>:j LSSKJrJr SSKJr SSKJr \"SS\55rg)) dataclassfield)PromptLearningConfig)PeftTypecJ^\rSrSr%Sr\"SSS0S9r\\S'U4Sjr S r U=r $) CartridgeConfigu- Configuration for CARTRIDGE, a KV-cache-parameterized prefix adapter. This is similar to prefix-tuning in how it is served (as `past_key_values`), but it stores the KV cache directly as trainable parameters instead of learning it via an MLP projection. Initialization: The Cartridges paper discusses multiple initialization options. In PEFT, initialization is a *separate* step from constructing the adapter config: - **Random KV initialization (paper option 2)**: Create the adapter via `get_peft_model(...)`. The CARTRIDGE prompt encoder parameters are randomly initialized by PyTorch. - **KV derived from the first tokens of a prompt/corpus (paper option 3)**: Run a no-grad prefill on the *base model* and copy the first `num_virtual_tokens` cached KV tokens into the adapter. PEFT provides utilities for this (importable from `peft` or from `peft.tuners.cartridge.utils`): - `initialize_kv_prefix_from_text(model, tokenizer, text=...)` - `initialize_kv_prefix_from_past_key_values(model, past_key_values=...)` If you already have a flattened KV-prefix tensor, you can load it directly via the prompt encoder’s `load_prompt_embeddings(...)` method. Args: num_frozen_tokens (`int`, defaults to 1): Number of *prefix* tokens at the start of the cartridge to keep frozen (no gradients). The Cartridges paper recommends freezing the first token as an attention sink for stability (set this to `1`), as many LLMs use early tokens as attention sinks and changing them can harm training. helpzNumber of initial virtual tokens to freeze (no gradients). The paper recommends freezing the first token as an attention sink for stability.)defaultmetadatanum_frozen_tokensc8>[TU]5 URS:a[SURS35eURb@URUR:a&[SURSURS35e[ R Ulg)Nrz&`num_frozen_tokens` must be >= 0, got .zK`num_frozen_tokens` must be <= `num_virtual_tokens`, got num_frozen_tokens=z and num_virtual_tokens=)super __post_init__r ValueErrornum_virtual_tokensr CARTRIDGE peft_type)self __class__s Y/home/mande/repo/quber/.venv/lib/python3.13/site-packages/peft/tuners/cartridge/config.pyrCartridgeConfig.__post_init__?s   ! !A %EdF\F\E]]^_` `  # # /d6L6LtOfOf6f]^b^t^t]uv**.*A*A)B!E "++)r) __name__ __module__ __qualname____firstlineno____doc__rrint__annotations__r__static_attributes__ __classcell__)rs@rrrs6<# < s , ,rrN) dataclassesrr peft.configr peft.utilsrrrrr)s,), 4,*4, 4,r