chore: add .gitignore

unsloth_compiled_cache/AqlmLoraLinear_peft_forward.py

@@ -0,0 +1,89 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from peft.tuners.lora.aqlm import (torch)
+
+
+torch_addmm = torch.addmm
+torch_add   = torch.add
+# @torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def lora_forward(result, lora_A, lora_B, dropout, x, scaling):
+    # Use result.dtype (bfloat16 from base layer) since x may have been cast to float32
+    # by _cast_input_dtype when autocast is disabled
+    target_dtype = result.dtype
+    xA = dropout(x).to(target_dtype) @ lora_A.weight.to(target_dtype).t()
+    # output = result + scaling * xA @ lora_B.weight.t()
+    shape = result.shape
+    output = torch_addmm(
+        result.view(-1, shape[-1]),
+        xA.view(-1, xA.shape[-1]),
+        lora_B.weight.to(target_dtype).t(),
+        alpha = scaling,
+        beta = 1,
+    ).view(shape)
+
+    bias = lora_B.bias
+    if bias is not None:
+        output = torch_add(
+            output,
+            bias.to(target_dtype),
+            alpha = scaling,
+        )
+    return output
+pass
+
+def unsloth_forward(self, x: torch.Tensor):
+    # note: logic differs from default Linear because merging is not supported
+    result = self.base_layer(x)
+
+    if self.disable_adapters:
+        return result
+
+    for active_adapter in self.active_adapters:
+        if active_adapter not in self.lora_A.keys():
+            continue
+        lora_A = self.lora_A[active_adapter]
+        lora_B = self.lora_B[active_adapter]
+        dropout = self.lora_dropout[active_adapter]
+        scaling = self.scaling[active_adapter]
+
+        requires_conversion = not torch.is_autocast_enabled()
+        if requires_conversion:
+            expected_dtype = result.dtype
+            x = self._cast_input_dtype(x, lora_A.weight.dtype)
+
+        output = lora_B(lora_A(dropout(x)))
+        if requires_conversion:
+            output = output.to(expected_dtype)
+        output = output * scaling
+        result += output
+    return result

unsloth_compiled_cache/AwqLoraLinear_peft_forward.py

@@ -0,0 +1,88 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from peft.tuners.lora.awq import (torch)
+
+
+torch_addmm = torch.addmm
+torch_add   = torch.add
+# @torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def lora_forward(result, lora_A, lora_B, dropout, x, scaling):
+    # Use result.dtype (bfloat16 from base layer) since x may have been cast to float32
+    # by _cast_input_dtype when autocast is disabled
+    target_dtype = result.dtype
+    xA = dropout(x).to(target_dtype) @ lora_A.weight.to(target_dtype).t()
+    # output = result + scaling * xA @ lora_B.weight.t()
+    shape = result.shape
+    output = torch_addmm(
+        result.view(-1, shape[-1]),
+        xA.view(-1, xA.shape[-1]),
+        lora_B.weight.to(target_dtype).t(),
+        alpha = scaling,
+        beta = 1,
+    ).view(shape)
+
+    bias = lora_B.bias
+    if bias is not None:
+        output = torch_add(
+            output,
+            bias.to(target_dtype),
+            alpha = scaling,
+        )
+    return output
+pass
+
+def unsloth_forward(self, x: torch.Tensor):
+    result = self.quant_linear_module(x)
+
+    if self.disable_adapters:
+        return result
+
+    for active_adapter in self.active_adapters:
+        if active_adapter not in self.lora_A.keys():
+            continue
+        lora_A = self.lora_A[active_adapter]
+        lora_B = self.lora_B[active_adapter]
+        dropout = self.lora_dropout[active_adapter]
+        scaling = self.scaling[active_adapter]
+
+        requires_conversion = not torch.is_autocast_enabled()
+        if requires_conversion:
+            expected_dtype = result.dtype
+            x = self._cast_input_dtype(x, lora_A.weight.dtype)
+
+        output = lora_B(lora_A(dropout(x)))
+        if requires_conversion:
+            output = output.to(expected_dtype)
+        output = output * scaling
+        result = result + output
+    return result

unsloth_compiled_cache/BatchNorm1d.py

@@ -0,0 +1,117 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.gemma3.modeling_gemma3 import (nn)
+
+def forward(self, input: Tensor) -> Tensor:
+    self._check_input_dim(input)
+
+    # exponential_average_factor is set to self.momentum
+    # (when it is available) only so that it gets updated
+    # in ONNX graph when this node is exported to ONNX.
+    if self.momentum is None:
+        exponential_average_factor = 0.0
+    else:
+        exponential_average_factor = self.momentum
+
+    if self.training and self.track_running_stats:
+        # TODO: if statement only here to tell the jit to skip emitting this when it is None
+        if self.num_batches_tracked is not None:  # type: ignore[has-type]
+            self.num_batches_tracked.add_(1)  # type: ignore[has-type]
+            if self.momentum is None:  # use cumulative moving average
+                exponential_average_factor = 1.0 / float(self.num_batches_tracked)
+            else:  # use exponential moving average
+                exponential_average_factor = self.momentum
+
+    r"""
+    Decide whether the mini-batch stats should be used for normalization rather than the buffers.
+    Mini-batch stats are used in training mode, and in eval mode when buffers are None.
+    """
+    if self.training:
+        bn_training = True
+    else:
+        bn_training = (self.running_mean is None) and (self.running_var is None)
+
+    r"""
+    Buffers are only updated if they are to be tracked and we are in training mode. Thus they only need to be
+    passed when the update should occur (i.e. in training mode when they are tracked), or when buffer stats are
+    used for normalization (i.e. in eval mode when buffers are not None).
+    """
+    return F.batch_norm(
+        input,
+        # If buffers are not to be tracked, ensure that they won't be updated
+        (
+            self.running_mean
+            if not self.training or self.track_running_stats
+            else None
+        ),
+        self.running_var if not self.training or self.track_running_stats else None,
+        self.weight,
+        self.bias,
+        bn_training,
+        exponential_average_factor,
+        self.eps,
+    ).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/BatchNorm2d.py

@@ -0,0 +1,117 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.gemma3.modeling_gemma3 import (nn)
+
+def forward(self, input: Tensor) -> Tensor:
+    self._check_input_dim(input)
+
+    # exponential_average_factor is set to self.momentum
+    # (when it is available) only so that it gets updated
+    # in ONNX graph when this node is exported to ONNX.
+    if self.momentum is None:
+        exponential_average_factor = 0.0
+    else:
+        exponential_average_factor = self.momentum
+
+    if self.training and self.track_running_stats:
+        # TODO: if statement only here to tell the jit to skip emitting this when it is None
+        if self.num_batches_tracked is not None:  # type: ignore[has-type]
+            self.num_batches_tracked.add_(1)  # type: ignore[has-type]
+            if self.momentum is None:  # use cumulative moving average
+                exponential_average_factor = 1.0 / float(self.num_batches_tracked)
+            else:  # use exponential moving average
+                exponential_average_factor = self.momentum
+
+    r"""
+    Decide whether the mini-batch stats should be used for normalization rather than the buffers.
+    Mini-batch stats are used in training mode, and in eval mode when buffers are None.
+    """
+    if self.training:
+        bn_training = True
+    else:
+        bn_training = (self.running_mean is None) and (self.running_var is None)
+
+    r"""
+    Buffers are only updated if they are to be tracked and we are in training mode. Thus they only need to be
+    passed when the update should occur (i.e. in training mode when they are tracked), or when buffer stats are
+    used for normalization (i.e. in eval mode when buffers are not None).
+    """
+    return F.batch_norm(
+        input,
+        # If buffers are not to be tracked, ensure that they won't be updated
+        (
+            self.running_mean
+            if not self.training or self.track_running_stats
+            else None
+        ),
+        self.running_var if not self.training or self.track_running_stats else None,
+        self.weight,
+        self.bias,
+        bn_training,
+        exponential_average_factor,
+        self.eps,
+    ).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/BatchNorm3d.py

