liger-kernel-nightly 0.6.2.dev20251010184612__tar.gz → 0.6.2.dev20251011154226__tar.gz

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251010184612
+Version: 0.6.2.dev20251011154226
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.2.dev20251010184612"
+version = "0.6.2.dev20251011154226"
 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-0.6.2.dev20251011154226/src/liger_kernel/transformers/model/falcon_h1.py

@@ -0,0 +1,108 @@
+from typing import TYPE_CHECKING
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+if TYPE_CHECKING:
+    from transformers.models.falcon_h1.modeling_falcon_h1 import FalconHybridMambaAttentionDynamicCache
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+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["FalconHybridMambaAttentionDynamicCache"] = 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,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[tuple, CausalLMOutputWithPast]:
+    r"""
+    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]`.
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, FalconH1ForCausalLM
+
+    >>> model = FalconH1ForCausalLM.from_pretrained("...")
+    >>> tokenizer = AutoTokenizer.from_pretrained("...")
+
+    >>> 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 conscious, but I can talk to you."
+    ```"""
+    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]
+    # 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
+    # if in training mode, don't materialize logits
+    if skip_logits and 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
+
+    if skip_logits:
+        loss = 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,
+        )
+    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, 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,
+    )

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/src/liger_kernel/transformers/monkey_patch.py

@@ -15,6 +15,7 @@ from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
 from liger_kernel.transformers.functional import liger_cross_entropy
 from liger_kernel.transformers.geglu import LigerGEGLUMLP
 from liger_kernel.transformers.layer_norm import LigerLayerNorm
+from liger_kernel.transformers.model.falcon_h1 import lce_forward as falcon_h1_lce_forward
 from liger_kernel.transformers.model.gemma import lce_forward as gemma_lce_forward
 from liger_kernel.transformers.model.gemma import lce_forward_deprecated as gemma_lce_forward_deprecated
 from liger_kernel.transformers.model.gemma2 import lce_forward as gemma2_lce_forward
@@ -2109,8 +2110,8 @@ def apply_liger_kernel_to_internvl(
 
 def apply_liger_kernel_to_falcon_h1(
     rope: bool = True,
-    cross_entropy: bool = True,
-    fused_linear_cross_entropy: bool = False,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
     rms_norm: bool = True,
     swiglu: bool = False,
     model: PreTrainedModel = None,
@@ -2144,7 +2145,7 @@ def apply_liger_kernel_to_falcon_h1(
         logger.info("Apply liger RMSNorm")
         modeling_falcon_h1.FalconH1RMSNorm = LigerRMSNorm
     if swiglu:
-        raise NotImplementedError("LigerSwiGLUMLP is not available for Falcon-H1 models.")
+        logger.warning("LigerSwiGLUMLP is not available for Falcon-H1 models. There will be no effect.")
 
     if cross_entropy:
         logger.info("Apply liger cross entropy")
@@ -2153,7 +2154,10 @@ def apply_liger_kernel_to_falcon_h1(
         nn.functional.cross_entropy = liger_cross_entropy
 
     if fused_linear_cross_entropy:
-        raise NotImplementedError("LigerFusedLinearCrossEntropy is not available for Falcon-H1 models.")
