Skip to content

Peft

Evo2LoRA

Bases: LoRA

LoRA adapter specifically for Evo2/Hyena models.

Source code in bionemo/evo2/run/peft.py 
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
class Evo2LoRA(LoRA):
    """LoRA adapter specifically for Evo2/Hyena models."""

    def __init__(
        self,
        peft_ckpt_path: Optional[str] = None,
        freeze_modules: List[str] = ["encoder", "embedding"],
        target_modules: List[str] = [
            "linear_qkv",
            "linear_proj",
            "linear_fc1",
            "linear_fc2",
            "short_filter",  # Short convolution filters
            "hyena_filter",  # Hyena layer filters
            "positional_encoding",  # ROPE or other position encodings
        ],
        *args,
        **kwargs,
    ):
        """Initialize the LoRA Adapter for Evo2.

        Args:
            peft_ckpt_path: Path to pre-trained LoRA checkpoint.
            freeze_modules: List of module names to freeze (Evo2-specific defaults).
            target_modules: Modules to apply LoRA to (uses Evo2 defaults if None).
            *args: placeholder.
            **kwargs:
                dim: LoRA rank dimension.
                alpha: LoRA scaling parameter.
                dropout: Dropout rate for LoRA layers.
                dropout_position: Where to apply dropout ('pre' or 'post').
                lora_A_init_method: Initialization for A matrix ('xavier', 'uniform', 'normal').
                lora_B_init_method: Initialization for B matrix ('zero', 'normal').
        """
        """Initialize the LoRA Adapter for Evo2."""
        super().__init__(target_modules=target_modules, *args, **kwargs)
        self.freeze_modules = freeze_modules
        self.peft_ckpt_path = peft_ckpt_path

        # CRITICAL: Set model_transform to self
        # The callback system expects this attribute
        self.model_transform = self

    def setup(self, trainer, pl_module, stage):
        """Setup callback - properly initialize transform."""
        super().setup(trainer, pl_module, stage)

        logging.info(f"Will attempt to apply to model if matches: \n{self.target_modules}")

        # Ensure model_transform is set
        if not hasattr(self, "model_transform") or self.model_transform is None:
            self.model_transform = self

        # Pass checkpoint path to wrapped IO if available
        if hasattr(self, "wrapped_io") and self.peft_ckpt_path:
            self.wrapped_io.adapter_ckpt_path = self.peft_ckpt_path

    def on_predict_epoch_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
        """Event hook.

        Apply transformations for prediction if needed.

        Args:
            trainer: The trainer object.
            pl_module: The LightningModule object.
        """
        self._maybe_apply_transform(trainer)

    def adapter_key_filter(self, key: str) -> bool:
        """Filter state dict keys to identify adapter parameters.

        Args:
            key: State dict key to check

        Returns:
            bool: True if key corresponds to an adapter parameter
        """
        if isinstance(key, tuple):
            return key[1].requires_grad

        if "_extra_state" in key:
            return False

        # Check if it's an adapter parameter or not in freeze list
        return (
            (not any(substring in key for substring in self.freeze_modules))
            or ".adapter." in key
            or key.endswith(".adapters")
            or "lora_A" in key
            or "lora_B" in key
        )

    def __call__(self, model: nn.Module) -> nn.Module:
        """Apply LoRA transformations to the model.

        Override to avoid fn.walk compatibility issues.
        """
        # First, manually freeze specified modules
        self._apply_selective_freeze(model)

        # Then apply LoRA transformations
        self._apply_lora_transform(model)

        # THEN freeze ALL base model parameters
        # This must happen AFTER LoRA is applied
        self._freeze_base_model_parameters(model)

        # Log summary
        self._log_lora_summary(model)

        return model

    def _apply_selective_freeze(self, model: nn.Module, prefix=""):
        """Manually walk model and freeze specified modules."""
        for name, child in model.named_children():
            full_name = f"{prefix}.{name}" if prefix else name

            # Check if this module should be frozen
            if name in self.freeze_modules:
                logging.info(f"Freezing module: {full_name}")
                for param in child.parameters():
                    param.requires_grad = False

