o
    IhW                    @  s  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mZmZmZmZmZ d dlmZ d dl	mZ d dlm Z m!Z!m"Z"m#Z#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z* d dl+m,Z,m-Z-m.Z.m/Z/m0Z0 d dlm1Z1 d dl2Z2d dl3Z3d d	l4m5Z5 d d
l6m7Z7 d dl8m9Z9 e(rd dlm:Z:m;Z;m<Z< d dl3m=Z=m>Z>m?Z? d dl@mAZA d dlBmCZC d dlDmEZE d dlFmGZG ddlHmIZI ddlJmKZK ddlLmMZM ddlNmOZOmPZPmQZQmRZRmSZSmTZT ddlUmVZV ddlWmXZXmYZY g dZZe)dZ[e\dd_ddZ]d dl^m_Z_ d d l`maZa d d!lbmcZc d d"ldmeZe d d#lfmgZg d d$lhmiZi d d%ljmkZkmlZlmmZmmnZnmoZo d d&lpmqZqmrZr d d'lsmtZtmuZu dd(lvmwZw dd)lxmyZz ejd*kZ{e|e}Z~e)d+Zee2je2jf Ze&e*e3jee3j?f  Zd,d-d.Zd/Zd/Zd/Zd0Zd1Zeed @ d kred2ksJ d3d`d6d7Zdad;d<ZG d=d> d>e2jZ	@dbdcdFdGZe\ddddHdIZdedMdNZdfdQdRZdgdVdWZdhdZd[Zdid_d`ZydjdcddZdkdhdiZdldldmZdmdpdqZdrds fdndxdyZdoddZdpdqddZ		drdsddZ					dtduddZdvddZdwddZdxddZdyddZdzddZe.dZe)dddZG dd de'e#eef Zd{ddZd|ddZd}ddZd~ddÄZ	ddddʄZdddτZddd҄ZdddքZdddلZdddބZdddZdddZdddZdddZdddZdddZdddZg Zd}ed< dddZdddZd dlZdddZej			ddddZdddZdddZe\d2dddZG dd de%ZejG dd dZG dd dZG dd deƃZejdddZG dd dZG dd deɃZe\dddd!d"Zej\dd#d$Z̐dd%d&Z͐dd)d*Zΐddd+d,Zϐdd1d2ZАdd4d5Zѐdd6d7ZҐd8d8d9dd<d=ZӐdd@dAZԐddEdFZe\dddGdHZe\dddJdKZאddLdMZؐddNdOZِddPdQZڐddRdSZېddWdXZܐ	8		8	ddd^d_Zݐddd`daZG dbdc dcZߐddhdiZddkdlZddmdnZddodpZddqdrZddsdtZddvdwZejddzd{Z	ddddZdddZdddZdddZdddZdddZdddZejdddZdddZe\ddddZe\ddddZe\ddddZdddZdddZdddZddddZddddZdddZdddZG dd dejZdddZdddZdddZ	ddddZ dddÄZddƐdǄZddȐdɄZdd̐d̈́Zddѐd҄Zdds fddڐdۄZdds fddݐdބZdddZdddZ	ejG dd dZ
ejdddZdddZdddZdddZdÐddZdĐddZdŐddZdƐddZdǐddZdȐddZdɐddZdʐddZdːddZd̐ddZd͐ddZdΐddZdϐddZdАd d!Zddd"d#Zdd$d%Zd&d'd(d)d*d+d,Zd-d. e  D Z!e"d/Z#dѐd0d1Z$dҐd2d3Z%dӐd6d7Z&dӐd8d9Z'e\ddԐd;d<Z(ejG d=d> d>Z)i Z*d?ed@< dՐdDdEZ+d֐dFdGZ,e)dHZ-e)dIZ.G dJdK dKee-e.f Z/e-ddLdddMdאdQdRZ0dؐdTdUZ1dِdWdXZ2G dYdZ dZejZ3e\ddڐd[d\Z4ddd]d^Z5dS (      )annotationsN)
CollectionIteratorMappingMutableMapping
MutableSet)datetime)StringIO)AnyCallablecastGenericLiteral
NamedTupleOptionalProtocolTYPE_CHECKINGTypeVarUnion)Concatenatedataclass_transform	ParamSpecSelf	TypeGuard)mock)DeviceProperties)
OrderedSet)tree_map_only)IterableSequence
ValuesView)SymBoolSymFloatSymInt)ELEMENTWISE_TYPE_PROMOTION_KIND)GraphModule)ShapeEnv)Node   )WorkspaceArgPythonWrapperCodegenGraphLowering)BufferExternKernelIRNodeLayout	OperationReinterpretViewCompiledFxGraph)BaseSchedulerNodeSchedulerBuffer)cudampsxpuTreturnstrc                  C  s>   dd t D } t| dksJ t| dkrd}|S |  }|S )Nc                 S  s   g | ]}t t| r|qS  )getattrtorchis_available.0xr>   r>   I/var/www/vscode/kcb/lib/python3.10/site-packages/torch/_inductor/utils.py
<listcomp>R   s    z get_gpu_type.<locals>.<listcomp>r(   r   r8   )	GPU_TYPESlenpop)
avail_gpusgpu_typer>   r>   rE   get_gpu_typeP   s   rL   )get_interface_for_device)detect_fake_mode)
DeviceType)	EventList)GraphTransformObserver)	ShapeProp)CeilDivCleanDivFloorDivIdentityModularIndexing)make_symbolSymT)bound_sympyValueRanges)config)ceildivwin32_Tz.cubinz.spv)r8   r:         @      zmust be power of 2nbytesintc                 C  s   | t  d t  @ S )z/Round up to the nearest multiple of ALIGN_BYTESr(   )ALIGN_BYTES)rd   r>   r>   rE   _align   s   rg   v
sympy.Exprboolc                 C  s<   t | tjtjfrttt| jS t | tpt	| t
t
kS )z:v can be statically proven to be a multiple of ALIGN_BYTES)
isinstancesympyAddMaxallmap_is_alignedargsaligngcdrf   )rh   r>   r>   rE   rq      s   rq   c                   @  s&   e Zd ZdZdZdZeddd	Zd
S )rs   z<Symbolically round up to the nearest multiple of ALIGN_BYTESr(   Tvalueri   r<   Optional[sympy.Expr]c                 C  s,   t |ttjfrtt|S t|r|S d S N)rk   re   rl   Integerrg   rq   )clsrv   r>   r>   rE   eval   s
   z
align.evalN)rv   ri   r<   rw   )__name__
__module____qualname____doc__nargs
is_integerclassmethodr{   r>   r>   r>   rE   rs      s    rs      d   fnCallable[[], Any]warmuprepfloatc                   s  |   t j  t jtdt jdd}t jjdd}t jjdd}|  tdD ]	}|  |   q)|  t j  |	|d }t
dt|| }t
dt|| }	t|D ]}|   qYt jjt jjjgd}
t|	D ]	}|  |   qot j  W d	   n1 sw   Y  td
 t|
 jddd tdd |
 D }t||	 dkrtdt||	t||	  t fddt|D }|  | }td t|jdd tdd |D d |	 }td| |S )aR  
    Returns benchmark results by examining torch profiler events.
    This could be more accurate as it doesn't count CPU side overhead.
    However, this also requires manually excluding irrelevant event, e.g.
    vectorized_elementwise_kernel which is used to fill L2 cache,
    various CUDA events, etc, so could also be fragile.
    g    Ar8   )dtypedeviceT)enable_timing   r(   )
activitiesNz
raw eventsself_device_time_total)sort_by	row_limitc                 S  s&   g | ]}|j tjkr|jd kr|qS )zContext Sync)device_typerO   CUDAnamerC   eventr>   r>   rE   rF      s
    z,do_bench_using_profiling.<locals>.<listcomp>r   zYFailed to divide all profiling events into #repeat groups. #CUDA events: %d, #repeats: %sc                   s    g | ]\}}|  d kr|qS r   r>   )rC   ir   num_event_per_groupr>   rE   rF      s
    zprofiling time breakdown)r   c                 s      | ]}|j V  qd S rx   )device_time_totalr   r>   r>   rE   	<genexpr>       z+do_bench_using_profiling.<locals>.<genexpr>g     @@zprofiling results: %s ms)r@   r8   synchronizeemptyre   Eventrecordrangezero_elapsed_timemaxprofilerprofileProfilerActivityr   logdebugkey_averagestablerP   eventsrH   RuntimeError	enumerate_build_treesum)r   r   r   cachestart_event	end_event_estimate_msn_warmupn_repeatpr   filtered_eventsactual_eventsresr>   r   rE   do_bench_using_profiling   sh   




r   c               
   C  s   zddl m}  tjdd | d uotttjdd dW S  ty&   Y dS  t	y@ } zdt
|v s5J W Y d }~dS d }~ww )	Nr   )	roi_alignztorchvision::nmsMetatorchvisionr   Fztorchvision::nms does not exist)torchvision.opsr   r@   _C%_dispatch_has_kernel_for_dispatch_keyhasattrr?   opsImportErrorr   r=   )r   er>   r>   rE   has_torchvision_roi_align   s   
r   r   "Union[Optional[torch.device], str]torch.devicec                 C  s`   | d u r
t djS t| trt | } | jdvr.| jd u r.t| j}t j| j|j	 dS | S )Ng        )cpumeta)index)
r@   tensorr   rk   r=   typer   rM   Workercurrent_devicer   device_interfacer>   r>   rE   decode_device   s   


r   itIterable[sympy.Expr]c                 C  s   t tj| tjjS rx   )	functoolsreduceoperatormulrl   SOner   r>   r>   rE   sympy_product	     r   seq1Sequence[sympy.Expr]seq2c                 C  s2   t | t |ks
J ttdd t| |D S )Nc                 s  s    | ]	\}}|| V  qd S rx   r>   )rC   abr>   r>   rE   r     s    zsympy_dot.<locals>.<genexpr>)rH   rl   expandr   zip)r   r   r>   r>   rE   	sympy_dot  s   r   Iterable[_T]ValuesView[_T]c                 C  s   dd | D   S )Nc                 S  s   i | ]}t ||qS r>   )idrB   r>   r>   rE   
<dictcomp>      zunique.<locals>.<dictcomp>)valuesr   r>   r>   rE   unique     r   numerUnion[int, sympy.Expr]denomc              	   C  sr   t | tjst |tjrtt| t|S t | tr!t |ts4J |  dt|  d| dt| t| |S )Nz: , )rk   rl   ExprrS   sympifyre   r   runtime_ceildiv)r   r   r>   r>   rE   r]     s    
r]   keyOptional[torch.dtype]c                 C  s   | d u rdS t | dd }i dddddd	d
ddddddd	ddddddddddddddddd d!d"dd#d$d%d&}t| D ]}|||< qPt| t r^| S d'||  S )(Nz*i8.r   rj   i1
float8e4nvfp8e4nvfloat8e5fp8e5float8e4b15fp8e4b15float8e4b15x4
fp8e4b15x4float8_e4m3fnfloat8_e5m2float8_e8m0fnuu8float16fp16bfloat16bf16float32fp32float64fp64int8i8int16i16int32i32int64i64uint8u16u32u64)uint16uint32uint64*)r=   splitlistr   rk   )r   	dtype_strtysrh   r>   r>   rE   _type_of$  sZ   

r%  lst"Iterable[Union[int, torch.SymInt]]list[sympy.Expr]c                 C  s   dd | D S )z
    Gets the shape and stride of a tensor. For non-symbolic tensors, this is
    trivial. But for symbolic tensors, we need to map from SymIntNode into
    sympy.Expr.
    c                 S  s   g | ]}t |qS r>   )rl   r   rC   r   r>   r>   rE   rF   R  r   z-convert_shape_to_inductor.<locals>.<listcomp>r>   r&  r>   r>   rE   convert_shape_to_inductorJ  s   r+   Iterable[Union[int, sympy.Expr]]list[Union[int, torch.SymInt]]c                   s   ddl m   fdd| D S )zz
    Takes a list of shapes from Inductor and converts them into symints (or just
    ints if all shapes are static).
    r(   Vc                   sB   g | ]}t |tr|nt |tjrt|n	 jjjj|d dqS )N)hint)rk   re   rl   ry   graphsizevars	shape_envcreate_symintnoder)  r.  r>   rE   rF   ^  s    


z+convert_shape_to_symint.<locals>.<listcomp>)virtualizedr/  r*  r>   r.  rE   convert_shape_to_symintU  s   

r6  optorch._ops.OpOverloadc                 C  s   t dd | jjD S )z-
    Does this op overload have aliasing
    c                 s  s    | ]}|j d uV  qd S rx   )
alias_inforC   r   r>   r>   rE   r   p      zis_view.<locals>.<genexpr>)any_schema	argumentsr7  r>   r>   rE   is_viewl  s   r@  c                 C     dS NFr>   )r   r>   r>   rE   <lambda>u      rC  user'   is_pointwise_fn'Callable[[torch._ops.OpOverload], bool]c                   s~   | j dksdS t| jtjjs| jtju sdS ttjj| j}|tju s(t	|r4t
 fdd| jD S tjj|jv p> |S )z
    Do all uses of this op have torch.Tag.pointwise or return True for optional `is_pointwise_fn`

    Uses in views ops will follow the views uses
    call_functionFc                 3  s    | ]}t | V  qd S rx   )is_pointwise_use)rC   urF  r>   rE   r     r;  z#is_pointwise_use.<locals>.<genexpr>)r7  rk   targetr@   _ops
OpOverloadr   getitemr   r@  ro   usersTag	pointwisetags)rE  rF  rL  r>   rK  rE   rI  s  s   

rI  rL  r
   rr   	list[Any]kwargsdict[str, Any]&tuple[GraphModule, list[torch.Tensor]]c                   s   t j  g d
 fdd} j| gtt j|||fR  }t| jjdkr5t	| jjd j
d	kr5|f} | t ji  }|fS )Nargtorch.Tensorr<   r'   c                   s    |   dt S )NrX  )appendplaceholderrH   )rX  g
graph_argsr>   rE   add_tensor_arg  s   
z)gen_gm_and_inputs.<locals>.add_tensor_argr(   r   Tensor)rX  rY  r<   r'   )r@   fxGraphrH  r   r`  rH   r=  returnsr=   r   outputr%   )rL  rr   rU  r_  nodegmr>   r\  rE   gen_gm_and_inputs  s   

rg  r8   Nonec                 C  s,   | dkrd S t | }| r|  d S d S Nr   )rM   rA   r   r   r>   r>   rE   r     s   r   modelCallable[..., Any]example_inputsSequence[Any]timesc                 C  sT   t | td t }t|D ]
}| | }t | qt }|d us&J || S )Ni9  )r   r@   manual_seedtimeperf_counterr   )rj  rl  rn  r   t0r   resultt1r>   r>   rE   timed  s   

ru  r>   
         ?repeatbaselinec                   sH   t  fddt|D }t | }t|| d | S )Nc                   s   g | ]	}t  qS r>   )ru  )rC   r   r   rl  rj  rn  r>   rE   rF         z%print_performance.<locals>.<listcomp>z.6f)r@   r   r   medianprintitem)rj  rl  rn  rx  ry  r   timingstookr>   rz  rE   print_performance  s   r  objmethodc                   s$   t | |  t| | fdd dS )zKReplace obj.method() with a new method that returns a precomputed constant.c                     s    S rx   r>   r>   rs  r>   rE   rC    rD  z#precompute_method.<locals>.<lambda>N)r?   setattr)r  r  r>   r  rE   precompute_method  s   r  methods	list[str]c                 C  s   |D ]}t | | qdS )zFReplace methods with new methods that returns a precomputed constants.N)r  )r  r  r  r>   r>   rE   precompute_methods  s   r  r   r   c                 C  s   t | |kt | |k  S rx   )re   )r   r   r>   r>   rE   cmp     r  rD   Union[int, Sequence[int]]sizeSequence[int]c                 C  s:   t | tr
| g| S t| dkrt| | d g| S | S )Nr(   r   )rk   re   rH   r   )rD   r  r>   r>   rE   pad_listlike  s
   

r  tuple[_T, ...]list[_T]c                 C  s&   t | dkrg S d	dd}t| |dS )
Nr   elemr_   r<   r=   c                 S  s0   t | tr| S ddlm} t | |sJ |  S )Nr(   )r6   )rk   r=   	schedulerr6   get_name)r  r6   r>   r>   rE   	sort_func  s
   
ztuple_sorted.<locals>.sort_funcr   )r  r_   r<   r=   )rH   sorted)rD   r  r>   r>   rE   tuple_sorted  s   
	r  PRVT)	covariantc                   @  s$   e Zd ZedddZdddZdS )CachedMethodr   r
   r<   rh  c                 C     d S rx   r>   )r   r>   r>   rE   clear_cache     zCachedMethod.clear_cacherr   P.argsrU  P.kwargsr  c                 O  r  rx   r>   selfrr   rU  r>   r>   rE   __call__  rD  zCachedMethod.__call__N)r   r
   r<   rh  )rr   r  rU  r  r<   r  )r|   r}   r~   staticmethodr  r  r>   r>   r>   rE   r    s    r  !Callable[Concatenate[Any, P], RV]CachedMethod[P, RV]c                   sl   | j }d| d d| i}td| d  d  d | t| || d }d fdd}||_|S )N___cacher   z        def zC_cache_on_self(self):
            try:
                return self.zy
            except AttributeError:
                pass
            rv = fn(self)
            object.__setattr__(self, "z%", rv)
            return rv
        _cache_on_selfr  r
   r<   rh  c                   s   t |  rt|   d S d S rx   )r   delattrr  r  r>   rE   r    s   
z"cache_on_self.<locals>.clear_cache)r  r
   r<   rh  )r|   execlstripr   wrapsr  )r   r   ctxwrapperr  r>   r  rE   cache_on_self  s$   	r  node_schedule0Union[Sequence[BaseSchedulerNode], ExternKernel]OrderedSet[Node]c                 C  sJ   ddl m} t| trttjdd | D t S t| |j	r"| j
S t S )Nr(   irc                 S  s$   g | ]}t |d r|jr|jjqS )re  )r   re  origins)rC   re  r>   r>   rE   rF   '  s    z%aggregate_origins.<locals>.<listcomp>) r  rk   r"  r   r   r   or_r   r/   r  )r  r  r>   r>   rE   aggregate_origins  s   
	r  Sequence[BaseSchedulerNode]descriptive_names8Literal[True, 'torch', 'original_aten', 'inductor_node']c                 C  s   t | }|dkrdd |D }tt|}nH|dkrPg }|D ]*}|jdkrHd|jv rH|jd d }t|d tr@||d  q||d j qtt|}n|d	kr\d
d |D }nt	|}d
dg| S )Noriginal_atenc                 S  s<   g | ]}|j d krd|jv r|jd dur|jd jjqS )rH  r  N)r7  r   _overloadpacketr|   rC   originr>   r>   rE   rF   ;  s    

z)get_fused_kernel_name.<locals>.<listcomp>r@   rH  source_fn_stackr   r(   inductor_nodec                 S  s   g | ]
}|j d kr|jqS rH  )r7  r   r  r>   r>   rE   rF   O  s    r   fused)r  r  r   r7  r   rk   r=   rZ  r|   NotImplementedErrorjoin)r  r  all_originssourcesr  	source_fnr>   r>   rE   get_fused_kernel_name4  s.   r  r  r+   tuple[str, str]c                   s  t | }dd |D }tt}tt}d  t|rQtdd |D }t|dkrQ|d j t dsGi }t j	D ]\}}	|||	< q;| _
|j fdd	d
 |D ]3}
d|
jv rq|
jd d urqt|
jd j}|| |
j d|
jv r|
jd d j}|| |
j qS d urdnd}|j d| dd|  dd|  d}|j dg}t| D ]\}}||j d| ddt|  q d ur||j d |D ]}	||j d|	   q|d|fS )Nc                 S  s   g | ]	}|j d kr|qS r  r?  r  r>   r>   rE   rF   ]  r{  z'get_kernel_metadata.<locals>.<listcomp>c                 s  r   rx   )r1  )rC   nr>   r>   rE   r   g  r   z&get_kernel_metadata.<locals>.<genexpr>r(   r   )_inductor_kernel_metadata_node_to_idx_mapc                   s
    j |  S rx   )r  r  single_graphr>   rE   rC  q  s   
 z%get_kernel_metadata.<locals>.<lambda>r  r  	from_nodezTopologically SortedUnsorted z Source Nodes: [r   z], Original ATen: []z" Source node to ATen node mapping:z   z => z Graph fragment:
)r  collectionsdefaultdictr"  rH   r   r1  r   r   nodesr  sortr   r=   r  rZ  r   commentr  keysr  itemsformat_node)r  r  r  inductor_nodesfrom_node_dictoriginal_aten_dictunique_graphsnode_to_idx_mapidxr  re  r   sort_strmetadatadetailed_metadataoriginal_noder  r>   r  rE   get_kernel_metadataX  sP   






r  initial_queueIterable[torch.fx.Node]skip_filterOptional[Callable[[Any], bool]]OrderedSet[torch.fx.Node]c                 C  sZ   t | } t| }| r+|  }|jD ]}|r||rq||vr(|| | | q| s
|S )zJReturns the set of nodes whose values depend on those within initial_queue)r"  r   rI   rP  addrZ  )r  r  dominated_setre  userr>   r>   rE   dominated_nodes  s   


	r  Sequence[IRNode]dict[str, IRNode]OrderedSet[IRNode]c                   sd   dd l }ddlm  d fdd	fd
d| D }fdd| D }t|jg ||R  S )Nr   r(   r  r  r0   r<   rj   c                   sD   t |  jr| jS t |  jr| jS t |  jo!t |  jS rx   )rk   	TensorBoxdata
StorageBoxr0   	Pointwiser  r  is_unrealized_noder>   rE   r    s
   

