Fix: modelloader handling of model_kwargs load_in*bit (#1999)

* fix: load_in_*bit not properly read

* fix: load_*bit check

* fix: typo

* refactor: load * bit handling

* feat: add test dpo lora multi-gpu

* fix: turn off sample packing for dpo

* fix: missing warmup_steps

* fix: test to load in 8bit for lora

* skip 8bit lora on h100, add 4bit lora on h100 to multi gpu tests

* chore: reduce max_steps

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>

1991.yml

@@ -1,295 +0,0 @@
-base_model: Qwen/Qwen2.5-14B-Instruct
-model_type: AutoModelForCausalLM #nohup accelerate launch -m axolotl.cli.train /home/ubuntu/qwen2.5_14B.yml > training_output.log 2>&1 &
-tokenizer_type: AutoTokenizer
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-
-chat_template: chatml
-dataset_prepared_path:
-val_set_size: 0
-output_dir: ./outputs/out
-
-sequence_len: 2048
-sample_packing: true
-eval_sample_packing: true
-pad_to_sequence_len: true
-
-unfrozen_parameters:
-- ^lm_head.weight$
-- ^model.embed_tokens.weight$
-# input_layernorm layers
-- model.layers.0.input_layernorm
-- model.layers.1.input_layernorm
-- model.layers.2.input_layernorm
-- model.layers.3.input_layernorm
-- model.layers.4.input_layernorm
-- model.layers.5.input_layernorm
-- model.layers.6.input_layernorm
-- model.layers.7.input_layernorm
-- model.layers.8.input_layernorm
-- model.layers.9.input_layernorm
-- model.layers.10.input_layernorm
-- model.layers.11.input_layernorm
-- model.layers.12.input_layernorm
-- model.layers.13.input_layernorm
-- model.layers.14.input_layernorm
-- model.layers.15.input_layernorm
-- model.layers.16.input_layernorm
-- model.layers.17.input_layernorm
-- model.layers.18.input_layernorm
-- model.layers.19.input_layernorm
-- model.layers.20.input_layernorm
-- model.layers.21.input_layernorm
-- model.layers.22.input_layernorm
-- model.layers.23.input_layernorm
-# lm_head layers
-# mlp.down_proj layers
-- model.layers.1.mlp.down_proj
-- model.layers.35.mlp.down_proj
-- model.layers.38.mlp.down_proj
-- model.layers.37.mlp.down_proj
-- model.layers.36.mlp.down_proj
-- model.layers.15.mlp.down_proj
-- model.layers.11.mlp.down_proj
-- model.layers.12.mlp.down_proj
-- model.layers.34.mlp.down_proj
-- model.layers.44.mlp.down_proj
-- model.layers.45.mlp.down_proj
-- model.layers.9.mlp.down_proj
-- model.layers.41.mlp.down_proj
-- model.layers.33.mlp.down_proj
-- model.layers.43.mlp.down_proj
-- model.layers.40.mlp.down_proj
-- model.layers.13.mlp.down_proj
-- model.layers.8.mlp.down_proj
-- model.layers.39.mlp.down_proj
-- model.layers.10.mlp.down_proj
-- model.layers.14.mlp.down_proj
-- model.layers.16.mlp.down_proj
-- model.layers.31.mlp.down_proj
-- model.layers.32.mlp.down_proj
-# mlp.gate_proj layers
-- model.layers.1.mlp.gate_proj
-- model.layers.44.mlp.gate_proj
-- model.layers.46.mlp.gate_proj
-- model.layers.45.mlp.gate_proj
-- model.layers.43.mlp.gate_proj
-- model.layers.47.mlp.gate_proj
-- model.layers.42.mlp.gate_proj
-- model.layers.32.mlp.gate_proj
-- model.layers.27.mlp.gate_proj
-- model.layers.33.mlp.gate_proj
-- model.layers.28.mlp.gate_proj
-- model.layers.39.mlp.gate_proj
-- model.layers.41.mlp.gate_proj
