Ë
    g^(hÖß ã                  óZ	  — U 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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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Zd dlZd dlZd dlmZ d dlmZ d dlmZmZmZ d dl m!Z! d dlm"Z" d dl#m$Z$ d d	l%m%Z%m&Z& d d
l'm(Z( d dl)m*Z*m+Z+m,Z,m-Z-m.Z.m/Z/m0Z0m1Z1 d dl2m3Z3 d dl4Z4d dl5m6Z7 d dl4m8Z8m9Z9 d dl:m;Z;m<Z<m=Z= d dl>m?Z?m@Z@mAZA d dlBmCZC d dlDmEZEmFZF d dlGmHZHmIZImJZJmKZKmLZLmMZMmNZNmOZOmPZPmQZQ d dlRmSZS d dlTmUZUmVZV d dlWmXZXmYZY d dlZm[Z[m\Z\ d dl]m^Z^m_Z_ d dl`maZambZbmcZcmdZd d dlemfZf d dlgmhZhmiZimjZj d dlkmlZl d dlmm?Zn d dlompZpmqZq d dlrmsZsmtZtmuZu d dlvmwZw d d!lxmyZy d d"lzm{Z{ d d#l|m}Z} d d$l~mZ d d%l€m�Z�  e?�j                  «       rd d&lƒm„Z„ d d'l…m†Z†m‡Z‡mˆZˆm‰Z‰ nd…d(„Z†d…d)„Z‡d…d*„Zˆd†d+„Z‰e/rPd d,lŠm‹Z‹mŒZŒm�Z� d d-lŽm�Z� d d.l�m‘Z‘m’Z’ d d/l“m”Z” d d0l•m–Z– d d1l—m˜Z˜m™Z™ d d2lzmšZšm›Z› d d3lœm�Z�mžZž d d4lŸm Z  d d5l¡m¢Z¢  e0d6«      Z£e�jH                  d7k(  Z¥d8Z¦e4�jN                  �jQ                  e©d9«      Zª e�jV                  e©«      Z¬d‡d:„Z­ e�j\                  d«      dˆd;„«       Z¯ G d<„ d=«      Z° G d>„ d?e°«      Z± G d@„ dAe°«      Z²d‡dB„Z³d‰dC„Z´dŠd‹dD„Zµ	 dŠ	 	 	 	 	 	 	 dŒdE„Z¶	 d�	 	 	 	 	 	 	 dŽdF„Z·	 	 	 d�	 	 	 	 	 	 	 	 	 	 	 d�dG„Z¸d‘dH„Z¹	 	 d’	 	 	 	 	 	 	 	 	 d“dI„Zºe�jv                   G dJ„ dK«      «       Z¼d”dL„Z½d•dM„Z¾ G dN„ dOe�j~                  «      ZÀ	 	 	 	 	 	 	 	 d–dP„ZÁ e�j\                  d«      d—dQ„«       ZÂd‡dR„ZÃe�jv                   G dS„ dT«      «       ZÄ G dU„ dVeÅ«      ZÆ G dW„ dX«      ZÇ	 	 	 	 	 	 	 	 	 	 d˜dY„ZÈd™dZ„ZÉ G d[„ d\«      ZÊ e�j\                  d«      dšd]„«       ZËeb G d^„ d_«      «       ZÌ G d`„ da«      ZÍebe�j\                  d‡db„«       «       ZÎd‡dc„ZÏdaÐddeÑde<   d›df„ZÒeb G dg„ dh«      «       ZÓ	 	 	 	 	 	 dœdi„ZÔeb G dj„ dkeÓ«      «       ZÕeb G dl„ dmeÕ«      «       ZÖeb G dn„ doeÕ«      «       Z×d�dp„ZØdždq„ZÙeb G dr„ ds«      «       ZÚ	 	 	 	 	 	 dŸdt„ZÛd du„ZÜd¡dv„ZÝd¡dw„ZÞd¡dx„Zßd¡dy„Zà	 d¢	 	 	 	 	 	 	 	 	 d£dz„Zá G d{„ d|«      Zâeb G d}„ d~«      «       Zãeb G d„ d€«      «       Zä G d�„ d‚«      Zå G dƒ„ d„eå«      Zæy)¤é    )ÚannotationsN)Úbisect_right)Úcopy)Úc_void_pÚCDLLÚcdll)Ú	timedelta)Úpartial)ÚPath)ÚtimeÚtime_ns)Ú
ModuleType)ÚAnyÚCallableÚcastÚNoReturnÚOptionalÚTYPE_CHECKINGÚTypeVarÚUnion)ÚSelf)ÚSymIntÚTensor)ÚCompileEventLoggerÚcountersÚdynamo_timed)ÚconfigÚexcÚmetrics)Úcuda_env)Úrocm_compile_commandÚrocm_compiler)
Ú_LINKER_SCRIPTÚ_set_gpu_runtime_envÚ_TORCH_PATHÚ_transform_cuda_pathsÚ
CppBuilderÚ
CppOptionsÚCppTorchDeviceOptionsÚget_compiler_version_infoÚ&get_name_and_dir_from_output_file_pathÚnormalize_path_separator)Úpick_vec_isa)ÚCustomGraphPassÚCustomGraphPassType)Úhas_frozen_paramsÚis_frozen_param)Ú_reload_python_moduleÚ _reload_python_module_in_subproc)Ú	cache_dirÚdefault_cache_dir)ÚALIGN_BYTESÚclear_on_fresh_inductor_cacheÚis_linuxÚ
is_windows)Útrace_structured)Úextract_tensor_metadataÚ
FakeTensorÚTensorMetadata)Úlog_cache_bypass)r   )ÚCacheArtifactManagerÚCacheArtifactType)Úhas_hintÚhint_intÚShapeEnv)Ú
OrderedSeté   )ÚCUSTOM_OBJ_FILENAME_PREFIX)Úcreate_cache)Úautotune_cache)ÚAutotuneCacheBundler)ÚTritonBundler)Úbuild_paths)Úlog_global_cache_errorsÚlog_global_cache_statsÚlog_global_cache_valsÚuse_global_cachec                  ó   — y ©N© ©ÚargsÚkwargss     úW/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/torch/_inductor/codecache.pyrL   rL   p   ó   € Øó    c                  ó   — y rQ   rR   rS   s     rV   rM   rM   s   rW   rX   c                  ó   — y rQ   rR   rS   s     rV   rN   rN   v   rW   rX   c                  ó   — y©NFrR   rR   rX   rV   rO   rO   y   s   € ØrX   )Ú	GeneratorÚKeysViewÚSequence)ÚFuture)Ú_CompileFxKwargsÚCompiledFxGraph©ÚGraphLowering)ÚChoiceCaller)ÚCompiledFxGraphConstantsÚ
OutputCode)Ú
JsonDataTyÚRemoteCache)ÚHalideInputSpecÚ
HalideMeta)ÚCachingAutotuner)Ú	InputTypeÚTÚwin32iX  Úoutput_codec                 ó>   — t         j                  j                  €dS dS )NÚ
cubin_pathÚ
hsaco_path)ÚtorchÚversionÚhiprR   rX   rV   Úget_cpp_wrapper_cubin_path_namerw   ”   s   € Ü Ÿ=™=×,Ñ,Ð4ˆ<ÐF¸,ÐFrX   c                ó†   — | �>t        t        j                  j                  | t        j                  «       d   «      «      S d S )NÚhash)r   ÚosÚpathÚjoinÚ	CacheBaseÚ
get_system)Úglobal_cache_dirs    rV   Úget_global_cache_path_implr€   ˜   sA   € ð Ð'ô 	ŒR�W‰W�\‰\Ð*¬I×,@Ñ,@Ó,BÀ6Ñ,JÓKÓLðð ðrX   c                  ó¸   — e Zd Ze ej
                  d«      dd„«       «       Zee ej
                  d«      d	d„«       «       «       Zed
d„«       Z	dd„Z
dd„Zdd„Zy)r}   Nc                 ó�  — 	 ddl m}   | «       }	 dd id|idœ}t        j                  j                  t        j                  j                  «       «      }t        j                  j                  �3|j                  |d   d<   t        j                  j                  |d   d<   n2|j                  |d   d<   t        j                  j                  |d   d	<   t        j                  t        j                   |d
¬«      j#                  d«      «      j%                  «       |d<   |S # t        $ r d }Y �Œw xY w# t        t        f$ r i }Y Œqw xY w)Nr   )Ú
triton_keyÚnameÚtriton)Údeviceru   r†   ru   Úcudarv   T)Ú	sort_keysúutf-8ry   )Útriton.compiler.compilerrƒ   ÚModuleNotFoundErrorrt   r‡   Úget_device_propertiesÚcurrent_deviceru   r„   ÚgcnArchNamerv   ÚAssertionErrorÚRuntimeErrorÚhashlibÚsha256ÚjsonÚdumpsÚencodeÚ	hexdigest)rƒ   Útriton_versionÚsystemÚdevice_propertiess       rV   r~   zCacheBase.get_system¢   s3  € ð	"Ý;ñ (›\ˆNð	à! 4˜.à˜nðñ&ˆFô !&§
¡
× @Ñ @Ü—
‘
×)Ñ)Ó+ó!Ðô �}‰}×!Ñ!Ð-Ø+<×+AÑ+A��xÑ  Ñ(Ü,1¯M©M×,>Ñ,>��yÑ! &Ò)à+<×+HÑ+H��xÑ  Ñ(Ü+0¯=©=×+<Ñ+<��yÑ! %Ñ(ô
 !Ÿ™Ü�J‰J�v¨Ô.×5Ñ5°gÓ>ó
ç
‰)‹+ð 	ˆv‰ð ˆøô7 #ò 	"Ø!‹Nð	"ûô& ¤Ð-ò 	àŠFð	ús#   ‚D �CD1 ÄD.Ä-D.Ä1EÅEc                 ó�   — t        t        j                  j                  t	        «       dt
        j                  «       d   «      «      S )NÚcachery   )r   rz   r{   r|   r4   r}   r~   rR   rX   rV   Úget_local_cache_pathzCacheBase.get_local_cache_pathÈ   s0   € ô ”B—G‘G—L‘L¤£¨g´y×7KÑ7KÓ7MÈfÑ7UÓVÓWÐWrX   c                 ó4   — t        t        j                  «      S rQ   )r€   r   r   rR   rX   rV   Úget_global_cache_pathzCacheBase.get_global_cache_pathÎ   s   € ä)¬&×*AÑ*AÓBÐBrX   c                ó6   — t         j                  «       | _        y rQ   )r}   r~   r˜   ©Úselfs    rV   Ú__init__zCacheBase.__init__Ò   s   € Ü×*Ñ*Ó,ˆ�rX   c                óÂ   — | j                  «       }|j                  «       si S t        |«      5 }t        j                  |«      }d d d «       |d   S # 1 sw Y   d   S xY w©Nr›   )rœ   Úis_fileÚopenr“   Úload)r¡   Úlocal_cache_pathÚlocal_cache_fpÚlocal_caches       rV   Úget_local_cachezCacheBase.get_local_cacheÕ   sa   € Ø×4Ñ4Ó6ÐØ×'Ñ'Ô)ØˆIÜÐ"Ó#ð 	4 ~ÜŸ)™) NÓ3ˆK÷	4à˜7Ñ#Ð#÷	4à˜7Ñ#Ð#ús   ®AÁAc                ó–   — | j                  «       }t        t        |«      t        j                  | j
                  |dœd¬«      d¬«       y )N)r˜   r›   é   )ÚindentT©Ú	make_dirs)rœ   Úwrite_atomicÚstrr“   r”   r˜   )r¡   rª   r¨   s      rV   Úupdate_local_cachezCacheBase.update_local_cacheÝ   s<   € Ø×4Ñ4Ó6ÐÜÜÐ Ó!Ü�J‰J $§+¡+¸ÑDÈQÔOØö	
rX   ©Úreturnúdict[str, Any])rµ   r   )rµ   úOptional[Path]©rµ   ÚNone)rª   r¶   rµ   r¹   )Ú__name__Ú
__module__Ú__qualname__ÚstaticmethodÚ	functoolsÚ	lru_cacher~   r7   rœ   rž   r¢   r«   r³   rR   rX   rV   r}   r}   ¡   s‚   „ ØØ€Y×Ñ˜Óò"ó ó ð"ðH Ø"Ø€Y×Ñ˜ÓòXó ó #ó ðXð òCó ðCó-ó$ô
rX   r}   c                  ó   — e Zd Zdd„Zdd„Zy)Ú
LocalCachec                óN   — | j                  «       }|}|D ]  }||v r||   }Œ y  |S rQ   )r«   )r¡   Úkeysr›   Ú	sub_cacheÚkeys        rV   ÚlookupzLocalCache.lookupç   s?   € Ø×$Ñ$Ó&ˆàˆ	Øò 	ˆCØ�e‰|Ø! #™J‘	áð		ð ÐrX   c               óœ   — | j                  «       }|}|dd D ]  }|j                  |i «       ||   }Œ |||d   <   | j                  |«       y )Nr   éÿÿÿÿ)r«   Ú
setdefaultr³   )r¡   ÚvaluerÃ   r›   rÄ   rÅ   s         rV   Ú	set_valuezLocalCache.set_valueó   sa   € Ø×$Ñ$Ó&ˆàˆ	Ø˜˜"�:ò 	'ˆCØ× Ñ   bÔ)Ø! #™‰Ið	'ð $ˆ	�$�r‘(Ñà×Ñ Õ&rX   N)rÃ   r²   rµ   úOptional[dict[str, Any]])rÃ   r²   rÊ   r   rµ   r¹   )rº   r»   r¼   rÆ   rË   rR   rX   rV   rÁ   rÁ   æ   s   „ ó
ô	'rX   rÁ   c                  óZ   — e Zd Z ej                  d«      dd„«       Z	 	 	 	 	 	 	 	 	 	 dd„Zy)ÚPersistentCacheNc                óÆ   — | j                  «       }|�|j                  «       si S t        |«      5 }t        j                  |«      }d d d «       |d   S # 1 sw Y   d   S xY wr¤   )rž   r¥   r¦   r“   r§   )r¡   Úglobal_cache_pathÚglobal_cache_fpÚglobal_caches       rV   Úget_global_cachez PersistentCache.get_global_cache   sg   € à ×6Ñ6Ó8ÐØÐ$Ð,=×,EÑ,EÔ,GØˆIÜÐ#Ó$ð 	6¨ÜŸ9™9 _Ó5ˆL÷	6à˜GÑ$Ð$÷	6à˜GÑ$Ð$ús   °AÁA c                óð  ‡‡‡‡‡— t        j                  «       Št        t        | j                  ‰‰‰«      }t        t
        | j                  ‰‰‰«      }t        t        | j                  ‰‰‰«      }i Šddˆˆˆˆˆfd„}t        j                  st        j                  �rt        j                  r| j                  «       ni }	 ||	«      sét        «       r || j                  «       |¬«      sÇ|�Å	  |‰«      Št        ˆfd„‰D «       «      sJ ‚|	j                  ‰i «       |	‰   j                  ‰i «      j                  ‰i «       ‰j!                  «       D ]!  \  }
}||	‰   ‰   ‰   |
j#                  «       <   Œ# 	 | j'                  |	«       ‰D �
ci c]  }
|
j#                  «       ‰|
   “Œ }}
 ||«       ‰S t        «       r || j                  «       |¬«       ‰S # t$        $ r} ||«       |‚d}~ww xY wc c}
w )aG  
        Check to see if we have benchmarked the given choice callers. For each
        choice caller:

            1. Check global_cache[op][inputs][choice][precision], return benchmark if cached.
            2. Check local_cache[op][inputs][choice][precision], return benchmark if cached.
            3. If benchmark is not None:
                a. `max_autotune_gemm=True`: benchmark the choice, update
                    local_cache[op][inputs][choice], and return the benchmark.
                b. `max_autotune_gemm=False`: don't benchmark the choice, return nothing.
        Nc                óâ   •— d}‰D ][  }|j                  «       }|| j                  ‰i «      j                  ‰i «      j                  ‰i «      v r| ‰   ‰   ‰   |   ‰	|<   ŒYd} n |r	 ||¬«       |S )z2Check if `cache` contains data for all the choicesTF)Úcached)Úhash_keyÚget)
r›   ÚcallbackÚhitÚchoiceÚchoice_hashÚchoicesÚinputsÚopÚ	precisionÚtimingss
        €€€€€rV   Úcheck_cachez+PersistentCache.lookup.<locals>.check_cache$  s‰   ø€ àˆCØ!ò �Ø$Ÿo™oÓ/�Ø %§)¡)¨B°Ó"3×"7Ñ"7¸ÀÓ"C×"GÑ"GÈ	ÐSUÓ"VÑVà&+¨B¡i°Ñ&7¸	Ñ&BÀ;Ñ&O�G˜F’Oð  �CÙðñ Ù Õ$ØˆJrX   )rÙ   c              3  ó&   •K  — | ]  }|‰v –— Œ
 y ­wrQ   rR   )Ú.0rÛ   rá   s     €rV   ú	<genexpr>z)PersistentCache.lookup.<locals>.<genexpr>B  s   øè ø€ ÒG°V˜v¨Ô0ÑGùs   ƒrQ   )r›   r¶   rÙ   r   rµ   Úbool)rt   Úget_float32_matmul_precisionr
   rM   r˜   rN   rL   r   Úmax_autotuneÚmax_autotune_gemmÚautotune_local_cacher«   rO   rÓ   ÚallrÉ   Úitemsr×   r�   r³   )r¡   rÝ   rß   rÞ   Ú	benchmarkÚ	log_statsÚlog_valsÚ
log_errorsrâ   rª   rÛ   ÚtimingÚeÚtimings_to_logrà   rá   s    ```          @@rV   rÆ   zPersistentCache.lookup	  sâ  ü€ ô$ ×6Ñ6Ó8ˆ	äÔ2°D·K±KÀÀVÈYÓWˆ	ÜÔ0°$·+±+¸rÀ6È9ÓUˆÜÜ# T§[¡[°"°f¸ió
ˆ
ð ˆ÷	ò 	ô  ×Ò¤&×":Ó":Ü4:×4OÒ4O˜$×.Ñ.Ô0ÐUWˆKñ   Ô,ä$Ô&Ù# D×$9Ñ$9Ó$;ÀiÕPàÐ)ðá'¨Ó0�GÜÓG¸wÔGÔGÐGÐGØ×*Ñ*¨2¨rÔ2Ø ‘O×.Ñ.¨v°rÓ:×EÑEÀiÐQSÔTØ*1¯-©-«/ò W™˜ ØPV˜ B™¨Ñ/°	Ñ:¸6¿?¹?Ó;LÒMñWð ×'Ñ'¨Ô4ð FMö"Ø;A�F—O‘OÓ% w¨v¡Ñ6ð"�ð "ñ ˜Ô(ð ˆô Ôá˜×-Ñ-Ó/¸)ÕDð ˆøô! $ò á˜q”MØ�Gûðüò"s   Ã.B	G ÆG3Ç	G0Ç!
G+Ç+G0r´   )
rÝ   zlist[ChoiceCaller]rß   r²   rÞ   r²   rí   z4Optional[Callable[[Any], dict[ChoiceCaller, float]]]rµ   zdict[ChoiceCaller, float])rº   r»   r¼   r¾   r¿   rÓ   rÆ   rR   rX   rV   rÎ   rÎ   ÿ   s]   „ Ø€Y×Ñ˜Óò%ó ð%ðNà#ðNð ðNð ð	Nð
 HðNð 
#ôNrX   rÎ   c                 óÂ   — t         j                  j                  t        «       d«      } t         j                  j	                  | «      st        j
                  | d¬«       | S )NÚlocksT©Úexist_ok)rz   r{   r|   r4   ÚexistsÚmakedirs)Úlock_dirs    rV   Úget_lock_dirrû   Z  s;   € Ü�w‰w�|‰|œI›K¨Ó1€HÜ�7‰7�>‰>˜(Ô#Ü
�‰�H tÕ,Ø€OrX   c                ó®   — t        j                  t        j                  | «      j	                  «       «      d d j                  d«      j                  «       S )Né3   r‰   )Úbase64Ú	b32encoder‘   r’   ÚdigestÚdecodeÚlower)Údatas    rV   Úsha256_hashr  a  s@   € ä×ÑœGŸN™N¨4Ó0×7Ñ7Ó9Ó:¸3¸BÐ?×FÑFÀwÓO×UÑUÓWÐWrX   c                ó¾   — t        | t        «      r| n| j                  d«      }|r+t        |t        «      r|n|j                  d«      }|dz   |z   }dt        |«      z   S )Nr‰   s   ||Úc)Ú
isinstanceÚbytesr•   r  )ÚcodeÚextraÚhashing_strÚextra_bs       rV   Ú	code_hashr  f  sV   € Ü$ T¬5Ô1‘$°t·{±{À7Ó7K€KÙÜ% e¬UÔ3‘%¸¿¹ÀgÓ9NˆØ! EÑ)¨GÑ3ˆØ”˜[Ó)Ñ)Ð)rX   c                óF  — |rKt         j                  j                  |«      r|}nTt         j                  j                  t	        «       |«      }n+t         j                  j                  t	        «       | dd «      }t         j                  j                  || › d|› �«      }| ||fS )NrE   é   ú.)rz   r{   Úisabsr|   r4   )ÚbasenameÚ	extensionÚspecified_dirÚsubdirr{   s        rV   Úget_pathr  n  sz   € ñ Ü�7‰7�=‰=˜Ô'Ø"‰Fä—W‘W—\‘\¤)£+¨}Ó=‰Fä—‘—‘œi›k¨8°A°a¨=Ó9ˆÜ�7‰7�<‰<˜ 8 *¨A¨i¨[Ð 9Ó:€DØ�V˜TÐ!Ð!rX   c                óp   — |dk(  rt        | |«      S |dv rt        t        | «      «      S t        d|› �«      ‚)Nr	  )ÚcubinÚhsacoÚspvzUnknown hash type )r  Úreprr�   )Úcontentr
  Ú	hash_types      rV   Úget_hashr  |  sD   € ð �FÒÜ˜ %Ó(Ð(ØÐ-Ñ-Üœ˜g›Ó'Ð'Ü
Ð-¨i¨[Ð9Ó
:Ð:rX   c                ó¼   — t        | j                  «       ||«      }t        |||«      \  }}}t        j                  j                  |«      st        || d¬«       ||fS )NTr¯   )r  Ústripr  rz   r{   rø   r±   )	r  r  r
  r  r  rÅ   r  Ú_subdirr{   s	            rV   Úwriter"  †  sS   € ô ˜Ÿ™›¨°	Ó:€CÜ& s¨I°}ÓEÑ€Hˆg�tÜ�7‰7�>‰>˜$ÔÜ�T˜7¨dÕ3Ø�Tˆ>ÐrX   c                ó    — t        | d«      d   S )zT
    Write the `text` to a file and return the path computed based on the hash.
    ÚtxtrE   ©r"  )Útexts    rV   Ú
write_textr'  —  s   € ô ��uÓ˜aÑ Ð rX   c                óR  — t        |t        t        f«      sJ d«       ‚t        | «      }|r|j                  j                  dd¬«       |j                  dt        j                  «       › dt        j                  «       › d�z  }t        |t        «      rdnd}|j                  ||rdnd ¬	«      5 }|j                  |«       d d d «       	 |j                  |¬
«       y # 1 sw Y   ŒxY w# t        $ r6 t        s‚ t        j                   ||¬«       t        j"                  |«       Y y w xY w)Nz6Only strings and byte arrays can be saved in the cacheT)Úparentsr÷   r  z.tmpÚwÚwbr‰   )Úencoding)Útarget)ÚsrcÚdst)r  r²   r  r   ÚparentÚmkdirrz   ÚgetpidÚ	threadingÚ	get_identr¦   r"  ÚrenameÚFileExistsErrorÚ_IS_WINDOWSÚshutilÚcopy2Úremove)Úpath_r  r°   Úencode_utf_8r{   Útmp_pathÚ
write_modeÚfs           rV   r±   r±   ž  s  € ô �g¤¤U˜|Ô,ð Ø@óÐ,ô �‹;€DÙØ�‰×Ñ $°ÐÔ6Ø�{‰{˜q¤§¡£ ¨Q¬y×/BÑ/BÓ/DÐ.EÀTÐJÑJ€HÜ" 7¬CÔ0‘°d€JØ	�‰�z±|©GÈˆÓ	Nð ÐRSØ	�‰�Ô÷ð
Ø�‰˜tˆÕ$÷ð ûô ò ÝØô 	�‰˜ tÕ,ä
�	‰	�(Öðús   Â-CÃC' ÃC$Ã'<D&Ä%D&c                  ó&   — e Zd ZU dZded<   ded<   y)ÚTensorMetadataAndValueszk
    TensorMetadata plus the elements as a list of raw values.
    Used for hashing inlined constants.
    r=   Útensor_metadataú	list[Any]ÚvaluesN©rº   r»   r¼   Ú__doc__Ú__annotations__rR   rX   rV   rA  rA  ½  s   … ñð
 $Ó#ØÔrX   rA  c                ó   — | S rQ   rR   ©Úxs    rV   Ú_identrK  È  s   € Ø€HrX   c                ód   — t        | «      }t        | d«      st        j                  |dd¬«      }|S )zs
    Extracts the tensor metadata and removes fields of the TensorMetadata
    that are not needed for caching
    Ú_is_inductor_staticr   N)Ústorage_offsetÚstorage_bytes)r;   ÚhasattrÚdataclassesÚreplace)ÚtÚmetas     rV   Ú%extract_tensor_metadata_for_cache_keyrU  Ì  s2   € ô
 # 1Ó%€DÜ�1Ð+Ô,Ü×"Ñ" 4¸ÈÔNˆà€KrX   c                  óŠ   ‡ — e Zd ZdZ	 d	 	 	 	 	 dˆ fd„Z	 	 	 	 dd„Z	 	 	 	 dd„Zdd„Zdd„Z	 	 	 	 dd„Z	dd„Z
dd	„Zdd
„Zˆ xZS )ÚFxGraphCachePicklera:  
    Custom pickler to customize the pickling of some objects (Tensors), only for the
    purpose of computing a hash for keying into the FxGraphCache. Tensors contain
    objects that don't pickle and/or vary between runs, and we want to capture the
    data that allow us to compute a stable, but safe hash.
    c                óv  •— t        j                  «       | _        t        ‰| �  | j                  «       t
        j                  j                  «       | _        | j                  j                  t        t        j                  | j                  «      t        j                  t        j                  | j                  «      t        j                   j"                  j$                  t        j                  | j                  «      t        j&                  t        j                  | j(                  «      t        j*                  j,                  j.                  j0                  t        j                  | j2                  «      i«       |r6t        j                  | j4                  «      | j                  |j6                  <   d| _        y)a2  
        Create an FX graph pickler. If include_non_inlined=True, then pickling will
        include the _values_ for all Tensors. (Note that any tensors are constants
        attached as attributes to the GraphModule). Otherwise, pickling will include
        only the metadata for these tensors.
        TN)ÚioÚBytesIOÚ_streamÚsuperr¢   ÚcopyregÚdispatch_tabler   Úupdater<   r¾   r
   Ú_reduce_fake_tensorrt   r   Ú_reduce_tensorÚnnÚ	parameterÚ	Parameterr   Ú_reduce_symintÚfxÚexperimentalÚ_backward_stateÚBackwardStateÚ_reduce_unsupportedÚ_reduce_graph_moduleÚ	__class__Úfast)r¡   ÚgmÚhas_user_defined_triton_kernelsrl  s      €rV   r¢   zFxGraphCachePickler.__init__à  s)  ø€ ô —z‘z“|ˆŒÜ‰Ñ˜Ÿ™Ô&ä%×4Ñ4×9Ñ9Ó;ˆÔØ×Ñ×"Ñ"äœI×-Ñ-¨d×.FÑ.FÓGÜ—‘œi×/Ñ/°×0CÑ0CÓDÜ—‘×"Ñ"×,Ñ,¬i×.?Ñ.?À×@SÑ@SÓ.TÜ—‘œi×/Ñ/°×0CÑ0CÓDÜ—‘×%Ñ%×5Ñ5×CÑCÄY×EVÑEVØ×,Ñ,óFðô
	
