pass additional info for fix untrained tokens when using distributed + offloading (#2388)

* pass additional info for fix untrained tokens when using distributed + offloading

* use latest version of vendored lib

* use v0.0.5 of contribs lgpl

* fix for no bad tokens and add tests

* use release

* add multigpu test too

* make sure the multigpu zero3 test actually uses zero3

tests/e2e/multigpu/test_llama.py

@@ -750,3 +750,66 @@ class TestMultiGPULlama:
         check_tensorboard(
             temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
         )
+
+    def test_fix_untrained_tokens(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "fix_untrained_tokens": True,
+                "sequence_len": 512,
+                "val_set_size": 0.0,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                    "bos_token": "<|custom_im_start|>",
+                    "eos_token": "<|custom_im_end|>",
+                },
+                "datasets": [
+                    {
+                        "chat_template": "jinja",
+                        "chat_template_jinja": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|custom_im_start|>' + message['role'] + '\n' + message['content'] + '<|custom_im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|custom_im_start|>assistant\n' }}{% endif %}",
+                        "path": "mlabonne/FineTome-100k",
+                        "type": "chat_template",
+                        "split": "train[:10%]",
+                        "field_messages": "conversations",
+                        "message_field_role": "from",
+                        "message_field_content": "value",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 5,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch_fused",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "sample_packing": True,
+                "bf16": True,
+                "save_safetensors": True,
+                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero3_bf16.json"),
+                "use_tensorboard": True,
+            }
+        )
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "axolotl",
+                "train",
+                str(Path(temp_dir) / "config.yaml"),
+                "--num-processes",
+                "2",
+                "--main-process-port",
+                f"{get_torch_dist_unique_port()}",
+            ]
+        )
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 4.0, "Train Loss is too high"
+        )

tests/e2e/test_llama.py

@@ -66,6 +66,54 @@ class TestLlama:
         check_model_output_exists(temp_dir, cfg)
 
     def test_fix_untrained_tokens(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "fix_untrained_tokens": True,
+                "sequence_len": 512,
+                "val_set_size": 0.0,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                    "bos_token": "<|custom_im_start|>",
+                    "eos_token": "<|custom_im_end|>",
+                },
+                "datasets": [
+                    {
+                        "chat_template": "jinja",
+                        "chat_template_jinja": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|custom_im_start|>' + message['role'] + '\n' + message['content'] + '<|custom_im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|custom_im_start|>assistant\n' }}{% endif %}",
+                        "path": "mlabonne/FineTome-100k",
+                        "type": "chat_template",
+                        "split": "train[:10%]",
+                        "field_messages": "conversations",
+                        "message_field_role": "from",
+                        "message_field_content": "value",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 5,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "sample_packing": True,
+                "bf16": True,
+                "save_safetensors": True,
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
+
+    def test_fix_untrained_tokens_already_trained(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {