oiSSKJr SSKrSSKJrJr SSKJrJrJ r SSK J r SSK J r SSKJr \"SS 55r\"S S 55r\"S S 55r\"SS55r\"SS55r\"SS\ 55rg)) annotationsN) dataclassfield)LiteralOptionalUnion)nn) PeftConfig)PeftTypec6\rSrSr%Sr\"SSS0S9rS\S'S rg ) LoraRuntimeConfigz This is the sub-configuration class to store the runtime configurations for the model. Args: ephemeral_gpu_offload (`bool`): Whether to use ephemeral GPU offloading for models partially kept in CPU memory. FhelpaWhether to use ephemeral GPU offloading for models partially kept in CPU memory. Ephemeral GPU offloading result in the data involved in intense operations being momentarily copied over to the GPU, and the results copied back to CPU. There is a momentary VRAM overhead, but operations are generally orders of magnitude faster compared to performing them on the CPU. This is useful when parts of the model and/or components (such as adapters) are kept in CPU memory until they are needed. Rather than perform expensive operations on small data, the data is transferred to the GPU on-demand, the operation(s) performed, and the results moved back to CPU memory. Currently only affects DoRA initialization.defaultmetadataboolephemeral_gpu_offloadN) __name__ __module__ __qualname____firstlineno____doc__rr__annotations____static_attributes__rQ/home/james-whalen/.local/lib/python3.13/site-packages/peft/tuners/lora/config.pyr r s,#( X   #4 rr cR\rSrSr%Sr\"SSS0S9rS\S'\"S SS 0S9rS\S 'S r g ) LoftQConfig5a This is the sub-configuration class to store the configuration of a [`LoraModel`]. Args: bits_pattern (`dict`): The mapping from layer names or regexp expression to bits which are different from the default bits specified by `bits`. For example, `{model.decoder.layers.0.encoder_attn.k_proj: 2`}. bits (`int`): Quantization bits for LoftQ. iter (`int`): Alternating iterations for LoftQ. fake (`bool`): True: use fp16/fp32; used for first time to save weights. False: use bitsandbytes 4bit linear models. weights can't be saved. Recommend to set to True, save the weights and load the saved weights in 4 bits. rzQuantization bits for LoftQrint loftq_bitsz Alternating iterations for LoftQ loftq_iterrN) rrrrrrr$rr&rrrrr r 5s7 A9V0WXJXA9[0\]J]rr c\rSrSr%Sr\"SSS0S9rS\S'\"S SS 0S9rS \S '\"S SS0S9r S\S'\"SS SS0S9r S\S'\"SS SS0S9r S\S'\"SSS0S9r S\S'Sr Srg) ArrowConfigHa This is the sub-configuration class to store the configuration for Arrow and GenKnowSub algorithm. Arrow is a routing algorithm to combine the trained LoRA modules to solve new tasks, proposed in 'https://huggingface.co/papers/2405.11157'. GenKnowSub is a refinement on the trained modules before being combined via Arrow, introduced in 'https://aclanthology.org/2025.acl-short.54/' rz7Number of top LoRA modules to combine in Arrow routing.rr#top_k?zExplained Variance Adaptation. Args: rho (`float`): Rho value for EVA redistribution (>= 1.0). The maximum rank for a layer is lora_r * rho. Default is 2.0, meaning the maximum rank allowed for a layer is 2r. Increasing rho will allow for a higher degree of redistribution of ranks across layers. Some pre-trained models might be more sensitive to a rank redistribution. It can therefore be beneficial to try rho=1.0 (no redistribution) if the performance is lower than expected. tau (`float`): Cosine similarity threshold for early stopping. Compares the cosine similarity of right-singular vectors between two consecutive SVD steps. If the cosine similarity is above this threshold, the SVD iteration is stopped. Default is 0.99. use_label_mask (`bool`): Use label mask for EVA initialization. This means that positions where labels=label_mask_value are ignored for the SVD computation. Setting use_label_mask=True is preferred in most cases and can be especially beneficial for multi-turn conversations. The default value is True. Filtering out items based on the label mask can sometimes lead to a small batch size and as a result instabilities in the SVD computation. For cases where a large share of batch items would be filtered out, set use_label_mask=False. label_mask_value (`int`): If use_label_mask=True the value to look for to mask out ignored tokens. Default is -100. whiten (`bool`): Apply whitening to singular vectors. Default is False. Whitening has been shown to be beneficial for EVA in the vision domain. adjust_scaling_factors (`bool`): Adjust LoRA scaling factors after the rank redistribution. Setting this to True means the scaling factors are adjusted so that all LoRA gradients have the same scale regardless of their rank. Default is True. g@rz Rho value for EVA redistributionrr-rhogGz?z.Cosine similarity threshold for early stoppingtauTz%Use label mask for EVA initializationruse_label_maskizGif use_label_mask=True the value to look for to mask out ignored tokensr#label_mask_valueFz#Apply whitening to singular vectorswhitenz9Adjust LoRA scaling factors after the rank redistributionadjust_scaling_factorscURS:a [S5eURS:dURS:a [S5eg)Nr,z`rho` must be >= 1.0z"`tau` must be between 0.0 and 1.0.)r>r6r?r7s rr9EvaConfig.__post_init__s? 88c>34 4 88c>TXX^AB B,rrN)rrrrrrr>rr?r@rArBrCr9rrrrr<r<zs<sf6X-YZCZtv7g.hiCi Ah8ijNDj!(qrc&:_1`aFDa#(UV$D Crr<c\rSrSr%Sr\"SSS0S9rS\S'\"SSS 0S9rS\S '\"S SS 0S9r S \S'\"SSS0S9r S\S'\"SSS0S9r S\S'\"SSS0S9r S\S'Sr g) CordaConfigaH This is the sub-configuration class to store the configuration of a [`LoraModel`]. Args: cache_file (`Optional[str]`): File to store the SVD cache. The SVD cache is much smaller than the residual model (for example, residual model of Llama-3-8b is 15GB, while SVD cache is 1.4GB), but with SVD cache and original model weights, residual model weights can be built quickly. If you need to reuse residual model weights with limited storage, you can store the SVD cache instead. covariance_file (`Optional[str]`): File to store the covariance matrix. If you wish to train multiple models with different ranks, but they sample from the same dataset, you can store the covariance matrix and reuse it for different ranks. Note that covariance file is usually large (comparable to model size), so you will need sufficient storage. corda_method (`Literal["ipm", "kpm"]`): Method to build adapter. The KPM (Knowledge-Preserved Mode) not only achieves better performance than LoRA on fine-tuning tasks, but also mitigates the catastrophic forgetting of pre-trained world knowledge. When preserving pre-trained knowledge is not a concern, the IPM (Instruction-Previewed Mode) is favored because it can further accelerate convergence and enhance the fine-tuning performance. Defaults to `'ipm'`. verbose (`bool`): If true, prints the progress of CorDA initialization. Defaults to `False`. use_float16_for_covariance (`bool`): If true, uses float16 for the covariance matrix. This can reduce the memory usage of the covariance matrix by half, but may lead to numerical instability. Defaults to `False`. prune_temporary_fields (`bool`): If true, temporary fields generated in CorDA preprocessing will be pruned. Defaults to `True`. NradFile to store the SVD cache. The SVD cache is much smaller than the residual model (for example, residual model of Llama-3-8b is 15GB, while SVD cache is 1.4GB), but with SVD cache and original model weights, residual model weights can be built quickly. If you need to reuse residual model weights with limited storage, you can store the SVD cache instead.r Optional[str] cache_filea8File to store the covariance matrix. If you wish to train multiple models with different ranks, but they sample from the same dataset, you can store the covariance matrix and reuse it for different ranks. Note that covariance file is usually large (comparable to model size), so you will need sufficient storage.covariance_fileipmaVMethod to build adapter. The KPM not only achieves better performance than LoRA on fine-tuning tasks, but also mitigates the catastrophic forgetting of pre-trained world knowledge. When preserving pre-trained knowledge is not a concern, the IPM is favored because it can further accelerate convergence and enhance the fine-tuning performance.zLiteral['ipm', 'kpm'] corda_methodFz5If true, prints the progress of CorDA initialization.rverbosezIf true, uses float16 for the covariance matrix. This can reduce the memory usage of the covariance matrix by half, but may lead to numerical instability.use_float16_for_covarianceTzJIf true, temporary fields generated in CorDA preprocessing will be pruned.prune_temporary_fieldsr)rrrrrrrKrrLrNrOrPrQrrrrrHrHs6!& H  !J &+ ~  &O] +0 /  +L' %6;r2stGTt', B ($)(tu$DrrHc^\rSrSr%Sr\"SSS0S9rS\S'\"S SS 0S9rS \S '\"S SS 0S9r S \S'\"SSS0S9r S\S'\"SSS0S9r S\S'\"SSS0S9r S\S'\"SSS0S9r S\S'\"SSS0S9rS\S'\"S SS0S9rS \S!'