liger-kernel-nightly 0.6.3.dev20251121195543__tar.gz → 0.6.3.dev20251121200119__tar.gz

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.3.dev20251121195543
+Version: 0.6.3.dev20251121200119
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -310,6 +310,7 @@ loss.backward()
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | 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 |
 | InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | HunyuanV1   | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_dense`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/README.md

@@ -262,6 +262,7 @@ loss.backward()
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | 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 |
 | InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | HunyuanV1   | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_dense`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.3.dev20251121195543"
+version = "0.6.3.dev20251121200119"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/src/liger_kernel/transformers/__init__.py

@@ -52,6 +52,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mixtral  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mllama  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_olmo2  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_olmo3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_paligemma  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_phi3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2  # noqa: F401
@@ -118,6 +119,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_mixtral",
         "apply_liger_kernel_to_mllama",
         "apply_liger_kernel_to_olmo2",
+        "apply_liger_kernel_to_olmo3",
         "apply_liger_kernel_to_paligemma",
         "apply_liger_kernel_to_phi3",
         "apply_liger_kernel_to_qwen2",
@@ -194,6 +196,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_mixtral",
             "apply_liger_kernel_to_mllama",
             "apply_liger_kernel_to_olmo2",
+            "apply_liger_kernel_to_olmo3",
             "apply_liger_kernel_to_paligemma",
             "apply_liger_kernel_to_phi3",
             "apply_liger_kernel_to_qwen2",

liger_kernel_nightly-0.6.3.dev20251121200119/src/liger_kernel/transformers/model/olmo3.py

@@ -0,0 +1,142 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def lce_forward(
+    self,
+    input_ids: 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,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
+    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:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Olmo3ForCausalLM
+
+    >>> model = Olmo3ForCausalLM.from_pretrained("allenai/Olmo-3-7B-Instruct")
+    >>> tokenizer = AutoTokenizer.from_pretrained("allenai/Olmo-3-7B-Instruct")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    'Hey, are you conscious? Can you talk to me?\nI’m not sure if you’re conscious of this, but I’m'
+    ```
+    """
+    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
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    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,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    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
+
+    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:
+        # By default, if in training mode, don't materialize logits
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    # Compute loss
+    if skip_logits:
+        result = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.hidden_size,
+            **kwargs,
+        )
+        loss, _, token_accuracy = 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,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    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,
+    )

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/src/liger_kernel/transformers/monkey_patch.py

@@ -1928,6 +1928,74 @@ def apply_liger_kernel_to_olmo2(
                 _patch_rms_norm_module(decoder_layer.post_feedforward_layernorm, in_place=False)
 
 
+def apply_liger_kernel_to_olmo3(
+    rope: bool = True,
+    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 Olmo3 models.
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is True.
+        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 Olmo3MLP. Default is True.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
+        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.olmo3 import modeling_olmo3
+    from transformers.models.olmo3.modeling_olmo3 import Olmo3Model
+
+    from liger_kernel.transformers.model.olmo3 import lce_forward as olmo3_lce_forward
+    from liger_kernel.transformers.rms_norm import LigerRMSNormForOlmo2
+
+    # Olmo3 arch is very similar to Olmo2, so we can reuse all these components in the same way.
+    if rope:
+        modeling_olmo3.apply_rotary_pos_emb = liger_rotary_pos_emb
+    if rms_norm:
+        modeling_olmo3.Olmo3RMSNorm = LigerRMSNormForOlmo2  # same as olmo2
+    if swiglu:
+        modeling_olmo3.Olmo3MLP = LigerSwiGLUMLP
+    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(olmo3_lce_forward, model)
+        else:
+            modeling_olmo3.Olmo3ForCausalLM.forward = olmo3_lce_forward
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        # get the base model from the model instance
+        base_model: Olmo3Model = 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:
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm, in_place=False)
+                _patch_rms_norm_module(decoder_layer.post_feedforward_layernorm, in_place=False)
+
+
 def apply_liger_kernel_to_glm4(
     rope: bool = False,
     cross_entropy: bool = False,
@@ -2589,7 +2657,7 @@ def apply_liger_kernel_to_hunyuan_v1_dense(
         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(hunyuan_v1_lce_forward, model)
@@ -2695,6 +2763,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "mistral": apply_liger_kernel_to_mistral,
     "mixtral": apply_liger_kernel_to_mixtral,
     "olmo2": apply_liger_kernel_to_olmo2,
+    "olmo3": apply_liger_kernel_to_olmo3,
     "qwen2": apply_liger_kernel_to_qwen2,
     "qwen3": apply_liger_kernel_to_qwen3,
     "qwen3_moe": apply_liger_kernel_to_qwen3_moe,

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/src/liger_kernel/transformers/swiglu.py

@@ -93,4 +93,4 @@ class LigerHunyuanV1SwiGLUMLP(nn.Module):
             raise ValueError(f"Activation function {config.hidden_act} not supported.")
