move import of llmcompressor to reset session inside test

Signed-off-by: Rahul Tuli <rtuli@redhat.com>

.github/workflows/main.yml

@@ -30,7 +30,7 @@ jobs:
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.0
-            axolotl_extras:
+            axolotl_extras: vllm
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/preview-docs.yml

@@ -4,12 +4,6 @@ on:
   pull_request:
     types: [opened, synchronize, reopened]
 
-    # Run the workflow only when one of these files changes
-    paths:
-      - '**/*.md'      # any Markdown file
-      - '**/*.qmd'     # any Quarto file
-      - '_quarto.yaml'
-
 permissions:
   checks: write
   contents: write

.runpod/tests.json

@@ -1,90 +0,0 @@
-{
-  "tests": [
-    {
-      "name": "quick_smoke_test_sft",
-      "input": {
-        "user_id": "user",
-        "model_id": "llama-test",
-        "run_id": "llama-test",
-        "credentials": {
-          "wandb_api_key": "",
-          "hf_token": ""
-        },
-        "args": {
-          "base_model": "HuggingFaceTB/SmolLM2-135M",
-          "model_type": "AutoModelForCausalLM",
-          "tokenizer_type": "AutoTokenizer",
-          "load_in_4bit": true,
-          "strict": false,
-          "datasets": [
-            {
-              "path": "mhenrichsen/alpaca_2k_test",
-              "type": "alpaca",
-              "split": "train[:10%]"
-            }
-          ],
-          "val_set_size": 0.02,
-          "output_dir": "./outputs/lora-out",
-          "sequence_len": 4096,
-          "sample_packing": true,
-          "eval_sample_packing": false,
-          "pad_to_sequence_len": true,
-          "adapter": "qlora",
-          "lora_r": 32,
-          "lora_alpha": 64,
-          "lora_dropout": 0.05,
-          "lora_target_linear": true,
-          "lora_modules_to_save": [
-            "embed_tokens",
-            "lm_head"
-          ],
-          "gradient_accumulation_steps": 2,
-          "micro_batch_size": 1,
-          "num_epochs": 1,
-          "optimizer": "adamw_torch_fused",
-          "lr_scheduler": "cosine",
-          "learning_rate": 0.0002,
-          "train_on_inputs": false,
-          "group_by_length": false,
-          "bf16": "auto",
-          "tf32": true,
-          "gradient_checkpointing": true,
-          "logging_steps": 1,
-          "flash_attention": true,
-          "warmup_steps": 1,
-          "evals_per_epoch": 1,
-          "eval_max_new_tokens": 128,
-          "saves_per_epoch": 1,
-          "weight_decay": 0.0,
-          "special_tokens": {
-            "pad_token": "<|endoftext|>"
-          },
-          "max_steps": 20
-        }
-      },
-      "timeout": 100000
-    }
-  ],
-  "config": {
-    "gpuTypeId": "NVIDIA GeForce RTX 4090",
-    "gpuCount": 1,
-    "containerDiskInGb": 200,
-    "env": [
-      {
-        "key": "TOKENIZER",
-        "value": ""
-      },
-      {
-        "key": "DISABLE_LOG_STATS",
-        "value": "true"
-      }
-    ],
-    "allowedCudaVersions": [
-      "12.8",
-      "12.7",
-      "12.6",
-      "12.5",
-      "12.4"
-    ]
-  }
-}

README.md

@@ -51,7 +51,7 @@ Features:
 
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
-- PyTorch ≥2.5.1
+- PyTorch ≥2.4.1
 
