liger-kernel-nightly 0.6.2.dev20251011152316__tar.gz → 0.6.2.dev20251011154427__tar.gz

{liger_kernel_nightly-0.6.2.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251011152316
+Version: 0.6.2.dev20251011154427
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/benchmark/scripts/benchmark_cross_entropy.py

@@ -70,6 +70,9 @@ def bench_speed_cross_entropy(
 
     if mode == "forward":
         ms_50, ms_20, ms_80 = triton.testing.do_bench(fwd, rep=100, quantiles=QUANTILES)
+    elif mode == "no-grad-forward":
+        with torch.no_grad():
+            ms_50, ms_20, ms_80 = triton.testing.do_bench(fwd, rep=100, quantiles=QUANTILES)
     elif mode == "backward":
         y = fwd()
 
@@ -109,7 +112,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_cross_entropy,
-        kernel_operation_modes=["forward", "backward", "full"],
+        kernel_operation_modes=["forward", "backward", "full", "no-grad-forward"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.6.2.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/benchmark/scripts/benchmark_fused_linear_cross_entropy.py

@@ -59,26 +59,26 @@ def bench_memory_fused_linear_cross_entropy(
     dtype = input.extra_benchmark_config["dtype"]
     provider = input.kernel_provider
 
-    torch_lm_head_ce = TorchLMHeadCE(H=H, V=V, dtype=dtype).to(device)
-    liger_lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype).to(device)
-    liger_lm_head_ce_fp32_accum = LigerLMHeadCE(H=H, V=V, dtype=dtype, accum_dtype=torch.float32).to(device)
+    lm_head_ce = None
+    if provider == "liger":
+        lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype).to(device)
+    elif provider == "liger-fp32-accum":
+        lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype, accum_dtype=torch.float32).to(device)
+    else:
+        lm_head_ce = TorchLMHeadCE(H=H, V=V, dtype=dtype).to(device)
 
     _input = torch.randn(BT, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (BT, 1), dtype=torch.long, device=device).squeeze(1)
 
     def fwd():
-        if provider == "liger":
-            return liger_lm_head_ce(_input, target)
-        elif provider == "liger-fp32-accum":
-            return liger_lm_head_ce_fp32_accum(_input, target)
-        elif provider == "huggingface":
-            return torch_lm_head_ce(_input, target)
+        return lm_head_ce(_input, target)
 
     def full():
         y = fwd()
         y.backward()
 
     mem_50, mem_20, mem_80 = _test_memory(full, _iter=10, quantiles=QUANTILES)
+
     return SingleBenchmarkRunOutput(
         y_20=mem_20,
         y_50=mem_50,
@@ -101,20 +101,19 @@ def bench_speed_fused_linear_cross_entropy(
     provider = input.kernel_provider
     mode = input.kernel_operation_mode
 
-    torch_lm_head_ce = TorchLMHeadCE(H=H, V=V, dtype=dtype).to(device)
-    liger_lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype).to(device)
-    liger_lm_head_ce_fp32_accum = LigerLMHeadCE(H=H, V=V, dtype=dtype, accum_dtype=torch.float32).to(device)
+    lm_head_ce = None
+    if provider == "liger":
+        lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype).to(device)
+    elif provider == "liger-fp32-accum":
+        lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype, accum_dtype=torch.float32).to(device)
+    else:
+        lm_head_ce = TorchLMHeadCE(H=H, V=V, dtype=dtype).to(device)
 
     _input = torch.randn(BT, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (BT, 1), dtype=torch.long, device=device).squeeze(1)
 
     def fwd():
-        if provider == "liger":
-            return liger_lm_head_ce(_input, target)
-        elif provider == "liger-fp32-accum":
-            return liger_lm_head_ce_fp32_accum(_input, target)
-        elif provider == "huggingface":
-            return torch_lm_head_ce(_input, target)
+        return lm_head_ce(_input, target)
 
     if mode == "forward":
         ms_50, ms_20, ms_80 = triton.testing.do_bench(
@@ -122,6 +121,13 @@ def bench_speed_fused_linear_cross_entropy(
             rep=100,
             quantiles=QUANTILES,
         )
+    elif mode == "no-grad-forward":
+        with torch.no_grad():
+            ms_50, ms_20, ms_80 = triton.testing.do_bench(
+                fwd,
+                rep=100,
+                quantiles=QUANTILES,
+            )
     elif mode == "backward":
         y = fwd()
 