+        if model is not None:
+            model.forward = MethodType(falcon_h1_lce_forward, model)
+        else:
+            modeling_falcon_h1.FalconH1ForCausalLM.forward = falcon_h1_lce_forward
 
     if model is not None:
         # The model instance already exists, so we need to additionally patch the

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251010184612
+Version: 0.6.2.dev20251011154226
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -187,6 +187,7 @@ src/liger_kernel/transformers/tvd.py
 src/liger_kernel/transformers/experimental/__init__.py
 src/liger_kernel/transformers/experimental/embedding.py
 src/liger_kernel/transformers/model/__init__.py
+src/liger_kernel/transformers/model/falcon_h1.py
 src/liger_kernel/transformers/model/gemma.py
 src/liger_kernel/transformers/model/gemma2.py
 src/liger_kernel/transformers/model/gemma3.py

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/test/convergence/bf16/test_mini_models.py

@@ -18,6 +18,7 @@ from transformers.models.phi3 import Phi3ForCausalLM
 from transformers.models.qwen2 import Qwen2Config
 from transformers.models.qwen2 import Qwen2ForCausalLM
 
+from liger_kernel.transformers import apply_liger_kernel_to_falcon_h1
 from liger_kernel.transformers import apply_liger_kernel_to_gemma
 from liger_kernel.transformers import apply_liger_kernel_to_gemma2
 from liger_kernel.transformers import apply_liger_kernel_to_gemma3_text
@@ -45,6 +46,7 @@ from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
 from test.utils import get_logprobs
 from test.utils import get_topk
+from test.utils import revert_liger_kernel_to_falcon_h1
 from test.utils import revert_liger_kernel_to_gemma
 from test.utils import revert_liger_kernel_to_gemma2
 from test.utils import revert_liger_kernel_to_gemma3_text
@@ -201,6 +203,15 @@ try:
 except ImportError:
     INTERNVL_AVAILABLE = False
 
+try:
+    # FalconH1 is only available in transformers>=4.53.0
+    from transformers.models.falcon_h1.configuration_falcon_h1 import FalconH1Config
+    from transformers.models.falcon_h1.modeling_falcon_h1 import FalconH1ForCausalLM
+
+    FALCONH1_AVAILABLE = True
+except ImportError:
+    FALCONH1_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -1065,6 +1076,36 @@ if INTERNVL_AVAILABLE:
         ),
     )
 
+if FALCONH1_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_falcon_h1"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_falcon_h1,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_falcon_h1,
+        model_class=FalconH1ForCausalLM,
+        mini_model_config=FalconH1Config(
+            model_type="falcon_h1",
+            vocab_size=32000,
+            hidden_size=256,  # 4096
+            num_hidden_layers=4,  # 24
+            num_attention_heads=4,  # 32
+            num_key_value_heads=2,  # 8
+            intermediate_size=1024,  # 11008
+            hidden_act="silu",
+            max_position_embeddings=4096,
+            initializer_range=0.02,
+            rms_norm_eps=1e-6,
+            use_cache=True,
+            pad_token_id=0,
+            bos_token_id=1,
+            eos_token_id=2,
+            tie_word_embeddings=False,
+            mamba_d_ssm=128,  # 1024
+            mamba_n_heads=16,  # 128
+            mamba_d_state=32,  # 245
+            mamba_d_conv=2,  # 4
+            attn_implementation="eager",
+        ),
+    )
+
 
 def create_model(model_name="mini_llama4"):
     """
@@ -1574,6 +1615,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_falcon_h1",
+            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 FALCONH1_AVAILABLE,
+                    reason="FalconH1 not available in this version of transformers",
+                ),
+            ],
+        ),
     ],
 )
 def test_mini_model(

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/test/convergence/bf16/test_mini_models_multimodal.py

@@ -1066,6 +1066,10 @@ def run_mini_model_multimodal(
                     not MLLAMA_AVAILABLE,
                     reason="Mllama not available in this version of transformers",
                 ),
+                pytest.mark.skipif(
+                    version.parse("4.51.0") > version.parse(transformers.__version__),
+                    reason="MllamaForConditionalGeneration doesn't accecpt `skip_logits` kwargs",
+                ),
             ],
         ),
         pytest.param(

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/test/convergence/bf16/test_mini_models_with_logits.py

@@ -18,6 +18,7 @@ from transformers.models.phi3 import Phi3ForCausalLM
 from transformers.models.qwen2 import Qwen2Config
 from transformers.models.qwen2 import Qwen2ForCausalLM
 
+from liger_kernel.transformers import apply_liger_kernel_to_falcon_h1