\"S"SS#0S9rS$\S%'\"S SS&0S9rS'\S('\"S SS)0S9rS \S*'\"\SS+0S,9rS-\S.'\"\SS/0S,9rS-\S0'\"S SS10S9rS-\S2'\"S3SS40S9rS5\S6'\"S SS70S9rS8\S9'\"\SS:0S,9rS;\S<'\"S SS=0S9rS>\S?'\"S SS@0S9rSA\SB'\"SSSC0S9rS\SD'\"S SSE0S9rSF\SG'\"SSSH0S9rS\SI'\"SJSSK0S9rS\SL'\"S SSM0S9r SN\SO'\"\!SSP0S,9r"SQ\SR'\"SSSS0S9r#S\ST'\"S SSU0S9r$S \SV'\"S SSW0S9r%SX\SY'\"SSSZ0S9r&S\S['U4S\jr'U4S]jr(S`S^jr)S_r*U=r+$)a LoraConfiga- This is the configuration class to store the configuration of a [`LoraModel`]. Args: r (`int`): Lora attention dimension (the "rank"). target_modules (`Optional[Union[List[str], str]]`): The names of the modules to apply the adapter to. If this is specified, only the modules with the specified names will be replaced. When passing a string, a regex match will be performed. When passing a list of strings, either an exact match will be performed or it is checked if the name of the module ends with any of the passed strings. If this is specified as 'all-linear', then all linear/Conv1D modules are chosen (if the model is a PreTrainedModel, the output layer excluded). If this is not specified, modules will be chosen according to the model architecture. If the architecture is not known, an error will be raised -- in this case, you should specify the target modules manually. To avoid targeting any modules (because you want to apply `target_parameters`), set `target_modules=[]`. exclude_modules (`Optional[Union[List[str], str]]`): The names of the modules to not apply the adapter. When passing a string, a regex match will be performed. When passing a list of strings, either an exact match will be performed or it is checked if the name of the module ends with any of the passed strings. lora_alpha (`int`): The alpha parameter for Lora scaling. lora_dropout (`float`): The dropout probability for Lora layers. fan_in_fan_out (`bool`): Set this to True if the layer to replace stores weight like (fan_in, fan_out). For example, gpt-2 uses `Conv1D` which stores weights like (fan_in, fan_out) and hence this should be set to `True`. bias (`str`): Bias type for LoRA. Can be 'none', 'all' or 'lora_only'. If 'all' or 'lora_only', the corresponding biases will be updated during training. Be aware that this means that, even when disabling the adapters, the model will not produce the same output as the base model would have without adaptation. use_rslora (`bool`): When set to True, uses [Rank-Stabilized LoRA](https://huggingface.co/papers/2312.03732) which sets the adapter scaling factor to `lora_alpha/math.sqrt(r)`, since it was proven to work better. Otherwise, it will use the original default value of `lora_alpha/r`. modules_to_save (`List[str]`): List of modules apart from adapter layers to be set as trainable and saved in the final checkpoint. init_lora_weights (`bool` | `Literal["gaussian", "eva", "olora", "pissa", "pissa_niter_[number of iters]", "corda", "loftq", "orthogonal"]`): How to initialize the weights of the adapter layers. Passing True (default) results in the default initialization from the reference implementation from Microsoft, with the LoRA B weight being set to 0. This means that without further training, the LoRA adapter will be a no-op. Setting the initialization to False leads to random initialization of LoRA A and B, meaning that LoRA is not a no-op before training; this setting is intended for debugging purposes. Passing 'gaussian' results in Gaussian initialization scaled by the LoRA rank for linear and layers. Pass `'loftq'` to use LoftQ initialization. Passing `'eva'` results in a data-driven initialization of Explained Variance Adaptation. EVA initializes LoRA based on the SVD of layer input activations and achieves SOTA performance due to its ability to adapt to the finetuning data. Pass `'olora'` to use OLoRA initialization. Passing `'pissa'` results in the initialization of Principal Singular values and Singular vectors Adaptation (PiSSA), which converges more rapidly than LoRA and ultimately achieves superior performance. Moreover, PiSSA reduces the quantization error compared to QLoRA, leading to further enhancements. Passing `'pissa_niter_[number of iters]'` initiates Fast-SVD-based PiSSA initialization, where `[number of iters]` indicates the number of subspace iterations to perform FSVD, and must be a nonnegative integer. When `[number of iters]` is set to 16, it