 ### Installation
 

requirements.txt

@@ -18,7 +18,7 @@ accelerate==1.6.0
 datasets==3.5.0
 deepspeed>=0.15.4
 trl==0.17.0
-hf_xet==1.1.0
+hf_xet==1.0.0
 hqq==0.2.5
 
 optimum==1.16.2

setup.py

@@ -67,11 +67,13 @@ def parse_requirements(extras_require_map):
             if (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 # _install_requires.append("xformers==0.0.29.post3")  # xformers seems to be hard pinned to 2.6.0
+                extras_require_map["vllm"] = ["vllm==0.8.5"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append(
                     "xformers==0.0.29.post2"
                 )  # vllm needs post2 w torch 2.6
+                extras_require_map["vllm"] = ["vllm==0.8.5"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
@@ -145,7 +147,7 @@ extras_require = {
         "ray[train]",
     ],
     "vllm": [
-        "vllm==0.8.5",
+        "vllm==0.7.2",
     ],
     "llmcompressor": [
         "llmcompressor==0.5.1",

src/axolotl/cli/__init__.py

@@ -2,7 +2,4 @@
 
 import os
 
-from axolotl.logging_config import configure_logging
-
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
-configure_logging()

src/axolotl/cli/checks.py

@@ -8,6 +8,9 @@ from accelerate.commands.config import config_args
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
 
+from axolotl.logging_config import configure_logging
+
+configure_logging()
 LOG = logging.getLogger(__name__)
 
 

src/axolotl/cli/config.py

@@ -5,7 +5,6 @@ import logging
 import os
 import tempfile
 from pathlib import Path
-from tempfile import NamedTemporaryFile
 from typing import Union
 from urllib.parse import urlparse
 
@@ -159,9 +158,7 @@ def plugin_set_cfg(cfg: DictDefault):
         plugin_manager.cfg = cfg
 
 
-def load_cfg(
+def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefault:
-    config: str | Path | DictDefault = Path("examples/"), **kwargs
-) -> DictDefault:
     """
     Loads the `axolotl` configuration stored at `config`, validates it, and performs
     various setup.
@@ -173,24 +170,13 @@ def load_cfg(
     Returns:
         `DictDefault` mapping configuration keys to values.
     """
-    if isinstance(config, (str, Path)):
+    config = check_remote_config(config)
-        config = check_remote_config(config)
+    if Path(config).is_dir():
-        if Path(config).is_dir():
+        config = choose_config(Path(config))
-            config = choose_config(Path(config))
 
-        # Load the config from the yaml file
+    # Load the config from the yaml file
-        with open(config, encoding="utf-8") as file:
+    with open(config, encoding="utf-8") as file:
-            cfg: DictDefault = DictDefault(yaml.safe_load(file))
+        cfg: DictDefault = DictDefault(yaml.safe_load(file))
-
-        cfg.axolotl_config_path = config
-    else:
-        cfg = config
-        with NamedTemporaryFile(
-            mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
-        ) as temp_file:
-            temp_file.write(yaml.dump(config.to_dict()))
-            temp_file.close()
-        cfg.axolotl_config_path = temp_file.name
 
     # If there are any options passed in the cli, if it is something that seems valid
     # from the yaml, then overwrite the value
@@ -204,6 +190,8 @@ def load_cfg(
             else:
                 cfg[k] = kwargs[k]
 
+    cfg.axolotl_config_path = config
+
     try:
         device_props = torch.cuda.get_device_properties("cuda")
         gpu_version = "sm_" + str(device_props.major) + str(device_props.minor)

src/axolotl/cli/utils.py

@@ -20,9 +20,11 @@ from transformers import (
     ProcessorMixin,
 )
 
+from axolotl.logging_config import configure_logging
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_processor, load_tokenizer
 
+configure_logging()
 LOG = logging.getLogger(__name__)
 
 

src/axolotl/common/datasets.py

@@ -47,7 +47,7 @@ def sample_dataset(dataset: Dataset, num_samples: int) -> Dataset:
 def load_datasets(
     *,
     cfg: DictDefault,
-    cli_args: PreprocessCliArgs | TrainerCliArgs | None = None,
+    cli_args: Union[PreprocessCliArgs, TrainerCliArgs],
 ) -> TrainDatasetMeta:
     """
     Loads one or more training or evaluation datasets, calling
@@ -64,8 +64,7 @@ def load_datasets(
     tokenizer = load_tokenizer(cfg)
     processor = load_processor(cfg, tokenizer=tokenizer) if cfg.processor_type else None
     preprocess_iterable = (
-        cli_args
+        hasattr(cli_args, "iterable")
-        and hasattr(cli_args, "iterable")
         and cli_args.iterable is not None
         and cli_args.iterable
     )
@@ -77,7 +76,7 @@ def load_datasets(
         preprocess_iterable=preprocess_iterable,
     )
 
