Merge branch 'main' into hymba_multipack2

* fix: allow trainer builder to use custom jinja chat template

* chore: use get_chat_template_from_config

Co-authored-by: Chirag Jain <jain.chirag925@gmail.com>

* fix: swap imports

---------

Co-authored-by: Chirag Jain <jain.chirag925@gmail.com>

docs/lr_groups.qmd

@@ -1,29 +0,0 @@
----
-title: Learning Rate Groups
-description: "Setting different learning rates by module name"
----
-
-## Background
-
-Inspired by LoRA+, Axolotl allows practitioners to specify separate learning rates for each module or groups of
-modules in a model.
-
-## Example
-
-```yaml
-lr_groups:
-  - name: o_proj
-    modules:
-      - self_attn.o_proj.weight
-    lr: 1e-6
-  - name: q_proj
-    modules:
-      - model.layers.2.self_attn.q_proj.weight
-    lr: 1e-5
-
-learning_rate: 2e-5
-```
-
-In this example, we have a default learning rate of 2e-5 across the entire model, but we have a separate learning rate
-of 1e-6 for all the self attention `o_proj` modules across all layers, and a learning are of 1e-5 to the 3rd layer's
-self attention `q_proj` module.

examples/hymba/fft-1.5b.yml

@@ -0,0 +1,58 @@
+base_model: nvidia/Hymba-1.5B-Base
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+trust_remote_code: true
+
+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: 5
+evals_per_epoch: 2
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: <|end_of_text|>

examples/hymba/qlora-1.5b.yml

@@ -0,0 +1,73 @@
+base_model: nvidia/Hymba-1.5B-Base
+
+load_in_8bit: false
+load_in_4bit: True
+strict: false
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+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: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+trust_remote_code: true
+
+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: 5
+evals_per_epoch: 2
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: <|end_of_text|>

requirements.txt

@@ -61,4 +61,4 @@ antlr4-python3-runtime==4.13.2
 torchao==0.7.0
 schedulefree==1.3.0
 
-axolotl-contribs-lgpl==0.0.1b2
+axolotl-contribs-lgpl==0.0.2

src/axolotl/cli/main.py

@@ -93,7 +93,7 @@ def evaluate(config: str, accelerate: bool, **kwargs):
 @click.argument("config", type=click.Path(exists=True, path_type=str))
 @click.option(
     "--accelerate/--no-accelerate",
-    default=True,
+    default=False,
     help="Use accelerate launch for multi-GPU inference",
 )
 @click.option(
@@ -124,7 +124,7 @@ def inference(
     if lora_model_dir:
         kwargs["lora_model_dir"] = lora_model_dir
     if base_model:
-        kwargs["output_dir"] = base_model
+        kwargs["base_model"] = base_model
 
     if accelerate:
         base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.inference"]

src/axolotl/core/trainer_builder.py

@@ -68,7 +68,7 @@ from axolotl.utils.callbacks import (
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
 from axolotl.utils.callbacks.profiler import PytorchProfilerCallback
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import get_chat_template_from_config
 from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
@@ -244,10 +244,6 @@ class AxolotlTrainingMixins:
         default=None,
         metadata={"help": "Scale the learning rate for the embedding layers."},
     )
-    lr_groups: Optional[list[dict]] = field(
-        default=None,
-        metadata={"help": "Specify learning rate groups for with different LRs."},
-    )
     embedding_lr: Optional[float] = field(
         default=None,
         metadata={"help": "absolute learning rate for the embedding layers."},
@@ -466,96 +462,11 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
             )
         return super()._wrap_model(model, training=training, dataloader=dataloader)
 