can complete the initialization of a 7B model within seconds, and the training effect is approximately equivalent to using SVD. Passing `'corda'` results in the initialization of Context-Oriented Decomposition Adaptation, which converges even more rapidly than PiSSA in Instruction-Previewed Mode, and preserves world knowledge better than LoRA in Knowledge-Preserved Mode. Passing `"orthogonal"` results in LoRA A and B being intialized orthogonally; in this, it resembles `"olora"`, but the base weights are left untouched (requires `r` to be even, only supported for linear layers for now). layers_to_transform (`Union[List[int], int]`): The layer indices to transform. If a list of ints is passed, it will apply the adapter to the layer indices that are specified in this list. If a single integer is passed, it will apply the transformations on the layer at this index. layers_pattern (`Optional[Union[List[str], str]]`): The layer pattern name, used only if `layers_to_transform` is different from `None`. This should target the `nn.ModuleList` of the model, which is often called `'layers'` or `'h'`. rank_pattern (`dict`): The mapping from layer names or regexp expression to ranks which are different from the default rank specified by `r`. For example, `{'^model.decoder.layers.0.encoder_attn.k_proj': 16}`. alpha_pattern (`dict`): The mapping from layer names or regexp expression to alphas which are different from the default alpha specified by `lora_alpha`. For example, `{'^model.decoder.layers.0.encoder_attn.k_proj': 16}`. megatron_config (`Optional[dict]`): The TransformerConfig arguments for Megatron. It is used to create LoRA's parallel linear layer. You can get it like this, `core_transformer_config_from_args(get_args())`, these two functions being from Megatron. The arguments will be used to initialize the TransformerConfig of Megatron. You need to specify this parameter when you want to apply LoRA to the ColumnParallelLinear and RowParallelLinear layers of megatron. megatron_core (`Optional[str]`): The core module from Megatron to use, defaults to `"megatron.core"`. trainable_token_indices (`Optional[Union[List[int], dict[str, List[int]]]]`) Lets you specify which token indices to selectively fine-tune without requiring to re-train the whole embedding matrix using the `peft.TrainableTokensModel` method. You can specify token indices in two ways. Either you specify a list of indices which will then target the model's input embedding layer (or, if not found, `embed_tokens`). Alternatively, you can specify a dictionary where the key is the name of the embedding module and the values are the list of token indices, e.g. `{'embed_tokens': [0, 1, ...]}`. Note that training with FSDP requires `use_orig_params=True` to avoid issues with non-uniform `requires_grad`. loftq_config (`Optional[LoftQConfig]`): The configuration of LoftQ. If this is not None, then LoftQ will be used to quantize the backbone weights and initialize Lora layers. Also pass `init_lora_weights='loftq'`. Note that you should not pass a quantized model in this case, as LoftQ will quantize the model itself. eva_config (`Optional[EvaConfig]`): The configuration of EVA. At a minimum the dataset argument needs to be set (use the same dataset as for finetuning). corda_config (`Optional[CordaConfig]`): The configuration of CorDA. If this is not None, then CorDA will be used to build the adapter layers. Also pass `init_lora_weights='corda'`. use_dora (`bool`): Enable 'Weight-Decomposed Low-Rank Adaptation' (DoRA). This technique decomposes the updates of the weights into two parts, magnitude and direction. Direction is handled by normal LoRA, whereas the magnitude is handled by a separate learnable parameter. This can improve the performance of LoRA especially at low ranks. Right now, DoRA only supports linear and Conv2D layers. DoRA introduces a bigger overhead than pure LoRA, so it is recommended to merge weights for inference. For more information, see https://huggingface.co/papers/2402.09353. alora_invocation_tokens (`List[int]`): If not None, enable 'Activated LoRA' (aLoRA), with alora_invocation_tokens being the tokenized invocation string for the adapter (must be present in all model input strings). This technique selectively activates the adapter weights only on tokens during and after the