CPT Tuner

docs/source/_toctree.yml

@@ -118,6 +118,8 @@
       title: VB-LoRA
     - local: package_reference/hra
       title: HRA
+    - local: package_reference/cpt
+      title: CPT
     - local: package_reference/bone
       title: Bone
 

docs/source/conceptual_guides/prompting.md

@@ -75,3 +75,19 @@ Take a look at [P-tuning for sequence classification](../task_guides/ptuning-seq
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/mpt-decomposition.png"/>
 </div>
 <small><a href="https://hf.co/papers/2103.10385">Prompt decomposition</a>.</small>
+
+
+## Context-Aware Prompt Tuning (CPT)
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/cpt.png"/>
+</div>
+<small>CPT optimizing only specific token embeddings while keeping the rest of the model frozen <a href="https://huggingface.co/papers/2410.17222">(image source)</a>.</small>
+
+[Context-Aware Prompt Tuning (CPT)](https://huggingface.co/papers/2410.17222) is designed to enhance few-shot classification by refining only context embeddings. 
+This approach combines ideas from In-Context Learning (ICL), Prompt Tuning (PT), and adversarial optimization, focusing on making model adaptation both parameter-efficient and effective.
+In CPT, only specific context token embeddings are optimized, while the rest of the model remains frozen. 
+To prevent overfitting and maintain stability, CPT uses controlled perturbations to limit the allowed changes to context embeddings within a defined range. 
+Additionally, to address the phenomenon of recency bias—where examples near the end of the context tend to be prioritized over earlier ones—CPT applies a decay loss factor.
+
+Take a look at [Context-Aware Prompt Tuning for few-shot classification](../task_guides/cpt-few-shot-classification) for a step-by-step guide on how to train a model with CPT.

docs/source/package_reference/cpt.md

@@ -0,0 +1,31 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Context-Aware Prompt Tuning (CPT)
+
+[Context-aware Prompt Tuning: Advancing In-Context Learning with Adversarial Methods (CPT)](https://huggingface.co/papers/2410.17222) combines In-Context Learning (ICL) with Prompt Tuning (PT) and adversarial optimization to improve few-shot learning by refining context embeddings. CPT optimizes only context tokens, which minimizes overfitting and enhances performance on classification tasks.
+
+The abstract from the paper is:
+
+*Traditional fine-tuning is effective but computationally intensive, as it requires updating billions of parameters. CPT, inspired by ICL, PT, and adversarial attacks, refines context embeddings in a parameter-efficient manner. By optimizing context tokens and applying a controlled gradient descent, CPT achieves superior accuracy across various few-shot classification tasks, showing significant improvement over existing methods such as LoRA, PT, and ICL.*
+
+## CPTConfig
+
+[[autodoc]] tuners.cpt.config.CPTConfig
+
+## CPTEmbedding
+
+[[autodoc]] tuners.cpt.model.CPTEmbedding
+

src/peft/__init__.py

@@ -94,6 +94,8 @@
     VBLoRAConfig,
     get_eva_state_dict,
     initialize_lora_eva_weights,
+    CPTEmbedding,
+    CPTConfig,
     BoneConfig,
     BoneModel,
 )

src/peft/mapping.py

@@ -40,6 +40,8 @@
     BOFTModel,
     BoneConfig,
     BoneModel,
+    CPTConfig,
+    CPTEmbedding,
     FourierFTConfig,
     FourierFTModel,
     HRAConfig,
@@ -106,6 +108,7 @@
     "XLORA": XLoraConfig,
     "HRA": HRAConfig,
     "VBLORA": VBLoRAConfig,
+    "CPT": CPTConfig,
     "BONE": BoneConfig,
 }
 
@@ -124,6 +127,7 @@
     "XLORA": XLoraModel,
     "HRA": HRAModel,
     "VBLORA": VBLoRAModel,
+    "CPT": CPTEmbedding,
     "BONE": BoneModel,
 }
 

src/peft/peft_model.py

@@ -47,6 +47,7 @@
     AdaptionPromptModel,
     BOFTModel,
     BoneModel,
+    CPTEmbedding,
     FourierFTModel,
     HRAModel,
     IA3Model,
@@ -105,6 +106,7 @@
     PeftType.XLORA: XLoraModel,
     PeftType.HRA: HRAModel,
     PeftType.VBLORA: VBLoRAModel,
+    PeftType.CPT: CPTEmbedding,
     PeftType.BONE: BoneModel,
 }
 
