nit [skip-e2e]

Co-authored-by: djsaunde <1245942+djsaunde@users.noreply.github.com>

.github/CONTRIBUTING.md

@@ -57,6 +57,13 @@ We welcome ideas for improvements and new features. To suggest an enhancement, o
 5. Push your branch to your fork on GitHub.
 6. Open a new pull request against the `main` branch of the axolotl repository. Include a clear and concise description of your changes, referencing any related issues.
 
+#### Skipping CI Checks
+
+You can skip certain CI checks by including specific keywords in your commit messages:
+
+- `[skip ci]` or `skip ci` - Skips all CI checks for that commit
+- `[skip-e2e]` or `skip-e2e` - Skips only end-to-end tests while running other CI checks
+
 ## Style Guidelines
 
 ### Code Style

.github/workflows/main.yml

@@ -98,6 +98,12 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.1
             axolotl_extras:
+            is_latest:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras: vllm
             is_latest: true
           - cuda: 128
             cuda_version: 12.8.1
@@ -151,6 +157,18 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras:
+            is_latest:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras: vllm
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/tests.yml

@@ -105,7 +105,8 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
           pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
           pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
@@ -179,8 +180,8 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v --durations=10 tests/patched/
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
           pytest -v --durations=10 tests/cli/
 
       - name: cleanup pip cache

.pre-commit-config.yaml

@@ -3,7 +3,7 @@ default_language_version:
 
 repos:
 -   repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+    rev: v6.0.0
     hooks:
     -   id: check-yaml
     -   id: end-of-file-fixer
@@ -23,7 +23,7 @@ repos:
     hooks:
     - id: flake8
 -   repo: https://github.com/pylint-dev/pylint
-    rev: v3.3.7
+    rev: v3.3.8
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy

CITATION.cff

@@ -0,0 +1,10 @@
+cff-version: 1.2.0
+type: software
+title: "Axolotl: Post-Training for AI Models"
+message: "If you use this software, please cite it as below."
+authors:
+  - name: "Axolotl maintainers and contributors"
+repository-code: "https://github.com/axolotl-ai-cloud/axolotl"
+url: "https://axolotl.ai/"
+license: Apache-2.0
+date-released: "2023-05-30"

README.md

@@ -149,6 +149,20 @@ Contributions are welcome! Please see our [Contributing Guide](https://github.co
 
 Interested in sponsoring? Contact us at [wing@axolotl.ai](mailto:wing@axolotl.ai)
 
+## 📝 Citing Axolotl
+
+If you use Axolotl in your research or projects, please cite it as follows:
+
+```bibtex
+@software{axolotl,
+  title = {Axolotl: Post-Training for AI Models},
+  author = {{Axolotl maintainers and contributors}},
+  url = {https://github.com/axolotl-ai-cloud/axolotl},
+  license = {Apache-2.0},
+  year = {2023}
+}
+```
+
 ## 📜 License
 
 This project is licensed under the Apache 2.0 License - see the [LICENSE](LICENSE) file for details.

TODO.md

@@ -1,10 +0,0 @@
-# todo list
-
-- [] Validation of parameters for combinations that won't work
-
-
-
-## things that are known not to work
-
-- FSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203
-- adamw_bnb_8bit doesn't play well with FSDP offload

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.12.0"
+__version__ = "0.13.0.dev"

src/axolotl/cli/config.py

@@ -153,15 +153,14 @@ def prepare_plugins(cfg: DictDefault):
         plugin_manager = PluginManager.get_instance()
         for plugin_name in cfg["plugins"]:
             plugin_manager.register(plugin_name)
+        for plugin in plugin_manager.plugins.values():
+            plugin.register(cfg)
 
 
 def plugin_set_cfg(cfg: DictDefault):
     if cfg.get("plugins"):
         plugin_manager = PluginManager.get_instance()
         plugin_manager.cfg = cfg
-        # now that we have the finalized cfg, register the plugins individually
-        for plugin in plugin_manager.plugins.values():
-            plugin.register(cfg)
 
 
 def load_cfg(

src/axolotl/cli/main.py

@@ -123,9 +123,10 @@ def train(
     _launcher = None if kwargs.get("use_ray") else launcher
 
