o
    ,h6                    @  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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/m0Z0m1Z1m2Z2 d dlm3Z3 d dl4Z4d dl5Z5d d	l6m7Z7 d d
l8m9Z9 d dl:m;Z; dgZ<e)r'd dlm=Z=m>Z>m?Z? d dl5m@Z@mAZAmBZB d dlCmDZD d dlEmFZF d dlGmHZH d dlImJZJ ddlKmLZL ddlMmNZN ddlOmPZP ddlQmRZRmSZSmTZTmUZUmVZVmWZWmXZX ddlYmZZZ ddl[m\Z\m]Z] g dZ^e*dZ_ej`dddZad d lbmcZc d d!ldmeZe d d"lfmgZg d d#lhmiZi d d$ljmkZk d d%llmmZm d d&lnmoZompZpmqZqmrZrmsZs d d'ltmuZumvZv d d(lwmxZxmyZy dd)lzm{Z{ dd*l|m}Z~ ejd+kZeeZe*d,Zee4je4jf Ze'e+e5jee5jBf  Zd-d.d/Zd0Zd0Zd0Zd1Zd2Zeed @ d kred3ksJ d4dd7d8Zdd<d=ZG d>d? d?e4jZejd@dAG dBdC dCZdddKdLZ	EdddMdNZej`ddOdPZddTdUZddXdYZdd]d^ZddadbZddfdgZ}ddjdkZddodpZddsdtZddwdxZdd{d|Zd}d~ fdddZdddZddddZ		ddddZ					ddddZdddZdddZdddZdddZdddZe/dZe*dd@dZG dd de(e$eef ZdddZdddÄZdddȄZddd̈́Z	ddddԄZdddلZddd܄ZdddZdddZdddZdddZdddZdddZdddZdddZdddZdddZdÐd dZd dlZdĐddZg ZdeĐd< dŐddZŐdĐddZej			@dƐdǐddZeZeZeZːdȐddZ̐dɐddZed3dʐddZG dd de&ZejG d d! d!ZG d"d# d#ZG d$d% d%e҃Zejǐdːd&d'ZG d(d) d)ZG d*d+ d+eՃZej`d̐d͐d.d/Zejΐdΐd0d1Zؐdΐd2d3Z	ddϐd8d9ZڐdАd>d?ZېdѐdAdBZܐdѐdCdDZݐdEdEdFdҐdIdJZސdӐdMdNZߐdԐdRdSZdՐdUdVZdWZdXZg dYZe+ee4jf ZdZeĐd[< d֐d\d]Zej`dאd^d_Zej`dؐd`daZej`dِdcddZdڐdedfZdԐdgdhZdԐdidjZdڐdkdlZdڐdmdnZdېdrdsZ	E	@	E	dܐdݐdxdyZddzd{ZG d|d} d}ZdސddZdސddZdߐddZdddZdddZdddZdddZejǐdddZ	ddddZdddZdddZdddZdddZdddZ dddZejǐdddZdΐddZej`dΐddZej`dʐddZej`dΐddZdΐddZdddZdddZ	dddZ
dddZddŐdƄZddǐdȄZG dɐdʄ dejZddΐdτZddҐdӄZddԐdՄZ	dddِdڄZddܐd݄ZdddZdddZdddZdddZdd~ fdddZdd~ fdddZdddZdddZejG d d dZejǐdddZdddZdddZ dd	d
Z!dddZ"dddZ#dddZ$dddZ%d 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d5d6Z,dd8d9Z-dd;d<Z.d	d?d@Z/ddAdBZ0ddCdDZ1dEdFdGdHdIdJdJdKZ2dLdM e23 D Z4e5dNZ6d
dOdPZ7ddQdRZ8ddUdVZ9ddWdXZ:ej`ddZd[Z;ejG d\d] d]Z<i Z=d^eĐd_< ddcddZ>e9 Z?deeĐdf< ddgdhZ@ddidjZAddkdlZBe*dmZCe*dnZDG dodp dpeeCeDf ZEe.d@dqdd@dAddudvZFddxdyZG	Eddd}d~ZHG dd dejZIej`dddZJdddZKdddZLdddZMdddZNdddZOdZPdddZQdddZRd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	TypeAlias	TypeGuard)mock)DeviceProperties)
OrderedSet)tree_map_only!activation_quantization_aten_pass)IterableSequence
ValuesView)SymBoolSymFloatSymInt)ELEMENTWISE_TYPE_PROMOTION_KIND)GraphModule)ShapeEnv)Node   )WorkspaceArgPythonWrapperCodegenGraphLowering)BufferExternKernelExternKernelOut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xrA   rA   P/var/www/html/scripts/venv/lib/python3.10/site-packages/torch/_inductor/utils.py
<listcomp>`   s    z get_gpu_type.<locals>.<listcomp>r*   r   r;   )	GPU_TYPESlenpop)
avail_gpusgpu_typerA   rA   rH   get_gpu_type^   s   rO   )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)r;   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)rg   rA   rA   rH   _align   s   rj   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gcdri   )rk   rA   rA   rH   rt      s   rt   c                   @  s&   e Zd ZdZdZdZeddd	Zd
S )rv   z<Symbolically round up to the nearest multiple of ALIGN_BYTESr*   Tvaluerl   r?   Optional[sympy.Expr]c                 C  s,   t |ttjfrtt|S t|r|S d S N)rn   rh   ro   Integerrj   rt   )clsry   rA   rA   rH   eval   s
   z
align.evalN)ry   rl   r?   rz   )__name__
__module____qualname____doc__nargs
is_integerclassmethodr~   rA   rA   rA   rH   rv      s    rv   Tfrozenc                   @  s2   e Zd ZU dZded< ded< ded< ded< d	S )
GraphPartitionMapzP
    Mapping from the partition info (e.g., input/output) to the graph info
    rh   idzlist[Optional[int]]input_index_mappingoutput_index_mapping	list[str]constant_namesN)r   r   r   r   __annotations__rA   rA   rA   rH   r      s   
 r      d   fnCallable[[], Any]warmuprepfloatc              
   C  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 ]}|   qYdd	 t|	D }d
d	 t|	D }t jjt jjjgdP}
t j  t|	D ],}|	  ||   t jjd |   W d   n1 sw   Y  ||   qt j  t dd	 t||D }W d   n1 sw   Y  t | }td t|
 jddd tdd	 |
 D }|r|tdd |D d 8 }td| |S )R  
    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.
        Ar;   dtypedeviceTenable_timing   r*   c                 S     g | ]	}t jjd dqS Tr   rC   r;   EventrF   _rA   rA   rH   rI          zfp8_bench.<locals>.<listcomp>c                 S  r   r   r   r   rA   rA   rH   rI      r   
activitiesRunCudaModuleNc                 S  s   g | ]	\}}| |qS rA   )elapsed_time)rF   serA   rA   rH   rI      r   
raw eventsself_device_time_totalsort_by	row_limitc                 S  s&   g | ]}|j tjkrd |jv r|qS )fused_abs_max_0device_typerR   CUDAnamerF   eventrA   rA   rH   rI      
    c                 s      | ]}|j V  qd S r{   device_time_totalr   rA   rA   rH   	<genexpr>       zfp8_bench.<locals>.<genexpr>     @@profiling results: %s ms)rC   r;   synchronizeemptyrh   float16r   recordrangezero_r   maxprofilerprofileProfilerActivityr   nvtxtensorzipmeanitemlogdebugkey_averagestablerS   events
statistics)r   r   r   cachestart_event	end_eventr   estimate_msn_warmupn_repeatpitimesresfiltered_eventsrA   rA   rH   	fp8_bench   sh   	