@@ -164,7 +170,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_fused_linear_cross_entropy,
-        kernel_operation_modes=["forward", "backward", "full"],
+        kernel_operation_modes=["forward", "backward", "full", "no-grad-forward"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel_nightly-0.6.2.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.2.dev20251011152316"
+version = "0.6.2.dev20251011154427"
 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.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/src/liger_kernel/ops/cross_entropy.py

@@ -45,6 +45,7 @@ def liger_cross_entropy_kernel(
     BLOCK_SIZE: tl.constexpr,
     HAS_WEIGHT: tl.constexpr,
     HAS_SOFTCAPPING: tl.constexpr,
+    HAS_GRADIENTS: tl.constexpr,
 ):
     """
     This kernel computes both cross entropy loss and the gradient of the input.
@@ -72,6 +73,7 @@ def liger_cross_entropy_kernel(
     BLOCK_SIZE (int): The block size for Triton operations.
     HAS_WEIGHT (bool): The boolean value to determine whether assigning weight to each of the classes.
     HAS_SOFTCAPPING (bool): The boolean value to determine whether applying soft-capping or not.
+    HAS_GRADIENTS (bool): The boolean value to determine whether calculating gradients in forward pass.
     """
 
     # https://github.com/triton-lang/triton/issues/1058
@@ -155,58 +157,58 @@ def liger_cross_entropy_kernel(
     # For 'sum' reduction, no normalization is applied:
     # dx_y = softmax(x_y) - 1
     # dx_i = softmax(x_i), for i ≠ y
-
-    for i in range(0, n_cols, BLOCK_SIZE):
-        X_offsets = i + tl.arange(0, BLOCK_SIZE)
-        X_block = tl.load(
-            X_ptr + X_offsets,
-            mask=X_offsets < n_cols,
-            other=float("-inf"),
-            # Ensure float32 precision for softmax calculation
-        ).cast(tl.float32)
-        if HAS_SOFTCAPPING:
-            intermediate = tanh(X_block / softcap)
-            X_block = softcap * intermediate
-
-        if not HAS_WEIGHT:
-            # softmax(x_i)
-            X_block = tl.exp(X_block - m) / d
-            # derivative of z-loss: 2 * lse_square_scale * lse * softmax(x_i)
-            X_block += 2 * lse_square_scale * lse * X_block
-            # smoothing term
-            X_block += -eps
-            # special handle dx_y
-            X_block = tl.where(X_offsets != y, X_block, X_block - (1 - label_smoothing))
-            # reduction scale
-            if reduction == "mean":
-                X_block = X_block / n_non_ignore
-        else:
-            weight_block = tl.load(weight_ptr + X_offsets, mask=X_offsets < n_cols)
-            softmax_X = tl.exp(X_block - m) / d
-            # derivative of original_loss
-            dloss_ori = (1 - label_smoothing) * softmax_X
-            # specially handle dx_y
-            dloss_ori = tl.where(X_offsets != y, dloss_ori, dloss_ori - (1 - label_smoothing))
-            dloss_ori = dloss_ori * weight_y
-            # derivative of smooth_loss
-            dloss_smooth = eps * (-weight_block + softmax_X * weight_sum)
-            # derivative of z-loss
-            dz_loss = 2 * lse_square_scale * lse * softmax_X
-            # reduction scale
-            if reduction == "mean":
-                dloss_ori = dloss_ori / sum_non_ignore_weight
-                dloss_smooth = dloss_smooth / sum_non_ignore_weight
-                # TODO: Implement weighted z_loss. Currently, z_loss is not scaled by weight.
-                dz_loss = dz_loss / n_non_ignore
-            # derivative of total_loss
-            X_block = dloss_ori + dloss_smooth + dz_loss
-
-        # chain rule softcapping
-        # d(softcap * tanh(x / softcap)) = (1 - tanh^2(x / softcap))
-        if HAS_SOFTCAPPING:
-            X_block = X_block * (1 - intermediate * intermediate)
-
-        tl.store(X_ptr + X_offsets, X_block, mask=X_offsets < n_cols)
+    if HAS_GRADIENTS:
+        for i in range(0, n_cols, BLOCK_SIZE):
+            X_offsets = i + tl.arange(0, BLOCK_SIZE)
+            X_block = tl.load(
+                X_ptr + X_offsets,
+                mask=X_offsets < n_cols,
+                other=float("-inf"),
+                # Ensure float32 precision for softmax calculation
+            ).cast(tl.float32)
+            if HAS_SOFTCAPPING:
+                intermediate = tanh(X_block / softcap)
+                X_block = softcap * intermediate
+
+            if not HAS_WEIGHT:
+                # softmax(x_i)
+                X_block = tl.exp(X_block - m) / d
+                # derivative of z-loss: 2 * lse_square_scale * lse * softmax(x_i)
+                X_block += 2 * lse_square_scale * lse * X_block
+                # smoothing term
+                X_block += -eps
+                # special handle dx_y
+                X_block = tl.where(X_offsets != y, X_block, X_block - (1 - label_smoothing))
+                # reduction scale
+                if reduction == "mean":
+                    X_block = X_block / n_non_ignore
+            else:
+                weight_block = tl.load(weight_ptr + X_offsets, mask=X_offsets < n_cols)
+                softmax_X = tl.exp(X_block - m) / d
+                # derivative of original_loss
+                dloss_ori = (1 - label_smoothing) * softmax_X
+                # specially handle dx_y
+                dloss_ori = tl.where(X_offsets != y, dloss_ori, dloss_ori - (1 - label_smoothing))
+                dloss_ori = dloss_ori * weight_y
+                # derivative of smooth_loss
+                dloss_smooth = eps * (-weight_block + softmax_X * weight_sum)
+                # derivative of z-loss
+                dz_loss = 2 * lse_square_scale * lse * softmax_X
+                # reduction scale
+                if reduction == "mean":
+                    dloss_ori = dloss_ori / sum_non_ignore_weight
+                    dloss_smooth = dloss_smooth / sum_non_ignore_weight
+                    # TODO: Implement weighted z_loss. Currently, z_loss is not scaled by weight.
+                    dz_loss = dz_loss / n_non_ignore
+                # derivative of total_loss
+                X_block = dloss_ori + dloss_smooth + dz_loss
+
+            # chain rule softcapping
+            # d(softcap * tanh(x / softcap)) = (1 - tanh^2(x / softcap))
+            if HAS_SOFTCAPPING:
+                X_block = X_block * (1 - intermediate * intermediate)
+
+            tl.store(X_ptr + X_offsets, X_block, mask=X_offsets < n_cols)
 