     # Process each configuration
-    for cfg_file in generate_config_files(config, sweep):
+    for cfg_file, is_group in generate_config_files(config, sweep):
         try:
-            launch_training(cfg_file, _launcher, cloud, kwargs, launcher_args)
+            use_exec = is_group is not True
+            launch_training(cfg_file, _launcher, cloud, kwargs, launcher_args, use_exec)
         except subprocess.CalledProcessError as exc:
             LOG.error(f"Failed to train/fine-tune config '{cfg_file}': {exc}")
             if not sweep:

src/axolotl/cli/utils/train.py

@@ -2,6 +2,7 @@
 
 import os
 import subprocess  # nosec
+import sys
 import tempfile
 from typing import Any, Iterator, Literal
 
@@ -64,10 +65,18 @@ def build_command(base_cmd: list[str], options: dict[str, Any]) -> list[str]:
     return cmd
 
 
-def generate_config_files(config: str, sweep: str | None) -> Iterator[str]:
-    """Generate list of configuration files to process."""
+def generate_config_files(config: str, sweep: str | None) -> Iterator[tuple[str, bool]]:
+    """
+    Generate list of configuration files to process. Yields a tuple of the configuration file name and a boolean indicating
+    whether this is a group of configurations (i.e., a sweep).
+
+    Args:
+        config: Base configuration file
+        sweep: Sweep configuration file
+    """
+
     if not sweep:
-        yield config
+        yield config, False
         return
 
     # Load sweep and base configurations
@@ -78,6 +87,7 @@ def generate_config_files(config: str, sweep: str | None) -> Iterator[str]:
 
     # Generate all possible configurations
     permutations = generate_sweep_configs(base_config, sweep_config)
+    is_group = len(permutations) > 1
     for permutation in permutations:
         # pylint: disable=consider-using-with
         temp_file = tempfile.NamedTemporaryFile(
@@ -88,7 +98,7 @@ def generate_config_files(config: str, sweep: str | None) -> Iterator[str]:
         )
         yaml.dump(permutation, temp_file)
         temp_file.close()
-        yield temp_file.name
+        yield temp_file.name, is_group
 
 
 def launch_training(
@@ -97,6 +107,7 @@ def launch_training(
     cloud: str | None,
     kwargs: dict,
     launcher_args: list[str] | None = None,
+    use_exec: bool = False,
 ) -> None:
     """Execute training with the given configuration."""
     launcher_args = launcher_args or []
@@ -105,11 +116,14 @@ def launch_training(
         _launch_cloud_training(cloud, cfg_file, launcher, kwargs, launcher_args)
     elif launcher:
         if launcher == "accelerate":
-            _launch_accelerate_training(cfg_file, kwargs, launcher_args)
+            _launch_accelerate_training(cfg_file, kwargs, launcher_args, use_exec)
         elif launcher == "torchrun":
-            _launch_torchrun_training(cfg_file, kwargs, launcher_args)
+            _launch_torchrun_training(cfg_file, kwargs, launcher_args, use_exec)
         elif launcher == "python":
             _launch_python_training(cfg_file, kwargs)
+    elif launcher is None:
+        # handle ray train launch
+        _launch_python_training(cfg_file, kwargs)
 
 
 def _launch_cloud_training(
@@ -136,7 +150,10 @@ def _launch_cloud_training(
 
 
 def _launch_accelerate_training(
-    cfg_file: str, kwargs: dict, launcher_args: list[str] | None = None
+    cfg_file: str,
+    kwargs: dict,
+    launcher_args: list[str] | None = None,
+    use_exec: bool = False,
 ) -> None:
     """Execute training via accelerate launcher."""
     launcher_args = launcher_args or []
@@ -161,11 +178,20 @@ def _launch_accelerate_training(
         base_cmd.append(cfg_file)
 
     cmd = build_command(base_cmd, kwargs)
-    subprocess.run(cmd, check=True)  # nosec B603
+    if use_exec:
+        # make sure to flush stdout and stderr before replacing the process
+        sys.stdout.flush()
+        sys.stderr.flush()
+        os.execvpe(cmd[0], cmd, os.environ)  # nosec B606
+    else:
+        subprocess.run(cmd, check=True)  # nosec B603
 