ñ +ä09×0AÑ0AØ×)Ñ)ó1ˆD×Ñ §¡Ñ-ð ˆ�	rX   c                ó*   — t        |«      }t        |ffS )z7
        Custom reducer to pickle FakeTensors.
        )rU  rK  )r¡   rS  Úmetadatas      rV   r`  z'FxGraphCachePickler._reduce_fake_tensor  s   € ô 9¸Ó;ˆÜ˜˜Ð$Ð$rX   c                óP  — ddl m} |j                  rt        d«      ‚t	        |«      }t        |«      r |j                  |«      s	t        |ffS t        «       }|j                  «       }t        «       |z
  }|dkD  rt        j                  d|d›d�«       t        t        ||«      ffS )z’
        Custom reducer to pickle Tensors.  If we see tensors, we know they're constants
        stored as attributes on the GraphModule.
        rE   rc   zmkldnn tensors unpickleableg      ð?z0FX graph cache copying of a large constant took z.1zs. Please file an issue.)Úgraphrd   Ú	is_mkldnnÚBypassFxGraphCacherU  r1   Úcan_inline_constantrK  r   ÚtolistÚwarningsÚwarnrA  )r¡   rS  rd   rq  ÚstartrD  Úelapseds          rV   ra  z"FxGraphCachePickler._reduce_tensor  s«   € õ 	)à�;Š;ô %Ð%BÓCÐCä8¸Ó;ˆô ˜1ÔÐ&G m×&GÑ&GÈÔ&JÜ˜X˜KÐ(Ð(ô “ˆØ—‘“ˆÜ“&˜5‘.ˆØ�SŠ=Ü�M‰MØBÀ7È2À,ð O(ð (ôô
 Ô0°¸6ÓBÐDÐEÐErX   c                ó&   — t         t        |«      ffS )z3
        Custom reducer to pickle SymInts.
        )rK  r²   ©r¡   Úss     rV   re  z"FxGraphCachePickler._reduce_symint/  s   € ô œ˜Q›˜	Ð"Ð"rX   c                ó   — t        d«      ‚)z{
        Custom reducer to handle any objects that we don't support and therefore
        raise to bypass caching.
        zReduce unsupported)ru  r}  s     rV   rj  z'FxGraphCachePickler._reduce_unsupported8  s   € ô
 !Ð!5Ó6Ð6rX   c                óª   — |j                  «       \  }\  }}|d   }t        j                  dd|«      }t        j                  dd|«      }||d<   |||ffS )aþ  
        Custom reducer for graph module to handle irrelevant data for user
        defined triton kernels
        Essentially what we are doing here is a huge hack where user defined
        triton kernel contain a dynamo time side table and the arguments to the
        call_function are indicies into this side table. These arguments are not
        for hashing purposes since we included the source code into the cache
        key and the numbers are prone to give false negatives due to ordering.
        Ú_codezkernel_idx = \d+Ú zconstant_args_idx = \d+)Ú
__reduce__ÚreÚsub)r¡   rn  Úfnr  Úimportsr	  s         rV   rk  z(FxGraphCachePickler._reduce_graph_module?  s_   € ð !Ÿm™m›oÑˆ‰OˆT�7Ø�G‰}ˆÜ�v‰vÐ)¨2¨tÓ4ˆÜ�v‰vÐ0°"°dÓ;ˆØˆˆW‰Ø�D˜'�?Ð"Ð"rX   c                ó°  — 	 | j                  |«       | j                  j                  «       | j                  j                  d«       | j                  j	                  d«       S # t
        t        f$ r(}t        j                  dd¬«       t        d«      |‚d}~ww xY w# | j                  j                  d«       | j                  j	                  d«       w xY w)z<
        Pickle an object and return a byte string.
        r   zFailed to pickle cache keyT©Úexc_infoN)
Údumpr[  ÚgetvalueÚseekÚtruncateÚ	TypeErrorÚAttributeErrorÚlogÚwarningru  )r¡   Úobjrò   s      rV   r”   zFxGraphCachePickler.dumpsR  s§   € ð
	%Ø�I‰I�cŒNØ—<‘<×(Ñ(Ó*ð �L‰L×Ñ˜aÔ Ø�L‰L×!Ñ! !Õ$øô œ>Ð*ò 	Jä�K‰KÐ4¸tˆKÔDÜ$Ð%AÓBÈÐIûð	Jûð �L‰L×Ñ˜aÔ Ø�L‰L×!Ñ! !Õ$ús#   ‚*A# Á#BÁ2#BÂBÂB Â8Cc                ó:   — | j                  |«      }t        |«      S )zE
        Serialize an object and return a hash of the bytes.
        )r”   r  )r¡   r“  Úserialized_datas      rV   r  zFxGraphCachePickler.get_hashb  s   € ð Ÿ*™* S›/ˆÜ˜?Ó+Ð+rX   c                ó>  ‡ — dˆ fd„}g }t        |«      j                  «       D ]÷  \  }}t        |t        «      rTt	        t        |«      «      D ]<  }‰ j                  ||   «      }|j                  d|› d|› d|› d |||   «      › �«       Œ> Œjt        |t        «      rM|j                  «       D ]9  \  }}	‰ j                  |	«      }|j                  d|› d|› d|› d ||	«      › �«       Œ; ŒÇ‰ j                  |«      }|j                  d|› d|› d ||«      › �«       Œù |S )zÉ
        Get a printable string describing in more detail all the attributes
        comprising an object. Useful for debugging when one graph hashes
        to a different value than another.
        c                ó  •— t        | t        j                  «      rt        t	        | «      «      S t        | t
        «      ryt        | «      ‰j                  v r*t         ‰j                  t        | «         | «      d   «      S t        | «      S )Nz<bytes>rE   )r  rt   r   r²   rU  r  Útyper^  )r“  r¡   s    €rV   Úget_strz0FxGraphCachePickler.debug_lines.<locals>.get_strp  sq   ø€ Ü˜#œuŸ|™|Ô,ÜÔ@ÀÓEÓFÐFÜ˜C¤Ô'Ø Ü�c“˜d×1Ñ1Ñ1äÐ9˜4×.Ñ.¬t°C«yÑ9¸#Ó>¸qÑAÓBÐBä˜3“x�rX   ú[z] z]: z: ©r“  r   rµ   r²   )	Úvarsrì   r  ÚlistÚrangeÚlenr  ÚappendÚdict)
r¡   Úinpr™  ÚlinesÚattrr“  ÚiiÚhÚkÚvs
   `         rV   Údebug_lineszFxGraphCachePickler.debug_linesi  s)  ø€ õ		 ð ˆÜ˜c›Ÿ™Ó*ò 	>‰IˆD�#Ü˜#œtÔ$Ü¤ C£›/ò L�BØŸ™ c¨"¡gÓ.�AØ—L‘L 1 Q C r¨$¨¨q°°°C¹ÀÀBÁÓ8HÐ7IÐ!JÕKñLô ˜C¤Ô&ØŸI™I›Kò E‘D�A�qØŸ™ aÓ(�AØ—L‘L 1 Q C r¨$¨¨q°°°3±w¸q³z°lÐ!CÕDñEð —M‘M #Ó&�Ø—‘˜q   2 d V¨2©g°c«l¨^Ð<Õ=ð	>ð ˆrX   ©F)rn  útorch.fx.GraphModulero  ræ   rµ   r¹   )rS  r   rµ   z.tuple[Callable[[T], T], tuple[TensorMetadata]])rS  r   rµ   zNtuple[Callable[[T], T], tuple[Union[TensorMetadata, TensorMetadataAndValues]]])r~  r   rµ   z#tuple[Callable[[T], T], tuple[str]])r~  r   rµ   r   )rn  r«  rµ   z&tuple[Any, tuple[dict[str, Any], str]])r“  r   rµ   r  r›  )r¢  ÚFxGraphHashDetailsrµ   ú	list[str])rº   r»   r¼   rF  r¢   r`  ra  re  rj  rk  r”   r  r©  Ú__classcell__)rl  s   @rV   rW  rW  Ø  s‹   ø„ ñð 16ð"à ð"ð *.ð"ð 
õ	"ðH%Øð%à	7ó%ð FØð Fà	Wó FóD#ó7ð#Ø&ð#à	/ó#ó&%ó ,÷rX   rW  c                óü  — t        t        j                  | |«      d„ ¬«      D ]Ë  }|j                  j	                  |j
                  d «      }|€J ‚|j                  }|€J ‚t        |d«      5 }|j                  |j
                  j                  d«      «       |j                  |j                  «       «       d d d «       |j                  sŒ¨t        |j                  |j
                  › d�|«       ŒÍ y # 1 sw Y   Œ=xY w)Nc                ó   — | j                   S rQ   )r„   rI  s    rV   ú<lambda>z!build_code_hash.<locals>.<lambda>Ž  s
   € ÈÏÉ€ rX   ©rÅ   Úrbr‰   r  )ÚsortedÚpkgutilÚiter_modulesÚmodule_finderÚ	find_specr„   Úoriginr¦   r_  r•   ÚreadÚispkgÚbuild_code_hashÚsubmodule_search_locations)ÚrootsÚprefixÚhasherÚlibÚspecÚmoduler?  s          rV   r¼  r¼  ‹  sÚ   € ô ”g×*Ñ*¨5°&Ó9Ñ?OÔPò 
VˆØ× Ñ ×*Ñ*¨3¯8©8°TÓ:ˆØÐÐÐØ—‘ˆØÐ!Ð!Ð!Ü�&˜$Óð 	$ 1Ø�M‰M˜$Ÿ)™)×*Ñ*¨7Ó3Ô4Ø�M‰M˜!Ÿ&™&›(Ô#÷	$ð �9‹9ä˜D×;Ñ;ÀÇ	Á	¸{È!¸_ÈfÕUñ
V÷
	$ð 	$ús   Á,A
C2Ã2C;	c                 ó
  — t        dd¬«      5  t        j                  «       sd
d„}  | t        «      cddd«       S ddlm} |j                  d«      j                  «       j                  d	«      cddd«       S # 1 sw Y   yxY w)zS
    Compute a key that contains relevant information about torch source files
    Úinductor_codecache_torch_keyT)Úlog_pt2_compile_eventc                ó>  — d}t         j                  j                  t        «      }|D �cg c]"  }t         j                  j	                  ||«      ‘Œ$ }}t        j                  «       }|j                  t        j                  j                  d«      «       t        | gd|«       |D ]V  }t         j                  j                  |«      sŒ#t        |d«      5 }|j                  |j                  «       «       d d d «       ŒX |j                  «       S c c}w # 1 sw Y   ŒwxY w)N)z"codegen/aoti_runtime/interface.cppúcodegen/cpp_prefix.hz	script.ldr‰   r‚  r³  )rz   r{   ÚdirnameÚ__file__r|   r‘   r’   r_  rt   Ú__version__r•   r¼  rø   r¦   rº  r   )ÚrootÚextra_filesÚinductor_rootrJ  rÀ  r{   r?  s          rV   Úget_code_hashz torch_key.<locals>.get_code_hash£  sá   € ð�ô
 !#§¡§¡´Ó 9�ØGRÖSÀ!œrŸw™wŸ|™|¨M¸1Õ=ÐS�ÐSÜ Ÿ™Ó)�Ø—‘œe×/Ñ/×6Ñ6°wÓ?Ô@Ü  ¨¨FÔ3Ø'ò 4�DÜ—w‘w—~‘~ dÕ+Ü! $¨Ó-ð 4°Ø"ŸM™M¨!¯&©&«(Ô3÷4ð 4ð4ð —}‘}“Ð&ùò T÷4ð 4ús   ª'DÃ DÄD	Nr   ©Úparutilztorch/src_hash.txtÚascii)rÌ  r²   rµ   r  )	r   r   Ú	is_fbcoder%   Úlibfb.pyrÑ  Úget_file_contentsÚrstripr•   )rÏ  rÑ  s     rV   Ú	torch_keyr×  ›  sw   € ô
 
Ð4ÈDÔ	Qñ XÜ×ÑÔ!ó'ñ( !¤Ó-÷/Xñ Xõ2 	%à×(Ñ(Ð)=Ó>×EÑEÓG×NÑNÈwÓW÷7X÷ Xò Xús   Ž$A9¼3A9Á9Bc                 óH   — t         j                  j                  t        «      S rQ   )rz   r{   rÉ  rÊ  rR   rX   rV   Úget_inductor_rootrÙ  ¾  s   € Ü�7‰7�?‰?œ8Ó$Ð$rX   c                  ó   — e Zd ZU dZded<   y)ÚOrderedSetHolderzb
    See FxGraphHashDetails. Holds a sorted list to support stable hashing
    of set kwargs.
    rC  rì   NrE  rR   rX   rV   rÛ  rÛ  Â  s   … ñð
 ÔrX   rÛ  c                  ó   — e Zd ZdZy)ru  zI
    Exception to indicate that the FxGraphCache should be bypassed.
    N)rº   r»   r¼   rF  rR   rX   rV   ru  ru  Ì  s   „ òrX   ru  c                  óB   — e Zd ZdZdgZ	 	 	 	 	 	 	 	 	 	 dd„Z	 	 	 	 dd„Zy)r¬  zz
    Object to capture all the details for a compiled FX graph relevant to computing
    a safe and stable cache key.
    Úgraph_idc                ó„  — || _         || _        t        j                  | _        i | _        t        |j                  «       «      D ]\  \  }}|| j                  vsŒt        |«      t        t        fv r"t        t        |«      «      | j                  |<   ŒN|| j                  |<   Œ^ ddlm}m}m}	 ddlm}
 g | _        |��@|j'                  «       D �],  }t)        |t*        j,                  j.                  «      sŒ)t1        j2                  |j4                  j7                  d|¬«      |j4                  j7                  d|	¬«      «      D ]µ  }ddlm} |j=                  |j>                  d   «      }d }t)        ||«      r=|j@                  r%tC        t        d„ |j@                  D «       «      «      }|jD                  } |
|«      }|jG                  |j>                  d	   «      }| j$                  jI                  |||f«       Œ· �Œ/ || _%        tM        d
„ |D «       «       }|rAt*        jN                  jQ                  «       r#t*        jN                  jS                  «       | _*        t+        jV                  «       t+        jX                  «       t*        jZ                  j\                  j^                  f| _0        t*        jb                  jd                  jf                  jh                  t*        jb                  jd                  jf                  jj                  t*        jb                  jd                  jf                  jl                  f| _7        tq        «       | _9        tt        jw                  «       | _<        t{        j|                  «       | _?        | j�                  tz        j‚                  «      | _A        | j�                  tz        j„                  «      | _B        y )Nr   )Úkernel_side_tableÚ triton_kernel_wrapper_functionalÚtriton_kernel_wrapper_mutation)Ú9user_defined_triton_kernel_transitive_closure_source_codeÚcall_function)rß   r-  )Ú	AutotunerÚ
kernel_idxc              3  óx   K  — | ]2  }t        d „ |j                  «       j                  «       D «       «      –— Œ4 y­w)c              3  ó2   K  — | ]  }t        |«      –— Œ y ­wrQ   )r²   )rä   Úkvs     rV   rå   z8FxGraphHashDetails.__init__.<locals>.<genexpr>.<genexpr>  s   è ø€ Ò*T°r¬3¨r¯7Ñ*Tùs   ‚N)r´  Ú
all_kwargsrì   )rä   r  s     rV   rå   z.FxGraphHashDetails.__init__.<locals>.<genexpr>  s2   è ø€ ò '"à()ô %+Ñ*T¸Q¿\¹\»^×=QÑ=QÓ=SÔ*T×$Tñ'"ùs   ‚8:Úconstant_args_idxc              3  óP   K  — | ]  }t        |t        j                  «      –— Œ  y ­wrQ   )r  rt   r   )rä   rJ  s     rV   rå   z.FxGraphHashDetails.__init__.<locals>.<genexpr>(  s   è ø€ Ò"WÀ1¤:¨a´·±×#>Ñ"Wùs   ‚$&)Crn  Úexample_inputsÚcconfigÚcache_key_tagÚ	fx_kwargsr´  rì   ÚEXCLUDED_KWARGSr˜  ÚsetrD   rÛ  Ú*torch._higher_order_ops.triton_kernel_wraprà  rá  râ  Útorch._inductor.codegen.wrapperrã  Úuser_defined_triton_sourceÚmodulesr  rt   rf  ÚGraphModuleÚ	itertoolsÚchainrs  Ú
find_nodesÚtriton.runtime.autotunerrå  Ú
get_kernelrU   Úconfigsr²   r†  Úget_constant_argsr   Úinputs_to_checkÚanyÚacceleratorÚis_availableÚcurrent_device_indexÚdefault_cuda_device_indexÚ$are_deterministic_algorithms_enabledÚ-is_deterministic_algorithms_warn_only_enabledÚutilsÚdeterministicÚfill_uninitialized_memoryÚ!deterministic_algorithms_settingsÚbackendsr‡   ÚmatmulÚ
allow_tf32Ú&allow_fp16_reduced_precision_reductionÚ&allow_bf16_reduced_precision_reductionÚcuda_matmul_settingsr×  Útorch_versionr}   r~   Úsystem_infor   Úsave_config_portableÚinductor_configÚ_get_custom_pass_detailÚpost_grad_custom_pre_passÚpost_grad_custom_post_pass)r¡   rn  rí  rð  rÿ  r§  r¨  rà  rá  râ  rã  rÃ  Únoderå  Úkernelrý  Úkernel_sourceÚconstant_argsÚno_tensor_inputss                      rV   r¢   zFxGraphHashDetails.__init__Û  s$  € ð ˆŒØ,ˆÔÜ$×2Ñ2ˆÔð
 -/ˆŒÜ˜9Ÿ?™?Ó,Ó-ò 	*‰DˆAˆqØ˜×,Ñ,Ò,Ü˜“7œs¤JÐ/Ñ/ô )9¼À»Ó(C�D—N‘N 1Ò%à()�D—N‘N 1Ò%ð	*÷	
ñ 	
õ
	