-    def create_optimizer_grouped_parameters(self, opt_model, optimizer_kwargs):
-        decay_parameters = self.get_decay_parameter_names(opt_model)
-        params = {
-            "to_weight_decay": {},  # LayerNorm and bias
-            "embeddings": {},  # lm_head, embed_tokens,
-            "no_weight_decay": {},
-        }
-        lr_groups_lookup = {}
-        lr_groups_learning_rates = {}
-        if self.args.lr_groups:
-            for lr_group in self.args.lr_groups:
-                group_name = lr_group["name"]
-                group_modules = lr_group["modules"]
-                for module in group_modules:
-                    lr_groups_lookup[module] = group_name
-                lr_groups_learning_rates[group_name] = lr_group["lr"]
-                params[f"to_weight_decay_{group_name}"] = {}
-
-        for name, param in opt_model.named_parameters():
-            if not param.requires_grad:
-                continue
-            if name.endswith("modules_to_save.default.weight") or any(
-                embed_name in name for embed_name in ["embed_tokens", "lm_head"]
-            ):
-                params["embeddings"][name] = param
-            elif name in decay_parameters:
-                if lr_groups_lookup and any(
-                    group_modules in name for group_modules in lr_groups_lookup
-                ):
-                    lr_group_module = [
-                        group_modules
-                        for group_modules in lr_groups_lookup
-                        if group_modules in name
-                    ][0]
-                    group_name = lr_groups_lookup[lr_group_module]
-                    params[f"to_weight_decay_{group_name}"][name] = param
-                else:
-                    params["to_weight_decay"][name] = param
-            else:
-                params["no_weight_decay"][name] = param
-        optimizer_grouped_parameters = []
-        if params["to_weight_decay"]:
-            optimizer_grouped_parameters.append(
-                {
-                    "params": list(params["to_weight_decay"].values()),
-                    "weight_decay": self.args.weight_decay,
-                    "lr": optimizer_kwargs["lr"],
-                }
-            )
-        if params["embeddings"]:
-            lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
-            if self.args.embedding_lr_scale:
-                lr *= self.args.embedding_lr_scale  # pylint: disable=invalid-name
-            elif self.args.embedding_lr:
-                lr = self.args.embedding_lr  # pylint: disable=invalid-name
-            optimizer_grouped_parameters.append(
-                {
-                    "params": list(params["embeddings"].values()),
-                    "weight_decay": 0.0,
-                    "lr": lr,
-                }
-            )
-        if params["no_weight_decay"]:
-            optimizer_grouped_parameters.append(
-                {
-                    "params": list(params["no_weight_decay"].values()),
-                    "weight_decay": 0.0,
-                    "lr": optimizer_kwargs["lr"],
-                }
-            )
-        for group_name, group_lr in lr_groups_learning_rates.items():
-            if params[f"to_weight_decay_{group_name}"]:
-                optimizer_grouped_parameters.append(
-                    {
-                        "params": list(
-                            params[f"to_weight_decay_{group_name}"].values()
-                        ),
-                        "weight_decay": self.args.weight_decay,
-                        "lr": group_lr,
-                    }
-                )
-
-        return optimizer_grouped_parameters
-
     def create_optimizer(self):
         if (
             self.args.loraplus_lr_ratio is None
             and self.args.embedding_lr_scale is None
             and self.args.embedding_lr is None
-            and self.args.lr_groups is None
             and self.args.alternate_optimizer
             not in [
                 "optimi_adamw",
@@ -569,13 +480,59 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
 
         opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
         if self.optimizer is None:  # pylint: disable=access-member-before-definition
+            decay_parameters = self.get_decay_parameter_names(opt_model)
+            params = {
+                "to_weight_decay": {},  # LayerNorm and bias
+                "embeddings": {},  # lm_head, embed_tokens,
+                "no_weight_decay": {},
+            }
+
             optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
                 self.args,
                 opt_model,
             )
-            optimizer_grouped_parameters = self.create_optimizer_grouped_parameters(
-                opt_model, optimizer_kwargs
-            )
+
+            for name, param in opt_model.named_parameters():
+                if not param.requires_grad:
+                    continue
+                if name.endswith("modules_to_save.default.weight") or any(
+                    embed_name in name for embed_name in ["embed_tokens", "lm_head"]
+                ):
+                    params["embeddings"][name] = param
+                elif name in decay_parameters:
+                    params["to_weight_decay"][name] = param
+                else:
+                    params["no_weight_decay"][name] = param
+            optimizer_grouped_parameters = []
+            if params["to_weight_decay"]:
+                optimizer_grouped_parameters.append(
+                    {
+                        "params": list(params["to_weight_decay"].values()),
+                        "weight_decay": self.args.weight_decay,
+                        "lr": optimizer_kwargs["lr"],
+                    }
+                )
+            if params["embeddings"]:
+                lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
+                if self.args.embedding_lr_scale:
+                    lr *= self.args.embedding_lr_scale  # pylint: disable=invalid-name
+                elif self.args.embedding_lr:
+                    lr = self.args.embedding_lr  # pylint: disable=invalid-name
+                optimizer_grouped_parameters.append(
+                    {
+                        "params": list(params["embeddings"].values()),
+                        "weight_decay": 0.0,
+                        "lr": lr,
+                    }
+                )
+            if params["no_weight_decay"]:
+                optimizer_grouped_parameters.append(
+                    {
+                        "params": list(params["no_weight_decay"].values()),
+                        "weight_decay": 0.0,
+                        "lr": optimizer_kwargs["lr"],
+                    }
+                )
 
