diff --git a/src/transformers/integrations/__init__.py b/src/transformers/integrations/__init__.py
index 2ecc31ae54cf..5d6fc98f1f25 100755
--- a/src/transformers/integrations/__init__.py
+++ b/src/transformers/integrations/__init__.py
@@ -80,6 +80,7 @@
         "replace_kernel_forward_from_hub",
         "use_kernel_forward_from_hub",
         "use_kernel_func_from_hub",
+        "use_kernel_func_from_hub_with_fallback",
         "use_kernelized_func",
     ],
     "integration_utils": [
@@ -241,6 +242,7 @@
         replace_kernel_forward_from_hub,
         use_kernel_forward_from_hub,
         use_kernel_func_from_hub,
+        use_kernel_func_from_hub_with_fallback,
         use_kernelized_func,
     )
     from .integration_utils import (
diff --git a/src/transformers/integrations/hub_kernels.py b/src/transformers/integrations/hub_kernels.py
index 25885dbb2b05..c69528c001c1 100644
--- a/src/transformers/integrations/hub_kernels.py
+++ b/src/transformers/integrations/hub_kernels.py
@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import functools
+import importlib
+import inspect
 import os
 import re
 import sys
@@ -31,6 +33,7 @@
     is_kernels_available,
     is_rocm_platform,
     is_torch_available,
+    resolve_internal_import,
 )
 from .flash_attention import flash_attention_forward
 
@@ -531,8 +534,6 @@ def lazy_load_kernel(kernel_name: str, mapping: dict[str, ModuleType | None] = _
 
     else:
         # Try to import is_{kernel_name}_available from ..utils
-        import importlib
-
         new_kernel_name = kernel_name.replace("-", "_")
         func_name = f"is_{new_kernel_name}_available"
 
@@ -655,6 +656,32 @@ def new_init(self, *args, **kwargs):
     return decorator
 
 
+def use_kernel_func_from_hub_with_fallback(package: str, func_name: str, internal_path: str | None = None):
+    # TODO: change when we sync to 0.16.x+
+    kernel_wrapper_decorator = use_kernel_func_from_hub(func_name)
+
+    def decorator(torch_function: Callable) -> Callable:
+        implementation = None
+        try:
+            module = importlib.import_module(package)
+            implementation = resolve_internal_import(module, internal_path or func_name)
+        except Exception:
+            implementation = torch_function
+        finally:
+            implementation = torch_function if implementation is None else implementation
+
+        applicable_params = inspect.signature(implementation).parameters
+
+        @functools.wraps(torch_function)
+        def wrapped(*args, **kwargs):
+            kwargs = {k: v for k, v in kwargs.items() if k in applicable_params}
+            return implementation(*args, **kwargs)
+
+        return kernel_wrapper_decorator(wrapped)
+
+    return decorator
+
+
 # Whether to allow hub kernels coming from untrusted repos, i.e. repos outside `kernels-community`
 ALLOW_ALL_KERNELS = False
 
diff --git a/src/transformers/models/qwen3_5/modeling_qwen3_5.py b/src/transformers/models/qwen3_5/modeling_qwen3_5.py
index 1f695935030c..dad105e8b954 100644
--- a/src/transformers/models/qwen3_5/modeling_qwen3_5.py
+++ b/src/transformers/models/qwen3_5/modeling_qwen3_5.py
@@ -32,7 +32,7 @@
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_forward_from_hub
+from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub_with_fallback
 from ...integrations.accelerate import force_accelerate_hooks
 from ...masking_utils import create_causal_mask, create_recurrent_attention_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
@@ -58,18 +58,13 @@
     maybe_autocast,
     merge_with_config_defaults,
 )
-from ...utils.import_utils import is_causal_conv1d_available, is_flash_linear_attention_available
+from ...utils.import_utils import is_flash_linear_attention_available
 from ...utils.output_capturing import capture_outputs
 from ...vision_utils import get_vision_bilinear_indices_and_weights, get_vision_cu_seqlens, get_vision_position_ids
 from ..auto.modeling_auto import AutoModel
 from .configuration_qwen3_5 import Qwen3_5Config, Qwen3_5TextConfig, Qwen3_5VisionConfig
 