 
 def _launch_torchrun_training(
-    cfg_file: str, kwargs: dict, launcher_args: list[str] | None = None
+    cfg_file: str,
+    kwargs: dict,
+    launcher_args: list[str] | None = None,
+    use_exec: bool = False,
 ) -> None:
     """Execute training via torchrun launcher."""
     launcher_args = launcher_args or []
@@ -178,7 +204,13 @@ def _launch_torchrun_training(
         base_cmd.append(cfg_file)
 
     cmd = build_command(base_cmd, kwargs)
-    subprocess.run(cmd, check=True)  # nosec B603
+    if use_exec:
+        # make sure to flush stdout and stderr before replacing the process
+        sys.stdout.flush()
+        sys.stderr.flush()
+        os.execvpe(cmd[0], cmd, os.environ)  # nosec B606
+    else:
+        subprocess.run(cmd, check=True)  # nosec B603
 
 
 def _launch_python_training(cfg_file: str, kwargs: dict) -> None:

src/axolotl/integrations/base.py

@@ -76,8 +76,8 @@ class BasePlugin:
     def __init__(self):
         """Initializes the BasePlugin."""
 
-    def register(self, cfg: DictDefault):  # pylint: disable=unused-argument
-        """Registers the plugin with the given configuration.
+    def register(self, cfg: dict):  # pylint: disable=unused-argument
+        """Registers the plugin with the given configuration as an unparsed dict.
 
         Args:
             cfg: The configuration for the plugin.

src/axolotl/loaders/patch_manager.py

@@ -73,9 +73,6 @@ class PatchManager:
         self._apply_voxtral_patches()
 
     def _apply_transformers_patches(self):
-        from axolotl.monkeypatch.transformers.modeling_flash_attention_utils import (
-            patch_prepare_from_posids,
-        )
         from axolotl.monkeypatch.transformers.trainer_loss_calc import (
             patch_evaluation_loop,
             patch_maybe_log_save_evaluate,
@@ -87,7 +84,6 @@ class PatchManager:
             and self.cfg.fsdp_version == 2
         )
 
-        patch_prepare_from_posids()
         patch_evaluation_loop(patch_fsdp2)
         patch_maybe_log_save_evaluate()
 

src/axolotl/monkeypatch/transformers/modeling_flash_attention_utils.py

@@ -1,87 +0,0 @@
-"""
-Monkey patch to fix transformers.modeling_flash_attention_utils.
-
-see https://github.com/huggingface/transformers/pull/39653/files
-"""
-
-import sys
-
-import torch
-
-
-def _prepare_from_posids(query, key, value, position_ids):
-    """
-    This function returns necessary arguments to call `flash_attn_varlen_func`.
-    All three query, key, value states will be flattened.
-    Cumulative lengths of each examples in the batch will be extracted from position_ids.
-    NOTE: ideally cumulative lengths should be prepared at the data collator stage
-    Arguments:
-        query (`torch.Tensor`):
-            Query state with padding. Shape: (batch_size, query_length, num_heads, head_dim).
-        key (`torch.Tensor`):
-            Key state with padding. Shape: (batch_size, kv_seq_len, num_key_value_heads, head_dim).
-        value (`torch.Tensor`):
-            Value state with padding. Shape: (batch_size, kv_seq_len, num_key_value_heads, head_dim).
-        position_ids (`torch.Tensor`):
-            Boolean or int tensor of shape (batch_size, sequence_length), 1 means valid and 0 means not valid.
-    Return:
-        query (`torch.Tensor`):
-            Query state without padding. Shape: (total_target_length, num_heads, head_dim).
-        key (`torch.Tensor`):
-            Key state with padding. Shape: (total_source_length, num_key_value_heads, head_dim).
-        value (`torch.Tensor`):
-            Value state with padding. Shape: (total_source_length, num_key_value_heads, head_dim).
-        indices_q (`torch.Tensor`):
-            The indices of non-masked tokens from the flattened input target sequence.
-        (cu_seqlens_q, cu_seqlens_k) (`tuple[int]`):
-            The cumulative sequence lengths for the target (query) and source (key, value), used to index into ragged (unpadded) tensors. `cu_seqlens` shape is (batch_size + 1,).
-        (max_seqlen_in_batch_q, max_seqlen_in_batch_k) (`tuple[int]`):
-            Maximum sequence length in batch (`max_seqlen_in_batch_q` for the target sequence i.e. query, `max_seqlen_in_batch_k` for the source sequence i.e. key/value).
-    """
-    query = query.contiguous().view(-1, query.size(-2), query.size(-1))
-    key = key.contiguous().view(-1, key.size(-2), key.size(-1))
-    value = value.contiguous().view(-1, value.size(-2), value.size(-1))
-
-    position_ids = position_ids.flatten()
-    indices_q = torch.arange(
-        position_ids.size(0), device=position_ids.device, dtype=torch.int32
-    )
-
-    cu_seq_lens = torch.cat(
-        (
-            indices_q[position_ids == 0],
-            torch.tensor(
-                position_ids.size(), device=position_ids.device, dtype=torch.int32
-            ),
-        )
-    )
-    # NOTE: With torch compile, this will cause a graph break if you don't set
-    # `TORCHDYNAMO_CAPTURE_SCALAR_OUTPUTS=1` in the environment or call
-    # `torch._dynamo.config.capture_scalar_outputs = True` before doing the forward pass.
-    # This is a limitation of flash attention API, as the function `flash_attn_varlen_func`
-    # requires `max_length_q`, `max_length_k` to be passed as `int` and not `torch.Tensor`.
-    # https://github.com/Dao-AILab/flash-attention/blob/2dd8078adc1d9b74e315ee99718c0dea0de8eeb6/flash_attn/flash_attn_interface.py#L1423-L1424
-    # We should use cu_seq_lens instead of position_ids to get the max length since position_ids is not always increasing
-    # for some models (e.g. qwen2-vl).
-    max_length = cu_seq_lens.diff().max().item()
-    return (
-        query,
-        key,
-        value,
-        indices_q,
-        (cu_seq_lens, cu_seq_lens),
-        (max_length, max_length),
-    )
-
-
-def patch_prepare_from_posids():
-    import transformers.modeling_flash_attention_utils
-
-    transformers.modeling_flash_attention_utils._prepare_from_posids = (  # pylint: disable=protected-access
-        _prepare_from_posids
-    )
-    setattr(
-        sys.modules["transformers.modeling_flash_attention_utils"],
-        "_prepare_from_posids",
-        _prepare_from_posids,
-    )