ð 68ˆÔ'Ø‰>ØŸ*™*›,ó #�Ü! &¬%¯(©(×*>Ñ*>Ô?ØÜ%ŸO™OØ—L‘L×+Ñ+Ø*Ð3Sð ,ó ð —L‘L×+Ñ+Ø*Ð3Qð ,ó ó	ò  �Dõ Cà.×9Ñ9¸$¿+¹+ÀlÑ:SÓT�FØ"�GÜ! &¨)Ô4Ø!Ÿ>š>Ü&)Ü &ñ '"à-3¯^©^ô'"ó !"ó'˜Gð "(§¡˜ñ RØ"óð "ð
 %6×$GÑ$GØŸ™Ð$7Ñ8ó%�Mð ×3Ñ3×:Ñ:Ø&¨°wÐ?õò= ð#ðL  /ˆÔä"Ñ"WÈÔ"WÓWÐWÐñ ¤× 1Ñ 1× >Ñ >Ô @Ü-2×->Ñ->×-SÑ-SÓ-UˆDÔ*ô ×6Ñ6Ó8Ü×?Ñ?ÓAÜ�K‰K×%Ñ%×?Ñ?ð2
ˆÔ.ô �N‰N×Ñ×&Ñ&×1Ñ1Ü�N‰N×Ñ×&Ñ&×MÑMÜ�N‰N×Ñ×&Ñ&×MÑMð%
ˆÔ!ô '›[ˆÔÜ$×/Ñ/Ó1ˆÔÜ%×:Ñ:Ó<ˆÔà)-×)EÑ)EÜ×,Ñ,ó*
ˆÔ&ð +/×*FÑ*FÜ×-Ñ-ó+
ˆÕ'rX   c                óL   — |sy t        |t        «      sJ ‚|j                  «       S rQ   )r  r.   Úuuid)r¡   Úcustom_passs     rV   r  z*FxGraphHashDetails._get_custom_pass_detailI  s)   € ñ ØÜ˜+¤Ô7Ð7Ð7Ø×ÑÓ!Ð!rX   N)
rn  r«  rí  úSequence[InputType]rð  ra   rÿ  úSequence[int]rµ   r¹   )r  r/   rµ   zOptional[Any])rº   r»   r¼   rF  rñ  r¢   r  rR   rX   rV   r¬  r¬  Ò  sa   „ ñð "�l€Oðl
à ðl
ð ,ðl
ð $ð	l
ð
 'ðl
ð 
ól
ð\"Ø.ð"à	ô"rX   r¬  c                ó  — t        | |||«      }t        |j                  «      dk7  }t        | |«      }d|j	                  |«      z   }|j                  |«      }dj                  |«      }	t        j                  d|› d|	› �«       ||fS )z=
    Generate a unique hash of the FX graph for caching.
    r   r?  ú
z$FX graph cache hash details for key z:
)	r¬  rŸ  rõ  rW  r  r©  r|   r‘  Údebug)
rn  rí  rð  rÿ  Údetailsro  ÚpicklerrÅ   r©  Ú	debug_strs
             rV   Úcompiled_fx_graph_hashr(  R  s’   € ô !  ^°YÀÓP€GÜ&)¨'×*LÑ*LÓ&MÐQRÑ&RÐ#Ü! "Ð&EÓF€Gð �× Ñ  Ó)Ñ
)€CØ×%Ñ% gÓ.€KØ—	‘	˜+Ó&€IÜ‡I�IÐ4°S°E¸¸Y¸KÐHÔIØ�ÐÐrX   c                óÖ  — t         j                  j                  «       rt         j                  j                  «       syt	        | dz  «      }t        j                  «       rJt         j                  j                  d«      }t        j                  d||«       |t	        ||z  dz  «      z  }t        j                  d|«       t        j                  j                  t        |¬«      «       |S )z}
    Ephemerally increases the NCCL timeout when compiling for a distributed job
    Returns amount of seconds increased
    r   g    eÍÍAz>pytorch/remote_cache:ephemeral_timeout_fudge_factor_percentagezNEphemeral NCCL timeout increase fudge factor %d and original increase value %déd   zIncreasing NCCL timeout by %d)Úseconds)rt   Údistributedr  Úis_initializedÚintr   rÓ  Ú_utils_internalÚjustknobs_getval_intr‘  ÚinfoÚdistÚdistributed_c10dÚ"_add_ephemeral_timeout_for_all_pgsr	   )Útime_saved_nsÚincreased_timeout_secÚfudge_factors      rV   Ú.add_ephemeral_timeout_increase_for_distributedr8  h  sÈ   € ô
 ×Ñ×)Ñ)Ô+´5×3DÑ3D×3SÑ3SÔ3UØä °Ñ 4Ó5Ðä×ÑÔÜ×,Ñ,×AÑAØLó
ˆô 	�‰Ø\ØØ!ô	
ð
 	¤Ð%:¸\Ñ%IÈCÑ%OÓ!PÑPÐä‡H�HÐ,Ð.CÔDÜ×Ñ×<Ñ<ÜÐ/Ô0ôð !Ð rX   c                  ób  — e Zd ZdZedd„«       Zedd„«       Zedd„«       Zedd„«       Ze	 	 	 	 	 	 	 	 	 	 	 	 dd„«       Z	edd„«       Z
e	 	 	 	 	 	 	 	 	 	 	 	 dd„«       Zedd	„«       Zedd
„«       Ze	 	 	 	 	 	 	 	 	 	 	 	 dd„«       Zedd„«       Ze	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 dd„«       Zedd„«       Zy)ÚFxGraphCachea7  
    Supports caching and reusing compiled Fx graphs.

    The overall strategy is as follows:
    - This cache stores entries on disk. When saving an entry, we can't
      serialize callables (that could be C++, Triton, etc.), so we serialize
      their own disk cache location. We then recreate the compiled artifact
      after fetching from disk.
    - For indexing the cache, we gather the fields relevant to identifying an
      FxGraph (the graph module, graph inputs, system settings etc.) into an
      FxGraphCacheDetails object, pickle it, and compute a hash for the key.
      See FxGraphCachePickler.
    - Among the metadata we store, we also include a guards expression that's
      appropriate for validating any symbols for Tensor arguments that have
      symbolic bounds. On cache lookup then, we evaluate those guards in the
      current context to validate that a cached entry can be served.
    - A given graph could have multiple compiled versions, corresponding to
      different sets of guards. Therefore, we store cache entries in the form:
          <temp dir>/<fx graph hash>/<serialized metatdata>
    - On lookup, we compute the key from the graph details, iterate over all
      leaf files in the corresponding subdirectory, deserialize the entry, and
      evaluate its guards expression. If the evaluation succeeds, we have a
      cache hit. If it fails, we compile the graph and store a new entry.
    - Finally, on a cache hit, we need to make sure any guards that would
      have been created during compilation are added to the current context.
    c                 óR   — t         j                  j                  t        «       d«      S )zS
        Get the toplevel temporary directory for storing compiled graphs.
        Úfxgraph)rz   r{   r|   r4   rR   rX   rV   Ú_get_tmp_dirzFxGraphCache._get_tmp_dir¢  s   € ô
 �w‰w�|‰|œI›K¨Ó3Ð3rX   c                ón   — t         j                  j                  t        j	                  «       | dd | «      S )zA
        Return the disk location for a given cache key.
        rE   r  )rz   r{   r|   r:  r=  r²  s    rV   Ú_get_tmp_dir_for_keyz!FxGraphCache._get_tmp_dir_for_key©  s*   € ô
 �w‰w�|‰|œL×5Ñ5Ó7¸¸Q¸q¸À3ÓGÐGrX   c                óz   — | D �cg c]+  }t        |t        j                  «      sŒt        |«      sŒ*|‘Œ- c}S c c}w )z�
        Get the backed SymInt objects from the input list. Note that we can never
        have guards that depend on unbacked symint.
        )r  rt   r   rA   )rÞ   r~  s     rV   Ú_filter_backed_symintsz#FxGraphCache._filter_backed_symints°  s+   € ð "ÖQ�a¤Z°´5·<±<Õ%@ÄXÈaÅ[’ÒQÐQùÒQs   …8¥8±8c                 ó„   — t         j                  j                  j                  «       } | sy| j                  j
                  S )zG
        Helper to get the shape env from the tracing context.
        N)rt   Ú_guardsÚTracingContextÚtry_getÚ	fake_modeÚ	shape_env)Úctxs    rV   Ú_get_shape_envzFxGraphCache._get_shape_env¸  s2   € ô
 �m‰m×*Ñ*×2Ñ2Ó4ˆÙØØ�}‰}×&Ñ&Ð&rX   c                óÚ  ‡ ‡‡‡‡— t         j                  «       }|€J ‚t         j                  |«      }|D �cg c]  }t        |«      ‘Œ }}dˆ ˆˆfd„}	d}
d}t	        «       } |	«       D ]c  \  }}|j
                  s|}
 nPt        |j                  |j
                  |«      «      }t        j                  d‰ |j
                  ||«       |sŒa|}
 n |
€d|fS |�%t        j                  t        j                  ‰ |«       |
j                  x}rut        j                   |«      }|x}�\t#        |«      |d<   t%        j&                  d|j(                  ¬«       t+        |j(                  «      dkD  rt%        j,                  d«       	 |
j/                  |«      Šd	d
lm} |j4                  � |j4                  |
j6                  «       t;        j<                  «       }|
j6                  Št?        j@                  |‰¬«       |
j
                  rLt        |j                  |
j
                  |«      «      }|du sJ ‚t        j                  d‰ |jB                  «       tD        jF                  jI                  |
jJ                  «       tL        dxx   |
jN                  z  cc<   tP        j                  d‰«       tP        j                  d‰«       tS        dˆfd„ˆfd„¬«       |
|fS c c}w # t8        $ r d|fcY S w xY w)z˜
        Lookup a compiled graph in the cache by key. On a hit, return the
        deserialized CompiledFxGraph object. On a miss, return None.
        Nc               3  ó  •K  — ‰r´t         j                  ‰«      } t        j                  j	                  | «      r€t        t        j                  | «      «      D ]_  }	 t        t        j                  j                  | |«      d«      5 }|j                  «       }t        j                  |«      |f–— d d d «       Œa ‰rs	 ‰j                  ‰«      x}�^t!        |t"        «      sJ ‚|d   }t!        |t$        t&        f«      sJ ‚t)        j*                  |«      }t        j                  |«      |f–— y y y # 1 sw Y   Œ�xY w# t        $ r t        j                  dd¬«       Y �Œw xY w# t        $ r t        j                  dd¬«       Y y w xY w­w)Nr³  z,fx graph cache unable to load compiled graphTr‰  r  )r:  r?  rz   r{   rø   r´  Úlistdirr¦   r|   rº  ÚpickleÚloadsÚ	Exceptionr‘  r’  rØ   r  r¡  r²   r  rþ   Ú	b64decode)	r  r{   r?  r  Ú
cache_datar  rÅ   ÚlocalÚremote_caches	         €€€rV   Úiterate_over_candidatesz;FxGraphCache._lookup_graph.<locals>.iterate_over_candidatesÔ  sl  øè ø€ ñ Ü%×:Ñ:¸3Ó?�Ü—7‘7—>‘> &Ô)Ü &¤r§z¡z°&Ó'9Ó :ò 	˜ðÜ!%¤b§g¡g§l¡l°6¸4Ó&@À$Ó!Gð EÈ1Ø*+¯&©&«( Ü&,§l¡l°7Ó&;¸WÐ&DÒ D÷Eøð	ñ ð
Ø&2×&6Ñ&6°sÓ&;Ð;˜
ÐHÜ)¨*´dÔ;Ð;Ð;Ø)¨&Ñ1˜Ü)¨$´´e°Ô=Ð=Ð=Ü"(×"2Ñ"2°4Ó"8˜Ü$Ÿl™l¨7Ó3°WÐ<Ó<ð Ið ÷Eð Eûô  )ò ÜŸK™KØ NØ)-ð (÷ ðûô !ò Ü—K‘KØFÐQUð  ö ðüsg   ƒAF	Á*D<Â*D0Â0D<Â8F	Â=A0E# Ä-F	Ä0D9	Ä5D<Ä< E ÅF	ÅE Å F	Å# FÆF	ÆFÆF	zEfx graph cache key %s evaluating guards [%s] with values %s => hit=%sÚtriton_bundler_metaÚinductor_compile)Úcached_kernel_namesr   Únum_triton_bundlesrE   rc   ©r	  Tz*fx graph cache key %s post-load guards: %sÚinductorzOutput code: 
%szOutput code written to: %sÚinductor_output_codec                 ó   •— d‰ iS )NÚfilenamerR   )Úartifact_paths   €rV   r±  z,FxGraphCache._lookup_graph.<locals>.<lambda>J  s   ø€ �Z Ð/€ rX   c                 ó   •— ‰ S rQ   rR   rY  s   €rV   r±  z,FxGraphCache._lookup_graph.<locals>.<lambda>K  s   ø€ ˜t€ rX   ©Ú
payload_fn)rµ   z4Generator[tuple[CompiledFxGraph, bytes], None, None])*r:  rI  rA  rB   r¡  Úguards_exprræ   Úevaluate_guards_expressionr‘  r$  r?   Úrecord_artifactr@   ÚINDUCTORÚ_triton_bundlerJ   Úread_and_emitr²   r   Útry_add_pt2_compilerW  rŸ  Úincrement_toplevelÚafter_deserializationrs  rd   Úsave_output_codeÚsource_codeÚOSErrorrH   Úinductor_meta_from_configrI   Úbegin_compileÚguardsr   ÚCachedMetricsHelperÚapply_deltasÚmetrics_deltasr   Úcounter_deltasÚoutput_code_logr:   )rÅ   rí  rR  rS  Ú	constantsrG  Úsymintsr~  ÚhintsrT  rs  Úpickled_contentÚ
cache_infoÚ	candidaterÚ   ÚbundlerU  rT  rd   Úinductor_metaÚcheckr^  r	  s   ` ``                 @@rV   Ú_lookup_graphzFxGraphCache._lookup_graphÂ  s×  ü€ ô !×/Ñ/Ó1ˆ	ØÐ$Ð$Ð$ä×5Ñ5°nÓEˆØ&-Ö. ”˜!•Ð.ˆÐ.÷	ð@ ˆØˆÜ%)£Vˆ
á*AÓ*Cò 	Ñ&ˆI�Ø×(Ò(à!�Ùô Ø×4Ñ4°Y×5JÑ5JÈEÓRóˆCô �I‰IØWØØ×%Ñ%ØØôò Ø!�Ùð-	ð0 ˆ=Ø˜Ð#Ð#àÐ&Ü ×0Ñ0Ü!×*Ñ*¨C°ôð ×)Ñ)Ð)ˆ6Ð)Ü"/×"=Ñ"=¸fÓ"EÐØ+Ð+�Ð8Ü47¸³I�
Ð0Ñ1ä"×6Ñ6Ø&¸D×<TÑ<Tõô �t×/Ñ/Ó0°1Ò4Ü&×9Ñ9Ð:NÔOð	$Ø!×7Ñ7¸	ÓBˆMå,ð ×-Ñ-Ð9Ø.�×.Ñ.¨u×/@Ñ/@ÔAô '×@Ñ@ÓBˆØ× Ñ ˆÜ×*Ñ*¨=¸tÕDð ×ÒÜØ×4Ñ4°U×5FÑ5FÈÓPóˆEð ˜D‘=Ð �=Ü�I‰IØ<¸cÀ9×CSÑCSôô 	×#Ñ#×0Ñ0°×1EÑ1EÔFÜ�Ó × 4Ñ 4Ñ4Óä×ÑÐ1°4Ô8Ü×ÑÐ:¸MÔJäØ"Û/Û#õ	
ð
 �jÐ Ð ùòw /øôt ò 	$ð ˜Ð#Ò#ð	$ús   ·KÅ??K ËK*Ë)K*c                ó  — t         j                  | «      }t        j                  j	                  |«      st        j
                  |d¬«       t        j                  j                  |t        |«      «      }t        ||d¬«       y )NTrö   r¯   )	r:  r?  rz   r{   rø   rù   r|   r  r±   )rÅ   r  r  r{   s       rV   Ú_write_to_local_cachez"FxGraphCache._write_to_local_cacheO  sW   € ä×2Ñ2°3Ó7ˆÜ�w‰w�~‰~˜fÔ%Ü�K‰K˜¨Õ.ô
 �w‰w�|‰|˜F¤K°Ó$8Ó9ˆÜ�T˜7¨dÖ3rX   c                ór  — ddl m} t        ||«      sJ dt        |«      › d�«       ‚t	        |«      }|j                  «        t        j                  «       }|€J ‚t        j                  |«      }|j                  |«      }	|j                  ||	¬«      |_        	 t        j                  |«      }
	 t%        j&                  t(        j*                  | |
«       |rt        j-                  | |
«       |rVt/        |j0                  xs ddz  «      }t3        j4                  |
«      j7                  d«      |dœ}|j9                  | |«       yy# t        $ r. t        j!                  dd¬	«       t"        d
   dxx   dz  cc<   Y yw xY w# t        $ r. t        j!                  dd¬	«       t"        d
   dxx   dz  cc<   Y yw xY w)z=
        Store a serialized CompiledFxGraph on disk.
        rE   )rb   zserialization for z NYIN)Úplaceholdersrp  z1fx graph cache unable to serialize compiled graphTr‰  rZ  Úfxgraph_cache_pickle_errorr   g    €„.ArÒ  )r  Útime_taken_msz!fx graph unable to write to cacheÚfxgraph_cache_write_error)Ú
compile_fxrb   r  r˜  r   Úprepare_for_serializationr:  rI  rA  Úget_pruned_guardsÚproduce_guards_expressionrb  rM  r”   rO  r‘  r’  r   r?   rd  r@   re  r�  r.  Ú_time_taken_nsrþ   Ú	b64encoder  Úput)rÅ   Úcompiled_graphrí  rR  rS  rb   Údisk_compiled_graphrG  rw  rp  r  r…  rQ  s                rV   Ú_save_graphzFxGraphCache._save_graph[  sÂ  € õ 	0ä˜.¨/Ô:ð 	
Ø ¤ nÓ!5Ð 6°dÐ;ó	
Ð:ô # >Ó2ÐØ×5Ñ5Ô7ô !×/Ñ/Ó1ˆ	ØÐ$Ð$Ð$Ü×5Ñ5°nÓEˆØ×,Ñ,¨WÓ5ˆØ*3×*MÑ*MØ ¨ð +Nó +
ÐÔ'ð	Ü—l‘lÐ#6Ó7ˆGð	CÜ ×0Ñ0Ü!×*Ñ*¨C°ôñ Ü×2Ñ2°3¸Ô@áÜ #Ð%8×%GÑ%GÒ%LÈ1ÐQTÑ$TÓ U�ä"×,Ñ,¨WÓ5×<Ñ<¸WÓEØ%2ñ*�
ð × Ñ   jÕ1ð øô ò 	Ü�K‰KØCÈdð ô ô �ZÑ Ð!=Ó>À!ÑCÓ>Ùð	ûô* ò 	CÜ�K‰KÐ;ÀdˆKÔKÜ�ZÑ Ð!<Ó=ÀÑBÕ=ð	Cús%   ÂE Â/BE? Å4E<Å;E<Å?4F6Æ5F6c                ó4  — | j                  «       D �]  }t        |t        j                  j                  «      sŒ)|j
                  j                  D ]Ã  }t        |j                  t        j                  j                  «      r@|j                  j                  «       s&t        d|j                  j                  «       › �«      ‚|j                  dk(  sŒ�t        t        | |j                  «      t        j                  j                   «      sŒºt        d«      ‚ �Œ y )Nz!Can't cache HigherOrderOperator: ÚgetattrzCan't cache torchbind objects)rö  r  rt   rf  r÷  rs  Únodesr-  Ú_opsÚHigherOrderOperatorÚ	cacheableru  r„   rß   r’  Ú_CÚScriptObject)rn  rÃ  r  s      rV   Ú_check_for_hopzFxGraphCache._check_for_hop–  sÕ   € à—j‘j“ló 	NˆFÜ˜f¤e§h¡h×&:Ñ&:Ô;ØØŸ™×*Ñ*ò N�ä˜tŸ{™{¬E¯J©J×,JÑ,JÔKØ ŸK™K×1Ñ1Ô3ä,Ø;¸D¿K¹K×<LÑ<LÓ<NÐ;OÐPóð ð —7‘7˜iÓ'¬JÜ˜B §¡Ó,¬e¯h©h×.CÑ.Cõ-ô -Ð-LÓMÐMòNñ	NrX   c                óJ  — t         j                  t         j                  fD ]0  }|sŒt        |t        «      r|j                  «       rŒ't        d«      ‚ t        | «      r*t        j                  j                  d«      st        d«      ‚t         j                  j                  rt        d«      ‚ddlm} |j                  rt         j#                  d«       t        ‚t$        j'                  «       € t         j#                  d	«       t        d
«      ‚t$        j)                  | «       y)z“
        Check some conditions that would preclude caching and raise BypassFxGraphCache
        to bypass in case caching is not possible.
        z!Unsupported post grad custom passz,pytorch/inductor:allow_freezing_with_cachingz$Skipping graph with frozen constantszORuntime constant folding can introduce constants that aren't static across runsr   )ÚCompilerBisectorz$dont cache graph when bisect enabledNzfx graph cache no shape envzNo shape env)r   r  r  r  r.   r  ru  r0   rt   r/  Újustknobs_checkÚaot_inductorÚuse_runtime_constant_foldingÚ!torch._inductor.compiler_bisectorr›  Úbisection_enabledr‘  r$  r:  rI  r™  )rn  Úpr›  s      rV   Ú_check_can_cachezFxGraphCache._check_can_cache¨  sð   € ô ×2Ñ2´F×4UÑ4UÐVò 	NˆAÚœ* Q¬Ô8ÀÇÁÅÜ(Ð)LÓMÐMð	Nô
 ˜RÔ ¬×)>Ñ)>×)NÑ)NØ:ô*
ô %Ð%KÓLÐLä×Ñ×;Ò;Ü$ð%óð õ
 	Gà×-Ò-Ü�I‰IÐ<Ô=Ü$Ð$ô ×&Ñ&Ó(Ð0Ü�I‰IÐ3Ô4Ü$ ^Ó4Ð4ô 	×#Ñ# BÕ'rX   c                ó@  — 	 t         j                  | «       t        | |||«      \  }}||fi fS # t        $ rf}t        d   dxx   dz  cc<   t
        j                  d|«       |rt        dt        |«      «       dt        |«      t        «       dœ}d|fcY d}~S d}~ww xY w)	aÐ  
        Checks that the inductor input is cacheable, then computes
        and returns the cache key for the input.
        Returns (key_info, cache_info) where:
        - key_info is (hash_key, debug_lines), and
        - cache_info will contain debug info in the event of BypassFxGraphCache.