z*gather_origins.<locals>.is_unrealized_nodec                      g | ]	} |r|j qS r>   r  )rC   valr  r>   rE   rF     r{  z"gather_origins.<locals>.<listcomp>c                   r  r>   r  )rC   rX  r   r>   rE   rF     r{  )r  r0   r<   rj   )	itertoolsr  r  r   r   chain)rr   rU  r  kwarg_originsarg_originsr>   r  rE   gather_origins  s   r  exprc                 C  s   t | tjr	| jS t | tjrdtt| jS t | tj	r'dtt| jS t | t
tttfrA| jj ddtt| j dS t| S )z
    Normal sympy str is very slow, this is a lot faster.  The result are
    somewhat worse, as it doesn't do as much simplification.  So don't
    use this for final codegen.
    z + z * (r   ))rk   rl   Symbolr   rm   r  rp   	sympy_strrr   MulrW   rT   rU   rV   funcr|   r=   r  r>   r>   rE   r
    s   "r
  r   ValueRanges[Any]c                 C  s>   ddl m} tjrt|jdd  }r|jdkrt| S t	 S )Nr(   r.  current_node
index_expr)
r5  r/  r\   compute_all_boundsr?   interpreterrL  rZ   r[   unknown)r   r/  fx_noder>   r>   rE   get_bounds_index_expr  s   
r  prefixc                 C  s   | d dkS )Nr   rr>   )r  r>   r>   rE   prefix_is_reduction     r  rY   r  sympy.Symbolc                 C  s   | t jksJ t| |dddS )9
    Used to generate an integer-nonnegative symbol.
    Tintegernonnegative)rY   SIZErX   )r  r  r>   r>   rE   sympy_index_symbol_with_prefix  s   r   checkc                 C  s   | st jot jS rx   )r\   debug_index_assertsassert_indirect_indexing)r!  r>   r>   rE   generate_assert     r$  r   c                 C  s    | d dksJ t j| dddS )r  r   sTr  )rl   r	  r   r>   r>   rE   sympy_index_symbol  s   r(  replacementsdict[sympy.Expr, Any]c                   s,   ddd t |  fd	d
| D S )z
    When the passed replacement symbol v is a string, it is converted to a symbol with name v that
    have the same replaced expression integer and nonnegative properties.
    replacedri   replacementUnion[sympy.Expr, str]r<   r  c                 S  s2   t | tjsJ t |trtj|| j| jdS |S )Nr  )rk   rl   r   r=   r	  r   is_nonnegative)r+  r,  r>   r>   rE   	to_symbol   s   
zsympy_subs.<locals>.to_symbolc                   s   i | ]
\}}| ||qS r>   r>   rC   krh   r/  r>   rE   r         zsympy_subs.<locals>.<dictcomp>N)r+  ri   r,  r-  r<   r  )rl   r   xreplacer  )r  r)  r>   r2  rE   
sympy_subs  s   

r5  ,TypeGuard[Union[torch.SymInt, torch.Tensor]]c                 C  s:   t | tjpt | tjotdd t|  |  D S )Nc                 s      | ]}t |V  qd S rx   is_symbolicrB   r>   r>   rE   r         zis_symbolic.<locals>.<genexpr>)	rk   r@   r#   r`  r<  r  r  r  stride)r   r>   r>   rE   r9    s    r9  c                  G     t dd | D S )Nc                 s  r7  rx   r8  r:  r>   r>   rE   r     r:  z"any_is_symbolic.<locals>.<genexpr>r<  )rr   r>   r>   rE   any_is_symbolic  r   r>  rf  torch.fx.GraphModuleOptional[torch.fx.Node]c                 C  sv   ddl m} tg d}t r|d | jjD ]}t|j	|v r&|  S |j
d }d ur8||r8|  S qd S )Nr   )free_unbacked_symbols)z,aten._fused_moving_avg_obs_fq_helper.defaultz7aten._fused_moving_avg_obs_fq_helper_functional.defaultzfbgemm.dense_to_jagged.defaultz%fbgemm.jagged_to_padded_dense.defaultrun_and_save_rng_staterun_with_rng_statezaten._local_scalar_densezaten._assert_scalar)zaten._unsafe_index_put.defaultz0aten._unsafe_masked_index_put_accumulate.defaultzaten.index_put.defaultzaten.index_put_.defaultzaten.scatter.srczaten.scatter.reducezaten.scatter.value_reducezaten.scatter_add_zaten.scatter_add.defaultzaten.scatter_reduce.twozaten.scatter_reduce_.twozaten.scatter_reduce.two_outr  )%torch.fx.experimental.symbolic_shapesrA  r   r@   $are_deterministic_algorithms_enabledupdater1  r  r=   rL  r   get)rf  rA  forbidden_setre  r  r>   r>   rE   %get_first_incompatible_cudagraph_node  s   rI  c                 C  s&   t tt| jj}|jdksJ |S )z$Get the output node from an FX graphrd  )nextiterreversedr1  r  r7  )rf  	last_noder>   r>   rE   output_nodeM  s   rN  _registered_cachesc                 C  s0   t | dr
t| jst|  dt|  | S )zq
    Use this decorator to register any caches that should be cache_clear'd
    with fresh_inductor_cache().
    cache_clearz# does not have a cache_clear method)r   callablerP  AttributeErrorrO  rZ  r  r>   r>   rE   clear_on_fresh_inductor_cacheW  s   
rT  c                  C  s   t D ]} |   qdS )z&
    Clear all registered caches.
    N)rO  rP  rS  r>   r>   rE   clear_inductor_cachesc  s   
rU  c                  C  s   t tj D ]9} | dsqtj|  }|j D ]"}|dr;t||}t|tj	j
jjr;|jD ]	}|jjj  q1qtj| = qdtjv rVtjd }t|jjj`|jj`t  d S )Nz&torch._inductor.runtime.compile_tasks.triton_ztriton.runtime.driver)r"  sysmodulesr  
startswith__dict__r?   rk   r@   	_inductorruntimetriton_heuristicsCachingAutotunercompile_resultskernelrunmod__del__r   driveractiveutilsinstancegccollect)module_namem	attr_namer`  rs  rb  r>   r>   rE   unload_xpu_triton_pydsn  s&   







rm  cache_entriesOptional[dict[str, Any]]dirOptional[str]deleteIterator[None]c              	   #  sP   t   tj|d zztjtjd iX t	d  tj
 dtjtjdi1 dV  t| trXt| dksAJ dtj
rXt}| fd	d
|D  W d   n1 sbw   Y  W d   n1 sqw   Y  |rt rtj rt  tj  fddd W n ty   td   w W t   dS t   w )z
    Contextmanager that provides a clean tmp cachedir for inductor.

    Optionally, pass a dict as 'cache_entries' to get a list of filenames and sizes
    generated with this cache instance.
    )rp  TORCHINDUCTOR_CACHE_DIRzUsing inductor cache dir %stritonTRITON_CACHE_DIRNr   z!expected empty cache_entries dictc              	     s,   i | ]}d |vr|t jt j |qS )z.lock)ospathgetsizer  )rC   f)triton_cache_dirr>   rE   r     s
    z(fresh_inductor_cache.<locals>.<dictcomp>c                   s   t jd |dS )Nz*Failed to remove temporary cache dir at %s)exc_info)r   warning)r  rx  r|  )inductor_cache_dirr>   rE   rC    s
    z&fresh_inductor_cache.<locals>.<lambda>)onerrorz(on error, temporary cache dir kept at %s)rU  tempfilemkdtempr   patchdictrw  environr   r   rx  r  rk   rH   existslistdirrF  
is_windowsr@   r:   rA   rm  shutilrmtree	Exceptionr}  )rn  rp  rr  filesr>   )r~  r{  rE   fresh_inductor_cache  sL   




r  seq	list[int]c                 C  s(   | j }tt| }ttt||ddS )NT)r   reverse)__getitem__r   rH   r"  rL  r  )r  gettera_rr>   r>   rE   argsort  s   r  r3  r&   .Sequence[Union[int, torch.SymInt, sympy.Expr]]c                   sD   d fdd}dd	 t |D }t|t|d
}dd	 |D }|S )Nr   tuple[int, sympy.Expr]r   r<   re   c                   sZ   | \}}|\}}d
 fdd}|||k rdS |||krdS ||k r%dS ||kr+dS d	S )Nr  %Union[bool, torch.SymInt, sympy.Expr]r<   rj   c                   s   t | tr| S  j| ddS )NT)size_oblivious)rk   rj   evaluate_exprr  r3  r>   rE   evaluate  s   
z*argsort_sym.<locals>.cmp.<locals>.evaluater   r(   r   )r  r  r<   rj   r>   )r   r   a_idxa_valb_idxb_valr  r  r>   rE   r    s   zargsort_sym.<locals>.cmpc                 S  s,   g | ]\}}|t |tjr|jjn|fqS r>   )rk   r@   r#   re  r  )rC   r  r&  r>   r>   rE   rF     s    zargsort_sym.<locals>.<listcomp>r  c                 S  s   g | ]\}}|qS r>   r>   )rC   r  r   r>   r>   rE   rF         )r   r  r   r  r<   re   )r   r  r   
cmp_to_key)r3  r  r  exprsrs  r>   r  rE   argsort_sym  s   r  r   torch.dtypec                 C  s    | t jkrdS t jd| d S )Nrc   r>   r   )r@   r  r   element_sizer  r>   r>   rE   get_dtype_size  s   
r  c                   @  s   e Zd ZU ded< dS )LineContextr
   contextNr|   r}   r~   __annotations__r>   r>   r>   rE   r    s   
 r  c                   @     e Zd ZU ded< ded< dS )ValueWithLineMapr=   rv   zlist[tuple[int, LineContext]]line_mapNr  r>   r>   r>   rE   r       
 r  c                   @  s   e Zd ZdZd<d=ddZd>d
dZd?ddZd?ddZd@ddZdAddZ	d?ddZ
d@ddZdBddZdCd d!ZdDdEd%d&ZdDdFd'd(ZdDdFd)d*Z	+dGdHd/d0ZdId3d4Zd?d5d6ZdJd9d:Zd;S )KIndentedBuffer   r   initial_indentre   r<   rh  c                 C  s   g | _ || _d S rx   )_lines_indent)r  r  r>   r>   rE   __init__      
zIndentedBuffer.__init__r  c                 C  s   t  }d}g }| jD ]:}t|tr| }|d u rq
nt|tr(|||jf q
|}t|ts1J || |d |d|	d 7 }q
t
| |S )Nr(   r  )r	   r  rk   DeferredLineBaser  rZ  r  r=   writecountr  getvalue)r  bufr   linemapliliner>   r>   rE   getvaluewithlinemap  s$   




z"IndentedBuffer.getvaluewithlinemapr=   c                 C  s
   |   jS rx   )r  rv   r  r>   r>   rE   r       
zIndentedBuffer.getvaluec                 C  s   t  }| jD ]8}t|tr| }|d u rqnt|trq|}t|ts%J |dr4||d d  q|| |d q| S )N\r   r  )	r	   r  rk   r  r  r=   endswithr  r  )r  r  r  r  r>   r>   rE   getrawvalue  s    