 
-if is_causal_conv1d_available():
-    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
-else:
-    causal_conv1d_update, causal_conv1d_fn = None, None
-
 if is_flash_linear_attention_available():
     from fla.modules import FusedRMSNormGated
     from fla.ops.gated_delta_rule import chunk_gated_delta_rule, fused_recurrent_gated_delta_rule
@@ -216,17 +211,16 @@ def apply_mask_to_padding_states(hidden_states, attention_mask):
     return hidden_states
 
 
-is_fast_path_available = all(
-    (causal_conv1d_fn, causal_conv1d_update, chunk_gated_delta_rule, fused_recurrent_gated_delta_rule)
-)
+is_fast_path_available = all((chunk_gated_delta_rule, fused_recurrent_gated_delta_rule))
 
 
-def torch_causal_conv1d_update(
-    hidden_states,
-    conv_state,
-    weight,
-    bias=None,
-    activation=None,
+@use_kernel_func_from_hub_with_fallback("causal_conv1d_update", "causal_conv1d")
+def causal_conv1d_update(
+    hidden_states: torch.Tensor,
+    conv_state: torch.Tensor,
+    weight: nn.Parameter,
+    bias: nn.Parameter | None = None,
+    activation: str | None = None,
 ):
     _, hidden_size, seq_len = hidden_states.shape
     state_len = conv_state.shape[-1]
@@ -234,9 +228,33 @@ def torch_causal_conv1d_update(
     hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
     conv_state.copy_(hidden_states_new[:, :, -state_len:])
     out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
-    out = F.silu(out[:, :, -seq_len:])
-    out = out.to(hidden_states.dtype)
-    return out
+    out = out[:, :, -seq_len:]
+    if activation is not None:
+        out = ACT2FN[activation](out)
+    return out.to(hidden_states.dtype)
+
+
+@use_kernel_func_from_hub_with_fallback("causal_conv1d_fn", "causal_conv1d")
+def causal_conv1d_fn(
+    hidden_states: torch.Tensor,
+    weight: nn.Parameter,
+    bias: nn.Parameter | None = None,
+    activation: str | None = None,
+    **kwargs,
+):
+    _, hidden_size, seq_len = hidden_states.shape
+    padding = weight.shape[-1] - 1
+
+    out = F.conv1d(
+        hidden_states.to(weight.dtype),
+        weight=weight.unsqueeze(1),
+        bias=bias,
+        padding=padding,
+        groups=hidden_size,
+    )[:, :, :seq_len]
+    if activation is not None:
+        out = ACT2FN[activation](out)
+    return out.to(hidden_states.dtype)
 
 
 def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6):
@@ -385,7 +403,6 @@ def __init__(self, config: Qwen3_5Config, layer_idx: int):
         self.conv_kernel_size = config.linear_conv_kernel_dim
         self.layer_idx = layer_idx
         self.activation = config.hidden_act
-        self.act = ACT2FN[config.hidden_act]
         self.layer_norm_epsilon = config.rms_norm_eps
 
         # QKV
@@ -418,16 +435,13 @@ def __init__(self, config: Qwen3_5Config, layer_idx: int):
 
         self.out_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
 
-        self.causal_conv1d_fn = causal_conv1d_fn
-        self.causal_conv1d_update = causal_conv1d_update or torch_causal_conv1d_update
         self.chunk_gated_delta_rule = chunk_gated_delta_rule or torch_chunk_gated_delta_rule
         self.recurrent_gated_delta_rule = fused_recurrent_gated_delta_rule or torch_recurrent_gated_delta_rule
 
         if not is_fast_path_available:
             logger.warning_once(
-                "The fast path is not available because one of the required library is not installed. Falling back to "
-                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and"
-                " https://github.com/Dao-AILab/causal-conv1d"
+                "The fast path is not available because the required library is not installed. Falling back to "
+                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation"
             )
 
