test

src/axolotl/core/trainer_builder.py

@@ -896,13 +896,13 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         for key, value in metrics.items():
             self._stored_metrics[train_eval][key].append(value)
 
-    def _save_checkpoint(self, model, trial):
+    def _save_checkpoint(self, model, trial, **kwargs):
         # make sure the checkpoint dir exists, since trainer is flakey
         checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
         run_dir = self._get_output_dir(trial=trial)
         output_dir = os.path.join(run_dir, checkpoint_folder)
         os.makedirs(output_dir, exist_ok=True)
-        return super()._save_checkpoint(model, trial)
+        return super()._save_checkpoint(model, trial, **kwargs)
 
 
 class AxolotlMambaTrainer(AxolotlTrainer):

test.yml

@@ -0,0 +1,76 @@
+base_model: JackFram/llama-68m
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_fused_linear_cross_entropy: true
+
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.5
+output_dir: ./outputs/out
+
+sequence_len: 1024
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_torch
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+early_stopping_patience:
+resume_from_checkpoint:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 100
+evals_per_epoch: 2
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+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_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
+  fsdp_backward_prefetch: BACKWARD_PRE
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot_id|>

tests/e2e/integrations/liger.py

@@ -63,6 +63,51 @@ class LigerIntegrationTestCase(unittest.TestCase):
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "model.safetensors").exists()
 
+    @with_temp_dir
+    def test_llama_wo_flce2(self, temp_dir):
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "plugins": [
+                    "axolotl.integrations.liger.LigerPlugin",
+                ],
+                "liger_rope": True,
+                "liger_rms_norm": True,
+                "liger_swiglu": True,
+                "liger_cross_entropy": True,
+                "liger_fused_linear_cross_entropy": False,
+                "sequence_len": 1024,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "save_safetensors": True,
+                "bf16": "auto",
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "model.safetensors").exists()
+    
     @with_temp_dir
     def test_llama_w_flce(self, temp_dir):
         cfg = DictDefault(

tests/test_datasets.py

@@ -306,6 +306,10 @@ class TestDatasetPreparation(unittest.TestCase):
         """Verify that processing data from the hub works with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:
             prepared_path = Path(tmp_dir) / "prepared"
+
+            # make sure prepared_path is empty
+            shutil.rmtree(prepared_path, ignore_errors=True)
+
             cfg = DictDefault(
                 {
                     "tokenizer_config": "huggyllama/llama-7b",