§
    l¬¸jQÚ  ã                  óø  — d dl mZ d dlZd dlZd dlZd dlZd dlZd dlZd dlm	Z	 d dl
mZ d dlmZ d dlmZ d dlmZ d dlZd dlZd dlmZ d dlmc mZ d dlmZ d d	lmZ d d
lmZ d dlm Z m!Z!m"Z"m#Z#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z*m+Z+ d dl,m-Z-m.Z. d dl/m0Z0 d dl1m2Z2  e.e(d¦  «        Z3 e.e(d¦  «        Z4 e.e(d¦  «        Z5 e.e(d¦  «        Z6 e.e(d¦  «        Z7 e.e(d¦  «        Z8 e.e(d¦  «        Z9 e.e(d¦  «        Z: e.e(d¦  «        Z; e.e(d¦  «        Z< e.e(d¦  «        Z= e.e(d¦  «        Z> e.e(d¦  «        Z? e.e(d¦  «        Z@ e.e(d¦  «        ZA e.e(d¦  «        ZB e.e)d¦  «        ZC e.e)d ¦  «        ZD e.e)d!¦  «        ZE e.e)d"¦  «        ZFe*r e.e(d#¦  «        r e$jG        d$¦  «         d{d'„ZHe	d|d*„¦   «         ZId}d,„ZJd~dd1„ZKejL        d2„ ¦   «         ZMejL        d3„ ¦   «         ZNd€d�d7„ZOd‚d9„ZPdƒd„d;„ZQd<„ ZRd=„ ZSd>d?gZTd…d†dD„ZUd‡dF„ZVdˆdH„ZWd‰dJ„ZXdŠdM„ZYd‹dO„ZZdP„ Z[dQ„ Z\dR„ Z]dS„ Z^dŒdT„Z_dU„ Z`d�dX„ZadŽdZ„Zbd�d[„Zcd\„ Zdd�d^„Zed‘da„Zfd’db„Zgdc„ Zh G dd„ de¦  «        Zid“d”dk„Zjdl„ Zke	dƒdm„¦   «         Zld•do„Zm G dp„ dq¦  «        Zn	 	 	 	 d–d—dz„ZodS )˜é    )ÚannotationsN)Úcontextmanager)Údeepcopy)Údatetime)ÚPath)ÚAny)Únn)Úfuse_conv_bn_weights)Ú__version__)ÚARM64ÚDEFAULT_CFG_DICTÚDEFAULT_CFG_KEYSÚ
LOCAL_RANKÚLOGGERÚMACOSÚNUM_THREADSÚPYTHON_VERSIONÚTORCH_VERSIONÚTORCHVISION_VERSIONÚWINDOWSÚcolorstr)Úcheck_requirementsÚcheck_version)ÚCPUInfo)Ú
torch_loadz1.9.0z1.10.0z1.11.0z1.13.0z2.0.0z2.1.0z2.2.0z2.3.0z2.4.0z2.5.0z2.7.0z2.8.0z2.9.0z2.10.0z2.12.0z2.13.0z0.10.0z0.11.0z0.13.0z0.18.0z==2.4.0z™Known issue with torch==2.4.0 on Windows with CPU, recommend upgrading to torch>=2.4.1 to resolve https://github.com/ultralytics/ultralytics/issues/15049Údeviceútorch.device | strc                óì   — t          | dt          | ¦  «                             d¦  «        d         ¦  «        }t          t          d¦  «        rt	          j        |¦  «        nt          t          |¦  «        S )z@Return the PyTorch module that owns the selected device backend.Útypeú:r   Úget_device_module)ÚgetattrÚstrÚsplitÚhasattrÚtorchr!   )r   Údevice_types     úm/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/ultralytics/utils/torch_utils.pyÚget_torch_device_backendr)   G   s_   € å˜& &­#¨f©+¬+×*;Ò*;¸CÑ*@Ô*@ÀÔ*CÑDÔD€KÝ3:½5ÐBUÑ3VÔ3VÐw�5Ô" ;Ñ/Ô/Ð/Õ\cÕdiÐkvÑ\wÔ\wÐwó    Ú
local_rankÚintc              #  óÐ  K  — t          j        ¦   «         ot          j        ¦   «         }|ot          j        ¦   «         dk    }|rK| dvrG|r2t          j        t
          j                             ¦   «         g¬¦  «        nt          j        ¦   «          dV — |rO| dk    rK|r2t          j        t
          j                             ¦   «         g¬¦  «        nt          j        ¦   «          dS dS dS )ziEnsure all processes in distributed training wait for the local master (rank 0) to complete a task first.Únccl>   r   éÿÿÿÿ)Ú
device_idsNr   )ÚdistÚis_availableÚis_initializedÚget_backendÚbarrierr&   ÚcudaÚcurrent_device)r+   ÚinitializedÚuse_idss      r(   Útorch_distributed_zero_firstr:   M   sá   è è € õ Ô#Ñ%Ô%Ð?­$Ô*=Ñ*?Ô*?€KØÐ:�dÔ.Ñ0Ô0°FÒ:€Gàð ^�z¨Ð0Ð0ØBIÐ]�Œ¥¤×!:Ò!:Ñ!<Ô!<Ð =Ð>Ñ>Ô>Ð>ÍtÌ|É~Ì~øØ	€E€E€EØð ^�z Q’�ØBIÐ]�Œ¥¤×!:Ò!:Ñ!<Ô!<Ð =Ð>Ñ>Ô>Ð>ÍtÌ|É~Ì~øè è ð^ð ^��r*   Tc                ó   ‡ — ˆ fd„}|S )zQApply or disable torch inference mode while supporting the minimum torch version.c                ó(  •‡ — ‰sDt           r!t          j        ‰ ¦  «        ˆ fd„¦   «         }|S  t          j        ¦   «         ‰ ¦  «        S t           rt          j        ¦   «         r‰ S  t          rt          j        nt          j        ¦   «         ‰ ¦  «        S )zLApply appropriate torch decorator for inference mode based on torch version.c                 óà   •— t          j        d¦  «        5  t          j        ¦   «         5   ‰| i |¤Žcd d d ¦  «         cd d d ¦  «         S # 1 swxY w Y   d d d ¦  «         d S # 1 swxY w Y   d S ©NF)r&   Úinference_modeÚno_grad)ÚargsÚkwargsÚfns     €r(   Údisablez7smart_inference_mode.<locals>.decorate.<locals>.disableb   s  ø€ åÔ-¨eÑ4Ô4ð 3ð 3µe´m±o´oð 3ð 3Ø!˜r 4Ð2¨6Ð2Ð2ð3ð 3ð 3ð 3ñ 3ô 3ð 3ð 3ð 3ð 3ð 3ñ 3ô 3ð 3ð 3ð 3ð 3ð 3ð 3øøøð 3ð 3ð 3ð 3ð 3ð 3ð 3ñ 3ô 3ð 3ð 3ð 3ð 3ð 3ð 3ð 3øøøð 3ð 3ð 3ð 3ð 3ð 3s3   –A#ªA²A#ÁA	ÁA#ÁA	ÁA#Á#A'Á*A')Ú	TORCH_1_9Ú	functoolsÚwrapsr&   r@   Úis_inference_mode_enabledÚ
TORCH_1_10r?   )rC   rD   Úmodes   ` €r(   Údecoratez&smart_inference_mode.<locals>.decorate]   s£   øø€ àð 		'Ýð å” Ñ$Ô$ð3ð 3ð 3ð 3ñ %Ô$ð3ð �Ø"•5”=‘?”? 2Ñ&Ô&Ð&Ýð 	Q�Ô8Ñ:Ô:ð 	QØˆIàL­JÐI•EÔ(Ð(½E¼MÑLÔLÈRÑPÔPÐPr*   © )rJ   rK   s   ` r(   Úsmart_inference_moderM   Z   s(   ø€ ðQð Qð Qð Qð Qð" €Or*   r6   Úenabledúbool | torch.dtyper#   c                óÜ  — t          | t          j        ¦  «        r| nd}t          | ¦  «        } |t          j        u r«|dk    ot
          }|r‹t          r t          j                             d¬¦  «        nct          j                             ¦   «         oEt          t          j	        j
        ¦  «        p't          j                             ¦   «         d         dk    }|st          d¦  «        ‚|�d|ini }|d	k    rddl} |j        j        j        dd
| i|¤ŽS t
          r+|dk    rt"          sd\  }} t          j        j        |fd
| i|¤ŽS t          j        j                             | o|dk    ¦  «        S )aS  Get the appropriate autocast context manager based on PyTorch version and AMP setting.

    This function returns a context manager for automatic mixed precision (AMP) training that is compatible with both
    older and newer versions of PyTorch. It handles the differences in the autocast API between PyTorch versions.

    Args:
        enabled (bool | torch.dtype): Whether to enable AMP, or the autocast dtype to enable.
        device (str, optional): Device type to use for autocast, e.g. "cuda" or "npu".

    Returns:
        (torch.amp.autocast): The appropriate autocast context manager.

    Examples:
        >>> with autocast(enabled=True):
        ...     # Your mixed precision operations here
        ...     pass

    Notes:
        Uses `torch.amp.autocast` on torch>=1.13 and the backend-specific AMP context on older releases.
    Nr6   F)Úincluding_emulationr   é   zLbfloat16 autocast requires CUDA with native bfloat16 support and torch>=1.13ÚdtypeÚnpurN   Úmps)ÚcpuFrL   )Ú
isinstancer&   rS   ÚboolÚbfloat16Ú
TORCH_1_13Ú	TORCH_2_4r6   Úis_bf16_supportedÚversionÚhipÚget_device_capabilityÚRuntimeErrorÚ	torch_npurT   ÚampÚautocastÚ	TORCH_2_5)rN   r   rS   Úbf16_supportedrB   ra   s         r(   rc   rc   q   s‚  € õ* " '­5¬;Ñ7Ô7ÐAˆGˆG¸T€EÝ�7‰mŒm€GØ•”ÐÐØ 6Ò)Ð8­jˆØð 	õ ð\•”
×,Ò,ÀÐ,ÑGÔGÐGå”Z×1Ò1Ñ3Ô3ð \Ý�%œ-Ô+Ñ,Ô,ÐZµ´
×0PÒ0PÑ0RÔ0RÐSTÔ0UÐYZÒ0Zð	 ð ð 	oÝÐmÑnÔnÐnØ!&Ð!2ˆg�uÐÐ¸€FØ�‚€ØÐÐÐà)ˆyŒ}Ô Ô)ÐDÐD°'ÐD¸VÐDÐDÐDÝð EØ�UŠ?ˆ?¥9ˆ?Ø*‰OˆF�GÝŒyÔ! &ÐDÐD°'ÐD¸VÐDÐDÐDåŒzŒ~×&Ò& wÐ'C°6¸VÒ3CÑDÔDÐDr*   c                 ó(   — t          j        ¦   «         S )z=Return a string with system CPU information, i.e. 'Apple M2'.)r   ÚnamerL   r*   r(   Úget_cpu_inforh       s   € õ Œ<‰>Œ>Ðr*   c                ól   — t           j                             | ¦  «        }|j        › d|j        dz  d›d�S )zGReturn a string with system GPU information, i.e. 'Tesla T4, 15102MiB'.ú, i   z.0fÚMiB)r&   r6   Úget_device_propertiesrg   Útotal_memory)ÚindexÚ
propertiess     r(   Úget_gpu_inforp   ¦   s=   € õ ”×1Ò1°%Ñ8Ô8€JØŒoÐMÐM Ô!8¸GÑ!DÐMÐMÐMÐMÐMr*   Ú ú'str | int | list | tuple | torch.deviceÚreturnc                ó¼  ‡
— t          | t          j        ¦  «        rI| j        dk    r	| j        €dS | j        dk    rdS | j        dv r| j        €| j        n| j        › d| j        › �S t          | ¦  «                             ¦   «         } dD ]}|                      |d¦  «        } Œ| dk    rd} d	D ]—}|                      |¦  «        r€| t          |¦  «        d…          
                    d¦  «                             |› d�d¦  «        }d
                     d„ |                     d
¦  «        D ¦   «         ¦  «        }|r|› d|› �n|c S Œ˜d
                     d„ |                      d
¦  «        D ¦   «         ¦  «        } t          j                             dd¦  «                             dd¦  «        }d„ |                     d
¦  «        D ¦   «         dt          j                             ¦   «         …         Š
d„ |                      d
¦  «        D ¦   «         }|rht%          ˆ
fd„|D ¦   «         ¦  «        rMt'          d„ |D ¦   «         ¦  «        r4d
                     ˆ
fd„|                      d
¦  «        D ¦   «         ¦  «        } d| v rãddlm} |                      d
¦  «        }‰
rd„ ‰
D ¦   «         nd} |¦   «                              |                     d¦  «        d|¬¦  «        }‰
rˆ
fd„|D ¦   «         n|}t1          t          |¦  «        ¦  «        D ]7}	||	         dk    r)|r"t          |                     d¦  «        ¦  «        nd||	<   Œ8d
                     d„ |D ¦   «         ¦  «        } | S )a/  Parse a device request of any form into a canonical device string.

    Args:
        device (str | int | list | tuple | torch.device, optional): Device request, e.g. 'cuda:0', '0,1', [0, 1], 'cpu',
            'mps', or '-1' to auto-select an idle GPU ('-1,-1' for two).

    Returns:
        (str): Canonical device string, e.g. '', 'cpu', 'mps', '0', or '0,1'.