@@ -0,0 +1,117 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.gemma3.modeling_gemma3 import (nn)
+
+def forward(self, input: Tensor) -> Tensor:
+    self._check_input_dim(input)
+
+    # exponential_average_factor is set to self.momentum
+    # (when it is available) only so that it gets updated
+    # in ONNX graph when this node is exported to ONNX.
+    if self.momentum is None:
+        exponential_average_factor = 0.0
+    else:
+        exponential_average_factor = self.momentum
+
+    if self.training and self.track_running_stats:
+        # TODO: if statement only here to tell the jit to skip emitting this when it is None
+        if self.num_batches_tracked is not None:  # type: ignore[has-type]
+            self.num_batches_tracked.add_(1)  # type: ignore[has-type]
+            if self.momentum is None:  # use cumulative moving average
+                exponential_average_factor = 1.0 / float(self.num_batches_tracked)
+            else:  # use exponential moving average
+                exponential_average_factor = self.momentum
+
+    r"""
+    Decide whether the mini-batch stats should be used for normalization rather than the buffers.
+    Mini-batch stats are used in training mode, and in eval mode when buffers are None.
+    """
+    if self.training:
+        bn_training = True
+    else:
+        bn_training = (self.running_mean is None) and (self.running_var is None)
+
+    r"""
+    Buffers are only updated if they are to be tracked and we are in training mode. Thus they only need to be
+    passed when the update should occur (i.e. in training mode when they are tracked), or when buffer stats are
+    used for normalization (i.e. in eval mode when buffers are not None).
+    """
+    return F.batch_norm(
+        input,
+        # If buffers are not to be tracked, ensure that they won't be updated
+        (
+            self.running_mean
+            if not self.training or self.track_running_stats
+            else None
+        ),
+        self.running_var if not self.training or self.track_running_stats else None,
+        self.weight,
+        self.bias,
+        bn_training,
+        exponential_average_factor,
+        self.eps,
+    ).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/Conv1d.py

@@ -0,0 +1,70 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+
+
+def forward(self, input: Tensor) -> Tensor:
+    return self._conv_forward(input, self.weight, self.bias).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/Conv2d.py

@@ -0,0 +1,70 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+
+
+def forward(self, input: Tensor) -> Tensor:
+    return self._conv_forward(input, self.weight, self.bias).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/Conv3d.py

@@ -0,0 +1,70 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+
+
+def forward(self, input: Tensor) -> Tensor:
+    return self._conv_forward(input, self.weight, self.bias).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/ConvTranspose1d.py

@@ -0,0 +1,97 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.gemma3.modeling_gemma3 import (nn)
+
+def forward(self, input: Tensor, output_size: list[int] | None = None) -> Tensor:
+    if self.padding_mode != "zeros":
+        raise ValueError(
+            "Only `zeros` padding mode is supported for ConvTranspose1d"
+        )
+
+    assert isinstance(self.padding, tuple)
+    # One cannot replace List by Tuple or Sequence in "_output_padding" because
+    # TorchScript does not support `Sequence[T]` or `Tuple[T, ...]`.
+    num_spatial_dims = 1
+    output_padding = self._output_padding(
+        input,
+        output_size,
+        self.stride,  # type: ignore[arg-type]
+        self.padding,  # type: ignore[arg-type]
+        self.kernel_size,  # type: ignore[arg-type]
+        num_spatial_dims,
+        self.dilation,  # type: ignore[arg-type]
+    )
+    return F.conv_transpose1d(
+        input,
+        self.weight,
+        self.bias,
+        self.stride,
+        self.padding,
+        output_padding,
+        self.groups,
+        self.dilation,
+    ).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/ConvTranspose2d.py

@@ -0,0 +1,106 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.gemma3.modeling_gemma3 import (nn)
+
+def forward(self, input: Tensor, output_size: list[int] | None = None) -> Tensor:
+    """
+    Performs the forward pass.
+
+    Attributes:
+        input (Tensor): The input tensor.
+        output_size (list[int], optional): A list of integers representing
+            the size of the output tensor. Default is None.
+    """
+    if self.padding_mode != "zeros":
+        raise ValueError(
+            "Only `zeros` padding mode is supported for ConvTranspose2d"
+        )
+
+    assert isinstance(self.padding, tuple)
+    # One cannot replace List by Tuple or Sequence in "_output_padding" because
+    # TorchScript does not support `Sequence[T]` or `Tuple[T, ...]`.
+    num_spatial_dims = 2
+    output_padding = self._output_padding(
+        input,
+        output_size,
+        self.stride,  # type: ignore[arg-type]
+        self.padding,  # type: ignore[arg-type]
+        self.kernel_size,  # type: ignore[arg-type]
+        num_spatial_dims,
+        self.dilation,  # type: ignore[arg-type]
+    )
+
+    return F.conv_transpose2d(
+        input,
+        self.weight,
+        self.bias,
+        self.stride,
+        self.padding,
+        output_padding,
+        self.groups,
+        self.dilation,
+    ).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/ConvTranspose3d.py

@@ -0,0 +1,98 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.gemma3.modeling_gemma3 import (nn)
+
+def forward(self, input: Tensor, output_size: list[int] | None = None) -> Tensor:
+    if self.padding_mode != "zeros":
+        raise ValueError(
+            "Only `zeros` padding mode is supported for ConvTranspose3d"
+        )
+
+    assert isinstance(self.padding, tuple)
+    # One cannot replace List by Tuple or Sequence in "_output_padding" because
+    # TorchScript does not support `Sequence[T]` or `Tuple[T, ...]`.
+    num_spatial_dims = 3
+    output_padding = self._output_padding(
+        input,
+        output_size,
+        self.stride,  # type: ignore[arg-type]
+        self.padding,  # type: ignore[arg-type]
+        self.kernel_size,  # type: ignore[arg-type]
+        num_spatial_dims,
+        self.dilation,  # type: ignore[arg-type]
+    )
+
+    return F.conv_transpose3d(
+        input,
+        self.weight,
+        self.bias,
+        self.stride,
+        self.padding,
+        output_padding,
+        self.groups,
+        self.dilation,
+    ).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/GPTQLoraLinear_peft_forward.py

@@ -0,0 +1,96 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from peft.tuners.lora.gptq import (torch)
+
+
+torch_addmm = torch.addmm
+torch_add   = torch.add
+# @torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def lora_forward(result, lora_A, lora_B, dropout, x, scaling):
+    # Use result.dtype (bfloat16 from base layer) since x may have been cast to float32
+    # by _cast_input_dtype when autocast is disabled
+    target_dtype = result.dtype
+    xA = dropout(x).to(target_dtype) @ lora_A.weight.to(target_dtype).t()
+    # output = result + scaling * xA @ lora_B.weight.t()
+    shape = result.shape
+    output = torch_addmm(
+        result.view(-1, shape[-1]),
+        xA.view(-1, xA.shape[-1]),
+        lora_B.weight.to(target_dtype).t(),
+        alpha = scaling,
+        beta = 1,
+    ).view(shape)
+
+    bias = lora_B.bias
+    if bias is not None:
+        output = torch_add(
+            output,
+            bias.to(target_dtype),
+            alpha = scaling,
+        )
+    return output
+pass
+
+def unsloth_forward(self, x: torch.Tensor):
+    # note: logic differs from default Linear because merging is not supported
+    result = self.quant_linear_module(x)
+
+    if self.disable_adapters:
+        return result
+
+    lora_A_keys = self.lora_A.keys()
+
+    for active_adapter in self.active_adapters:
+        if active_adapter not in lora_A_keys:
+            continue
+        torch_result_dtype = result.dtype
+
+        lora_A = self.lora_A[active_adapter]
+        lora_B = self.lora_B[active_adapter]
+        dropout = self.lora_dropout[active_adapter]
+        scaling = self.scaling[active_adapter]
+
+        if not torch.is_autocast_enabled(): result, x = result.to(lora_A.weight.dtype), x.to(lora_A.weight.dtype)
+
+        if active_adapter not in self.lora_variant:  # vanilla LoRA
+            return lora_forward(result, lora_A, lora_B, dropout, x, scaling)
+        else:
+            result = self.lora_variant[active_adapter].forward(
+                self,
+                active_adapter=active_adapter,
+                x=x,
+                result=result,
+            )
+
+        result = result.to(torch_result_dtype)
+    return result

unsloth_compiled_cache/GroupNorm.py

@@ -0,0 +1,70 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+
+
+def forward(self, input: Tensor) -> Tensor:
+    return F.group_norm(input, self.num_groups, self.weight, self.bias, self.eps).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/LayerNorm.py

@@ -0,0 +1,72 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+
+
+def forward(self, input: Tensor) -> Tensor:
+    return F.layer_norm(
+        input, self.normalized_shape, self.weight, self.bias, self.eps
+    ).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/Linear4bit_peft_forward.py