     # We need tl.debug_barrier() to ensure the new result of X_ptr is written as mentioned in
     # https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/ops/cross_entropy.py#L34
@@ -332,6 +334,7 @@ def cross_entropy_forward(
         BLOCK_SIZE=BLOCK_SIZE,
         HAS_WEIGHT=True if weight is not None else False,
         HAS_SOFTCAPPING=True if softcap is not None else False,
+        HAS_GRADIENTS=_input.requires_grad,
         # TODO: 32 seems to give the best performance
         # Performance is quite sensitive to num_warps
         num_warps=32 if not is_hip() else 16,

{liger_kernel_nightly-0.6.2.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/src/liger_kernel/ops/fused_linear_cross_entropy.py

@@ -150,6 +150,7 @@ def fused_linear_cross_entropy_forward(
             RETURN_Z_LOSS=return_z_loss,
             HAS_WEIGHT=True if ce_weight is not None else False,
             HAS_SOFTCAPPING=True if softcap is not None else False,
+            HAS_GRADIENTS=_input.requires_grad,
             BLOCK_SIZE=BLOCK_SIZE,
             num_warps=32 if not is_hip() else 16,
         )
@@ -173,10 +174,10 @@ def fused_linear_cross_entropy_forward(
 
         grad_input[start_idx:end_idx] = grad_logits_chunk @ weight
 
-        if grad_weight is not None:
+        if grad_weight is not None and _input.requires_grad:
             grad_weight += torch.mm(grad_logits_chunk.t(), _input_chunk).float()
 
-        if bias is not None:
+        if bias is not None and _input.requires_grad:
             torch.add(
                 input=grad_bias,
                 other=grad_logits_chunk.sum(dim=0),

liger_kernel_nightly-0.6.2.dev20251011154427/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.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/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.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251011152316
+Version: 0.6.2.dev20251011154427
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/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.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/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.dev20251011152316 → liger_kernel_nightly-0.6.2.dev20251011154427}/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(