zIndentedBuffer.getrawvaluec                 C  s   | j   d S rx   )r  clearr  r>   r>   rE   r  /     zIndentedBuffer.clearrj   c                 C  
   t | jS rx   )rj   r  r  r>   r>   rE   __bool__2  r  zIndentedBuffer.__bool__c                 C  s   d| j | j  S )Nr  )r  tabwidthr  r>   r>   rE   r  5  r%  zIndentedBuffer.prefixc                 C  s   |  d d S )Nr  	writeliner  r>   r>   rE   newline8  r  zIndentedBuffer.newliner  )Union[LineContext, DeferredLineBase, str]c                 C  sr   t |tr| j| d S t |tr| j||   d S | r1| j|   |  d S | jd d S Nr  )rk   r  r  rZ  r  with_prefixr  stripr  r  r>   r>   rE   r  ;  s   

zIndentedBuffer.writelinelines3Sequence[Union[LineContext, DeferredLineBase, str]]c                 C  s   |D ]}|  | qd S rx   r  )r  r  r  r>   r>   rE   
writelinesE  s   zIndentedBuffer.writelinesr(   offset'contextlib.AbstractContextManager[None]c                   s   t jd fdd}| S )Nr<   rs  c                	   3  s<     j  7  _ zd V  W  j  8  _ d S  j  8  _ w rx   r  r>   r  r  r>   rE   r  L  
   "z"IndentedBuffer.indent.<locals>.ctxr<   rs  )
contextlibcontextmanager)r  r  r  r>   r  rE   indentK  s   zIndentedBuffer.indentc                 C  s   |  j |7  _ d S rx   r  r  r  r>   r>   rE   	do_indentV  r   zIndentedBuffer.do_indentc                 C  s   |  j |8  _ d S rx   r  r  r>   r>   rE   do_unindentY  r   zIndentedBuffer.do_unindentF
other_codeUnion[IndentedBuffer, str]r  c                 C  s   t |trJtd}|jD ]}t |ts"|r"t|t|t|  }qt	|r*d}|jD ]}t |tr;| j
| q-t| |t|d   q-d S t|}|rU| }|sYd S | }|dD ]}| | qbd S )Ninfr   r  )rk   r  r   r  r  minrH   r  mathisinfrZ  r  re   textwrapdedentrstripr!  )r  r  r  r  r  r&  r>   r>   rE   splice\  s,   





zIndentedBuffer.splicer  Callable[[Any], Any]c                   s&   t | jd} fdd| jD |_|S )Nr  c                   s   g | ]} |qS r>   r>   )rC   r  r  r>   rE   rF   w  r  z&IndentedBuffer.map.<locals>.<listcomp>)r  r  r  )r  r  r   r>   r  rE   rp   u  s   zIndentedBuffer.mapc                 C  s   t |  d|   dS )Nr  r  )r   r  r  r>   r>   rE   __repr__z  r  zIndentedBuffer.__repr__otherr   c                 C  s8   | j |j ksJ t| j d}|| j ||j |S )Nr  )r  r  r  r  )r  r  r   r>   r>   rE   __add__}  s
   zIndentedBuffer.__add__Nr   )r  re   r<   rh  )r<   r  r<   r=   r<   rh  r<   rj   )r  r  r<   rh  )r  r  r<   rh  ru   )r  re   r<   r  )r  re   r<   rh  )F)r  r  r  rj   r<   rh  )r  r  r<   r  )r  r   r<   r  )r|   r}   r~   r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  rp   r  r  r>   r>   r>   rE   r    s(    











r  c                      s(   e Zd Zd
 fddZddd	Z  ZS )FakeIndentedBufferr<   rh  c                   s   t    d S rx   )superr  r  	__class__r>   rE   r    r  zFakeIndentedBuffer.__init__r   r=   r
   c                 C  s$   |dkr
t | |S td| d)Nr  zTried to call self.z on FakeIndentedBuffer. This bufferis currently used on TritonTemplateKernel to prevent actualwrites to the body without explicitly specifying the body with`TritonTemplateKernel.set_subgraph_body(name)`)object__getattribute__r   )r  r   r>   r>   rE   r    s
   
z#FakeIndentedBuffer.__getattribute__r  )r   r=   r<   r
   )r|   r}   r~   r  r  __classcell__r>   r>   r  rE   r    s    r  c               	   c  s<    t jt j} }zd V  W | |t _t _d S | |t _t _w rx   )rW  stdoutstderr)initial_stdoutinitial_stderrr>   r>   rE   restore_stdout_stderr  r  r  c                   @  s`   e Zd ZdZdddZddd	ZdddZd ddZd!ddZd"ddZ	d#ddZ
d$ddZdS )%r  z.A line that can be 'unwritten' at a later timer  r=   c                 C  s   |  sd}|| _d S r  )r  r  r  r>   r>   rE   r    s   
zDeferredLineBase.__init__r<   Union[str, None]c                 C     t )zJReturns either self.line or None to indicate the line has been 'unwritten'r  r  r>   r>   rE   r       zDeferredLineBase.__call__r   c                 C  r  )z3Returns a new deferred line with the same conditionr   r  r>   r>   rE   	_new_line  r  zDeferredLineBase._new_liner  c                 C  s   |  | | j S rx   r  r  )r  r  r>   r>   rE   r    r   zDeferredLineBase.with_prefixc                 C  s   |  | j S rx   )r  r  r  r  r>   r>   rE   r    r%  zDeferredLineBase.lstripr   Union[int, slice]c                 C  s   |  | j| S rx   r  )r  r   r>   r>   rE   r    r%  zDeferredLineBase.__getitem__rj   c                 C  r  rx   )rj   r  r  r>   r>   rE   r    r  zDeferredLineBase.__bool__re   c                 C  r  rx   )rH   r  r  r>   r>   rE   __len__  r  zDeferredLineBase.__len__N)r  r=   )r<   r  )r  r=   r<   r   )r  r=   r<   r   )r<   r   )r   r  r<   r   r  r<   re   )r|   r}   r~   r   r  r  r  r  r  r  r  r  r>   r>   r>   rE   r    s    






r  c                      s6   e Zd ZdZd fddZdd
dZdddZ  ZS )DelayReplaceLinez6At end of codegen call `line.replace(key, value_fn())`r   r=   value_fnCallable[[], str]r  c                   s   t  | || _|| _d S rx   )r  r  r   r  )r  r   r  r  r  r>   rE   r    s   
zDelayReplaceLine.__init__r<   c                 C  s   | j | j|  S rx   )r  replacer   r  r  r>   r>   rE   r    r   zDelayReplaceLine.__call__c                 C  s   t | j| j|S rx   )r  r   r  r  r>   r>   rE   r    r%  zDelayReplaceLine._new_line)r   r=   r  r	  r  r=   r  )r  r=   r<   r  )r|   r}   r~   r   r  r  r  r  r>   r>   r  rE   r    s
    
r  index_or_deviceUnion[int, torch.device]c                 C  s   t | tjr	| }ntt | }t|}tjjr3|jd us J |jdk s*|jdkr1t	
d dS dS |jdkr:dnd}|j}||k rOt	j
d	||d
d dS dS )N	   rv  z6GPU arch does not support max_autotune_gemm mode usageFTr:   r`   D   z,Not enough SMs to use max_autotune_gemm mode)min_sms	avail_sms)extra)rk   r@   r   rL   r   createversionhipmajorr   r}  r   multi_processor_count)r  r   propr  r  r>   r>   rE   
is_big_gpu  s&   

r  c                   C  s   t jdjS )Nr8   )r@   r8   get_device_propertiesr  r>   r>   r>   rE   get_max_num_sms     r  c                  C  s"   t j } t | dur|  S d S )zFHandle experimental carveout if set otherwise return hardware SM countNr   )r@   r   _get_sm_carveout_experimentalr  )carveoutr>   r>   rE   get_num_sms  s   
r  num_tma_descriptorsr)   c                 C  s<   ddl m}m} |d}t |  t }||||| dS )zKBuilds and returns a WorkspaceArg for the device side TMA workspace buffer.r(   )r)   WorkspaceZeroModeF)r  	zero_moder   
outer_name)codegen.commonr)   r   	from_boolr  TMA_DESCRIPTOR_SIZEunique_name)r  r   r)   r   r!  r  r>   r>   rE   get_tma_workspace_arg  s   
r'  c                   C  s   t jpt jpt jS rx   )r\   max_autotunemax_autotune_gemmsearch_autotune_cacher>   r>   r>   rE   use_max_autotune  s   r+  layoutr1   allowed_layout_dtypeslist[torch.dtype]c                 C  s    t | jjo| j|v ot| jS rx   )is_gpur   r   r   r  )r,  r-  r>   r>   rE   _use_template_for_gpu  s
   r0  backendc                 C  "   |   dd tj  dD v S )Nc                 S     g | ]}|  qS r>   r  rB   r>   r>   rE   rF         z)_use_autotune_backend.<locals>.<listcomp>,)upperr\   max_autotune_gemm_backendsr!  r1  r>   r>   rE   _use_autotune_backend     r:  c                 C  r2  )Nc                 S  r3  r>   r4  rB   r>   r>   rE   rF     r5  z._use_conv_autotune_backend.<locals>.<listcomp>r6  )r7  r\   max_autotune_conv_backendsr!  r9  r>   r>   rE   _use_conv_autotune_backend  r;  r=  F)enable_int32enable_float8r>  r?  c                C  s   ddl m}m} tjtjtjg}|rtjtjtjtjg}|r'|tj	tj
g t| jjo1t| |p<| jjdko<| j|v oJt oJtdoJ|| j|jS )Nr(   )BackendFeaturehas_backend_featurer   TRITON)r#  r@  rA  r@   r	  r  r  r  extendr  r  r/  r   r   r0  r   r+  r:  TRITON_TEMPLATES)r,  r>  r?  r@  rA  layout_dtypesr>   r>   rE   use_triton_template#  s"   	rF  matricesr0   c                    sJ   ddl m} ddlm  d fd	d
tjjo$| o$tfdd| D S )Nr   )has_triton_tma_devicer(   r.  rD   r0   r<   rj   c                   s   t |  dkr
dS |  }|tjtjfvrdS |  }| }| s(|s(dS |j	d }|r4|j	d }||j
 } jj|tS )N   Fr(   r   )rH   get_size	get_dtyper@   r	  r  
get_layoutis_transposedis_contiguousr  itemsizer1  r2  statically_known_multiple_ofTMA_ALIGNMENT)rD   r   r,  
transposed	inner_diminner_bytesr.  r>   rE   _is_tma_compatible@  s   