 
     def forward(self, x):
-        return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))
+        return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.3.dev20251121195543
+Version: 0.6.3.dev20251121200119
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -310,6 +310,7 @@ loss.backward()
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | 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 |
 | InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | HunyuanV1   | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_dense`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -210,6 +210,7 @@ src/liger_kernel/transformers/model/mistral.py
 src/liger_kernel/transformers/model/mixtral.py
 src/liger_kernel/transformers/model/mllama.py
 src/liger_kernel/transformers/model/olmo2.py
+src/liger_kernel/transformers/model/olmo3.py
 src/liger_kernel/transformers/model/output_classes.py
 src/liger_kernel/transformers/model/paligemma.py
 src/liger_kernel/transformers/model/phi3.py

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/test/convergence/bf16/test_mini_models.py

@@ -40,6 +40,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_mistral
 from liger_kernel.transformers import apply_liger_kernel_to_mixtral
 from liger_kernel.transformers import apply_liger_kernel_to_mllama
 from liger_kernel.transformers import apply_liger_kernel_to_olmo2
+from liger_kernel.transformers import apply_liger_kernel_to_olmo3
 from liger_kernel.transformers import apply_liger_kernel_to_phi3
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_5_vl
@@ -74,6 +75,7 @@ from test.utils import revert_liger_kernel_to_mistral
 from test.utils import revert_liger_kernel_to_mixtral
 from test.utils import revert_liger_kernel_to_mllama
 from test.utils import revert_liger_kernel_to_olmo2
+from test.utils import revert_liger_kernel_to_olmo3
 from test.utils import revert_liger_kernel_to_phi3
 from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_5_vl
@@ -194,6 +196,15 @@ try:
 except ImportError:
     OLMO2_AVAILABLE = False
 
+try:
+    # OLMO3 is only available in transformers>=4.57.0
+    from transformers.models.olmo3.configuration_olmo3 import Olmo3Config
+    from transformers.models.olmo3.modeling_olmo3 import Olmo3ForCausalLM
+
+    OLMO3_AVAILABLE = True
+except ImportError:
+    OLMO3_AVAILABLE = False
+
 try:
     # Glm4 is only available in transformers>=4.51.3
     from transformers.models.glm4.configuration_glm4 import Glm4Config
@@ -1009,6 +1020,35 @@ if OLMO2_AVAILABLE:
         ),
     )
 
+if OLMO3_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_olmo3"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_olmo3,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_olmo3,
+        model_class=Olmo3ForCausalLM,
+        mini_model_config=Olmo3Config(
+            bos_token_id=1,  # 128000
+            eos_token_id=2,  # 128001
+            pad_token_id=2,
+            cross_attention_layers=None,
+            dropout=0,
+            hidden_act="silu",
+            hidden_size=1024,  # 4096
+            initializer_range=0.02,
+            intermediate_size=2048,  # 14336
+            max_position_embeddings=4096,
+            num_attention_heads=8,  # 32
+            num_hidden_layers=4,  # 40
+            num_key_value_heads=2,  # 8
+            rms_norm_eps=1e-5,
+            rope_scaling=None,
+            rope_theta=500_000,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32000,  # 128256,
+            attn_implementation="sdpa",  # default value, pytorch native attention
+        ),
+    )
+
 if GLM4_AVAILABLE:
     MINI_MODEL_SETUPS["mini_glm4"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_glm4,
@@ -1351,7 +1391,7 @@ if HUNYUAN_V1_AVAILABLE:
         liger_kernel_patch_func=apply_liger_kernel_to_hunyuan_v1_moe,
         liger_kernel_patch_revert_func=revert_liger_kernel_to_hunyuan_v1_moe,
         model_class=HunYuanMoEV1ForCausalLM,