    Examples:
        >>> parse_device("cuda:0")
        '0'

        >>> parse_device([0, 1])
        '0,1'

    Notes:
        Each '-1' is replaced with an idle GPU index. Requested ids exceeding the torch device count that match
        physical GPU ids visible under an external CUDA_VISIBLE_DEVICES restriction are translated to the
        corresponding torch indices, e.g. '3' -> '0' when CUDA_VISIBLE_DEVICES='3'; in-range ids are always torch
        indices, keeping parsing idempotent. Returned indices are relative to the active restriction, so strings
        persisted under one environment (e.g. resumed checkpoint args) address the same physical GPUs only in that
        environment.
    r6   Nrq   rV   >   rT   Úxpur    )úcuda:Únoneú(ú)ú[ú]ú'ú Ú0©rT   ru   ú,c              3  ó~   K  — | ]8}|¯|                      ¦   «         rt          t          |¦  «        ¦  «        n|V — Œ9d S ©N©Úisdigitr#   r,   ©Ú.0Úxs     r(   ú	<genexpr>zparse_device.<locals>.<genexpr>Õ   sE   è è € ÐbÐbÀQÐ`aÐb¨a¯iªi©k¬kÐ@�s¥3 q¡6¤6™{œ{˜{¸qÐbÐbÐbÐbÐbÐbr*   c              3  ó~   K  — | ]8}|¯|                      ¦   «         rt          t          |¦  «        ¦  «        n|V — Œ9d S r‚   rƒ   r…   s     r(   rˆ   zparse_device.<locals>.<genexpr>×   sE   è è € ÐXÐX¸AÐVWÐX Q§Y¢Y¡[¤[Ð7•c�#˜a™&œ&‘k”k�k°aÐXÐXÐXÐXÐXÐXr*   ÚCUDA_VISIBLE_DEVICESr}   c                óv   — g | ]6}|¯|                      ¦   «         rt          t          |¦  «        ¦  «        n|‘Œ7S rL   rƒ   r…   s     r(   ú
<listcomp>z parse_device.<locals>.<listcomp>Û   s:   € ÐNÐNÐN°QÈAÐN˜aŸiši™kœkÐ0�s•3�q‘6”6‰{Œ{ˆ{¨qÐNÐNÐNr*   c                ó:   — g | ]}|                      ¦   «         ¯|‘ŒS rL   )r„   r…   s     r(   rŒ   z parse_device.<locals>.<listcomp>Ü   s%   € Ð;Ð;Ð;�Q¨q¯yªy©{¬{Ð;ˆqÐ;Ð;Ð;r*   c              3  ó    •K  — | ]}|‰v V — Œ	d S r‚   rL   ©r†   r‡   Úvisibles     €r(   rˆ   zparse_device.<locals>.<genexpr>Ý   s'   øè è € Ð5Ð5¨�q˜G�|Ð5Ð5Ð5Ð5Ð5Ð5r*   c              3  ót   K  — | ]3}t          |¦  «        t          j                             ¦   «         k    V — Œ4d S r‚   )r,   r&   r6   Údevice_countr…   s     r(   rˆ   zparse_device.<locals>.<genexpr>Ý   s<   è è € Ð=sÐ=sÐfg½cÀ!¹f¼fÍÌ
×H_ÒH_ÑHaÔHaÒ>aÐ=sÐ=sÐ=sÐ=sÐ=sÐ=sr*   c              3  óˆ   •K  — | ]<}|                      ¦   «         r"t          ‰                     |¦  «        ¦  «        n|V — Œ=d S r‚   )r„   r#   rn   r�   s     €r(   rˆ   zparse_device.<locals>.<genexpr>à   sI   øè è € ÐaÐaÈ!°1·9²9±;´;ÐE�#˜gŸmšm¨AÑ.Ô.Ñ/Ô/Ð/ÀAÐaÐaÐaÐaÐaÐar*   z-1r   )ÚGPUInfoc                óT   — g | ]%}|                      ¦   «         ¯t          |¦  «        ‘Œ&S rL   ©r„   r,   r…   s     r(   rŒ   z parse_device.<locals>.<listcomp>ç   s+   € Ð=Ð=Ð= °·²±´Ð=•c˜!‘f”fÐ=Ð=Ð=r*   gš™™™™™É?)ÚcountÚmin_memory_fractionÚindicesc                óT   •— g | ]$}‰                      t          |¦  «        ¦  «        ‘Œ%S rL   )rn   r#   r�   s     €r(   rŒ   z parse_device.<locals>.<listcomp>é   s+   ø€ Ð<Ð<Ð<¨a�G—M’M¥# a¡&¤&Ñ)Ô)Ð<Ð<Ð<r*   c              3  ó   K  — | ]}|¯|V — Œ	d S r‚   rL   )r†   Úps     r(   rˆ   zparse_device.<locals>.<genexpr>í   s'   è è € Ð0Ð0 ¨aÐ0˜!Ð0Ð0Ð0Ð0Ð0Ð0r*   )rW   r&   r   r   rn   r#   ÚlowerÚreplaceÚ
startswithÚlenÚlstripÚjoinr$   ÚosÚenvironÚgetr6   r’   ÚallÚanyÚultralytics.utils.autodevicer”   Úselect_idle_gpur—   ÚrangeÚpop)r   ÚremoveÚbackendr™   Úcvdr”   ÚpartsÚ
candidatesÚselectedÚir�   s             @r(   Úparse_devicer³   ­   sÁ  ø€ õ2 �&�%œ,Ñ'Ô'ð \ØŒ;˜&Ò Ð  V¤\Ð%9Ø�2ØŒ;˜%ÒÐØ�5ØŒ;˜.Ð(Ð(Ø"(¤,Ð"6�6”;�;¸v¼{Ð<[Ð<[ÈVÌ\Ð<[Ð<[Ð[Ý�‰[Œ[×ÒÑ Ô €FØ?ð ,ð ,ˆØ—’ ¨Ñ+Ô+ˆˆØ�ÒÐØˆØ!ð Bð BˆØ×Ò˜WÑ%Ô%ð 	BØ�S ™\œ\˜^˜^Ô,×3Ò3°CÑ8Ô8×@Ò@ÀGÀÀÀÐPRÑSÔSˆGØ—h’hÐbÐbÈ'Ï-Ê-ÐX[ÑJ\ÔJ\ÐbÑbÔbÑbÔbˆGØ-4ÐA�gÐ)Ð) Ð)Ð)Ð)¸'ÐAÐAÐAð	Bð �XŠXÐXÐXÀÇÂÈcÑARÔARÐXÑXÔXÑXÔX€Fõ Œ*�.Š.Ð/°Ñ
4Ô
4×
<Ò
<¸SÀ"Ñ
EÔ
E€CØNÐN¸#¿)º)ÀC¹.¼.ÐNÑNÔNÐOjÕQVÔQ[×QhÒQhÑQjÔQjÐOjÔk€GØ;Ð;˜&Ÿ,š, sÑ+Ô+Ð;Ñ;Ô;€GØð b•3Ð5Ð5Ð5Ð5¨WÐ5Ñ5Ô5Ñ5Ô5ð b½#Ð=sÐ=sÐkrÐ=sÑ=sÔ=sÑ:sÔ:sð bð —’ÐaÐaÐaÐaÈvÏ|Ê|Ð\_ÑO`ÔO`ÐaÑaÔaÑaÔaˆØˆv€~€~Ø8Ð8Ð8Ð8Ð8Ð8ð —’˜SÑ!Ô!ˆØAHÐRÐ=Ð= gÐ=Ñ=Ô=Ð=Èdˆ
Ø�7‘9”9×,Ò,°5·;²;¸tÑ3DÔ3DÐZ]ÐgqÐ,ÑrÔrˆØ@GÐUÐ<Ð<Ð<Ð<°8Ð<Ñ<Ô<Ð<ÈXˆÝ•s˜5‘z”zÑ"Ô"ð 	Dð 	DˆAØ�QŒx˜4ÒÐØ3;ÐC�3˜xŸ|š|¨A™œÑ/Ô/Ð/À��a‘øØ—’Ð0Ð0 UÐ0Ñ0Ô0Ñ0Ô0ˆØ€Mr*   Fc                ó¦  ‡— t          | t          j        ¦  «        r| j        dvr| S n$t	          | ¦  «                             d¦  «        r| S dt          › dt          › dt          › d�}t          | ¦  «        } |                      d¦  «        �r¶|  
                    dd	¦  «        d
         }|dk    r'	 d
dl}n!# t          $ r t          d| › d�¦  «        ‚w xY wt          t          |¦  «        s(t          d|                     ¦   «         › d| › d�¦  «        ‚t!          |¦  «        }|                     ¦   «         s(t          d|                     ¦   «         › d| › d�¦  «        ‚| |k    rdgn| dd…          
                    d¦  «        }d„ |D ¦   «         }|rMt%          |¦  «        t%          |¦  «        k    s-t%          |¦  «        t%          t'          |¦  «        ¦  «        k    r1t          d|                     ¦   «         › d| › d|› d|› d|› d�¦  «        ‚|                     ¦   «         Št+          ˆfd„|D ¦   «         ¦  «        r+t          d|                     ¦   «         › d| › d‰› d�¦  «        ‚t%          |¦  «        d	k    r|                     |d
         ¦  «         |rŽdt%          |¦  «        z  }	t/          |¦  «        D ]B\  }
}||
d
k    rdn|	› |                     ¦   «         › d|› d|                     |¦  «        › d�z  }ŒCt3          j        |r|n|                     ¦   «         ¦  «         t          j        ||d
         ¦  «        S | d k    }| d!v }|sý|sû| rùt9          d"„ |  
                    d¦  «        D ¦   «         ¦  «        }t          j                             ¦   «         r|s­t3          j        |¦  «         t          j                             ¦   «         d
k    rd#nd}t          d$| › d%t          j                             ¦   «         › d&t          j                             ¦   «         › d't<          j                              d(¦  «        › d)|› �
¦  «        ‚|�s|�st          j                             ¦   «         rí| r|  
                    d¦  «        n+t	          t          j         !                    ¦   «         ¦  «        g}dt%          |¦  «        z  }	t/          |¦  «        D ]5\  }
}||
d
k    rdn|	› d*|› dtE          tG          |¦  «        ¦  «        › d�z  }Œ6d+|d
         › �}| rEt%          |¦  «        d	k    r2t          j                             tG          |d
         ¦  «        ¦  «         n[|rBtH          r;t          j%        j&                             ¦   «         r|d,tO          ¦   «         › d�z  }d-}n|d.tO          ¦   «         › d�z  }d }|d/v rt          j(        tR          ¦  «         |d k    r9tT          r2tV          r+tX          r$t          j%        j-         .                    d0¦  «         |r*t3          j        |r|n|                     ¦   «         ¦  «         t          j        |¦  «        S )1a  Select the appropriate PyTorch device based on the provided arguments.

    The function takes a string specifying the device or a torch.device object and returns a torch.device object
    representing the selected device. The function also validates the number of available devices and raises an
    exception if the requested device(s) are not available.

    Args:
        device (str | torch.device, optional): Device string or torch.device object. Options include 'cpu', 'cuda', '0',
            '0,1,2,3', 'mps', 'npu:0', 'npu:0,1', 'xpu:0', 'xpu:0,1', or '-1' for auto-select. Defaults to auto-selecting
            the first available GPU, or CPU if no GPU is available.
        newline (bool, optional): If True, adds a newline at the end of the log string.
        verbose (bool, optional): If True, logs the device information.

    Returns:
        (torch.device): Selected device.

    Examples:
        >>> select_device("cuda:0")
        device(type='cuda', index=0)

        >>> select_device("cpu")
        device(type='cpu')

    Notes:
        CUDA indices are torch device indices, which reflect any externally set CUDA_VISIBLE_DEVICES. This function
        never modifies CUDA_VISIBLE_DEVICES; an explicit single-GPU request is made the default CUDA device with
        torch.cuda.set_device() so that indexless 'cuda' operations land on it, while default '' requests (resolved
        to the current device) and multi-GPU requests (DDP ranks pin their own device in trainer._setup_ddp()) leave
        the current device untouched.
    >   rV   rT   ru   r6   )ÚtpuÚintelÚvulkanzUltralytics u    ðŸš€ Python-z torch-r}   r   r    é   r   rT   NzInvalid NPU 'device=z;'. Install 'torch_npu' at https://github.com/Ascend/pytorchzInvalid z	 'device=z&' requested. Backend is not available.r~   é   r€   c                óT   — g | ]%}|                      ¦   «         ¯t          |¦  «        ‘Œ&S rL   r–   r…   s     r(   rŒ   z!select_device.<locals>.<listcomp>)  s+   € Ð<Ð<Ð<˜a°·	²	±´Ð<•3�q‘6”6Ð<Ð<Ð<r*   z' format. Use 'z', 'z	:0', or 'z:0,1'.c              3  ó$   •K  — | ]
}|‰k    V — Œd S r‚   rL   )r†   ÚidxÚns     €r(   rˆ   z select_device.<locals>.<genexpr>0  s'   øè è € Ð+Ð+˜Cˆs�aŠxÐ+Ð+Ð+Ð+Ð+Ð+r*   z' requested. Only z device(s) available.rq   z (z)
rV   >   úmps:0rU   c              3  óœ   K  — | ]G}|                      ¦   «         o.t          |¦  «        t          j                             ¦   «         k     V — ŒHd S r‚   )r„   r,   r&   r6   r’   r…   s     r(   rˆ   z select_device.<locals>.<genexpr>A  sH   è è € ÐbÐbÈ1�A—I’I‘K”KÐF¥C¨¡F¤F­U¬Z×-DÒ-DÑ-FÔ-FÒ$FÐbÐbÐbÐbÐbÐbr*   z}See https://pytorch.org/get-started/locally/ for up-to-date torch install instructions if no CUDA devices are seen by torch.
zInvalid CUDA 'device=z˜' requested. Use 'device=cpu' or pass valid CUDA device(s) if available, i.e. 'device=0' or 'device=0,1,2,3' for Multi-GPU.

