FSDP+LORA on multiple gpu(A100 80gb*4) ValueError: Cannot flatten integer dtype tensors

[Paxwell-Paxwell]

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

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

Expected Behavior

The LoRA configuration should work with fsdp

Current behaviour

[rank0]: raise ValueError("Cannot flatten integer dtype tensors")
[rank0]: ValueError: Cannot flatten integer dtype tensors
[rank1]: Traceback (most recent call last):

Steps to reproduce

Run Axolotl on multiple GPUs using LoRA+ FSDP, 4 NVIDIA A100 GPUs with 80GB
-torch Version: 2.5.1
-axolotl Version: 0.6.0

Config yaml

base_model: meta-llama/Llama-3.3-70B-Instruct
model_type: LlamaForCausalLM
processing_class: AutoTokenizer

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

load_in_8bit: true
load_in_4bit: false
strict: false

chat_template: llama3
datasets:

dataset_prepared_path: ./workspace/aiLawData/last_run_prepared
val_set_size: 0
output_dir: ./workspace/aiLawData/outputs/Llama-3.3-70B-memo-law-Instruct-lora-r256-v3
hub_model_id: PaxwellPaxwell/Llama-3.3-70B-Memo-law-Instruct-adapter-lora-r256-v3
sequence_len: 12000
sample_packing: false
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 256
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project: Ai-Law
wandb_entity:
wandb_watch:
wandb_name: Llama-3.3-70B-memo-law-Instruct-adapter-lora-r256-v3
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 10
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.0002
auto_resume_from_checkpoints: true

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing_kwargs:
  use_reentrant: true
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 100
evals_per_epoch: 4
eval_table_size:
saves_per_epoch: 2
debug:
deepspeed:
weight_decay: 0.0

fsdp:
  - full_shard
  - auto_wrap

fsdp_config:
  activation_checkpointing: true
  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: LlamaDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD

special_tokens:
  pad_token: <|finetune_right_pad_id|>
  eos_token: <|eot_id|>

Possible solution

No response

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

3.10.13

axolotl branch-commit

main/latest

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.

[NJordan72]

Can you post the stack trace so we can see what is throwing the error? I had a similar problem earlier this week and depending on where it is coming from I either had to set gradient_accumulation_steps to 1 or turn off the liger cross entropy kernel