@@ -654,6 +656,8 @@ def _setup_prompt_encoder(self, adapter_name: str):
             if any(getattr(module, "gradient_checkpointing", False) for module in self.get_base_model().modules()):
                 raise ValueError("Prefix tuning does not work with gradient checkpointing.")
             prompt_encoder = PrefixEncoder(config)
+        elif config.peft_type == PeftType.CPT:
+            prompt_encoder = CPTEmbedding(config, self.word_embeddings)
         else:
             raise ValueError("Not supported")
 
@@ -1746,6 +1750,8 @@ def forward(
             # overwrite past_kv in kwargs
             kwargs["past_key_values"] = self.get_prompt(batch_size)
             return self.base_model(input_ids=input_ids, inputs_embeds=inputs_embeds, **kwargs)
+        elif peft_config.peft_type == PeftType.CPT:
+            return self._cpt_forward(input_ids, inputs_embeds, peft_config, task_ids, batch_size, **kwargs)
         else:
             if inputs_embeds is None:
                 inputs_embeds = self.word_embeddings(input_ids)
@@ -1758,6 +1764,63 @@ def forward(
             inputs_embeds = torch.cat((prompts, inputs_embeds), dim=1)
             return self.base_model(inputs_embeds=inputs_embeds, **kwargs)
 
+    def _cpt_forward(
+        self, input_ids=None, inputs_embeds=None, peft_config=None, task_ids=None, batch_size=None, **kwargs
+    ):
+        # Extract labels from kwargs
+        labels = kwargs.pop("labels")
+        device = [i.device for i in [input_ids, inputs_embeds, labels] if i is not None][0]
+        # Extract input_type_mask from kwargs and move it to the same device as labels
+        if "input_type_mask" in kwargs.keys():
+            input_type_mask = kwargs.pop("input_type_mask").to(device)
+        else:
+            if input_ids is None:
+                N_tokens = inputs_embeds.shape[1]
+            else:
+                N_tokens = input_ids.shape[1]
+            input_type_mask = torch.ones((batch_size, N_tokens)).to(device) * 4
+
+        cpt_token_ids = peft_config.cpt_token_ids
+        cpt_tokens_type_mask = peft_config.cpt_tokens_type_mask
+
+        # Generate embeddings if not provided
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        # Get prompt and concatenate with input embeddings
+        prompts = self.get_prompt(batch_size=batch_size, task_ids=task_ids)
+        prompts = prompts.to(inputs_embeds.dtype)
+        inputs_embeds = torch.cat((prompts, inputs_embeds), dim=1)
+        # If labels are provided, generate prefix labels and type mask
+        cpt_labels = None
+        if labels is not None:
+            # Generate prefix labels and concatenate with the input labels
+            prefix_labels = torch.Tensor(cpt_token_ids).long().view(1, -1)
+            prefix_labels = prefix_labels.repeat(batch_size, 1).to(labels.device)
+            cpt_labels = torch.cat((prefix_labels, labels), dim=1)
+            # Generate prefix type mask and shift input type mask values to avoid conflicts
+            prefix_type_mask = torch.Tensor(cpt_tokens_type_mask).long().view(1, -1)
+            prefix_type_mask = prefix_type_mask.repeat(batch_size, 1).to(labels.device)
+            adjusted_input_type_mask = input_type_mask
+            adjusted_input_type_mask[adjusted_input_type_mask > 0] += prefix_type_mask.max()
+            # Concatenate prefix and shifted input type masks
+            cpt_type_mask = torch.cat((prefix_type_mask, adjusted_input_type_mask), dim=1)
+            # Identify valid label positions and mask invalid ones with -100
+            labels_idx = (cpt_type_mask > 0) & (cpt_type_mask % 4 == 0)
+            cpt_labels[~labels_idx] = -100
+            # Update kwargs with the modified labels
+
+        kwargs["labels"] = cpt_labels
+        # Pass the modified inputs to the base model
+        base_model_output = self.base_model(inputs_embeds=inputs_embeds, **kwargs)
+        if labels is None:
+            return base_model_output
+        else:
+            # Calculate the loss using the custom CPT loss function
+            base_model_output = CPTEmbedding.calculate_loss(
+                base_model_output, cpt_labels, cpt_type_mask, self.peft_config["default"]
+            )
+            return base_model_output
+
     def generate(self, *args, **kwargs):
         peft_config = self.active_peft_config
         self.base_model.prepare_inputs_for_generation = self.prepare_inputs_for_generation