alora_invocation_tokens. When used in a CausalLM, this means that the KV cache prior to invocation is interchangeable with that of the base model (and other aLoRA adapters operating this way). As a result, in inference pipelines involving switching between base model inference and adapter inference (e.g. agentic pipelines, see paper for examples), significant savings are realized (relative to LoRA) by saving prefill operations. Overall adapter inference speedups of an order of magnitude or more can occur on vLLM, depending on the length of the shared context. Note that merging is not possible due to the selective application of the weights. layer_replication (`List[Tuple[int, int]]`): Build a new stack of layers by stacking the original model layers according to the ranges specified. This allows expanding (or shrinking) the model without duplicating the base model weights. The new layers will all have separate LoRA adapters attached to them. runtime_config (`LoraRuntimeConfig`): Runtime configurations (which are not saved or restored). lora_bias (`bool`): Defaults to `False`. Whether to enable the bias term for the LoRA B parameter. Typically, this should be disabled. The main use case for this is when the LoRA weights were extracted from fully fine-tuned parameters so the bias of those parameters can be taken into account. target_parameters (`List[str]`, *optional*) List of parameter names or regex expression of the parameter names to replace with LoRA. This argument behaves similarly to `target_modules`, except that the parameter name should be passed. Generally, you should use `target_modules` to target the module (e.g. `nn.Linear`). However, in some circumstances, this is not possible. E.g., in many mixture of expert (MoE) layers in HF Transformers, instead of using `nn.Linear`, an `nn.Parameter` is used. PEFT normally overwrites the `forward` method for LoRA, but for `nn.Parameter`, there is none. Therefore, to apply LoRA to that parameter, it needs to be targeted with `target_parameters`. As an example, for Llama4, you can pass: `target_parameters=['feed_forward.experts.gate_up_proj', 'feed_forward.experts.down_proj]`. Passing a string for regex matching is not implemented yet. rzLora attention dimensionrr#rNa^List of module names or regex expression of the module names to replace with LoRA. For example, ['q', 'v'] or '.*decoder.*(SelfAttention|EncDecAttention).*(q|v)$'. This can also be a wildcard 'all-linear' which matches all linear/Conv1D (if the model is a PreTrainedModel, the output layer excluded). If not specified, modules will be chosen according to the model architecture, If the architecture is not known, an error will be raised -- in this case, you should specify the target modules manually. To avoid targeting any modules (because you want to apply `target_parameters`), set `target_modules=[]`.zOptional[Union[list[str], str]]target_moduleszRList of module names or regex expression of the module names to exclude from Lora.exclude_modulesz Lora alpha lora_alpharEz Lora dropoutr- lora_dropoutFzMSet this to True if the layer to replace stores weight like (fan_in, fan_out)rfan_in_fan_outnonez7Bias type for Lora. Can be 'none', 'all' or 'lora_only'z#Literal['none', 'all', 'lora_only']biasaWhen set to True, uses [Rank-Stabilized LoRA](https://huggingface.co/papers/2312.03732) which sets the adapter scaling factor to `lora_alpha/math.sqrt(r)`, since it was proven to work better. Otherwise, it will use the original default value of `lora_alpha/r`. use_rsloraaList of modules apart from LoRA layers to be set as trainable and saved in the final checkpoint. For example, in Sequence Classification or Token Classification tasks, the final layer `classifier/score` are randomly initialized and as such need to be trainable and saved.r1modules_to_saveTadHow to initialize the weights of the LoRA layers. Passing True (default) results in the default initialization from the reference implementation from Microsoft, with the LoRA B weight being set to 0. This means that without further training, the LoRA adapter will be a no-op. Setting the initialization to False leads to random initialization of LoRA A and B, meaning that LoRA is not a no-op before training; this setting is intended for debugging purposes. Passing `'gaussian'` results in Gaussian initialization