 from liger_kernel.transformers import apply_liger_kernel_to_gemma
 from liger_kernel.transformers import apply_liger_kernel_to_gemma2
 from liger_kernel.transformers import apply_liger_kernel_to_gemma3_text
@@ -45,6 +46,7 @@ from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
 from test.utils import get_logprobs
 from test.utils import get_topk
+from test.utils import revert_liger_kernel_to_falcon_h1
 from test.utils import revert_liger_kernel_to_gemma
 from test.utils import revert_liger_kernel_to_gemma2
 from test.utils import revert_liger_kernel_to_gemma3_text
@@ -201,6 +203,15 @@ try:
 except ImportError:
     INTERNVL_AVAILABLE = False
 
+try:
+    # FalconH1 is only available in transformers>=4.53.0
+    from transformers.models.falcon_h1.configuration_falcon_h1 import FalconH1Config
+    from transformers.models.falcon_h1.modeling_falcon_h1 import FalconH1ForCausalLM
+
+    FALCONH1_AVAILABLE = True
+except ImportError:
+    FALCONH1_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -1063,6 +1074,35 @@ if INTERNVL_AVAILABLE:
         ),
     )
 
+if FALCONH1_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_falcon_h1"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_falcon_h1,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_falcon_h1,
+        model_class=FalconH1ForCausalLM,
+        mini_model_config=FalconH1Config(
+            model_type="falcon_h1",
+            vocab_size=32000,
+            hidden_size=256,  # 4096
+            num_hidden_layers=4,  # 24
+            num_attention_heads=4,  # 32
+            num_key_value_heads=2,  # 8
+            intermediate_size=1024,  # 11008
+            hidden_act="silu",
+            max_position_embeddings=4096,
+            initializer_range=0.02,
+            rms_norm_eps=1e-6,
+            use_cache=True,
+            pad_token_id=0,
+            bos_token_id=1,
+            eos_token_id=2,
+            tie_word_embeddings=False,
+            mamba_d_ssm=128,  # 1024
+            mamba_n_heads=16,  # 128
+            mamba_d_state=32,  # 245
+            mamba_d_conv=2,  # 4
+        ),
+    )
+
 
 def create_model(model_name="mini_llama3"):
     """
@@ -1547,6 +1587,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_falcon_h1",
+            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 FALCONH1_AVAILABLE,
+                    reason="FalconH1 not available in this version of transformers",
+                ),
+            ],
+        ),
     ],
 )
 def test_mini_model(

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/test/convergence/fp32/test_mini_models.py

@@ -18,6 +18,7 @@ from transformers.models.phi3 import Phi3ForCausalLM
 from transformers.models.qwen2 import Qwen2Config
 from transformers.models.qwen2 import Qwen2ForCausalLM
 
+from liger_kernel.transformers import apply_liger_kernel_to_falcon_h1
 from liger_kernel.transformers import apply_liger_kernel_to_gemma
 from liger_kernel.transformers import apply_liger_kernel_to_gemma2
 from liger_kernel.transformers import apply_liger_kernel_to_gemma3_text
@@ -45,6 +46,7 @@ from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
 from test.utils import get_logprobs
 from test.utils import get_topk
+from test.utils import revert_liger_kernel_to_falcon_h1
 from test.utils import revert_liger_kernel_to_gemma
 from test.utils import revert_liger_kernel_to_gemma2
 from test.utils import revert_liger_kernel_to_gemma3_text
@@ -200,6 +202,15 @@ try:
 except ImportError:
     INTERNVL_AVAILABLE = False
 
+try:
+    # FalconH1 is only available in transformers>=4.53.0
+    from transformers.models.falcon_h1.configuration_falcon_h1 import FalconH1Config
+    from transformers.models.falcon_h1.modeling_falcon_h1 import FalconH1ForCausalLM
+
+    FALCONH1_AVAILABLE = True
+except ImportError:
+    FALCONH1_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -1061,6 +1072,35 @@ if INTERNVL_AVAILABLE:
         ),
     )
 
+if FALCONH1_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_falcon_h1"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_falcon_h1,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_falcon_h1,
+        model_class=FalconH1ForCausalLM,
+        mini_model_config=FalconH1Config(
+            model_type="falcon_h1",
+            vocab_size=32000,
+            hidden_size=256,  # 4096
+            num_hidden_layers=4,  # 24
+            num_attention_heads=4,  # 32
+            num_key_value_heads=2,  # 8
+            intermediate_size=1024,  # 11008
+            hidden_act="silu",
+            max_position_embeddings=4096,
+            initializer_range=0.02,
+            rms_norm_eps=1e-6,
+            use_cache=True,