-        mini_model_config = HunYuanMoEV1Config(
+        mini_model_config=HunYuanMoEV1Config(
             vocab_size=32000,
             hidden_size=128,
             intermediate_size=512,
@@ -1751,6 +1791,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_olmo3",
+            32,
+            1e-5,
+            torch.bfloat16,
+            1e-2,
+            1e-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 OLMO3_AVAILABLE,
+                    reason="OLMO3 not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_glm4",
             32,

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/test/convergence/bf16/test_mini_models_with_logits.py

@@ -40,6 +40,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_mistral
 from liger_kernel.transformers import apply_liger_kernel_to_mixtral
 from liger_kernel.transformers import apply_liger_kernel_to_mllama
 from liger_kernel.transformers import apply_liger_kernel_to_olmo2
+from liger_kernel.transformers import apply_liger_kernel_to_olmo3
 from liger_kernel.transformers import apply_liger_kernel_to_phi3
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_5_vl
@@ -74,6 +75,7 @@ from test.utils import revert_liger_kernel_to_mistral
 from test.utils import revert_liger_kernel_to_mixtral
 from test.utils import revert_liger_kernel_to_mllama
 from test.utils import revert_liger_kernel_to_olmo2
+from test.utils import revert_liger_kernel_to_olmo3
 from test.utils import revert_liger_kernel_to_phi3
 from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_5_vl
@@ -186,6 +188,15 @@ try:
 except ImportError:
     OLMO2_AVAILABLE = False
 
+try:
+    # OLMO3 is only available in transformers>=4.57.0
+    from transformers.models.olmo3.configuration_olmo3 import Olmo3Config
+    from transformers.models.olmo3.modeling_olmo3 import Olmo3ForCausalLM
+
+    OLMO3_AVAILABLE = True
+except ImportError:
+    OLMO3_AVAILABLE = False
+
 try:
     # Glm4 is only available in transformers>=4.51.3
     from transformers.models.glm4.configuration_glm4 import Glm4Config
@@ -1005,6 +1016,35 @@ if OLMO2_AVAILABLE:
         ),
     )
 
+if OLMO3_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_olmo3"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_olmo3,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_olmo3,
+        model_class=Olmo3ForCausalLM,
+        mini_model_config=Olmo3Config(
+            bos_token_id=1,  # 128000
+            eos_token_id=2,  # 128001
+            pad_token_id=2,
+            cross_attention_layers=None,
+            dropout=0,
+            hidden_act="silu",
+            hidden_size=1024,  # 4096
+            initializer_range=0.02,
+            intermediate_size=2048,  # 14336
+            max_position_embeddings=4096,
+            num_attention_heads=8,  # 32
+            num_hidden_layers=4,  # 40
+            num_key_value_heads=2,  # 8
+            rms_norm_eps=1e-5,
+            rope_scaling=None,
+            rope_theta=500_000,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32000,  # 128256,
+            attn_implementation="sdpa",  # default value, pytorch native attention
+        ),
+    )
+
 if GLM4_AVAILABLE:
     MINI_MODEL_SETUPS["mini_glm4"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_glm4,
@@ -1346,7 +1386,7 @@ if HUNYUAN_V1_AVAILABLE:
         liger_kernel_patch_func=apply_liger_kernel_to_hunyuan_v1_moe,
         liger_kernel_patch_revert_func=revert_liger_kernel_to_hunyuan_v1_moe,
         model_class=HunYuanMoEV1ForCausalLM,
-        mini_model_config = HunYuanMoEV1Config(
+        mini_model_config=HunYuanMoEV1Config(
             vocab_size=32000,
             hidden_size=128,
             intermediate_size=512,
@@ -1768,6 +1808,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_olmo3",
+            32,
+            1e-4,
+            torch.bfloat16,
+            1e-2,
+            1e-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 OLMO3_AVAILABLE,