scaled by the LoRA rank for linear and layers. Passing `'eva'` results in a data-driven initialization of Explained Variance Adaptation. Passing `'olora'` results in OLoRA initialization. Passing `'pissa'` results in PiSSA initialization. Passing `'pissa_niter_[number of iters]'` initiates Fast-SVD-based PiSSA initialization, where [number of iters] indicates the number of subspace iterations to perform fsvd, and must be a nonnegative integer. Passing `'corda'` results in CorDA initialization. Pass `'loftq'` to use LoftQ initialization. Pass `'orthogonal'` for orthogonal initialization of LoRA A and B.ztbool | Literal['gaussian', 'eva', 'olora', 'pissa', 'pissa_niter_[number of iters]', 'corda', 'loftq', 'orthogonal']init_lora_weightsaThe layer indexes to transform, is this argument is specified, PEFT will transform only the layers indexes that are specified inside this list. If a single integer is passed, PEFT will transform only the layer at this index. This only works when target_modules is a list of str.zOptional[Union[list[int], int]]layers_to_transforma"The layer pattern name, used only if `layers_to_transform` is different to None and if the layer pattern is not in the common layers pattern.This only works when target_modules is a list of str. This should target the `nn.ModuleList` of the model, which is often called `'layers'` or `'h'`.layers_patternzThe mapping from layer names or regexp expression to ranks which are different from the default rank specified by `r`. For example, `{'^model.decoder.layers.0.encoder_attn.k_proj': 16}`.)default_factoryrzOptional[dict] rank_patternzThe mapping from layer names or regexp expression to alphas which are different from the default alpha specified by `lora_alpha`. For example, `{'^model.decoder.layers.0.encoder_attn.k_proj': 16}`. alpha_patternaThe TransformerConfig from Megatron. It is used to create LoRA's parallel linear layer.You can get it like this, `core_transformer_config_from_args(get_args())`, these two functions being from Megatron.You need to specify this parameter when you want to apply LoRA to the ColumnParallelLinear and RowParallelLinear layers of megatron.It should be noted that we may not be able to use the `save_pretrained` and `from_pretrained` functions, because TransformerConfig may not necessarily be serialized.But when using megatron, we can use `get_peft_model_state_dict` function and megatron's framework, they can also save and load models and configurations.megatron_configz megatron.corezThe core module from Megatron, it is used to create LoRA's parallel linear layer. It only needs to be passed in when you need to use your own modified megatron core module. Otherwise, it will use the default value `megatron.core`. rJ megatron_corearLets you specify which token indices to selectively fine-tune without requiring to re-train the whole embedding matrix using the `peft.TrainableTokensModel` method. You can specify token indices in two ways. Either you specify a list of indices which will then target the model's input embedding layer (or, if not found, `embed_tokens`). Alternatively, you can specify a dictionary where the key is the name of the embedding module and the values are the list of token indices, e.g. `{'embed_tokens': [0, 1, ...]}`. Note that training with FSDP requires `use_orig_params=True` to avoid issues with non-uniform `requires_grad`.z0Optional[Union[list[int], dict[str, list[int]]]]trainable_token_indiceszThe configuration of LoftQ. If this is passed, then LoftQ will be used to quantize the backbone weights and initialize Lora layers. Also set `init_lora_weights='loftq'` in this case.zUnion[LoftQConfig, dict] loftq_configzThe configuration of EVA. If this is passed, then EVA will be used to initialize the LoRA layers. Also set `init_lora_weights='eva'` in this case. zOptional[EvaConfig] eva_configzThe configuration of CorDA. If this is passed, then CorDA will be used to build the adapter layers. Also set `init_lora_weights='corda'` in this case.zOptional[CordaConfig] corda_configaEnable 'Weight-Decomposed Low-Rank Adaptation' (DoRA). This technique decomposes the updates of the weights into two parts, magnitude and direction. Direction is handled by normal LoRA, whereas the magnitude is handled by a separate learnable parameter. This can improve the