@@ -0,0 +1,126 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+try:
+    from peft.tuners.lora.layer import VARIANT_KWARG_KEYS
+except ImportError:
+    VARIANT_KWARG_KEYS = ['alora_offsets']
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from peft.tuners.lora.bnb import (VARIANT_KWARG_KEYS, torch)
+
+
+torch_addmm = torch.addmm
+torch_add   = torch.add
+# @torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def lora_forward(result, lora_A, lora_B, dropout, x, scaling):
+    # Use result.dtype (bfloat16 from base layer) since x may have been cast to float32
+    # by _cast_input_dtype when autocast is disabled
+    target_dtype = result.dtype
+    xA = dropout(x).to(target_dtype) @ lora_A.weight.to(target_dtype).t()
+    # output = result + scaling * xA @ lora_B.weight.t()
+    shape = result.shape
+    output = torch_addmm(
+        result.view(-1, shape[-1]),
+        xA.view(-1, xA.shape[-1]),
+        lora_B.weight.to(target_dtype).t(),
+        alpha = scaling,
+        beta = 1,
+    ).view(shape)
+
+    bias = lora_B.bias
+    if bias is not None:
+        output = torch_add(
+            output,
+            bias.to(target_dtype),
+            alpha = scaling,
+        )
+    return output
+pass
+
+def unsloth_forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+    
+    adapter_names = kwargs.pop("adapter_names", None)
+    variant_kwargs = {k: kwargs.pop(k, None) for k in VARIANT_KWARG_KEYS}  # don't pass these to base_layer
+
+    if self.disable_adapters:
+        if self.merged:
+            self.unmerge()
+        if not torch.is_autocast_enabled() and hasattr(self.base_layer, 'weight') and self.base_layer.weight is not None and not hasattr(self.base_layer.weight, 'quant_state') and x.dtype != self.base_layer.weight.dtype:
+            x = x.to(self.base_layer.weight.dtype)
+        result = self.base_layer(x, *args, **kwargs)
+    elif adapter_names is not None:
+        result = self._mixed_batch_forward(x, *args, adapter_names=adapter_names, **variant_kwargs, **kwargs)
+    elif self.merged:
+        if not torch.is_autocast_enabled() and hasattr(self.base_layer, 'weight') and self.base_layer.weight is not None and not hasattr(self.base_layer.weight, 'quant_state') and x.dtype != self.base_layer.weight.dtype:
+            x = x.to(self.base_layer.weight.dtype)
+        result = self.base_layer(x, *args, **kwargs)
+    else:
+        if not torch.is_autocast_enabled() and hasattr(self.base_layer, 'weight') and self.base_layer.weight is not None and not hasattr(self.base_layer.weight, 'quant_state') and x.dtype != self.base_layer.weight.dtype:
+            x = x.to(self.base_layer.weight.dtype)
+        result = self.base_layer(x, *args, **kwargs)
+        # As per Tim Dettmers, for 4bit, we need to defensively clone here.
+        # The reason is that in some cases, an error can occur that backprop
+        # does not work on a manipulated view. This issue may be solved with
+        # newer PyTorch versions but this would need extensive testing to be
+        # sure.
+        
+
+        for active_adapter in self.active_adapters:
+            if active_adapter not in self.lora_A.keys():
+                continue
+            lora_A = self.lora_A[active_adapter]
+            lora_B = self.lora_B[active_adapter]
+            dropout = self.lora_dropout[active_adapter]
+            scaling = self.scaling[active_adapter]
+
+            requires_conversion = not torch.is_autocast_enabled()
+            if requires_conversion:
+                expected_dtype = result.dtype
+                x = self._cast_input_dtype(x, lora_A.weight.dtype)
+
+            if active_adapter not in self.lora_variant:  # vanilla LoRA
+                return lora_forward(result, lora_A, lora_B, dropout, x, scaling)
+                if requires_conversion:
+                    output = output.to(expected_dtype)
+                result = result + output
+            else:
+                result = self.lora_variant[active_adapter].forward(
+                    self,
+                    active_adapter=active_adapter,
+                    x=x,
+                    result=result,
+                    **variant_kwargs,
+                    **kwargs,
+                )
+                if requires_conversion:
+                    result = result.to(expected_dtype)
+
+    return result

unsloth_compiled_cache/Linear8bitLt_peft_forward.py

@@ -0,0 +1,118 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+try:
+    from peft.tuners.lora.layer import VARIANT_KWARG_KEYS
+except ImportError:
+    VARIANT_KWARG_KEYS = ['alora_offsets']
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+
+import torch._dynamo
+@torch._dynamo.disable
+def _call_8bit_base_layer(base_layer, x, *args, **kwargs):
+    return base_layer(x, *args, **kwargs)
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from peft.tuners.lora.bnb import (VARIANT_KWARG_KEYS, torch)
+
+
+torch_addmm = torch.addmm
+torch_add   = torch.add
+# @torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def lora_forward(result, lora_A, lora_B, dropout, x, scaling):
+    # Use result.dtype (bfloat16 from base layer) since x may have been cast to float32
+    # by _cast_input_dtype when autocast is disabled
+    target_dtype = result.dtype
+    xA = dropout(x).to(target_dtype) @ lora_A.weight.to(target_dtype).t()
+    # output = result + scaling * xA @ lora_B.weight.t()
+    shape = result.shape
+    output = torch_addmm(
+        result.view(-1, shape[-1]),
+        xA.view(-1, xA.shape[-1]),
+        lora_B.weight.to(target_dtype).t(),
+        alpha = scaling,
+        beta = 1,
+    ).view(shape)
+
+    bias = lora_B.bias
+    if bias is not None:
+        output = torch_add(
+            output,
+            bias.to(target_dtype),
+            alpha = scaling,
+        )
+    return output
+pass
+
+def unsloth_forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+    
+    adapter_names = kwargs.pop("adapter_names", None)
+    variant_kwargs = {k: kwargs.pop(k, None) for k in VARIANT_KWARG_KEYS}  # don't pass these to base_layer
+
+    if self.disable_adapters:
+        if self.merged:
+            self.unmerge()
+        result = _call_8bit_base_layer(self.base_layer, x, *args, **kwargs)
+    elif adapter_names is not None:
+        result = self._mixed_batch_forward(x, *args, adapter_names=adapter_names, **variant_kwargs, **kwargs)
+    elif self.merged:
+        result = _call_8bit_base_layer(self.base_layer, x, *args, **kwargs)
+    else:
+        result = _call_8bit_base_layer(self.base_layer, x, *args, **kwargs)
+        for active_adapter in self.active_adapters:
+            if active_adapter not in self.lora_A.keys():
+                continue
+            lora_A = self.lora_A[active_adapter]
+            lora_B = self.lora_B[active_adapter]
+            dropout = self.lora_dropout[active_adapter]
+            scaling = self.scaling[active_adapter]
+
+            requires_conversion = not torch.is_autocast_enabled()
+            if requires_conversion:
+                expected_dtype = result.dtype
+                x = self._cast_input_dtype(x, lora_A.weight.dtype)
+
+            if active_adapter not in self.lora_variant:  # vanilla LoRA
+                return lora_forward(result, lora_A, lora_B, dropout, x, scaling)
+                if requires_conversion:
+                    output = output.to(expected_dtype)
+                result = result + output
+            else:
+                result = self.lora_variant[active_adapter].forward(
+                    self,
+                    active_adapter=active_adapter,
+                    x=x,
+                    result=result,
+                    **variant_kwargs,
+                    **kwargs,
+                )
+                if requires_conversion:
+                    result = result.to(expected_dtype)
+
+    return result

unsloth_compiled_cache/Linear_peft_forward.py

@@ -0,0 +1,115 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+try:
+    from peft.tuners.lora.layer import VARIANT_KWARG_KEYS
+except ImportError:
+    VARIANT_KWARG_KEYS = ['alora_offsets']
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from peft.tuners.lora.variants import (Any, torch)
+
+
+torch_addmm = torch.addmm
+torch_add   = torch.add
+# @torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def lora_forward(result, lora_A, lora_B, dropout, x, scaling):
+    # Use result.dtype (bfloat16 from base layer) since x may have been cast to float32
+    # by _cast_input_dtype when autocast is disabled
+    target_dtype = result.dtype
+    xA = dropout(x).to(target_dtype) @ lora_A.weight.to(target_dtype).t()
+    # output = result + scaling * xA @ lora_B.weight.t()
+    shape = result.shape
+    output = torch_addmm(
+        result.view(-1, shape[-1]),
+        xA.view(-1, xA.shape[-1]),
+        lora_B.weight.to(target_dtype).t(),
+        alpha = scaling,
+        beta = 1,
+    ).view(shape)
+
+    bias = lora_B.bias
+    if bias is not None:
+        output = torch_add(
+            output,
+            bias.to(target_dtype),
+            alpha = scaling,
+        )
+    return output
+pass
+
+def unsloth_forward(self, x: torch.Tensor, *args: Any, **kwargs: Any) -> torch.Tensor:
+    
+    adapter_names = kwargs.pop("adapter_names", None)
+    variant_kwargs = {k: kwargs.pop(k, None) for k in VARIANT_KWARG_KEYS}  # don't pass these to base_layer
+
+    if self.disable_adapters:
+        if self.merged:
+            self.unmerge()
+        if not torch.is_autocast_enabled() and hasattr(self.base_layer, 'weight') and self.base_layer.weight is not None and not hasattr(self.base_layer.weight, 'quant_state') and x.dtype != self.base_layer.weight.dtype:
+            x = x.to(self.base_layer.weight.dtype)
+        result = self.base_layer(x, *args, **kwargs)
+    elif adapter_names is not None:
+        result = self._mixed_batch_forward(x, *args, adapter_names=adapter_names, **variant_kwargs, **kwargs)
+    elif self.merged:
+        if not torch.is_autocast_enabled() and hasattr(self.base_layer, 'weight') and self.base_layer.weight is not None and not hasattr(self.base_layer.weight, 'quant_state') and x.dtype != self.base_layer.weight.dtype:
+            x = x.to(self.base_layer.weight.dtype)
+        result = self.base_layer(x, *args, **kwargs)
+    else:
+        if not torch.is_autocast_enabled() and hasattr(self.base_layer, 'weight') and self.base_layer.weight is not None and not hasattr(self.base_layer.weight, 'quant_state') and x.dtype != self.base_layer.weight.dtype:
+            x = x.to(self.base_layer.weight.dtype)
+        result = self.base_layer(x, *args, **kwargs)
+        torch_result_dtype = result.dtype
+
+        lora_A_keys = self.lora_A.keys()
+        for active_adapter in self.active_adapters:
+            if active_adapter not in lora_A_keys:
+                continue
+
+            lora_A = self.lora_A[active_adapter]
+            lora_B = self.lora_B[active_adapter]
+            dropout = self.lora_dropout[active_adapter]
+            scaling = self.scaling[active_adapter]
+            if not torch.is_autocast_enabled(): result, x = result.to(lora_A.weight.dtype), x.to(lora_A.weight.dtype)
+            if active_adapter not in self.lora_variant:  # vanilla LoRA
+                return lora_forward(result, lora_A, lora_B, dropout, x, scaling)
+            else:
+                result = self.lora_variant[active_adapter].forward(
+                    self,
+                    active_adapter=active_adapter,
+                    x=x,
+                    result=result,
+                    **variant_kwargs,
+                    **kwargs,
+                )
+
+        result = result.to(torch_result_dtype)
+
+    return result