         self.layer_type = config.layer_types[layer_idx]
@@ -472,7 +486,7 @@ def forward(
 
         if use_precomputed_states and seq_len == 1:
             # Single-token cached decode: the fused per-step kernel updates the conv state in-place.
-            mixed_qkv = self.causal_conv1d_update(
+            mixed_qkv = causal_conv1d_update(
                 mixed_qkv,
                 conv_state,
                 self.conv1d.weight.squeeze(1),
@@ -489,16 +503,15 @@ def forward(
             if cache_params is not None:
                 new_conv_state = F.pad(mixed_qkv, (self.conv_kernel_size - mixed_qkv.shape[-1], 0))
                 cache_params.update_conv_state(new_conv_state, self.layer_idx)
-            if self.causal_conv1d_fn is not None:
-                mixed_qkv = self.causal_conv1d_fn(
-                    x=mixed_qkv,
-                    weight=self.conv1d.weight.squeeze(1),
-                    bias=self.conv1d.bias,
-                    activation=self.activation,
-                    seq_idx=kwargs.get("seq_idx"),
-                )
-            else:
-                mixed_qkv = F.silu(self.conv1d(mixed_qkv)[:, :, : mixed_qkv.shape[-1]])
+
+            mixed_qkv = causal_conv1d_fn(
+                mixed_qkv,
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                activation=self.activation,
+                seq_idx=kwargs.get("seq_idx"),
+            )
+
             if use_precomputed_states:
                 mixed_qkv = mixed_qkv[:, :, -seq_len:]
 
diff --git a/src/transformers/models/qwen3_5/modular_qwen3_5.py b/src/transformers/models/qwen3_5/modular_qwen3_5.py
index 2b88938d533d..d81edf5015a3 100644
--- a/src/transformers/models/qwen3_5/modular_qwen3_5.py
+++ b/src/transformers/models/qwen3_5/modular_qwen3_5.py
@@ -46,6 +46,8 @@
     Qwen3NextPreTrainedModel,
     Qwen3NextRMSNorm,
     apply_mask_to_padding_states,
+    causal_conv1d_fn,
+    causal_conv1d_update,
 )
 from ..qwen3_vl.configuration_qwen3_vl import Qwen3VLConfig, Qwen3VLVisionConfig
 from ..qwen3_vl.modeling_qwen3_vl import (
@@ -251,7 +253,7 @@ def forward(
 
         if use_precomputed_states and seq_len == 1:
             # Single-token cached decode: the fused per-step kernel updates the conv state in-place.
-            mixed_qkv = self.causal_conv1d_update(
+            mixed_qkv = causal_conv1d_update(
                 mixed_qkv,
                 conv_state,
                 self.conv1d.weight.squeeze(1),
@@ -268,16 +270,15 @@ def forward(
             if cache_params is not None:
                 new_conv_state = F.pad(mixed_qkv, (self.conv_kernel_size - mixed_qkv.shape[-1], 0))
                 cache_params.update_conv_state(new_conv_state, self.layer_idx)
-            if self.causal_conv1d_fn is not None:
-                mixed_qkv = self.causal_conv1d_fn(
-                    x=mixed_qkv,
-                    weight=self.conv1d.weight.squeeze(1),
-                    bias=self.conv1d.bias,
-                    activation=self.activation,
-                    seq_idx=kwargs.get("seq_idx"),
-                )
-            else:
-                mixed_qkv = F.silu(self.conv1d(mixed_qkv)[:, :, : mixed_qkv.shape[-1]])
+
+            mixed_qkv = causal_conv1d_fn(
+                mixed_qkv,
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                activation=self.activation,
+                seq_idx=kwargs.get("seq_idx"),
+            )
+
             if use_precomputed_states:
                 mixed_qkv = mixed_qkv[:, :, -seq_len:]
 