z3use_triton_tma_template.<locals>._is_tma_compatiblec                 3      | ]} |V  qd S rx   r>   )rC   rk  )rU  r>   rE   r   V  r:  z*use_triton_tma_template.<locals>.<genexpr>rD   r0   r<   rj   )torch.utils._tritonrH  r5  r/  r\   ru  enable_persistent_tma_matmulro   )rG  rH  r>   )r/  rU  rE   use_triton_tma_template;  s   rZ  rk  r  r1  c           	      C  s   ddl m} |jjj|| | dd}|dks|tjjk rdS ddlm	} t
jjr+dS t
jt
jt
jt
jg}t| |o@t o@td}|rM| sMtd	 dS |S )
Nr(   r.  r   fallbackr   F)try_import_cutlassCUTLASSzFailed to import CUTLASS lib. Please check whether _inductor.config.cuda.cutlass_dir is set correctly. Skipping CUTLASS backend for now.)r5  r/  r1  r2  	size_hintr\   r8   cutlass_backend_min_gemm_sizecodegen.cuda.cutlass_utilsr]  r@   r  r  r	  r  r  r  r0  r+  r:  r   r}  )	r,  rk  r  r1  r/  	gemm_sizer]  rE  r   r>   r>   rE   use_cutlass_templateZ  s(   
rc  c                 C  s   t j| jS rx   )r@   r8   r  gcnArchNamer   r>   r>   rE   _rocm_native_device_arch_namex  r  rf  Qtuple[Optional[str], Callable[[], list[Any]], Callable[[], list[Any]], type[Any]]c                  C  s|   zdd l } ddlm}m} ddlm} tj| j	}W n t
y7   ddd}ddd	}G d
d d}d }Y nw ||||fS )Nr   )gen_ops_librarygen_ops_preselected)CKGemmOperationr<   rT  c                   S     g S rx   r>   r>   r>   r>   rE   rh    r  z*try_import_ck_lib.<locals>.gen_ops_libraryc                   S  rk  rx   r>   r>   r>   r>   rE   ri    r  z.try_import_ck_lib.<locals>.gen_ops_preselectedc                   @  s   e Zd ZdS )z*try_import_ck_lib.<locals>.CKGemmOperationN)r|   r}   r~   r>   r>   r>   rE   rj    s    rj  )r<   rT  )ck4inductor(ck4inductor.universal_gemm.gen_instancesrh  ri  ck4inductor.universal_gemm.oprj  rw  rx  dirname__file__r   )rl  rh  ri  rj  package_dirnamer>   r>   rE   try_import_ck_lib}  s   

rr  c                   s   t  sdS tjjsdS | jjdksdS t| j}dd tjj	D p)|
dd |i  fdd  tjj@ D }|s=dS | jtjtjtjfvrJdS t \}}}}|sZtd	 dS t rb|tj_tjjsmtd
 dS |tjjkrztd dS dS )NFr8   c                 S  s   i | ]
}| d d |qS ):r   )r!  rC   r1  r>   r>   rE   r     r3  z#use_ck_template.<locals>.<dictcomp>rs  r   c                   s   g | ]} | qS r>   r>   rt  requested_archsr>   rE   rF     s    z#use_ck_template.<locals>.<listcomp>z,Please pip install Composable Kernel packagez,Please set TORCHINDUCTOR_CK_DIR env variablezInvalid path to CK libraryT)r+  r@   r  r  r   r   rf  r\   rocmarchr!  r  ck_supported_archr   r	  r  r  rr  r   r}  	is_fbcodeck_dir)r,  native_archrequested_supported_archsck_package_dirnamer   r>   ru  rE   use_ck_template  s<   




r  c                 C  s:   ddl m} tdot| o|jjj|| | dddkS )Nr(   r.  CKr   r[  r   )r5  r/  r:  r  r1  r2  r_  )r,  rk  r  r1  r/  r>   r>   rE   use_ck_gemm_template  s   r  c                 C  s   t dot| S )Nr  )r=  r  r,  r>   r>   rE   use_ck_conv_template  r%  r  c                 C  s   t  o| jjdkS ri  )r+  r   r   r  r>   r>   rE   _use_template_for_cpu  r   r  mat1Union[ReinterpretView, Buffer]mat2c                 C  s6   ddl m} t|j|sJ t| ||ddo|j S )Nr(   )r1   F)require_constant_mat2)r  r1   rk   r,  use_cpp_gemm_templaterN  )r,  r  r  r1   r>   r>   rE   use_cpp_bmm_template  s
   r  mat2_transposedr  is_woq_int4q_group_sizeOptional[int]c                 C  s:  ddl m} ddlm} ddlm}	 ddlm}
 t| r t	ds"dS t
jjs(dS | tjtjfv }tjtjtjtjg}|
|||rD| jnd ||d\}}}} }}t||frXdS t||jrb| }|	| \}}|d	|||| | |t | |d

}ddd}| j|v o|d uo||ot||jo| p| S )Nr(   r  )create_micro_gemm)*get_gemm_template_output_and_compute_dtype)mm_argsCPPF)	out_dtyper  use_4x2_dim
micro_gemm)input_dtypeinput2_dtypeoutput_dtypenum_threadsuse_refr  rD   r0   r<   rj   c                 S  s   |    |  d dkS )Nr   r(   )freeze_layout
get_striderD   r>   r>   rE   is_last_dim_stride1  s   z2use_cpp_gemm_template.<locals>.is_last_dim_stride1rW  )r  r  codegen.cpp_micro_gemmr  codegen.cpp_utilsr  kernel.mm_commonr  r  r:  r\   cppweight_prepackrK  r@   r  r  r  r  halfr   has_free_symbolsrk   BaseViewunwrap_viewparallel_num_threadsr  is_module_buffer)r,  r  r  r  r  r  r  r  r  r  r  	int8_gemmrE  rk  r  r1  r  r   r  r  r>   r>   rE   r    sX   		


r  c                   C  s   t   ptdS )NATEN)r+  r:  r>   r>   r>   rE   use_aten_gemm_kernels'  r%  r  c                   @  s>   e Zd ZU edZded< dddZddd	ZdddZ	dS )DebugDirManagerr   r=   prev_debug_namer<   rh  c                 C  s   t tj| _d S rx   )rJ  r  counterr   r  r>   r>   rE   r  /  r%  zDebugDirManager.__init__c                 C  s0   t jjj| _| j d| j | _| jt jj_d S )N_tmp_)r@   _dynamor\   debug_dir_rootr  r   new_namer  r>   r>   rE   	__enter__2  s   zDebugDirManager.__enter__rr   r
   c                 G  s   t | j | jtjj_d S rx   )r  r  r  r  r@   r  r\   r  )r  rr   r>   r>   rE   __exit__7  s   zDebugDirManager.__exit__Nr  )rr   r
   r<   rh  )
r|   r}   r~   r  r  r  r  r  r  r  r>   r>   r>   rE   r  +  s   
 


r  Callable[P, _T]r  r  tuple[_T, list[str]]c                   st   ddl m} g  d
 fdd}tj|d	| tj  | |i |}W d    | fS 1 s1w   Y  | fS )Nr(   r,   coder=   r<   rh  c                        |  d S rx   rZ  r  source_codesr>   rE   save_output_codeE  r  z*run_and_get_code.<locals>.save_output_coder  r  r=   r<   rh  r1  r-   r   r  r  r@   r  reset)r   rr   rU  r-   r  rs  r>   r  rE   run_and_get_code<  s   

r  tuple[Any, list[str]]c                 O  sF   t | g|R i |\}}g }|D ]}|td|tj q||fS )Nz	'''.*?''')r  rC  refindallDOTALL)r   rr   rU  rs  r  kernelsr  r>   r>   rE   run_and_get_kernelsN  s
   r  c                   s   d fdd}t |S )Nr<   r
   c                    s     } |     | S rx   )r   backwardr  r   r>   rE   run_with_backwardY  s   z1run_fw_bw_and_get_code.<locals>.run_with_backward)r<   r
   )r  )r   r  r>   r  rE   run_fw_bw_and_get_codeX  s   r  c              	     s   ddl m} g dfdd d fdd}tj|d|5 tj|d  tj  | |i |}W d   n1 s>w   Y  W d   S W d   S 1 sVw   Y  S )zLGet the inductor-generated code, but skip any actual compilation or running.r(   r,   r  r=   r<   rh  c                   r  rx   r  r  r  r>   rE   r  g  r  z"get_code.<locals>.save_output_coder  r-   r
   c                   sF   G dd d}| j r|  n|  \}} |j |r  |j | S )Nc                   @  s$   e Zd ZdZdddZdd	d
ZdS )z@get_code.<locals>.patched_compile_to_module.<locals>.DummyModulez4This is empty to replace the generated triton moduler<   rh  c                 S  r  rx   r>   r  r>   r>   rE   r  n  r  zIget_code.<locals>.patched_compile_to_module.<locals>.DummyModule.__init__rr   r
   rU  c                 _  r  rx   r>   r  r>   r>   rE   callq  r  zEget_code.<locals>.patched_compile_to_module.<locals>.DummyModule.callNr  rr   r
   rU  r
   r<   rh  )r|   r}   r~   r   r  r  r>   r>   r>   rE   DummyModulek  s    
r  )cpp_wrappercodegen_with_cpp_wrappercodegenrv   )r  r  wrapper_codekernel_code)r  r>   rE   patched_compile_to_modulej  s   

z+get_code.<locals>.patched_compile_to_modulecompile_to_moduler  Nr  )r  r-   r<   r
   r  )r   rr   rU  r-   r  r   r>   )r  r  rE   get_codea  s$   
(


r  c                 O  sJ   t | g|R i |}dt|  krdks!n J dt| |d S Nr(   rI  z%expected one or two code outputs got r   )r  rH   )r   rr   rU  r  r>   r>   rE   get_triton_code  s
   r  c                 O  sN   t | g|R i |\}}dt|  krdks#n J dt| |d S r  )r  rH   )r   rr   rU  r   r  r>   r>   rE   run_and_get_triton_code  s
   r  tuple[Any, list[GraphLowering]]c                   s   ddl m  ddlm} |jg d fd	d
}tj|d| | |i |}W d    |fS 1 s7w   Y  |fS )Nr   r,   r4   rr   r
   rU  r<   rh  c                    s2   | i | | d }t | sJ | d S )NrI  )rk   rZ  )rr   rU  r1  r-   graph_lowerings	real_initr>   rE   	fake_init  s   z-run_and_get_graph_lowering.<locals>.fake_initr  r  )torch._inductor.graphr-   torch._inductor.output_coder5   r  r   r  r  )r   rr   rU  r5   r  rs  r>   r  rE   run_and_get_graph_lowering  s   
r  aten_opoverride_fnc              	   c  sN    ddl m} |j|  }zt|||j| < dV  W ||j| < dS ||j| < w )z
    Override the lowering of aten_op with override_fn.
    The first argument of override_fn is the original lowering fn.
    r   )loweringN)torch._inductorr  	loweringsr   partial)r  r  r  orig_fnr>   r>   rE   override_lowering  s   
r  pre_fnpost_fnOptional[Callable[..., Any]]c                   s6   ddl m} |j d fdd}tjj|d	|S )zr
    Add hook functions to be called at the beginning and end of Scheduler.__init__.
    Used for unit tests.
    r   )	Schedulerr  r
   r  r<   c                   s&   | |  | |}r| | |S rx   r>   )r  r  outr  r  r  r>   rE   r    s
   


z(add_scheduler_init_hook.<locals>.wrapperr  N)r  r
   r  r
   r<   r
   )torch._inductor.schedulerr  r  unittestr   r  r  )r  r  r  r  r>   r  rE   add_scheduler_init_hook  s   r  msgc                 C  s"   t jr
t|  dS t|  dS )z
    Warnings that will be actionable for PyTorch developers, but not
    end users.  Allows us to easily disable them in stable releases but
    keep them on for nightly builds.
    N)r\   developer_warningsr   r}  info)r  r>   r>   rE   developer_warning  s   r  c                  C  s   z/t jd} | d tt jk r.tt j| d  dkr.t j| d  d dkr.t j| d  W S W n	 ty8   Y nw t jD ]}|drM|tdd   S q<dS )a  
    An experimental API used only when config.benchmark_kernel is true.