torch.cuda.is_available(): z
torch.cuda.device_count(): z%
os.environ['CUDA_VISIBLE_DEVICES']: rŠ   ú
zCUDA:rv   zMPS (rU   zCPU (>   rV   rU   F)/rW   r&   r   r   r#   rŸ   r   r   r   r³   r$   ra   ÚImportErrorÚ
ValueErrorr%   Úupperr)   r2   r    Úsetr’   r§   Ú
set_deviceÚ	enumerateÚget_device_namer   ÚinfoÚrstripr¦   r6   r£   r¤   r¥   r7   rp   r,   Ú	TORCH_2_0ÚbackendsrU   rh   Úset_num_threadsr   r   r   Ú	TORCH_2_3ÚnnpackÚ	set_flags)r   ÚnewlineÚverboseÚsr'   ra   r­   Ú	requestedr™   Úspacer²   r¼   rV   rU   ÚvalidÚinstallÚdevicesÚdÚargr½   s                      @r(   Úselect_devicerÚ   ñ   s  ø€ õ> �&�%œ,Ñ'Ô'ð ØŒ;Ð;Ð;Ð;ØˆMð <å	ˆV‰Œ×	Ò	Ð :Ñ	;Ô	;ð ØˆàX•{ÐXÐXµÐXÐXÍÐXÐXÐX€AÝ˜&Ñ!Ô!€Fà×Ò˜Ñ(Ô(ñ #5Ø—l’l 3¨Ñ*Ô*¨1Ô-ˆØ˜%ÒÐðØ Ð Ð Ð Ð øÝð ð ð Ý Øn¨6ÐnÐnÐnñô ð ðøøøõ •u˜kÑ*Ô*ð 	vÝÐt¨×(9Ò(9Ñ(;Ô(;ÐtÐtÀfÐtÐtÐtÑuÔuÐuÝ*¨;Ñ7Ô7ˆØ×#Ò#Ñ%Ô%ð 	vÝÐt¨×(9Ò(9Ñ(;Ô(;ÐtÐtÀfÐtÐtÐtÑuÔuÐuà# {Ò2Ð2�S�E�E¸¸q¸r¸r¼
×8HÒ8HÈÑ8MÔ8Mˆ	Ø<Ð< 9Ð<Ñ<Ô<ˆØð 	�#˜g™,œ,­#¨i©.¬.Ò8Ð8½CÀ¹L¼LÍCÕPSÐT[ÑP\ÔP\ÑL]ÔL]Ò<]Ð<]ÝðS˜;×,Ò,Ñ.Ô.ð Sð S¸ð Sð SØ#ðSð SØ)4ðSð SØ?JðSð Sð Sñô ð ð × Ò Ñ"Ô"ˆÝÐ+Ð+Ð+Ð+ 7Ð+Ñ+Ô+Ñ+Ô+ð 	ÝØk˜;×,Ò,Ñ.Ô.ÐkÐk¸ÐkÐkÐSTÐkÐkÐkñô ð õ ˆw‰<Œ<˜1ÒÐØ×Ò˜w qœzÑ*Ô*Ð*Øð 	6Ø�#˜a™&œ&‘LˆEÝ# GÑ,Ô,ð pð p‘��3Ø˜a 1šf˜f˜˜¨%Ðo°×1BÒ1BÑ1DÔ1DÐoÐoÀsÐoÐoÈg×NeÒNeÐfiÑNjÔNjÐoÐoÐoÑo��ÝŒK˜WÐ4˜˜¨!¯(ª(©*¬*Ñ5Ô5Ð5ÝŒ|˜K¨°¬Ñ4Ô4Ð4à
�EŠ/€CØ
Ð$Ð
$€CØð �sð ˜vð ÝÐbÐbÐPV×P\ÒP\Ð]`ÑPaÔPaÐbÑbÔbÑbÔbˆÝ”
×'Ò'Ñ)Ô)ð 	¨eð 	ÝŒK˜‰NŒNˆNõ ”:×*Ò*Ñ,Ô,°Ò1Ð1ð4ð 4ð ð	 õ ð¨ð ð õ 16´
×0GÒ0GÑ0IÔ0Iðð õ 16´
×0GÒ0GÑ0IÔ0Ið	ð õ
 :<¼¿ºÐH^Ñ9_Ô9_ðð ð ðð ñô ð ð ñ �sñ �uœz×6Ò6Ñ8Ô8ð Ø'-ÐU�&—,’,˜sÑ#Ô#Ð#µC½¼
×8QÒ8QÑ8SÔ8SÑ4TÔ4TÐ3UˆØ•c˜!‘f”f‘ˆÝ˜gÑ&Ô&ð 	Qð 	Q‰DˆAˆqØ˜!˜qš&˜&�B�B eÐPÐP°!ÐPÐPµ|ÅCÈÁFÄFÑ7KÔ7KÐPÐPÐPÑPˆAˆAØ"�g˜a”jÐ"Ð"ˆØð 	3•c˜'‘l”l aÒ'Ð'ÝŒJ×!Ò!¥# g¨a¤j¡/¤/Ñ2Ô2Ð2øØ	ð •ð �uœ~Ô1×>Ò>Ñ@Ô@ð à	Ð(•\‘^”^Ð(Ð(Ð(Ñ(ˆØˆˆà	Ð(•\‘^”^Ð(Ð(Ð(Ñ(ˆØˆà
ˆnÐÐÝÔ�kÑ*Ô*Ð*Ø
ˆe‚|€|�€|¥%€|­I€|ÝŒÔ×'Ò'¨Ñ.Ô.Ð.Øð 2ÝŒ˜Ð0�A�A a§h¢h¡j¤jÑ1Ô1Ð1ÝŒ<˜ÑÔÐs   Â/B4 Â4Cútorch.device | Nonec                óÐ   — | �	| j         dvrIt          | pd¦  «        }|                     ¦   «         r$t          |d¦  «        r|                     ¦   «          t          j        ¦   «         S )zReturn PyTorch-accurate time.N>   rV   rU   r6   Úsynchronize)r   r)   r2   r%   rÝ   ÚtimeÚperf_counter)r   Úaccelerators     r(   Ú	time_syncrá   m  sk   € à€~˜œ¨NÐ:Ð:Ý.¨vÐ/?¸Ñ@Ô@ˆØ×#Ò#Ñ%Ô%ð 	&­'°+¸}Ñ*MÔ*Mð 	&Ø×#Ò#Ñ%Ô%Ð%ÝÔÑÔÐr*   c           	     óœ  — t          | j        | j        |j        |j        |j        |j        |j        ¦  «        \  | _        | _        t          | dd¦  «        }|�`|j        |j        t          j	        |j        |j        z   ¦  «        z   
                    ¦   «                              |j        ¦  «        z  |_        |                      d¦  «        S )a­  Fuse Conv2d and BatchNorm2d layers for inference optimization.

    Args:
        conv (nn.Conv2d): Convolutional layer to fuse.
        bn (nn.BatchNorm2d): Batch normalization layer to fuse.

    Returns:
        (nn.Conv2d): The fused convolutional layer with gradients disabled.

    Examples:
        >>> conv = nn.Conv2d(3, 16, 3)
        >>> bn = nn.BatchNorm2d(16)
        >>> fused_conv = fuse_conv_and_bn(conv, bn)
    Úweight_quantizerNF)r
   ÚweightÚbiasÚrunning_meanÚrunning_varÚepsr"   Úamaxr&   ÚsqrtÚabsÚview_asÚrequires_grad_)ÚconvÚbnÚqs      r(   Úfuse_conv_and_bnrñ   v  s®   € õ 2ØŒ�T”Y ¤°´ÀÄÈÌÐTVÔT[ñô Ñ€D„K�”õ 	�Ð(¨$Ñ/Ô/€AØ€}Ø”˜2œ9¥u¤z°"´.À2Ä6Ñ2IÑ'JÔ'JÑJ×OÒOÑQÔQ×YÒYÐZ[ÔZ`ÑaÔaÑaˆŒØ×Ò˜uÑ%Ô%Ð%r*   c           	     óB  — t          |t          j        ¦  «        r|                      d¦  «        S | j        | j        j        c}^}}}  | j        j        |||z  |g|¢R Ž                      dd¦  «        j	        ||z  ||z  g|¢R Ž }t          || j        |j        |j        |j        |j        |j        ¦  «        \  }| _        t          j          |j        ||||z  g|¢R Ž                      dd¦  «        j	        ||g|¢R Ž ¦  «        | _        |                      d¦  «        S )að  Fuse ConvTranspose2d and BatchNorm2d layers for inference optimization.

    Args:
        deconv (nn.ConvTranspose2d): Transposed convolutional layer to fuse.
        bn (nn.BatchNorm2d): Batch normalization layer to fuse.

    Returns:
        (nn.ConvTranspose2d): The fused transposed convolutional layer with gradients disabled.

    Examples:
        >>> deconv = nn.ConvTranspose2d(16, 3, 3)
        >>> bn = nn.BatchNorm2d(3)
        >>> fused_deconv = fuse_deconv_and_bn(deconv, bn)
    Fr¸   é   )rW   r	   ÚIdentityrí   Úgroupsrä   ÚshapeÚviewÚ	transposeÚreshaper
   rå   ræ   rç   rè   Ú	Parameter)Údeconvrï   ÚgÚciÚcoÚkrä   s          r(   Úfuse_deconv_and_bnr   Ž  s@  € õ �"•b”kÑ"Ô"ð ,Ø×$Ò$ UÑ+Ô+Ð+ð ”m V¤]Ô%8€O€A€|ˆˆB�ØKÐˆVŒ]Ô  2¨¡7¨BÐ3°Ð3Ð3Ð3×=Ò=¸aÀÑCÔCÔKÈAÐPRÉFÐTVÐZ[ÑT[Ð`Ð^_Ð`Ð`Ð`€FÝ.Ø�”˜Rœ_¨b¬n¸b¼fÀbÄiÐQSÔQXñô Ñ€FˆFŒKõ ”LÐ!X  ¤¨Q°°B¸!±GÐ!@¸aÐ!@Ð!@Ð!@×!JÒ!JÈ1ÈaÑ!PÔ!PÔ!XÐY[Ð]_Ð!dÐbcÐ!dÐ!dÐ!dÑeÔe€F„MØ× Ò  Ñ'Ô'Ð'r*   znvidia-modelopt>=0.44Úhuggingface_hubrR   Úmodelú	nn.ModuleÚbatchesc                ó4  ‡‡‡‡— t          t          ¦  «        5  t          t          ¦  «         ddlmc m} ddd¦  «         n# 1 swxY w Y   ˆˆˆfd„}t          j        d‰› d�¦  «         | j	        }|  
                    ¦   «          t          j        ¦   «         5  |                     | |j        |¦  «        } dt          | j        ¦  «        dz
  › d�Š|                     | ˆfd	„¦  «         ddd¦  «         n# 1 swxY w Y   |                      |¦  «         | S )
až  Insert INT8 fake-quantization into a model for quantization-aware training (QAT).

    Swaps Conv and Linear layers for ModelOpt equivalents that fake-quantize their input and weight, so training learns
    weights that survive INT8 export and `torch.onnx.export` emits those ranges as Q/DQ nodes. Activation and weight
    ranges are calibrated once from `batches` batches and then held fixed (ModelOpt's INT8 config keeps `amax` as a
    buffer, not a learnable parameter), so training adapts the weights to them.

    Training keeps BatchNorm unfused; `fuse()` folds it at export and rescales the weight ranges along. The head's
    output layers, the bare convolutions and linears outside its `Conv` blocks, are left in float to limit INT8 accuracy
    loss.

    Args:
        model (nn.Module): Model to prepare, modified in place.
        dataloader (Iterable): Loader yielding Ultralytics batches for the initial range calibration.
        preprocess (Callable): Task trainer preprocessing applied to each calibration batch.
        batches (int): Number of calibration batches.

    Returns:
        (nn.Module): The prepared model, carrying fake-quantization modules.
    r   Nc                óv   •— t          ‰t          ‰¦  «        ¦  «        D ]\  }} |  ‰|¦  «        ¦  «         ŒdS )z&Calibrate through the task batch path.N)Úziprª   )ÚmÚbatchÚ_r  Ú
dataloaderÚ
preprocesss      €€€r(   Úforward_loopz!prepare_qat.<locals>.forward_loopÈ  sM   ø€ å˜J­¨g©¬Ñ7Ô7ð 	!ð 	!‰HˆE�1ØˆAˆjˆj˜ÑÔÑ Ô Ð Ð ð	!ð 	!r*   z0Preparing INT8 quantization-aware training from z calibration batches...zmodel.r¸   ú.c                ó>   •— |                       ‰¦  «        od| vpd| v S )Nz.conv.z.dfl.)rŸ   )r½   Úheads    €r(   ú<lambda>zprepare_qat.<locals>.<lambda>Õ  s*   ø€ ¨q¯|ª|¸DÑ/AÔ/AÐ/iÀxÐWXÐGXÐGhÐ\cÐghÐ\h€ r*   )r:   r   r   ÚMODELOPT_REQUIREMENTSÚmodelopt.torch.quantizationr&   Úquantizationr   rÈ   ÚtrainingÚevalr@   ÚquantizeÚINT8_DEFAULT_CFGr    r  Údisable_quantizerÚtrain)r  r  r  r  Úmtqr  r  r  s    ```   @r(   Úprepare_qatr  ¯  sÎ  øøøø€ õ* 
&¥jÑ	1Ô	1ð 2ð 2ÝÕ0Ñ1Ô1Ð1Ø1Ð1Ð1Ð1Ð1Ð1Ð1Ð1Ð1ð2ð 2ð 2ñ 2ô 2ð 2ð 2ð 2ð 2ð 2ð 2øøøð 2ð 2ð 2ð 2ð!ð !ð !ð !ð !ð !ð !õ
 „KÐcÀ7ÐcÐcÐcÑdÔdÐdØŒ~€HØ	‡J‚J�L„L€LÝ	Œ‰Œð kð kØ—’˜U CÔ$8¸,ÑGÔGˆð 0�˜EœKÑ(Ô(¨1Ñ,Ð/Ð/Ð/ˆØ×Ò˜eÐ%iÐ%iÐ%iÐ%iÑjÔjÐjðkð kð kñ kô kð kð kð kð kð kð køøøð kð kð kð kð 
‡K‚K�ÑÔÐØ€Ls$   ™AÁAÁ
AÂAC8Ã8C<Ã?C<rX   c                ó¼   — t          t          | dd¦  «        t          j        ¦  «        r| j        n| } t          d„ |                      ¦   «         D ¦   «         ¦  «        S )zÀReturn True if the model carries fake-quantization modules inserted by `prepare_qat`.

