feat: add wip fft offload config

examples/glm4/glm4.5-fft-fsdp2-offload.yaml

@@ -0,0 +1,73 @@
+base_model: zai-org/GLM-4.5-Air
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding
+
+datasets:
+  - path: winglian/pirate-ultrachat-10k
+    type: chat_template
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/qlora-out
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: true
+
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+# gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: true
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Glm4MoeDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/glm4/glm4.5-qlora.yaml

@@ -2,6 +2,9 @@ base_model: zai-org/GLM-4.5-Air
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
 load_in_4bit: true
 
 datasets:
@@ -40,7 +43,7 @@ wandb_log_model:
 gradient_accumulation_steps: 2
 micro_batch_size: 2
 num_epochs: 1
-optimizer: adamw_8bit
+optimizer: adamw_torch_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002