src/peft/tuners/__init__.py

@@ -45,4 +45,5 @@
 from .xlora import XLoraConfig, XLoraModel
 from .hra import HRAConfig, HRAModel
 from .vblora import VBLoRAConfig, VBLoRAModel
+from .cpt import CPTConfig, CPTEmbedding
 from .bone import BoneConfig, BoneModel

src/peft/tuners/cpt/__init__.py

@@ -0,0 +1,20 @@
+# Copyright 2024-present the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from .config import CPTConfig
+from .model import CPTEmbedding
+
+
+__all__ = ["CPTConfig", "CPTEmbedding"]

src/peft/tuners/cpt/config.py

@@ -0,0 +1,98 @@
+# Copyright 2024-present the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+from typing import Literal, Optional
+
+from peft.config import PromptLearningConfig
+from peft.utils import PeftType
+
+
+@dataclass
+class CPTConfig(PromptLearningConfig):
+    """
+    CPT Configuration class extending PeftConfig for Context-aware Prompt Tuning (CPT).
+
+    This class introduces additional parameters required for CPT, such as:
+    - Token type masks
+    - Prompt tuning initialization
+    - Loss weighting
+    - Projection settings
+
+    For more details, see the paper: https://arxiv.org/abs/2410.17222
+    """
+
+    # Token-related configurations
+    cpt_token_ids: Optional[list[int]] = field(
+        default=None, metadata={"help": "Tensor of token IDs used for CPT prompts."}
+    )
+    cpt_mask: Optional[list[int]] = field(default=None, metadata={"help": "Tensor mask applied to CPT tokens."})
+    cpt_tokens_type_mask: Optional[list[int]] = field(
+        default=None, metadata={"help": "Mask indicating the type of each CPT token."}
+    )
+
+    # Loss-related configurations
+    opt_weighted_loss_type: Optional[Literal["none", "decay"]] = field(
+        default="none", metadata={"help": "Type of weighted loss: 'none' or 'decay'."}
+    )
+    opt_loss_decay_factor: Optional[float] = field(
+        default=1.0, metadata={"help": "Factor for exponential decay in loss weighting."}
+    )
+
+    # Projection-related configurations
+    opt_projection_epsilon: Optional[float] = field(
+        default=0.1, metadata={"help": "Epsilon value for input projection."}
+    )
+    opt_projection_format_epsilon: Optional[float] = field(
+        default=0.1, metadata={"help": "Epsilon value for format projection."}
+    )
+
+    # Tokenizer configuration
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The tokenizer to use for prompt tuning initialization. Only used if prompt_tuning_init is `TEXT`"
+        },
+    )
+    # Neet to define CPT-specific static attributes
+    is_prompt_learning = True  # Indicates that CPT is a prompt-learning method.
+
+    def __post_init__(self):
+        """
+        Post-initialization hook to set additional attributes after the config is initialized.
+        """
+        # CPT-specific static attributes
+        self.is_prompt_learning = True  # Indicates that CPT is a prompt-learning method.
+        self.num_layers = None  # Number of layers (optional, not always required).
+        self.token_dim = None  # Dimension of token embeddings.
+        self.num_attention_heads = None  # Number of attention heads (if applicable).
+        self.num_transformer_submodules = 1  # Number of transformer submodules used.
+        self.peft_type = PeftType.CPT  # Specifies that the PEFT type is CPT.
+        self.task_type = "CAUSAL_LM"  # Ensures task type is causal language modeling.
+
+        if self.cpt_token_ids is None:
+            self.cpt_token_ids = [0]
+
+        self.num_virtual_tokens = len(self.cpt_token_ids)
+
+        if self.cpt_mask is None:
+            self.cpt_mask = [1 for _ in self.cpt_token_ids]
+
+        if self.cpt_tokens_type_mask is None:
+            self.cpt_tokens_type_mask = [1 for _ in self.cpt_token_ids]
+
+        if not (
+            len(self.cpt_token_ids) == len(self.cpt_mask) == len(self.cpt_tokens_type_mask) == self.num_virtual_tokens
+        ):
+            raise ValueError("cpt_token_ids, cpt_mask and cpt_tokens_type_mask must have the same length.")