    Matched by class name so that non-QAT models, i.e. every ordinary export, never import ModelOpt.
    r  Nc              3  óF   K  — | ]}t          |¦  «        j        d k    V — ŒdS )ÚTensorQuantizerN)r   Ú__name__)r†   r  s     r(   rˆ   zis_qat.<locals>.<genexpr>à  s1   è è € ÐNÐN¸�t�A‰wŒwÔÐ#4Ò4ÐNÐNÐNÐNÐNÐNr*   )rW   r"   r	   ÚModuler  r§   Úmodules©r  s    r(   Úis_qatr$  Ú  sR   € õ
 &¥g¨e°W¸dÑ&CÔ&CÅRÄYÑOÔOÐZˆEŒKˆKÐUZ€EÝÐNÐN¸e¿mºm¹o¼oÐNÑNÔNÑNÔNÐNr*   údict[str, Any] | Nonec                óä   — t          | d| ¦  «        } t          | ¦  «        sdS ddlmc m} |                     | ¦  «        d„ |                      ¦   «                              ¦   «         D ¦   «         dœS )aa  Return the state that reproduces a model's fake-quantization, or None if it carries none.

    Ultralytics checkpoints are pickled modules, but ModelOpt builds its quantized layers as classes created at runtime,
    which pickle cannot look up on load. The quantization therefore travels beside the module as data. The checkpoint
    writers read it here and `restore_qat` reconstructs it at load and resume.

    Args:
        model (nn.Module): Model to read, left untouched.

    Returns:
        (dict | None): ModelOpt conversion state and the calibrated quantizer ranges, or None for a plain model.
    Ústudent_modelNr   c                ó"   — i | ]\  }}d |v ¯	||“ŒS )Ú	quantizerrL   ©r†   rÿ   Úvs      r(   ú
<dictcomp>zqat_state.<locals>.<dictcomp>÷  s)   € ÐSÐSÐS™D˜A˜qÀ+ÐQRÐBRÐBR�1�aÐBRÐBRÐBRr*   )ÚmodeloptÚranges)r"   r$  Úmodelopt.torch.optr&   ÚoptÚmodelopt_stateÚ
state_dictÚitems)r  Úmtos     r(   Ú	qat_stater5  ã  sŽ   € õ �E˜?¨EÑ2Ô2€EÝ�%‰=Œ=ð ØˆtØ$Ð$Ð$Ð$Ð$Ð$Ð$Ð$Ð$ð ×&Ò& uÑ-Ô-ØSÐS E×$4Ò$4Ñ$6Ô$6×$<Ò$<Ñ$>Ô$>ÐSÑSÔSðð ð r*   ÚNonec                ó8  — t          | d| ¦  «        } t          | ¦  «        sdS ddlm} ddlm} |                      ¦   «         D ]A}t          ||¦  «        r/|                     ¦   «          |j	         
                    dd¦  «         ŒB|                     | ¦  «         dS )a  Revert a model's fake-quantization in place, leaving the plain layers it wraps.

    Checkpoint writers call this on the copy they are about to pickle, after `qat_state` has read the quantization out
    of it, since the runtime-generated layer classes cannot be pickled.
    r'  Nr   )ÚModeloptStateManager)ÚDynamicModuleÚ_parallel_state)r"   r$  Úmodelopt.torch.opt.conversionr8  Úmodelopt.torch.opt.dynamicr9  r"  rW   ÚexportÚ__dict__r«   Úremove_state)r  r8  r9  r  s       r(   Ú	strip_qatr@  û  s´   € õ �E˜?¨EÑ2Ô2€EÝ�%‰=Œ=ð ØˆØBÐBÐBÐBÐBÐBØ8Ð8Ð8Ð8Ð8Ð8à�]Š]‰_Œ_ð 4ð 4ˆÝ�a˜Ñ'Ô'ð 	4Ø�HŠH‰JŒJˆJØŒJ�NŠNÐ,¨dÑ3Ô3Ð3øØ×%Ò% eÑ,Ô,Ð,Ð,Ð,r*   Ústateúdict[str, Any]c                óF  — t          | d| ¦  «        } t          t          ¦  «         ddlmc m} |                     | |d         ¦  «         |                      t          |  	                    ¦   «         ¦  «        j
        ¦  «         |                      |d         d¬¦  «         dS )zLRe-apply the fake-quantization captured by `qat_state` to a model, in place.r'  r   Nr-  r.  F)Ústrict)r"   r   r  r/  r&   r0  Úrestore_from_modelopt_stateÚtoÚnextÚ
parametersr   Úload_state_dict)r  rA  r4  s      r(   Úrestore_qatrJ    s¤   € å�E˜?¨EÑ2Ô2€EÝÕ,Ñ-Ô-Ð-Ø$Ð$Ð$Ð$Ð$Ð$Ð$Ð$Ð$à×#Ò# E¨5°Ô+<Ñ=Ô=Ð=Ø	‡H‚H�T�%×"Ò"Ñ$Ô$Ñ%Ô%Ô,Ñ-Ô-Ð-Ø	×Ò˜% œ/°%ÐÑ8Ô8Ð8Ð8Ð8r*   é€  c                óx  — |sdS t          | ¦  «        }t          | ¦  «        }t          d¦  «                             d„ |                      ¦   «         D ¦   «         ¦  «        }t          |¦  «        }|�r‚dd›dd›dd	›d
d›dd›dd	›dd›dd›�}t          j        |¦  «         t          | 	                    ¦   «         ¦  «        D �]1\  }	\  }
}|
 
                    dd¦  «        }
|j        j        }t          |j        ¦  «        rÃ|                     ¦   «         D ]¬\  }}t          j        |	d›|
› d|› �d›|d	›|j        d›|                     ¦   «         d›t#          |j        ¦  «        d	›|                     ¦   «         d›|                     ¦   «         d›t+          |j        ¦  «         
                    dd¦  «        d›�	¦  «         Œ­�Œt          j        |	d›|
d›|d	›dd›dd›g d	›dd›dd›dd›�	¦  «         �Œ3t/          | |¦  «        } t1          | dd„ ¦  «        ¦   «         rdnd}|rd|d ›d!�nd}t1          | d"d¦  «        p$t1          | d#i ¦  «                             d"d¦  «        }t5          |¦  «        j         
                    d$d%¦  «        pd&}t          j        |› d'|› d(|d)›d*|d)›d+|d)›d,|› �¦  «         ||||fS )-a$  Print and return detailed model information layer by layer.

    Args:
        model (nn.Module): Model to analyze.
        detailed (bool, optional): Whether to print detailed layer information.
        verbose (bool, optional): Whether to print model information.
        imgsz (int | list, optional): Input image size.

    Returns:
        (tuple): Tuple containing:
            - n_l (int): Number of layers.
            - n_p (int): Number of parameters.
            - n_g (int): Number of gradients.
            - flops (float): GFLOPs.
    NÚcollectionsc              3  óT   K  — | ]#\  }}t          |j        ¦  «        d k    ¯||fV — Œ$dS )r   N)r    Ú_modules)r†   r½   r  s      r(   rˆ   zmodel_info.<locals>.<genexpr>-  sB   è è € Ð2tÐ2t¹d¸aÀÕ_bÐcdÔcmÑ_nÔ_nÐrsÒ_sÐ_s°A°q°6Ð_sÐ_sÐ_sÐ_sÐ2tÐ2tr*   Úlayerz>5rg   z>40r   z>20Úgradientz>10rH  z>12rö   ÚmuÚsigmazmodule_list.rq   z>5gr  z>12gz>10.3gztorch.z>15Fr   ú-Úis_fusedc                 ó   — dS r>   rL   rL   r*   r(   r  zmodel_info.<locals>.<lambda>>  s   € ¸U€ r*   z (fused)rj   ú.1fz GFLOPsÚ	yaml_fileÚyamlÚyoloÚYOLOÚModelz summaryz: r€   z	 layers, z parameters, z
 gradients)Úget_num_paramsÚget_num_gradientsÚ
__import__ÚOrderedDictÚnamed_modulesr    r   rÈ   rÆ   r3  rž   Ú	__class__r   Ú_parametersÚnamed_parametersÚrequires_gradÚnumelÚliströ   ÚmeanÚstdr#   rS   Ú	get_flopsr"   r¥   r   Ústem)r  ÚdetailedrÑ   ÚimgszÚn_pÚn_gÚlayersÚn_lÚhr²   Úmnr  ÚmtÚpnrœ   ÚflopsÚfusedÚfsrX  Ú
model_names                       r(   Ú
model_inforz    sî  € ð  ð ØˆÝ
˜Ñ
Ô
€CÝ
˜EÑ
"Ô
"€CÝ˜Ñ&Ô&×2Ò2Ð2tÐ2tÀe×FYÒFYÑF[ÔF[Ð2tÑ2tÔ2tÑtÔt€FÝ
ˆf‰+Œ+€CØñ rØÐyÐy˜6ÐyÐy vÐyÐy°:ÐyÐyÀ<ÐyÐyÐU\ÐyÐyÐbfÐyÐyÐlsÐyÐyÐyˆÝŒ�A‰ŒˆÝ# F§L¢L¡N¤NÑ3Ô3ð 		rñ 		r‰JˆA‰w��AØ—’˜N¨BÑ/Ô/ˆBØ”Ô%ˆBÝ�1”=Ñ!Ô!ð rØ×/Ò/Ñ1Ô1ð ð ‘E�B˜Ý”KØð  Bð  B R , ,¨" , ,ð  Bð  B°Bð  Bð  B¸A¼Oð  Bð  BÐST×SZÒSZÑS\ÔS\ð  Bð  BÕcgÐhiÔhoÑcpÔcpð  Bð  BÐxy×x~Òx~ñ  yAô  yAð  Bð  Bð  JK÷  JOò  JOñ  JQô  JQð  Bð  Bõ  Z]ð  ^_ô  ^eñ  Zfô  Zf÷  Znò  Znð  owð  y{ñ  Z|ô  Z|ð  Bð  Bð  Bñô ð ð ñõ
 ”˜qÐpÐp bÐpÐp¨bÐpÐp°eÐpÐpÀ1ÐpÐpÈ2ÐpÐpÐUXÐpÐpÐ^aÐpÐpÐgjÐpÐpÐpÑqÔqÐqÑqå�e˜UÑ#Ô#€EØC�' %¨°]°]ÑCÔCÑEÔEÐMˆJˆJÈ2€EØ$)Ð	1Ð	 ˆeÐ	 Ð	 Ð	 Ð	 Ð	 ¨r€BÝ˜˜{¨BÑ/Ô/Ðbµ7¸5À&È"Ñ3MÔ3M×3QÒ3QÐR]Ð_aÑ3bÔ3b€IÝ�i‘”Ô%×-Ò-¨f°fÑ=Ô=ÐHÀ€JÝ
„K�:ÐiÐi uÐiÐi°ÐiÐiÐiÀÐiÐiÐiÐTWÐiÐiÐiÐegÐiÐiÑjÔjÐjØ��S˜%ÐÐr*   c                óX   — t          d„ |                      ¦   «         D ¦   «         ¦  «        S )z6Return the total number of parameters in a YOLO model.c              3  ó>   K  — | ]}|                      ¦   «         V — Œd S r‚   ©rf  r…   s     r(   rˆ   z!get_num_params.<locals>.<genexpr>H  s*   è è € Ð5Ð5˜Qˆq�wŠw‰yŒyÐ5Ð5Ð5Ð5Ð5Ð5r*   ©ÚsumrH  r#  s    r(   r]  r]  F  s+   € åÐ5Ð5 %×"2Ò"2Ñ"4Ô"4Ð5Ñ5Ô5Ñ5Ô5Ð5r*   c                óX   — t          d„ |                      ¦   «         D ¦   «         ¦  «        S )zEReturn the total number of parameters with gradients in a YOLO model.c              3  óL   K  — | ]}|j         ¯	|                     ¦   «         V — Œ d S r‚   )re  rf  r…   s     r(   rˆ   z$get_num_gradients.<locals>.<genexpr>M  s1   è è € ÐHÐH˜Q¸¼ÐHˆq�wŠw‰yŒyÐHÐHÐHÐHÐHÐHr*   r~  r#  s    r(   r^  r^  K  s+   € åÐHÐH %×"2Ò"2Ñ"4Ô"4ÐHÑHÔHÑHÔHÐHr*   c                ó–  — | j         j        rWddlm}  || j        g| j        | j         j        ¬¦  «                             ¦   «         d         }|                     d¦  «         nBt          | j
        ¦  «        t          t          | j
        | j         j        ¦  «        d¦  «        dœ}t          | j        j        d         d¦  «        |d<   |S )	a\  Return model info dict with useful model information.

    Args:
        trainer (ultralytics.engine.trainer.BaseTrainer): The trainer object containing model and validation data.