            # Recursively apply to children
            self._apply_selective_freeze(child, full_name)

    def _freeze_base_model_parameters(self, model: nn.Module):
        """Freeze all parameters except LoRA adapters and critical layers."""
        logging.info("\nFreezing base model parameters...")
        frozen_count = 0
        kept_trainable = []

        for name, param in model.named_parameters():
            # Keep LoRA/adapter parameters trainable
            if any(adapter_term in name for adapter_term in ["adapter", "lora_A", "lora_B", "lora"]):
                param.requires_grad = True
                kept_trainable.append(name)
            # CRITICAL: Keep output layer trainable to maintain gradient flow
            elif "output_layer" in name or "lm_head" in name:
                param.requires_grad = True
                kept_trainable.append(name)
                logging.info(f"  Keeping output layer trainable: {name}")
            # CRITICAL: Keep final layer norm trainable
            elif "final_norm" in name or ("decoder" in name and "norm" in name and "24" in name):
                param.requires_grad = True
                kept_trainable.append(name)
                logging.info(f"  Keeping final norm trainable: {name}")
            else:
                param.requires_grad = False
                frozen_count += 1

        logging.info(f"Froze {frozen_count} parameter tensors")
        logging.info(f"Kept {len(kept_trainable)} parameters trainable")

    def _apply_lora_transform(self, model: nn.Module, prefix=""):
        """Apply LoRA with better tracking."""
        # Get all modules in a flat list first
        modules_to_transform = []

        for name, module in model.named_modules():
            # Skip if has children (not a leaf module)
            if list(module.children()):
                continue

            # Check if this matches our target modules
            module_type = name.split(".")[-1] if "." in name else name
            if module_type in self.target_modules:
                modules_to_transform.append((name, module))

        logging.info(f"\nFound {len(modules_to_transform)} modules to apply LoRA to")

        # Apply transformations
        for full_name, module in modules_to_transform:
            # Get parent and attribute name
            parts = full_name.split(".")
            parent = model
            for part in parts[:-1]:
                parent = getattr(parent, part)

            # Apply transform
            attr_name = parts[-1]
            transformed = self.transform(module, name=attr_name, prefix="")

            if transformed is not module:
                setattr(parent, attr_name, transformed)
                logging.info(f"Applied LoRA to: {full_name}")

                # Verify LoRA was applied
                if hasattr(transformed, "adapter") or hasattr(transformed, "lora_A"):
                    logging.info(f"  ✓ LoRA adapter confirmed on {full_name}")

    def selective_freeze(self, m: nn.Module, name=None, prefix=None):
        """Selectively freeze modules based on freeze_modules list.

        Args:
            m: Module to potentially freeze.
            name: Name of the module.
            prefix: Prefix for the module name.

        Returns:
            nn.Module: The module (frozen or not).
        """
        if name in self.freeze_modules:
            FNMixin.freeze(m)
            logging.info(f"Freezing module: {prefix}.{name}" if prefix else f"Freezing module: {name}")

        return m

    # Deepcopy compatibility
    def __deepcopy__(self, memo):
        """Custom deepcopy to handle unpickleable objects."""
        # Create a new instance with the same parameters
        cls = self.__class__
        result = cls.__new__(cls)

        # Copy all attributes except problematic ones
        memo[id(self)] = result
        for k, v in self.__dict__.items():
            if k not in ["_metadata", "_fields"]:  # Skip dataclass internals
                try:
                    setattr(result, k, deepcopy(v, memo))
                except Exception:
                    # If deepcopy fails, just use the original reference
                    setattr(result, k, v)

        return result

    def __getstate__(self):
        """Prepare object for pickling."""
        state = self.__dict__.copy()
        # Remove unpickleable entries
        state.pop("_metadata", None)
        state.pop("_fields", None)
        return state

    def __setstate__(self, state):
        """Restore object from pickle."""
        self.__dict__.update(state)

    # Debug module
    def _log_lora_summary(self, model: nn.Module):
        """Log a summary of LoRA modifications."""
        total_params = sum(p.numel() for p in model.parameters())
        trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
        adapter_params = sum(p.numel() for n, p in model.named_parameters() if p.requires_grad and "adapter" in n)

        logging.info(f"\n{'=' * 50}")
        logging.info("LoRA Summary:")
        logging.info(f"  Total parameters: {total_params:,}")
        logging.info(f"  Trainable parameters: {trainable_params:,}")
        logging.info(f"  Adapter parameters: {adapter_params:,}")  # Changed from "LoRA parameters"
        logging.info(f"  Percentage trainable: {100 * trainable_params / total_params:.2f}%")
        logging.info(f"  Percentage adapters: {100 * adapter_params / total_params:.2f}%")
        logging.info(f"{'=' * 50}\n")

__call__(model)

Apply LoRA transformations to the model.