diff --git a/src/transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py b/src/transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py
index 643af3afd48b..2ad29909ec05 100644
--- a/src/transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py
+++ b/src/transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py
@@ -32,7 +32,11 @@
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_experts_implementation, use_kernel_forward_from_hub
+from ...integrations import (
+    use_experts_implementation,
+    use_kernel_forward_from_hub,
+    use_kernel_func_from_hub_with_fallback,
+)
 from ...integrations.accelerate import force_accelerate_hooks
 from ...masking_utils import create_causal_mask, create_recurrent_attention_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
@@ -55,18 +59,13 @@
     maybe_autocast,
     merge_with_config_defaults,
 )
-from ...utils.import_utils import is_causal_conv1d_available, is_flash_linear_attention_available
+from ...utils.import_utils import is_flash_linear_attention_available
 from ...utils.output_capturing import OutputRecorder, capture_outputs
 from ...vision_utils import get_vision_bilinear_indices_and_weights, get_vision_cu_seqlens, get_vision_position_ids
 from ..auto.modeling_auto import AutoModel
 from .configuration_qwen3_5_moe import Qwen3_5MoeConfig, Qwen3_5MoeTextConfig, Qwen3_5MoeVisionConfig
 
 
-if is_causal_conv1d_available():
-    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
-else:
-    causal_conv1d_update, causal_conv1d_fn = None, None
-
 if is_flash_linear_attention_available():
     from fla.modules import FusedRMSNormGated
     from fla.ops.gated_delta_rule import chunk_gated_delta_rule, fused_recurrent_gated_delta_rule
@@ -213,17 +212,16 @@ def apply_mask_to_padding_states(hidden_states, attention_mask):
     return hidden_states
 
 
-is_fast_path_available = all(
-    (causal_conv1d_fn, causal_conv1d_update, chunk_gated_delta_rule, fused_recurrent_gated_delta_rule)
-)
+is_fast_path_available = all((chunk_gated_delta_rule, fused_recurrent_gated_delta_rule))
 
 
-def torch_causal_conv1d_update(
-    hidden_states,
-    conv_state,
-    weight,
-    bias=None,
-    activation=None,
+@use_kernel_func_from_hub_with_fallback("causal_conv1d_update", "causal_conv1d")
+def causal_conv1d_update(
+    hidden_states: torch.Tensor,
+    conv_state: torch.Tensor,
+    weight: nn.Parameter,
+    bias: nn.Parameter | None = None,
+    activation: str | None = None,
 ):
     _, hidden_size, seq_len = hidden_states.shape
     state_len = conv_state.shape[-1]
@@ -231,9 +229,33 @@ def torch_causal_conv1d_update(
     hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
     conv_state.copy_(hidden_states_new[:, :, -state_len:])
     out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
-    out = F.silu(out[:, :, -seq_len:])
-    out = out.to(hidden_states.dtype)
-    return out
+    out = out[:, :, -seq_len:]
+    if activation is not None:
+        out = ACT2FN[activation](out)
+    return out.to(hidden_states.dtype)
+
+
+@use_kernel_func_from_hub_with_fallback("causal_conv1d_fn", "causal_conv1d")
+def causal_conv1d_fn(
+    hidden_states: torch.Tensor,
+    weight: nn.Parameter,
+    bias: nn.Parameter | None = None,
+    activation: str | None = None,
+    **kwargs,
+):
+    _, hidden_size, seq_len = hidden_states.shape
+    padding = weight.shape[-1] - 1
+
+    out = F.conv1d(
+        hidden_states.to(weight.dtype),
+        weight=weight.unsqueeze(1),
+        bias=bias,
+        padding=padding,
+        groups=hidden_size,
+    )[:, :, :seq_len]
+    if activation is not None:
+        out = ACT2FN[activation](out)
+    return out.to(hidden_states.dtype)
 
 
 def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6):
@@ -382,7 +404,6 @@ def __init__(self, config: Qwen3_5MoeConfig, layer_idx: int):
         self.conv_kernel_size = config.linear_conv_kernel_dim
         self.layer_idx = layer_idx
         self.activation = config.hidden_act
-        self.act = ACT2FN[config.hidden_act]
         self.layer_norm_epsilon = config.rms_norm_eps
 