+            pad_token_id=0,
+            bos_token_id=1,
+            eos_token_id=2,
+            tie_word_embeddings=False,
+            mamba_d_ssm=128,  # 1024
+            mamba_n_heads=16,  # 128
+            mamba_d_state=32,  # 245
+            mamba_d_conv=2,  # 4
+        ),
+    )
+
 
 def create_model(model_name="mini_llama3"):
     """
@@ -1432,6 +1472,22 @@ def run_mini_model(
                 reason="InternVL not available in this version of transformers",
             ),
         ),
+        pytest.param(
+            "mini_falcon_h1",
+            32,
+            1e-5,
+            torch.float32,
+            1e-8,
+            1e-4,
+            4e-2,
+            1e-5,
+            5e-3,
+            1e-5,
+            marks=pytest.mark.skipif(
+                not FALCONH1_AVAILABLE,
+                reason="FalconH1 not available in this version of transformers",
+            ),
+        ),
     ],
 )
 def test_mini_model(

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/test/convergence/fp32/test_mini_models_multimodal.py

@@ -1061,10 +1061,16 @@ def run_mini_model_multimodal(
             1e-5,
             5e-3,
             1e-5,
-            marks=pytest.mark.skipif(
-                not MLLAMA_AVAILABLE,
-                reason="Mllama not available in this version of transformers",
-            ),
+            marks=[
+                pytest.mark.skipif(
+                    not MLLAMA_AVAILABLE,
+                    reason="Mllama not available in this version of transformers",
+                ),
+                pytest.mark.skipif(
+                    version.parse("4.51.0") > version.parse(transformers.__version__),
+                    reason="MllamaForConditionalGeneration doesn't accecpt `skip_logits` kwargs",
+                ),
+            ],
         ),
         pytest.param(
             "mini_paligemma",

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/test/convergence/fp32/test_mini_models_with_logits.py

@@ -18,6 +18,7 @@ from transformers.models.phi3 import Phi3ForCausalLM
 from transformers.models.qwen2 import Qwen2Config
 from transformers.models.qwen2 import Qwen2ForCausalLM
 
+from liger_kernel.transformers import apply_liger_kernel_to_falcon_h1
 from liger_kernel.transformers import apply_liger_kernel_to_gemma
 from liger_kernel.transformers import apply_liger_kernel_to_gemma2
 from liger_kernel.transformers import apply_liger_kernel_to_gemma3_text
@@ -45,6 +46,7 @@ from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
 from test.utils import get_logprobs
 from test.utils import get_topk
+from test.utils import revert_liger_kernel_to_falcon_h1
 from test.utils import revert_liger_kernel_to_gemma
 from test.utils import revert_liger_kernel_to_gemma2
 from test.utils import revert_liger_kernel_to_gemma3_text
@@ -200,6 +202,15 @@ try:
 except ImportError:
     INTERNVL_AVAILABLE = False
 
+try:
+    # FalconH1 is only available in transformers>=4.53.0
+    from transformers.models.falcon_h1.configuration_falcon_h1 import FalconH1Config
+    from transformers.models.falcon_h1.modeling_falcon_h1 import FalconH1ForCausalLM
+
+    FALCONH1_AVAILABLE = True
+except ImportError:
+    FALCONH1_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -1061,6 +1072,35 @@ if INTERNVL_AVAILABLE:
         ),
     )
 
+if FALCONH1_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_falcon_h1"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_falcon_h1,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_falcon_h1,
+        model_class=FalconH1ForCausalLM,
+        mini_model_config=FalconH1Config(
+            model_type="falcon_h1",
+            vocab_size=32000,
+            hidden_size=256,  # 4096
+            num_hidden_layers=4,  # 24
+            num_attention_heads=4,  # 32
+            num_key_value_heads=2,  # 8
+            intermediate_size=1024,  # 11008
+            hidden_act="silu",
+            max_position_embeddings=4096,
+            initializer_range=0.02,
+            rms_norm_eps=1e-6,
+            use_cache=True,
+            pad_token_id=0,
+            bos_token_id=1,
+            eos_token_id=2,
+            tie_word_embeddings=False,
+            mamba_d_ssm=128,  # 1024
+            mamba_n_heads=16,  # 128
+            mamba_d_state=32,  # 245
+            mamba_d_conv=2,  # 4
+        ),
+    )
+
 
 def create_model(model_name="mini_llama3"):
     """
@@ -1398,6 +1438,22 @@ def run_mini_model(
                 reason="InternVL not available in this version of transformers",
             ),
         ),
+        pytest.param(
+            "mini_falcon_h1",
+            32,
+            1e-5,
+            torch.float32,
+            1e-8,