unsloth_compiled_cache/LoraParallelLinear_peft_forward.py

@@ -0,0 +1,92 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from peft.tuners.lora.tp_layer import (Any, __name__, torch)
+
+
+torch_addmm = torch.addmm
+torch_add   = torch.add
+# @torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def lora_forward(result, lora_A, lora_B, dropout, x, scaling):
+    # Use result.dtype (bfloat16 from base layer) since x may have been cast to float32
+    # by _cast_input_dtype when autocast is disabled
+    target_dtype = result.dtype
+    xA = dropout(x).to(target_dtype) @ lora_A.weight.to(target_dtype).t()
+    # output = result + scaling * xA @ lora_B.weight.t()
+    shape = result.shape
+    output = torch_addmm(
+        result.view(-1, shape[-1]),
+        xA.view(-1, xA.shape[-1]),
+        lora_B.weight.to(target_dtype).t(),
+        alpha = scaling,
+        beta = 1,
+    ).view(shape)
+
+    bias = lora_B.bias
+    if bias is not None:
+        output = torch_add(
+            output,
+            bias.to(target_dtype),
+            alpha = scaling,
+        )
+    return output
+pass
+
+def unsloth_forward(self, x: torch.Tensor, *args: Any, **kwargs: Any):
+    
+    adapter_names = kwargs.pop("adapter_names", None)
+    # If weight is used for matrix multiplication here, the final aggregation operation of the original
+    # parallel_linear layer will be missing, so we need to directly call its forward function to obtain the
+    # output of the original parallel_linear layer.
+    if self.disable_adapters:
+        if self.merged:
+            self.unmerge()
+        result, bias = self.base_layer(x, *args, **kwargs)
+    elif adapter_names is not None:
+        raise ValueError(f"{self.__class__.__name__} does not support mixed_batch_forward yet.")
+    elif self.merged:
+        result, bias = self.base_layer(x, *args, **kwargs)
+    else:
+        result, bias = self.base_layer(x, *args, **kwargs)
+        torch_result_dtype = result.dtype
+        for active_adapter in self.active_adapters:
+            if active_adapter not in self.lora_A.keys():
+                continue
+            lora_A = self.lora_A[active_adapter]
+            lora_B = self.lora_B[active_adapter]
+            dropout = self.lora_dropout[active_adapter]
+            scaling = self.scaling[active_adapter]
+            if not torch.is_autocast_enabled(): result, x = result.to(lora_A.weight.dtype), x.to(lora_A.weight.dtype)
+            return lora_forward(result, lora_A, lora_B, dropout, x, scaling)
+
+        result = result.to(torch_result_dtype)
+    return result, bias

unsloth_compiled_cache/RMSNorm.py

@@ -0,0 +1,73 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.gemma3.modeling_gemma3 import (torch)
+
+def forward(self, x: torch.Tensor) -> torch.Tensor:
+    """
+    Runs the forward pass.
+    """
+    return F.rms_norm(x, self.normalized_shape, self.weight, self.eps).to(input.dtype).to(input.dtype)