         # QKV
@@ -415,16 +436,13 @@ def __init__(self, config: Qwen3_5MoeConfig, layer_idx: int):
 
         self.out_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
 
-        self.causal_conv1d_fn = causal_conv1d_fn
-        self.causal_conv1d_update = causal_conv1d_update or torch_causal_conv1d_update
         self.chunk_gated_delta_rule = chunk_gated_delta_rule or torch_chunk_gated_delta_rule
         self.recurrent_gated_delta_rule = fused_recurrent_gated_delta_rule or torch_recurrent_gated_delta_rule
 
         if not is_fast_path_available:
             logger.warning_once(
-                "The fast path is not available because one of the required library is not installed. Falling back to "
-                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and"
-                " https://github.com/Dao-AILab/causal-conv1d"
+                "The fast path is not available because the required library is not installed. Falling back to "
+                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation"
             )
 
         self.layer_type = config.layer_types[layer_idx]
@@ -469,7 +487,7 @@ def forward(
 
         if use_precomputed_states and seq_len == 1:
             # Single-token cached decode: the fused per-step kernel updates the conv state in-place.
-            mixed_qkv = self.causal_conv1d_update(
+            mixed_qkv = causal_conv1d_update(
                 mixed_qkv,
                 conv_state,
                 self.conv1d.weight.squeeze(1),
@@ -486,16 +504,15 @@ def forward(
             if cache_params is not None:
                 new_conv_state = F.pad(mixed_qkv, (self.conv_kernel_size - mixed_qkv.shape[-1], 0))
                 cache_params.update_conv_state(new_conv_state, self.layer_idx)
-            if self.causal_conv1d_fn is not None:
-                mixed_qkv = self.causal_conv1d_fn(
-                    x=mixed_qkv,
-                    weight=self.conv1d.weight.squeeze(1),
-                    bias=self.conv1d.bias,
-                    activation=self.activation,
-                    seq_idx=kwargs.get("seq_idx"),
-                )
-            else:
-                mixed_qkv = F.silu(self.conv1d(mixed_qkv)[:, :, : mixed_qkv.shape[-1]])
+
+            mixed_qkv = causal_conv1d_fn(
+                mixed_qkv,
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                activation=self.activation,
+                seq_idx=kwargs.get("seq_idx"),
+            )
+
             if use_precomputed_states:
                 mixed_qkv = mixed_qkv[:, :, -seq_len:]
 
diff --git a/src/transformers/models/qwen3_next/modeling_qwen3_next.py b/src/transformers/models/qwen3_next/modeling_qwen3_next.py
index 95c506aaf8e1..e618ad1462e8 100644
--- a/src/transformers/models/qwen3_next/modeling_qwen3_next.py
+++ b/src/transformers/models/qwen3_next/modeling_qwen3_next.py
@@ -29,7 +29,7 @@
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_experts_implementation
+from ...integrations import use_experts_implementation, use_kernel_func_from_hub_with_fallback, use_kernelized_func
 from ...integrations.accelerate import force_accelerate_hooks
 from ...masking_utils import create_causal_mask, create_recurrent_attention_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
@@ -45,16 +45,11 @@
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
 from ...utils.generic import maybe_autocast, merge_with_config_defaults
-from ...utils.import_utils import is_causal_conv1d_available, is_flash_linear_attention_available
+from ...utils.import_utils import is_flash_linear_attention_available
 from ...utils.output_capturing import OutputRecorder, capture_outputs
 from .configuration_qwen3_next import Qwen3NextConfig
 
 
-if is_causal_conv1d_available():
-    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
-else:
-    causal_conv1d_update, causal_conv1d_fn = None, None
-
 if is_flash_linear_attention_available():
     from fla.modules import FusedRMSNormGated
     from fla.ops.gated_delta_rule import chunk_gated_delta_rule, fused_recurrent_gated_delta_rule
@@ -342,17 +337,16 @@ def apply_mask_to_padding_states(hidden_states, attention_mask):
     return hidden_states
 
 
-is_fast_path_available = all(
-    (causal_conv1d_fn, causal_conv1d_update, chunk_gated_delta_rule, fused_recurrent_gated_delta_rule)
-)
+is_fast_path_available = all((chunk_gated_delta_rule, fused_recurrent_gated_delta_rule))
 