+            1e-4,
+            4e-2,
+            1e-5,
+            5e-3,
+            1e-5,
+            marks=pytest.mark.skipif(
+                not FALCONH1_AVAILABLE,
+                reason="FalconH1 not available in this version of transformers",
+            ),
+        ),
     ],
 )
 def test_mini_model(

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/test/transformers/test_monkey_patch.py

@@ -32,6 +32,7 @@ SUPPORTED_TRANSFORMER_VERSION = "4.46.1"
 
 # Import forward functions based on transformer version
 if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
+    from liger_kernel.transformers.model.falcon_h1 import lce_forward as falcon_h1_lce_forward
     from liger_kernel.transformers.model.gemma import lce_forward as gemma_lce_forward
     from liger_kernel.transformers.model.gemma2 import lce_forward as gemma2_lce_forward
     from liger_kernel.transformers.model.llama import lce_forward as llama_lce_forward
@@ -156,6 +157,15 @@ def is_paligemma_available():
         return False
 
 
+def is_falcon_h1_available():
+    try:
+        import transformers.models.falcon_h1  # noqa: F401
+
+        return True
+    except ImportError:
+        return False
+
+
 def test_import_from_root():
     try:
         from liger_kernel.transformers import AutoLigerKernelForCausalLM  # noqa: F401
@@ -383,6 +393,47 @@ def test_apply_liger_kernel_to_instance_for_llama():
             pytest.fail(f"An exception occured in extra_expr: {type(e).__name__} - {e}")
 
 
+@pytest.mark.skipif(not is_falcon_h1_available(), reason="falcon_h1 module not available")
+def test_apply_liger_kernel_to_falcon_h1_for_causal_lm():
+    with patch("transformers.models.falcon_h1.modeling_falcon_h1"):
+        from transformers.models.falcon_h1.modeling_falcon_h1 import FalconH1ForCausalLM
+
+        # Instantiate a dummy model
+        config = transformers.models.falcon_h1.configuration_falcon_h1.FalconH1Config(
+            hidden_size=256,
+            num_hidden_layers=2,
+            num_attention_heads=4,
+            num_key_value_heads=2,
+            intermediate_size=1024,
+        )
+        dummy_model_instance = FalconH1ForCausalLM(config)
+
+        # Check that model instance variables are not yet patched with Liger modules
+        assert inspect.getsource(dummy_model_instance.forward) != inspect.getsource(falcon_h1_lce_forward)
+        assert inspect.getsource(dummy_model_instance.model.final_layernorm.forward) != inspect.getsource(
+            LigerRMSNorm.forward
+        )
+        for layer in dummy_model_instance.model.layers:
+            assert inspect.getsource(layer.input_layernorm.forward) != inspect.getsource(LigerRMSNorm.forward)
+            assert inspect.getsource(layer.pre_ff_layernorm.forward) != inspect.getsource(LigerRMSNorm.forward)
+
+        # Test applying kernels to the model instance
+        _apply_liger_kernel_to_instance(model=dummy_model_instance)
+
+        # Check that the model's instance variables were correctly patched with Liger modules
+        assert inspect.getsource(dummy_model_instance.forward) == inspect.getsource(falcon_h1_lce_forward)
+        assert inspect.getsource(dummy_model_instance.model.final_layernorm.forward) == inspect.getsource(
+            LigerRMSNorm.forward
+        )
+        for layer in dummy_model_instance.model.layers:
+            assert inspect.getsource(layer.input_layernorm.forward) == inspect.getsource(LigerRMSNorm.forward)
+            assert inspect.getsource(layer.pre_ff_layernorm.forward) == inspect.getsource(LigerRMSNorm.forward)
+        try:
+            print(dummy_model_instance)
+        except Exception as e:
+            pytest.fail(f"An exception occured in extra_expr: {type(e).__name__} - {e}")
+
+
 @pytest.mark.skipif(not is_mllama_available(), reason="mllama module not available")
 def test_apply_liger_kernel_to_instance_for_mllama_for_conditional_generation():
     # Ensure any monkey patching is cleaned up for subsequent tests

{liger_kernel_nightly-0.6.2.dev20251010184612 → liger_kernel_nightly-0.6.2.dev20251011154226}/test/utils.py

@@ -582,6 +582,18 @@ def revert_liger_kernel_to_internvl(model_config: MiniModelConfig):
     print("Liger kernel patches have been reverted.")
 
 
+def revert_liger_kernel_to_falcon_h1(model_config: MiniModelConfig):
+    """
+    Revert all Liger kernel patches applied to FalconH1.
+    """
+
+    from transformers.models.falcon_h1 import modeling_falcon_h1
+
+    importlib.reload(modeling_falcon_h1)
+    model_config.model_class = modeling_falcon_h1.FalconH1ForCausalLM
+    print("Liger kernel patches have been reverted.")
+
+
 class HFAlignmentLoss:
     def __init__(
         self,