tests/cli/test_cli_base.py

@@ -47,7 +47,9 @@ class BaseCliTest:
         config_path = tmp_path / "config.yml"
         config_path.write_text(valid_test_config)
 
-        with patch("subprocess.run") as mock:
+        mock_fn = "os.execvpe" if command == "train" else "subprocess.run"
+
+        with patch(mock_fn) as mock:
             result = cli_runner.invoke(cli, [command, str(config_path)])
 
             assert mock.called
@@ -65,8 +67,12 @@ class BaseCliTest:
             if train:
                 expected.append("--shard=False")
 
-            assert mock.call_args.args[0] == expected
-            assert mock.call_args.kwargs == {"check": True}
+            if command == "train":
+                assert mock.call_args.args[0] == "accelerate"
+                assert mock.call_args.args[1] == expected
+            else:
+                assert mock.call_args.args[0] == expected
+                assert mock.call_args.kwargs == {"check": True}
             assert result.exit_code == 0
 
     def _test_cli_overrides(self, tmp_path: Path, valid_test_config: str):

tests/cli/test_cli_train.py

@@ -85,7 +85,7 @@ class TestTrainCommand(BaseCliTest):
         config_path = tmp_path / "config.yml"
         config_path.write_text(valid_test_config)
 
-        with patch("subprocess.run") as mock_subprocess:
+        with patch("os.execvpe") as mock_subprocess:
             result = cli_runner.invoke(
                 cli,
                 [
@@ -104,7 +104,7 @@ class TestTrainCommand(BaseCliTest):
             mock_subprocess.assert_called_once()
 
             # Verify launcher args are passed to torchrun
-            called_cmd = mock_subprocess.call_args.args[0]
+            called_cmd = mock_subprocess.call_args.args[1]
             assert called_cmd[0] == "torchrun"
             assert "--nproc_per_node=2" in called_cmd
             assert "--nnodes=1" in called_cmd
@@ -118,7 +118,7 @@ class TestTrainCommand(BaseCliTest):
         config_path = tmp_path / "config.yml"
         config_path.write_text(valid_test_config)
 
-        with patch("subprocess.run") as mock_subprocess:
+        with patch("os.execvpe") as mock_subprocess:
             result = cli_runner.invoke(
                 cli,
                 [
@@ -137,7 +137,8 @@ class TestTrainCommand(BaseCliTest):
             mock_subprocess.assert_called_once()
 