        NB: It is possible to have this function return a union instead. But
        I personally believe it is more annoying/difficult to read in that format.
        rZ  Úfxgraph_cache_bypassrE   z%Bypassing FX Graph Cache because '%s'Úbypass_fx_graphÚbypass)Úcache_stateÚcache_bypass_reasonÚcache_event_timeN)
r:  r¢  r(  ru  r   r‘  r1  r>   r²   r   )	rn  rí  rð  rÿ  ÚremoterÅ   r©  rò   rz  s	            rV   Úprepare_keyzFxGraphCache.prepare_keyÏ  s«   € ð$	$Ü×)Ñ)¨"Ô-Ü5Ø�N I¨ó ÑˆC�ð �[Ð! 2Ð%Ð%øô "ò 
	$Ü�ZÑ Ð!7Ó8¸AÑ=Ó8Ü�H‰HÐ<¸aÔ@ÙÜ Ð!2´C¸³FÔ;à'Ü'*¨1£vÜ$+£IñˆJð
 ˜Ð#Õ#ûð
	$ús   ‚&. ®	B·ABÂBÂBc                 óF   — d} t        | t        j                  «       dd«      S )zK
        Attempts to load the remote cache, returns None on error.
        zfx-graph-v1ÚFbRemoteFxGraphCacheÚRemoteFxGraphCache)rG   r   rÓ  )Úcache_ids    rV   Úget_remote_cachezFxGraphCache.get_remote_cacheô  s+   € ð
 !ˆÜØÜ×ÑÓØ"Ø ó	
ð 	
rX   c                ód  — t         j                  | ||||«      \  }}i |¥| |t        «       dœ¥}|�¡t        j	                  d| «       t
        d   dxx   dz  cc<   d|d<   |r+t        j                  d«       t        j                  d	| «       |j                  x}	�3|	|d
<   t        j                  d|	dz  «       t        |	«      x}
dk7  r|
|d<   ||fS |r+t        j                  d«       t        j                  d| «       t        j	                  d| «       t
        d   dxx   dz  cc<   d|d<   ||fS )zÒ
        Lookup the graph with the given key, and return results and metadata.
        Doesn't do any logging on its own, because AOTAutograd handles a cache miss
        differently from FXGraphCache.
        )rÅ   Ú
componentsr©  zfx graph cache hit for key %srZ  Úfxgraph_cache_hitrE   rÚ   r§  Ú"inductor_fx_remote_cache_hit_countÚ!inductor_fx_remote_cache_hit_keysr5  Ú distributed_ephemeral_timeout_usiè  r   Úephemeral_timeout_increaseÚ#inductor_fx_remote_cache_miss_countÚ"inductor_fx_remote_cache_miss_keyszfx graph cache miss for key %sÚfxgraph_cache_missÚmiss)r:  r  r   r‘  r1  r   r   ri  Úadd_to_set_toplevelr‹  r8  )rÅ   r©  rí  rR  rS  Úis_backwardrv  rŽ  rz  r5  Úephemeral_increases              rV   Úload_with_keyzFxGraphCache.load_with_key  sw  € ô &2×%?Ñ%?Ø� ¨°ió&
Ñ"ˆ˜
ð
Øð
àØ%Ü '£	ò	
ˆ
ð Ð%Ü�H‰HÐ4°cÔ:Ü�ZÑ Ð!4Ó5¸Ñ:Ó5Ø(-ˆJ�}Ñ%Ùä"×5Ñ5Ø8ôô #×6Ñ6Ø7¸ôð "0×!>Ñ!>Ð>�ÐKØ.;�
˜?Ñ+Ü"×5Ñ5Ø6¸ÈÑ8Môô
 FÀmÓTðUÐ&àòð @R�JÐ;Ñ<ð ˜zÐ)Ð)ñ ä"×5Ñ5Ø9ôô #×6Ñ6Ø8¸#ôô �H‰HÐ5°sÔ;Ü�ZÑ Ð!5Ó6¸!Ñ;Ó6Ø(.ˆJ�}Ñ%à˜zÐ)Ð)rX   c                 ór   — 	 t        j                  t        j                  «       «       y# t        $ r Y yw xY w)z.
        Clear out the on-disk cache.
        N)r8  Úrmtreer:  r=  ÚFileNotFoundErrorrR   rX   rV   ÚclearzFxGraphCache.clear?  s.   € ð
	Ü�M‰Mœ,×3Ñ3Ó5Õ6øÜ ò 	Ùð	ús   ‚'* ª	6µ6N©rµ   r²   )rÅ   r²   rµ   r²   )rÞ   r   rµ   zlist[torch.SymInt])rµ   zOptional[ShapeEnv])rÅ   r²   rí  r   rR  ræ   rS  ú!Optional[RemoteCache[JsonDataTy]]rv  rf   rµ   ú0tuple[Optional[CompiledFxGraph], dict[str, Any]])rÅ   r²   r  r  rµ   r¹   )rÅ   r²   rŽ  rg   rí  r   rR  ræ   rS  rÅ  rµ   r¹   )rn  r«  rµ   r¹   )rn  r«  rí  r   rð  ra   rÿ  r!  rª  ræ   rµ   z6tuple[Optional[tuple[str, list[str]]], dict[str, Any]])rµ   rÅ  )rÅ   r²   r©  r­  rí  r   rR  ræ   rS  rÅ  r½  ræ   rv  rf   rµ   rÆ  r¸   )rº   r»   r¼   rF  r½   r=  r?  rA  rI  r  r�  r�  r™  r¢  r«  r°  r¿  rÃ  rR   rX   rV   r:  r:  „  s  „ ñð: ò4ó ð4ð òHó ðHð òRó ðRð ò'ó ð'ð ðJ!ØðJ!à+ðJ!ð ðJ!ð 8ð	J!ð
 ,ðJ!ð 
:òJ!ó ðJ!ðX ò	4ó ð	4ð ð8CØð8Cà"ð8Cð ,ð8Cð ð	8Cð
 8ð8Cð 
ò8Có ð8Cðt òNó ðNð" ò$(ó ð$(ðL ð"&Ø ð"&à+ð"&ð $ð"&ð 'ð	"&ð
 ð"&ð 
@ò"&ó ð"&ðH ò