unsloth_compiled_cache/unsloth_compiled_module_gemma3.py

@@ -0,0 +1,984 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+
+from unsloth_zoo.loss_utils import (
+    fused_linear_cross_entropy,
+    unsloth_fused_ce_loss,
+)
+
+if UNSLOTH_STUDIO_ENABLED:
+    from unsloth_zoo.loss_utils import fast_linear_cross_entropy
+
+scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
+@torch.compiler.disable(recursive = False)
+def disable_compile_scaled_dot_product_attention(*args, **kwargs):
+    return scaled_dot_product_attention(*args, **kwargs)
+pass
+
+
+from transformers.modeling_flash_attention_utils import is_flash_attn_available
+
+if is_flash_attn_available():
+    try:
+        from transformers.modeling_flash_attention_utils import flash_attn_supports_top_left_mask
+    except:
+        flash_attn_supports_top_left_mask = None
+    try:
+        from transformers.modeling_flash_attention_utils import _flash_attention_forward
+    except:
+        _flash_attention_forward = None
+    try:
+        from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+    except:
+        FlashAttentionKwargs = None
+    try:
+        from transformers.modeling_flash_attention_utils import flash_attn_varlen_func
+    except:
+        flash_attn_varlen_func = None
+else:
+    flash_attn_supports_top_left_mask = None
+    _flash_attention_forward = None
+    FlashAttentionKwargs = None
+    flash_attn_varlen_func = None
+pass
+
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+
+from torch.nn import CrossEntropyLoss
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def normal_cross_entropy_loss(self, hidden_states, labels):
+    logits = self.lm_head(hidden_states)
+    logits = logits.float()
+    # Shift so that tokens < n predict n
+    shift_logits = logits[..., :-1, :].contiguous()
+    shift_labels = labels[..., 1:].contiguous()
+    # Flatten the tokens
+    loss_fct = CrossEntropyLoss()
+    shift_logits = shift_logits.view(-1, self.config.vocab_size)
+    shift_labels = shift_labels.view(-1)
+    # Enable model parallelism
+    shift_labels = shift_labels.to(shift_logits.device)
+    loss = loss_fct(shift_logits, shift_labels)
+    return loss, logits
+pass
+
+# We need an empty logits flag to warn people logits will not be returned anymore unless asked ie
+# os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
+LOGITS_ERROR_STRING = \
+    "Unsloth: Logits are empty from 2024.11 onwards. To get raw logits again, please "\
+    'set the environment variable `UNSLOTH_RETURN_LOGITS` to `"1" BEFORE starting to train ie before `trainer.train()`. For example:\n'\
+    "```\nimport os\n"\
+    "os.environ['UNSLOTH_RETURN_LOGITS'] = '1'\n"\
+    "trainer.train()\n```\n"\
+    "No need to restart your console - just add `os.environ['UNSLOTH_RETURN_LOGITS'] = '1'` before trainer.train() and re-run the cell!"
+
+def raise_logits_error(*args, **kwargs): raise NotImplementedError(LOGITS_ERROR_STRING)
+def return_none(*args, **kwargs): return None
+class EmptyLogits:
+    def __init__(self): return
+    def raise_getattr_error(self, attr): return return_none if attr == "to" else raise_logits_error
+    __getitem__ = raise_logits_error
+    __getattr__ = raise_getattr_error
+    def __repr__(self): return LOGITS_ERROR_STRING
+    def __str__ (self): return LOGITS_ERROR_STRING
+pass
+EMPTY_LOGITS = EmptyLogits()
+functions = dir(torch.Tensor)
+for j, function in enumerate(functions):
+    if function.startswith("__") and function.endswith("__"):
+        exec(f"def raise_{j}(*args, **kwargs): print('{function}')", globals(), locals())
+        try: exec(f"EMPTY_LOGITS.{function} = raise_{j}", globals(), locals())
+        except: continue
+pass
+
+
+def mask_attention_mask_out(labels = None, attention_mask = None):
+    if labels is not None and attention_mask is not None:
+        attention_mask = attention_mask.to(device = labels.device)
+        labels[attention_mask == 0] = -100
+    return labels
+pass
+
+
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.gemma3.modeling_gemma3 import (Callable, Optional, Union, torch, nn, ACT2FN, Cache, PretrainedConfig, GenerationMixin, BaseModelOutputWithPast, ModelOutput, CausalLMOutputWithPast, ROPE_INIT_FUNCTIONS, dynamic_rope_update, PreTrainedModel, can_return_tuple, Gemma3Config, Gemma3TextConfig, logger, __name__, Gemma3Model, Gemma3CausalLMOutputWithPast, Gemma3PreTrainedModel, Gemma3TextModel, Gemma3ForCausalLM, Gemma3ForConditionalGeneration, create_masks_for_generate)
+
+@torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def Gemma3MLP_forward(self, x):
+    down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+    return down_proj
+
+class Gemma3MLP(nn.Module):
+    def __init__(self, config: Gemma3TextConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_activation]
+
+    def forward(self, x):
+        return Gemma3MLP_forward(self, x)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def Gemma3RMSNorm_forward(self, x):
+    x_fp32 = x.to(torch.float32)
+    variance = x_fp32.pow(2).mean(-1, keepdim=True)
+    hidden_states_fp32 = x_fp32 * torch.rsqrt(variance + self.eps)
+    output_fp32 = hidden_states_fp32 * (1.0 + self.weight.to(torch.float32))
+    return output_fp32.to(x.dtype)
+
+class Gemma3RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.zeros(dim))
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        output = self._norm(x.float())
+        # Llama does x.to(float16) * w whilst Gemma3 is (x * w).to(float16)
+        # See https://github.com/huggingface/transformers/pull/29402
+        output = output * (1.0 + self.weight.float())
+        return output.type_as(x)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+@torch.no_grad()
+@dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+def Gemma3RotaryEmbedding_forward(self, x, position_ids):
+    inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+    position_ids_expanded = position_ids[:, None, :].float()
+
+    device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+    with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        cos = emb.cos() * self.attention_scaling
+        sin = emb.sin() * self.attention_scaling
+
+    return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+class Gemma3RotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Gemma3TextConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+
+    def forward(self, x, position_ids):
+        return Gemma3RotaryEmbedding_forward(self, x, position_ids)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def apply_rotary_pos_emb(q, k, cos, sin,  unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    dropout: float = 0.0,
+    scaling: Optional[float] = None,
+    softcap: Optional[float] = None,
+    **kwargs,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    if scaling is None:
+        scaling = module.head_dim**-0.5
+
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+
+    if softcap is not None:
+        attn_weights = attn_weights / softcap
+        attn_weights = torch.tanh(attn_weights)
+        attn_weights = attn_weights * softcap
+    if attention_mask is not None:  # no matter the length, we just slice it
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    # upcast attention to fp32
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype = torch.float32).to(attn_weights.dtype).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    return attn_output, attn_weights
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def Gemma3MultiModalProjector_forward(self, vision_outputs: torch.Tensor):
+    batch_size, _, seq_length = vision_outputs.shape
+
+    reshaped_vision_outputs = vision_outputs.transpose(1, 2)
+    reshaped_vision_outputs = reshaped_vision_outputs.reshape(
+        batch_size, seq_length, self.patches_per_image, self.patches_per_image
+    )
+    reshaped_vision_outputs = reshaped_vision_outputs.contiguous()
+
+    pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs)
+    pooled_vision_outputs = pooled_vision_outputs.flatten(2)
+    pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2)
+
+    normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs)
+
+    projected_vision_outputs = torch.matmul(normed_vision_outputs, self.mm_input_projection_weight)
+    return projected_vision_outputs.type_as(vision_outputs)
+
+class Gemma3MultiModalProjector(nn.Module):
+    def __init__(self, config: Gemma3Config):
+        super().__init__()
+
+        self.mm_input_projection_weight = nn.Parameter(
+            torch.zeros(config.vision_config.hidden_size, config.text_config.hidden_size)
+        )
+
+        self.mm_soft_emb_norm = Gemma3RMSNorm(
+            config.vision_config.hidden_size, eps=config.vision_config.layer_norm_eps
+        )
+
+        self.patches_per_image = int(config.vision_config.image_size // config.vision_config.patch_size)
+        self.tokens_per_side = int(config.mm_tokens_per_image**0.5)
+        self.kernel_size = self.patches_per_image // self.tokens_per_side
+        self.avg_pool = nn.AvgPool2d(kernel_size=self.kernel_size, stride=self.kernel_size)
+
+    def forward(self, vision_outputs: torch.Tensor):
+        return Gemma3MultiModalProjector_forward(self, vision_outputs)
+
+
+def _bidirectional_window_overlay(sliding_window: int) -> Callable[[int, int, int, int], bool]:
+    """
+    Enables a bidirectional mask within the sliding window.
+    """
+
+    def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
+        """A token can attend to any other token if their absolute distance is within
+        the (exclusive) sliding window size (distance < sliding_window)."""
+        return abs(q_idx - kv_idx) < sliding_window
+
+    return inner_mask
+
+
+@torch.compiler.disable(recursive = False)
+@can_return_tuple
+def Gemma3ForCausalLM_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Cache] = 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,
+    **kwargs,
+) -> CausalLMOutputWithPast:
+    r"""
