liger-kernel-nightly 0.5.3.dev20250221233257__tar.gz → 0.5.3.dev20250224212643__tar.gz

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.3.dev20250221233257
+Version: 0.5.3.dev20250224212643
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -314,6 +314,7 @@ loss.backward()
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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 |
 
 
 ## Low-level APIs

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/README.md

@@ -266,6 +266,7 @@ loss.backward()
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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 |
 
 
 ## Low-level APIs

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.3.dev20250221233257"
+version = "0.5.3.dev20250224212643"
 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-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/src/liger_kernel/ops/layer_norm.py

@@ -57,13 +57,14 @@ def _layer_norm_forward_kernel(
     B_row = tl.load(B_ptr + col_offsets, mask=mask, other=0)
 
     mean = tl.sum(X_row, axis=0) / n_cols
-    var = tl.sum((X_row - mean) * (X_row - mean), axis=0) / n_cols
+    Xmm = tl.where(mask, X_row - mean, 0)
+    var = tl.sum(Xmm * Xmm, axis=0) / n_cols
     rstd = rsqrt(var + eps)
 
     tl.store(Mean_ptr, mean)
     tl.store(RSTD_ptr, rstd)
 
-    Y_row = (X_row - mean) * rstd * W_row + B_row
+    Y_row = Xmm * rstd * W_row + B_row
 
     tl.store(Y_ptr + col_offsets, Y_row, mask=mask)
 
@@ -147,9 +148,11 @@ def layer_norm_forward(X, W, B, eps):
     Y = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
     Mean = torch.empty(n_rows, dtype=X.dtype, device=X.device)
     RSTD = torch.empty(n_rows, dtype=X.dtype, device=X.device)
-    assert X.shape[1] == W.shape[0], (
-        f"Incompatible hidden size dimension between input tensor with shape[1] = {X.shape[1]} and weight tensor with shape[0] = {W.shape[0]}"
-    )
+    if X.shape[1] != W.shape[0]:
+        raise ValueError(
+            f"Incompatible dimensions: input feature size (X.shape[1]={X.shape[1]}) "
+            f"must match weight size (W.shape[0]={W.shape[0]})"
+        )
 
     _layer_norm_forward_kernel[(n_rows,)](
         Y,
@@ -190,11 +193,21 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
 
     BLOCK_SIZE, num_warps = calculate_settings(n_cols)
     if n_cols > BLOCK_SIZE:
-        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+        raise RuntimeError(
+            f"Feature dimension {n_cols} exceeds maximum supported size of {BLOCK_SIZE}. Consider using a smaller feature dimension."
+        )
 
     rows_per_program = math.ceil(n_rows / sm_count)
     grid = (sm_count,)
-    triton_dtype = tl.float32 if X.dtype == torch.float32 else tl.bfloat16
+    triton_dtype = (
+        tl.float32
+        if X.dtype == torch.float32
+        else tl.bfloat16
+        if X.dtype == torch.bfloat16
+        else tl.float16
+        if X.dtype == torch.float16
+        else tl.float32  # fallback to float32 for other types
+    )
     _layer_norm_backward_kernel[grid](
         X,
         W,

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/src/liger_kernel/transformers/__init__.py

@@ -14,6 +14,7 @@ from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llama
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mistral  # noqa: F401
 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_phi3  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_vl  # noqa: F401

liger_kernel_nightly-0.5.3.dev20250224212643/src/liger_kernel/transformers/model/olmo2.py

@@ -0,0 +1,124 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.olmo2.modeling_olmo2 import _CONFIG_FOR_DOC
+from transformers.models.olmo2.modeling_olmo2 import OLMO2_INPUTS_DOCSTRING
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
+
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+
+
+@add_start_docstrings_to_model_forward(OLMO2_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+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,
+    num_logits_to_keep: int = 0,
+    **loss_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]`.
+
+        num_logits_to_keep (`int`, *optional*):
+            Calculate logits for the last `num_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.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Olmo2ForCausalLM
+
+    >>> model = Olmo2ForCausalLM.from_pretrained("allenai/Olmo2-1B-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("allenai/Olmo2-1B-hf")
+
+    >>> 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 = 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,
+    )
+
+    hidden_states = outputs[0]
+
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if self.training and (labels is not None):
+        # We do the same thing as ForCausalLMLoss but using Liger FLCE
+
+        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+
+        # flatten tokens
+        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
+        shift_labels = shift_labels.view(-1)
+
+        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
+        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
+
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        if reduction == "sum":
+            loss /= loss_kwargs["num_items_in_batch"]
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_kwargs,
+            )
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/src/liger_kernel/transformers/monkey_patch.py