    Returns:
        (dict): Dictionary containing model parameters, GFLOPs, and inference speeds.

    Examples:
        YOLOv8n info for loggers
        >>> results = {
        ...     "model/parameters": 3151904,
        ...     "model/GFLOPs": 8.746,
        ...     "model/speed_ONNX(ms)": 41.244,
        ...     "model/speed_TensorRT(ms)": 3.211,
        ...     "model/speed_PyTorch(ms)": 18.755,
        ... }
    r   )ÚProfileModels)r   rm  z
model/nameé   )zmodel/parameterszmodel/GFLOPsÚ	inferencezmodel/speed_PyTorch(ms))rA   ÚprofileÚultralytics.utils.benchmarksrƒ  Úlastr   rm  Úrunr«   r]  r  Úroundrj  Ú	validatorÚspeed)Útrainerrƒ  Úresultss      r(   Úmodel_info_for_loggersr�  P  sÍ   € ð& „|Ôð 	
Ø>Ð>Ð>Ð>Ð>Ð>à�- ¤ °w´~ÈWÌ\ÔM_Ð`Ñ`Ô`×dÒdÑfÔfÐghÔiˆØ�Š�LÑ!Ô!Ð!Ð!õ !/¨w¬}Ñ =Ô =Ý!¥)¨G¬M¸7¼<Ô;MÑ"NÔ"NÐPQÑRÔRð
ð 
ˆõ */¨wÔ/@Ô/FÀ{Ô/SÐUVÑ)WÔ)W€GÐ%Ñ&Ø€Nr*   c                óà   — |d         j         \  }}}}t          | dd¦  «        }||z  |z  }t          | d| j        ¦  «        }	| xj        ||z  | j        z  |z  |z  |	| j        z   z  z  c_        dS )a�  Count the query-key and attention-value matmuls of an attention block for THOP.

    Both run functionally on reshaped tensors, so no child-module hook observes them and the block would otherwise be
    charged only for its qkv/proj/pe convolutions. Each output element of the two products costs one multiply-add over
    the contracted axis, giving `tokens**2 * (key_dim + head_dim)` per head.
    r   Úarear¸   Úkey_dimN)rö   r"   Úhead_dimÚ	total_opsÚ	num_heads)
r  r‡   ÚyÚbr
  rr  Úwr‘  Útokensr’  s
             r(   Ú_attention_opsrš  q  s}   € ð �1””�J€A€qˆ!ˆQÝ�1�f˜aÑ Ô €DØ�‰U�d‰]€FÝ�a˜ A¤JÑ/Ô/€GØ€K„K�1�t‘8˜aœkÑ)¨FÑ2°VÑ;¸wÈÌÑ?SÑTÑT€K„K€K€Kr*   c                ó„  ‡	‡
‡— 	 ddl }n# t          $ r d}Y nw xY w|sdS 	 ddlmŠ	mŠ
 ddlmŠ t          | ¦  «        } t          |  	                    ¦   «         ¦  «        }t          |t          ¦  «        s||g}t          ˆ	ˆ
fd„|                      ¦   «         D ¦   «         ¦  «        }t          ˆfd„|                      ¦   «         D ¦   «         ¦  «        }|rdnEt          | d¦  «        r4t!          t#          | j                             ¦   «         ¦  «        d	¦  «        nd	}t'          j        d
|j        d
         g|¢R |j        |j        ¬¦  «        }|r‰
t0          ‰	t0          ind}|r&|                     | |g|d¬¦  «        d         dz  dz  S |                     | |g||d¬¦  «        d         dz  dz  S # t4          $ r Y dS w xY w)aÄ  Calculate FLOPs (floating point operations) for a model in GFLOPs.

    Uses THOP's stride-aware image profiling for efficiency and accurate size-independent operations. Returns 0.0 if
    thop is unavailable or profiling fails.

    Args:
        model (nn.Module): The model to calculate FLOPs for.
        imgsz (int | list, optional): Input image size.

    Returns:
        (float): The model's GFLOPs (billions of floating point operations).
    r   Nç        )ÚAAttnÚ	Attention)ÚRTDETRDecoderc              3  ó@   •K  — | ]}t          |‰‰f¦  «        ¯|V — Œd S r‚   ©rW   )r†   r  r�  rž  s     €€r(   rˆ   zget_flops.<locals>.<genexpr>œ  s7   øè è € ÐUÐU˜1µ:¸aÀ)ÈUÐASÑ3TÔ3TÐU�QÐUÐUÐUÐUÐUÐUr*   c              3  ó8   •K  — | ]}t          |‰¦  «        V — Œd S r‚   r¡  )r†   r  rŸ  s     €r(   rˆ   zget_flops.<locals>.<genexpr>�  s-   øè è € ÐKÐK°a•Z  =Ñ1Ô1ÐKÐKÐKÐKÐKÐKr*   Ústrideé    r¸   ©r   rS   F)ÚinputsÚ
custom_opsrÑ   ç    eÍÍAró   )r¦  r£  r§  rÑ   )ÚthoprÁ   Úultralytics.nn.modules.blockr�  rž  Úultralytics.nn.modules.headrŸ  Úunwrap_modelrG  rH  rW   rg  Útupler"  r§   r%   Úmaxr,   r£  r&   Úemptyrö   r   rS   rš  r†  Ú	Exception)r  rm  r©  rœ   ÚattnÚrtdetrr£  Úimr§  r�  rž  rŸ  s            @@@r(   rj  rj    s$  øøø€ ðØˆˆˆˆøÝð ð ð Øˆˆˆðøøøð ð ØˆsðØAÐAÐAÐAÐAÐAÐAÐAØ=Ð=Ð=Ð=Ð=Ð=å˜UÑ#Ô#ˆÝ�×!Ò!Ñ#Ô#Ñ$Ô$ˆÝ˜%¥Ñ&Ô&ð 	#Ø˜E�NˆEÝÐUÐUÐUÐUÐU §¢¡¤ÐUÑUÔUÑUÔUˆÝÐKÐKÐKÐK¸5¿=º=¹?¼?ÐKÑKÔKÑKÔKˆàÐg��ÅwÈuÐV^ÑG_ÔG_Ð#g¥3¥s¨5¬<×+;Ò+;Ñ+=Ô+=Ñ'>Ô'>ÀÑ#CÔ#CÐ#CÐegˆÝŒ[˜!˜QœW QœZÐ0¨%Ð0Ð0¸¼ÈÌÐQÑQÔQˆØKOÐY�i¥°½ÐGÐGÐUYˆ
Øð 	gØ—<’< ¨r¨d¸zÐSX�<ÑYÔYÐZ[Ô\Ð_bÑbÐefÑfÐfØ�|Š|˜E¨2¨$°vÈ*Ð^cˆ|ÑdÔdÐefÔgÐjmÑmÐpqÑqÐqøÝð ð ð Øˆsˆsðøøøs$   …
 Š˜¡E(F1 Æ
&F1 Æ1
F?Æ>F?c                ó¨   — |                       ¦   «         D ]<}t          |¦  «        }|t          j        u rŒ |t          j        u rd|_        d|_        Œ=dS )zHInitialize model weights, biases, and module settings to default values.gü©ñÒMbP?g¸…ëQ¸ž?N)r"  r   r	   ÚConv2dÚBatchNorm2drè   Úmomentum)r  r  Úts      r(   Úinitialize_weightsr¹  ©  sX   € à�]Š]‰_Œ_ð ð ˆÝ�‰GŒGˆØ•”	ˆ>ˆ>ØØ•"”.Ð Ð ØˆAŒEØˆAŒJøðð r*   ç      ð?r¤  c           	     ó8  ‡‡— ‰dk    r| S | j         dd…         \  }}t          |‰z  ¦  «        t          |‰z  ¦  «        f}t          j        | |dd¬¦  «        } |sˆˆfd„||fD ¦   «         \  }}t          j        | d||d	         z
  d||d         z
  gd
¬¦  «        S )a�  Scale and pad an image tensor, optionally maintaining aspect ratio and padding to gs multiple.

    Args:
        img (torch.Tensor): Input image tensor.
        ratio (float, optional): Scaling ratio.
        same_shape (bool, optional): Whether to maintain the same shape.
        gs (int, optional): Grid size for padding.

    Returns:
        (torch.Tensor): Scaled and padded image tensor.
    rº  ró   NÚbilinearF)ÚsizerJ   Úalign_cornersc              3  óR   •K  — | ]!}t          j        |‰z  ‰z  ¦  «        ‰z  V — Œ"d S r‚   )ÚmathÚceil)r†   r‡   ÚgsÚratios     €€r(   rˆ   zscale_img.<locals>.<genexpr>Æ  s:   øè è € Ð?Ð?°1•”	˜!˜e™) b™.Ñ)Ô)¨BÑ.Ð?Ð?Ð?Ð?Ð?Ð?r*   r   r¸   gÏ÷Sã¥›Ü?)Úvalue)rö   r,   ÚFÚinterpolateÚpad)ÚimgrÃ  Ú
same_shaperÂ  rr  r˜  rÒ   s    ` `   r(   Ú	scale_imgrÊ  ´  sº   øø€ ð �‚|€|Øˆ
ØŒ9�Q�R�RŒ=�D€A€qÝ	ˆQ�‰Y‰Œ�˜Q ™Y™œÐ(€AÝ
Œ-˜ !¨*ÀEÐ
JÑ
JÔ
J€CØð @Ø?Ð?Ð?Ð?Ð?¸¸A¸Ð?Ñ?Ô?‰ˆˆ1ÝŒ5��q˜!˜a œd™( A q¨1¨Q¬4¡xÐ0¸Ð>Ñ>Ô>Ð>r*   rL   c                óÀ   — |j                              ¦   «         D ]C\  }}t          |¦  «        r||vs|                     d¦  «        s||v rŒ2t	          | ||¦  «         ŒDdS )a€  Copy attributes from object 'b' to object 'a', with options to include/exclude certain attributes.

    Args:
        a (Any): Destination object to copy attributes to.
        b (Any): Source object to copy attributes from.
        include (tuple, optional): Attributes to include. If empty, all attributes are included.
        exclude (tuple, optional): Attributes to exclude.
    r
  N)r>  r3  r    rŸ   Úsetattr)Úar—  ÚincludeÚexcluderÿ   r+  s         r(   Ú	copy_attrrÐ  Ê  sx   € ð ”
× Ò Ñ"Ô"ð ð ‰ˆˆ1Ý�‰LŒLð 	˜Q gÐ-Ð-°!·,²,¸sÑ2CÔ2CÐ-ÀqÈGÀ|À|Øå�A�q˜!ÑÔÐÐð	ð r*   c                óH   ‡‡— ˆˆfd„|                       ¦   «         D ¦   «         S )aL  Return a dictionary of intersecting keys with matching shapes, excluding 'exclude' keys, using da values.

    Args:
        da (dict): First dictionary.
        db (dict): Second dictionary.
        exclude (tuple, optional): Keys to exclude.

    Returns:
        (dict): Dictionary of intersecting keys with matching shapes.
    c                óˆ   •‡— i | ]=\  Š}‰‰v ¯	t          ˆfd „‰D ¦   «         ¦  «        ¯$|j        ‰‰         j        k    ¯:‰|“Œ>S )c              3  ó    •K  — | ]}|‰vV — Œ	d S r‚   rL   )r†   r‡   rÿ   s     €r(   rˆ   z-intersect_dicts.<locals>.<dictcomp>.<genexpr>å  s'   øè è € Ð:WÐ:WÈ!¸1ÀA¸:Ð:WÐ:WÐ:WÐ:WÐ:WÐ:Wr*   )r¦   rö   )r†   r+  rÿ   ÚdbrÏ  s     @€€r(   r,  z#intersect_dicts.<locals>.<dictcomp>å  sa   øø€ ÐsÐsÐs‘T�Q˜¨1°¨7¨7µsÐ:WÐ:WÐ:WÐ:WÈwÐ:WÑ:WÔ:WÑ7WÔ7W¨7Ð\]Ô\cÐgiÐjkÔglÔgrÒ\rÐ\rˆAˆqÐ\rÐ\rÐ\rr*   )r3  )ÚdarÔ  rÏ  s    ``r(   Úintersect_dictsrÖ  Ú  s,   øø€ ð tÐsÐsÐsÐs˜RŸXšX™ZœZÐsÑsÔsÐsr*   c                ób   — t          | t          j        j        t          j        j        f¦  «        S )z½Return True if model is of type DP or DDP.

    Args:
        model (nn.Module): Model to check.

    Returns:
        (bool): True if model is DataParallel or DistributedDataParallel.
    )rW   r	   ÚparallelÚDataParallelÚDistributedDataParallelr#  s    r(   Úis_parallelrÛ  è  s#   € õ �e�bœkÔ6½¼Ô8[Ð\Ñ]Ô]Ð]r*   r  c                óæ   — 	 t          | d¦  «        r't          | j        t          j        ¦  «        r| j        } n9t          | d¦  «        r't          | j        t          j        ¦  «        r| j        } n| S Œq)aV  Unwrap compiled and parallel models to get the base model.

    Args:
        m (nn.Module): A model that may be wrapped by torch.compile (._orig_mod) or parallel wrappers such as
            DataParallel/DistributedDataParallel (.module).

    Returns:
        (nn.Module): The unwrapped base model without compile or parallel wrappers.
    TÚ	_orig_modÚmodule)r%   rW   rÝ  r	   r!  rÞ  )r  s    r(   r¬  r¬  ô  sq   € ðÝ�1�kÑ"Ô"ð 	¥z°!´+½r¼yÑ'IÔ'Ið 	Ø”ˆAˆAÝ�Q˜Ñ!Ô!ð 	¥j°´½2¼9Ñ&EÔ&Eð 	Ø”ˆAˆAàˆHðr*   rœ  éd   c                ó   ‡ ‡‡— ˆˆ ˆfd„S )aM  Return a lambda function for sinusoidal ramp from y1 to y2 https://arxiv.org/pdf/1812.01187.pdf.