    The benchmark name is only available at codegen time. So we can not
    directly call it in benchmark_all_kernels which is run after codegen.

    The function assumes the argument after --only is the benchmark name.
    It works for torchbench.py/hugginface.py/timm_models.py. But for ad-hoc
    scripts, this function may return None.

    There are 2 flavors of --only argument we need handle:
    1. --only model_name
    2. --only=model_name
    z--onlyr(   r   -z--only=N)rW  argvr   rH   
ValueErrorrY  )r  rX  r>   r>   rE   get_benchmark_name  s   

r  r  c                 C  r<  )Nc                 s      | ]}|d kV  qdS r(   Nr>   rB   r>   r>   rE   r   	  r:  zis_ones.<locals>.<genexpr>ro   r  r>   r>   rE   is_ones  r   r   c                 C  r<  )Nc                 s  r  )r   Nr>   rB   r>   r>   rE   r     r:  zis_zeros.<locals>.<genexpr>r  r  r>   r>   rE   is_zeros  r   r  inputsSequence[torch.Tensor]c                 C  r<  )Nc                 s  s,    | ]}t |tjr|jtd kV  qdS )r   N)rk   r@   r`  r   )rC   r~  r>   r>   rE   r     s    

z is_cpu_device.<locals>.<genexpr>r  )r  r>   r>   rE   is_cpu_device  s   r  r  c                 C  s&   t | tjs
J d| jrtjS tjS )Nz8only support sympy.Expr as input to get_sympy_Expr_dtype)rk   rl   r   r   r@   r  r  )r  r>   r>   rE   get_sympy_Expr_dtype  s   r  should_profileIterator[Any]c                 o  sN    | r"t jj|i |}|V  W d    d S 1 sw   Y  d S d V  d S rx   )r@   r   r   )r  rr   rU  r   r>   r>   rE   maybe_profile"  s   "
r  c                  C  s   t jj} | dk rt } | S Nr(   )r\   r  threadsr@   get_num_threads)r
  r>   r>   rE   r  +  s   r  c                  C  s,   ddl m}  |  }|dtjjrdS dS )Nr(   )get_backend_options
num_stagesrI     )runtime.triton_helpersr  rG  r@   r  r  )r  optionsr>   r>   rE   get_backend_num_stages2  s   r  c                 C  s   ddl m}m} | tjtjtjfv sJ t|j	
drEddlm} | }| tjtjfv r3|| |S tjjjjr?|tj|S |tj|S | tjtjfv rQ|| S tjjjjr\|tjS |tjS )Nr   )get_max_simd_tflopsget_max_tensorcore_tflops
clock_rate)max_clock_rate)triton.testingr  r  r@   r	  r  r  inspect	signature
parametersrG  torch._utils_internalr  backendsr8   matmul
allow_tf32)r   r  r  r  sm_clockr>   r>   rE   get_device_tflops:  s   


r  c                  C  s   ddl m}  |  S )Nr   get_dram_gbps)r  r!  r   r>   r>   rE   get_gpu_dram_gbpsV  s   r"  c                  C  s"   ddl m}  | jjdddS )Nr   rd  max_shared_mem)triton.runtimerd  re  rf  r  rG  r#  r>   r>   rE   get_gpu_shared_memory]  s   r&  reduction_typec                 C  s
   |  dS )Nwelford)rY  r'  r>   r>   rE   is_welford_reductionc  r  r*  c                 C  s   t | rdS | dkrdS dS )Nr  online_softmax_reducerI  r(   )r*  r)  r>   r>   rE   reduction_num_outputsg  s
   r,  c                   C  s   t  dkS )NLinux)platformsystemr>   r>   r>   rE   is_linuxp  r  r0  c                   C  s
   t jdkS )Nr^   )rW  r.  r>   r>   r>   rE   r  t  r  r  itrIterable[Any]c                 C  r<  )Nc                 s  s$    | ]}t |tjo|j V  qd S rx   )rk   rl   r   	is_numberrB   r>   r>   rE   r   y  s   " z#has_free_symbols.<locals>.<genexpr>r=  )r1  r>   r>   rE   r  x  r   r  c                  G  s~   ddl m} | D ]4}t||j|j|j|j|jfr-t|	 pds)t|
 p'dr, dS qt||js4qtdt| dS )Nr(   r  r>   Tzunexpected type for is_dynamic F)r  r  rk   r  r  r  ComputedBufferr.   r  maybe_get_sizemaybe_get_strider0   	TypeErrorr   )rr   r  tr>   r>   rE   
is_dynamic|  s   
r9  c                   @  s   e Zd ZdZdZdS )PlaceholderKERNEL_NAMEDESCRIPTIVE_NAMEN)r|   r}   r~   r;  r<  r>   r>   r>   rE   r:    s    r:  r  r%   inpc              	   C  s4  ddl m} tjdddd}t }t }t|t|dj|  t	d|j
 |d	 t	|j
|d	 t }t|| | |j
 W d    n1 sLw   Y  t | }	||j
 |j
  |  t	d
|j
 |d	 t	|j
|d	 | | k}
td||j|
|	 W d    d S 1 sw   Y  d S )Nr(   )stable_topological_sortwzutf-8F)modeencodingrr  )rf  	fake_modezBefore:
)filezAfter:
zZ%s, save before/after graph to %s, graph before/after are the same = %s, time elapsed = %s)pattern_matcherr>  r  NamedTemporaryFileior	   rR   rN   	propagater}  r1  r   nowrQ   lint	recompiler  r   r  r   )r  rf  r=  r  r>  rz  	before_ioafter_io
start_timetime_elapsedr8  r>   r>   rE   pass_execution_and_save  s>   

"rO  	input_buf"Optional[Union[Buffer, Operation]]c                 C  s&   ddl m} t| |jot| j|jS )zB
    Check if input buffer is a multi-outputs template buffer
    r(   r  )r  r  rk   CppTemplateBufferr,  MultiOutputLayoutrP  r  r>   r>   rE   is_multi_outputs_template  s   rU  c                 C  s4   ddl m} t| |jot| jdkot| jd S )zL
    Check if input buffer is a output of multi-outputs template buffer
    r(   r  r   )r  r  rk   MultiOutputrH   r  rU  rT  r>   r>   rE   #is_output_of_multi_outputs_template  s   rW  re   Optional[Union[Node, Operation]]!Optional[torch._ops.OperatorBase]c                 C  s   | d u rdS ddl m} t| |jko|d u p| j|u pRt| |jkoRttjj	do2| jtjj	j
jkpRttjj	doB| jtjj	jjkpRttjj	doR| jtjj	jjkS )NFr(   r  all_to_all_singleall_gather_into_tensorreduce_scatter_tensor)r  r  r   _CollectiveKernelop_overloadFallbackKernelr   r@   r   torchrecrZ  defaultr[  r\  re  r7  r  r>   r>   rE   is_collective  s"   

rc  "Optional[Union[IRNode, Operation]]c                 C  s   ddl m} t| |jkS Nr(   r  )r  r  r   _WaitKernel)re  r  r>   r>   rE   is_wait  s   rg  snoder6   c                 C  4   ddl m} t| |rtdd | jD S t| jS )Nr   GroupedSchedulerNodec                 s  r7  rx   )contains_collectiverB   r>   r>   rE   r     r:  z&contains_collective.<locals>.<genexpr>)r  rk  rk   r<  snodesrc  re  rh  rk  r>   r>   rE   rl       

rl  c                 C  ri  )Nr   rj  c                 s  r7  rx   )contains_waitrB   r>   r>   rE   r     r:  z contains_wait.<locals>.<genexpr>)r  rk  rk   r<  rm  rg  re  rn  r>   r>   rE   rp    ro  rp  Optional[Operation]?Union[torch._ops.OpOverload, Collection[torch._ops.OpOverload]]c                 C  s6   ddl m} t|tjjr|g}t| |jo| j|v S re  )r  r  rk   r@   rM  rN  r_  r^  rb  r>   r>   rE   is_fallback_op  s   rs  buf_namename_to_bufname_to_fused_nodec                 C  s   |||  j   S rx   )defining_opr  )rt  ru  rv  r>   r>   rE   buf_name_to_fused_snode  s   rx  c                 C  rA  rB  r>   rh  r>   r>   rE   rC  *  rD  collected_node_setMutableSet[BaseSchedulerNode]dict[str, SchedulerBuffer]dict[str, BaseSchedulerNode]criteria_cbCallable[[Any], bool]c                 C  sP   || rd S | |  | jD ]}t|j||}||v rqt|||||d qd S )Nr~  )r  unmet_dependenciesrx  r   find_recursive_deps_of_node)rh  rz  ru  rv  r~  depdefining_op_for_depr>   r>   rE   r  %  s"   

r  c                 C  rA  rB  r>   ry  r>   r>   rE   rC  C  rD  c              	   C  s   || rd S | |  |  D ]4}|jD ].}|jd usJ |j dkr%q|j |vr-q||j  }||v r9qt|||||d qqd S )NOUTPUTr  )r  get_outputsrP  re  r  find_recursive_users_of_node)rh  rz  ru  rv  r~  or  user_opr>   r>   rE   r  >  s,   

r  dynamo_gm_num_inputsaot_fw_gm_num_inputsc                 C  s   t jjjrdnd}||  | S )zaComputes the number of inputs to the aot fw graph which have fixed addresses (params and buffers)rI  r   )r@   
_functorchr\   functionalize_rng_ops)r  r  num_rng_seed_offset_inputsr>   r>   rE   num_fw_fixed_arguments[  s   r  fx_gc                 C  sd   ddd}d}g }| j jD ]}|jdkr!||r|| |d	7 }q|ttt|ks.J t|S )z>
    Infers which inputs are static for a backwards graph
    rD   r'   r<   rj   c                 S  s(   d| j vod| j vod| j vod| j vS )Ntangentsbwd_seedbwd_base_offsetbwd_rng_stater'  r  r>   r>   rE   is_saved_tensork  s   
z'count_tangents.<locals>.is_saved_tensorr   r[  r(   N)rD   r'   r<   rj   )r1  r  r7  rZ  r"  r   rH   )r  r  	arg_countstatic_arg_idxsr  r>   r>   rE   count_tangentsf  s   