 
-def torch_causal_conv1d_update(
-    hidden_states,
-    conv_state,
-    weight,
-    bias=None,
-    activation=None,
+@use_kernel_func_from_hub_with_fallback("causal_conv1d_update", "causal_conv1d")
+def causal_conv1d_update(
+    hidden_states: torch.Tensor,
+    conv_state: torch.Tensor,
+    weight: nn.Parameter,
+    bias: nn.Parameter | None = None,
+    activation: str | None = None,
 ):
     _, hidden_size, seq_len = hidden_states.shape
     state_len = conv_state.shape[-1]
@@ -360,9 +354,33 @@ def torch_causal_conv1d_update(
     hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
     conv_state.copy_(hidden_states_new[:, :, -state_len:])
     out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
-    out = F.silu(out[:, :, -seq_len:])
-    out = out.to(hidden_states.dtype)
-    return out
+    out = out[:, :, -seq_len:]
+    if activation is not None:
+        out = ACT2FN[activation](out)
+    return out.to(hidden_states.dtype)
+
+
+@use_kernel_func_from_hub_with_fallback("causal_conv1d_fn", "causal_conv1d")
+def causal_conv1d_fn(
+    hidden_states: torch.Tensor,
+    weight: nn.Parameter,
+    bias: nn.Parameter | None = None,
+    activation: str | None = None,
+    **kwargs,
+):
+    _, hidden_size, seq_len = hidden_states.shape
+    padding = weight.shape[-1] - 1
+
+    out = F.conv1d(
+        hidden_states.to(weight.dtype),
+        weight=weight.unsqueeze(1),
+        bias=bias,
+        padding=padding,
+        groups=hidden_size,
+    )[:, :, :seq_len]
+    if activation is not None:
+        out = ACT2FN[activation](out)
+    return out.to(hidden_states.dtype)
 
 
 def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6):
@@ -496,6 +514,7 @@ def torch_recurrent_gated_delta_rule(
     return core_attn_out, last_recurrent_state
 
 
+@use_kernelized_func([causal_conv1d_update, causal_conv1d_fn])
 class Qwen3NextGatedDeltaNet(nn.Module):
     def __init__(self, config: Qwen3NextConfig, layer_idx: int):
         super().__init__()
@@ -510,7 +529,6 @@ def __init__(self, config: Qwen3NextConfig, layer_idx: int):
         self.conv_kernel_size = config.linear_conv_kernel_dim
         self.layer_idx = layer_idx
         self.activation = config.hidden_act
-        self.act = ACT2FN[config.hidden_act]
         self.layer_norm_epsilon = config.rms_norm_eps
 
         # QKV
@@ -549,16 +567,13 @@ def __init__(self, config: Qwen3NextConfig, layer_idx: int):
 
         self.out_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
 
-        self.causal_conv1d_fn = causal_conv1d_fn
-        self.causal_conv1d_update = causal_conv1d_update or torch_causal_conv1d_update
         self.chunk_gated_delta_rule = chunk_gated_delta_rule or torch_chunk_gated_delta_rule
         self.recurrent_gated_delta_rule = fused_recurrent_gated_delta_rule or torch_recurrent_gated_delta_rule
 
         if not is_fast_path_available:
             logger.warning_once(
-                "The fast path is not available because one of the required library is not installed. Falling back to "
-                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and"
-                " https://github.com/Dao-AILab/causal-conv1d"
+                "The fast path is not available because the required library is not installed. Falling back to "
+                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation"
             )
 