performance of LoRA, especially at low ranks. Right now, DoRA only supports linear and Conv2D layers. DoRA introduces a biggeroverhead than pure LoRA, so it is recommended to merge weights for inference.use_doraaIf not None, enable 'Activated LoRA' (aLoRA), with alora_invocation_tokens being the tokenized invocation string for the adapter (must be present in all model input strings). This technique selectively activates the adapter weights only on tokens during and after the alora_invocation_tokens. When used in a CausalLM, this means that the KV cache prior to invocation is interchangeable with that of the base model (and other aLoRA adapters operating this way). As a result, in inference pipelines involving switching between base model inference and adapter inference (e.g. agentic pipelines, see paper for examples), significant savings are realized (relative to LoRA) by saving prefill operations. Overall adapter inference speedups of an order of magnitude or more can occur on vLLM, depending on the length of the shared context. Note that merging is not possible due to the selective application of the weights.zOptional[list[int]]alora_invocation_tokensarIt is only implemented in GPTQ for now. Enable Quantization-Aware Low-Rank Adaptation (QALoRA).This technique combines quantization-aware training with LoRA to improve performance for quantized models. This can improve the performance of LoRA, especially at low ranks. Right now, QALoRA only supports linear layers. use_qaloraa{Group size parameter for QALoRA pooling, controlling the dimension reduction factor. Input dimensions are pooled into groups of this size, reducing the computational cost. Higher values provide more compression but may reduce model quality. This parameter determines how many original features are averaged together to create one pooled feature. Only used when `use_qalora=True`.qalora_group_sizea(This enables using LoRA to effectively expand a transformer model to a larger size by repeating some layers. The transformation handles models (currently Llama, Bert or Falcon compatible architectures) with a module list in the model which it modifies to expand the number of modules. Base weights are shared so the memory usage is close to the original model. The intended use is these base weights remain fixed during finetuning but each layer has a separate LoRA adapter so the layers can be specialed via the adapter layers fit during fine tuning.The format is a list of [start, end) pairs which specify the layer ranges to stack. For example: Original model has 5 layers labelled by their position in the model: `[0, 1, 2, 3, 4]` layer_replication: `[[0, 4], [2, 5]]` Final model will have this arrangement of original layers: `[0, 1, 2, 3, 2, 3, 4]` This format is based on what is used for pass-through merges in mergekit. It makes it simple to select sequential ranges of a model and stack them while reusing layers at either end of each sequence.zOptional[list[tuple[int, int]]]layer_replicationzRuntime configurationsr runtime_configzWhether to enable the bias term for the LoRA B parameter. Typically, this should be disabled. The main use case for this is when the LoRA weights were extracted from fully fine-tuned parameters so the bias of those parameters can be taken into account. lora_biasa@List of parameter names or regex expression of the parameter names to replace with LoRA. This argument behaves similarly to `target_modules`, except that the parameter name should be passed. Generally, you should use `target_modules` to target the module (e.g. `nn.Linear`). However, in some circumstances, this is not possible. E.g., in many mixture of expert (MoE) layers in HF Transformers, instead of using `nn.Linear`, an `nn.Parameter` is used. PEFT normally overwrites the `forward` method for LoRA, but for `nn.Parameter`, there is none. Therefore, to apply LoRA to that parameter, it needs to be targeted with `target_parameters`. As an example, for Llama4, you can pass: `target_parameters=['feed_forward.experts.gate_up_proj', 'feed_forward.experts.down_proj]`. Passing a string for regex matching is not implemented yet.target_parametersz9The necessary config to apply arrow routing on the model.zOptional[ArrowConfig] arrow_configzWhether to tie weights or not after peft initialization. This will ensure that the