Override to avoid fn.walk compatibility issues.

Source code in bionemo/evo2/run/peft.py 
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
def __call__(self, model: nn.Module) -> nn.Module:
    """Apply LoRA transformations to the model.

    Override to avoid fn.walk compatibility issues.
    """
    # First, manually freeze specified modules
    self._apply_selective_freeze(model)

    # Then apply LoRA transformations
    self._apply_lora_transform(model)

    # THEN freeze ALL base model parameters
    # This must happen AFTER LoRA is applied
    self._freeze_base_model_parameters(model)

    # Log summary
    self._log_lora_summary(model)

    return model

__deepcopy__(memo)

Custom deepcopy to handle unpickleable objects.

Source code in bionemo/evo2/run/peft.py 
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
def __deepcopy__(self, memo):
    """Custom deepcopy to handle unpickleable objects."""
    # Create a new instance with the same parameters
    cls = self.__class__
    result = cls.__new__(cls)

    # Copy all attributes except problematic ones
    memo[id(self)] = result
    for k, v in self.__dict__.items():
        if k not in ["_metadata", "_fields"]:  # Skip dataclass internals
            try:
                setattr(result, k, deepcopy(v, memo))
            except Exception:
                # If deepcopy fails, just use the original reference
                setattr(result, k, v)

    return result

__getstate__()

Prepare object for pickling.

Source code in bionemo/evo2/run/peft.py 
267
268
269
270
271
272
273
def __getstate__(self):
    """Prepare object for pickling."""
    state = self.__dict__.copy()
    # Remove unpickleable entries
    state.pop("_metadata", None)
    state.pop("_fields", None)
    return state

__init__(peft_ckpt_path=None, freeze_modules=['encoder', 'embedding'], target_modules=['linear_qkv', 'linear_proj', 'linear_fc1', 'linear_fc2', 'short_filter', 'hyena_filter', 'positional_encoding'], *args, **kwargs)

Initialize the LoRA Adapter for Evo2.

Parameters:

Name Type Description Default
peft_ckpt_path Optional[str]

Path to pre-trained LoRA checkpoint.

 None
freeze_modules List[str]

List of module names to freeze (Evo2-specific defaults).

 ['encoder', 'embedding']
target_modules List[str]

Modules to apply LoRA to (uses Evo2 defaults if None).

 ['linear_qkv', 'linear_proj', 'linear_fc1', 'linear_fc2', 'short_filter', 'hyena_filter', 'positional_encoding']
*args

placeholder.

 ()
**kwargs

dim: LoRA rank dimension. alpha: LoRA scaling parameter. dropout: Dropout rate for LoRA layers. dropout_position: Where to apply dropout ('pre' or 'post'). lora_A_init_method: Initialization for A matrix ('xavier', 'uniform', 'normal'). lora_B_init_method: Initialization for B matrix ('zero', 'normal').

 {}
Source code in bionemo/evo2/run/peft.py 
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
def __init__(
    self,
    peft_ckpt_path: Optional[str] = None,
    freeze_modules: List[str] = ["encoder", "embedding"],
    target_modules: List[str] = [
        "linear_qkv",
        "linear_proj",
        "linear_fc1",
        "linear_fc2",
        "short_filter",  # Short convolution filters
        "hyena_filter",  # Hyena layer filters
        "positional_encoding",  # ROPE or other position encodings
    ],
    *args,
    **kwargs,
):
    """Initialize the LoRA Adapter for Evo2.