ó ð

ð ð;*Øð;*àð;*ð ,ð;*ð ð	;*ð
 8ð;*ð ð;*ð ,ð;*ð 
:ò;*ó ð;*ðz òó ñrX   r:  c                óÊ   — | j                  d«      rt        j                  j                  | «      S | j                  d«      rt        j                  j                  | «      S | dfS )zDReturns the path where the AOT Inductor compiled kernels are stored.ú.soz.pt2r‚  )Úendswithrz   r{   Úsplit)r{   s    rV   Úsplit_aot_inductor_output_pathrË  J  sK   € ð ‡}�}�UÔÜ�w‰w�}‰}˜TÓ"Ð"Ø	�‰�vÔ	Ü�w‰w�}‰}˜TÓ"Ð"à�RˆxˆrX   c                  óv   — e Zd ZU i Zded<    eej                  «      Zedd„«       Z	edd„«       Z
ed	d„«       Zy)
ÚCudaKernelParamCachezdict[str, dict[str, Any]]r›   c                ó¨   — t        |||t        t        j                  j                  «      d   ¬«      \  }}||t        «       <   || j                  |<   y )Nr   )r  r  )r"  rË  r   r�  Úoutput_pathrw   r›   )ÚclsrÅ   Úparamsr  Úbin_typeÚ_r{   s          rV   rò  zCudaKernelParamCache.setZ  sU   € äØØØÜ8Ü×#Ñ#×/Ñ/óàñô	
‰ˆˆ4ð 59ˆÔ.Ó0Ñ1àˆ�	‰	�#ŠrX   c                ó:   — | j                   j                  |d «      S rQ   )r›   rØ   )rÐ  rÅ   s     rV   rØ   zCudaKernelParamCache.geth  s   € à�y‰y�}‰}˜S $Ó'Ð'rX   c                ó6   — | j                   j                  «       S rQ   )r›   rÃ   )rÐ  s    rV   Úget_keyszCudaKernelParamCache.get_keysl  s   € à�y‰y�~‰~ÓÐrX   N)
rÅ   r²   rÑ  zdict[str, str]r  r²   rÒ  r²   rµ   r¹   )rÅ   r²   rµ   rÌ   )rµ   zKeysView[str])rº   r»   r¼   r›   rG  r½   rÃ  Úcache_clearÚclassmethodrò  rØ   rÖ  rR   rX   rV   rÍ  rÍ  U  sU   … à')€EÐ$Ó)Ù˜uŸ{™{Ó+€Kàò ó ð ð ò(ó ð(ð ò ó ñ rX   rÍ  c                  ó:   — e Zd Ze	 	 	 	 	 	 	 	 	 	 	 	 	 	 dd„«       Zy)ÚAotCodeCompilerc          
     óz  ‡‡‡‡‡H‡I‡J‡K‡L‡M— |}t         j                  dk(  rt        d«      ‚t        «        t	        «       }t        ddt        |‰‰j                  ¬«      ¬«      }	t        |	j                  «       «      }
t        j                  «       xr ‰dk(  xr ‰j                  ŠLt        t        j                  j                  «      \  }}t        j                  j                  rdj!                  ‰‰f«      Šd	Št#        ‰d
|
|¬«      \  }ŠMt#        ‰d|
|¬«      \  }ŠJt        j                  j$                  r<|j'                  ‰M«       t        j                  j                  s|j'                  ‰J«       t(        j+                  d‰M«       t(        j+                  d‰J«       t-        dˆMfd„ˆfd„¬«       t-        dˆJfd„ˆfd„¬«       t/        ‰M«      }t/        ‰J«      }|j0                  |z  ŠK‰Kj3                  «       s‰Kj5                  d¬«       t7        t/        ‰K«      dz  «      }dCˆˆˆKˆLfd„}ddlm} t=        «       } |t>        j@                  j!                  ||dz   «      tB        ¬«      }|5  |rkt7        |jE                  d«      «      }tG        |d«      5 }|j#                  |«       ddd«       t        j                  j$                  r|j'                  |«       t        j                  jH                  }‰|d <   t7        |jK                  |jL                  › d!�«      «      }t        j                  jH                  jO                  «       D ],  \  }}tQ        |t6        «      rtQ        |t6        «      rŒ'J d"«       ‚ tG        |d«      5 }|j#                  tS        jT                  t        j                  jH                  «      «       ddd«       t7        |jK                  |jL                  › d!�«      «      }tW        jX                  ||«       t        j                  j$                  r<|j'                  |«       t        j                  j                  s|j'                  |«       |rt        j                  j                  nt7        |jE                  d#«      «      }t[        ˆfd$„‰j\                  j_                  «       D «       «      ŠIdDd%„ŠHt        j                  j`                  r5d&j!                  ˆHˆIˆfd'„‰j\                  j_                  «       D «       «      }nd&}tc        |«      }t        j                  «        xr |d(kD  } t        j                  jd                  rd} ‰j                  ‰| t        j                  «       |d)œ}!t        dEdt        j                  j                   d*œ|!¤Ž}"t        dEd+di|!¤Ž}#t        t7        |jL                  «      ‰Mt7        |j0                  «      |"¬,«      }$|$j                  «       }%|$jg                  «       }&t        t7        |jL                  «      ‰Jt7        |j0                  «      |#¬,«      }'|'j                  «       }(|'jg                  «       })th        jk                  d-|%«       th        jk                  d.|(«       t        j                  j                  r~t7        |jK                  |jL                  › d/�«      «      }*|"jm                  |*«       |j'                  |*«       |$jo                  |«       |$jq                  |‰M«       |j'                  |«       n |$js                  «        |'js                  «        | s|}+d},nxtu        tv        ty        jz                  dty        j|                  tx        j~                  «      j€                  d0«      jƒ                  «       «      },t…        j†                  d1|d2z   |,«      }+ ||+t         j                  «      }-d}.i }/t‰        ‰jŠ                  jO                  «       «      D ]»  \  }.\  }0}1tQ        |1tx        jŒ                  jŽ                  «      sJ ‚t�        › |.› �}2th        jk                  d3|0|2«       |2|/|0<   tx        jŒ                  j“                  |1«      }3t>        j@                  j!                  |j0                  |2«      }4t•        |4|3d«       |j'                  |4«       Œ½ t>        j@                  j!                  |j0                  d4«      }5tG        |5d«      5 }|j#                  tS        jT                  |/«      «       ddd«       |j'                  |5«       tx        j–                  j˜                  r
t›        «       n	t�        «       }6|6jž                  j¡                  «       D �7cg c])  }7|7j                  j£                  d5«      r|7j                  ‘Œ+ }8}7d6j!                  |8«      }8t¥        |«      \  }9}:t        |‰‰j                  ‰L¬7«      };t        |9|8r|&|)|-|8gn|&|)|-g|:|;¬,«      }<|<j                  «       }=|<jg                  «       }th        jk                  d8|=«       tG        ‰Md9«      5 }|j#                  d«       |j#                  d:|%› d�«       |j#                  d;|=› d�«       ddd«       tG        ‰Jd9«      5 }|j#                  d«       |j#                  d:|(› d�«       |j#                  d;|=› d�«       ddd«       t        j                  j                  �rPt7        |jK                  |jL                  › d<�«      «      }>|;jm                  |>«       |j'                  |>«       |j'                  t¦        «       | rƒt7        |jK                  |jL                  › d=�«      «      }?tG        |?d>«      5 }@|@j#                  |«       |@j#                  t…        j†                  d?|,«      «       ddd«       |j'                  |?«       |j'                  |-«       |j'                  |8«       |<jq                  ||-«       |8j©                  «       D ]  }A|<jq                  ||A«       Œ |<j«                  |«       nñ|<js                  «        |&|)|-fD ]  }Bt?        j¬                  |B«       Œ | r•ddlW}C|Cj±                  «       }Dt�        d@|D«      }EtG        |dA«      5 }F|Fj³                  «       }G|Fj#                  dBE|G|Ez  z
  z  «       |Fj#                  |«       |Fj#                  t…        j†                  d?|,«      «       ddd«       t        j                  j$                  r|j'                  |«       ddd«       t        j                  j$                  r|S S # 1 sw Y   �
Œ©xY w# 1 sw Y   �	Œ�xY w# 1 sw Y   �ŒYxY wc c}7w # 1 sw Y   �ŒxY w# 1 sw Y   �ŒÅxY w# 1 sw Y   �ŒåxY w# 1 sw Y   Œ­xY w# 1 sw Y   Œ†xY w)Fzƒ
        Returns the .so path, or returns a list of files that were generated if
        config.aot_inductor.package=True.
        ro   z.AotCodeCompiler not yet supported for inductorÚoÚi)Úvec_isaÚdevice_typeÚaot_mode©r„   ÚsourcesÚBuildOptionÚcpur#  r‚  zwrapper.cpp)r
  r  z
kernel.cppzWrapper code written to: %szKernel code written to: %sÚ
graph_dumpc                 ó   •— dd‰ dœS )NÚinductor_aot_wrapper_codeÚcpp©r„   r˜  r]  rR   )Úwrapper_paths   €rV   r±  z)AotCodeCompiler.compile.<locals>.<lambda>À  s   ø€ Ø3ØØ(ñ€ rX   c                 ó   •— ‰ S rQ   rR   )Úwrapper_codes   €rV   r±  z)AotCodeCompiler.compile.<locals>.<lambda>Å  s   ø€ ˜|€ rX   r`  c                 ó   •— dd‰ dœS )NÚinductor_aot_kernel_coderè  ré  rR   )Úkernel_paths   €rV   r±  z)AotCodeCompiler.compile.<locals>.<lambda>É  s   ø€ Ø2ØØ'ñ€ rX   c                 ó   •— ‰ S rQ   rR   )Úkernel_codes   €rV   r±  z)AotCodeCompiler.compile.<locals>.<lambda>Î  s   ø€ ˜{€ rX   Trö   zCMakeLists.txtc                óŽ  •— |dk(  rQ‰j                   t        ‰j                  j                  «       «      z  rt	        | «      dkD  rt        d«      ‚d}nd}d}n|dk(  rd}d	}nt        d
|› �«      ‚t	        | «      dkD  }d|› d�}|dt        › d�z  }|d|› d�z  }||› d�z  }|s| D ]  }|d|› d�z  }Œ | s |dz  }n|dz  }|dt	        | «      dz
  › d�z  }|d|› d�z  }||› d�z  }t        |dt        ‰«      ¬«      \  }}t        |«      }t        ‰dk7  r‰nd‰j                  d‰¬«      }	t        t        |j                  «      t        |«      t        |j                  «      |	¬ «      }
|
j!                  «       }|
j#                  «        |r˜t%        |d!«      5 }|j'                  d"«       |j)                  d«      }|j+                  d#«      }|d$k7  sJ ‚|j'                  |«       d"}|t	        | «      k  r(|j                  | |d  «      }||z  }|t	        | «      k  rŒ(d d d «       t-        j.                  |«       |S # 1 sw Y   Œ xY w)%NÚlinuxé ”5wzPModels with buffer mutation included doesn't support constants greater than 2GB!z.ldata, "aw"z.lrodata, "a"r‚  Údarwinz__DATA,__datarÓ  zUnsupported platform: i   z
	.section	r#  z		.balign z	.globl	z_binary_constants_bin_start
z_binary_constants_bin_start:
z	.byte z
	.space 1
z	.quad 0x1234567899abcdef
z	.space é   z.globl	z_binary_constants_bin_end
z_binary_constants_bin_end:
ÚS)r  Úxpurä  T)rß  rà  Úcompile_onlyÚuse_relative_path©r„   râ  Ú
output_dirrã  zr+br   s   ïÍ«™xV4rÈ   )Úmutated_buffersrD   rv  rÃ   rŸ  Ú
ValueErrorr�   r6   r"  r²   r   r)   rà  r'   Ústemr0  Úget_target_file_pathÚbuildr¦   r�  rº  Úfindrz   r:  )ÚconstsÚplatformÚsection_attrÚsymbol_prefixÚis_large_constsÚ
consts_asmr  rÓ  Úconsts_sÚobject_build_optionsÚobject_builderÚconsts_or?  ÚhdrÚ	start_idxÚposÚrcrß  rs  Úspecified_sub_dirrú  s                    €€€€rV   Ú_compile_constsz0AotCodeCompiler.compile.<locals>._compile_constsÛ  sŸ  ø€ Ø˜7Ò"Ø×(Ñ(¬:°e·o±o×6JÑ6JÓ6LÓ+MÒMô ˜6“{ ]Ò2Ü(Ønóð ð $2‘Là#2�LØ "‘Ø˜XÒ%Ø.�Ø #‘ä"Ð%;¸H¸:Ð#FÓGÐGä! &›k¨DÑ0ˆOØ'¨ ~°RÐ8ˆJØ˜J¤{ m°2Ð6Ñ6ˆJØ˜J } oÐ5RÐSÑSˆJØ˜]˜OÐ+IÐJÑJˆJÙ"Øò 3�AØ H¨Q¨C¨rÐ"2Ñ2‘Jð3ñ Ø .Ñ0‘JàÐ<Ñ<�
Ø 	¬#¨f«+¸©/Ð):¸"Ð=Ñ=�
Ø˜H ] OÐ3NÐOÑOˆJØ˜]˜OÐ+GÐHÑHˆJÜØØÜ!Ð"3Ó4ô‰KˆAˆxô
 ˜H“~ˆHÜ#8ð ,7¸%Ò+?™KÀUØŸ™Ø!Ø"3ô$Ð ô (Ü˜Ÿ™Ó'Ü˜H›Ü˜xŸ™Ó/Ø0ô	ˆNð &×:Ñ:Ó<ˆHØ× Ñ Ô"áÜ˜( EÓ*ð 
"¨aØ—F‘F˜1”IØŸ&™& ›,�Cà #§¡Ð)LÓ M�IØ$¨š?Ð*˜?Ø—F‘F˜9Ô%Ø�CØ¤ F£Ò+ØŸW™W V¨C¨D \Ó2˜Ø˜r™	˜ð ¤ F£Ó+÷
"ô �I‰I�hÔàˆO÷
"ð 
"ús   ÆBH;È;Ir   ©ÚFileLockú.lock©Útimeoutz.jsonr*  NÚAOTI_DEVICE_KEYz_metadata.jsonz"Metadata must only contain stringsrÈ  c              3  óp   •K  — | ]-  }|‰j                   vr‰j                  |«      j                  –— Œ/ y ­wrQ   )Úfolded_constantsÚget_original_value_of_constantÚis_cuda)rä   r„   rs  s     €rV   rå   z*AotCodeCompiler.compile.<locals>.<genexpr>\  s8   øè ø€ ò àØ˜u×5Ñ5Ñ5ð ×4Ñ4°TÓ:×BÕBñùs   ƒ36c                ó  — dd„}dd l }| j                  «       dk(  ry| j                  rSt        j                  j
                  j                  | «      }t        j                  j
                  j                  | «      }n>| j                  «       j                  «       }|j                  «       }|j                  «       }|j                  ||j                  |j                  |z  «      «      }t        |j                  «      }|r|S  ||«      S )Nc                ól   — | j                  t        | «      t        z   dz
  t        z  t        z  d«      }|S )NrE   ó    )ÚljustrŸ  r6   )Ú	raw_bytesÚpadded_bytess     rV   Ú_pad_to_alignmentzEAotCodeCompiler.compile.<locals>._to_bytes.<locals>._pad_to_alignmentc  s6   € Ø#,§?¡?Ü˜Y›¬+Ñ5¸Ñ9¼kÑIÌKÑWØó$�Lð (Ð'rX   r   rX   )r!  r  rµ   r  )ÚctypesÚnumelrt  rt   ÚopsÚmkldnnÚdata_ptrÚ_nbytesÚuntyped_storagerä  Únbytesr   ÚPOINTERÚc_ubyter  Úcontents)	rS  Úall_cudar#  r$  r(  r+  Út_cpuÚ	raw_arrayr!  s	            rV   Ú	_to_bytesz*AotCodeCompiler.compile.<locals>._to_bytesb  sÐ   € ó(ó à—7‘7“9 ’>Øà—;’;Ü$Ÿy™y×/Ñ/×8Ñ8¸Ó;�HÜ"ŸY™Y×-Ñ-×5Ñ5°aÓ8‘Fà×-Ñ-Ó/×3Ñ3Ó5�EØ$Ÿ~™~Ó/�HØ"Ÿ\™\›^�Fà"ŸK™KØØ—N‘N 6§>¡>°FÑ#:Ó;ó�	ô " )×"4Ñ"4Ó5�	Ù$,�yÐNÑ2CÀIÓ2NÐNrX   rX   c              3  ój   •K  — | ]*  }|‰j                   vr ‰‰j                  |«      ‰«      –— Œ, y ­wrQ   )r  r  )rä   r„   r2  r/  rs  s     €€€rV   rå   z*AotCodeCompiler.compile.<locals>.<genexpr>�  s9   øè ø€ ò .àØ 5×#9Ñ#9Ñ9ñ ˜e×BÑBÀ4ÓHÈ(×Sñ.ùs   ƒ03rô  )rà  rß  Úuse_mmap_weightsrú  rÞ  )rù  Úmin_optimizerù  rû  z#aot wrapper compilation command: %sz"aot kernel compilation command: %sz_compile_flags.json)rE   Úqqrö  zsaving script object %s as %szcustom_objs_config.jsonz.oú )rÞ  rß  rà  rú  zaot linkage command: %sÚaz// Compile cmd
// z// Link cmd
// z_linker_flags.jsonz_serialized_weights.binr+  Úqi @  za+bó    )r  r  r  r²   rµ   r²   )rS  ztorch.Tensorr/  ræ   rµ   r  rR   )ZÚsysr  r�   r$   r-   r'   r)   rà  r  Úget_command_liner   rÓ  rË  r�  rÏ  Úpackage_cpp_onlyr|   r"  Úpackager   ru  r1  r:   r   r0  rø   r1  r²   Útorch.utils._filelockr  rû   rz   r{   ÚLOCK_TIMEOUTÚwith_suffixr¦   rq  Ú	with_namerÿ  rì   r  r“   r”   r8  r   rë   rv  rÃ   Úpackage_constants_in_sorŸ  Úforce_mmap_weightsr   r‘  r$  Úsave_flags_to_jsonÚsave_compile_cmd_to_cmakeÚsave_src_to_cmaker  r   r.  rt   ÚrandintÚiinfoÚint64ÚmaxÚitemÚstructÚpackÚ	enumerateÚtorchbind_constantsr—  r˜  rF   Ú_pickle_saver±   ru   rv   ÚROCmCodeCacheÚCUDACodeCacher›   rD  rÉ  r+   r#   rÊ  Úsave_link_cmd_to_cmaker:  ÚresourceÚgetpagesizeÚtell)NrÐ  rs  rì  rñ  Úserialized_extern_kernel_nodesrß  Úadditional_filesÚgenerated_filesÚpicked_vec_isaÚvec_isa_cmd_genÚcpp_commandÚspecified_output_pathÚspecified_artifact_nameÚwrapper_keyrÓ  Úwrapper_path_operatorÚkernel_path_operatorÚ
cmake_pathr  r  rú   ÚlockÚextern_kernel_nodes_jsonr?  rq  Ú	meta_jsonr§  r¨  Úkernel_meta_jsonÚ	output_soÚserialized_weightsÚconsts_sizer4  Úcompile_commandÚwrapper_build_optionsÚkernel_build_optionsÚwrapper_builderÚwrapper_compile_cmdÚ	wrapper_oÚkernel_builderÚkernel_compile_cmdÚkernel_oÚcompile_flagsÚaot_constantsÚmagic_numberr  Úcustom_obj_idxÚqual_name_to_idr„   ÚconstantÚcustom_obj_nameÚcustom_obj_bytesÚcustom_obj_pathÚconstants_config_jsonÚgpu_codecacheÚentryÚgpu_kernels_oÚoutput_namerü  Úso_build_optionsÚ
so_builderÚlink_cmdÚlinker_flagsÚweight_fileÚ	f_weightsÚgpu_oÚo_filerU  Ú
page_size_Ú	page_sizeÚf_soÚso_sizer2  r/  rï  r  rú  rê  sN    ``` `                                                                  @@@@@@rV   ÚcompilezAotCodeCompiler.compiler  sm  ÿù€ ð +ˆä�<‰<˜7Ò"ÜÐOÓPÐPäÔä%›ˆÜ$ØØÜ-Ø&Ø'ØŸ™ôô
ˆô ˜?×;Ñ;Ó=Ó>ˆô ×ÑÓÒJ ;°%Ñ#7ÒJ¸E¿N¹Nð 	ô +¬6×+>Ñ+>×+JÑ+JÓKñ	
Ø!Ø#ô ×Ñ×/Ò/ØŸ9™9 l°KÐ%@ÓAˆLØˆKä$)ØØØØ/ô	%
Ñ!ˆ�\ô ØØØØ/ô	
‰ˆˆ;ô ×Ñ×&Ò&Ø×"Ñ" <Ô0Ü×&Ñ&×7Ò7Ø×&Ñ& {Ô3ä×ÑÐ:¸LÔIÜ×ÑÐ9¸;ÔGÜØóó
 ,õ	
ô 	Øóó
 +õ	
ô !% \Ó 2ÐÜ# KÓ0ÐØ1×8Ñ8¸;ÑFÐØ ×'Ñ'Ô)Ø×#Ñ#¨TÐ#Ô2ÜœÐ/Ó0Ð3CÑCÓDˆ
÷L	ð L	õ\ 	3ä“>ˆÙÜ�G‰G�L‰L˜ ;°Ñ#8Ó9Ä<ô
ˆð ñ _	6Ù-Ü+.Ø)×5Ñ5°gÓ>ó,Ð(ô Ð2°CÓ8ð <¸AØ—G‘GÐ:Ô;÷<ô ×&Ñ&×.Ò.Ø#×*Ñ*Ð+CÔDä×*Ñ*×3Ñ3ˆHØ*5ˆHÐ&Ñ'ô Ø%×/Ñ/Ø,×1Ñ1Ð2°.ÐAóóˆIô
 ×+Ñ+×4Ñ4×:Ñ:Ó<ò ‘��1Ü! !¤SÔ)¬j¸¼SÕ.Bð Ø8óÐBðô
 �i Ó%ð B¨Ø—‘œŸ
™
¤6×#6Ñ#6×#?Ñ#?Ó@ÔA÷Bô  #Ø$×.Ñ.Ø+×0Ñ0Ð1°Ð@óó Ðô
 �K‰K˜	Ð#3Ô4ä×"Ñ"×*Ò*Ø×&Ñ& yÔ1Ü×*Ñ*×;Ò;Ø#×*Ñ*Ð+;Ô<ñ +ô ×#Ñ#×/Ò/äÐ.×:Ñ:¸5ÓAÓBð ô
 ó à!ŸO™O×0Ñ0Ó2ôó ˆHóOô< ×"Ñ"×:Ò:Ø%(§X¡Xõ .à %§¡× 4Ñ 4Ó 6ô.ó &Ñ"ð &)Ð"äÐ0Ó1ˆKô $*×#3Ñ#3Ó#5Ð5ÒU¸+ÈÑ:UÐÜ×"Ñ"×5Ò5Ø#'Ð ð "ŸN™NØ*Ø$4Ü%+×%5Ñ%5Ó%7Ø)ñ/ˆOô %:ð %Ø!Ü!'×!4Ñ!4×!EÑ!EÐEñ%ð "ñ%Ð!ô
 $9ñ $Ø!ð$à!ñ$Ð ô
 )ÜÐ.×3Ñ3Ó4Ø$ÜÐ4×;Ñ;Ó<Ø1ô	ˆOð #2×"BÑ"BÓ"DÐØ'×<Ñ<Ó>ˆIä'ÜÐ-×2Ñ2Ó3Ø#ÜÐ4×;Ñ;Ó<Ø0ô	ˆNð "0×!@Ñ!@Ó!BÐØ%×:Ñ:Ó<ˆHä�I‰IÐ;Ð=PÔQÜ�I‰IÐ:Ð<NÔOÜ×"Ñ"×3Ò3ä #Ø)×3Ñ3Ø0×5Ñ5Ð6Ð6IÐJóó!�ð
 &×8Ñ8¸ÔGØ×&Ñ& }Ô5Ø×9Ñ9¸*ÔEØ×1Ñ1°*¸lÔKØ×&Ñ& zÕ2à×%Ñ%Ô'Ø×$Ñ$Ô&á#Ø 2�Ø ‘ä#ÜœŸ™ q¬%¯+©+´e·k±kÓ*B×*FÑ*FÈÓM×RÑRÓTó �ô !'§¡¨D°+À±/À<Ó P�á& }´c·l±lÓCˆHØˆNð !ˆOÜ4=Ø×)Ñ)×/Ñ/Ó1ó5ò 8Ñ0�Ñ 0  xô " (¬E¯H©H×,AÑ,AÔBÐBÐBÜ%?Ð$@ÀÐ@PÐ"Q�ä—	‘	Ð9¸4ÀÔQà(7� Ñ%Ü#(§8¡8×#8Ñ#8¸Ó#BÐ Ü"$§'¡'§,¡,Ø)×0Ñ0°/ó#�ô ˜_Ð.>ÀÔEØ×&Ñ& Õ7ð8ô" %'§G¡G§L¡LØ%×,Ñ,Ð.Gó%Ð!ô Ð+¨SÓ1ð 5°QØ—‘œŸ
™
 ?Ó3Ô4÷5à×"Ñ"Ð#8Ô9ô $)§=¡=×#4Ò#4””¼-»/ð ð
 +×0Ñ0×7Ñ7Ó9öàØ×$Ñ$×-Ñ-¨dÔ3ð ×!Ó!ðˆMð ð
  ŸH™H ]Ó3ˆMä&LÈYÓ&WÑ#ˆK˜Ü4Ø&Ø'ØŸ™Ø"3ô	 Ðô $Ø á ð # H¨h¸ÑFà ¨8Ð4Ø%Ø,ôˆJð "×2Ñ2Ó4ˆHØ"×7Ñ7Ó9ˆIä�I‰IÐ/°Ô:ô �l CÓ(ð 9¨AØ—‘˜”Ø—‘Ð-Ð.AÐ-BÀ"ÐEÔFØ—‘Ð*¨8¨*°BÐ7Ô8÷9ô
 �k 3Ó'ð 9¨1Ø—‘˜”Ø—‘Ð-Ð.@Ð-AÀÐDÔEØ—‘Ð*¨8¨*°BÐ7Ô8÷9ô
 ×"Ñ"×3Ó3Ü"Ø)×3Ñ3Ø0×5Ñ5Ð6Ð6HÐIóó �ð
 !×3Ñ3°LÔAØ×&Ñ& |Ô4Ø×&Ñ&¤~Ô6ñ
 $Ü"%Ø-×7Ñ7Ø4×9Ñ9Ð:Ð:QÐRóó#�Kô
 ˜k¨4Ó0ð H°IØ!Ÿ™Ð(:Ô;Ø!Ÿ™¬¯©°C¸Ó(FÔG÷Hð $×*Ñ*¨;Ô7à×&Ñ& xÔ0Ø×&Ñ& }Ô5à×,Ñ,¨Z¸ÔBØ*×0Ñ0Ó2ò D�EØ×0Ñ0°¸UÕCðDà×1Ñ1°*Õ=à× Ñ Ô"à(¨(°HÐ=ò &�Fä—I‘I˜fÕ%ð&ñ $Û#à!)×!5Ñ!5Ó!7�JÜ # E¨:Ó 6�Iä˜i¨Ó/ð C°4Ø"&§)¡)£+˜àŸ
™
 4¨9°wÀÑ7JÑ+JÑ#KÔLØŸ
™
Ð#5Ô6ØŸ
™
¤6§;¡;¨s°LÓ#AÔB÷Cô ×&Ñ&×.Ò.Ø#×*Ñ*¨9Ô5÷_	6ôB	 ×Ñ×&Ò&ð #Ð"àÐ÷E	<ñ <ú÷(Bñ Bú÷F5ñ 5üò÷:9ñ 9ú÷
9ñ 9ú÷0Hñ Hú÷4Cð Cú÷o_	6ð _	6úsÍ   É0)v1ÊuÊ+Cv1Í3v1Î=uÏT,v1ã0%u,äA#v1å8.u9æ&Bv1è<<u>é8v1ê<vëB6v1í>7vî5C#v1òA!v%ó93v1õu	õv1õu)	õ$v1õ,u6	õ1v1õ>v	öv1öv	öv1öv"	öv1ö%v.	ö*v1ö1v:N)rs  rd   rì  r²   rñ  r²   rX  úOptional[str]rß  r²   rY  r­  rµ   zUnion[list[str], str])rº   r»   r¼   rØ  rŽ  rR   rX   rV   rÚ  rÚ  q  sa   „ Øðcàðcð ðcð ð	cð
 )6ðcð ðcð $ðcð 
òcó ñcrX   rÚ  c                 óì   — t        t        «      j                  dz  } | j                  «       5 }|j	                  «       }t        |d«      \  }}d d d «       t        |«      S # 1 sw Y   t        «      S xY w)NrÈ  r¦  )r   rÊ  r0  r¦   rº  r"  r,   )r{   r?  r  rÓ  r]  s        rV   Úcpp_prefix_pathr‘  a  sm   € ô ”‹>× Ñ Ð#9Ñ9€DØ	�‰‹ð 
˜Ø—&‘&“(ˆÜØØó
‰ˆˆ8÷
ô $ HÓ-Ð-÷
ô $ HÓ-Ð-ús   ­ A Á A3c                 ó�   — t        «       } t        j                  «       r#dt        j                  j                  | «      › d�S d| › d�S )Nz
#include "ú")r‘  r   rÓ  rz   r{   r  ©r]  s    rV   Ú
cpp_prefixr•  n  sF   € ÜÓ €HÜ×ÑÔð œBŸG™G×,Ñ,¨XÓ6Ð7°qÐ9Ð9à˜H˜: QÐ'Ð'rX   zOptional[CDLL]Ú_libgompc                óâ  ‡— d	ˆfd„Š|D �cg c]
  } ‰|«      ‘Œ }}| j                  d«      s
J | dz   «       ‚d }t        | j                  d«      «      D ]+  \  }}|dk(  rt        j                  |«      }t        ||«      }Œ- t        |«      s
J | dz   «       ‚t        «       }t        |j                  j                  |«      D ]!  \  }}	|j                  sŒ|	||j                  <   Œ# |r|t        |«       d …=  ||i |¤Ž}
t        |
t        t         f«      r†|
D �cg c]  }|€t#        j$                  g «      n|‘Œ }
}t        |
«      D ])  \  }}t        |t"        j&                  «      rŒ!J | dz   «       ‚ t"        j(                  j*                  j-                  |
«      S t        |
t"        j&                  «      s
J | dz   «       ‚t"        j(                  j*                  j/                  |
«      S c c}w c c}w )
Nc                óè   •— t        t        | «      «      dk(  r)t        j                  j                  j                  | «      S t        | t        t        f«      r t        | «      ˆfd„| D «       «      S | S )Nz<class 'PyCapsule'>c              3  ó.   •K  — | ]  } ‰|«      –— Œ y ­wrQ   rR   )rä   r8  Úconvert_args     €rV   rå   z9custom_op_wrapper.<locals>.convert_arg.<locals>.<genexpr>ƒ  s   øè ø€ Ò9°™[¨Ÿ^Ñ9ùs   ƒ)	r²   r˜  rt   r—  Ú_aotiÚ&alloc_tensor_by_stealing_from_void_ptrr  r�  Útuple)Úargrš  s    €rV   rš  z&custom_op_wrapper.<locals>.convert_arg~  sW   ø€ ÜŒt�C‹y‹>Ð2Ò2ä—8‘8—>‘>×HÑHÈÓMÐMÜ˜œd¤E˜]Ô+Ø”4˜“9Ó9°SÔ9Ó9Ð9àˆJrX   z
torch.ops.z, can not be called through custom_op_wrapperr  r   z, can not be loaded through custom_op_wrapperz returns a list of non-tensorsz returns a non-tensor)rž  r   rµ   r   )Ú
startswithrO  rÊ  Ú	importlibÚimport_moduler’  Úcallabler¡  ÚzipÚ_schemaÚ	argumentsÚ
kwarg_onlyr„   rŸ  r  r�  r�  rt   Útensorr   r—  r›  Ú#unsafe_alloc_void_ptrs_from_tensorsÚ!unsafe_alloc_void_ptr_from_tensor)rß   rT   rž  Úconverted_argsÚfuncrÝ  r~  rU   Úfunc_argÚconv_argÚresultÚrrš  s               @rV   Úcustom_op_wrapperr°  {  sØ  ø€ õð 37Ö7¨3‘k #Õ&Ð7€NÐ7à�=‰=˜Ô&ð Ø
Ð;Ñ;óÐ&ð €DÜ˜"Ÿ(™( 3›-Ó(ò  ‰ˆˆ1Ø�Š6Ü×*Ñ*¨1Ó-ˆDÜ�t˜QÓ‰ð ô
 �DŒ>ÐN˜2Ð NÑNÓNˆ>ô ‹V€FÜ! $§,¡,×"8Ñ"8¸.ÓIò -Ñˆ�(Ø×ÓØ$,ˆF�8—=‘=Ò!ð-ñ ØœC ›K˜<™>Ð*á�>Ð, VÑ,€FÜ�&œ4¤˜-Ô(à@FÖG¸1 a i”%—,‘,˜rÔ"°QÑ6ÐGˆÐGÜ˜fÓ%ò 	V‰DˆAˆqÜ˜a¤§¡Õ.ÐU°Ð5UÑ0UÓUÐ.ð	Vä�x‰x�~‰~×AÑAÀ&ÓIÐIä˜&¤%§,¡,Ô/ÐM°Ð6MÑ1MÓMÐ/Ü�x‰x�~‰~×?Ñ?ÀÓGÐGùò= 8ùò0 Hs   ŒG'Ä G,c                  ó²   — e Zd ZU i Zded<    eej                  «      Zi Zded<   ed
d„«       Z	e
d
d„«       Ze
	 	 	 d	 	 	 	 	 	 	 	 	 dd„«       Ze
ddd	„«       Zy)ÚCppCodeCacheú0dict[str, Callable[[], Union[CDLL, ModuleType]]]r›   r¶   Úcpp_compile_command_flagsc                ó,   — t        j                  | «      S rQ   )r   ÚLoadLibrary)r{   rÅ   s     rV   Ú_load_library_innerz CppCodeCache._load_library_inner®  s   € ä×Ñ Ó%Ð%rX   c           	     ó¸  — 	 | j                  ||«      }||_        |S # t        t        f$ r­}dt	        |«      v rTt
        j                  j                  d«      r5t        j                  d«      a
| j                  ||«      }||_        |cY d }~S dt	        |«      v r9t        |› dt        j                  «       › dt        j                  «       › d�«      |‚‚ d }~ww xY w)NÚgompz/usr/lib64/libgomp.so.1z(failed to map segment from shared objectz3.  The most common reason this may occur is if the zl folder is mounted with noexec (e.g., by default Docker mounts tmp file systems as noexec).  Please remount zi with exec enabled, or set another temporary directory with TORCHINDUCTOR_CACHE_DIR environment variable.)r·  rÅ   ÚImportErrorrm  r²   rz   r{   rø   r   r¶  r–  ÚtempfileÚ
gettempdir)rÐ  r{   rÅ   r®  rò   s        rV   Ú_load_libraryzCppCodeCache._load_library²  sâ   € ð	Ø×,Ñ,¨T°3Ó7ˆFØˆFŒJØˆMøÜœWÐ%ò 	Øœ˜Q›Ñ¤B§G¡G§N¡NÐ3LÔ$Mô  ×+Ñ+Ð,EÓF�Ø×0Ñ0°°sÓ;�Ø �”
Ø•Ø9¼SÀ»VÑCÜØ�cÐLÌX×M`ÑM`ÓMbÐLcð d3ä3;×3FÑ3FÓ3HÐ2Ið J]ð]óð
 ðð ûð	ús"   ‚ �C¬ACÂCÂACÃCNc           	     ó@  ‡ ‡‡‡‡‡— i ‰ j                   ¥|t        «       |dœ¥}t        «        t        ddt	        di |¤Ž¬«      }t        |j                  «       «      }t        |d|¬«      \  Š}‰‰ j                  v�rddl	m
}	 t        j                  j                  t        «       ‰d	z   «      }
t        |«      \  }}d Šd Št	        di |¤d
t!        j"                  «       xr |dk(  i¤Ž}t        ||||¬«      }t%        j&                  t(        |
|«      Št+        |j-                  «       «      Šdˆˆ ˆˆˆˆfd„}|�> |	|
t.        ¬«      5  t        j                  j1                  ‰«      s |‰«      Šd d d «       |‰ j                  ‰<   ‰ j                  ‰   S # 1 sw Y   Œ'xY w)N)rß  rÞ  Úextra_flagsrÜ  rÝ  rá  rè  ©r
  r   r  r  rú  rä  rû  c                 ór   •— ‰€3‰�‰j                  «         ‰«       } | �J ‚‰j                  ‰‰«      Š‰€J ‚‰S rQ   )r®  r½  )r®  Úbinary_pathrÐ  ÚfuturerÅ   rÁ  Ú	worker_fns    €€€€€€rV   Úload_fnz(CppCodeCache.load_async.<locals>.load_fnþ  sI   ø€ à�;ØÐ)ØŸ™œÙ&›[�FØ!˜>Ð)˜>Ø×+Ñ+¨K¸Ó=�CØ˜?Ð*˜?Ø�
rX   r  rR   ©rµ   r   )r´  r-   r$   r'   r)   r  r<  r"  r›   r?  r  rz   r{   r|   rû   r+   r   rÓ  r¾   r
   Ú_worker_compile_cppr,   r   r@  rø   )rÐ  rl  rß  Ú	submit_fnr¿  rk  Úcommand_genÚvec_isa_cmdÚ
input_pathr  Ú	lock_pathr�  rü  Úcpp_build_optionÚcpp_builderrÅ  rÂ  rÃ  rÅ   rÁ  rÄ  s   `               @@@@@rV   Ú
load_asynczCppCodeCache.load_asyncÉ  s‘  ý€ ð
Ø×+Ñ+ð
à&Ü#“~Ø&ò	
ˆô 	Ôä Ø˜cÔ/DÑ/WÀÑ/Wô
ˆô ˜;×7Ñ7Ó9Ó:ˆÜ ¨U¸+ÔF‰ˆˆZà�c—i‘iÒÝ6äŸ™Ÿ™¤\£^°S¸7±]ÓCˆIÜ&LÈZÓ&XÑ#ˆK˜Ø,0ˆFØˆCä4ñ  Ø!ñ ä#)×#3Ñ#3Ó#5Ò#N¸+ÈÑ:Nò Ðô %Ø Ø"Ø%Ø,ô	ˆKô "×)Ñ)Ü#ØØóˆIô
 3°;×3SÑ3SÓ3UÓVˆK÷	ò 	ð Ð$Ù˜i´Ô>ñ 6ÜŸ7™7Ÿ>™>¨+Ô6Ù!*¨9Ó!5˜÷6ð %ˆC�I‰I�c‰Nà�y‰y˜‰~Ð÷6ð 6ús   Å(FÆFc                ó0   —  | j                  ||«      «       S rQ   )rÏ  )rÐ  rl  rß  s      rV   r§   zCppCodeCache.load  s   € à7ˆs�~‰~˜k¨;Ó7Ó9Ð9rX   )r{   r²   rÅ   r²   rµ   zUnion[CDLL, ModuleType])rä  NrR   )
rl  r²   rß  r²   rÈ  r   r¿  úSequence[str]rµ   r   )rä  )rl  r²   rß  r²   rµ   r   )rº   r»   r¼   r›   rG  r½   rÃ  r×  r´  r·  rØ  r½  rÏ  r§   rR   rX   rV   r²  r²  ¨  s·   … à>@€EÐ;Ó@Ù˜uŸ{™{Ó+€KØ02Ð˜~Ó2àò&ó ð&ð òó ðð, ð !ØØ%'ðFàðFð ðFð ð	Fð
 #ðFð 
òFó ðFðP ó:ó ñ:rX   r²  c                óÐ   — ddl m}  || t        ¬«      5  t        j                  j                  |j                  «       «      s|j                  «        d d d «       y # 1 sw Y   y xY w)Nr   r  r  )r?  r  r@  rz   r{   rø   r   r  )rÌ  rÎ  r  s      rV   rÇ  rÇ    sK   € õ /á	�)¤\Ô	2ñ  Ü�w‰w�~‰~˜k×>Ñ>Ó@ÔAØ×ÑÔ÷ ÷  ñ  ús   •>AÁA%c                  óÔ   — e Zd ZU i Zded<    eej                  «      ZdddœZdZ	dZ
dZ ej                  d	«      Zedd
„«       Ze	 	 	 	 d	 	 	 	 	 	 	 	 	 	 	 	 	 dd„«       Zedd„«       Zy)ÚCppPythonBindingsCodeCacher³  r›   FT©Úinclude_pytorchÚsharedr  zkernel({}); Py_RETURN_NONE;r‚  aR  
        // Python bindings to call {entry_func}():
        #define PY_SSIZE_T_CLEAN
        #include <Python.h>
        #include <sstream>
        #include <cstdlib>

