fix llama modeling

src/axolotl/cli/config.py

@@ -159,6 +159,9 @@ def plugin_set_cfg(cfg: DictDefault):
     if cfg.get("plugins"):
         plugin_manager = PluginManager.get_instance()
         plugin_manager.cfg = cfg
+        # now that we have the finalized cfg, register the plugins individually
+        for plugin in plugin_manager.plugins.values():
+            plugin.register(cfg)
 
 
 def load_cfg(

src/axolotl/integrations/modeling/__init__.py

@@ -0,0 +1,11 @@
+"""
+Axolotl custom modeling module
+"""
+
+from .args import AxolotlModelingArgs
+from .plugin import AxolotlModelingPlugin
+
+__all__ = [
+    "AxolotlModelingArgs",
+    "AxolotlModelingPlugin",
+]

src/axolotl/integrations/modeling/args.py

@@ -0,0 +1,13 @@
+"""
+Args for using Axolotl custom modeling
+"""
+
+from pydantic import BaseModel
+
+
+class AxolotlModelingArgs(BaseModel):
+    """
+    Args for using Axolotl custom modeling
+    """
+
+    use_liger_fused_rms_add: bool = False

src/axolotl/integrations/modeling/gemma3/__init__.py

@@ -0,0 +1,9 @@
+"""
+Gemma3 modeling
+"""
+
+from .modeling_gemma3 import patch_gemma3
+
+__all__ = [
+    "patch_gemma3",
+]

src/axolotl/integrations/modeling/gemma3/modeling_gemma3.py

@@ -0,0 +1,110 @@
+"""
+Gemma3 custom decoder layer using liger fused add rms norm kernels
+"""
+
+import sys
+
+import torch
+from liger_kernel.transformers.fused_add_rms_norm import LigerFusedAddRMSNorm
+from transformers import Cache, GradientCheckpointingLayer
+from transformers.models.gemma3.configuration_gemma3 import Gemma3TextConfig
+from transformers.models.gemma3.modeling_gemma3 import (
+    Gemma3Attention,
+    Gemma3MLP,
+    Gemma3RMSNorm,
+)
+
+
+class Gemma3AddRMSNorm(LigerFusedAddRMSNorm):
+    """
+    Fused add rms norm
+    """
+
+    def __init__(self, hidden_size: int, eps: float = 1e-6):
+        super().__init__(hidden_size, eps, offset=1.0, casting_mode="gemma")
+
+
+class Gemma3DecoderLayer(GradientCheckpointingLayer):
+    """
+    Gemma3 decoder layer using liger fused add rms norm
+    """
+
+    def __init__(self, config: Gemma3TextConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.layer_idx = layer_idx
+        self.attention_type = config.layer_types[layer_idx]
+        self.self_attn = Gemma3Attention(config=config, layer_idx=layer_idx)
+        self.mlp = Gemma3MLP(config)
+        self.input_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Gemma3RMSNorm(
+            self.hidden_size, eps=config.rms_norm_eps
+        )
+        self.pre_feedforward_layernorm = Gemma3AddRMSNorm(
+            self.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_feedforward_layernorm = Gemma3RMSNorm(
+            self.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings_global: torch.Tensor,
+        position_embeddings_local: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_value: Cache | None = None,
+        output_attentions: bool | None = False,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs,
+    ) -> tuple[
+        torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor | None] | None
+    ]:
+        # pylint: disable=duplicate-code
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # apply global RoPE to non-sliding layer only
+        if self.self_attn.is_sliding:
+            position_embeddings = position_embeddings_local
+        else:
+            position_embeddings = position_embeddings_global
+
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            position_embeddings=position_embeddings,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states, residual = self.pre_feedforward_layernorm(
+            hidden_states, residual
+        )
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = self.post_feedforward_layernorm(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)  # type: ignore
+
+        return outputs  # type: ignore
+
+
+def patch_gemma3():
+    import transformers.models.gemma3.modeling_gemma3
+
+    transformers.models.gemma3.modeling_gemma3.Gemma3DecoderLayer = Gemma3DecoderLayer
+    sys.modules["transformers.models.gemma3.modeling_gemma3"].Gemma3DecoderLayer = (
+        Gemma3DecoderLayer
+    )

src/axolotl/integrations/modeling/llama/__init__.py

@@ -0,0 +1,9 @@
+"""
+Llama modeling
+"""
+
+from modeling_llama import patch_llama
+
+__all__ = [
+    "patch_llama",
+]

src/axolotl/integrations/modeling/llama/modeling_llama.py

@@ -0,0 +1,86 @@
+"""
+Custom modeling for Llama for fused rms add kernels
+"""
+
+import sys
+
+import torch
+from liger_kernel.transformers.fused_add_rms_norm import LigerFusedAddRMSNorm
+from transformers import Cache, GradientCheckpointingLayer
+from transformers.models.llama.configuration_llama import LlamaConfig
+from transformers.models.llama.modeling_llama import (
+    LlamaAttention,
+    LlamaMLP,
+    LlamaRMSNorm,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs
+
+
+class LlamaAddRMSNorm(LigerFusedAddRMSNorm):
+    """
+    Fused add rms norm
+    """
+
+    def __init__(self, hidden_size: int, eps: float = 1e-6):
+        super().__init__(hidden_size, eps, casting_mode="llama")
+
+
+class LlamaDecoderLayer(GradientCheckpointingLayer):
+    """
+    Llama decoder layer using liger fused add rms norm
+    """
+
+    def __init__(self, config: LlamaConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = LlamaAttention(config=config, layer_idx=layer_idx)
+
+        self.mlp = LlamaMLP(config)
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaAddRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_value: Cache | None = None,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        position_embeddings: (
+            tuple[torch.Tensor, torch.Tensor] | None
+        ) = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
+        # pylint: disable=duplicate-code
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+def patch_llama():
+    import transformers.models.llama.modeling_llama
+
+    transformers.models.llama.modeling_llama.LlamaDecoderLayer = LlamaDecoderLayer
+    sys.modules["transformers.models.llama.modeling_llama"].LlamaDecoderLayer = (
+        LlamaDecoderLayer
+    )

src/axolotl/integrations/modeling/plugin.py

@@ -0,0 +1,22 @@
+"""
+Axolotl custom modeling plugin
+"""
+
+from axolotl.integrations.base import BasePlugin
+
+
+class AxolotlModelingPlugin(BasePlugin):
+    """
+    Axolotl custom modeling plugin
+    """
+
+    def get_input_args(self) -> str | None:
+        return "axolotl.integrations.modeling.AxolotlModelingArgs"
+
+    def register(self, cfg):  # pylint: disable=unused-argument
+        if cfg.use_liger_fused_rms_add:
+            from .gemma3 import patch_gemma3
+            from .llama import patch_llama
+
+            patch_gemma3()
+            patch_llama()