-    if cli_args and (
+    if (
         cli_args.debug
         or cfg.debug
         or cli_args.debug_text_only

src/axolotl/core/trainer_builder.py

@@ -488,7 +488,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         # these are all the "standard" kwargs that are def used
         training_arguments_kwargs["max_steps"] = (
-            self.cfg.max_steps if self.cfg.max_steps else -1
+            total_num_steps if self.cfg.max_steps else -1
         )
         training_arguments_kwargs["max_seq_length"] = self.cfg.sequence_len
         training_arguments_kwargs["per_device_train_batch_size"] = (

src/axolotl/core/trainers/dpo/trainer.py

@@ -177,8 +177,12 @@ class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
             # dpo trainer may incorrectly prepend the bos_token_id to the dpo outputs
             if res["chosen_input_ids"][0] == processing_class.bos_token_id:
                 res["chosen_input_ids"] = res["chosen_input_ids"][1:]
+                res["chosen_labels"] = res["chosen_labels"][1:]
+                res["chosen_attention_mask"] = res["chosen_attention_mask"][1:]
             if res["rejected_input_ids"][0] == processing_class.bos_token_id:
                 res["rejected_input_ids"] = res["rejected_input_ids"][1:]
+                res["rejected_labels"] = res["rejected_labels"][1:]
+                res["rejected_attention_mask"] = res["rejected_attention_mask"][1:]
 
         return res
 

src/axolotl/core/trainers/grpo/__init__.py

@@ -63,7 +63,6 @@ class GRPOStrategy:
 
         grpo_args_kwargs["max_completion_length"] = trl.max_completion_length
         grpo_args_kwargs["log_completions"] = trl.log_completions
-        grpo_args_kwargs["num_completions_to_print"] = trl.num_completions_to_print
 
         if trl.reward_weights:
             grpo_args_kwargs["reward_weights"] = trl.reward_weights

src/axolotl/evaluate.py

@@ -11,6 +11,7 @@ from accelerate.logging import get_logger
 from datasets import Dataset
 from transformers.trainer import Trainer
 
+from axolotl.logging_config import configure_logging
 from axolotl.train import (
     TrainDatasetMeta,
     setup_model_and_tokenizer,
@@ -23,6 +24,7 @@ project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
 src_dir = os.path.join(project_root, "src")
 sys.path.insert(0, src_dir)
 
+configure_logging()
 LOG = get_logger(__name__)
 
 

src/axolotl/integrations/liger/__init__.py

@@ -151,30 +151,6 @@ class LigerPlugin(BasePlugin):
                 rms_norm=cfg.liger_rms_norm,
                 layer_norm=cfg.liger_layer_norm,
             )
-        elif cfg.model_config_type == "qwen3":
-            from axolotl.integrations.liger.models.qwen3 import (
-                apply_liger_kernel_to_qwen3,
-            )
-
-            apply_liger_kernel_to_qwen3(
-                cross_entropy=cfg.liger_cross_entropy,
-                fused_linear_cross_entropy=cfg.liger_fused_linear_cross_entropy,
-                glu_activation=cfg.liger_glu_activation,
-                rms_norm=cfg.liger_rms_norm,
-                layer_norm=cfg.liger_layer_norm,
-            )
-        elif cfg.model_config_type == "qwen3_moe":
-            from axolotl.integrations.liger.models.qwen3_moe import (
-                apply_liger_kernel_to_qwen3_moe,
-            )
-
-            apply_liger_kernel_to_qwen3_moe(
-                cross_entropy=cfg.liger_cross_entropy,
-                fused_linear_cross_entropy=cfg.liger_fused_linear_cross_entropy,
-                glu_activation=cfg.liger_glu_activation,
-                rms_norm=cfg.liger_rms_norm,
-                layer_norm=cfg.liger_layer_norm,
-            )
         else:
             logging.warning(
                 f"Unsupported model config type: {cfg.model_config_type}. Liger not applied."