        #ifndef _MSC_VER
        #if __cplusplus < 202002L
        // C++20 (earlier) code
        // https://en.cppreference.com/w/cpp/language/attributes/likely
        #define likely(x)       __builtin_expect(!!(x), 1)
        #define unlikely(x)     __builtin_expect(!!(x), 0)
        #endif
        #else
        #define likely(x) (x)
        #define unlikely(x) (x)
        #endif

        // This is defined in guards.cpp so we don't need to import PyTorch headers that are slooow.
        // We manually link it below to workaround issues with fbcode build.
        static void* (*_torchinductor_pyobject_tensor_data_ptr)(PyObject* obj);

        template <typename T> static inline T parse_arg(PyObject* args, size_t n) {{
            static_assert(std::is_pointer_v<T>, "arg type must be pointer or long");
            return static_cast<T>(_torchinductor_pyobject_tensor_data_ptr(PyTuple_GET_ITEM(args, n)));
        }}
        template <> inline int64_t parse_arg<int64_t>(PyObject* args, size_t n) {{
            auto result = PyLong_AsSsize_t(PyTuple_GET_ITEM(args, n));
            if(unlikely(result == -1 && PyErr_Occurred()))
                throw std::runtime_error("expected int arg");
            return result;
        }}
        template <> inline uintptr_t parse_arg<uintptr_t>(PyObject* args, size_t n) {{
            auto result = PyLong_AsVoidPtr(PyTuple_GET_ITEM(args, n));
            if(unlikely(result == reinterpret_cast<void*>(-1) && PyErr_Occurred()))
                throw std::runtime_error("expected int arg");
            return reinterpret_cast<uintptr_t>(result);
        }}

        {extra_parse_arg}

        static PyObject* {entry_func}_py(PyObject* self, PyObject* args) {{
            try {{
                if(unlikely(!PyTuple_CheckExact(args)))
                    throw std::runtime_error("tuple args required");
                if(unlikely(PyTuple_GET_SIZE(args) != {arg_len}))
                    throw std::runtime_error("requires {arg_len} args");
                {call_entry_func}
            }} catch(std::exception const& e) {{
                PyErr_SetString(PyExc_RuntimeError, e.what());
                return nullptr;
            }} catch(...) {{
                PyErr_SetString(PyExc_RuntimeError, "unhandled error");
                return nullptr;
            }}
        }}

        static PyMethodDef py_methods[] = {{
            {{"{entry_func}", {entry_func}_py, METH_VARARGS, ""}},
            {{NULL, NULL, 0, NULL}}}};

        static struct PyModuleDef py_module =
            {{PyModuleDef_HEAD_INIT, "{entry_func}", NULL, -1, py_methods}};

        PyMODINIT_FUNC PyInit_{entry_func}(void) {{
            const char* str_addr = std::getenv("_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTR");
            if(!str_addr) {{
                PyErr_SetString(PyExc_RuntimeError, "_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTR must be set");
                return nullptr;
            }}
            std::istringstream iss(str_addr);
            uintptr_t addr = 0;
            iss >> addr;
            _torchinductor_pyobject_tensor_data_ptr =
                reinterpret_cast<decltype(_torchinductor_pyobject_tensor_data_ptr)>(addr);
            PyObject* module = PyModule_Create(&py_module);
            if (module == NULL) {{
                return NULL;
            }}
            #ifdef Py_GIL_DISABLED
                PyUnstable_Module_SetGIL(mod, Py_MOD_GIL_NOT_USED);
            #endif
            return module;
        }}
        c                óà  — t        t        j                  j                  j                  j
                  «      t        j                  d<   |› d| j                  › �}	 t        j                  |   S # t        $ r Y nw xY wt        j                  j                  ||«      }|€J ‚t        j                  j                  |«      }|t        j                  |<   |j                   j#                  |«       |S )NÚ'_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTRr  )r²   rt   r—  Ú_dynamorp  Ú'_torchinductor_pyobject_tensor_data_ptrrz   ÚenvironÚentry_functionr;  rö  ÚKeyErrorr   ÚutilÚspec_from_file_locationÚmodule_from_specÚloaderÚexec_module)rÐ  r{   rÅ   Úmodule_namerÂ  rÃ  s         rV   r·  z.CppPythonBindingsCodeCache._load_library_innerˆ  sÊ   € ä@CÜ�H‰H×Ñ×#Ñ#×KÑKóA
Œ�
‰
Ð<Ñ=ð ˜˜Q˜s×1Ñ1Ð2Ð3ˆð	Ü—;‘;˜{Ñ+Ð+øÜò 	Ùð	úä�~‰~×5Ñ5°kÀ4ÓHˆØÐÐÐÜ—‘×0Ñ0°Ó6ˆØ#)Œ�‰�KÑ Ø�‰×Ñ Ô'Øˆs   ÁA. Á.	A:Á9A:Nc                ó\  ‡ ‡
‡— dj                  d„ t        |«      D «       «      }‰ j                  j                  t	        |«      ‰ j
                  j                  |«      ‰ j                  ‰ j                  j                  |¬«      ¬«      }‰ j                  ||z   |||¬«      Š
dŠdˆ ˆ
ˆfd„}	|	S )	a5  
        Wrap a C++ function in fast Python bindings.

        Args:
            argtypes: The types of args to ENTRY_FUNCTION(), e.g. ["float*", "long"]
            source_code: C++ source code containing a ENTRY_FUNCTION() function

        Returns:
            A python version of ENTRY_FUNCTION()
        ú, c              3  óT   K  — | ]   \  }}d |j                  dd«      › d|› d�–— Œ" y­w)z
parse_arg<zconst r‚  z>(args, ú)N)rR  )rä   ÚnÚargtypes      rV   rå   zBCppPythonBindingsCodeCache.load_pybinding_async.<locals>.<genexpr>­  s7   è ø€ ò 
á��7ð ˜Ÿ™¨°2Ó6Ð7°xÀ¸sÀ!ÔDñ
ùs   ‚&()Ú	array_len)Úarg_lenÚcall_entry_funcÚ
entry_funcÚextra_parse_arg)rÈ  r¿  Nc                 óf   •— ‰€ ‰«       Št        ‰t        «      sJ ‚t        ‰‰ j                  «      S rQ   )r  r   r’  rÝ  )rÐ  Ú
get_resultr®  s   €€€rV   rÃ  z?CppPythonBindingsCodeCache.load_pybinding_async.<locals>.future¿  s2   ø€ àˆ~Ù#›�Ü! &¬*Ô5Ð5Ð5Ü˜6 3×#5Ñ#5Ó6Ð6rX   rÆ  )	r|   rO  Úsuffix_templateÚformatrŸ  Úcall_entry_functionrÝ  rï  rÏ  )rÐ  Úargtypesrl  rß  Únum_outputsrÈ  r¿  Ú	parseargsÚsuffixrÃ  rñ  r®  s   `         @@rV   Úload_pybinding_asyncz/CppPythonBindingsCodeCache.load_pybinding_async™  s¶   ú€ ð( —I‘Iñ 
ä'¨Ó1ô
ó 
ˆ	ð ×$Ñ$×+Ñ+Ü˜“MØ×3Ñ3×:Ñ:¸9ÓEØ×)Ñ)Ø×/Ñ/×6Ñ6ÀÐ6ÓMð	 ,ó 
ˆð —^‘^Ø˜&Ñ ØØØ#ð	 $ó 
ˆ
ð ˆ÷	7ð ˆrX   c                ó0   —   | j                   |i |¤Ž«       S rQ   )rù  ©rÐ  rT   rU   s      rV   Úload_pybindingz)CppPythonBindingsCodeCache.load_pybindingÈ  s    € à8Ð'ˆs×'Ñ'¨Ð8°Ñ8Ó:Ð:rX   )r{   r²   rÅ   r²   rµ   r   )rä  rÈ   NrR   )rõ  r­  rl  r²   rß  r²   rö  r.  rÈ  r   r¿  rÑ  rµ   r   )rT   r   rU   r   rµ   r   )rº   r»   r¼   r›   rG  r½   rÃ  r×  r´  rÝ  rô  rï  ÚtextwrapÚdedentrò  rØ  r·  rù  rü  rR   rX   rV   rÔ  rÔ  #  sä   … à>@€EÐ;Ó@Ù˜uŸ{™{Ó+€Kð !Øñ!Ðð
 €NØ7ÐØ€OØ%�h—o‘oðU	óW€Oðr òó ðð  ð
 !ØØØ%'ð,àð,ð ð,ð ð	,ð
 ð,ð ð,ð #ð,ð 
ò,ó ð,ð\ ò;ó ñ;rX   rÔ  c                  óv   — e Zd ZU i Zded<    eej                  «      ZdddœZdZ	dZ
 ej                  d«      Zy)	ÚCppWrapperCodeCacher³  r›   TrÕ  Úinductor_entry_cppzreturn inductor_entry_cpp({});a	  
        #include <torch/csrc/inductor/aoti_torch/c/shim.h>

        static inline std::vector<AtenTensorHandle> unpack_tensor_handle_list(PyObject* pyvec) {{
            std::vector<AtenTensorHandle> result;
            size_t result_len = PyList_GET_SIZE(pyvec);
            result.reserve(result_len);
            for (size_t i = 0; i < result_len; i++) {{
                // AtenTensorHandle is essentially a pointer
                void* elem = PyCapsule_GetPointer(PyList_GET_ITEM(pyvec, i), NULL);
                result.push_back(reinterpret_cast<AtenTensorHandle>(elem));
            }}
            return result;
        }}

        static inline PyObject* pack_tensor_handle_list(const std::array<AtenTensorHandle, {array_len}>& arr) {{
            PyObject* result = PyList_New({array_len});
            for (size_t i = 0; i < {array_len}; i++) {{
                PyObject *elem =
                    arr[i] == nullptr
                        ? Py_None
                        // Store AtenTensorHandle as PyCapsulate
                        : PyCapsule_New(reinterpret_cast<void*>(arr[i]), NULL, NULL);
                PyList_SET_ITEM(result, i, elem);
            }}
            return result;
        }}

        template <> inline std::vector<AtenTensorHandle> parse_arg<std::vector<AtenTensorHandle>>(PyObject* args, size_t n) {{
            return unpack_tensor_handle_list(PyTuple_GET_ITEM(args, n));
        }}

        PyObject* inductor_entry_cpp(std::vector<AtenTensorHandle>&& input_handles) {{
            // For outputs, we only allocate an array to hold returned tensor handles,
            // not the actual output tensor storage.
            std::array<AtenTensorHandle, {array_len}> output_handles{{}};
            try {{
                inductor_entry_impl(input_handles.data(), output_handles.data());
                if (PyErr_Occurred()) {{
                    return nullptr;
                }}
                return pack_tensor_handle_list(output_handles);
            }} catch(std::exception const& e) {{
                PyErr_SetString(PyExc_RuntimeError, e.what());
                return nullptr;
            }} catch(...) {{
                PyErr_SetString(PyExc_RuntimeError, "unhandled error");
                return nullptr;
            }}
        }}
        N)rº   r»   r¼   r›   rG  r½   rÃ  r×  r´  rÝ  rô  rý  rþ  rï  rR   rX   rV   r   r   Í  sK   … à>@€EÐ;Ó@Ù˜uŸ{™{Ó+€KàØñ!Ðð *€NØ:ÐØ%�h—o‘oð2	ó4�OrX   r   c                  ó  — e Zd ZU i Zded<    eej                  «      ZdZded<    e	j                  d«      Ze e	j                  d«      z   Ze e	j                  d«      z   Z e	j                  d	«      Zedd
„«       Zedd„«       Ze ej&                  d«      dd„«       «       Zedd„«       Ze ej&                  d«      dd„«       «       Ze ej&                  d«      dd„«       «       Ze	 d	 	 	 	 	 	 	 dd„«       Zedd„«       Zedd„«       Zy)ÚHalideCodeCachez0dict[str, Callable[[], Union[ModuleType, CDLL]]]r›   Nr�  Ú_standalone_runtime_pathaØ  
        #include "{halideruntime_h}"
        #include "{headerfile}"
        #include <stdexcept>
        #include <cmath>

        namespace c10 {{
            inline long div_floor_integer(long a, long b) {{
                if ((a<0) != (b<0)) {{
                    const auto quot = a / b;
                    const auto rem = a % b;
                    return rem ? quot - 1 : quot;
                }}
                return a / b;
            }}
        }}
        zÌ
        void kernel({argdefs}) {{
            {buffers}
            int err = halide_kernel({buffer_names});
            if(err != 0) throw std::runtime_error("halide_kernel failed");
        }}
        a{  
        #include <cuda.h>
        static const halide_device_interface_t* cuda_interface = halide_cuda_device_interface();

        void kernel({argdefs}, uintptr_t stream) {{
            {buffers}
            int err = halide_kernel(reinterpret_cast<void*>(stream), {buffer_names});
            if(err != 0) throw std::runtime_error("halide_kernel failed");
        }}
        aÎ  
        #include "{}"
        #include <cuda.h>

        static int acquire_context(void* user_context,
                                   void** cuda_context_out,
                                   bool create) {{
            return cuCtxGetCurrent(reinterpret_cast<CUcontext*>(cuda_context_out));
        }}

        static int release_context(void* user_context) {{
            return 0;
        }}

        static int get_stream(void* user_context,
                              void* cuda_context,
                              void** stream_out) {{
            *stream_out = user_context;
            return 0;
        }}

        static int register_halide_hooks() {{
            halide_set_cuda_acquire_context(&acquire_context);
            halide_set_cuda_release_context(&release_context);
            halide_set_cuda_get_stream(&get_stream);
            return 0;
        }}