-- model.layers.40.mlp.gate_proj
-- model.layers.30.mlp.gate_proj
-- model.layers.29.mlp.gate_proj
-- model.layers.31.mlp.gate_proj
-- model.layers.26.mlp.gate_proj
-- model.layers.37.mlp.gate_proj
-- model.layers.10.mlp.gate_proj
-- model.layers.38.mlp.gate_proj
-- model.layers.12.mlp.gate_proj
-- model.layers.36.mlp.gate_proj
-- model.layers.13.mlp.gate_proj
-# mlp.up_proj layers
-- model.layers.1.mlp.up_proj
-- model.layers.13.mlp.up_proj
-- model.layers.11.mlp.up_proj
-- model.layers.14.mlp.up_proj
-- model.layers.15.mlp.up_proj
-- model.layers.12.mlp.up_proj
-- model.layers.8.mlp.up_proj
-- model.layers.16.mlp.up_proj
-- model.layers.9.mlp.up_proj
-- model.layers.19.mlp.up_proj
-- model.layers.10.mlp.up_proj
-- model.layers.7.mlp.up_proj
-- model.layers.17.mlp.up_proj
-- model.layers.20.mlp.up_proj
-- model.layers.21.mlp.up_proj
-- model.layers.18.mlp.up_proj
-- model.layers.38.mlp.up_proj
-- model.layers.37.mlp.up_proj
-- model.layers.39.mlp.up_proj
-- model.layers.42.mlp.up_proj
-- model.layers.41.mlp.up_proj
-- model.layers.27.mlp.up_proj
-- model.layers.28.mlp.up_proj
-- model.layers.34.mlp.up_proj
-# model.norm layers
-# post_attention_layernorm layers
-- model.layers.0.post_attention_layernorm
-- model.layers.1.post_attention_layernorm
-- model.layers.2.post_attention_layernorm
-- model.layers.3.post_attention_layernorm
-- model.layers.4.post_attention_layernorm
-- model.layers.5.post_attention_layernorm
-- model.layers.6.post_attention_layernorm
-- model.layers.7.post_attention_layernorm
-- model.layers.8.post_attention_layernorm
-- model.layers.9.post_attention_layernorm
-- model.layers.10.post_attention_layernorm
-- model.layers.11.post_attention_layernorm
-- model.layers.12.post_attention_layernorm
-- model.layers.13.post_attention_layernorm
-- model.layers.14.post_attention_layernorm
-- model.layers.15.post_attention_layernorm
-- model.layers.16.post_attention_layernorm
-- model.layers.17.post_attention_layernorm
-- model.layers.18.post_attention_layernorm
-- model.layers.19.post_attention_layernorm
-- model.layers.20.post_attention_layernorm
-- model.layers.21.post_attention_layernorm
-- model.layers.22.post_attention_layernorm
-- model.layers.23.post_attention_layernorm
-# self_attn.k_proj layers
-- model.layers.47.self_attn.k_proj
-- model.layers.39.self_attn.k_proj
-- model.layers.41.self_attn.k_proj
-- model.layers.37.self_attn.k_proj
-- model.layers.35.self_attn.k_proj
-- model.layers.44.self_attn.k_proj
-- model.layers.38.self_attn.k_proj
-- model.layers.14.self_attn.k_proj
-- model.layers.7.self_attn.k_proj
-- model.layers.12.self_attn.k_proj
-- model.layers.11.self_attn.k_proj
-- model.layers.32.self_attn.k_proj
-- model.layers.10.self_attn.k_proj
-- model.layers.8.self_attn.k_proj
-- model.layers.9.self_attn.k_proj
-- model.layers.6.self_attn.k_proj
-- model.layers.45.self_attn.k_proj
-- model.layers.42.self_attn.k_proj
-- model.layers.5.self_attn.k_proj
-- model.layers.40.self_attn.k_proj
-- model.layers.33.self_attn.k_proj
-- model.layers.0.self_attn.k_proj
-- model.layers.34.self_attn.k_proj
-- model.layers.13.self_attn.k_proj
-# self_attn.o_proj layers
-- model.layers.12.self_attn.o_proj