    Args:
        y1 (float, optional): Initial value.
        y2 (float, optional): Final value.
        steps (int, optional): Number of steps.

    Returns:
        (function): Lambda function for computing the sinusoidal ramp.
    c                ó†   •— t          dt          j        | t          j        z  ‰z  ¦  «        z
  dz  d¦  «        ‰‰z
  z  ‰z   S )Nr¸   ró   r   )r®  rÀ  ÚcosÚpi)r‡   ÚstepsÚy1Úy2s    €€€r(   r  zone_cycle.<locals>.<lambda>  s?   ø€ •S˜!�dœh q­4¬7¡{°UÑ':Ñ;Ô;Ñ;¸qÑ@À!ÑDÔDÈÈRÉÑPÐSUÑU€ r*   rL   )rå  ræ  rä  s   ```r(   Ú	one_cyclerç    s   øøø€ ð VÐUÐUÐUÐUÐUÐUr*   c                óÚ  — t          j        | ¦  «         t          j                              | ¦  «         t          j        | ¦  «         |r“t
          rvt          j        dd¬¦  «         dt          j        j        _	        t          rdt          j        j	        _        dt          j        d<   t          | ¦  «        t          j        d<   dS t!          j        d¦  «         dS t%          ¦   «          dS )	zñInitialize random number generator (RNG) seeds https://pytorch.org/docs/stable/notes/randomness.html.

    Args:
        seed (int, optional): Random seed.
        deterministic (bool, optional): Whether to set deterministic algorithms.
    T)Ú	warn_onlyFz:4096:8ÚCUBLAS_WORKSPACE_CONFIGÚPYTHONHASHSEEDz3Upgrade to torch>=2.0.0 for deterministic training.N)ÚrandomÚseedÚnpr&   Úmanual_seedrÊ   Úuse_deterministic_algorithmsrË   ÚcudnnÚdeterministicÚ	TORCH_2_2ÚutilsÚfill_uninitialized_memoryr£   r¤   r#   r   ÚwarningÚunset_deterministic)rí  rò  s     r(   Ú
init_seedsrø    sÒ   € õ „K�ÑÔÐÝ„I‡N‚N�4ÑÔÐÝ	Ô�dÑÔÐàð Ýð 	RÝÔ.¨t¸tÐDÑDÔDÐDØ15�EŒNÔ Ô.Ýð LØFK•”Ô)ÔCØ4=�BŒJÐ0Ñ1Ý+.¨t©9¬9�BŒJÐ'Ñ(Ð(Ð(åŒNÐPÑQÔQÐQÐQÐQåÑÔÐÐÐr*   c                 ó  — t          j        d¦  «         dt           j        j        _        t
          rdt           j        j        _        t          j	         
                    dd¦  «         t          j	         
                    dd¦  «         dS )z@Unset all the configurations applied for deterministic training.FTrê  Nrë  )r&   rð  rË   rñ  rò  ró  rô  rõ  r£   r¤   r«   rL   r*   r(   r÷  r÷  .  sj   € å	Ô& uÑ-Ô-Ð-Ø).…E„NÔÔ&Ýð CØ>B�ŒÔ!Ô;Ý„J‡N‚NÐ,¨dÑ3Ô3Ð3Ý„J‡N‚NÐ# TÑ*Ô*Ð*Ð*Ð*r*   c                  ó(   — e Zd ZdZd
d„Zd„ Zdd„Zd	S )ÚModelEMAa”  Updated Exponential Moving Average (EMA) implementation.

    Keeps a moving average of everything in the model state_dict (parameters and buffers). For EMA details see
    References.

    To disable EMA set the `enabled` attribute to `False`.

    Attributes:
        ema (nn.Module): Copy of the model in evaluation mode.
        updates (int): Number of EMA updates.
        decay (function): Decay function that determines the EMA weight.
        enabled (bool): Whether EMA is enabled.

    References:
        - https://github.com/rwightman/pytorch-image-models
        - https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage
    ç§èH.ÿï?éÐ  r   c                óT  ‡‡— t          t          |¦  «        ¦  «                             ¦   «         | _        t	          | j        d¦  «        rd| j        _        || _        ˆˆfd„| _        | j                             ¦   «         D ]}| 	                    d¦  «         Œd| _
        d| _        dS )a7  Initialize EMA for 'model' with given arguments.

        Args:
            model (nn.Module): Model to create EMA for.
            decay (float, optional): Maximum EMA decay rate.
            tau (int, optional): EMA decay time constant.
            updates (int, optional): Initial number of updates.
        Úteacher_modelNc                ó@   •— ‰dt          j        |  ‰z  ¦  «        z
  z  S )Nr¸   )rÀ  Úexp)r‡   ÚdecayÚtaus    €€r(   r  z#ModelEMA.__init__.<locals>.<lambda>Y  s!   ø€ ˜u¨­D¬H°a°R¸#±XÑ,>Ô,>Ñ(>Ñ?€ r*   FT)r   r¬  r  Úemar%   rÿ  Úupdatesr  rH  rí   rN   Ú_pairs)Úselfr  r  r  r  rœ   s     ``  r(   Ú__init__zModelEMA.__init__K  s§   øø€ õ �L¨Ñ/Ô/Ñ0Ô0×5Ò5Ñ7Ô7ˆŒÝ�4”8˜_Ñ-Ô-ð 	*à%)ˆDŒHÔ"ØˆŒØ?Ð?Ð?Ð?Ð?ˆŒ
Ø”×$Ò$Ñ&Ô&ð 	$ð 	$ˆAØ×Ò˜UÑ#Ô#Ð#Ð#ØˆŒØˆŒˆˆr*   c                óú  — | j         �rp| xj        dz  c_        |                      | j        ¦  «        }| j        €›t	          |¦  «                             ¦   «         }g g }}| j                             ¦   «                              ¦   «         D ]A\  }}|j        j	        r0| 
                    |¦  «         | 
                    ||         ¦  «         ŒB||f| _        | j        \  }}|rUt          rN|d         j        j        dk    r8t          s|d         j        j        dk    rt          j        ||d|z
  ¦  «         dS t#          ||¦  «        D ]4\  }}|                     |¦  «                             |d|z
  ¬¦  «         Œ3dS dS )zgUpdate EMA parameters.

        Args:
            model (nn.Module): Model to update EMA from.
        r¸   Nr   rT   rU   )Úalpha)rN   r  r  r  r¬  r2  r  r3  rS   Úis_floating_pointÚappendrÊ   r   r   r[   r&   Ú_foreach_lerp_r  Úmul_Úadd_)	r  r  rØ   ÚmsdÚema_vÚmodel_vrÿ   r+  r  s	            r(   ÚupdatezModelEMA.update_  s‰  € ð Œ<ñ 	3ØˆLŒL˜AÑˆLŒLØ—
’
˜4œ<Ñ(Ô(ˆAàŒ{Ð"Ý" 5Ñ)Ô)×4Ò4Ñ6Ô6�Ø!# R�w�Ø œH×/Ò/Ñ1Ô1×7Ò7Ñ9Ô9ð /ð /‘D�A�qØ”wÔ0ð /ØŸš Q™œ˜ØŸš s¨1¤vÑ.Ô.Ð.øØ# W˜n�”Ø!œ[‰NˆE�7àð3Ý#ð3Ø(-¨a¬¬Ô(<ÀÒ(EÐ(EÍ9Ð(EÐX]Ð^_ÔX`ÔXgÔXlÐpuÒXuÐXuåÔ$ U¨G°Q¸±UÑ;Ô;Ð;Ð;Ð;å  wÑ/Ô/ð 3ð 3‘D�A�qØ—F’F˜1‘I”I—N’N 1¨A°©E�NÑ2Ô2Ð2Ð2ð'	3ð 	3ð$3ð 3r*   rL   c                ó`   — | j         r&t          | j        t          |¦  «        ||¦  «         dS dS )a[  Copy attributes from model to EMA, with options to include/exclude certain attributes.

        Args:
            model (nn.Module): Model to copy attributes from; compile and parallel wrappers are unwrapped first.
            include (tuple, optional): Attributes to include.
            exclude (tuple, optional): Attributes to exclude.
        N)rN   rÐ  r  r¬  )r  r  rÎ  rÏ  s       r(   Úupdate_attrzModelEMA.update_attrz  s?   € ð Œ<ð 	GÝ�d”h¥¨UÑ 3Ô 3°W¸gÑFÔFÐFÐFÐFð	Gð 	Gr*   N)rü  rý  r   ©rL   rL   )r   Ú
__module__Ú__qualname__Ú__doc__r  r  r  rL   r*   r(   rû  rû  8  s[   € € € € € ðð ð$ð ð ð ð(3ð 3ð 3ð6	Gð 	Gð 	Gð 	Gð 	Gð 	Gr*   rû  úbest.ptÚfú
str | PathrÒ   r  c           	     ó  — 	 t          | t          j        d¦  «        ¬¦  «        }t          |t          ¦  «        s
J d¦   «         ‚d|v s
J d¦   «         ‚n3# t
          $ r&}t          j        d| › d|› �¦  «         i cY d}~S d}~ww xY wt          j	        ¦   «          
                    ¦   «                              ¦   «         t          d	d
dœ}|                     d¦  «        r|d         |d<   ddlm} t          |d         |¦  «        r*|d                              ¦   «          |d         j        |d<   t%          |d         d¦  «        r%t	          |d         j        ¦  «        |d         _        t%          |d         d¦  «        rd|d         _        |d                              ¦   «          |d                              ¦   «         D ]	}d|_        Œ
i t0          ¥|                     di ¦  «        ¥}dD ]}	d||	<   Œd|d<   d„ |                     ¦   «         D ¦   «         |d<   i |¥|¥|pi ¥}
t          j        |
|p| ¦  «         t6          j                             |p| ¦  «        dz  }t          j        d| › d|rd|› d�nd› d|d›d�¦  «         |
S )a´  Strip optimizer from 'f' to finalize training, optionally save as 's'.

    Args:
        f (str | Path): File path to model to strip the optimizer from.
        s (str, optional): File path to save the model with stripped optimizer to. If not provided, 'f' will be
            overwritten.
        updates (dict, optional): A dictionary of updates to overlay onto the checkpoint before saving.

    Returns:
        (dict): The combined checkpoint dictionary.

    Examples:
        >>> from pathlib import Path
        >>> from ultralytics.utils.torch_utils import strip_optimizer
        >>> for f in Path("path/to/model/checkpoints").rglob("*.pt"):
        ...     strip_optimizer(f)
    rV   )Úmap_locationz%checkpoint is not a Python dictionaryr  z'model' missing from checkpointz	Skipping z!, not a valid Ultralytics model: Nz2AGPL-3.0 License (https://ultralytics.com/license)zhttps://docs.ultralytics.com)Údater]   ÚlicenseÚdocsr  r   )ÚDistillationModelrA   Ú	criterionFÚ
train_args)Ú	optimizerÚbest_fitnessr  r  Úscalerr/   Úepochc                ó,   — i | ]\  }}|t           v ¯||“ŒS rL   )r   r*  s      r(   r,  z#strip_optimizer.<locals>.<dictcomp>¿  s)   € ÐNÐNÐN¡  1¸Õ=MÐ8MÐ8M�q˜!Ð8MÐ8MÐ8Mr*   g    €„.AzOptimizer stripped from r€   z
 saved as rq   r}   rW  ÚMB)r   r&   r   rW   Údictr°  r   rö  r   ÚnowÚ
astimezoneÚ	isoformatr   r¥   Úultralytics.nn.distill_modelr"  Ú_remove_feature_hooksr'  r%   rA   r#  ÚhalfrH  re  r   r3  Úsaver£   ÚpathÚgetsizerÈ   )r  rÒ   r  r‡   ÚeÚmetadatar"  rœ   rA   rÿ   ÚcombinedÚmbs               r(   Ústrip_optimizerr9  †  s×  € ð$Ý�q¥u¤|°EÑ':Ô':Ð;Ñ;Ô;ˆÝ˜!�TÑ"Ô"ÐKÐKÐ$KÑKÔKÐ"Ø˜!ˆ|ˆ|ˆ|Ð>‰|Œ|ˆ|ˆ|øÝð ð ð ÝŒÐJ 1ÐJÐJÀqÐJÐJÑKÔKÐKØˆ	ˆ	ˆ	ˆ	ˆ	ˆ	øøøøðøøøõ
 ”‘”×)Ò)Ñ+Ô+×5Ò5Ñ7Ô7ÝØGØ.ð	ð €Hð 	‡u‚uˆU�|„|ð Ø�u”Xˆˆ'‰
ð ?Ð>Ð>Ð>Ð>Ð>å�!�G”*Ð/Ñ0Ô0ð .Ø	ˆ'Œ
×(Ò(Ñ*Ô*Ð*Ø�w”ZÔ-ˆˆ'‰
åˆq�Œz˜6Ñ"Ô"ð 0Ý˜q œzœÑ/Ô/ˆˆ'Œ
ŒÝˆq�Œz˜;Ñ'Ô'ð $Ø#ˆˆ'Œ
ÔØ€g„J‡O‚OÑÔÐØˆwŒZ×"Ò"Ñ$Ô$ð  ð  ˆØˆŒˆð ;ÕÐ: !§%¢%¨°bÑ"9Ô"9Ð:€DØDð ð ˆØˆˆ!‰ˆØ€A€g�JØNÐN¨¯
ª
©¬ÐNÑNÔN€A€l�Oð 4�(Ð3˜aÐ3 G M¨rÐ3€HÝ	„Jˆx˜˜˜aÑ Ô Ð Ý	Œ�Š˜˜˜aÑ	 Ô	  3Ñ	&€BÝ
„KÐZ¨1ÐZÐZÀAÐ/MÐ/@¸AÐ/@Ð/@Ð/@Ð/@È2ÐZÐZÐPRÐZÐZÐZÐZÑ[Ô[Ð[Ø€Os   ‚AA Á
BÁA>Á8BÁ>Bc                ó  — | d                               ¦   «         D ]d}|                     ¦   «         D ]M\  }}|dvrDt          |t          j        ¦  «        r*|j        t          j        u r|                     ¦   «         ||<   ŒNŒe| S )a  Convert the state_dict of a given optimizer to FP16, focusing on the 'state' key for tensor conversions.