src/axolotl/integrations/liger/models/qwen3.py

@@ -1,160 +0,0 @@
-"""
-Liger FLCE for Qwen3. Based on transformers v4.51.3.
-"""
-
-import sys
-from typing import Optional, Tuple, Union
-
-import torch
-from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
-from transformers.cache_utils import Cache
-from transformers.modeling_outputs import CausalLMOutputWithPast
-
-
-def lce_forward(
-    self,
-    input_ids: Optional[torch.LongTensor] = None,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    past_key_values: Optional[Cache] = None,
-    inputs_embeds: Optional[torch.FloatTensor] = None,
-    labels: Optional[torch.LongTensor] = None,
-    use_cache: Optional[bool] = None,
-    output_attentions: Optional[bool] = None,
-    output_hidden_states: Optional[bool] = None,
-    cache_position: Optional[torch.LongTensor] = None,
-    logits_to_keep: Union[int, torch.Tensor] = 0,
-    **kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
-    r"""
-    Args:
-        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
-            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
-            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
-
-        logits_to_keep (`int` or `torch.Tensor`, *optional*):
-            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
-            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
-            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
-            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
-            This is useful when using packed tensor format (single dimension for batch and sequence length).
-
-    Returns:
-    """
-
-    # pylint: disable=duplicate-code
-    output_attentions = (
-        output_attentions
-        if output_attentions is not None
-        else self.config.output_attentions
-    )
-    output_hidden_states = (
-        output_hidden_states
-        if output_hidden_states is not None
-        else self.config.output_hidden_states
-    )
-
-    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-    outputs = self.model(
-        input_ids=input_ids,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        past_key_values=past_key_values,
-        inputs_embeds=inputs_embeds,
-        use_cache=use_cache,
-        output_attentions=output_attentions,
-        output_hidden_states=output_hidden_states,
-        cache_position=cache_position,
-        **kwargs,
-    )
-
-    hidden_states = outputs[0]
-
-    logits = None
-    loss = None
-    # if in training mode, don't materialize logits
-    if self.training and (labels is not None):
-        loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
-            lm_head_weight=self.lm_head.weight,
-            labels=labels,
-            hidden_size=self.config.hidden_size,
-            **kwargs,
-        )
-
-    else:  # if in inference mode materialize logits
-        slice_indices = (
-            slice(-logits_to_keep, None)
-            if isinstance(logits_to_keep, int)
-            else logits_to_keep
-        )
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
-        if labels is not None:
-            loss = self.loss_function(
-                logits=logits,
-                labels=labels,
-                vocab_size=self.config.vocab_size,
-                **kwargs,
-            )
-
-    return CausalLMOutputWithPast(
-        loss=loss,
-        logits=logits,
-        past_key_values=outputs.past_key_values,
-        hidden_states=outputs.hidden_states,
-        attentions=outputs.attentions,
-    )
-
-
-def apply_liger_kernel_to_qwen3(
-    cross_entropy: bool = False,
-    fused_linear_cross_entropy: bool = False,
-    rms_norm: bool = False,
-    glu_activation: bool = False,
-    layer_norm: bool = False,
-    **kwargs,  # pylint: disable=unused-argument
-) -> None:
-    # pylint: disable=duplicate-code
-    """
-    Apply Liger kernels to replace original implementation in HuggingFace Llama models (2 and 3)
-
-    Args:
-        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
-        fused_linear_cross_entropy (bool):
-            Whether to apply Liger's fused linear cross entropy loss. Default is False.
-            `cross_entropy` and `fused_linear_cross_entropy` cannot both be False.
-            If `fused_linear_cross_entropy` is True, the logits will not be materialized but more memory efficient.
-        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is False.
-        glu_activation (bool): Whether to apply Liger's SwiGLU MLP. Default is False.
-        layer_norm (bool): Whether to apply Liger's LayerNorm. Default is False.
-    """
-
-    import transformers.models.qwen3.modeling_qwen3  # noqa: F401  # pylint: disable=unused-import
-    from liger_kernel.transformers.functional import liger_cross_entropy
-    from liger_kernel.transformers.layer_norm import LigerLayerNorm
-    from liger_kernel.transformers.rms_norm import LigerRMSNorm
-    from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
-
-    assert not (
-        cross_entropy and fused_linear_cross_entropy
-    ), "cross_entropy and fused_linear_cross_entropy cannot both be True."
-
-    modeling_qwen3 = sys.modules["transformers.models.qwen3.modeling_qwen3"]
-
-    if rms_norm:
-        modeling_qwen3.Qwen3RMSNorm = LigerRMSNorm
-
-    if glu_activation:
-        modeling_qwen3.Qwen3MLP = LigerSwiGLUMLP
-
-    if layer_norm:
-        modeling_qwen3.nn.LayerNorm = LigerLayerNorm
-
-    if cross_entropy:
-        from transformers.loss.loss_utils import nn
-
-        nn.functional.cross_entropy = liger_cross_entropy
-
-    if fused_linear_cross_entropy:
-        modeling_qwen3.Qwen3ForCausalLM.forward = lce_forward