-- model.layers.5.self_attn.o_proj
-- model.layers.14.self_attn.o_proj
-- model.layers.16.self_attn.o_proj
-- model.layers.20.self_attn.o_proj
-- model.layers.13.self_attn.o_proj
-- model.layers.11.self_attn.o_proj
-- model.layers.4.self_attn.o_proj
-- model.layers.6.self_attn.o_proj
-- model.layers.19.self_attn.o_proj
-- model.layers.7.self_attn.o_proj
-- model.layers.18.self_attn.o_proj
-- model.layers.8.self_attn.o_proj
-- model.layers.38.self_attn.o_proj
-- model.layers.15.self_attn.o_proj
-- model.layers.17.self_attn.o_proj
-- model.layers.9.self_attn.o_proj
-- model.layers.10.self_attn.o_proj
-- model.layers.21.self_attn.o_proj
-- model.layers.28.self_attn.o_proj
-- model.layers.32.self_attn.o_proj
-- model.layers.35.self_attn.o_proj
-- model.layers.39.self_attn.o_proj
-- model.layers.3.self_attn.o_proj
-# self_attn.q_proj layers
-- model.layers.1.self_attn.q_proj
-- model.layers.2.self_attn.q_proj
-- model.layers.3.self_attn.q_proj
-- model.layers.44.self_attn.q_proj
-- model.layers.29.self_attn.q_proj
-- model.layers.45.self_attn.q_proj
-- model.layers.43.self_attn.q_proj
-- model.layers.32.self_attn.q_proj
-- model.layers.38.self_attn.q_proj
-- model.layers.19.self_attn.q_proj
-- model.layers.42.self_attn.q_proj
-- model.layers.34.self_attn.q_proj
-- model.layers.36.self_attn.q_proj
-- model.layers.40.self_attn.q_proj
-- model.layers.26.self_attn.q_proj
-- model.layers.20.self_attn.q_proj
-- model.layers.39.self_attn.q_proj
-- model.layers.28.self_attn.q_proj
-- model.layers.35.self_attn.q_proj
-- model.layers.41.self_attn.q_proj
-- model.layers.33.self_attn.q_proj
-- model.layers.25.self_attn.q_proj
-- model.layers.30.self_attn.q_proj
-- model.layers.27.self_attn.q_proj
-# self_attn.v_proj layers
-- model.layers.0.self_attn.v_proj
-- model.layers.7.self_attn.v_proj
-- model.layers.39.self_attn.v_proj
-- model.layers.31.self_attn.v_proj
-- model.layers.15.self_attn.v_proj
-- model.layers.10.self_attn.v_proj
-- model.layers.32.self_attn.v_proj
-- model.layers.41.self_attn.v_proj
-- model.layers.6.self_attn.v_proj
-- model.layers.33.self_attn.v_proj
-- model.layers.42.self_attn.v_proj
-- model.layers.29.self_attn.v_proj
-- model.layers.14.self_attn.v_proj
-- model.layers.9.self_attn.v_proj
-- model.layers.35.self_attn.v_proj
-- model.layers.38.self_attn.v_proj
-- model.layers.13.self_attn.v_proj
-- model.layers.30.self_attn.v_proj
-- model.layers.5.self_attn.v_proj
-- model.layers.34.self_attn.v_proj
-- model.layers.28.self_attn.v_proj
-- model.layers.37.self_attn.v_proj
-- model.layers.27.self_attn.v_proj
-- model.layers.11.self_attn.v_proj
-# model.embed_tokens layers
-
-
-gradient_accumulation_steps: 2
-micro_batch_size: 2
-num_epochs: 3
-optimizer: adamw_torch_fused
-lr_scheduler: linear
-learning_rate: 5e-6
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-liger_rope: true
-liger_rms_norm: true
-liger_swiglu: true
-liger_fused_linear_cross_entropy: true
-
-gradient_checkpointing: unsloth
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch: 2
-saves_per_epoch: 1
-save_total_limit: 4
-debug:
-deepspeed: deepspeed_configs/zero3_bf16.json
-weight_decay: 0.05
-special_tokens:
-  eos_token: <|im_end|>

