qi rddlmZddlZddlZddlmZddlmZmZddl m Z erddl m Z d dZ d Zd Zd Zy) ) annotationsN) TYPE_CHECKING)is_torch_available strtobool)QuantizationMethod)nnctsyddl}|jjsyddl}t ||jj jxs t|ddS)NFr_is_fsdp_managed_module) rtorch distributed is_availabletorch.distributed.fsdp isinstancefsdpFullyShardedDataParallelgetattr)moduler s P/opt/pipecat/venv/lib/python3.12/site-packages/transformers/integrations/fsdp.pyis_fsdp_managed_modulers]      ) ) +! fe//44MM N RY)5ScDtrddl}|jjxrv|jj xrZt t jjdddk(xr,t t jjdddk(Sy)NrACCELERATE_USE_FSDPFalseFSDP_CPU_RAM_EFFICIENT_LOADINGF) rr r ris_initializedrosenvironget)r s ris_fsdp_enabledr!,s    * * , Z!!002 Z"**..)>HIQN Z"**..)I7STXYY   rcnddlm}dttj|j vrddiSiS)z Returns checkpoint kwargs for FSDP model saving. Checks if the `adapter_only` parameter is supported by `save_fsdp_model` from accelerate and returns the appropriate kwargs. rsave_fsdp_model adapter_onlyT)accelerate.utilsr$listinspect signature parametersr#s rget_fsdp_ckpt_kwargsr+:s51g//@KKLL%% rcddlm}ddlm}t |j |r!|||j j_t|ddtjk(rq|jjjjrF|j jj!|jjjdyyy)aG Updates the FSDP plugin for PEFT LoRA/QLoRA compatibility. When using FSDP with PEFT LoRA, the auto wrap policy needs to be updated to additionally wrap LoRA trainable layers separately. When using FSDP with QLoRA, the mixed precision policy needs to be updated to use the quantization storage data type. r) PeftConfig)fsdp_auto_wrap_policyquantization_methodNT)override)peftr-peft.utils.otherr.ractive_peft_configstate fsdp_pluginauto_wrap_policyrrBITS_AND_BYTES hf_quantizerquantization_configbnb_4bit_quant_storageis_floating_pointset_mixed_precision)model acceleratorr-r.s rupdate_fsdp_plugin_peftr?Is 6%**J79Nu9U %%6,d37I7X7XX    2 2 I I [ [%%99    2 2 I ITX :  \ Yr)rz nn.Modulereturnbool) __future__rr(rtypingrutilsrrutils.quantization_configrr r rr!r+r?rrrGs5# 1:    r