             if self.args.loraplus_lr_ratio is not None:
                 loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
@@ -592,7 +549,6 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
             elif (
                 self.args.embedding_lr_scale is not None
                 or self.args.embedding_lr is not None
-                or self.args.lr_groups is not None
             ):
                 self.optimizer = (  # pylint: disable=attribute-defined-outside-init
                     optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
@@ -1808,7 +1764,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         ] = self.cfg.loraplus_lr_embedding
         training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
         training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
-        training_arguments_kwargs["lr_groups"] = self.cfg.lr_groups
 
         if self.cfg.lr_scheduler in ["one_cycle", "log_sweep"]:
             training_arguments_kwargs["lr_scheduler_type"] = "cosine"
@@ -1879,8 +1834,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["model_type"] = self.cfg.model_config_type
         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,
+            training_arguments_kwargs["chat_template"] = get_chat_template_from_config(
+                cfg=self.cfg,
                 tokenizer=self.tokenizer,
             )
 

src/axolotl/monkeypatch/multipack.py

@@ -25,6 +25,7 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "gemmoe",
     "starcoder2",
     "deepseek_v2",
+    "hymba",
 ]
 
 

src/axolotl/train.py

@@ -1,5 +1,6 @@
 """Prepare and train a model on a dataset. Can also infer from a model or merge lora"""
 
+import inspect
 import os
 import signal
 import sys
@@ -126,7 +127,20 @@ def train(
     )
 
     if cfg.fix_untrained_tokens:
-        fix_untrained_tokens(model, tokenizer, train_dataset)
+        # check if the `token_ids_to_fix` kwarg exists in the fix_untrained_tokens args
+        sig = inspect.signature(fix_untrained_tokens)
+        # if the function has the `token_ids_to_fix` arg, and fix_untrained_tokens is a list
+        if "token_ids_to_fix" in sig.parameters and isinstance(
+            cfg.fix_untrained_tokens, list
+        ):
+            fix_untrained_tokens(
+                model,
+                tokenizer,
+                train_dataset,
+                token_ids_to_fix=cfg.fix_untrained_tokens,
+            )
+        else:
+            fix_untrained_tokens(model, tokenizer, train_dataset)
         if cfg.local_rank == 0:
             model.save_pretrained(
                 str(Path(cfg.output_dir)), safe_serialization=safe_serialization

src/axolotl/utils/chat_templates.py

@@ -31,6 +31,7 @@ _CHAT_TEMPLATES = {
     "qwen_25": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
     "exaone": "{% for message in messages %}{% if loop.first and message['role'] != 'system' %}{{ '[|system|][|endofturn|]\n' }}{% endif %}{{ '[|' + message['role'] + '|]' + message['content'] }}{% if message['role'] == 'user' %}{{ '\n' }}{% else %}{{ '[|endofturn|]\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '[|assistant|]' }}{% endif %}",
     "metharme": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% set system_message = 'Enter RP mode. You shall reply to the user while staying in character. Your responses must be detailed, creative, immersive, and drive the scenario forward.' %}{% endif %}{{ '<|system|>' + system_message }}{% for message in loop_messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|user|>' + content.strip() }}{% elif message['role'] == 'assistant' %}{{ '<|model|>'  + content.strip() }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|model|>' }}{% else %}{{ eos_token }}{% endif %}",
+    "hymba": "{{'<extra_id_0>System'}}{% for message in messages %}{% if message['role'] == 'system' %}{{'\n' + message['content'].strip()}}{% if tools or contexts %}{{'\n'}}{% endif %}{% endif %}{% endfor %}{% if tools %}{% for tool in tools %}{{ '\n<tool> ' + tool|tojson + ' </tool>' }}{% endfor %}{% endif %}{% if contexts %}{% if tools %}{{'\n'}}{% endif %}{% for context in contexts %}{{ '\n<context> ' + context.strip() + ' </context>' }}{% endfor %}{% endif %}{{'\n\n'}}{% for message in messages %}{% if message['role'] == 'user' %}{{ '<extra_id_1>User\n' + message['content'].strip() + '\n' }}{% elif message['role'] == 'assistant' %}{{ '<extra_id_1>Assistant\n' + message['content'].strip() + '\n' }}{% elif message['role'] == 'tool' %}{{ '<extra_id_1>Tool\n' + message['content'].strip() + '\n' }}{% endif %}{% endfor %}{%- if add_generation_prompt %}{{'<extra_id_1>Assistant\n'}}{%- endif %}",
 }
 