@@ -814,6 +814,69 @@ def apply_liger_kernel_to_phi3(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_olmo2(
+    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 OLMO2 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 Olmo2MLP. 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.olmo2 import modeling_olmo2
+    from transformers.models.olmo2.modeling_olmo2 import Olmo2Model
+
+    from liger_kernel.transformers.model.olmo2 import lce_forward as olmo2_lce_forward
+
+    if rope:
+        modeling_olmo2.apply_rotary_pos_emb = liger_rotary_pos_emb
+    if rms_norm:
+        modeling_olmo2.Olmo2RMSNorm = partial(LigerRMSNorm, in_place=False)
+    if swiglu:
+        modeling_olmo2.Olmo2MLP = LigerSwiGLUMLP
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+    if fused_linear_cross_entropy:
+        modeling_olmo2.Olmo2ForCausalLM.forward = olmo2_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: Olmo2Model = 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:
+                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+            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)
+
+
 # Model type corresponds to the keys defined in transformers/models/auto/modeling_auto.py
 MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma": apply_liger_kernel_to_gemma,
@@ -824,6 +887,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "mllama_text_model": apply_liger_kernel_to_mllama,
     "mistral": apply_liger_kernel_to_mistral,
     "mixtral": apply_liger_kernel_to_mixtral,
+    "olmo2": apply_liger_kernel_to_olmo2,
     "qwen2": apply_liger_kernel_to_qwen2,
     "qwen2_vl": apply_liger_kernel_to_qwen2_vl,
     "phi3": apply_liger_kernel_to_phi3,

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.3.dev20250221233257
+Version: 0.5.3.dev20250224212643
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -314,6 +314,7 @@ loss.backward()
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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 |
 
 
 ## Low-level APIs

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -162,6 +162,7 @@ src/liger_kernel/transformers/model/llama.py
 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/phi3.py
 src/liger_kernel/transformers/model/qwen2.py
 src/liger_kernel/transformers/model/qwen2_vl.py

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/test/convergence/bf16/test_mini_models.py

@@ -25,6 +25,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_llama
 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_phi3
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl
@@ -38,6 +39,7 @@ from test.utils import revert_liger_kernel_to_llama
 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_phi3
 from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_vl
@@ -71,6 +73,15 @@ try:
 except ImportError:
     GRANITE_AVAILABLE = False
 
+try:
+    # OLMO2 is only available in transformers>=4.47.0
+    from transformers.models.olmo2.configuration_olmo2 import Olmo2Config
+    from transformers.models.olmo2.modeling_olmo2 import Olmo2ForCausalLM
+
+    OLMO2_AVAILABLE = True
+except ImportError:
+    OLMO2_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -426,6 +437,35 @@ if GRANITE_AVAILABLE:
         ),
     )
 
+if OLMO2_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_olmo2"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_olmo2,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_olmo2,
+        model_class=Olmo2ForCausalLM,
+        mini_model_config=Olmo2Config(
+            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
+        ),
+    )
+
 
 def create_model(model_name="mini_llama3"):
     """
@@ -612,6 +652,25 @@ def run_mini_model(
             1e-2,
             marks=pytest.mark.skipif(not supports_bfloat16(), reason="bfloat16 not supported on this GPU"),
         ),
+        pytest.param(
+            "mini_olmo2",
+            32,
+            1e-4,
+            torch.bfloat16,
+            1e-3,
+            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 OLMO2_AVAILABLE,
+                    reason="OLMO2 not available in this version of transformers",
+                ),
+            ],
+        ),
         # TODO: mixtral is flaky so disable the test for now
         # pytest.param(
         #     "mini_mixtral",