    Args:
        state_dict (dict): Optimizer state dictionary.

    Returns:
        (dict): Converted optimizer state dictionary with FP16 tensors.
    rA  >   ÚstepÚ
exp_avg_sq)Úvaluesr3  rW   r&   ÚTensorrS   Úfloat32r1  )r2  rA  rÿ   r+  s       r(   Ú$convert_optimizer_state_dict_to_fp16r@  Ê  s�   € ð ˜GÔ$×+Ò+Ñ-Ô-ð $ð $ˆØ—K’K‘M”Mð 	$ð 	$‰DˆAˆqØÐ.Ð.Ð.µ:¸aÅÄÑ3NÔ3NÐ.ÐSTÔSZÕ^cÔ^kÐSkÐSkØŸ6š6™8œ8��a‘øð	$ð Ðr*   c              #  óF  K  — ddi}| �| j         dv r|V — dS t          | pd¦  «        }|                     ¦   «         r`t          |d¦  «        rP|                     ¦   «          	 |V — |                     | ¦  «        |d<   dS # |                     | ¦  «        |d<   w xY w|V — dS )a¹  Monitor and manage accelerator memory usage.

    This function empties the active accelerator cache, yields a dictionary containing memory usage information, and
    then records the reserved memory on the specified device.

    Args:
        device (torch.device, optional): The accelerator device to query memory usage for.

    Yields:
        (dict): A dictionary with a key 'memory' initialized to 0, updated with reserved memory.
    Úmemoryr   N>   rV   rU   r6   Úmemory_reserved)r   r)   r2   r%   Úempty_cacherC  )r   rÈ   rà   s      r(   Úcuda_memory_usagerE  Û  sÕ   è è € ð �aˆ=€DØÐ˜fœk¨^Ð;Ð;Øˆ
ˆ
ˆ
ØˆÝ*¨6Ð+;°VÑ<Ô<€KØ×ÒÑ!Ô!ð ¥g¨kÐ;LÑ&MÔ&Mð Ø×ÒÑ!Ô!Ð!ð	AØˆJˆJˆJà(×8Ò8¸Ñ@Ô@ˆD�‰NˆNˆNø˜[×8Ò8¸Ñ@Ô@ˆD�‰NÐ@Ð@Ð@Ð@àˆ
ˆ
ˆ
ˆ
ˆ
s   Á"B  Â Bé
   c                ó
  ‡— 	 ddl }n# t          $ r d}Y nw xY wg }t          |d¬¦  «        }t          j        dd›dd›dd	›d
d	›dd	›dd›dd›�¦  «         t          j        ¦   «          |j        dvrt          |¦  «        nd}|�| 	                    ¦   «          t          | t          ¦  «        r| n| gD �]WŠ‰                     |¦  «        Šd‰_        t          |t          ¦  «        r|n|gD �]}t          |d¦  «        r|                     |¦  «        n|}t          |d¦  «        rAt          ‰t          j        ¦  «        r'‰j        t          j        u r|                     ¦   «         n|}ddg d¢}}
}		 |r%|                     |‰gd¬¦  «        d         dz  dz  nd}n# t*          $ r d}Y nw xY w	 d}t-          |¦  «        D �]}t/          |¦  «        5 }t1          |¦  «        |d<    |‰¦  «        }t1          |¦  «        |d<   	 t          |t          ¦  «        rt3          d„ |D ¦   «         ¦  «        n|                     ¦   «                              ¦   «          t1          |¦  «        |d<   n"# t*          $ r t7          d¦  «        |d<   Y nw xY wddd¦  «         n# 1 swxY w Y   ||d         dz  z  }|	|d         |d         z
  dz  |z  z  }	|
|d         |d         z
  dz  |z  z  }
|rét/          |¦  «        5 }t9          t3          ˆfd„|j                             ¦   «         D ¦   «         ¦  «        ¦  «        }t          j        ‰j         d         d|z  ||t          j!        ¬¦  «        t          j        ‰j         d         |dtE          |j#        ¦  «        z  |t          j!        ¬¦  «        f}ddd¦  «         n# 1 swxY w Y   ~||d         dz  z  }�Œd„ ‰|fD ¦   «         \  }}t          |tH          j%        ¦  «        r+t3          d „ | &                    ¦   «         D ¦   «         ¦  «        nd}t          j        |d!›|d"›|d#›|	d$›|
d$›|d›|d›�¦  «         | '                    ||||	|
||g¦  «         n@# t*          $ r3}t          j        |¦  «         | '                    d¦  «         Y d}~nd}~ww xY wt          j        ¦   «          |�| 	                    ¦   «          �Œï# t          j        ¦   «          |�| 	                    ¦   «          w w xY w�ŒY|S )%aô  Ultralytics speed, memory and FLOPs profiler.

    Args:
        input (torch.Tensor | list): Input tensor(s) to profile.
        ops (nn.Module | list): Model or list of operations to profile.
        n (int, optional): Number of iterations to average.
        device (str | torch.device, optional): Device to profile on.
        max_num_obj (int, optional): Maximum number of objects for simulation.

    Returns:
        (list): Profile results for each operation.

    Examples:
        >>> from ultralytics.utils.torch_utils import profile_ops
        >>> input = torch.randn(16, 3, 640, 640)
        >>> m1 = lambda x: x * torch.sigmoid(x)
        >>> m2 = nn.SiLU()
        >>> profile_ops(input, [m1, m2], n=100)  # profile over 100 iterations
    r   NF)rÑ   ÚParamsz>12sÚGFLOPszGPU_mem (GB)z>14szforward (ms)zbackward (ms)Úinputz>24sÚoutput>   rV   rU   TrF  r1  )r   r   r   )r¦  rÑ   r¨  ró   r¸   c              3  ó>   K  — | ]}|                      ¦   «         V — Œd S r‚   )r  )r†   Úyis     r(   rˆ   zprofile_ops.<locals>.<genexpr>.  s*   è è € Ð 6Ð 6¨b §¢¡¤Ð 6Ð 6Ð 6Ð 6Ð 6Ð 6r*   ÚnanrB  iè  c              3  óZ   •K  — | ]%}‰j         d          |z  ‰j         d         |z  z  V — Œ&dS )r/   éþÿÿÿN)rö   )r†   rÒ   r‡   s     €r(   rˆ   zprofile_ops.<locals>.<genexpr>8  s>   øè è € Ð-oÐ-oÐXY¨q¬w°r¬{¸Q©À1Ä7È2Ä;ÐQRÁ?Ñ.SÐ-oÐ-oÐ-oÐ-oÐ-oÐ-or*   é   r¥  c              3  óv   K  — | ]4}t          |t          j        ¦  «        rt          |j        ¦  «        nd V — Œ5dS )rg  N)rW   r&   r>  r­  rö   r…   s     r(   rˆ   zprofile_ops.<locals>.<genexpr>B  sA   è è € ÐiÐiÐ]^µ¸A½u¼|Ñ1LÔ1LÐX�u Q¤W™~œ~˜~ÐRXÐiÐiÐiÐiÐiÐir*   c              3  ó>   K  — | ]}|                      ¦   «         V — Œd S r‚   r}  r…   s     r(   rˆ   zprofile_ops.<locals>.<genexpr>C  s*   è è € Ð:Ð: a˜Ÿš™	œ	Ð:Ð:Ð:Ð:Ð:Ð:r*   Ú12z12.4gz>14.3fz14.4g)(r©  rÁ   rÚ   r   rÈ   ÚgcÚcollectr   r)   rD  rW   rg  rF  re  r%   r&   r>  rS   Úfloat16r1  r†  r°  rª   rE  rá   r  ÚbackwardÚfloatr,   r£  ÚtolistÚrandnrö   r?  r    Únamesr	   r!  rH  r  )rJ  Úopsr½   r   Úmax_num_objr©  rŽ  rà   r  ÚtfÚtbr¸  rv  Úmemr
  Ú	cuda_infor–  ÚanchorsÚsimÚs_inÚs_outrœ   r5  r‡   s                          @r(   Úprofile_opsrg  ÷  sU  ø€ ð(ØˆˆˆˆøÝð ð ð Øˆˆˆðøøøð €GÝ˜6¨5Ð1Ñ1Ô1€FÝ
„KØÐið 	)˜(Ðið 	)¨Ðið 	)¸nÐið 	)ÐSbÐið 	)ØÐ(ð	)Ø!Ð(ð	)ð 	)ñô ð õ „J�L„L€LØ6<´kÈÐ6WÐ6WÕ*¨6Ñ2Ô2Ð2Ð]a€KØÐØ×ÒÑ!Ô!Ð!Ý  ­Ñ-Ô-Ð:ˆUˆU°E°7ð 2.ñ 2.ˆØ�DŠD�‰LŒLˆØˆŒÝ" 3­Ñ-Ô-Ð8��°C°5ð /	.ñ /	.ˆAÝ '¨¨4Ñ 0Ô 0Ð7�—’�V‘”�°aˆAÝ# A vÑ.Ô.Ðrµ:¸aÅÄÑ3NÔ3NÐrÐSTÔSZÕ^cÔ^kÐSkÐSk�—’‘”�ÐqrˆAØ˜1˜i˜i˜i�A�ˆBðØSWÐ^˜Ÿš Q°¨s¸E˜ÑBÔBÀ1ÔEÈÑKÈaÑOÐOÐ]^��øÝð ð ð Ø���ðøøøð&.Ø�Ý˜q™œð 9ñ 9�AÝ*¨6Ñ2Ô2ð 	0°iÝ(¨Ñ0Ô0˜˜!™Ø˜A˜a™DœD˜Ý(¨Ñ0Ô0˜˜!™ð0Ý:DÀQÍÑ:MÔ:MÐT�SÐ 6Ð 6°AÐ 6Ñ 6Ô 6Ñ6Ô6Ð6ÐST×YÒYÑ[Ô[×dÒdÑfÔfÐfÝ#,¨VÑ#4Ô#4˜A˜a™D˜DøÝ(ð 0ð 0ð 0å#(¨¡<¤<˜A˜a™D˜D˜Dð0øøøð	0ð 	0ð 	0ñ 	0ô 	0ð 	0ð 	0ð 	0ð 	0ð 	0ð 	0øøøð 	0ð 	0ð 	0ð 	0ð ˜9 XÔ.°Ñ4Ñ4�CØ˜1˜Qœ4 ! A¤$™;¨$Ñ.°Ñ2Ñ2�BØ˜1˜Qœ4 ! A¤$™;¨$Ñ.°Ñ2Ñ2�BØ"ð 9Ý.¨vÑ6Ô6ð ¸)Ý&)­#Ð-oÐ-oÐ-oÐ-oÐ]^Ô]e×]lÒ]lÑ]nÔ]nÐ-oÑ-oÔ-oÑ*oÔ*oÑ&pÔ&p˜Gõ
 !&¤¨A¬G°A¬J¸¸K¹ÈÐY_ÕglÔgtÐ uÑ uÔ uÝ %¤¨A¬G°A¬J¸ÀÅSÈÌÁ\Ä\ÑAQÐZ`ÕhmÔhuÐ vÑ vÔ vð#˜Cðð ð ñ ô ð ð ð ð ð ð øøøð ð ð ð ð  Ø˜y¨Ô2°SÑ8Ñ8˜ùØiÐiÐcdÐfgÐbhÐiÑiÔi‘��eÝ>HÈÍBÌIÑ>VÔ>VÐ]•CÐ:Ð:¨1¯<ª<©>¬>Ð:Ñ:Ô:Ñ:Ô:Ð:Ð\]�Ý”˜qÐmÐm UÐmÐm°#ÐmÐm¸bÐmÐmÈÐmÐmÐRVÐmÐmÐ_dÐmÐmÐmÑnÔnÐnØ—’  5¨#¨r°2°t¸UÐCÑDÔDÐDÐDøÝð %ð %ð %Ý”˜A‘”�Ø—’˜tÑ$Ô$Ð$Ð$Ð$Ð$Ð$Ð$øøøøð%øøøõ ”
‘”�ØÐ*Ø×+Ò+Ñ-Ô-Ð-ùøõ ”
‘”�ØÐ*Ø×+Ò+Ñ-Ô-Ð-Ð-ð +øøøñ]/	.ð` €Ns¾   ƒ ˆ–Å2)FÆF+Æ*F+Æ/#Q0Ç0J	ÈA&I*É)J	É*J	ÊJ	ÊJ	Ê	J	ÊQ0ÊJÊQ0ÊJÊ AQ0Ë6B4N6	Î*Q0Î6N:Î:Q0Î=N:Î>B1Q0Ñ/SÑ0
R-Ñ:)R(Ò#SÒ(R-Ò-SÓ,Tc                  ó    — e Zd ZdZdd„Zd„ ZdS )ÚEarlyStoppinga­  Early stopping class that stops training when a specified number of epochs have passed without improvement.