r  c                   @  s.   e Zd ZU ded< dddZedd	d
ZdS )	BoxedBoolrj   rv   r<   c                 C  s   | j S rx   )rv   r  r>   r>   rE   r    s   zBoxedBool.__bool__r  r
   Union[BoxedBool, bool]c                 C  s   t | tr
d| _| S dS rB  )rk   r  rv   rS  r>   r>   rE   disable  s   
zBoxedBool.disableNr  )r  r
   r<   r  )r|   r}   r~   r  r  r  r  r>   r>   r>   rE   r    s
   
 
r  kernel_listc                 #  sh    ddl m} |j	 		 dd fdd}tj|d| d V  W d    d S 1 s-w   Y  d S )Nr(   r*   Tr  r+   kernel_namer=   r  r  rq  gpurj   cpp_definitionr<   r
   c                   s     | | |||||S rx   r  )r  r  r  r  r  r  r  orig_define_kernelr>   rE   define_kernel  s   
z.collect_defined_kernels.<locals>.define_kernelr  )NTN)r  r+   r  r=   r  r=   r  rq  r  rj   r  rq  r<   r
   )codegen.wrapperr+   r  r   r  r  )r  r+   r  r>   r  rE   collect_defined_kernels  s   "r  c                 C  s   | d S )N__original__r>   r'  r>   r>   rE    get_cloned_parameter_buffer_name     r  c                 C  s   | t v S rx   )rG   re  r>   r>   rE   r/    r  r/  c                 C  s   t | S rx   )r/  re  r>   r>   rE   device_need_guard  r  r  c                 C  sF   t  r| tjkrtj rtj dkrdS | ttjtj	tjgv S )N)r  r   F)
r\   rz  r@   r  r8   rA   get_device_capabilityr   r  rj   r  r>   r>   rE   ,needs_fallback_due_to_atomic_add_limitations  s   
r  r^  
self_dtype	src_dtypesrc_device_typesrc_is_tensorc                 C  s   | j tjjjtjjjfv r|d u rdS | j tjjjkrdnd}|d |fvp]|o.t|o.t|p]| j tjjjkoM|dkoM|oM|dkoMt	j
joMt	j
jpMt dkp]||koY|tjtjfv p]t S )NFr  r   r   r(   )overloadpacketr@   r   atenscatter_reduce_scatter_reducescatter_r/  r  r\   r  fallback_scatter_reduce_sumdynamic_threadsr  rj   r  rE  )r^  r'  r  r  r  r  	reduce_tyr>   r>   rE   use_scatter_fallback  s8   	r  c                 C  s  ddl m}m} ddlm} tdt|  d t| D ]m\}}td|dd ||u r2td	 q||u r;td
 qt||r|	 }t|rIdnd d |rb|j
dusXJ td|j
jj  td |jjD ]}t| qjtd |jjD ]}t| qyqtdt| dS )z
    An API that can be used in pdb to dump a node_schedule.
    Right mainly dump the read/write dependencies but can add more as needed.
    r   DisableReductionEnableReduction)SchedulerNodezNode schedule with z nodesr  3rs  zenable reductionzdisable reductionredpwz scheduler nodeNzoriginal reduction hint zReadDep:z	WriteDep:zUnrecognized node type: )torch._inductor.codegen.simdr  r  r  r  r}  rH   r   rk   is_reductionre  r  reduction_hintread_writesreadswritesr   r   )r  r  r  r  r  re  is_redr  r>   r>   rE   dump_node_schedule  s0   




r  r   rY  c                 C  s*   ddl m} ||  t| j t dkS )Nr   )statically_known_true)rD  r  storage_offsetr  r   GPU_ALIGN_BYTES)r   r  r>   r>   rE   tensor_is_aligned	  s   r  example_inputc                 C  s   t | jjsdS tjpt| S rB  )r/  r   r   r\   assume_aligned_inputsr  )r  r>   r>   rE   should_assume_input_aligned	  s   r  r  c                  C  s4   t jj } | st S | jj}|st S | S rx   )	r@   _guardsTracingContexttry_getr  nullcontextrB  r3  suppress_guards)tracing_contextr3  r>   r>   rE   #maybe_get_suppress_shape_guards_ctx	  s   r  tuple[Any, str]c                 O  s   t jjtddJ tj  dd l}dd l	}|
 }||}ddlm} || |j}||j | |i |}	| }
|| || W d    |	|
fS 1 sVw   Y  |	|
fS )Nr   Tr   )output_code_log)r  r   r  r  r\   r@   r  r  rF  loggingr	   StreamHandlertorch._inductor.codecacher  
addHandlerlevelsetLevelDEBUGr  removeHandler)r   rr   rU  rF  r  log_capture_stringchr  
prev_levelrs  r&  r>   r>   rE   run_and_get_cpp_code.	  s$   




r  Sequence[InputType]Optional[ShapeEnv]c                 C  s<   t | }|d ur|jS | D ]}t|tjr|jj  S qd S rx   )rN   r3  rk   r@   r#   re  )r  rB  inputr>   r>   rE   shape_env_from_inputsG	  s   r   Callable[[list[InputType]], Any]inputs_to_checkc                   s$   t  dkrS d fdd}|S )	Nr   
new_inputslist[InputType]r<   r
   c                   s   t |   | S rx   )copy_misaligned_inputs)r  r  rj  r>   rE   ra  b	  s   
z)align_inputs_from_check_idxs.<locals>.run)r  r  r<   r
   )rH   )rj  r  ra  r>   r  rE   align_inputs_from_check_idxs[	  s   r  c                 C  s`   d|   v r	d}ntdd t|   |  D d }t| |fd }t||   |  S )Nr   c                 s  s     | ]\}}|d  | V  qdS r  r>   )rC   shaper;  r>   r>   rE   r   o	  s    z)clone_preserve_strides.<locals>.<genexpr>r(   ru   )r  r   r   r;  r@   
as_stridedclone)rD   needed_sizebufferr>   r>   rE   clone_preserve_stridesi	  s   "r  r  r  check_inputs_idxsc                 C  s>   |D ]}| | }t |tjsJ | t rt|| |< qd S rx   )rk   r@   r`  data_ptr	ALIGNMENTr  )r  r  r   _inpr>   r>   rE   r  u	  s   r  static_input_idxsc                 C  sT   g }|D ]}| | }t |tjr| t dkr|| qt|t|kr(|S |S )z[
    We require all inputs to be aligned, so introduce a copy for any
    that aren't.
    r   )rk   r@   r`  r  r  rZ  rH   )r  r  aligned_static_input_idxsr  r  r>   r>   rE   remove_unaligned_input_idxs	  s   
r  r   c                 C  sZ   ddl m} ttjj}|jjj}|jjj	j
}|jj| |kr#dS || o,|| |kS )Nr(   r.  T)r5  r/  r@   iinfor  r   r1  r2  r_  r3  has_hintis_expr_static_and_true)r   r/  int_maxr_  r  r>   r>   rE   expr_fits_within_32bit	  s   
r  compiled_graphr5   c                   s   t jj }|d urX|jd urZt|jdksJ t| |jd us#J |jD ]5}|d u r3|jd  q&d t jj  }r@|j d fdd|jt	fd	d
|D  q&d S d S d S )Nr   Fr   r
   r<   ,Union[float, int, SymInt, SymFloat, SymBool]c                   s(   d u rt | S  r| S | S rx   )re   deserialize_symexprevaluate_symexpr)r   )fakify_first_callr3  r>   rE   map_expr	  s
   

z4set_tracing_context_output_strides.<locals>.map_exprc                 3  rV  rx   r>   )rC   r   )r  r>   rE   r   	  r:  z5set_tracing_context_output_strides.<locals>.<genexpr>)r   r
   r<   r  )
r@   r  r  r  output_stridesrH   r  rZ  r  tuple)rl  r   r  r  r  r>   )r  r  r3  rE   "set_tracing_context_output_strides	  s"   
r  c                  C  s`   t jd urt jS t  sdS tj rdS zddlm}  W n
 ty'   Y dS w | tj	dkS )NFr   REMOTE_CACHE_VERSIONz.pytorch/remote_cache:fx_graph_memcache_version)
r\   fx_graph_remote_cacherz  r@   _utils_internalis_fb_unit_testtorch._inductor.fb.remote_cacher
  ModuleNotFoundErrorjustknobs_getval_intr	  r>   r>   rE    should_use_remote_fx_graph_cache	  s   

r  c                 C  s   t dd| S )Nz[^a-zA-Z0-9_]r   )r  subr'  r>   r>   rE   normalize_name	  r  r  ztl.int1ztl.float8e4nvztl.float8e5ztl.float8e4b8ztl.float8e5b16ztl.uint8)ztl.boolztl.float8_e4m3fnztl.float8_e5m2ztl.float8_e4m3fnuzztl.float8_e5m2fnuzztl.float8_e8m0fnuc                 C  s   i | ]\}}||qS r>   r>   r0  r>   r>   rE   r   	  r   r   z^.*[.]c                 C  s   t dt| }t||S )z"Convert torch.dtype to triton typetl.)_triton_type_rer  r=   _triton_type_mappingrG  )r   triton_type_namer>   r>   rE   triton_type	  s   r  c                 C  s6   t | | }|dd}tt|}t|tjsJ |S )Nr  r  )_torch_triton_mappingrG  r
  r?   r@   rk   r   )r   adjusted_type	type_namer  r>   r>   rE   triton_type_to_torch	  s
   
r  r  rv   c                 C  sh   | j  o3|  | ko3|  | ko3| j|jko3| j|jko3|   |  ko3|  | kS rx   )	is_mkldnnr  r;  r   r   untyped_storager  r  r  rv   r>   r>   rE   is_same_tensor	  s   

r   c                 C  sJ   | j o$|  | ko$| j|jko$| j|jko$tjj| tjj|kS rx   )r  r  r   r   r@   r   mkldnnr  r  r>   r>   rE   is_same_mkldnn_tensor	  s   

r"  tuple[str, ...]c                   C  rA  )N)r  isnanlogical_notlogical_andsignbitand_leltgegteqner  xorr>   r>   r>   r>   rE   boolean_ops	
  r  r0  c                   @  r  )OpDtypeRuler$   type_promotion_kindr   override_return_dtypeNr  r>   r>   r>   rE   r1  
  r  r1  zdict[str, OpDtypeRule]op_dtype_propagation_rulesr2  r$   r3  c                 C  s   t ||t| < d S rx   )r1  r4  )r   r2  r3  r>   r>   rE   #register_op_dtype_propagation_rules&
  s   r5  c                 C  s"   t jjr| tjtjfv rtjS | S )z"Maybe upcast [b]float16 to float32)r\   ru  codegen_upcast_to_fp32r@   r	  r  r  r  r>   r>   rE   upcast_compute_type0
  s   r7  KeyTypeValTypec                   @  sl   e Zd ZdZd#ddZd$d
dZd%ddZd&ddZd'd(ddZd)ddZ	d*ddZ
d+dd Zd,d!d"ZdS )-
ScopedDictz
    A dictionary-like object that allows for scoped updates. It maintains
    an original dictionary and a set of new items that can override
    the original items within the scope.  The original dictionary is
    unmodified.
    original_dictMapping[KeyType, ValType]c                 C  s   || _ i | _d S rx   r;  	new_items)r  r;  r>   r>   rE   r  E
  r  zScopedDict.__init__r   r8  r<   r9  c                 C  s   || j v r
| j | S | j| S rx   r>  r;  r  r   r>   r>   rE   r  I
  s   