 

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

@@ -145,14 +145,6 @@ class UserDefinedPrompterType(BaseModel):
     field: Optional[str] = None
 
 
-class LrGroup(BaseModel):
-    """Custom learning rate group configuration"""
-
-    name: str
-    modules: List[str]
-    lr: float
-
-
 class SFTDataset(BaseModel):
     """SFT configuration subset"""
 
@@ -474,7 +466,6 @@ class HyperparametersConfig(BaseModel):
     cosine_min_lr_ratio: Optional[float] = None
     cosine_constant_lr_ratio: Optional[float] = None
     lr_div_factor: Optional[float] = None
-    lr_groups: Optional[List[LrGroup]] = None
 
     adam_epsilon: Optional[float] = None
     adam_beta1: Optional[float] = None
@@ -803,7 +794,7 @@ class AxolotlInputConfig(
     chat_template_jinja: Optional[str] = None
     default_system_message: Optional[str] = None
 
-    fix_untrained_tokens: Optional[bool] = None
+    fix_untrained_tokens: Optional[Union[int, List[int]]] = None
 
     # INTERNALS - document for now, generally not set externally
     is_preprocess: Optional[bool] = None
@@ -1638,3 +1629,19 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             else:
                 data["torch_compile"] = False
         return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_hymba_torch_version(cls, data):
+        if "hymba" in data.get("base_model", {}).lower():
+            env_capabilities = data.get("env_capabilities", {})
+            torch_version = env_capabilities.get("torch_version")
+
+            if torch_version is None:
+                import torch
+
+                torch_version = str(torch.__version__).split("+", maxsplit=1)[0]
+
+            if version.parse(torch_version) < version.parse("2.5.0"):
+                raise ValueError("Hymba requires torch version >= 2.5")
+        return data

src/axolotl/utils/data/pretraining.py

@@ -28,8 +28,10 @@ def encode_pretraining(
     )
     # Convert to PyTorch tensors
     input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
+    targets = [torch.tensor(seq) for seq in res["input_ids"]]
     attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
     new_input_ids = []
+    new_labels = []
     new_attention_mask = []
     # Append EOS and PAD tokens to input_ids, and correct attention_mask
     for i, _ in enumerate(input_ids):
@@ -40,22 +42,34 @@ def encode_pretraining(
             ),
             dim=0,
         )
+        targets[i] = torch.cat(
+            (
+                targets[i],
+                torch.tensor([tokenizer.eos_token_id, -100]),
+            ),
+            dim=0,
+        )
         attention_mask[i] = torch.cat((attention_mask[i], torch.tensor([1, 0])), dim=0)
 
     # Concatenate tokens so that their lengths are less than max_tokens
     buffer_input_ids = torch.tensor([], dtype=torch.long)
+    buffer_labels = torch.tensor([], dtype=torch.long)
     buffer_attention_mask = torch.tensor([], dtype=torch.long)
 
-    for ids, mask in zip(input_ids, attention_mask):
+    for ids, labels, mask in zip(input_ids, targets, attention_mask):
         if buffer_input_ids.numel() == max_tokens:
             new_input_ids.append(buffer_input_ids)
+            new_labels.append(buffer_labels)
             new_attention_mask.append(buffer_attention_mask)
             buffer_input_ids = torch.tensor([], dtype=torch.long)
+            buffer_labels = torch.tensor([], dtype=torch.long)
             buffer_attention_mask = torch.tensor([], dtype=torch.long)
             buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
+            buffer_labels = torch.cat((buffer_labels, labels), dim=0)
             buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
         elif buffer_input_ids.numel() + ids.numel() <= max_tokens:
             buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
+            buffer_labels = torch.cat((buffer_labels, labels), dim=0)
             buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
         else:
             buffer_input_ids = torch.cat(
@@ -69,6 +83,17 @@ def encode_pretraining(
                 ),
                 dim=0,
             )
+            buffer_labels = torch.cat(
+                (
+                    buffer_labels,
+                    torch.full(
+                        (max_tokens - buffer_labels.numel(),),
+                        -100,
+                        dtype=torch.long,
+                    ),
+                ),
+                dim=0,
+            )
             buffer_attention_mask = torch.cat(
                 (
                     buffer_attention_mask,
@@ -81,11 +106,14 @@ def encode_pretraining(
                 dim=0,
             )
             new_input_ids.append(buffer_input_ids)
+            new_labels.append(buffer_labels)
             new_attention_mask.append(buffer_attention_mask)
             buffer_input_ids = torch.tensor([], dtype=torch.long)
+            buffer_labels = torch.tensor([], dtype=torch.long)
             buffer_attention_mask = torch.tensor([], dtype=torch.long)
 
             buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
+            buffer_labels = torch.cat((buffer_labels, labels), dim=0)
             buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
 
     if buffer_input_ids.numel() > 0:  # for any leftover tokens
@@ -101,6 +129,17 @@ def encode_pretraining(
                 ),
                 dim=0,
             )
+            buffer_labels = torch.cat(
+                (
+                    buffer_labels,
+                    torch.full(
+                        (max_tokens - buffer_labels.numel(),),
+                        -100,
+                        dtype=torch.long,
+                    ),
+                ),
+                dim=0,
+            )
             buffer_attention_mask = torch.cat(
                 (
                     buffer_attention_mask,
@@ -113,11 +152,12 @@ def encode_pretraining(
                 dim=0,
             )
         new_input_ids.append(buffer_input_ids)
+        new_labels.append(buffer_labels)
         new_attention_mask.append(buffer_attention_mask)
 
     ret = {
         "input_ids": [seq.tolist() for seq in new_input_ids],
-        "labels": [seq.tolist() for seq in new_input_ids],
+        "labels": [seq.tolist() for seq in new_labels],
         "attention_mask": [seq.tolist() for seq in new_attention_mask],
     }
 

src/axolotl/utils/models.py

@@ -409,6 +409,7 @@ class ModelLoader:
             and self.cfg.sample_packing
         ):
             if "auto_map" in self.model_config:
+                # some model config objects are not subscriptable
                 try:
                     auto_map_config = self.model_config["auto_map"]
                 except TypeError:

tests/e2e/test_optimizers.py

@@ -67,8 +67,8 @@ class TestCustomOptimizers(unittest.TestCase):
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.bin").exists()
 
-    @with_temp_dir
     @require_torch_2_5_1
+    @with_temp_dir
     def test_adopt_adamw(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(

tests/e2e/test_packing_loss.py

@@ -14,7 +14,7 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_tensorboard, with_temp_dir
+from .utils import check_tensorboard, require_torch_2_5_1, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -68,3 +68,129 @@ class TestPackedLlama(unittest.TestCase):
         check_tensorboard(
             temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
         )
+
+
+class TestUnpackedHymba(unittest.TestCase):
+    """
+    Test case for Unpacked training of hymba models
+    """
+
+    @require_torch_2_5_1
+    @with_temp_dir
+    def test_loss_unpacked(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "nvidia/Hymba-1.5B-Base",
+                "trust_remote_code": True,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 32,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_modules": [
+                    "gate_proj",
+                    "down_proj",
+                    "up_proj",
+                    "q_proj",
+                    "v_proj",
+                    "k_proj",
+                    "o_proj",
+                ],
+                "sequence_len": 1024,
+                "sample_packing": False,
+                "flash_attention": True,
+                "val_set_size": 0.0,
+                "datasets": [
+                    {
+                        "path": "vicgalle/alpaca-gpt4",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "use_tensorboard": True,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+        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)
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+        )
+
+
+class TestPackedHymba(unittest.TestCase):
+    """
+    Test case for Packed training of hymba models
+    """
+
+    @require_torch_2_5_1
+    @with_temp_dir
+    def test_loss_packed(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "nvidia/Hymba-1.5B-Base",
+                "trust_remote_code": True,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 32,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_modules": [
+                    "gate_proj",
+                    "down_proj",
+                    "up_proj",
+                    "q_proj",
+                    "v_proj",
+                    "k_proj",
+                    "o_proj",
+                ],
+                "sequence_len": 1024,
+                "sample_packing": True,
+                "flash_attention": True,
+                "val_set_size": 0.0,
+                "datasets": [
+                    {
+                        "path": "vicgalle/alpaca-gpt4",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "use_tensorboard": True,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+        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)
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+        )