adapters added to the tied layers are also tied. This is only applicable for layers passed via `modules_to_save`.ensure_weight_tyingcF>[TU]5nURS5 U$)zc Returns the configuration for your adapter model as a dictionary. Removes runtime configurations. rr)superto_dictpop)r8rv __class__s rryLoraConfig.to_dicts#W_    rcL >[TU]5 [RUl[ UR [5(a[UR 5O UR Ul[ UR[5(a[UR5O URUl UR(aSUl [ UR[5(a [S5e[ UR [5(aURb [!S5e[ UR [5(aUR"b [!S5eUR"(aUR(d [!S5eUR$(aUR&(a [!S5eUR(S:XaSSKnUR,R/S5(d [1S 5eUR2(d [!S 5e[ UR2[45(d[7UR25UlGOUR2(a0Ul[8R:"S 5 OUR(S :Xa3UR<c&[8R:"S 5 [?5UlOUR(S :wa$UR<b[8R:"S5 OvUR(S:Xa3UR@c&[8R:"S5 [C5Ul O3UR(S:wa#UR@b[8R:"S5 URD(aEUR(S;a[!SUR(S35eUR$(a [!S5eURFb&URHS:wa[8R:"S5 URJ(aURL(dURN(aw[ UR([5(a UR(RQS5(d UR(S:XdUR(S:XaSn[8R:"U5 SUl)g)Nz6`target_parameters` must be a list of strings or None.zD`layers_to_transform` cannot be used when `target_modules` is a str.z?`layers_pattern` cannot be used when `target_modules` is a str.zRWhen `layers_pattern` is specified, `layers_to_transform` must also be specified. zADoRA does not support megatron_core, please set `use_dora=False`.loftqrscipyzMThe required package 'scipy' is not installed. Please install it to continue.zE`loftq_config` must be specified when `init_lora_weights` is 'loftq'.zU`loftq_config` specified but will be ignored when `init_lora_weights` is not 'loftq'.evazY`init_lora_weights` is 'eva' but `eva_config` is not specified. Using default EVA config.zQ`eva_config` specified but will be ignored when `init_lora_weights` is not 'eva'.cordaz_`init_lora_weights` is 'corda' but `corda_config` is not specified. Using default CorDA config.zU`corda_config` specified but will be ignored when `init_lora_weights` is not 'corda'.)TFzjThe argument lora_bias=True is only supported with init_lora_weights=True or False, got init_lora_weights=z instead.zQThe argument lora_bias=True is not supported for DoRA, please pass use_dora=False CAUSAL_LMz5aLoRA is currently only supported for CAUSAL_LM task.pissaoloraasUsing Rank-Stabilized LoRA with rank_pattern/alpha_pattern and post-training conversion of modified base weights PiSSA/CorDA/OLoRA means that you won't be able to pass `path_initial_model_for_weight_conversion` to `save_pretrained` to restore the initial values of the base weights; if you intend to do this, please ensure not to use rslora or rank_pattern/alpha_pattern.)*rxr9r LORA peft_type isinstancerWlistsetrXrvmodules_to_tiertstr TypeErrorrar6rbrlrfr` importlibutil find_spec ImportErrorridictvarswarningswarnrjr<rkrHrsrm task_typer^rdre startswith_custom_modules)r8rmsgr|s rr9LoraConfig.__post_init__so !(243F3F(M(MC## $SWSfSf *4D4H4H$)O)OC$$ %UYUiUi   # #"&D  d,,c 2 2TU U d))3 / /D4L4L4Xcd d d))3 / /D4G4G4S^_ _   t'?'?qr r ==T11`a a  ! !W , >>++G44!"qrr$$ !hiid//66$():):$;!    "D  MMq r  # #u ,1H MMu v'kDO  # #u ,1L MMm n  # #w .43D3D3L MMq !, D   # #w .43D3D3P MMq r >>%%]: ))-)?)?(@ K}} !tuu  ' ' 3+8U MMQ R OO""d&8&8D22C88d>T>T>_>_`g>h>h**g5**g5y  MM# QUrcbURc0UlURRU5 g)a Experimental API to support providing custom LoRA layers. This API is subject to change, you should carefully read the docs before deciding to use it: https://huggingface.co/docs/peft/developer_guides/custom_models To register custom LoRA module types, call this method with a `mapping` argument that is a dict that maps from the target layer type to the custom LoRA layer type. The dict can contain multiple items if you wish to target multiple layer types. The target layer type can be any nn.Module that we currently don't support in PEFT, whether that is an official PyTorch layer type or a custom layer type. The custom LoRA module class has to be implemented by the user and follow the PEFT conventions for LoRA layers. N)rupdate)r8mappings r_register_custom_module"LoraConfig._register_custom_module s,    '#%D  ##G,r)rrkrjrXrirrrW)rz&dict[type[nn.Module], type[nn.Module]]returnNone),rrrrrrrVrrWrXrYrZr[r]r^r_r`rarbrrdrerfrgrhrirjrkrlrmrnrprqr rrrsrtrurvryr9rr __classcell__)r|s@rrSrSs1FP10J'K LAsL6; '  7N38=no8O4A0FGJGv~6NOL%O ijND16&*c!d1D - ,  J ,1 v ,O(  U   x4', _  'O^ $) M  $M= QV A   QM .3 i .L*', D 'J#+0 E +L' `  Hd 49  .  40" Z  J # H  s :?  h :6&).)V=U4V)N% J  It .3 D  .* +0(cd+L'!& %  ! ]V~--rrS) __future__rr dataclassesrrtypingrrrtorchr peft.configr peft.utilsr r r r(r<rHrSrrrrs#(++"  2 ^^ ^$ .K.K .Kb /C/C /Cd HH HV d-d- d-r