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Gemma3ForCausalLM
+
+    >>> model = Gemma3ForCausalLM.from_pretrained("google/gemma-2-9b")
+    >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")
+
+    >>> prompt = "What is your favorite condiment?"
+    >>> 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]
+    "What is your favorite condiment?"
+    ```"""
+
+    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: BaseModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    # 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
+    logits = self.lm_head(hidden_states[:, slice_indices, :]) if os.environ.get('UNSLOTH_RETURN_LOGITS', '0') == '1' else EMPTY_LOGITS
+    loss = None
+    NOT_RETURN_LOGITS = os.environ.get('UNSLOTH_RETURN_LOGITS', '0') == '0'
+    RETURN_HIDDEN_STATES = os.environ.get("UNSLOTH_RETURN_HIDDEN_STATES", "0") == "1"
+    
+    n_items = None
+    if (kwargs) != () and type(kwargs) is dict:
+        n_items = (kwargs).get("num_items_in_batch", None)
+        if n_items is None: n_items = (kwargs).get("n_items", None)
+    if n_items is None:
+        all_locals = locals()
+        if 'loss_kwargs' in all_locals:
+            __kwargs = all_locals['loss_kwargs']
+            if type(__kwargs) is dict:
+                n_items = __kwargs.get("num_items_in_batch", None)
+                if n_items is None: n_items = __kwargs.get("n_items", None)
+        if n_items is None and 'kwargs' in all_locals:
+            __kwargs = all_locals['kwargs']
+            if type(__kwargs) is dict:
+                n_items = __kwargs.get("num_items_in_batch", None)
+                if n_items is None: n_items = __kwargs.get("n_items", None)
+        if n_items is None:
+            all_locals = all_locals.values()
+            for __kwargs in all_locals:
+                if type(__kwargs) is dict:
+                    n_items = __kwargs.get("num_items_in_batch", None)
+                    if n_items is None: n_items = __kwargs.get("n_items", None)
+                    break
+    pass
+    
+    requires_grad_ = self.lm_head.weight.requires_grad
+    requires_grad_ = requires_grad_ or self.lm_head.weight.dtype == torch.float32
+    
+    if RETURN_HIDDEN_STATES:
+        logits = hidden_states[:, slice_indices, :]
+    elif labels is None:
+        
+    
+        # Set compiler stance to fail on recompiles for inference
+        global INFERENCE_RUNS
+        if torch_dynamo_eval_frame is not None:
+            old_stance = torch_dynamo_eval_frame._stance.stance
+        else:
+            old_stance = None
+        if old_stance is not None and INFERENCE_RUNS == 1:
+            # Skip guards and return to eager -> we still need guards!
+            torch_compiler_set_stance(stance = "eager_on_recompile", skip_guard_eval_unsafe = False)
+            if UNSLOTH_ENABLE_LOGGING:
+                logger_compiler.info(
+                    f"Unsloth: Removing compiler guards after 1 inference run. " \
+                    f"DYNAMO_STANCE.stance = {torch_dynamo_eval_frame._stance.stance} " \
+                    f"DYNAMO_STANCE.skip_guard_eval_unsafe = {torch_dynamo_eval_frame._stance.skip_guard_eval_unsafe}"
+                )
+        elif old_stance == "eager_on_recompile":
+            pass
+        elif old_stance == "default" and INFERENCE_RUNS > 1:
+            # Reset compiler stance
+            torch_compiler_set_stance(stance = "default", skip_guard_eval_unsafe = False)
+            if UNSLOTH_ENABLE_LOGGING:
+                logger_compiler.info(
+                    f"Unsloth: Reseting guards. " \
+                    f"DYNAMO_STANCE.stance = {torch_dynamo_eval_frame._stance.stance} " \
+                    f"DYNAMO_STANCE.skip_guard_eval_unsafe = {torch_dynamo_eval_frame._stance.skip_guard_eval_unsafe}"
+                )
+            INFERENCE_RUNS = 0
+        INFERENCE_RUNS += 1
+    
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+    elif (() == () and () == ()) and (UNSLOTH_ENABLE_CCE) and NOT_RETURN_LOGITS and self.loss_function.__name__.endswith("ForCausalLMLoss") and labels is not None and not requires_grad_:
+        loss = fused_linear_cross_entropy(
+            hidden_states      = hidden_states[:, slice_indices, :],
+            lm_weight          = self.lm_head.weight,
+            labels             = labels.to(self.lm_head.weight.device),
+            num_items_in_batch = n_items,
+            logit_softcapping  = None if (self.config.final_logit_softcapping) == () else (self.config.final_logit_softcapping),
+        )
+    elif self.loss_function.__name__.endswith("ForCausalLMLoss") and labels is not None:
+        lm_head_weight = self.lm_head.weight
+        lm_head_bias   = getattr(self.lm_head, "bias", None)
+    
+        # ========= NEW fused =========
+        _hidden_states = hidden_states[:, slice_indices, :]
+        torch._dynamo.mark_dynamic(_hidden_states, 1)
+        torch._dynamo.mark_dynamic(labels, 1)
+        loss = unsloth_fused_ce_loss(
+            trainer              = None,
+            hidden_states        = _hidden_states,
+            lm_head_weight       = lm_head_weight,
+            lm_head_bias         = lm_head_bias,
+            labels               = labels,
+            mask                 = None,
+            n_items              = n_items,
+            scaling              = getattr(self, "accelerator_scaler", None),
+            target_gb            = None,
+            torch_compile        = not UNSLOTH_COMPILE_DISABLE,
+            logit_scale_multiply = () if () != () else 0,
+            logit_scale_divide   = () if () != () else 0,
+            logit_softcapping    = (self.config.final_logit_softcapping) if (self.config.final_logit_softcapping) != () else 0,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        if () != ():
+            logits = logits * ()
+        if () != ():
+            logits = logits / ()
+        if (self.config.final_logit_softcapping) not in (None, (),):
+            logits = logits / (self.config.final_logit_softcapping)
+            logits = torch.tanh(logits)
+            logits = logits * (self.config.final_logit_softcapping)
+        loss = self.loss_function(logits, labels.to(self.lm_head.weight.device), vocab_size=self.vocab_size, **kwargs)
+
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+class Gemma3ForCausalLM(Gemma3PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+    config: Gemma3TextConfig
+    base_model_prefix = "language_model"
+
+    def __init__(self, config: Gemma3TextConfig):
+        super().__init__(config)
+        self.model = Gemma3TextModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = 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,
+        **kwargs,
+    ) -> CausalLMOutputWithPast:
+        return Gemma3ForCausalLM_forward(self, input_ids, attention_mask, position_ids, past_key_values, inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, cache_position, logits_to_keep, **kwargs)
+
+
+def token_type_ids_mask_function(
+    token_type_ids: Optional[torch.Tensor],
+    image_group_ids: Optional[torch.Tensor],
+    tokens_per_image: int,
+) -> Optional[Callable]:
+    """
+    This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths,
+    not start and end indices.
+    """
+    # Do not return an additional mask in this case
+    if token_type_ids is None:
+        return None
+
+    def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
+        # If it's 1 for both query and key/value, we are in an image block
+        # NOTE: static cache shape goes beyond input seq length, while token_type_ids.shape[1] == input seq length
+        # Since vmap doesn't support `if statement` we workaround it with `torch.where`
+        safe_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0)
+        token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_idx]
+        token_type_ids_at_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], token_type_ids_at_kv_idx, 0)
+
+        image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_idx]
+        image_group_ids_at_kv_idx = torch.where(kv_idx < image_group_ids.shape[1], image_group_ids_at_kv_idx, -1)
+
+        is_image_block = (token_type_ids[batch_idx, q_idx] == 1) & (token_type_ids_at_kv_idx == 1)
+        same_image_block = image_group_ids[batch_idx, q_idx] == image_group_ids_at_kv_idx
+
+        # This is bidirectional attention whenever we are dealing with image tokens
+        return is_image_block & same_image_block
+
+    return inner_mask
+
+
+@torch.compiler.disable(recursive = False)
+def Gemma3ForConditionalGeneration_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    pixel_values: Optional[torch.FloatTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Cache] = None,
+    token_type_ids: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = 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,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **lm_kwargs,
+) -> Union[tuple, Gemma3CausalLMOutputWithPast]:
+    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.text_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.text_config.vocab_size]`.
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Gemma3ForConditionalGeneration
+
+    >>> model = Gemma3ForConditionalGeneration.from_pretrained("google/gemma-3-4b-it")
+    >>> processor = AutoProcessor.from_pretrained("google/gemma-3-4b-it")
+
+    >>> messages = [
+    ...     {
+    ...         "role": "system",
+    ...         "content": [
+    ...             {"type": "text", "text": "You are a helpful assistant."}
+    ...         ]
+    ...     },
+    ...     {
+    ...         "role": "user", "content": [
+    ...             {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"},
+    ...             {"type": "text", "text": "Where is the cat standing?"},
+    ...         ]
+    ...     },
+    ... ]
+
+    >>> inputs = processor.apply_chat_template(
+    ...     messages,
+    ...     tokenize=True,
+    ...     return_dict=True,
