Add DeepSeek V3 support

README.md

@@ -387,6 +387,7 @@ loss.backward()
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | Olmo3   | `liger_kernel.transformers.apply_liger_kernel_to_olmo3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | GLM-4   | `liger_kernel.transformers.apply_liger_kernel_to_glm4`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
+| DeepSeek-V3   | `liger_kernel.transformers.apply_liger_kernel_to_deepseek_v3`     | RMSNorm, SwiGLU (dense/shared MLPs), CrossEntropyLoss, FusedLinearCrossEntropy |
 | DeepSeek-V4   | `liger_kernel.transformers.apply_liger_kernel_to_deepseek_v4`     | RMSNorm, CrossEntropyLoss, FusedLinearCrossEntropy |
 | GPT-OSS   | `liger_kernel.transformers.apply_liger_kernel_to_gpt_oss`     | RoPE, RMSNorm, CrossEntropyLoss, FusedLinearCrossEntropy |
 | InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |

src/liger_kernel/transformers/__init__.py

@@ -39,6 +39,7 @@
     from liger_kernel.transformers.auto_model import AutoLigerKernelForCausalLM  # noqa: F401
     from liger_kernel.transformers.monkey_patch import _apply_liger_kernel  # noqa: F401
     from liger_kernel.transformers.monkey_patch import _apply_liger_kernel_to_instance  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_deepseek_v3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_deepseek_v4  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_exaone4  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_falcon_h1  # noqa: F401
@@ -158,6 +159,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_smolvlm",
         "apply_liger_kernel_to_hunyuan_v1_dense",
         "apply_liger_kernel_to_hunyuan_v1_moe",
+        "apply_liger_kernel_to_deepseek_v3",
         "apply_liger_kernel_to_deepseek_v4",
         "apply_liger_kernel_to_exaone4",
     }
@@ -249,6 +251,7 @@ def __getattr__(name: str):
             "apply_liger_kernel_to_smolvlm",
             "apply_liger_kernel_to_hunyuan_v1_dense",
             "apply_liger_kernel_to_hunyuan_v1_moe",
+            "apply_liger_kernel_to_deepseek_v3",
             "apply_liger_kernel_to_deepseek_v4",
             "apply_liger_kernel_to_exaone4",
         ]

src/liger_kernel/transformers/model/deepseek_v3.py

@@ -0,0 +1,81 @@
+import torch
+
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.utils import can_return_tuple
+
+from liger_kernel.transformers.model.llama import lce_maybe_trainable_lm_head
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
+
+
+@can_return_tuple
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    attention_mask: torch.Tensor | None = None,
+    position_ids: torch.LongTensor | None = None,
+    past_key_values: Cache | None = None,
+    inputs_embeds: torch.FloatTensor | None = None,
+    labels: torch.LongTensor | None = None,
+    use_cache: bool | None = None,
+    logits_to_keep: int | torch.Tensor = 0,
+    skip_logits: bool | None = None,
+    **kwargs,
+) -> LigerCausalLMOutputWithPast:
+    outputs: BaseModelOutputWithPast = 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,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+    token_accuracy = None
+    predicted_tokens = None
+
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are None")
+
+    if skip_logits is None:
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        result = lce_maybe_trainable_lm_head(
+            self,
+            hidden_states=kept_hidden_states,
+            hidden_size=self.config.hidden_size,
+            labels=labels,
+            shift_labels=shift_labels,
+            **kwargs,
+        )
+        loss, _, token_accuracy, predicted_tokens = unpack_cross_entropy_result(result)
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    return LigerCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
+        predicted_tokens=predicted_tokens,
+    )

src/liger_kernel/transformers/monkey_patch.py

@@ -3377,6 +3377,90 @@ def apply_liger_kernel_to_hunyuan_v1_moe(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_deepseek_v3(
+    rope: bool = False,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace DeepSeek-V3 models.
+
+    NOTE: RoPE is not supported for DeepSeek-V3. Its attention uses interleaved partial RoPE,
+    which is incompatible with ``liger_rotary_pos_emb``. Routed experts are intentionally left
+    unchanged; SwiGLU is only applied to dense and shared-expert MLPs.
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is False.
+            Currently unsupported; emits a warning and is a no-op.
+        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 True.
+            `cross_entropy` and `fused_linear_cross_entropy` cannot both be True.
+            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 True.
+        swiglu (bool): Whether to apply Liger's SwiGLU to dense and shared-expert MLPs. Default is True.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if already loaded.
+            Default is None.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.deepseek_v3 import modeling_deepseek_v3
+    from transformers.models.deepseek_v3.modeling_deepseek_v3 import DeepseekV3Model
+
+    from liger_kernel.transformers.model.deepseek_v3 import lce_forward as deepseek_v3_lce_forward
+    from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP
+
+    if rope:
+        logger.warning_once(
+            "rope=True is not supported for DeepSeek-V3: interleaved partial RoPE is "
+            "incompatible with liger_rotary_pos_emb. Skipping rope kernel swap."
+        )
+
+    if rms_norm:
+        modeling_deepseek_v3.DeepseekV3RMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        if model is not None:
+            model.forward = MethodType(deepseek_v3_lce_forward, model)
+        else:
+            modeling_deepseek_v3.DeepseekV3ForCausalLM.forward = deepseek_v3_lce_forward
+
+    if swiglu:
+        modeling_deepseek_v3.DeepseekV3MLP = LigerQwen3MoeSwiGLUMLP
+
+    if model is not None:
+        base_model: DeepseekV3Model = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                shared_experts = getattr(decoder_layer.mlp, "shared_experts", None)
+                if shared_experts is not None:
+                    _patch_swiglu_module(shared_experts, LigerQwen3MoeSwiGLUMLP)
+                elif not hasattr(decoder_layer.mlp, "experts"):
+                    _patch_swiglu_module(decoder_layer.mlp, LigerQwen3MoeSwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+                q_a_layernorm = getattr(decoder_layer.self_attn, "q_a_layernorm", None)
+                if q_a_layernorm is not None:
+                    _patch_rms_norm_module(q_a_layernorm)
+                kv_a_layernorm = getattr(decoder_layer.self_attn, "kv_a_layernorm", None)
+                if kv_a_layernorm is not None:
+                    _patch_rms_norm_module(kv_a_layernorm)
+
+
 def apply_liger_kernel_to_deepseek_v4(
     rope: bool = False,
     cross_entropy: bool = False,
@@ -3532,6 +3616,7 @@ def __init__(self, hidden_size, eps=1e-6, **kwargs):
 