devtools/dev_chat_template.yml

@@ -7,8 +7,8 @@ load_in_8bit: true
 load_in_4bit: false
 
 datasets:
-  - path: philschmid/guanaco-sharegpt-style
-    type: sharegpt
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
     shards: 10
 val_set_size: 0
 output_dir: temp_debug/axolotl_outputs/model

docs/debugging.qmd

@@ -51,12 +51,12 @@ While debugging it's helpful to simplify your test scenario as much as possible.
 
 ### Background
 
-The below example shows how to configure VSCode to debug data preprocessing of the `sharegpt` format.  This is the format used when you have the following in your axolotl config:
+The below example shows how to configure VSCode to debug data preprocessing of the `chat_template` format.  This is the format used when you have the following in your axolotl config:
 
 ```yaml
 datasets:
-  - path: <path to your sharegpt formatted dataset> # example on HF Hub: philschmid/guanaco-sharegpt-style
-    type: sharegpt
+  - path: <path to your chat_template formatted dataset> # example on HF Hub: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
 ```
 
 >[!Important]
@@ -83,7 +83,7 @@ If you developing on a remote host, you can easily use VSCode to debug remotely.
 
 The easiest way to get started is to modify the [.vscode/launch.json](../.vscode/launch.json) file in this project.  This is just an example configuration, so you may need to modify or copy it to suit your needs.
 