+    ...     return_tensors="pt",
+    ...     add_generation_prompt=True
+    ... )
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "user\nYou are a helpful assistant.\n\n\n\n\n\nWhere is the cat standing?\nmodel\nBased on the image, the cat is standing in a snowy area, likely outdoors. It appears to"
+    ```
+    """
+
+    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
+
+    outputs = self.model(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        token_type_ids=token_type_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        labels=mask_attention_mask_out(labels = labels, attention_mask = attention_mask),
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **lm_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
+    logits = self.lm_head(hidden_states[:, slice_indices, :]) if os.environ.get('UNSLOTH_RETURN_LOGITS', '0') == '1' else EMPTY_LOGITS
+    loss = None
+    NOT_RETURN_LOGITS = os.environ.get('UNSLOTH_RETURN_LOGITS', '0') == '0'
+    RETURN_HIDDEN_STATES = os.environ.get("UNSLOTH_RETURN_HIDDEN_STATES", "0") == "1"
+    
+    all_locals = locals()
+    n_items = None
+    if 'loss_kwargs' in all_locals:
+        __kwargs = all_locals['loss_kwargs']
+        if type(__kwargs) is dict:
+            n_items = __kwargs.get("num_items_in_batch", None)
+            if n_items is None: n_items = __kwargs.get("n_items", None)
+    if n_items is None and 'kwargs' in all_locals:
+        __kwargs = all_locals['kwargs']
+        if type(__kwargs) is dict:
+            n_items = __kwargs.get("num_items_in_batch", None)
+            if n_items is None: n_items = __kwargs.get("n_items", None)
+    if n_items is None:
+        all_locals = all_locals.values()
+        for __kwargs in all_locals:
+            if type(__kwargs) is dict:
+                n_items = __kwargs.get("num_items_in_batch", None)
+                if n_items is None: n_items = __kwargs.get("n_items", None)
+                break
+    pass
+    
+    requires_grad_ = self.lm_head.weight.requires_grad
+    requires_grad_ = requires_grad_ or self.lm_head.weight.dtype == torch.float32
+    
+    if RETURN_HIDDEN_STATES:
+        logits = hidden_states[:, slice_indices, :]
+    elif labels is None:
+        
+    
+        # Set compiler stance to fail on recompiles for inference
+        global INFERENCE_RUNS
+        if torch_dynamo_eval_frame is not None:
+            old_stance = torch_dynamo_eval_frame._stance.stance
+        else:
+            old_stance = None
+        if old_stance is not None and INFERENCE_RUNS == 1:
+            # Skip guards and return to eager -> we still need guards!
+            torch_compiler_set_stance(stance = "eager_on_recompile", skip_guard_eval_unsafe = False)
+            if UNSLOTH_ENABLE_LOGGING:
+                logger_compiler.info(
+                    f"Unsloth: Removing compiler guards after 1 inference run. " \
+                    f"DYNAMO_STANCE.stance = {torch_dynamo_eval_frame._stance.stance} " \
+                    f"DYNAMO_STANCE.skip_guard_eval_unsafe = {torch_dynamo_eval_frame._stance.skip_guard_eval_unsafe}"
+                )
+        elif old_stance == "eager_on_recompile":
+            pass
+        elif old_stance == "default" and INFERENCE_RUNS > 1:
+            # Reset compiler stance
+            torch_compiler_set_stance(stance = "default", skip_guard_eval_unsafe = False)
+            if UNSLOTH_ENABLE_LOGGING:
+                logger_compiler.info(
+                    f"Unsloth: Reseting guards. " \
+                    f"DYNAMO_STANCE.stance = {torch_dynamo_eval_frame._stance.stance} " \
+                    f"DYNAMO_STANCE.skip_guard_eval_unsafe = {torch_dynamo_eval_frame._stance.skip_guard_eval_unsafe}"
+                )
+            INFERENCE_RUNS = 0
+        INFERENCE_RUNS += 1
+    
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+    else:
+        lm_head_weight = self.lm_head.weight
+        lm_head_bias   = getattr(self.lm_head, "bias", None)
+    
+        # ========= NEW fused =========
+        _hidden_states = hidden_states[:, slice_indices, :]
+        torch._dynamo.mark_dynamic(_hidden_states, 1)
+        torch._dynamo.mark_dynamic(labels, 1)
+        if attention_mask is not None:
+            torch._dynamo.mark_dynamic(attention_mask, 1)
+        loss = unsloth_fused_ce_loss(
+            trainer              = None,
+            hidden_states        = _hidden_states,
+            lm_head_weight       = lm_head_weight,
+            lm_head_bias         = lm_head_bias,
+            labels               = labels,
+            mask                 = attention_mask,
+            n_items              = n_items,
+            scaling              = getattr(self, "accelerator_scaler", None),
+            target_gb            = None,
+            torch_compile        = not UNSLOTH_COMPILE_DISABLE,
+            logit_scale_multiply = () if () != () else 0,
+            logit_scale_divide   = () if () != () else 0,
+            logit_softcapping    = () if () != () else 0,
+        )
+
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Gemma3CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=outputs.image_hidden_states,
+    )
+
+
+class Gemma3ForConditionalGeneration(Gemma3PreTrainedModel, GenerationMixin):
+    _checkpoint_conversion_mapping = {
+        "^language_model.model": "model.language_model",
+        "^vision_tower": "model.vision_tower",
+        "^multi_modal_projector": "model.multi_modal_projector",
+        "^language_model.lm_head": "lm_head",
+    }
+    _tied_weights_keys = ["lm_head.weight"]
+    # we are filtering the logits/labels so we shouldn't divide the loss based on num_items_in_batch
+    # Fix: https://github.com/huggingface/transformers/issues/40564
+    accepts_loss_kwargs = False
+
+    def __init__(self, config: Gemma3Config):
+        super().__init__(config)
+        self.model = Gemma3Model(config)
+        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.model.set_input_embeddings(value)
+
+    def set_decoder(self, decoder):
+        self.model.set_decoder(decoder)
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def get_image_features(self, pixel_values):
+        return self.model.get_image_features(pixel_values)
+
+    # Make modules available through conditional class for BC
+    @property
+    def language_model(self):
+        return self.model.language_model
+
+    @property
+    def vision_tower(self):
+        return self.model.vision_tower
+
+    @property
+    def multi_modal_projector(self):
+        return self.model.multi_modal_projector
+
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = 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,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **lm_kwargs,
+    ) -> Union[tuple, Gemma3CausalLMOutputWithPast]:
+        return Gemma3ForConditionalGeneration_forward(self, input_ids, pixel_values, attention_mask, position_ids, past_key_values, token_type_ids, cache_position, inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, return_dict, logits_to_keep, **lm_kwargs)
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        pixel_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        use_cache=True,
+        logits_to_keep=None,
+        labels=None,
+        **kwargs,
+    ):
+        # Overwritten -- custom `position_ids` and `pixel_values` handling
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            cache_position=cache_position,
+            use_cache=use_cache,
+            logits_to_keep=logits_to_keep,
+            token_type_ids=token_type_ids,
+            **kwargs,
+        )
+
+        # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+        # Otherwise we need pixel values to be passed to model. NOTE: use_cache=False needs pixel_values always
+        if cache_position[0] == 0:
+            model_inputs["pixel_values"] = pixel_values
+
+        return model_inputs
+
+    @staticmethod
+    def create_masks_for_generate(
+        config: PretrainedConfig,
+        input_embeds: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        cache_position: torch.Tensor,
+        past_key_values: Optional[Cache],
+        position_ids: Optional[torch.Tensor],
+        token_type_ids: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> dict:
+        # Prepare mask arguments
+        mask_kwargs = {
+            "config": config.get_text_config(),
+            "input_embeds": input_embeds,
+            "attention_mask": attention_mask,
+            "cache_position": cache_position,
+            "past_key_values": past_key_values,
+            "position_ids": position_ids,
+        }
+        # Add the token type ids mask for generate as well
+        if token_type_ids is not None and input_embeds.shape[1] != 1:
+            # We need to pass an additional mask function to account for token type ids, and it needs to be an `or`
+
+            # First find where a new image block starts: 1 if image and previous not image
+            # The images cannot attend to future images, but can attend to all prev images and to itself bidirectionally
+            is_image = (token_type_ids == 1).to(cache_position.device)
+            new_image_start = is_image & ~nn.functional.pad(is_image, (1, 0), value=0)[:, :-1]
+            image_group_ids = torch.cumsum(new_image_start.int(), dim=1) - 1
+            image_group_ids = torch.where(is_image, image_group_ids, torch.full_like(token_type_ids, -1))
+            mask_kwargs["or_mask_function"] = token_type_ids_mask_function(
+                token_type_ids.to(cache_position.device), image_group_ids, config.mm_tokens_per_image
+            )
+
+        return create_masks_for_generate(**mask_kwargs)
+
+
+if hasattr(logger, "addFilter"):
+    import logging
+    class HideLoggingMessage(logging.Filter):
+        def __init__(self, text): self.text = text
+        def filter(self, x): return not (self.text in x.getMessage())
+    pass
+    logger.addFilter(HideLoggingMessage("`use_cache=True`"))
+

