fix(doc): update min torch version

* qwen3 and qwen3_moe support for liger kernels

* fix moe module path

* fix: qwen3 liger input args and mlp

* fix: qwen3 input args and output class

---------

Co-authored-by: NanoCode012 <nano@axolotl.ai>

.github/workflows/preview-docs.yml

@@ -4,6 +4,12 @@ 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

README.md

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

setup.py

@@ -67,13 +67,11 @@ 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:
@@ -147,7 +145,7 @@ extras_require = {
         "ray[train]",
     ],
     "vllm": [
-        "vllm==0.7.2",
+        "vllm==0.8.5",
     ],
     "llmcompressor": [
         "llmcompressor==0.5.1",

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

@@ -177,12 +177,8 @@ 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/integrations/liger/__init__.py

@@ -151,6 +151,30 @@ 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

@@ -0,0 +1,160 @@
+"""
+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

@@ -0,0 +1,191 @@
+"""
+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/kernels/quantize.py

@@ -55,16 +55,13 @@ def dequantize(
     target_device = W.device
 
     # Extract quantization state
-    nested = False
     if not isinstance(quant_state, list):
         # New style quant_state class
         absmax = quant_state.absmax.to(target_device)
         shape = quant_state.shape
         dtype = quant_state.dtype
         blocksize = quant_state.blocksize
-        if quant_state.nested:
+        offset = quant_state.offset.to(target_device)
-            nested = True
-            offset = quant_state.offset.to(target_device)
         state2 = quant_state.state2
         absmax2 = state2.absmax.to(target_device)
         code2 = state2.code.to(target_device)
@@ -118,8 +115,7 @@ def dequantize(
             ctypes.c_int(n_elements_absmax),
         )
 
-    if nested:
+    out_absmax += offset
-        out_absmax += offset
 
     # Choose appropriate dequantization function
     fx = (

src/axolotl/utils/schemas/config.py

@@ -512,10 +512,17 @@ class AxolotlInputConfig(
     @model_validator(mode="before")
     @classmethod
     def hint_sample_packing_padding(cls, data):
-        if data.get("sample_packing") and not data.get("pad_to_sequence_len"):
+        if data.get("sample_packing"):
-            LOG.warning(
+            pad_to_sequence_len = data.get("pad_to_sequence_len")
-                "`pad_to_sequence_len: true` is recommended when using sample_packing"
+            if pad_to_sequence_len is False:
-            )
+                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")

tests/patched/test_validation.py

@@ -648,7 +648,7 @@ class TestValidation(BaseValidation):
             DictDefault(
                 {
                     "sample_packing": True,
-                    "pad_to_sequence_len": None,
+                    "pad_to_sequence_len": False,
                     "flash_attention": True,
                 }
             )
@@ -662,6 +662,26 @@ 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.