src/axolotl/integrations/liger/models/qwen3_moe.py

@@ -1,191 +0,0 @@
-"""
-Liger FLCE for Qwen3 MoE. Based on transformers v4.51.3.
-"""
-
-import sys
-from copy import deepcopy
-from typing import List, Optional, Union
-
-import torch
-from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
-from transformers.modeling_outputs import MoeCausalLMOutputWithPast
-from transformers.models.qwen3_moe.modeling_qwen3_moe import load_balancing_loss_func
-
-
-def lce_forward(
-    self,
-    input_ids: Optional[torch.LongTensor] = None,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    past_key_values: Optional[List[torch.FloatTensor]] = None,
-    inputs_embeds: Optional[torch.FloatTensor] = None,
-    labels: Optional[torch.LongTensor] = None,
-    use_cache: Optional[bool] = None,
-    output_attentions: Optional[bool] = None,
-    output_hidden_states: Optional[bool] = None,
-    output_router_logits: Optional[bool] = None,
-    cache_position: Optional[torch.LongTensor] = None,
-    logits_to_keep: Union[int, torch.Tensor] = 0,
-    **kwargs,
-) -> MoeCausalLMOutputWithPast:
-    r"""
-    Args:
-        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
-            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
-            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
-
-        logits_to_keep (`int` or `torch.Tensor`, *optional*):
-            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
-            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
-            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
-            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
-            This is useful when using packed tensor format (single dimension for batch and sequence length).
-
-    Returns:
-    """
-
-    # pylint: disable=duplicate-code
-    output_attentions = (
-        output_attentions
-        if output_attentions is not None
-        else self.config.output_attentions
-    )
-    output_router_logits = (
-        output_router_logits
-        if output_router_logits is not None
-        else self.config.output_router_logits
-    )
-    output_hidden_states = (
-        output_hidden_states
-        if output_hidden_states is not None
-        else self.config.output_hidden_states
-    )
-
-    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-    outputs = self.model(
-        input_ids=input_ids,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        past_key_values=past_key_values,
-        inputs_embeds=inputs_embeds,
-        use_cache=use_cache,
-        output_attentions=output_attentions,
-        output_hidden_states=output_hidden_states,
-        output_router_logits=output_router_logits,
-        cache_position=cache_position,
-        **kwargs,
-    )
-
-    hidden_states = outputs[0]
-
-    logits = None
-    loss = None
-    # if in training mode, don't materialize logits
-    if self.training and (labels is not None):
-        loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
-            lm_head_weight=self.lm_head.weight,
-            labels=labels,
-            hidden_size=self.config.hidden_size,
-            **kwargs,
-        )
-
-    else:  # if in inference mode materialize logits
-        slice_indices = (
-            slice(-logits_to_keep, None)
-            if isinstance(logits_to_keep, int)
-            else logits_to_keep
-        )
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
-        if labels is not None:
-            loss = self.loss_function(
-                logits=logits,
-                labels=labels,
