OPRO, DPO don't work with Mixtral-8x22B. FSDP + QLORA & bigstral-ds-zero3

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

[X] I searched previous Bug Reports didn't find any similar reports.

Expected Behavior

DPO/ORPO training should run successfully

Current behaviour

Models tested (don't have any issues with inference)

mistral-community/Mixtral-8x22B-v0.1
0-hero/Matter-0.2-8x22B (FSDP + QLORA finetune of the above)

Machines tested (tried each type from multiple providers)

8xA100 (80GB)
8xH100 (SXM)

Images tested

winglian/axolotl-runpod:main-latest (as of 18 Apr 2024)
winglian/axolotl-runpod:main-py3.11-cu121-2.1.2
winglian/axolotl-runpod:main-py3.10-cu118-2.1.2
runpod/pytorch:2.0.1-py3.10-cuda11.8.0-devel-ubuntu22.04

Ran tests with all the combinations mentioned above

Both issues mentioned below happen for both ORPO & DPO

Issue 1 - FSDP + QLORA

https://github.com/OpenAccess-AI-Collective/axolotl/issues/1494

Issue 2 - bigstral-ds-zero3

Happened anytime before the first 20 steps. Tried reducing the below to 1 but the issue persists.

gradient_accumulation_steps: 1
micro_batch_size: 1

Training hangs and eventually stops with a NCCL timeout https://github.com/huggingface/accelerate/issues/314 GPU util also falls once it hangs, example below

Steps to reproduce

Start training with any of the configs below

FSDP + QLORA config

base_model: mistral-community/Mixtral-8x22B-v0.1
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

rl: dpo
datasets:
  - path: argilla/ultrafeedback-binarized-preferences-cleaned
    split: train
    type: chatml.ultra

dpo_beta: 0.1

chat_template: chatml
default_system_message: You are a helpful assistant

dataset_prepared_path: data
val_set_size: 0
output_dir: output

sequence_len: 8192
sample_packing: false
pad_to_sequence_len: false

adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_modules_to_save:
- embed_tokens
- lm_head

gradient_accumulation_steps: 8
micro_batch_size: 4
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint: 
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
weight_decay: 0.0
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_transformer_layer_cls_to_wrap: MixtralSparseMoeBlock
 fsdp_state_dict_type: FULL_STATE_DICT
 fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP

special_tokens:
  bos_token: "<s>"
  eos_token: "<|im_end|>"
  unk_token: "<unk>"
tokens:
  - "<|begin_func|>"
  - "<|end_func|>"
  - "<|begin_func_response|>"
  - "<|end_func_response|>"
  - "<|im_start|>"
  - "<|im_end|>"

bigstral-ds-zero3 config

base_model: 0-hero/Matter-0.2-8x22B
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

unfrozen_parameters:
  - ^lm_head.weight$
  - ^model.embed_tokens.weight$
  - model.layers.4[4-9]+.block_sparse_moe.gate
  - model.layers.4[4-9]+.block_sparse_moe.experts
  - model.layers.5[0-5]+.block_sparse_moe.gate
  - model.layers.5[0-5]+.block_sparse_moe.experts

model_config:
  output_router_logits: true

rl: orpo
datasets:
  - path: mlabonne/orpo-mix-40k
    split: train
    type: orpo.chat_template

chat_template: chatml
default_system_message: You are a helpful assistant

dataset_prepared_path: data
val_set_size: 0
output_dir: output

sequence_len: 8192
sample_packing: false
pad_to_sequence_len: false

gradient_accumulation_steps: 8
micro_batch_size: 4
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint: 
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
save_total_limit: 1
save_steps:
debug:
deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_params.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "<|im_end|>"
  unk_token: "<unk>"
tokens:
  - "<|begin_func|>"
  - "<|end_func|>"
  - "<|begin_func_response|>"
  - "<|end_func_response|>"
  - "<|im_start|>"
  - "<|im_end|>"

Config yaml

No response

Possible solution