             # Verify launcher args are passed to accelerate
-            called_cmd = mock_subprocess.call_args.args[0]
+            assert mock_subprocess.call_args.args[0] == "accelerate"
+            called_cmd = mock_subprocess.call_args.args[1]
             assert called_cmd[0] == "accelerate"
             assert called_cmd[1] == "launch"
             assert "--config_file=accelerate_config.yml" in called_cmd
@@ -152,7 +153,7 @@ class TestTrainCommand(BaseCliTest):
         config_path = tmp_path / "config.yml"
         config_path.write_text(valid_test_config)
 
-        with patch("subprocess.run") as mock_subprocess:
+        with patch("os.execvpe") as mock_subprocess:
             result = cli_runner.invoke(
                 cli,
                 [
@@ -170,7 +171,8 @@ class TestTrainCommand(BaseCliTest):
             mock_subprocess.assert_called_once()
 
             # Verify no launcher args contamination
-            called_cmd = mock_subprocess.call_args.args[0]
+            assert mock_subprocess.call_args.args[0] == "accelerate"
+            called_cmd = mock_subprocess.call_args.args[1]
             assert called_cmd[0] == "accelerate"
             assert called_cmd[1] == "launch"
             # Should not contain any extra launcher args
@@ -186,7 +188,7 @@ class TestTrainCommand(BaseCliTest):
         config_path = tmp_path / "config.yml"
         config_path.write_text(valid_test_config)
 
-        with patch("subprocess.run") as mock_subprocess:
+        with patch("os.execvpe") as mock_subprocess:
             result = cli_runner.invoke(
                 cli,
                 [
@@ -207,7 +209,8 @@ class TestTrainCommand(BaseCliTest):
             assert result.exit_code == 0
             mock_subprocess.assert_called_once()
 
-            called_cmd = mock_subprocess.call_args.args[0]
+            assert mock_subprocess.call_args.args[0] == "torchrun"
+            called_cmd = mock_subprocess.call_args.args[1]
             # Verify launcher args
             assert "--nproc_per_node=8" in called_cmd
             # Verify axolotl args are also present

tests/core/test_builders.py

@@ -281,7 +281,9 @@ class TestHFRLTrainerBuilder:
         # Other settings
         assert training_arguments.dataloader_num_workers == 1
         assert training_arguments.dataloader_pin_memory is True
-        assert training_arguments.gradient_checkpointing is False
+
+        # TODO(wing): restore once trl releases 0.22.0
+        # assert training_arguments.gradient_checkpointing is True
 
     def test_dpo_training_arguments(self, dpo_cfg, model, tokenizer):
         builder = HFRLTrainerBuilder(dpo_cfg, model, tokenizer)

tests/e2e/multigpu/test_ray.py

@@ -10,7 +10,11 @@ from accelerate.test_utils import execute_subprocess_async
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import check_tensorboard, require_torch_lt_2_6_0
+from tests.e2e.utils import (
+    check_tensorboard,
+    require_torch_2_7_0,
+    require_torch_lt_2_6_0,
+)
 
 AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 
@@ -139,3 +143,71 @@ class TestMultiGPURay:
         check_tensorboard(
             temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
         )
+
+    @require_torch_2_7_0
+    @pytest.mark.parametrize(
+        "gradient_accumulation_steps",
+        [1, 2],
+    )
+    def test_sft_fsdp2_packed(self, temp_dir, gradient_accumulation_steps):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
+                "sequence_len": 1024,
+                "val_set_size": 0.01,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "tatsu-lab/alpaca",
+                        "type": "alpaca",
+                        "split": "train[:10%]",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 2,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": gradient_accumulation_steps,
+                "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "fsdp_version": 2,
+                "fsdp_config": {
+                    "offload_params": False,
+                    "cpu_ram_efficient_loading": False,
+                    "transformer_layer_cls_to_wrap": "LlamaDecoderLayer",
+                    "state_dict_type": "FULL_STATE_DICT",
+                    "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+                    "reshard_after_forward": True,
+                },
+                "use_tensorboard": True,
+                "save_first_step": False,
+            }
+        )
+
+        # 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"),
+                "--use-ray",
+                "--ray-num-workers",
+                "2",
+            ]
+        )
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
+        )