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/test/convergence/bf16/test_mini_models_with_logits.py

@@ -25,6 +25,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_llama
 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_phi3
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl
@@ -38,6 +39,7 @@ from test.utils import revert_liger_kernel_to_llama
 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_phi3
 from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_vl
@@ -71,6 +73,15 @@ try:
 except ImportError:
     GRANITE_AVAILABLE = False
 
+try:
+    # OLMO2 is only available in transformers>=4.47.0
+    from transformers.models.olmo2.configuration_olmo2 import Olmo2Config
+    from transformers.models.olmo2.modeling_olmo2 import Olmo2ForCausalLM
+
+    OLMO2_AVAILABLE = True
+except ImportError:
+    OLMO2_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -390,6 +401,34 @@ if QWEN2_VL_AVAILABLE:
             attn_implementation="sdpa",
         ),
     )
+if OLMO2_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_olmo2"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_olmo2,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_olmo2,
+        model_class=Olmo2ForCausalLM,
+        mini_model_config=Olmo2Config(
+            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 GRANITE_AVAILABLE:
     MINI_MODEL_SETUPS["mini_granite3"] = MiniModelConfig(
@@ -655,6 +694,25 @@ def run_mini_model(
             1e-2,
             marks=pytest.mark.skipif(not supports_bfloat16(), reason="bfloat16 not supported on this GPU"),
         ),
+        pytest.param(
+            "mini_olmo2",
+            32,
+            1e-4,
+            torch.bfloat16,
+            1e-3,
+            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 OLMO2_AVAILABLE,
+                    reason="OLMO2 not available in this version of transformers",
+                ),
+            ],
+        ),
         # TODO: Gemma2 test for bf16 is not passing within the tolerance range, might be casting issue, need to investigate
         # pytest.param(
         #     "mini_gemma2",
@@ -699,6 +757,8 @@ def test_mini_model(
         rtol=loss_rtol,
     )
 
+    # No logits are materialized
+    # import pdb; pdb.set_trace()
     # Compare the logits from the last step
     assert_verbose_allclose(
         expected_output["logits"],

{liger_kernel_nightly-0.5.3.dev20250221233257 → liger_kernel_nightly-0.5.3.dev20250224212643}/test/convergence/fp32/test_mini_models.py

@@ -25,6 +25,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_llama
 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_phi3
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl
@@ -38,6 +39,7 @@ from test.utils import revert_liger_kernel_to_llama
 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_phi3
 from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_vl
@@ -70,6 +72,15 @@ try:
 except ImportError:
     GRANITE_AVAILABLE = False
 
+try:
+    # OLMO2 is only available in transformers>=4.47.0
+    from transformers.models.olmo2.configuration_olmo2 import Olmo2Config
+    from transformers.models.olmo2.modeling_olmo2 import Olmo2ForCausalLM
+
+    OLMO2_AVAILABLE = True
+except ImportError:
+    OLMO2_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -425,6 +436,35 @@ if GRANITE_AVAILABLE:
         ),
     )
 
+if OLMO2_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_olmo2"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_olmo2,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_olmo2,
+        model_class=Olmo2ForCausalLM,
+        mini_model_config=Olmo2Config(
+            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
+        ),
+    )
+
 
 def create_model(model_name="mini_llama3"):
     """
@@ -536,6 +576,22 @@ def run_mini_model(
                 reason="Qwen2-VL not available in this version of transformers",
             ),
         ),
+        pytest.param(
+            "mini_olmo2",
+            32,
+            1e-4,
+            torch.float32,
+            1e-8,
+            1e-5,
+            5e-3,
+            1e-5,
+            5e-3,
+            1e-5,
+            marks=pytest.mark.skipif(
+                not OLMO2_AVAILABLE,
+                reason="OLMO2 not available in this version of transformers",
+            ),
+        ),
         ("mini_phi3", 32, 1e-4, torch.float32, 1e-8, 1e-5, 5e-3, 1e-5, 5e-3, 1e-5),
         ("mini_mistral", 32, 1e-4, torch.float32, 1e-8, 1e-5, 5e-3, 1e-5, 5e-3, 1e-5),
         # TODO: mixtral is flaky so disable the test for now