         self.layer_type = config.layer_types[layer_idx]
@@ -626,7 +641,7 @@ def forward(
 
         if use_precomputed_states and seq_len == 1:
             # Single-token cached decode: the fused per-step kernel updates the conv state in-place.
-            mixed_qkv = self.causal_conv1d_update(
+            mixed_qkv = causal_conv1d_update(
                 mixed_qkv,
                 conv_state,
                 self.conv1d.weight.squeeze(1),
@@ -643,16 +658,15 @@ def forward(
             if cache_params is not None:
                 new_conv_state = F.pad(mixed_qkv, (self.conv_kernel_size - mixed_qkv.shape[-1], 0))
                 cache_params.update_conv_state(new_conv_state, self.layer_idx)
-            if self.causal_conv1d_fn is not None:
-                mixed_qkv = self.causal_conv1d_fn(
-                    x=mixed_qkv,
-                    weight=self.conv1d.weight.squeeze(1),
-                    bias=self.conv1d.bias,
-                    activation=self.activation,
-                    seq_idx=kwargs.get("seq_idx"),
-                )
-            else:
-                mixed_qkv = F.silu(self.conv1d(mixed_qkv)[:, :, : mixed_qkv.shape[-1]])
+
+            mixed_qkv = causal_conv1d_fn(
+                mixed_qkv,
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                activation=self.activation,
+                seq_idx=kwargs.get("seq_idx"),
+            )
+
             if use_precomputed_states:
                 mixed_qkv = mixed_qkv[:, :, -seq_len:]
 
diff --git a/src/transformers/models/qwen3_next/modular_qwen3_next.py b/src/transformers/models/qwen3_next/modular_qwen3_next.py
index 86f31ad1431d..6e9e7bffba37 100644
--- a/src/transformers/models/qwen3_next/modular_qwen3_next.py
+++ b/src/transformers/models/qwen3_next/modular_qwen3_next.py
@@ -23,6 +23,7 @@
 from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
+from ...integrations import use_kernel_func_from_hub_with_fallback, use_kernelized_func
 from ...integrations.accelerate import force_accelerate_hooks
 from ...masking_utils import create_causal_mask, create_recurrent_attention_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
@@ -31,10 +32,7 @@
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, logging
 from ...utils.generic import merge_with_config_defaults, no_inherit_decorator
-from ...utils.import_utils import (
-    is_causal_conv1d_available,
-    is_flash_linear_attention_available,
-)
+from ...utils.import_utils import is_flash_linear_attention_available
 from ...utils.output_capturing import OutputRecorder, capture_outputs
 from ..bamba.modeling_bamba import apply_mask_to_padding_states, apply_rotary_pos_emb
 from ..gemma2.modeling_gemma2 import Gemma2RotaryEmbedding
@@ -55,11 +53,6 @@
 from .configuration_qwen3_next import Qwen3NextConfig
 
 
-if is_causal_conv1d_available():
-    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
-else:
-    causal_conv1d_update, causal_conv1d_fn = None, None
-
 if is_flash_linear_attention_available():
     from fla.modules import FusedRMSNormGated
     from fla.ops.gated_delta_rule import chunk_gated_delta_rule, fused_recurrent_gated_delta_rule
@@ -68,9 +61,7 @@
     FusedRMSNormGated = None
 
 
-is_fast_path_available = all(
-    (causal_conv1d_fn, causal_conv1d_update, chunk_gated_delta_rule, fused_recurrent_gated_delta_rule)
-)
+is_fast_path_available = all((chunk_gated_delta_rule, fused_recurrent_gated_delta_rule))
 
 logger = logging.get_logger(__name__)
 
@@ -188,12 +179,13 @@ def forward(
         return attn_output, attn_weights
 
 
-def torch_causal_conv1d_update(
-    hidden_states,
-    conv_state,
-    weight,
-    bias=None,
-    activation=None,
+@use_kernel_func_from_hub_with_fallback("causal_conv1d_update", "causal_conv1d")
+def causal_conv1d_update(
+    hidden_states: torch.Tensor,
+    conv_state: torch.Tensor,
+    weight: nn.Parameter,
+    bias: nn.Parameter | None = None,
+    activation: str | None = None,
 ):
     _, hidden_size, seq_len = hidden_states.shape
     state_len = conv_state.shape[-1]
@@ -201,9 +193,33 @@ def torch_causal_conv1d_update(
     hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
     conv_state.copy_(hidden_states_new[:, :, -state_len:])
     out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
-    out = F.silu(out[:, :, -seq_len:])
-    out = out.to(hidden_states.dtype)
-    return out
+    out = out[:, :, -seq_len:]
+    if activation is not None:
+        out = ACT2FN[activation](out)
+    return out.to(hidden_states.dtype)
+
+
+@use_kernel_func_from_hub_with_fallback("causal_conv1d_fn", "causal_conv1d")
+def causal_conv1d_fn(
+    hidden_states: torch.Tensor,
+    weight: nn.Parameter,
+    bias: nn.Parameter | None = None,
+    activation: str | None = None,
+    **kwargs,
+):
+    _, hidden_size, seq_len = hidden_states.shape
+    padding = weight.shape[-1] - 1
+
+    out = F.conv1d(
+        hidden_states.to(weight.dtype),
+        weight=weight.unsqueeze(1),
+        bias=bias,
+        padding=padding,
+        groups=hidden_size,
+    )[:, :, :seq_len]
+    if activation is not None:
+        out = ACT2FN[activation](out)
+    return out.to(hidden_states.dtype)
 
