a Adh@stUddlZddlZddlmZddlZddlmZmZejdkrLddlm Z e dddZ d d Z d d Z d dZddZddZddZeededZejed<dZejed<dZejed<dZejed<dZejed<dZejed<ejZejed<ejZejed<ejZejed<dZd d!Zd"d#Z dZ!d.d$d%Z"d/d&d'Z#Gd(d)d)e$Z%d*d+Z&ed,d-Z'dS)0N) lru_cache)errorsshareddarwin) mac_specific)returncCstjdkrdStjSdS)NrF)sysplatformrhas_mpsr r 7/var/www/html/stable-diffusion-webui/modules/devices.pyr s r cCstjjdurdtjjSdS)Nzcuda:cuda)rcmd_opts device_idr r r r get_cuda_device_strings rcCstjrtStrdSdS)Nmpscpu)torchr is_availablerr r r r r get_optimal_device_names  rcCs ttSN)rdevicerr r r r get_optimal_device$srcCs|tjjvrtStSr)rruse_cpurr)taskr r r get_device_for(s rcCs^tjrLtjt$tjtjWdn1sB0YtrZt dSr) rr rrr empty_cache ipc_collectr rZ torch_mps_gcr r r r torch_gc/s   (rcCsLtjrHtddtdtjDr2dtjj_dtjjj _ dtjj_ dS)Ncss|]}tj|dkVqdS))N)rr get_device_capability).0Zdevidr r r ?zenable_tf32..rT) rr ranyrange device_countbackendscudnn benchmarkmatmul allow_tf32r r r r enable_tf32:s    r-z Enabling TF32rrdevice_interrogate device_gfpgan device_esrgandevice_codeformerdtype dtype_vae dtype_unetFcCstr|tS|Sr)unet_needs_upcasttor4inputr r r cond_cast_unetTsr9cCstr |S|Sr)r5floatr7r r r cond_cast_floatXsr;cCs4|r tSttjks"tjjdkr*tStdS)Nfullr ) contextlib nullcontextr2rfloat32rr precisionautocastdisabler r r rA_s rAcCs"tr|stjdddStS)Nr F)enabled)ris_autocast_enabledrAr=r>rBr r r without_autocastisrFc@s eZdZdS) NansExceptionN)__name__ __module__ __qualname__r r r r rGmsrGcCstjjr dStt|s$dS|dkrBd}tjjsl|d7}n*|dkrhd}tjjsltjjsl|d7}nd}|d7}t |dS) NZunetz,A tensor with all NaNs was produced in Unet.a$ This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this.vaez+A tensor with all NaNs was produced in VAE.z This could be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this.z$A tensor with all NaNs was produced.zD Use --disable-nan-check commandline argument to disable this check.) rrdisable_nan_checkrallisnanitemno_half no_half_vaerG)xwheremessager r r test_for_nansqs  rUcCsftdtt}tjddtt}||tdtt}tjdddtt}||dS)z just do any calculation with pytorch layers - the first time this is done it allocaltes about 700MB of memory and spends about 2.7 seconds doing that, at least wih NVidia. )rVrV)rVrVrW)rWrWN)rzerosr6rr2nnLinearConv2d)rRlinearconv2dr r r first_time_calculations r^)F)F)(rr= functoolsrrmodulesrrr rboolr rrrrrr-runrr__annotations__r.r/r0r1float16r2r3r4r5r9r;Znv_rngrArF ExceptionrGrUr^r r r r sB