 # Model type corresponds to the keys defined in transformers/models/auto/modeling_auto.py
 MODEL_TYPE_TO_APPLY_LIGER_FN = {
+    "deepseek_v3": apply_liger_kernel_to_deepseek_v3,
     "deepseek_v4": apply_liger_kernel_to_deepseek_v4,
     "gemma": apply_liger_kernel_to_gemma,
     "gemma2": apply_liger_kernel_to_gemma2,

test/convergence/bf16/test_mini_models.py

@@ -24,6 +24,7 @@
 from transformers.models.qwen2 import Qwen2Config
 from transformers.models.qwen2 import Qwen2ForCausalLM
 
+from liger_kernel.transformers import apply_liger_kernel_to_deepseek_v3
 from liger_kernel.transformers import apply_liger_kernel_to_deepseek_v4
 from liger_kernel.transformers import apply_liger_kernel_to_exaone4
 from liger_kernel.transformers import apply_liger_kernel_to_falcon_h1
@@ -68,6 +69,7 @@
 from test.utils import get_logprobs
 from test.utils import get_topk
 from test.utils import require_deterministic
+from test.utils import revert_liger_kernel_to_deepseek_v3
 from test.utils import revert_liger_kernel_to_deepseek_v4
 from test.utils import revert_liger_kernel_to_exaone4
 from test.utils import revert_liger_kernel_to_falcon_h1
@@ -340,6 +342,14 @@
 except ImportError:
     HUNYUAN_V1_AVAILABLE = False
 
+try:
+    from transformers.models.deepseek_v3.configuration_deepseek_v3 import DeepseekV3Config
+    from transformers.models.deepseek_v3.modeling_deepseek_v3 import DeepseekV3ForCausalLM
+
+    DEEPSEEK_V3_AVAILABLE = True
+except ImportError:
+    DEEPSEEK_V3_AVAILABLE = False
+
 try:
     from transformers.models.deepseek_v4.configuration_deepseek_v4 import DeepseekV4Config
     from transformers.models.deepseek_v4.modeling_deepseek_v4 import DeepseekV4ForCausalLM
@@ -1648,6 +1658,35 @@
         ),
     )
 
+if DEEPSEEK_V3_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_deepseek_v3"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_deepseek_v3,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_deepseek_v3,
+        model_class=DeepseekV3ForCausalLM,
+        mini_model_config=DeepseekV3Config(
+            vocab_size=32000,
+            hidden_size=32,
+            intermediate_size=64,
+            moe_intermediate_size=16,
+            num_hidden_layers=4,
+            num_attention_heads=2,
+            num_key_value_heads=2,
+            q_lora_rank=8,
+            kv_lora_rank=8,
+            qk_rope_head_dim=8,
+            qk_nope_head_dim=8,
+            v_head_dim=16,
+            num_experts_per_tok=2,
+            n_routed_experts=4,
+            n_shared_experts=1,
+            n_group=2,
+            topk_group=1,
+            first_k_dense_replace=1,
+            max_position_embeddings=128,
+            attn_implementation="sdpa",
+        ),
+    )
+
 if DEEPSEEK_V4_AVAILABLE:
     MINI_MODEL_SETUPS["mini_deepseek_v4"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_deepseek_v4,
@@ -1767,7 +1806,13 @@ def run_mini_model(
             "rms_norm": True,
         }
 
-        if "glm4" in model_name or "qwen3_next" in model_name or "qwen3_5" in model_name or "deepseek_v4" in model_name:
+        if (
+            "glm4" in model_name
+            or "qwen3_next" in model_name
+            or "qwen3_5" in model_name
+            or "deepseek_v3" in model_name
+            or "deepseek_v4" in model_name
+        ):
             kwargs["rope"] = False
 
         model_supports_layer_norm = "qwen2_vl" in model_name
@@ -2467,6 +2512,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_deepseek_v3",
+            32,
+            1e-5,
+            torch.bfloat16,
+            1e-2,
+            5e-2,
+            1e-1,
+            1e-2,
+            1e-2,
+            1e-2,
+            marks=[
+                pytest.mark.skipif(not supports_bfloat16(), reason="bfloat16 not supported on this GPU"),
+                pytest.mark.skipif(
+                    not DEEPSEEK_V3_AVAILABLE,
+                    reason="DeepSeek-V3 not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_deepseek_v4",
             32,