        int inductor_register_halide_hooks_result = register_halide_hooks();
        c                óŒ  — |j                   €J ‚|j                  �+t        |j                   «      t        |j                  «      k(  sJ ‚|j                  €J ‚|j                  xs |j
                  › d|j                  › �}|rd|› d�}d}d}d}nd}d}d|› d�}d	}g }	t        |j                   |j                  «      D ]  \  }
}|	j                  d
|
› d|› d�«       Œ d|› d�d|› ddj                  |	«      › d�|› d|› d�|› d|› d�|› d|› d�|› d|› d�|› d|j                  «       › d�|› dt        |	«      › d�|› d|› d�|› d�g
S )Nz + zreinterpret_cast<uint64_t>(rè  Úcuda_interfaceÚnullptrÚhalide_buffer_flag_device_dirtyÚ0zreinterpret_cast<uint8_t*>(Úhalide_buffer_flag_host_dirtyzhalide_dimension_t(0, ræ  zhalide_buffer_t ú;zhalide_dimension_t z_dims[] = {z};z
.device = z.device_interface = z.host = z	.flags = z.type = z.dimensions = z.dim = z_dims;z.padding = nullptr;)
ÚshapeÚstriderŸ  ÚoffsetÚalias_ofr„   r£  r   r|   Úhalide_type)rÐ  r„   rž  r‡   r(  r†   Údevice_interfaceÚhostÚflagsÚdimsÚsizer  s               rV   Ú_codegen_bufferzHalideCodeCache._codegen_buffer\	  s¬  € à�y‰yÐ$Ð$Ð$Ø�z‰zÐ%¬#¨c¯i©i«.¼CÀÇ
Á
»OÒ*KÐKÐKØ�z‰zÐ%Ð%Ð%Ø—l‘lÒ. c§h¡hÐ/¨s°3·:±:°,Ð?ˆÙØ2°8°*¸AÐ>ˆFØ/ÐØˆDØ5‰EàˆFØ(ÐØ0°°
¸!Ð<ˆDØ3ˆEàˆÜ §	¡	¨3¯:©:Ó6ò 	D‰LˆD�&Ø�K‰KÐ0°°°b¸¸ÀÐBÕCð	Dð ˜t˜f AÐ&Ø! $  |°D·I±I¸d³OÐ3DÀCÐHØˆf�J˜v˜h aÐ(ØˆfÐ(Ð)9Ð(:¸!Ð<Øˆf�H˜T˜F !Ð$Øˆf�I˜e˜W AÐ&Øˆf�H˜SŸ_™_Ó.Ð/¨qÐ1Øˆf�N¤3 t£9 +¨QÐ/Øˆf�G˜D˜6 Ð(ØˆfÐ'Ð(ð
ð 	
rX   c           	     óä  — |j                  «       }|d|j                  v u sJ ‚d|j                  v sJ ‚g }g }t        |j                  «      D ]z  \  }}|j	                  «       r:|j                  d|› �«       |j                  | j                  d|› �||«      «       ŒPd|j                  vsJ ‚|j                  |j                  «       Œ| dj                  |D �cg c]  }d|› �‘Œ	 c}«      j                  «       }|r| j                  n| j                  }	|	j                  | j                  |rdnd	«      |d
j                  d„ |j                  D «       «      |d
j                  |«      ¬«      }
|
S c c}w )NÚuser_contextÚ
no_runtimez&hl_buf_Úhl_buf_Ú*r#  ú    úHalideRuntimeCuda.hzHalideRuntime.hræ  c              3  ór   K  — | ]/  }|j                   €!|j                  «       › d|j                  › �–— Œ1 y ­w)Nr7  )r  Úbindings_typer„   )rä   r8  s     rV   rå   z0HalideCodeCache._codegen_glue.<locals>.<genexpr>”	  s9   è ø€ ò àØ—:‘:Ð%ð —?‘?Ó$Ð% Q q§v¡v hÔ/ñùs   ‚57)Úhalideruntime_hÚ
headerfileÚargdefsÚbuffersÚbuffer_names)r  r-  rO  rõ  Ú	is_bufferr   Úextendr  Úctyper„   r|   ÚlstripÚglue_template_cudaÚglue_template_cppró  Úfind_header)rÐ  rT  r!  r  r#  r$  rÝ  rž  ÚlineÚglue_templateÚ	glue_codes              rV   Ú_codegen_gluezHalideCodeCache._codegen_glue~	  sh  € à—,‘,“.ˆØ˜>¨T¯[©[Ð8Ñ9Ð9Ð9Ø˜tŸ{™{Ñ*Ð*Ð*ØˆØˆÜ §¡Ó.ò 	.‰FˆAˆsØ�}‰}ŒØ×#Ñ# h¨q¨c NÔ3Ø—‘˜s×2Ñ2°W¸Q¸C°=À#ÀwÓOÕPà #§)¡)Ñ+Ð+Ð+Ø×#Ñ# C§H¡HÕ-ð	.ð —)‘)°wÖ?¨t˜t D 6š]Ò?Ó@×GÑGÓIˆá29˜×.Ò.¸s×?TÑ?TˆØ!×(Ñ(ØŸO™OÙ)0Ñ%Ð6Góð "Ø—I‘Iñ àŸ™ôó ð
 ØŸ™ <Ó0ð )ó 
ˆ	ð Ðùò! @s   ÃE-c                óä   — t        ddt        «       ¬«      }|j                  «       }t        dj	                  | j
                  | j                  | j                  |g«      j                  d«      «      S )NÚOÚIrá  r#  r‰   )	r'   r(   r<  r  r|   r*  r)  Ústandalone_runtime_cuda_initr•   )rÐ  rÉ  Úcommand_lines      rV   Úconfig_hashzHalideCodeCache.config_hashž	  sp   € ô !ØØÜ"›ô
ˆð
 #×3Ñ3Ó5ˆÜØ�I‰Ià×)Ñ)Ø×*Ñ*Ø×4Ñ4Ø ð	ó÷ ‰f�W‹oó	
ð 		
rX   c                óX  — t         j                  j                  j                  d«      }|�|j                  st        d«      ‚	 |j                  d   }t        j                  |«      D ]þ  }|j                  d«      sŒ	 t        j                  dt        j                  j                  ||«      g«      }t        j                  d|j!                  d«      «      }|sŒst        j                  j                  t        j                  j#                  |j%                  d«      «      | «      }t        j                  j'                  |«      sŒßt        j                  j#                  |«      c S  	 t        |«      ‚# t        j                  $ r Y �Œ#w xY w# t(        $ r}t        |«      |‚d }~ww xY w)	NÚhalidez$halide python bindings not installedr   rÈ  Úlddz(/.*)/libHalide.sor‰   rE   )r   Ú	machineryÚ
PathFinderr¸  r½  r�   rz   rL  rÉ  Ú
subprocessÚcheck_outputr{   r|   ÚSubprocessErrorr„  Úsearchr  ÚabspathÚgrouprø   rO  )	rø  ÚerrmsgrÂ  r>  ÚfileÚoutÚmr{   rò   s	            rV   Ú_search_for_filez HalideCodeCache._search_for_file²	  sL  € ä×"Ñ"×-Ñ-×7Ñ7¸ÓAˆØˆ<˜t×>Ò>ÜÐEÓFÐFð	.Ø×4Ñ4°QÑ7ˆFÜŸ
™
 6Ó*ò 9�Ø—=‘= Õ'ð!Ü(×5Ñ5Ø"¤B§G¡G§L¡L°¸Ó$>Ð?ó˜ô
 Ÿ	™	Ð"7¸¿¹ÀGÓ9LÓM�AÚÜ!Ÿw™wŸ|™|¬B¯G©G¯O©O¸A¿G¹GÀA»JÓ,GÈÓP˜ÜŸ7™7Ÿ>™>¨$Õ/Ü#%§7¡7§?¡?°4Ó#8Ò8ñ9ô ˜6Ó"Ð"øô &×5Ñ5ò !Ú ð!ûô ò 	.Ü˜vÓ&¨AÐ-ûð	.úsO   Á8F Á>5E5Â3'F ÃA+F Å F Å(F Å5FÆF ÆFÆF Æ	F)ÆF$Æ$F)c                ó*  — d| j                  «       › d�}dt        j                  v rRt        j                  j	                  t        j                  d   |«      }t        j                  j                  |«      r|S d|› d�}t        j                  ||«      S )NÚlibautoschedule_rÈ  Ú
HALIDE_LIBúCan't find z3, set env HALIDE_LIB to the directory containing it)r  rz   rÜ  r{   r|   rø   r  rE  )r„   Úsofiler{   rA  s       rV   Úfind_libautoschedulez$HalideCodeCache.find_libautoscheduleÊ	  s}   € ð $ D§J¡J£L >°Ð5ˆØœ2Ÿ:™:Ñ%Ü—7‘7—<‘<¤§
¡
¨<Ñ 8¸&ÓAˆDÜ�w‰w�~‰~˜dÔ#Ø�à˜&˜Ð!TÐUð 	ô ×/Ñ/°¸Ó?Ð?rX   c                ó  — dt         j                  v rRt         j                  j                  t         j                  d   | «      }t         j                  j	                  |«      r|S dt         j                  v rrt         j                  j                  t         j                  j                  t         j                  d   d| › �«      «      }t         j                  j	                  |«      r|S d| › d�}t        j                  d| › �|«      S )NÚHALIDE_INCLUDErH  z../include/rI  z7, set env HALIDE_INCLUDE to the directory containing it)rz   rÜ  r{   r|   rø   r?  r  rE  )r„   r{   rA  s      rV   r+  zHalideCodeCache.find_header×	  sÈ   € ð œrŸz™zÑ)Ü—7‘7—<‘<¤§
¡
Ð+;Ñ <¸dÓCˆDÜ�w‰w�~‰~˜dÔ#Ø�Øœ2Ÿ:™:Ñ%Ü—7‘7—?‘?Ü—‘—‘œRŸZ™Z¨Ñ5¸ÀTÀFÐ7KÓLóˆDô �w‰w�~‰~˜dÔ#Ø�à˜$˜ÐVÐWð 	ô ×/Ñ/°+¸d¸VÐ0DÀfÓMÐMrX   c                ó  ‡‡— t        t        t        |t        | j	                  «       |f«      ¬«      d«      d   «      }t        j                  |d¬«       d Št        |dz  «      }t        |dz  «      }t        |dz  «      }t        |d	z  «      }t        |d
z  «      }	t
        j                  j                  |«       }
g }|
r±t        ||«       t        j                  |ddd|› ddddg
}|j                  r,|j                  d| j                  |j                  «      g«       |j                  |j!                  «       «       |j#                  t%        j&                  t(        j*                  |«      «       |j,                  D �cg c]  }|j.                  �Œ|j1                  «       ‘Œ! }}|j3                  «       r|j#                  d«       | j5                  || j7                  ||«      || j9                  «       f|
r|j"                  nd |j3                  «       rdnd¬«      Š|
r`|j#                  t%        j&                  t:        |«      «       t%        j&                  t<        |	|«      }|r ||«      j>                  Šn |«        dˆˆfd„}|S c c}w )NrÀ  r7  é   Trö   zgenerate_kernel.pyzhalide_kernel.azhalide_kernel.hÚdonerd  ú-gr  ú-oz-fÚhalide_kernelz-ezstatic_library,h,schedulez-pÚ	uintptr_tr‡   rä  )r¿  rÈ  rß  c                 ó$   •— ‰r ‰«         ‰ «       S rQ   rR   )Úbindings_futureÚwait_for_compiles   €€rV   r§   z3HalideCodeCache.generate_halide_async.<locals>.load'
  s   ø€ ÙÙ Ô"Ù"Ó$Ð$rX   )rµ   úCallable[[], Any]) r   r  r  r  r5  rz   rù   r²   r{   rø   r±   r;  Ú
executableÚ	schedulerr&  rK  rT   r   r¾   r
   r;  Ú
check_callrõ  r  r  r  rù  r/  Úbuild_standalone_runtimeÚtouchÚ_worker_task_halider®  )rÐ  rT  rl  rÈ  ÚdirpathÚgenfileÚlibfiler!  ÚdonefileÚlockfileÚneed_compileÚjobsÚcmdrž  Úbinding_typesÚtaskr§   rV  rW  s                    @@rV   Úgenerate_halide_asyncz%HalideCodeCache.generate_halide_asyncé	  sM  ù€ ô ÜÜØÜ §¡Ó 1°4Ð8Ó9ôð óð ñó
ˆô 	�‰�G dÕ+ØÐÜ�gÐ 4Ñ4Ó5ˆÜ�gÐ 1Ñ1Ó2ˆÜ˜Ð#4Ñ4Ó5ˆ
Ü�w Ñ'Ó(ˆÜ�w Ñ'Ó(ˆÜŸ7™7Ÿ>™>¨(Ó3Ð3ˆØˆÙÜ˜ +Ô.ä—‘ØØØØØ�)ØØØØ+ðˆCð �~Š~Ø—
‘
˜D #×":Ñ":¸4¿>¹>Ó"JÐKÔLØ�J‰J�t—y‘y“{Ô#Ø�K‰Kœ	×)Ñ)¬*×*?Ñ*?ÀÓEÔFð ,0¯=©=ö
Ø$'¸C¿L¹LÑ<PˆC×ÑÕð
ˆð 
ð �<‰<Œ>Ø× Ñ  Ô-Ø×2Ñ2ØØ×Ñ˜d JÓ/Ø  #×">Ñ">Ó"@ÐAÙ%1�d—k’k°tØ"&§,¡,¤.™°eð 3ó 
ˆñ Ø�K‰Kœ	×)Ñ)¬%°Ó:Ô;Ü×$Ñ$Ô%8¸(ÀDÓIˆDÙÙ#,¨T£?×#9Ñ#9Ñ á”ö	%ð
 ˆùò5
s   ÆJ	ÆJ	c                ó0   —   | j                   |i |¤Ž«       S rQ   )ri  rû  s      rV   Úgenerate_halidezHalideCodeCache.generate_halide.
  s    € à9Ð(ˆs×(Ñ(¨$Ð9°&Ñ9Ó;Ð;rX   c           	     ór  — | j                   r5t        j                  j                  | j                   «      r| j                   S t        j
                  j                  «       rdnd}d}|dk(  rdnd}| j                   r6t        j                  j                  | j                   «      rJ ‚t        «       }n
t        «       }t        |«      d|› d| j                  «       › �z  }t        j                  |d¬	«       t        |d
z  «      }t        |dz  «      }t        |dz  «      }t        |dz  «      }	t        ||z  «      }
t        j                  j                  |«      �sdd l}ddlm}  ||t         «      5  t        j                  j                  |«      sÞt#        |d«      5 }|dk(  r9|j%                  | j&                  j)                  | j+                  d«      «      «       d d d «       |j-                  |	|j/                  |«      «       t1        |
«      \  }}t3        |||	g|t5        |¬«      ¬«      }t7        j8                  t;        j<                  |j?                  «       «      «       tA        |«       d d d «       t        j                  j                  |
«      sJ ‚|
| _         |
S # 1 sw Y   ŒÆxY w# 1 sw Y   Œ?xY w)Nr‡   rä  zlibStandaloneHalideRuntime.soz	host-cudar  zhalide-runtime-ú-Trö   rP  rd  z	hooks.cppzstandalone_halide_runtime.ar   r  r*  r  ©rß  rû  )!r  rz   r{   rø   rt   r‡   r  r5   r4   r   r5  rù   r²   r7  r?  r  r@  r¦   r"  r3  ró  r+  Úcompile_standalone_runtimeÚTargetr+   r'   r)   r;  r[  ÚshlexrÊ  r<  r]  )rÐ  rß  Úlibnamer-  Úbaser_  rb  rc  ÚhookfileÚafilerJ  Úhlr  r?  r„   rü  Úhalide_cmd_gens                    rV   r\  z(HalideCodeCache.build_standalone_runtime2
  sH  € à×'Ò'¬B¯G©G¯N©NØ×(Ñ(ô-
ð ×/Ñ/Ð/Ü %§
¡
× 7Ñ 7Ô 9‘f¸uˆØ1ˆØ +¨vÒ 5‘¸6ˆØ×'Ò'Ü—w‘w—~‘~ c×&BÑ&BÔCÐCÐCô
 %Ó&‰Dä“;ˆDÜ�t“* °°¸¸#¿/¹/Ó:KÐ9LÐMÑMˆÜ
�‰�G dÕ+Ü�w Ñ'Ó(ˆÜ�w Ñ'Ó(ˆÜ�w Ñ,Ó-ˆÜ�GÐ;Ñ;Ó<ˆÜ�W˜wÑ&Ó'ˆÜ�w‰w�~‰~˜hÕ'Ûå6á˜(¤LÓ1ñ $Ü—w‘w—~‘~ hÔ/Ü˜h¨Ó,ð °Ø&¨&Ò0ØŸG™GØ #× @Ñ @× GÑ GØ$'§O¡OÐ4IÓ$Jó!"ô÷ð ×1Ñ1°%¸¿¹À6Ó9JÔKä'MÈfÓ'UÑ$�D˜*Ü%/Ø!Ø!)¨5Ð 1Ø#-Ü$9Ø(3ô%ô	&�Nô ×)Ñ)ÜŸ™ N×$CÑ$CÓ$EÓFôô ˜(”O÷1$ô2 �w‰w�~‰~˜fÔ%Ð%Ð%Ø'-ˆÔ$Øˆ÷3ð ú÷$ð $ús%   Å1,J-Æ?J!ÇBJ-Ê!J*	Ê&J-Ê-J6)r„   r²   rž  rj   r‡   ræ   rµ   r­  )rT  rk   r!  Úobjectrµ   r²   rÄ  )rø  r²   rA  r²   rµ   r²   )r„   r²   rµ   r²   rQ   )rT  rk   rl  r²   rÈ  r   rµ   rX  )rT   r   rU   r   rµ   rX  )rº   r»   r¼   r›   rG  r½   rÃ  r×  r  rý  rþ  r¿  r*  r)  r3  rØ  r  r/  r¾   r¿   r5  rE  rK  r+  ri  rk  r\  rR   rX   rV   r  r  	  s´  … à>@€EÐ;Ó@Ù˜uŸ{™{Ó+€KØ.2Ð˜mÓ2ØˆX�_‰_ð	ó€Fð&   §¡ð	ó"ñ Ðð   / (§/¡/ð		ó#ñ Ðð $3 8§?¡?ð	ó$Ð ðB ò
ó ð
ðB òó ðð> Ø€Y×Ñ˜Óò
ó ó ð
ð$ ò#ó ð#ð. Ø€Y×Ñ˜Óò	@ó ó ð	@ð Ø€Y×Ñ˜ÓòNó ó ðNð  àBFðBØðBØ,/ðBØ<?ðBà	òBó ðBðH ò<ó ð<ð ò8ó ñ8rX   r  c                óB  — ddl m} 	  || t        «      5  |D ]	  } |«        Œ 	 d d d «       y # 1 sw Y   y xY w# t        j                  $ �rP}t
        j                  j                  d«      dk(  �r't        |dd«      ^}}}t
        j                  j                  |«      j                  d«      rét        |«      j                  «       }d}	|j                  |	«      d	k(  sJ ‚ G d
„ d«      }
 |
«       ||j                  d«      d	z   <   t!        j"                  t!        j$                  ddg|¢›d�«      d«      }|j'                  |	|«      }t        dd«      5 }|j)                  |j+                  «       «       d d d «       n# 1 sw Y   nxY wt-        d|› �«      |‚‚ d }~ww xY w)Nr   r  ÚHALIDE_REPROÚ1rf  )r‚  r‚  r‚  Úpythonz    hl.main()rE   c                  ó   — e Zd Zdd„Zy)ú _worker_task_halide.<locals>.Outc                 ó   — y)NrC  rR   r    s    rV   Ú__repr__z)_worker_task_halide.<locals>.Out.__repr__~
  s   € Ø$rX   NrÄ  )rº   r»   r¼   r€  rR   rX   rV   ÚOutr~  }
  s   „ ô%rX   r�  rR  z—                        import sys, tempfile
                        with tempfile.TemporaryDirectory() as out:
                            sys.argv = zrepro.pyz?
                            hl.main()
                        r  r*  zwrote repro.py: )r?  r  r@  r;  r=  rz   rÜ  rØ   r’  r{   r  rŸ  r¦   rº  ÚcountÚindexrý  r®   rþ  rR  r"  r(  r�   )rc  re  r  Újobrò   r|  Úscriptrf  r	  Úmainr�  ÚreplÚfds                rV   r^  r^  n
  s‰  € Ý.ð Ù�h¤Ó-ñ 	Øò �Ù•ñ÷	÷ 	ñ 	ûô ×%Ñ%ó Ü�:‰:�>‰>˜.Ó)¨SÓ0Ü#*¨1¨e°\Ó#BÐ ˆF�F˜SÜ�w‰w×Ñ Ó'×2Ñ2°8Ô<Ü˜F“|×(Ñ(Ó*�Ø&�Ø—z‘z $Ó'¨1Ò,Ð,Ð,÷%ñ %ñ ,/«5��C—I‘I˜d“O aÑ'Ñ(Ü—‘Ü—O‘Oð(ð *4Ð(:°cÐ(:Ð'=ð >ðóð ó
�ð —|‘| D¨$Ó/�Ü˜* cÓ*ð ,¨bØ—H‘H˜TŸ[™[›]Ô+÷,÷ ,ñ ,úä"Ð%5°a°SÐ#9Ó:ÀÐAØûð9úsB   ˆ: •.¥: ®7³: ·: ºFÁDFÅ E=Å4	FÅ=F	ÆFÆFc                ó8   — t        | d«      j                  «        y )Nr8  )r¦   Úcloser”  s    rV   r]  r]  ”
  s   € Üˆ�3Ó×ÑÕrX   c                  óð   — e Zd ZU g Zded<   i Zded<   eddd„«       Ze	 	 	 d	 	 	 	 	 	 	 	 	 dd„«       Ze	 	 d	 	 	 	 	 	 	 	 	 dd„«       Z	eddd	„«       Z
e ej                  d«      	 	 	 	 	 	 dd
„«       «       Zy)ÚPyCodeCachezlist[ModuleType]rö  z dict[str, list[tuple[Any, ...]]]Úlinemapsc                ó   — t        |d|¬«      S ©NÚpyrÀ  r%  )rÐ  rl  r
  s      rV   r"  zPyCodeCache.write¡
  s   € ä�[ $¨eÔ4Ð4rX   Nc                óL   — t        |d|¬«      \  }}| j                  ||||«      S r�  )r"  Úload_by_key_path)rÐ  rl  r
  ÚlinemapÚattrsrÅ   r{   s          rV   r§   zPyCodeCache.load¥
  s-   € ô ˜+ t°5Ô9‰	ˆˆTØ×#Ñ# C¨¨w¸Ó>Ð>rX   c                ó.  — |€g }t        ||«      }t        t        |Ž «      | j                  |<   |�%|j	                  «       D ]  \  }}t        |||«       Œ |s"|s t        j                  t        ||«      |_	        | j                  j                  |«       |S rQ   )r2   r�  r£  r�  rì   Úsetattrr¾   r
   r3   Ú_reload_in_subprocrö  r   )rÐ  rÅ   r{   r“  r”  Úmodr§  r¨  s           rV   r’  zPyCodeCache.load_by_key_path°
  s˜   € ð ˆ?ØˆGä# C¨Ó.ˆô "¤# w -Ó0ˆ�‰�TÑàÐØŸ™›ò #‘��1Ü˜˜Q Õ"ð#ñ ™5Ü%.×%6Ñ%6Ü0°#°tó&ˆCÔ"ð 	�‰×Ñ˜3ÔØˆ
rX   c                óØ   — |r?| j                   D ]0  }	 |j                  sJ ‚t        j                  |j                  «       Œ2 | j                   j                  «        y# t        $ r Y ŒYw xY w)z‚
        Clear the in-memory module cache. If purge=True, also delete all the
        corresponding on-disk source files.
        N)rö  rÊ  rz   r:  rÂ  rÃ  )rÐ  Úpurger˜  s      rV   r×  zPyCodeCache.cache_clearÌ
  sa   € ñ Ø—{‘{ò �ðØŸ<š<Ð'˜<Ü—I‘I˜cŸl™lÕ+ðð 	�‰×ÑÕøô )ò Ùðús   “-AÁ	A)Á(A)c                ó–   — || j                   vry | j                   |   \  }}t        ||«      }|dk(  ry ||dz
     }|sy dd„} ||«      S )Nr   rE   c           	     óœ   — d}t        j                  || «      }t        |«      D ���cg c]  \  }}}|t        |«      |dœ‘Œ c}}}S c c}}}w )Nz"File "(.+)", line (\d+), in (.+)\n)r]  r,  r„   )r„  ÚfindallÚreversedr.  )Ústack_traceÚregexÚmatchesr?  Úlré  s         rV   Úparse_stack_tracez<PyCodeCache.stack_frames_for_code.<locals>.parse_stack_traceë
  sV   € ð :ˆEÜ—j‘j ¨Ó4ˆGô  (¨Ó0÷ð á�A�q˜!ð ¬¨A«¸Ó:ôð ùô s   ¨A)rŸ  r²   rµ   zlist[dict[str, Any]])r�  r   )rÐ  r{   Úlinenor£  r“  r¡  r  r£  s           rV   Ústack_frames_for_codez!PyCodeCache.stack_frames_for_codeÛ
  s`   € ð
 �s—|‘|Ñ#Øà—|‘| DÑ)‰ˆˆuÜ˜ Ó'ˆØ�Š6ØØ�a˜!‘e‘ˆÙØó	ñ ! Ó'Ð'rX   ©r‚  )rl  r²   r
  r²   rµ   útuple[str, str])r‚  NN)
rl  r²   r
  r²   r“  úOptional[list[tuple[int, str]]]r”  rÌ   rµ   r   )NN)
rÅ   r²   r{   r²   r“  r¨  r”  rÌ   rµ   r   rª  )rš  ræ   rµ   r¹   )r{   r²   r¤  r.  rµ   zOptional[list[dict[str, Any]]])rº   r»   r¼   rö  rG  r�  rØ  r"  r§   r’  r×  r¾   r¿   r¥  rR   rX   rV   rŒ  rŒ  ˜
  s  … ð !#€GÐÓ"Ø13€HÐ.Ó3àó5ó ð5ð ð Ø37Ø*.ð?àð?ð ð?ð 1ð	?ð
 (ð?ð 
ò?ó ð?ð ð
 48Ø*.ðàðð ðð 1ð	ð
 (ðð 
òó ðð6 óó ðð Ø€Y×Ñ˜Óð(Øð(Ø #ð(à	'ò(ó ó ñ(rX   rŒ  c                ó@   — t        t        j                  |«      | «      S rQ   )r’  rŒ  r§   )Úkernel_namerl  s     rV   Ú_load_triton_kernel_from_sourcer«  ø
  s   € ô ”;×#Ñ# KÓ0°+Ó>Ð>rX   c                 ó†  — t        j                  t        j                  j                  «      rt        j                  j                  S t        j
                  «       r/t        j                  j                  t        j                  dd«      S t        j                  t        j                  d«      «      rt        j                  dd«      S t        j                  t        j                  d«      «      rQt        j                  j                  t        j                  j                  t        j                  dd«      d«      «      S y)NÚbinÚnvccÚCUDACXXr‚  Ú	CUDA_HOMEzbin/nvcc)r    Ú
nvcc_existr   r‡   Úcuda_cxxrÓ  rz   r{   r|   rK   Úsdk_homeÚgetenvÚrealpathrR   rX   rV   Ú_cuda_compilerr¶  þ