 
 def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6):
@@ -337,6 +353,7 @@ def torch_recurrent_gated_delta_rule(
     return core_attn_out, last_recurrent_state
 
 
+@use_kernelized_func([causal_conv1d_update, causal_conv1d_fn])
 class Qwen3NextGatedDeltaNet(nn.Module):
     def __init__(self, config: Qwen3NextConfig, layer_idx: int):
         super().__init__()
@@ -351,7 +368,6 @@ def __init__(self, config: Qwen3NextConfig, layer_idx: int):
         self.conv_kernel_size = config.linear_conv_kernel_dim
         self.layer_idx = layer_idx
         self.activation = config.hidden_act
-        self.act = ACT2FN[config.hidden_act]
         self.layer_norm_epsilon = config.rms_norm_eps
 
         # QKV
@@ -390,16 +406,13 @@ def __init__(self, config: Qwen3NextConfig, layer_idx: int):
 
         self.out_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
 
-        self.causal_conv1d_fn = causal_conv1d_fn
-        self.causal_conv1d_update = causal_conv1d_update or torch_causal_conv1d_update
         self.chunk_gated_delta_rule = chunk_gated_delta_rule or torch_chunk_gated_delta_rule
         self.recurrent_gated_delta_rule = fused_recurrent_gated_delta_rule or torch_recurrent_gated_delta_rule
 
         if not is_fast_path_available:
             logger.warning_once(
-                "The fast path is not available because one of the required library is not installed. Falling back to "
-                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and"
-                " https://github.com/Dao-AILab/causal-conv1d"
+                "The fast path is not available because the required library is not installed. Falling back to "
+                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation"
             )
 
         self.layer_type = config.layer_types[layer_idx]
@@ -467,7 +480,7 @@ def forward(
 
         if use_precomputed_states and seq_len == 1:
             # Single-token cached decode: the fused per-step kernel updates the conv state in-place.
-            mixed_qkv = self.causal_conv1d_update(
+            mixed_qkv = causal_conv1d_update(
                 mixed_qkv,
                 conv_state,
                 self.conv1d.weight.squeeze(1),
@@ -484,16 +497,15 @@ def forward(
             if cache_params is not None:
                 new_conv_state = F.pad(mixed_qkv, (self.conv_kernel_size - mixed_qkv.shape[-1], 0))
                 cache_params.update_conv_state(new_conv_state, self.layer_idx)
-            if self.causal_conv1d_fn is not None:
-                mixed_qkv = self.causal_conv1d_fn(
-                    x=mixed_qkv,
-                    weight=self.conv1d.weight.squeeze(1),
-                    bias=self.conv1d.bias,
-                    activation=self.activation,
-                    seq_idx=kwargs.get("seq_idx"),
-                )
-            else:
-                mixed_qkv = F.silu(self.conv1d(mixed_qkv)[:, :, : mixed_qkv.shape[-1]])
+
+            mixed_qkv = causal_conv1d_fn(
+                mixed_qkv,
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                activation=self.activation,
+                seq_idx=kwargs.get("seq_idx"),
+            )
+
             if use_precomputed_states:
                 mixed_qkv = mixed_qkv[:, :, -seq_len:]