-                vocab_size=self.config.vocab_size,
-                **kwargs,
-            )
-
-    aux_loss = None
-    if output_router_logits:
-        aux_loss = load_balancing_loss_func(
-            outputs.router_logits,
-            self.num_experts,
-            self.num_experts_per_tok,
-            attention_mask,
-        )
-        if labels is not None:
-            loss += self.router_aux_loss_coef * aux_loss.to(
-                loss.device
-            )  # make sure to reside in the same device
-
-    return MoeCausalLMOutputWithPast(
-        loss=loss,
-        aux_loss=aux_loss,
-        logits=logits,
-        past_key_values=outputs.past_key_values,
-        hidden_states=outputs.hidden_states,
-        attentions=outputs.attentions,
-    )
-
-
-def apply_liger_kernel_to_qwen3_moe(
-    cross_entropy: bool = False,
-    fused_linear_cross_entropy: bool = False,
-    rms_norm: bool = False,
-    glu_activation: bool = False,
-    layer_norm: bool = False,
-    **kwargs,  # pylint: disable=unused-argument
-) -> None:
-    # pylint: disable=duplicate-code
-    """
-    Apply Liger kernels to replace original implementation in HuggingFace Llama models (2 and 3)
-
-    Args:
-        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
-        fused_linear_cross_entropy (bool):
-            Whether to apply Liger's fused linear cross entropy loss. Default is False.
-            `cross_entropy` and `fused_linear_cross_entropy` cannot both be False.
-            If `fused_linear_cross_entropy` is True, the logits will not be materialized but more memory efficient.
-        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is False.
-        glu_activation (bool): Whether to apply Liger's SwiGLU MLP. Default is False.
-        layer_norm (bool): Whether to apply Liger's LayerNorm. Default is False.
-    """
-
-    import transformers.models.qwen3_moe.modeling_qwen3_moe  # noqa: F401  # pylint: disable=unused-import
-    from liger_kernel.transformers.functional import liger_cross_entropy
-    from liger_kernel.transformers.layer_norm import LigerLayerNorm
-    from liger_kernel.transformers.rms_norm import LigerRMSNorm
-    from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
-
-    assert not (
-        cross_entropy and fused_linear_cross_entropy
-    ), "cross_entropy and fused_linear_cross_entropy cannot both be True."
-
-    modeling_qwen3_moe = sys.modules["transformers.models.qwen3_moe.modeling_qwen3_moe"]
-
-    if rms_norm:
-        modeling_qwen3_moe.Qwen3MoeRMSNorm = LigerRMSNorm
-
-    if glu_activation:
-
-        def _liger_swiglu_mlp_wrapper(config, intermediate_size=None, **kwargs):
-            "Accepts intermediate_size to pass to LigerSwiGLUMLP"
-            # clone config to avoid modifying the original
-            config = deepcopy(config)
-            if intermediate_size:
-                setattr(config, "intermediate_size", intermediate_size)
-            return LigerSwiGLUMLP(config, **kwargs)
-
-        modeling_qwen3_moe.Qwen3MoeMLP = _liger_swiglu_mlp_wrapper
-
-    if layer_norm:
-        modeling_qwen3_moe.nn.LayerNorm = LigerLayerNorm
-
-    if cross_entropy:
-        from transformers.loss.loss_utils import nn
-
-        nn.functional.cross_entropy = liger_cross_entropy
-
-    if fused_linear_cross_entropy:
-        modeling_qwen3_moe.Qwen3MoeForCausalLM.forward = lce_forward

src/axolotl/monkeypatch/attention/ring_attn/patch.py

@@ -12,8 +12,10 @@ import torch
 import torch.distributed as dist
 from accelerate.logging import get_logger
 