unsloth_compiled_cache/unsloth_compiled_module_siglip.py

@@ -0,0 +1,535 @@
+"""
+2026.2.1
+2026.2.1
+4.57.6
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+
+import os
+import torch
+import importlib.util
+import math
+if importlib.util.find_spec("unsloth_studio") is None:
+    UNSLOTH_STUDIO_ENABLED = False
+else:
+    UNSLOTH_STUDIO_ENABLED = os.environ.get("UNSLOTH_STUDIO_DISABLED", "0") == "0"
+pass
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+import math
+
+UNSLOTH_ENABLE_LOGGING = os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1"
+UNSLOTH_ENABLE_CCE = os.environ.get("UNSLOTH_ENABLE_CCE", "1") == "1"
+UNSLOTH_COMPILE_DISABLE = os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") in ("1", "partial",)
+
+import logging
+logger_compiler = logging.getLogger(__name__)
+if UNSLOTH_ENABLE_LOGGING:
+    logger_compiler.setLevel(logging.DEBUG)
+
+global INFERENCE_RUNS
+INFERENCE_RUNS = 0
+
+try:
+    import torch._dynamo.eval_frame as torch_dynamo_eval_frame
+    torch_dynamo_eval_frame._stance.stance
+    torch_compiler_set_stance = torch.compiler.set_stance
+except:
+    torch_dynamo_eval_frame = None
+    torch_compiler_set_stance = None
+pass
+
+from unsloth_zoo import DEVICE_TYPE_TORCH, DEVICE_COUNT
+
+
+from unsloth_zoo.loss_utils import (
+    fused_linear_cross_entropy,
+    unsloth_fused_ce_loss,
+)
+
+if UNSLOTH_STUDIO_ENABLED:
+    from unsloth_zoo.loss_utils import fast_linear_cross_entropy
+
+scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
+@torch.compiler.disable(recursive = False)
+def disable_compile_scaled_dot_product_attention(*args, **kwargs):
+    return scaled_dot_product_attention(*args, **kwargs)
+pass
+
+
+from transformers.modeling_flash_attention_utils import is_flash_attn_available
+
+if is_flash_attn_available():
+    try:
+        from transformers.modeling_flash_attention_utils import flash_attn_supports_top_left_mask
+    except:
+        flash_attn_supports_top_left_mask = None
+    try:
+        from transformers.modeling_flash_attention_utils import _flash_attention_forward
+    except:
+        _flash_attention_forward = None
+    try:
+        from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+    except:
+        FlashAttentionKwargs = None
+    try:
+        from transformers.modeling_flash_attention_utils import flash_attn_varlen_func
+    except:
+        flash_attn_varlen_func = None
+else:
+    flash_attn_supports_top_left_mask = None
+    _flash_attention_forward = None
+    FlashAttentionKwargs = None
+    flash_attn_varlen_func = None
+pass
+
+
+torch_compile_options = {'epilogue_fusion': True, 'max_autotune': False, 'shape_padding': True, 'trace.enabled': False, 'triton.cudagraphs': False, 'debug': False, 'dce': True, 'memory_planning': True, 'coordinate_descent_tuning': False, 'trace.graph_diagram': False, 'compile_threads': 32, 'group_fusion': True, 'disable_progress': True, 'verbose_progress': False, 'triton.multi_kernel': 0, 'triton.use_block_ptr': False, 'triton.enable_persistent_tma_matmul': True, 'triton.autotune_at_compile_time': False, 'triton.cooperative_reductions': False, 'cuda.compile_opt_level': '-O2', 'cuda.enable_cuda_lto': True, 'combo_kernels': False, 'benchmark_combo_kernel': True, 'combo_kernel_foreach_dynamic_shapes': True}
+
+from torch.nn import CrossEntropyLoss
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def normal_cross_entropy_loss(self, hidden_states, labels):
+    logits = self.lm_head(hidden_states)
+    logits = logits.float()
+    # Shift so that tokens < n predict n
+    shift_logits = logits[..., :-1, :].contiguous()
+    shift_labels = labels[..., 1:].contiguous()
+    # Flatten the tokens
+    loss_fct = CrossEntropyLoss()
+    shift_logits = shift_logits.view(-1, self.config.vocab_size)
+    shift_labels = shift_labels.view(-1)
+    # Enable model parallelism
+    shift_labels = shift_labels.to(shift_logits.device)
+    loss = loss_fct(shift_logits, shift_labels)
+    return loss, logits
+pass
+
+# We need an empty logits flag to warn people logits will not be returned anymore unless asked ie
+# os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
+LOGITS_ERROR_STRING = \
+    "Unsloth: Logits are empty from 2024.11 onwards. To get raw logits again, please "\
+    'set the environment variable `UNSLOTH_RETURN_LOGITS` to `"1" BEFORE starting to train ie before `trainer.train()`. For example:\n'\
+    "```\nimport os\n"\
+    "os.environ['UNSLOTH_RETURN_LOGITS'] = '1'\n"\
+    "trainer.train()\n```\n"\
+    "No need to restart your console - just add `os.environ['UNSLOTH_RETURN_LOGITS'] = '1'` before trainer.train() and re-run the cell!"
+
+def raise_logits_error(*args, **kwargs): raise NotImplementedError(LOGITS_ERROR_STRING)
+def return_none(*args, **kwargs): return None
+class EmptyLogits:
+    def __init__(self): return
+    def raise_getattr_error(self, attr): return return_none if attr == "to" else raise_logits_error
+    __getitem__ = raise_logits_error
+    __getattr__ = raise_getattr_error
+    def __repr__(self): return LOGITS_ERROR_STRING
+    def __str__ (self): return LOGITS_ERROR_STRING
+pass
+EMPTY_LOGITS = EmptyLogits()
+functions = dir(torch.Tensor)
+for j, function in enumerate(functions):
+    if function.startswith("__") and function.endswith("__"):
+        exec(f"def raise_{j}(*args, **kwargs): print('{function}')", globals(), locals())
+        try: exec(f"EMPTY_LOGITS.{function} = raise_{j}", globals(), locals())
+        except: continue
+pass
+
+
+def mask_attention_mask_out(labels = None, attention_mask = None):
+    if labels is not None and attention_mask is not None:
+        attention_mask = attention_mask.to(device = labels.device)
+        labels[attention_mask == 0] = -100
+    return labels
+pass
+
+
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from unsloth_zoo.temporary_patches.common import torch_compile
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from transformers.models.siglip.modeling_siglip import (math, warnings, Callable, Optional, np, torch, nn, _calculate_fan_in_and_fan_out, ACT2FN, ALL_ATTENTION_FUNCTIONS, torch_int, SiglipTextConfig, SiglipVisionConfig)
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    tensor.clamp_(min=a, max=b)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+) -> torch.Tensor:
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsequently scaled and shifted by the mean and std args.
+
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+
+@torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def SiglipVisionEmbeddings_forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
+    _, _, height, width = pixel_values.shape
+    target_dtype = self.patch_embedding.weight.dtype
+    patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
+    embeddings = patch_embeds.flatten(2).transpose(1, 2)
+
+    if interpolate_pos_encoding:
+        embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
+    else:
+        embeddings = embeddings + self.position_embedding(self.position_ids)
+    return embeddings
+
+class SiglipVisionEmbeddings(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
+
+    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        """
+        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
+        images. This method is also adapted to support torch.jit tracing and no class embeddings.
+
+        Adapted from:
+        - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
+        - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
+        """
+
+        num_patches = embeddings.shape[1]
+        num_positions = self.position_embedding.weight.shape[0]
+
+        # always interpolate when tracing to ensure the exported model works for dynamic input shapes
+        if not torch.jit.is_tracing() and num_patches == num_positions and height == width:
+            return self.position_embedding(self.position_ids)
+
+        patch_pos_embed = self.position_embedding.weight.unsqueeze(0)
+
+        dim = embeddings.shape[-1]
+
+        new_height = height // self.patch_size
+        new_width = width // self.patch_size
+
+        sqrt_num_positions = torch_int(num_positions**0.5)
+        patch_pos_embed = patch_pos_embed.reshape(1, sqrt_num_positions, sqrt_num_positions, dim)
+        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
+
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed,
+            size=(new_height, new_width),
+            mode="bicubic",
+            align_corners=False,
+        )
+
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return patch_pos_embed
+
+    def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
+        return SiglipVisionEmbeddings_forward(self, pixel_values, interpolate_pos_encoding)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def SiglipTextEmbeddings_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+) -> torch.Tensor:
+    seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
+    max_position_embedding = self.position_embedding.weight.shape[0]
+
+    if seq_length > max_position_embedding:
+        raise ValueError(
+            f"Sequence length must be less than max_position_embeddings (got `sequence length`: "
+            f"{seq_length} and max_position_embeddings: {max_position_embedding}"
+        )
+
+    if position_ids is None:
+        position_ids = self.position_ids[:, :seq_length]
+
+    if inputs_embeds is None:
+        inputs_embeds = self.token_embedding(input_ids)
+
+    position_embeddings = self.position_embedding(position_ids)
+    embeddings = inputs_embeds + position_embeddings
+
+    return embeddings
+
+class SiglipTextEmbeddings(nn.Module):
+    def __init__(self, config: SiglipTextConfig):
+        super().__init__()
+        embed_dim = config.hidden_size
+
+        self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
+        self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
+        )
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+    ) -> torch.Tensor:
+        return SiglipTextEmbeddings_forward(self, input_ids, position_ids, inputs_embeds)
+
+
+@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    attn_weights = torch.matmul(query, key.transpose(-1, -2)) * scaling
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype = torch.float32).to(attn_weights.dtype).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+
+    attn_output = torch.matmul(attn_weights, value)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+@torch.compiler.disable(recursive = False)
+def SiglipAttention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    **kwargs,
+) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+    """Input shape: Batch x Time x Channel"""
+
+    batch_size, seq_length, embed_dim = hidden_states.shape
+
+    queries = self.q_proj(hidden_states)
+    keys = self.k_proj(hidden_states)
+    values = self.v_proj(hidden_states)
+
+    queries = queries.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
+    keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
+    values = values.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
+
+    attention_interface: Callable = eager_attention_forward
+    if self.config._attn_implementation != "eager":
+        attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+    attn_output, attn_weights = attention_interface(
+        self,
+        queries,
+        keys,
+        values,
+        attention_mask,
+        is_causal=self.is_causal,
+        scaling=self.scale,
+        dropout=0.0 if not self.training else self.dropout,
+    )
+
+    attn_output = attn_output.reshape(batch_size, seq_length, embed_dim).contiguous()
+    attn_output = self.out_proj(attn_output)
+
+    return attn_output, attn_weights
+
+class SiglipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+        self.is_causal = False
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+        return SiglipAttention_forward(self, hidden_states, attention_mask, **kwargs)
+
+
+@torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def SiglipMLP_forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+    hidden_states = self.fc1(hidden_states)
+    hidden_states = self.activation_fn(hidden_states)
+    hidden_states = self.fc2(hidden_states)
+    return hidden_states
+
+class SiglipMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        return SiglipMLP_forward(self, hidden_states)
+
+
+@torch.compile(fullgraph = False, dynamic = True, options = torch_compile_options)
+def SiglipMultiheadAttentionPoolingHead_forward(self, hidden_state):
+    batch_size = hidden_state.shape[0]
+    probe = self.probe.repeat(batch_size, 1, 1)
+
+    hidden_state = self.attention(probe, hidden_state, hidden_state)[0]
+
+    residual = hidden_state
+    hidden_state = self.layernorm(hidden_state)
+    hidden_state = residual + self.mlp(hidden_state)
+
+    return hidden_state[:, 0]
+
+class SiglipMultiheadAttentionPoolingHead(nn.Module):
+    """Multihead Attention Pooling."""
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+
+        self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
+        self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.mlp = SiglipMLP(config)
+
+    def forward(self, hidden_state):
+        return SiglipMultiheadAttentionPoolingHead_forward(self, hidden_state)