-For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_sharegpt.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
+For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
 
 ```jsonc
 // .vscode/launch.json
@@ -91,12 +91,12 @@ For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 acceler
     "version": "0.2.0",
     "configurations": [
         {
-            "name": "Debug axolotl prompt - sharegpt",
+            "name": "Debug axolotl prompt - chat_template",
             "type": "python",
             "module": "accelerate.commands.launch",
             "request": "launch",
             "args": [
-                "-m", "axolotl.cli.train", "dev_sharegpt.yml",
+                "-m", "axolotl.cli.train", "dev_chat_template.yml",
                 // The flags below simplify debugging by overriding the axolotl config
                 // with the debugging tips above.  Modify as needed.
                 "--dataset_processes=1",      // limits data preprocessing to one process
@@ -240,6 +240,6 @@ style="border-radius: 10px; display: block; margin: auto;" width="560" height="3
 </div>
 <br>
 
-[^1]: The config actually mimics the command `CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/sharegpt.yml`, but this is the same thing.
+[^1]: The config actually mimics the command `CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/chat_template.yml`, but this is the same thing.
 
 [^2]: Many of the below flags are recommended best practices by Nvidia when using nvidia-container-toolkit.  You can read more about these flags [here](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html).

examples/deepseek-v2/qlora-fsdp-2_5.yaml

@@ -16,7 +16,10 @@ chat_template: deepseek_v2
 datasets:
   - path: mlabonne/FineTome-100k
     type: chat_template
-    split: train
+    split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

examples/gemma2/qlora.yml

@@ -11,8 +11,11 @@ chat_template: gemma
 datasets:
   - path: cgato/SlimOrcaDedupCleaned
     type: chat_template
-    chat_template: gemma
     drop_system_message: true
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+
 val_set_size: 0.0
 output_dir: ./outputs/out
 

examples/jamba/qlora_fsdp_large.yaml

@@ -4,11 +4,15 @@ tokenizer_type: AutoTokenizer
 load_in_4bit: true
 strict: false
 use_tensorboard: true
+chat_template: jamba
 datasets:
   - path: cgato/SlimOrcaDedupCleaned
     type: chat_template
-    chat_template: jamba
     drop_system_message: true
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: jamba-large-fsdp-qlora-ft

examples/llama-3/fft-8b-liger-fsdp.yaml

@@ -14,6 +14,10 @@ datasets:
   - path: mlabonne/FineTome-100k
     type: chat_template
     split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.02
 output_dir: ./outputs/out

examples/phi/lora-3.5.yaml

@@ -10,7 +10,6 @@ chat_template: phi_3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    chat_template: phi_3
     field_messages: messages
     message_field_role: role
     message_field_content: content

src/axolotl/cli/__init__.py

@@ -272,7 +272,7 @@ def do_inference_gradio(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template)
+        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 

src/axolotl/core/trainer_builder.py

@@ -895,13 +895,13 @@ 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, metrics=None):
         # 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, metrics=metrics)
 
 
 class AxolotlMambaTrainer(AxolotlTrainer):
@@ -1595,7 +1595,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
         if self.cfg.chat_template:
             training_arguments_kwargs["chat_template"] = get_chat_template(
-                self.cfg.chat_template
+                self.cfg.chat_template,
+                tokenizer=self.tokenizer,
             )
 
         if self.cfg.rl == "orpo":

src/axolotl/monkeypatch/multipack.py

@@ -27,18 +27,15 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
 ]
 
 
-# def patch_for_multipack(model_type, model_name=None, is_remote_code=False):
-def patch_for_multipack(model_type, model_name=None, has_remote_code=False):
+def patch_for_multipack(model_type, model_name=None, is_remote_code=False):
     if model_type == "gemmoe":
         patch_remote(model_name, ".configuration_gemmoe", ".modeling_gemmoe")
     elif model_type == "deepseek_v2":
         patch_remote(model_name, ".configuration_deepseek", ".modeling_deepseek")
-    # elif hasattr(transformers, "modeling_flash_attention_utils") and not is_remote_code:
-    elif hasattr(transformers, "modeling_flash_attention_utils"):
-        if not has_remote_code:
-            transformers.modeling_flash_attention_utils._get_unpad_data = (  # pylint: disable=protected-access
-                get_unpad_data
-            )
+    elif hasattr(transformers, "modeling_flash_attention_utils") and not is_remote_code:
+        transformers.modeling_flash_attention_utils._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
         if model_type == "mixtral" and is_deepspeed_zero3_enabled():
             patch_mixtral_moe_forward_zero3()
         return

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -57,6 +57,7 @@ class ChatTemplate(str, Enum):
     jinja = "jinja"  # pylint: disable=invalid-name
     qwen_25 = "qwen_25"  # pylint: disable=invalid-name
     tokenizer_default = "tokenizer_default"  # pylint: disable=invalid-name
+    exaone = "exaone"  # pylint: disable=invalid-name
 
 
 class DeprecatedParameters(BaseModel):

src/axolotl/utils/models.py

@@ -394,15 +394,10 @@ class ModelLoader:
             and self.cfg.flash_attention
             and self.cfg.sample_packing
         ):
-            has_remote_code = (
-                "auto_map" in self.model_config
-                and self.model_type in self.model_config["auto_map"]
-            )
-
             patch_for_multipack(
                 self.cfg.model_config_type,
                 model_name=self.cfg.base_model,
-                has_remote_code=has_remote_code,
+                is_remote_code=self.cfg.trust_remote_code,
             )
 
             if self.cfg.is_llama_derived_model:
@@ -645,9 +640,7 @@ class ModelLoader:
                 self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