    Args:
        peft_ckpt_path: Path to pre-trained LoRA checkpoint.
        freeze_modules: List of module names to freeze (Evo2-specific defaults).
        target_modules: Modules to apply LoRA to (uses Evo2 defaults if None).
        *args: placeholder.
        **kwargs:
            dim: LoRA rank dimension.
            alpha: LoRA scaling parameter.
            dropout: Dropout rate for LoRA layers.
            dropout_position: Where to apply dropout ('pre' or 'post').
            lora_A_init_method: Initialization for A matrix ('xavier', 'uniform', 'normal').
            lora_B_init_method: Initialization for B matrix ('zero', 'normal').
    """
    """Initialize the LoRA Adapter for Evo2."""
    super().__init__(target_modules=target_modules, *args, **kwargs)
    self.freeze_modules = freeze_modules
    self.peft_ckpt_path = peft_ckpt_path

    # CRITICAL: Set model_transform to self
    # The callback system expects this attribute
    self.model_transform = self

__setstate__(state)

Restore object from pickle.

Source code in bionemo/evo2/run/peft.py 
275
276
277
def __setstate__(self, state):
    """Restore object from pickle."""
    self.__dict__.update(state)

adapter_key_filter(key)

Filter state dict keys to identify adapter parameters.

Parameters:

Name Type Description Default
key str

State dict key to check

 required

Returns:

Name Type Description
bool bool

True if key corresponds to an adapter parameter
Source code in bionemo/evo2/run/peft.py 
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
def adapter_key_filter(self, key: str) -> bool:
    """Filter state dict keys to identify adapter parameters.

    Args:
        key: State dict key to check

    Returns:
        bool: True if key corresponds to an adapter parameter
    """
    if isinstance(key, tuple):
        return key[1].requires_grad

    if "_extra_state" in key:
        return False

    # Check if it's an adapter parameter or not in freeze list
    return (
        (not any(substring in key for substring in self.freeze_modules))
        or ".adapter." in key
        or key.endswith(".adapters")
        or "lora_A" in key
        or "lora_B" in key
    )

on_predict_epoch_start(trainer, pl_module)

Event hook.

Apply transformations for prediction if needed.

Parameters:

Name Type Description Default
trainer Trainer

The trainer object.

 required
pl_module LightningModule

The LightningModule object.

 required
Source code in bionemo/evo2/run/peft.py 
 97
 98
 99
100
101
102
103
104
105
106
def on_predict_epoch_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
    """Event hook.

    Apply transformations for prediction if needed.

    Args:
        trainer: The trainer object.
        pl_module: The LightningModule object.
    """
    self._maybe_apply_transform(trainer)

selective_freeze(m, name=None, prefix=None)

Selectively freeze modules based on freeze_modules list.

Parameters:

Name Type Description Default
m Module

Module to potentially freeze.

 required
name

Name of the module.

 None
prefix

Prefix for the module name.

 None

Returns:

Type Description

nn.Module: The module (frozen or not).
Source code in bionemo/evo2/run/peft.py 
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
def selective_freeze(self, m: nn.Module, name=None, prefix=None):
    """Selectively freeze modules based on freeze_modules list.

    Args:
        m: Module to potentially freeze.
        name: Name of the module.
        prefix: Prefix for the module name.

    Returns:
        nn.Module: The module (frozen or not).
    """
    if name in self.freeze_modules:
        FNMixin.freeze(m)
        logging.info(f"Freezing module: {prefix}.{name}" if prefix else f"Freezing module: {name}")

    return m

setup(trainer, pl_module, stage)

Setup callback - properly initialize transform.

Source code in bionemo/evo2/run/peft.py 
83
84
85
86
87
88
89
90
91
92
93
94
95
def setup(self, trainer, pl_module, stage):
    """Setup callback - properly initialize transform."""
    super().setup(trainer, pl_module, stage)

    logging.info(f"Will attempt to apply to model if matches: \n{self.target_modules}")

    # Ensure model_transform is set
    if not hasattr(self, "model_transform") or self.model_transform is None:
        self.model_transform = self

    # Pass checkpoint path to wrapped IO if available
    if hasattr(self, "wrapped_io") and self.peft_ckpt_path:
        self.wrapped_io.adapter_ckpt_path = self.peft_ckpt_path