+from axolotl.logging_config import configure_logging
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 
+configure_logging()
 LOG = get_logger(__name__)
 
 

src/axolotl/train.py

@@ -30,6 +30,7 @@ from axolotl.core.trainers.mixins.sequence_parallel import (
     SequenceParallelContextManager,
 )
 from axolotl.integrations.base import PluginManager
+from axolotl.logging_config import configure_logging
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import cleanup_distributed
 from axolotl.utils.freeze import freeze_layers_except
@@ -41,6 +42,7 @@ try:
 except ImportError:
     BetterTransformer = None
 
+configure_logging()
 LOG = get_logger(__name__)
 
 

src/axolotl/utils/config/__init__.py

@@ -67,7 +67,7 @@ def resolve_dtype(cfg):
         else:
             LOG.debug("bf16 support not detected, disabling for this configuration.")
             cfg.bf16 = False
-            if cfg.fp16 is None and not cfg.float16:
+            if cfg.fp16 is None:
                 cfg.fp16 = True
 
     if cfg.device == "mps":

src/axolotl/utils/schemas/config.py

@@ -512,17 +512,10 @@ class AxolotlInputConfig(
     @model_validator(mode="before")
     @classmethod
     def hint_sample_packing_padding(cls, data):
-        if data.get("sample_packing"):
+        if data.get("sample_packing") and not data.get("pad_to_sequence_len"):
-            pad_to_sequence_len = data.get("pad_to_sequence_len")
+            LOG.warning(
-            if pad_to_sequence_len is False:
+                "`pad_to_sequence_len: true` is recommended when using sample_packing"
-                LOG.warning(
+            )
-                    "`pad_to_sequence_len: true` is recommended when using sample_packing"
-                )
-            elif pad_to_sequence_len is None:
-                LOG.info(
-                    "Setting `pad_to_sequence_len: true` to prevent memory leaks when sample_packing"
-                )
-                data["pad_to_sequence_len"] = True
         return data
 
     @model_validator(mode="before")

src/axolotl/utils/schemas/trl.py

@@ -67,12 +67,6 @@ class TRLConfig(BaseModel):
         default=False,
         json_schema_extra={"description": "Whether to log completions"},
     )
-    num_completions_to_print: int | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "Number of completions to print. If `log_completions` is `True`, this will be the number of completions logged."
-        },
-    )
     sync_ref_model: bool | None = Field(
         default=False,
         json_schema_extra={

src/axolotl/utils/trainer.py

@@ -597,8 +597,6 @@ def prepare_optim_env(cfg):
         os.environ["ACCELERATE_MIXED_PRECISION"] = "bf16"
     elif cfg.fp16:
         os.environ["ACCELERATE_MIXED_PRECISION"] = "fp16"
-    else:
-        os.environ["ACCELERATE_MIXED_PRECISION"] = "no"
 
 
 def prepare_opinionated_env(cfg):

tests/patched/test_validation.py

@@ -648,7 +648,7 @@ class TestValidation(BaseValidation):
             DictDefault(
                 {
                     "sample_packing": True,
-                    "pad_to_sequence_len": False,
+                    "pad_to_sequence_len": None,
                     "flash_attention": True,
                 }
             )
@@ -662,26 +662,6 @@ class TestValidation(BaseValidation):
                 for record in self._caplog.records
             )
 
-    def test_packing_autoset(self, minimal_cfg):
-        cfg = (
-            DictDefault(
-                {
-                    "sample_packing": True,
-                    "pad_to_sequence_len": None,
-                    "flash_attention": True,
-                }
-            )
-            | minimal_cfg
-        )
-        with self._caplog.at_level(logging.INFO):
-            cfg = validate_config(cfg)
-            assert any(
-                "Setting `pad_to_sequence_len: true` to prevent memory leaks when sample_packing"
-                in record.message
-                for record in self._caplog.records
-            )
-            assert cfg.pad_to_sequence_len is True
-
     def test_merge_lora_no_bf16_fail(self, minimal_cfg):
         """
         This is assumed to be run on a CPU machine, so bf16 is not supported.