                     **self.model_config.quantization_config
                 )
-        elif self.cfg.adapter == "qlora" and (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        ):
+        elif self.cfg.adapter == "qlora" and self.model_kwargs["load_in_4bit"]:
             bnb_config = {
                 "load_in_4bit": True,
                 "llm_int8_threshold": 6.0,
@@ -670,9 +663,7 @@ class ModelLoader:
             self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
                 **bnb_config,
             )
-        elif self.cfg.adapter == "lora" and (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        ):
+        elif self.cfg.adapter == "lora" and self.model_kwargs["load_in_8bit"]:
             bnb_config = {
                 "load_in_8bit": True,
             }
@@ -685,10 +676,8 @@ class ModelLoader:
 
         # no longer needed per https://github.com/huggingface/transformers/pull/26610
         if "quantization_config" in self.model_kwargs or self.cfg.gptq:
-            if "load_in_8bit" in self.model_kwargs:
-                del self.model_kwargs["load_in_8bit"]
-            if "load_in_4bit" in self.model_kwargs:
-                del self.model_kwargs["load_in_4bit"]
+            self.model_kwargs.pop("load_in_8bit", None)
+            self.model_kwargs.pop("load_in_4bit", None)
 
     def set_attention_config(self) -> None:
         """
@@ -973,17 +962,10 @@ class ModelLoader:
         if is_deepspeed_zero3_enabled():
             skip_prepare_model_for_kbit_training = True
 
-        is_load_in_8bit = (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        )
-        is_load_in_4bit = (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        )
-
         if (
             not skip_prepare_model_for_kbit_training
             and self.cfg.adapter in ["lora", "qlora"]
-            and (is_load_in_8bit or is_load_in_4bit)
+            and (self.cfg.load_in_8bit or self.cfg.load_in_4bit)
         ):
             LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
             self.model = prepare_model_for_kbit_training(
@@ -1121,16 +1103,10 @@ class ModelLoader:
         # ---------------------------------------------------------
         #  put model to accelerator
         # ---------------------------------------------------------
-        is_load_in_8bit = (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        )
-        is_load_in_4bit = (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        )
         if (
             self.cfg.ddp
-            and not is_load_in_8bit
-            and not (self.cfg.rl and is_load_in_4bit)
+            and not self.cfg.load_in_8bit
+            and not (self.cfg.rl and self.cfg.load_in_4bit)
             and not skip_move_to_device
         ):
             # TODO revaldate this conditional

tests/e2e/multigpu/test_llama.py

@@ -14,7 +14,7 @@ from huggingface_hub import snapshot_download
 
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import is_hopper, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
@@ -59,7 +59,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -116,7 +116,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 50,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -144,6 +144,146 @@ class TestMultiGPULlama(unittest.TestCase):
             ]
         )
 
+    @pytest.mark.skipif(is_hopper(), reason="h100 doesn't support 8-bit lora")
+    @with_temp_dir
+    def test_dpo_lora_ddp(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "TinyLlama/TinyLlama_v1.1",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 2048,
+                "sample_packing": False,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "rl": "dpo",
+                "chat_template": "llama3",
+                "datasets": [
+                    {
+                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
+                        "type": "chat_template.default",
+                        "field_messages": "conversation",
+                        "field_chosen": "chosen",
+                        "field_rejected": "rejected",
+                        "message_field_role": "role",
+                        "message_field_content": "content",
+                        "roles": {
+                            "system": ["system"],
+                            "user": ["user"],
+                            "assistant": ["assistant"],
+                        },
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 15,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "warmup_steps": 0,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": 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(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
+    @with_temp_dir
+    def test_dpo_qlora_ddp(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM-135M",
+                "sequence_len": 2048,
+                "sample_packing": False,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "rl": "dpo",
+                "chat_template": "chatml",
+                "datasets": [
+                    {
+                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
+                        "type": "chat_template.default",
+                        "field_messages": "conversation",
+                        "field_chosen": "chosen",
+                        "field_rejected": "rejected",
+                        "message_field_role": "role",
+                        "message_field_content": "content",
+                        "roles": {
+                            "system": ["system"],
+                            "user": ["user"],
+                            "assistant": ["assistant"],
+                        },
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 15,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "warmup_steps": 0,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": 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(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
     @with_temp_dir
     def test_fsdp(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -165,7 +305,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -231,7 +371,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -273,7 +413,6 @@ class TestMultiGPULlama(unittest.TestCase):