+                    reason="OLMO3 not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_glm4",
             32,

{liger_kernel_nightly-0.6.3.dev20251121195543 → liger_kernel_nightly-0.6.3.dev20251121200119}/test/convergence/fp32/test_mini_models.py

@@ -40,6 +40,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_mistral
 from liger_kernel.transformers import apply_liger_kernel_to_mixtral
 from liger_kernel.transformers import apply_liger_kernel_to_mllama
 from liger_kernel.transformers import apply_liger_kernel_to_olmo2
+from liger_kernel.transformers import apply_liger_kernel_to_olmo3
 from liger_kernel.transformers import apply_liger_kernel_to_phi3
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_5_vl
@@ -74,6 +75,7 @@ from test.utils import revert_liger_kernel_to_mistral
 from test.utils import revert_liger_kernel_to_mixtral
 from test.utils import revert_liger_kernel_to_mllama
 from test.utils import revert_liger_kernel_to_olmo2
+from test.utils import revert_liger_kernel_to_olmo3
 from test.utils import revert_liger_kernel_to_phi3
 from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_5_vl
@@ -182,6 +184,15 @@ try:
 except ImportError:
     OLMO2_AVAILABLE = False
 
+try:
+    # OLMO3 is only available in transformers>=4.57.0
+    from transformers.models.olmo3.configuration_olmo3 import Olmo3Config
+    from transformers.models.olmo3.modeling_olmo3 import Olmo3ForCausalLM
+
+    OLMO3_AVAILABLE = True
+except ImportError:
+    OLMO3_AVAILABLE = False
+
 try:
     # Glm4 is only available in transformers>=4.51.3
     from transformers.models.glm4.configuration_glm4 import Glm4Config
@@ -945,6 +956,35 @@ if OLMO2_AVAILABLE:
         ),
     )
 
+if OLMO3_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_olmo3"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_olmo3,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_olmo3,
+        model_class=Olmo3ForCausalLM,
+        mini_model_config=Olmo3Config(
+            bos_token_id=1,  # 128000
+            eos_token_id=2,  # 128001
+            pad_token_id=2,
+            cross_attention_layers=None,
+            dropout=0,
+            hidden_act="silu",
+            hidden_size=1024,  # 4096
+            initializer_range=0.02,
+            intermediate_size=2048,  # 14336
+            max_position_embeddings=4096,
+            num_attention_heads=8,  # 32
+            num_hidden_layers=4,  # 40
+            num_key_value_heads=2,  # 8
+            rms_norm_eps=1e-5,
+            rope_scaling=None,
+            rope_theta=500_000,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32000,  # 128256,
+            attn_implementation="sdpa",  # default value, pytorch native attention
+        ),
+    )
+
 if GLM4_AVAILABLE:
     MINI_MODEL_SETUPS["mini_glm4"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_glm4,
@@ -1343,7 +1383,7 @@ if HUNYUAN_V1_AVAILABLE:
         liger_kernel_patch_func=apply_liger_kernel_to_hunyuan_v1_moe,
         liger_kernel_patch_revert_func=revert_liger_kernel_to_hunyuan_v1_moe,
         model_class=HunYuanMoEV1ForCausalLM,
-        mini_model_config = HunYuanMoEV1Config(
+        mini_model_config=HunYuanMoEV1Config(
             hidden_act="silu",
             attention_dropout=0.0,
             num_hidden_layers=4,
@@ -1649,6 +1689,22 @@ def run_mini_model(
                 reason="OLMO2 not available in this version of transformers",
             ),
         ),
+        pytest.param(
+            "mini_olmo3",
+            32,
+            1e-4,
+            torch.float32,
+            1e-8,
+            1e-5,
+            5e-3,
+            1e-5,
+            5e-3,
+            1e-5,
+            marks=pytest.mark.skipif(
+                not OLMO3_AVAILABLE,
+                reason="OLMO3 not available in this version of transformers",
+            ),
+        ),
         pytest.param(
             "mini_glm4",
             32,