  sÄ   € Ü×Ñœ6Ÿ;™;×/Ñ/Ô0Ü�{‰{×#Ñ#Ð#Ü×ÑÔÜ�w‰w�|‰|œK×0Ñ0°%¸Ó@Ð@Ü×Ñœ2Ÿ9™9 YÓ/Ô0Ü�y‰y˜ BÓ'Ð'Ü×Ñœ2Ÿ9™9 [Ó1Ô2Ü�w‰w×Ñ¤§¡§¡¬R¯Y©Y°{ÀBÓ-GÈÓ TÓUÐUØrX   c            	     ór  — t        j                  «       rddlm}  | j	                  d«      }nt         j
                  j                  }t        j                  j                  t        j                  j                  |d«      «      t        j                  j                  t        j                  j                  |d«      «      t        j                  j                  t        j                  j                  |d«      «      t        j                  j                  t        j                  j                  |d«      «      gS )Nr   rÐ  zcutlass-3-headersÚincludeztools/library/includeztools/library/srcztools/util/include)r   rÓ  rÔ  rÑ  Úget_dir_pathr‡   Úcutlass_dirrz   r{   rµ  r|   )rÑ  Úcutlass_paths     rV   Ú_cutlass_include_pathsr¼  
  sÄ   € Ü×ÑÔÝ$à×+Ñ+Ð,?Ó@‰ä—{‘{×.Ñ.ˆô 	�‰×ÑœŸ™Ÿ™ l°IÓ>Ó?Ü
�‰×ÑœŸ™Ÿ™ lÐ4KÓLÓMÜ
�‰×ÑœŸ™Ÿ™ lÐ4GÓHÓIÜ
�‰×ÑœŸ™Ÿ™ lÐ4HÓIÓJðð rX   c                 ó  — t        «        ddlm}  | j                  d¬«      }g }t	        «       rPt        |«       |D ]  }|j                  d|› �dd|› �g«       Œ |j                  d«       |j                  d	«       |S t        d
«      ‚)Nr   )Úcpp_extensionr‡   rn  z-Lz-Xlinkerz-rpath=z-lcudaz-lcudartzMUnsupported env, failed to find cuda libs! Currently only Linux is supported.)	r$   Útorch.utilsr¾  Úlibrary_pathsr8   r&   r&  r   ÚNotImplementedError)r¾  ÚlpathsÚextra_ldflagsr{   s       rV   Ú_cuda_lib_optionsrÄ    s�   € ÜÔÝ)à×(Ñ(°VÐ(Ó<€FØ!€MÜ„zÜ˜fÔ%Øò 	NˆDð × Ñ  B t f +¨z¸WÀTÀFÐ;KÐ!LÕMð	Nð 	×Ñ˜XÔ&Ø×Ñ˜ZÔ(ð
 Ðô "Ø[ó
ð 	
rX   c                 ó
   — g d¢S )N)z-fPICz-fno-strict-aliasingz-fvisibility=hiddenz-WconversionrR   rR   rX   rV   Ú_nvcc_host_compiler_optionsrÆ  /  s   € òð rX   c                 ó¼  — t        j                  «       } | dk(  rd} d| › �d| › �g}t        j                  j                  r	|d| › �gz  }dddd	d
d| › ddj                  |«      › d�t        j                  j                  dddg
}t        j                  «       r>|j                  dt        j                  j                  t        j                  «      g«       t        j                  j                  r|j                  g d¢«       t        j                  j                  r|j                  g d¢«       t        j                  j                   r|j                  ddg«       |S )NÚ90Ú90aÚsm_Úcompute_Últo_z-t=0z"-DCUTLASS_ENABLE_TENSOR_CORE_MMA=1z+-DCUTLASS_ENABLE_SM90_EXTENDED_MMA_SHAPES=1z'-DCUTE_SM90_EXTENDED_MMA_SHAPES_ENABLEDz-wz-gencode=arch=compute_z,code=[ú,ú]z
-std=c++17z--expt-relaxed-constexprz-DNDEBUGz-ccbin)z	-lineinforQ  z-DCUTLASS_DEBUG_TRACE_LEVEL=1)z--keepz,--ptxas-options=--warn-on-local-memory-usagez --ptxas-options=--warn-on-spillsz--resource-usagez--source-in-ptxz--use_fast_mathz -DCUTLASS_USE_TANH_FOR_SIGMOID=1)r    Úget_cuda_archr   r‡   Úenable_cuda_ltor|   Úcompile_opt_levelrÓ  r&  rz   r{   rÉ  rK   ÚgccÚenable_debug_infoÚenable_ptxas_infoÚuse_fast_math)Úarchr	  Úoptionss      rV   Ú_nvcc_compiler_optionsrØ  8  s.  € Ü×!Ñ!Ó#€DØˆt‚|àˆØ�$�ˆL˜H T FÐ+Ð,€DÜ‡{�{×"Ò"Ø�4˜�v��ÑˆàØ,Ø5Ø1ØØ
    g¨c¯h©h°t«nÐ-=¸QÐ?Ü�‰×%Ñ%ØØ"Øð€Gô ×ÑÔØ�‰˜¤"§'¡'§/¡/´+·/±/Ó"BÐCÔDÜ‡{�{×$Ò$Ø�‰ÒKÔLÜ‡{�{×$Ò$Ø�‰òô	
ô ‡{�{× Ò Ø�‰à!Ø2ðô	
ð €NrX   c                ór  — |€g }t        «       }t        «       }t        «       }t        «       }||z   |D �cg c]  }d|v rd|› �nd|› �‘Œ c}z   |D �	cg c]  }	d|	z   ‘Œ	 c}	z   |z   }
dj	                  | «      }d}|dk(  r%t        «       › ddj	                  |
«      › d|› d|› �}nt|d	k(  r6|
j                  d
«       t        «       › ddj	                  |
«      › d|› d|› �}n9|dk(  r%t        «       › ddj	                  |
«      › d|› d|› �}nt        d|› d�«      ‚t        j                  d|«       |S c c}w c c}	w )Nú=z-Xcompiler z-Xcompiler=z-Ir7  r‚  rÜ  z -c -o Úsoz-sharedz -o ÚexezUnsupported output file suffix ú!zCUDA command: %s)
r¼  rÄ  rÆ  rØ  r|   r¶  r   rÁ  r‘  r$  )Ú	src_filesÚdst_fileÚdst_file_extÚ
extra_argsÚinclude_pathsÚcuda_lib_optionsÚnvcc_host_compiler_optionsÚnvcc_compiler_optionsÚoptr{   r×  Úsrc_fileÚress                rV   Úcuda_compile_commandré  d  s†  € ð ÐØˆ
Ü*Ó,€MÜ(Ó*ÐÜ!<Ó!>ÐÜ2Ó4ÐàØ
ñ	ð 2ö
àð $'¨#¡:ˆk˜#˜Ñ°[ÀÀÐ3FÑFò
ñ	
ð $1Ö
1˜4ˆ4�$‹;Ò
1ñ	2ð ñ	ð ð �x‰x˜	Ó"€HØ
€CØ�sÒÜÓ!Ð" ! C§H¡H¨WÓ$5Ð#6°g¸h¸ZÀqÈÈ
ÐS‰Ø	˜Ò	Ø�‰�yÔ!ÜÓ!Ð" ! C§H¡H¨WÓ$5Ð#6°d¸8¸*ÀAÀhÀZÐP‰Ø	˜Ò	ÜÓ!Ð" ! C§H¡H¨WÓ$5Ð#6°d¸8¸*ÀAÀhÀZÐP‰ä!Ð$CÀLÀ>ÐQRÐ"SÓTÐTÜ‡I�IÐ  #Ô&Ø€Jùò'
ùò 2s   µD/ÁD4c                  óP   — e Zd ZdZ	 	 	 	 d
d„Zdd„Zdd„Zdd„Zdd„Zdd„Z	dd„Z
y	)Ú
DLLWrapperz A wrapper for a dynamic library.c                ób   — || _         d| _        t        j                  |«      | _        d| _        y )NFT)Úlib_pathÚis_openr   r¶  ÚDLL)r¡   rí  s     rV   r¢   zDLLWrapper.__init__Œ  s+   € ð !ˆŒØˆŒÜ×#Ñ# HÓ-ˆŒØˆ�rX   c                óL   — | j                   r| j                  «        d| _         y y r\   )rî  Ú_dlcloser    s    rV   rŠ  zDLLWrapper.close•  s   € Ø�<Š<Ø�M‰MŒOØ ˆD�Lð rX   c                ó   — d }t        «       r;t        d «      }t        |d«      st        d«      }t        |d«      rF|j                  }n9t	        «       r$dd l}|j                  dd¬«      }|j                  }nt        d«      ‚|�wt        «       r)t        g|_	         || j                  j                  «       y t	        «       r9dd l}ddlm} |j                  g|_	         || j                  j                  «       y y t        j                  d	«       y )
NÚdlclosezlibc.sor   Úkernel32T)Úuse_last_errorz&Unsupported env, failed to do dlclose!)ÚwintypeszKdll unloading function was not found, library may not be unloaded properly!)r8   r   rP  ró  r9   r$  ÚFreeLibraryrÁ  r   rõ  rï  Ú_handlerö  ÚHMODULEr‘  r’  )r¡   Ú	f_dlcloseÚsymsr$  rô  rö  s         rV   rñ  zDLLWrapper._dlcloseš  s×   € Øˆ	äŒ:Ü˜“:ˆDÜ˜4 Ô+ä˜I“�ä�t˜YÔ'Ø ŸL™L‘	ÜŒ\Ûà—{‘{ :¸d�{ÓCˆHà ×,Ñ,‰Iä%Ð&NÓOÐOàÐ ÜŒzÜ&. Z�	Ô"Ù˜$Ÿ(™(×*Ñ*Õ+Ü”ÛÝ+à&.×&6Ñ&6Ð%7�	Ô"Ù˜$Ÿ(™(×*Ñ*Õ+ð ô �K‰KØ]õrX   c                óˆ   ‡— | j                   st        d| j                  › �«      ‚t        | j                  |«      Šdˆfd„}|S )NzCannot use closed DLL library: c                 óD   •—  ‰| Ž }|rt        d‰j                  › �«      ‚y )NzError in function: )r�   rº   )rT   ÚerrÚmethods     €rV   Ú_wrapped_funcz-DLLWrapper.__getattr__.<locals>._wrapped_funcÃ  s,   ø€ Ù˜$�-ˆCÙÜ"Ð%8¸¿¹Ð8IÐ#JÓKÐKð rX   ©rT   r   rµ   r¹   )rî  r�   rí  r’  rï  )r¡   r„   r   rÿ  s      @rV   Ú__getattr__zDLLWrapper.__getattr__½  s?   ø€ Ø�|Š|ÜÐ!@ÀÇÁÀÐPÓQÐQä˜Ÿ™ 4Ó(ˆõ	Lð
 ÐrX   c                ó   — | S rQ   rR   r    s    rV   Ú	__enter__zDLLWrapper.__enter__Ê  s   € ØˆrX   c                ó$   — | j                  «        y rQ   ©rŠ  )r¡   rT   s     rV   Ú__exit__zDLLWrapper.__exit__Í  ó   € Ø�
‰
�rX   c                ó$   — | j                  «        y rQ   r  r    s    rV   Ú__del__zDLLWrapper.__del__Ð  r  rX   N)rí  r²   rµ   r¹   r¸   )r„   r²   rµ   zCallable[..., None])rµ   r   r  )rº   r»   r¼   rF  r¢   rŠ  rñ  r  r  r  r
  rR   rX   rV   rë  rë  ‰  s;   „ Ù*ðàðð 
óó!ó
!óFóóôrX   rë  c                  ó¾   — e Zd ZU ej                   G d„ d«      «       Zi Zded<    eej                  «      Z
dZed
d„«       Ze	 d	 	 	 	 	 	 	 dd„«       Zedd	„«       Zy)rS  c                  ó"   — e Zd ZU ded<   ded<   y)úCUDACodeCache.CacheEntryr²   rË  rÏ  N©rº   r»   r¼   rG  rR   rX   rV   Ú
CacheEntryr  Ö  ó   … à‹ØÔrX   r  údict[str, CacheEntry]r›   Úcuc                ón   — t        t        dgd|«      «      }t        || j                  |¬«      \  }}||fS ©z 
        Writes source code into a file with dst_file_ext as the file extension.
        Returns the hash key of source code, and the path to the file.
        Údummy_inputÚdummy_outputrÀ  )r  ré  r"  Ú_SOURCE_CODE_SUFFIX©rÐ  rl  rà  Úcuda_commandrÅ   rË  s         rV   r"  zCUDACodeCache.writeß  óE   € ô Ü  - °.À,ÓOó
ˆô  Ø˜×0Ñ0¸ô
‰ˆˆZð �JˆÐrX   Nc                ó\  — | j                  ||«      \  }}|| j                  v�r™ddlm} t	        «       } |t
        j                  j                  ||dz   «      t        ¬«      }|5  |dt        | j                  «        |z   }	t
        j                  j                  |	«      sÛt        |g|	||«      }
t        |d«      5 }|j                  d«       |j                  d|
› d�«       ddd«       t        «       }t        j!                  d	|
«       |
j#                  d
«      }	 t%        j&                  |t$        j(                  t
        j*                  ¬«       t        «       }d||z
  › d|
› �}t        j5                  |«       nt        j!                  d|«       t6        j9                  ||	«      | j                  |<   ddd«       | j                  |   j:                  ||fS # 1 sw Y   Œ÷xY w# t$        j,                  $ r&}t/        j0                  ||j2                  «      |‚d}~ww xY w# 1 sw Y   ŒmxY w)z™
        Compiles CUDA source_code into a file with dst_file_ext extension.
        Returns a tuple of dst_file_path, hash_key, source_code_path
        r   r  r  r  Nr8  r#  z// CUDA Compile cmd
// zCUDA Compilation: %sr7  )ÚstderrÚenvzCUDA Compilation took ú seconds. Compile command: z8CUDA Compilation skipped: %s since output already exists)r"  r›   r?  r  rû   rz   r{   r|   r@  rŸ  r  rø   ré  r¦   r   r‘  r$  rÊ  r;  r<  ÚSTDOUTrÜ  ÚCalledProcessErrorr   ÚCUDACompileErrorÚoutputr1  rS  r  rÏ  )rÐ  rl  rà  rá  rÅ   rË  r  rú   rd  rÏ  rf  r?  Ú
start_timeÚ	cmd_partsÚerrorÚend_timeÚlog_duration_msgs                    rV   rŽ  zCUDACodeCache.compileî  sð  € ð Ÿ)™) K°Ó>‰ˆˆZØ�c—i‘iÒÝ6ä#“~ˆHÙœBŸG™GŸL™L¨°3¸±=ÓAÌ<ÔXˆDØñ SØ(Ð)H¬C°×0GÑ0GÓ,HÐ+HÐIÈLÑX�Ü—w‘w—~‘~ kÔ2Ü.Ø#˜ k°<Àó�Cô ˜j¨#Ó.ð D°!ØŸ™ œØŸ™Ð":¸3¸%¸rÐ BÔC÷Dô "&£�JÜ—I‘IÐ4°cÔ:Ø #§	¡	¨#£�IðWÜ"×/Ñ/Ø%¬j×.?Ñ.?ÄRÇZÁZõô
  $›v�HØ)?ÀÈ:Ñ@UÐ?VÐVqÐruÐqvÐ'wÐ$Ü—H‘HÐ-Õ.ä—I‘IØRØ"ôô "/×!9Ñ!9¸*ÀkÓ!R�—	‘	˜#‘÷5Sð8 —	‘	˜#‘×*Ñ*¨C°Ð<Ð<÷-Dð Dûô &×8Ñ8ò WÜ!×2Ñ2°9¸e¿l¹lÓKÐQVÐVûðWú÷!Sð SúsJ   Á&AH"Â='GÃ$9H"Ä4G&ÅA$H"ÇG#	ÇH"Ç&HÇ9!HÈHÈH"È"H+c                óv   — |dk7  rt        d|› d|› �«      ‚| j                  ||«      \  }}}t        |«      ||fS ©zŠ
        Compiles source code and loads the generated .so file.
        Returns a tuple of DLLWrapper, hash_key, source_code_path
        rÛ  zCOnly support loading a .so file for now. Requested file extension: z. Source code: ©r�   rŽ  rë  ©rÐ  rl  rà  Údst_file_pathr×   Úsource_code_paths         rV   r§   zCUDACodeCache.load  óa   € ð ˜4ÒÜð-Ø-9¨N¸/È+ÈðXóð ð 58·K±KØ˜ó5
Ñ1ˆ�xÐ!1ô ˜=Ó)¨8Ð5EÐFÐFrX   ©rl  r²   rà  r²   rµ   r§  rQ   ©rl  r²   rà  r²   rá  úOptional[list[str]]rµ   útuple[str, str, str]©rl  r²   rà  r²   rµ   ztuple[DLLWrapper, str, str])rº   r»   r¼   rQ  Ú	dataclassr  r›   rG  r½   rÃ  r×  r  rØ  r"  rŽ  r§   rR   rX   rV   rS  rS  Ô  s    … à×Ñ÷ð ó ðð $&€EÐ Ó%Ù˜uŸ{™{Ó+€KØÐàòó ðð àTXð)=Øð)=Ø-0ð)=Ø>Qð)=à	ò)=ó ð)=ðV òGó ñGrX   rS  c                  óÂ   — e Zd ZU ej                   G d„ d«      «       Zi Zded<    eej                  «      Z
dZdZedd„«       Ze	 d	 	 	 	 	 	 	 dd	„«       Zedd
„«       Zy)rR  c                  ó"   — e Zd ZU ded<   ded<   y)úROCmCodeCache.CacheEntryr²   rË  rÏ  Nr  rR   rX   rV   r  r7  .  r  rX   r  r  r›   rè  Fc                ón   — t        t        dgd|«      «      }t        || j                  |¬«      \  }}||fS r  )r  r!   r"  r  r  s         rV   r"  zROCmCodeCache.write8  r  rX   Nc                óV  — | j                   s6d| _         t        j                  t        t	        t        «       «      «      «       | j                  ||«      \  }}|| j                  v�r`ddlm	} t        «       } |t        j                  j                  ||dz   «      t        ¬«      }|5  |dt        | j                   «        |z   }	t        j                  j#                  |	«      s¡t%        |g|	||«      }
t'        «       }|
j)                  d«      }	 t+        j,                  |t*        j.                  dt        j0                  ¬«      }t        j                  d	|«       t'        «       }d
||z
  › d|
› �}t        j;                  |«       nt        j                  d||	«       t<        j?                  ||	«      | j                  |<   ddd«       | j                  |   j@                  ||fS # t*        j2                  $ r&}t5        j6                  ||j8                  «      |‚d}~ww xY w# 1 sw Y   ŒaxY w)zÖ
        Compiles source_code into a file with dst_file_ext extension,
        using the compile command specific for the ROCm platform.
        Returns a tuple of dst_file_path, hash_key, source_code_path
        Tr   r  r  r  Nr7  )r  r&  r  zCompilation output: %szCompilation took r  z+Skip compiling %s: output %s already exists)!Ú_logged_compiler_versionr‘  r$  r*   r²   r"   r"  r›   r?  r  rû   rz   r{   r|   r@  rŸ  r  rø   r!   r   rÊ  r;  r<  r  rÜ  r   r   r!  r"  r1  rR  r  rÏ  )rÐ  rl  rà  rá  rÅ   rË  r  rú   rd  rÏ  rf  r#  r$  r"  r%  r&  r'  s                    rV   rŽ  zROCmCodeCache.compileG  sß  € ð ×+Ò+Ø+/ˆCÔ(Ü�I‰IÔ/´´M³OÓ0DÓEÔFàŸ)™) K°Ó>‰ˆˆZØ�c—i‘iÒÝ6ä#“~ˆHÙœBŸG™GŸL™L¨°3¸±=ÓAÌ<ÔXˆDØñ SØ(Ð)H¬C°×0GÑ0GÓ,HÐ+HÐIÈLÑX�Ü—w‘w—~‘~ kÔ2Ü.Ø#˜ k°<Àó�Cô "&£�JØ #§	¡	¨#£�Ið	WÜ!+×!8Ñ!8Ø%Ü#-×#4Ñ#4Ø!%Ü "§
¡
ô	"˜ô Ÿ	™	Ð":¸FÔCô  $›v�HØ):¸8ÀjÑ;PÐ:QÐQlÐmpÐlqÐ'rÐ$Ü—H‘HÐ-Õ.ä—I‘IØEØ"Ø#ôô
 "/×!9Ñ!9¸*ÀkÓ!R�—	‘	˜#‘÷7Sð: —	‘	˜#‘×*Ñ*¨C°Ð<Ð<øô &×8Ñ8ò WÜ!×2Ñ2°9¸e¿l¹lÓKÐQVÐVûðWú÷!Sð Sús3   Â(A&HÄAG#ÅA%HÇ#HÇ6!HÈHÈHÈH(c                óv   — |dk7  rt        d|› d|› �«      ‚| j                  ||«      \  }}}t        |«      ||fS r)  r*  r+  s         rV   r§   zROCmCodeCache.loady  r.  rX   r/  rQ   r0  r3  )rº   r»   r¼   rQ  r4  r  r›   rG  r½   rÃ  r×  r  r:  rØ  r"  rŽ  r§   rR   rX   rV   rR  rR  ,  s¦   … à×Ñ÷ð ó ðð $&€EÐ Ó%Ù˜uŸ{™{Ó+€KØÐØ$Ðàòó ðð àTXð/=Øð/=Ø-0ð/=Ø>Qð/=à	ò/=ó ð/=ðb òGó ñGrX   rR  c                  ó   — e Zd Zdd„Zy)ÚCodeCacheFuturec                ó   — t         ‚rQ   )rÁ  r    s    rV   r®  zCodeCacheFuture.resultŒ  s   € Ü!Ð!rX   N©rµ   úCallable[..., Any])rº   r»   r¼   r®  rR   rX   rV   r=  r=  ‹  s   „ ô"rX   r=  c                  ó*   — e Zd Z	 d	 	 	 	 	 dd„Zdd„Zy)ÚLambdaFutureNc                ó    — || _         || _        y rQ   )Ú	result_fnrÃ  )r¡   rD  rÃ  s      rV   r¢   zLambdaFuture.__init__‘  s   € ð #ˆŒØˆ�rX   c                ó"   — | j                  «       S rQ   )rD  r    s    rV   r®  zLambdaFuture.result—  s   € Ø�~‰~ÓÐrX   rQ   )rD  r@  rÃ  zOptional[Future[Any]]rµ   r¹   r?  )rº   r»   r¼   r¢   r®  rR   rX   rV   rB  rB  �  s'   „ àMQðØ+ðØ5Jðà	óô rX   rB  )rT   r   rU   r   rµ   r¹   )rµ   ræ   rÄ  )r   r²   rµ   r·   )r  r  rµ   r²   r¦  )r	  úUnion[str, bytes]r
  rF  rµ   r²   )r  r²   r  r²   r  r²   rµ   r2  )r‚  r	  )r  rF  r
  r²   r  r²   rµ   r²   )r‚  r	  r‚  )r  rF  r  r²   r
  r²   r  r²   r  r²   rµ   r§  )r&  r²   rµ   r²   )FF)
r;  r²   r  rF  r°   ræ   r<  ræ   rµ   r¹   )rJ  rn   rµ   rn   )rS  r   rµ   r=   )r¾  zlist[str] | Noner¿  r²   rÀ  zhashlib._Hashrµ   r¹   )rµ   r  )
rn  r«  rí  r   rð  ra   rÿ  r!  rµ   ztuple[str, list[str]])r5  r.  rµ   r.  )r{   r²   rµ   r§  )rß   r²   rT   r   rµ   zUnion[list[c_void_p], c_void_p])rÌ  r²   rÎ  r'   rµ   r¹   )rc  r²   re  zlist[partial[Any]]rµ   r¹   )r]  r²   rµ   r¹   )rª  r²   rl  r²   rµ   rl   )rµ   r�  )rµ   r­  rQ   )
rÞ  r­  rß  r²   rà  r²   rá  r1  rµ   r²   )çÚ
__future__r   rþ   r]  rQ  r¾   r‘   r   rY  rø  r“   Úloggingrz   rM  rµ  r„  rq  r8  rM  r;  r;  r»  rý  r3  rx  Úbisectr   r   r$  r   r   r   Údatetimer	   r
   Úpathlibr   r   r   Útypesr   Útypingr   r   r   r   r   r   r   r   Útyping_extensionsr   rt   Útorch.distributedr,  r2  r   r   Útorch._dynamo.utilsr   r   r   Útorch._inductorr   r   r   Útorch._inductor.codegen.cudar    Ú,torch._inductor.codegen.rocm.compile_commandr!   r"   Útorch._inductor.cpp_builderr#   r$   r%   r&   r'   r(   r)   r*   r+   r,   Útorch._inductor.cpu_vec_isar-   Ú!torch._inductor.custom_graph_passr.   r/   Útorch._inductor.freezing_utilsr0   r1   Ú%torch._inductor.runtime.compile_tasksr2   r3   Ú%torch._inductor.runtime.runtime_utilsr4   r5   Útorch._inductor.utilsr6   r7   r8   r9   Útorch._loggingr:   Útorch._subclasses.fake_tensorr;   r<   r=   Útorch._utils_internalr>   Útorch.compilerrî  Útorch.compiler._cacher?   r@   Ú%torch.fx.experimental.symbolic_shapesrA   rB   rC   Útorch.utils._ordered_setrD   Úpackage.pt2_archive_constantsrF   rS  rG   ÚruntimerH   Úruntime.autotune_cacherI   Útriton_bundlerrJ   rÓ  Ú	triton.fbrK   Útorch._inductor.fb.utilsrL   rM   rN   rO   Úcollections.abcr]   r^   r_   Úconcurrent.futuresr`   r‡  ra   rb   rs  rd   Úirre   rp   rf   rg   rh   ri   Úruntime.hintsrj   rk   Úruntime.triton_heuristicsrl   r  rm   rn   r  r7  r@  Ú_loggingÚgetArtifactLoggerrº   ru  Ú	getLoggerr‘  rw   r¿   r€   r}   rÁ   rÎ   rû   r  r  r  r  r"  r'  r±   r4  rA  rK  rU  ÚPicklerrW  r¼  r×  rÙ  rÛ  rO  ru  r¬  r(  r8  r:  rË  rÍ  rÚ  r‘  r•  r–  rG  r°  r²  rÇ  rÔ  r   r  r^  r]  rŒ  r«  r¶  r¼  rÄ  rÆ  rØ  ré  rë  rS  rR  r=  rB  rR   rX   rV   ú<module>rq     sô  ðÞ "ã Û Û Û Û Û Û 	Û Û Û Û 	Û Û Û 	Û Û Û Û Û 
Û Û Û Û Ý Ý ß 'Ñ 'Ý Ý Ý ß Ý ÷	÷ 	ó 	õ #ã Ý  ß  ß JÑ Jß 0Ñ 0Ý 1÷÷÷ ÷ õ 5ß Rß M÷÷ O÷ó õ ,÷ñ õ
 3Ý ,ß Iß NÑ NÝ /å EÝ &Ý #Ý 8Ý )ð €6×ÒÔÝ%÷ô óóóóñ ß=Ñ=Ý)ç=Ý$Ý ßAß5ß:Ý;Ý á�‹€Að �lŠl˜gÑ%€Ø€à—.’.×2Ò2°8¸]ÓK€Ø€g×Ò˜Ó!€óGð €×Ò�TÓòó ð÷B
ñ B
ôJ'�ô 'ô2X�iô XóvóXô
*ð 9;ð"Øð"Ø!ð"Ø25ð"àó"ð CIð;Øð;Ø'*ð;Ø<?ð;àó;ð ØØðØðàðð ðð ð	ð
 ðð óó"!ð Øð	Øðàðð ðð ð	ð
 
óð> ×Ò÷ð ó ðóó	ôp˜&Ÿ.š.ô pðfVØðVØ%(ðVØ2?ðVà	óVð  €×Ò�TÓòXó ðXóD%ð ×Ò÷ð ó ðô˜ô ÷}"ñ }"ð@Øðà'ðð  ðð #ð	ð
 óó,!÷8Cñ CðL €×Ò�TÓòó ðð ÷ ð  ó ð ÷6eñ eð` Ø
×Òò.ó ó ð.ó(ð  €ˆ.Ó ó*HðZ ÷k:ð k:ó ðk:ð\ Øð àð ð 
ó ð ôf; ó f;ó ðf;ðR ô=Ð4ó =ó ð=ð@ ô\Ð0ó \ó ð\ó~
#óL ð ÷\(ð \(ó ð\(ð~?Øð?Ø#&ð?àó?ó	óó ó*ó)ð` '+ð	"Øð"àð"ð ð"ð $ð	"ð
 	ó"÷JHñ HðV ÷TGð TGó ðTGðn ÷[Gð [Gó ð[G÷|"ñ "ô
 �?õ  rX   