumulation_steps: 16
micro_batch_size: 1
gradient_checkpointing: unsloth

# wandb
wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
wandb_project: mistral-nemo-v1
wandb_entity: # A wandb Team name if using a Team
wandb_watch:
wandb_name: loraplus-nemo--8192
wandb_run_id: # Set the ID of your wandb run
wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training

# Optimizer
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00001

# Misc
auto_resume_from_checkpoints: true
logging_steps: 1
weight_decay: 0.0

special_tokens:
  pad_token: <pad>

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true

# Multi-GPU
deepspeed:
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: MistralDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD

[fblgit]

the high grad_norm is because there is a big mismatch of whatever the model originally fitted and what is being presented now.
This could be a wrong pad/eos/bos.. it could be obviously the entire template change that OP is performing 😄
i guess that changing template with train_on_inputs: false this.. may not be ideal right?..

[fizzAI]

FWIW, an MN lora trained fine for me on 1xGPU but I'm still seeing people occasionally complaining about this being a bug. Possibly a multi-GPU issue? Seems to persist across FSDP and Deepspeed though