    Attributes:
        best_fitness (float): Best fitness value observed.
        best_epoch (int): Epoch where best fitness was observed.
        patience (int): Number of epochs to wait after fitness stops improving before stopping.
        possible_stop (bool): Flag indicating if stopping may occur next epoch.
    é2   c                ó\   — d| _         d| _        |pt          d¦  «        | _        d| _        dS )z§Initialize early stopping object.

        Args:
            patience (int, optional): Number of epochs to wait after fitness stops improving before stopping.
        rœ  r   ÚinfFN)r&  Ú
best_epochrY  ÚpatienceÚpossible_stop)r  rn  s     r(   r  zEarlyStopping.__init__Z  s3   € ð  ˆÔØˆŒØ Ð0¥E¨%¡L¤LˆŒØ"ˆÔÐÐr*   c           
     ó&  — |€dS || j         k    s| j         dk    r|| _        || _         || j        z
  }|| j        dz
  k    | _        || j        k    }|r>t	          d¦  «        }t          j        |› d| j        › d| j        › d| j        › d	�¦  «         |S )
zýCheck whether to stop training.

        Args:
            epoch (int): Current epoch of training.
            fitness (float): Fitness value of current epoch.

        Returns:
            (bool): True if training should stop, False otherwise.
        NFr   r¸   zEarlyStopping: z:Training stopped early as no improvement observed in last z( epochs. Best results observed at epoch z@, best model saved as best.pt.
To update EarlyStopping(patience=z^) pass a new patience value, i.e. `patience=300` or use `patience=0` to disable EarlyStopping.)r&  rm  rn  ro  r   r   rÈ   )r  r(  ÚfitnessÚdeltaÚstopÚprefixs         r(   Ú__call__zEarlyStopping.__call__e  sÞ   € ð ˆ?Ø�5à�TÔ&Ò&Ð&¨$Ô*;¸qÒ*@Ð*@Ø#ˆDŒOØ 'ˆDÔØ˜œÑ'ˆØ" t¤}°qÑ'8Ò9ˆÔØ˜œÒ%ˆØð 	ÝÐ/Ñ0Ô0ˆFÝŒKØð Uð UÐUYÔUbð Uð UØ26´/ðUð Uà48´MðUð Uð Uñô ð ð ˆr*   N)rj  )r   r  r  r  r  ru  rL   r*   r(   ri  ri  P  sA   € € € € € ðð ð	#ð 	#ð 	#ð 	#ðð ð ð ð r*   ri  Údefaultútorch.nn.Moduleútorch.devicerm  Úuse_autocastÚwarmuprJ   ú
bool | strc           
     ó  — t          t          d¦  «        r|s| S t          | ¦  «        rt          d¦  «        ‚|du rd}t	          d¦  «        }|j        dk    rO	 ddlm}  |¦   «          n=# t          $ r Y n1t          $ r%}t          j        |› d	|› �¦  «         | cY d
}~S d
}~ww xY wt          j        |› d|› d�¦  «         t          j        ¦   «         }		 t          j        | |d¬¦  «        } n2# t          $ r%}t          j        |› d|› �¦  «         | cY d
}~S d
}~ww xY wt          j        ¦   «         |	z
  }
d}|�rt          j        dd|||¬¦  «        }|r|j        dk    r|                     ¦   «         }t          j        ¦   «         }t          j        ¦   «         5  |rF|j        dv r=t          j        |j        ¦  «        5   | |¦  «        }d
d
d
¦  «         n# 1 swxY w Y   n | |¦  «        }d
d
d
¦  «         n# 1 swxY w Y   |j        dk    rt          j                             |¦  «         t          j        ¦   «         |z
  }|
|z   }|r$t          j        |› d|d›d|
d›d|d›d�¦  «         nt          j        |› d|
d›d�¦  «         | S )uì  Compile a model with torch.compile and optionally warm up the graph to reduce first-iteration latency.

    This utility attempts to compile the provided model using the inductor backend. If compilation is unavailable or
    fails, the original model is returned unchanged. An optional warmup performs a single forward pass on a dummy input
    to prime the compiled graph and measure compile/warmup time.

    Args:
        model (torch.nn.Module): Model to compile.
        device (torch.device): Inference device used for warmup and autocast decisions.
        imgsz (int, optional): Square input size to create a dummy tensor with shape (1, 3, imgsz, imgsz) for warmup.
        use_autocast (bool, optional): Whether to run warmup under autocast on CUDA or MPS devices.
        warmup (bool, optional): Whether to execute a single dummy forward pass to warm up the compiled model.
        mode (bool | str, optional): torch.compile mode. True â†’ "default", False â†’ no compile, or a string like
            "default", "reduce-overhead", "max-autotune-no-cudagraphs".

    Returns:
        (torch.nn.Module): Compiled model if compilation succeeds, otherwise the original unmodified model.

    Examples:
        >>> device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
        >>> # Try to compile and warm up a model with a 640x640 input
        >>> model = attempt_compile(model, device=device, imgsz=640, use_autocast=True, warmup=True)

    Notes:
        - If the current PyTorch build does not provide torch.compile, the function returns the input model immediately.
        - Compilation is lazy and runs at the first forward pass, so the inductor CPU prerequisite of a host C++
          compiler is verified up front and the original model is returned if none is available.
        - Warmup runs under torch.inference_mode and may use torch.autocast for CUDA/MPS to align compute precision.
        - CUDA devices are synchronized after warmup to account for asynchronous kernel execution.
    Úcompilez;QAT models do not support torch.compile. Use compile=False.Trv  zcompile:rV   r   )Úget_cpp_compilerzH no C++ compiler found for the inductor backend, continuing uncompiled: Nz starting torch.compile with 'z	' mode...Úinductor)rJ   r­   z. torch.compile failed, continuing uncompiled: rœ  r¸   r„  )r   r6   >   rU   r6   z complete in rW  zs (compile zs + warmup zs)z compile complete in zs (no warmup))r%   r&   r$  rÂ   r   r   Útorch._inductor.cpp_builderr~  rÁ   r°  r   rö  rÈ   rÞ   rß   r}  Úzerosr1  r?   rc   r6   rÝ   )r  r   rm  ry  rz  rJ   rt  r~  r5  Út0Ú	t_compileÚt_warmÚdummyÚt1r
  Útotals                   r(   Úattempt_compilerˆ  ƒ  sª  € õL •5˜)Ñ$Ô$ð ¨Dð ØˆÝˆe�}„}ð XÝÐVÑWÔWÐWàˆt€|€|ØˆÝ�jÑ!Ô!€FØ„{�eÒÐð	ØDÐDÐDÐDÐDÐDàÐÑÔÐÐøÝð 	ð 	ð 	ØˆDÝð 	ð 	ð 	ÝŒN˜fÐqÐqÐnoÐqÐqÑrÔrÐrØˆLˆLˆLˆLˆLˆLøøøøð	øøøõ „K�6ÐHÐH¸ÐHÐHÐHÑIÔIÐIÝ	Ô	Ñ	Ô	€BðÝ”˜e¨$¸
ÐCÑCÔCˆˆøÝð ð ð ÝŒ˜&ÐSÐSÐPQÐSÐSÑTÔTÐTØˆˆˆˆˆˆøøøøðøøøõ Ô!Ñ#Ô# bÑ(€Ià€FØñ *å”˜A˜q %¨°vÐ>Ñ>Ô>ˆØð 	!˜FœK¨6Ò1Ð1Ø—J’J‘L”LˆEÝÔÑ Ô ˆÝÔ!Ñ#Ô#ð 	!ð 	!Øð ! ¤¨Ð >Ð >Ý”^ F¤KÑ0Ô0ð %ð %Ø˜˜e™œ�Að%ð %ð %ñ %ô %ð %ð %ð %ð %ð %ð %øøøð %ð %ð %ð %øð �E˜%‘L”L�ð	!ð 	!ð 	!ñ 	!ô 	!ð 	!ð 	!ð 	!ð 	!ð 	!ð 	!øøøð 	!ð 	!ð 	!ð 	!ð Œ;˜&Ò Ð ÝŒJ×"Ò" 6Ñ*Ô*Ð*ÝÔ"Ñ$Ô$ rÑ)ˆà˜Ñ€EØð RÝŒ�vÐjÐj¨EÐjÐjÐjÀ)ÐjÐjÐjÐ\bÐjÐjÐjÐjÑkÔkÐkÐkåŒ�vÐPÐP°IÐPÐPÐPÐPÑQÔQÐQØ€Lsx   ÁA* Á*
B$Á6	B$Á?BÂB$ÂB$ÃC- Ã-
DÃ7DÄDÄDÆ%G:Æ?GÇG:ÇG	ÇG:ÇG	ÇG:Ç:G>ÈG>)r   r   )r+   r,   )T)r6   )rN   rO   r   r#   )rq   )r   rr   rs   r#   )rq   FTr‚   )r   rÛ   )rR   )r  r  r  r,   rs   r  )r  r  rs   rX   )r  r  rs   r%  )r  r  rs   r6  )r  r  rA  rB  rs   r6  )FTrK  )rK  )rº  Fr¤  r  )rL   )r  r  rs   r  )rœ  rº  rß  )r   F)r  rq   N)r  r  rÒ   r#   r  r%  rs   rB  )rF  Nr   )rK  FFrv  )r  rw  r   rx  rm  r,   ry  rX   rz  rX   rJ   r{  rs   rw  )pÚ
__future__r   rF   rU  rÀ  r£   rì  rÞ   Ú
contextlibr   Úcopyr   r   Úpathlibr   Útypingr   Únumpyrî  r&   Útorch.distributedÚdistributedr1   Útorch.nn.functionalr	   Ú
functionalrÅ  Útorch.nn.utils.fusionr
   Úultralyticsr   Úultralytics.utilsr   r   r   r   r   r   r   r   r   r   r   r   Úultralytics.utils.checksr   r   Úultralytics.utils.cpur   Úultralytics.utils.patchesr   rE   rI   Ú
TORCH_1_11rZ   rÊ   Ú	TORCH_2_1ró  rÍ   r[   rd   Ú	TORCH_2_7Ú	TORCH_2_8Ú	TORCH_2_9Ú
TORCH_2_10Ú
TORCH_2_12Ú
TORCH_2_13ÚTORCHVISION_0_10ÚTORCHVISION_0_11ÚTORCHVISION_0_13ÚTORCHVISION_0_18rö  r)   r:   rM   rc   Ú	lru_cacherh   rp   r³   rÚ   rá   rñ   r   r  r  r$  r5  r@  rJ  rz  r]  r^  r�  rš  rj  r¹  rÊ  rÐ  rÖ  rÛ  r¬  rç  rø  r÷  rû  r9  r@  rE  rg  ri  rˆ  rL   r*   r(   ú<module>r¦     s=  ðð #Ð "Ð "Ð "Ð "Ð "à Ð Ð Ð Ø 	€	€	€	Ø €€€Ø 	€	€	€	Ø €€€Ø €€€Ø %Ð %Ð %Ð %Ð %Ð %Ø Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð à Ð Ð Ð Ø €€€Ø  Ð  Ð  Ð  Ð  Ð  Ø Ð Ð Ð Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð Ø 6Ð 6Ð 6Ð 6Ð 6Ð 6à #Ð #Ð #Ð #Ð #Ð #ðð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð GÐ FÐ FÐ FÐ FÐ FÐ FÐ FØ )Ð )Ð )Ð )Ð )Ð )Ø 0Ð 0Ð 0Ð 0Ð 0Ð 0ð ˆM˜-¨Ñ1Ô1€	Øˆ]˜=¨(Ñ3Ô3€
Øˆ]˜=¨(Ñ3Ô3€
Øˆ]˜=¨(Ñ3Ô3€
ØˆM˜-¨Ñ1Ô1€	ØˆM˜-¨Ñ1Ô1€	ØˆM˜-¨Ñ1Ô1€	ØˆM˜-¨Ñ1Ô1€	ØˆM˜-¨Ñ1Ô1€	ØˆM˜-¨Ñ1Ô1€	ØˆM˜-¨Ñ1Ô1€	ØˆM˜-¨Ñ1Ô1€	ØˆM˜-¨Ñ1Ô1€	Øˆ]˜=¨(Ñ3Ô3€
Øˆ]˜=¨(Ñ3Ô3€
Øˆ]˜=¨(Ñ3Ô3€
Ø �=Ð!4°hÑ?Ô?Ð Ø �=Ð!4°hÑ?Ô?Ð Ø �=Ð!4°hÑ?Ô?Ð Ø �=Ð!4°hÑ?Ô?Ð Ø
ð ˆ}ˆ}˜]¨IÑ6Ô6ð Ø€F„Nð	Bñô ð ðxð xð xð xð ð	^ð 	^ð 	^ñ „ð	^ðð ð ð ð.,Eð ,Eð ,Eð ,Eð ,Eð^ Ôðð ñ Ôðð
 ÔðNð Nñ ÔðNðAð Að Að Að AðHyð yð yð yðxð ð ð ð ð&ð &ð &ð0(ð (ð (ð< 1Ð2CÐDÐ ð(ð (ð (ð (ð (ðVOð Oð Oð Oðð ð ð ð0-ð -ð -ð -ð&9ð 9ð 9ð 9ð* ð * ð * ð * ðZ6ð 6ð 6ð
Ið Ið Ið
ð ð ðBUð Uð Uð'ð 'ð 'ð 'ðTð ð ð?ð ?ð ?ð ?ð,ð ð ð ð tð tð tð tð	^ð 	^ð 	^ðð ð ð ð&Vð Vð Vð Vðð ð ð ð2+ð +ð +ðKGð KGð KGð KGð KGñ KGô KGð KGð\Að Að Að Að AðHð ð ð" ðð ð ñ „ðð6Vð Vð Vð Vðr0ð 0ð 0ð 0ð 0ñ 0ô 0ð 0ðl ØØØ ðWð Wð Wð Wð Wð Wð Wr*   