zScopedDict.__getitem__rv   rh  c                 C  s   || j |< d S rx   )r>  )r  r   rv   r>   r>   rE   __setitem__N
  r  zScopedDict.__setitem__r  rj   c                 C  s   || j v p	|| jv S rx   r?  r@  r>   r>   rE   __contains__Q
  r   zScopedDict.__contains__Nra  Optional[ValType]c                 C  s"   || j v r
| j | S | j||S rx   )r>  r;  rG  )r  r   ra  r>   r>   rE   rG  T
  s   

zScopedDict.getre   c                 C  s,   t | j}| jD ]}|| jvr|d7 }q|S r	  )rH   r;  r>  )r  r  r1  r>   r>   rE   r  Y
  s   


zScopedDict.__len__Iterator[KeyType]c                 c  s.    | j E d H  | jD ]
}|| j vr|V  q
d S rx   r=  )r  r1  r>   r>   rE   __iter__`
  s   

zScopedDict.__iter__c                 C  s   t | jp| jS rx   )rj   r;  r>  r  r>   r>   rE   r  f
  r%  zScopedDict.__bool__c                 C  r  rx   r   r@  r>   r>   rE   __delitem__i
  r  zScopedDict.__delitem__)r;  r<  )r   r8  r<   r9  )r   r8  rv   r9  r<   rh  )r   r  r<   rj   rx   )r   r8  ra  rC  r<   rC  r  )r<   rD  r  )r   r8  r<   rh  )r|   r}   r~   r   r  r  rA  rB  rG  r  rE  r  rF  r>   r>   r>   rE   r:  =
  s    






r:  )frozen_defaultfrozenrz   Optional[type[Any]]rI  c                 s"   d fdd}| d u r|S || S )Nrz   r_   r<   c                   s(   t jdkrtj| d dS tj|  dS )N)r  rv  T)kw_onlyrI  rH  )rW  version_infodataclasses	dataclass)rz   rH  r>   rE   wrapo
  s   
zir_dataclass.<locals>.wrap)rz   r_   r<   r_   r>   )rz   rI  rO  r>   rH  rE   ir_dataclassm
  s   rP  Optional[list[int]]c                  C  s&   t jj } | d ur| jr| jjS d S rx   )r@   r  r  r  fw_metadatabw_donated_idxs)r  r>   r>   rE   get_donated_idxs|
  s   rT  r  c                 C  sZ   ddl m}m} ddlm} | D ]}|||fvr*|jd ur*dd |jjD |jj|< qd S )Nr(   r  r.  c                 S  s   g | ]}|j qS r>   r'  r  r>   r>   rE   rF   
  s    z;set_kernel_post_grad_provenance_tracing.<locals>.<listcomp>)	codegen.simd_kernel_featuresr  r  r5  r/  re  r  r   ._inductor_triton_kernel_to_post_grad_node_info)r  r  r  r  r/  re  r>   r>   rE   'set_kernel_post_grad_provenance_tracing
  s   
rW  c                   @  s    e Zd ZdZdZdZdZdZdS )TritonAttrsDescriptorVersionr   r(   rI  r  r  N)r|   r}   r~   V0_NO_TRITONV1_COMPILERV2_BACKENDSV3_BACKENDS_TUPLEV4_DICTr>   r>   r>   rE   rX  
  s    rX  c                  C  sT   t jdd u rtjS dd l} dd l} t| jj	drtj
S t| j	j	dr'tjS tjS )Nru  r   AttrsDescriptor)	importlibutil	find_specrX  rY  triton.backends.compilertriton.compiler.compilerr   r  compilerr[  rZ  r]  )ru  r>   r>   rE   #get_triton_attrs_descriptor_version
  s   re  c                   C  s   t  tjkS rx   )re  rX  r]  r>   r>   r>   rE   triton_version_uses_attrs_dict
  r  rf  r  )rd   re   r<   re   )rh   ri   r<   rj   )r   r   )r   r   r   re   r   re   r<   r   r  )r   r   r<   r   )r   r   r<   ri   )r   r   r   r   r<   ri   )r   r   r<   r   )r   r   r   r   r<   r   )r   r   r<   r=   )r&  r'  r<   r(  )r&  r,  r<   r-  )r7  r8  r<   rj   )rE  r'   rF  rG  r<   rj   )rL  r
   rr   rT  rU  rV  r<   rW  )r8   )r   r=   r<   rh  )r(   r8   )
rj  rk  rl  rm  rn  re   r   r=   r<   r   )r>   rv  rv  rw  r8   )rj  rk  rl  rm  rn  re   rx  re   ry  r   r   r=   r<   r   )r  r
   r  r=   r<   rh  )r  r
   r  r  r<   rh  )r   re   r   re   r<   re   )rD   r  r  re   r<   r  )rD   r  r<   r  )r   r  r<   r  )r  r  r<   r  )r  r  r  r  r<   r=   )r  r  r  r+   r<   r  rx   )r  r  r  r  r<   r  )rr   r  rU  r  r<   r  )r  ri   r<   r=   )r   ri   r<   r  )r  r=   r<   rj   )r  rY   r  re   r<   r  )r!  rj   r<   rj   )r   r=   r<   r  )r  ri   r)  r*  r<   ri   )r   r
   r<   r6  )rr   r
   r<   rj   )rf  r?  r<   r@  )rf  r?  r<   r'   )r  r
   r<   r
   r  )NNT)rn  ro  rp  rq  rr  rj   r<   rs  )r  rm  r<   r  )r3  r&   r  r  r<   r  )r   r  r<   re   r  r   )r  r  r<   rj   r  )r  re   r   r   r<   r)   )r,  r1   r-  r.  r<   rj   )r1  r=   r<   rj   )r,  r1   r>  rj   r?  rj   r<   rj   )rG  r0   r<   rj   )
r,  r1   rk  re   r  re   r1  re   r<   rj   )r   r=   r<   r=   )r<   rg  )r,  r1   r<   rj   )r,  r1   r  r  r  r0   r<   rj   )FTFN)r,  r1   r  r0   r  r0   r  rj   r  rj   r  rj   r  r  r<   rj   )r   r  rr   r  rU  r  r<   r  )r   rk  rr   r
   rU  r
   r<   r  )r   rk  r<   r  )r   rk  rr   r
   rU  r
   r<   r  )r   rk  rr   r
   rU  r
   r<   r=   )r   rk  rr   r
   rU  r
   r<   r  )r  rk  r  rk  r<   rs  )r  rk  r  r  r<   r
   )r  r=   r<   rh  )r<   rq  )r  rm  r<   rj   )r  r  r<   rj   )r  ri   r<   r  )r  rj   rr   r
   rU  r
   r<   r  )r'  r=   r<   rj   )r'  r=   r<   re   )r1  r2  r<   rj   )
r  rk  rf  r%   r=  rm  r  r=   r<   rh  )rP  rQ  r<   rj   )re  rX  r7  rY  r<   rj   )re  rd  r<   rj   )rh  r6   r<   rj   )re  rq  r7  rr  r<   rj   )rt  r=   ru  rV  rv  rV  r<   r
   )rh  r6   rz  r{  ru  r|  rv  r}  r~  r  r<   rh  )r  re   r  re   r<   re   )r  r?  r<   re   )r  r  r<   rs  )r   r=   r<   r=   )r   rq  r<   rj   )r   r=   r<   rj   )r   r  r<   rj   )r^  r8  r'  rq  r  r  r  r  r  r=   r  rj   r<   rj   )r  r  r<   rh  )r   rY  r<   rj   )r  rY  r<   rj   )r<   r  )r   rk  rr   r
   rU  r
   r<   r  )r  r  r<   r  )rj  r  r  r  r<   r  )rD   rY  r<   rY  )r  r  r  r  r<   rh  )r  r  r  r  r<   r  )r   ri   r<   rj   )rl  rm  r   r5   r<   rh  )r   r  r<   r=   )r   r=   r<   r  )r  rY  rv   rY  r<   rj   )r<   r#  )r   r=   r2  r$   r3  r   r<   rh  )r   r  r<   r  )rz   rJ  rI  rj   r<   r
   )r<   rQ  )r  r  r  r=   r<   rh  )r<   rX  (6  
__future__r   r  r  rM  enumr   r_  r  rF  r  r  r  r   rw  r.  r  r  rW  r  r  rp  r  collections.abcr   r   r   r   r   r   r	   typingr
   r   r   r   r   r   r   r   r   r   r   typing_extensionsr   r   r   r   r   r   rl   r@   torch._inductor.runtime.hintsr   torch.utils._ordered_setr   torch.utils._pytreer   r   r   r    r!   r"   r#   torch._prims_commonr$   torch.fxr%   rD  r&   torch.fx.noder'   r#  r)   r  r+   r1  r-   r  r.   r/   r0   r1   r2   r3   output_coder5   r  r6   r7   rG   r;   	lru_cacherL   torch._dynamo.device_interfacerM   torch._dynamo.utilsrN   torch.autogradrO   torch.autograd.profiler_utilrP   (torch.fx.passes.graph_transform_observerrQ   torch.fx.passes.shape_proprR   torch.utils._sympy.functionsrS   rT   rU   rV   rW   torch.utils._sympy.symbolrX   rY   torch.utils._sympy.value_rangesrZ   r[   r  r\   runtime.runtime_utilsr]   r   _IS_WINDOWS	getLoggerr|   r   r_   r  r   	VarRangesr`  re   	InputTypeGPU_KERNEL_BIN_EXTSr  r  rQ  r%  rf   rg   rq   Functionrs   r   r   r   r   r   r   r%  r+  r6  r@  rI  rg  r   ru  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r
  r  r  r   r$  r(  r5  r9  r>  rI  rN  rO  r  rT  rU  rh  rm  r  r  r  r  r  r  rN  r  r  r  r  r  r  r  r  r  r'  r+  r0  r:  r=  rF  rZ  rc  rf  rr  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r   r  r  r  r  r  r  r  r"  r&  r*  r,  r0  r  r  r9  Enumr:  rO  rU  rW  rc  rg  rl  rp  rs  rx  r  r  r  r  r  r  r  r/  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  compiler  r  r  r   r"  r0  r1  r4  r5  r7  r8  r9  r:  rP  rT  rW  rX  re  rf  r>   r>   r>   rE   <module>   s<   4 


$T&		$=/7$ 
 
.
?
	,		!
	
$$		'	




	$
0 