             ]
         )
 
-    @pytest.mark.skip("disabled due to upstream issue")
     @with_temp_dir
     def test_fsdp_qlora_prequant_packed(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -282,6 +421,7 @@ class TestMultiGPULlama(unittest.TestCase):
                 "base_model": "axolotl-ai-co/TinyLlama_v1.1-bnb-nf4-bf16",
                 "tokenizer_type": "AutoTokenizer",
                 "adapter": "qlora",
+                "mean_resizing_embeddings": True,
                 "load_in_4bit": True,
                 "lora_r": 8,
                 "lora_alpha": 16,
@@ -297,7 +437,7 @@ class TestMultiGPULlama(unittest.TestCase):
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "pad_token": "<|end_of_text|>",
+                    "pad_token": "</s>",
                 },
                 "datasets": [
                     {
@@ -307,7 +447,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -373,7 +513,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -432,7 +572,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,

tests/e2e/multigpu/test_qwen2.py

@@ -47,7 +47,7 @@ class TestMultiGPUQwen2(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "warmup_steps": 20,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 2,

tests/e2e/patched/test_4d_multipack_llama.py

@@ -13,7 +13,7 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import require_torch_2_1_1, with_temp_dir
+from ..utils import require_torch_2_3_1, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -24,7 +24,7 @@ class Test4dMultipackLlama(unittest.TestCase):
     Test case for Llama models using 4d attention with multipack
     """
 
-    @require_torch_2_1_1
+    @require_torch_2_3_1
     @with_temp_dir
     def test_sdp_lora_packing(self, temp_dir):
         # pylint: disable=duplicate-code

tests/e2e/utils.py

@@ -9,6 +9,8 @@ from functools import wraps
 from importlib.metadata import version
 from pathlib import Path
 
+import torch
+
 
 def with_temp_dir(test_func):
     @wraps(test_func)
@@ -35,13 +37,18 @@ def most_recent_subdir(path):
     return subdir
 
 
-def require_torch_2_1_1(test_case):
+def require_torch_2_3_1(test_case):
     """
-    Decorator marking a test that requires torch >= 2.1.1
+    Decorator marking a test that requires torch >= 2.3.1
     """
 
-    def is_min_2_1_1():
+    def is_min_2_3_1():
         torch_version = version("torch")
-        return torch_version >= "2.1.1"
+        return torch_version >= "2.3.1"
 
-    return unittest.skipUnless(is_min_2_1_1(), "test torch 2.1.1")(test_case)
+    return unittest.skipUnless(is_min_2_3_1(), "test torch 2.3.1")(test_case)
+
+
+def is_hopper():
+    compute_capability = torch.cuda.get_device_capability()
+    return compute_capability == (9, 0)

tests/test_datasets.py

@@ -367,43 +367,44 @@ class TestDatasetPreparation(unittest.TestCase):
     def test_load_local_hub_with_revision(self):
         """Verify that a local copy of a hub dataset can be loaded with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:
-            tmp_ds_path = Path("mhenrichsen/alpaca_2k_test")
-            tmp_ds_path.mkdir(parents=True, exist_ok=True)
-            snapshot_download(
-                repo_id="mhenrichsen/alpaca_2k_test",
-                repo_type="dataset",
-                local_dir=tmp_ds_path,
-                revision="d05c1cb",
-            )
+            with tempfile.TemporaryDirectory() as tmp_dir2:
+                tmp_ds_path = Path(tmp_dir2) / "mhenrichsen/alpaca_2k_test"
+                tmp_ds_path.mkdir(parents=True, exist_ok=True)
+                snapshot_download(
+                    repo_id="mhenrichsen/alpaca_2k_test",
+                    repo_type="dataset",
+                    local_dir=tmp_ds_path,
+                    revision="d05c1cb",
+                )
 
-            prepared_path = Path(tmp_dir) / "prepared"
-            cfg = DictDefault(
-                {
-                    "tokenizer_config": "huggyllama/llama-7b",
-                    "sequence_len": 1024,
-                    "datasets": [
-                        {
-                            "path": "mhenrichsen/alpaca_2k_test",
-                            "ds_type": "parquet",
-                            "type": "alpaca",
-                            "data_files": [
-                                "mhenrichsen/alpaca_2k_test/alpaca_2000.parquet",
-                            ],
-                            "revision": "d05c1cb",
-                        },
-                    ],
-                }
-            )
+                prepared_path = Path(tmp_dir) / "prepared"
+                cfg = DictDefault(
+                    {
+                        "tokenizer_config": "huggyllama/llama-7b",
+                        "sequence_len": 1024,
+                        "datasets": [
+                            {
+                                "path": "mhenrichsen/alpaca_2k_test",
+                                "ds_type": "parquet",
+                                "type": "alpaca",
+                                "data_files": [
+                                    f"{tmp_ds_path}/alpaca_2000.parquet",
+                                ],
+                                "revision": "d05c1cb",
+                            },
+                        ],
+                    }
+                )
 
-            dataset, _ = load_tokenized_prepared_datasets(
-                self.tokenizer, cfg, prepared_path
-            )
+                dataset, _ = load_tokenized_prepared_datasets(
+                    self.tokenizer, cfg, prepared_path
+                )
 
-            assert len(dataset) == 2000
-            assert "input_ids" in dataset.features
-            assert "attention_mask" in dataset.features
-            assert "labels" in dataset.features
-            shutil.rmtree(tmp_ds_path)
+                assert len(dataset) == 2000
+                assert "input_ids" in dataset.features
+                assert "attention_mask" in dataset.features
+                assert "labels" in dataset.features
+                shutil.rmtree(tmp_ds_path)
 
 
 if __name__ == "__main__":