r   c                   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  t jjt jjjgd}
t|	D ]	}|  |   qtt 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 )r   r   r;   r   Tr   r   r*   r   Nr   r   r   r   c                 S  s&   g | ]}|j tjkr|jd kr|qS )zContext Syncr   r   rA   rA   rH   rI   7  r   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   rA   )rF   r   r   num_event_per_grouprA   rH   rI   F  s
    zprofiling time breakdown)r   c                 s  r   r{   r   r   rA   rA   rH   r   R  r   z+do_bench_using_profiling.<locals>.<genexpr>r   r   )rC   r;   r   r   rh   r   r   r   r   r   r   r   r   r   r   r   r   r   r   rS   r   rK   RuntimeError	enumerate_build_treesum)r   r   r   r   r   r   r   r   r   r   r   r   r   actual_eventsr   rA   r   rH   do_bench_using_profiling  sj   





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   rC   _C%_dispatch_has_kernel_for_dispatch_keyhasattrrB   opsImportErrorr   r@   )r   r   rA   rA   rH   has_torchvision_roi_alignW  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)
rC   r   r   rn   r@   typer   rP   Workercurrent_devicer   device_interfacerA   rA   rH   decode_deviceg  s   


r   itIterable[sympy.Expr]c                 C  s   t tj| tjjS r{   )	functoolsreduceoperatormulro   SOner   rA   rA   rH   sympy_productr     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 r{   rA   )rF   abrA   rA   rH   r   x  s    zsympy_dot.<locals>.<genexpr>)rK   ro   expandr   r   )r
  r  rA   rA   rH   	sympy_dotv  s   r  Iterable[_T]ValuesView[_T]c                 C  s   dd | D   S )Nc                 S  s   i | ]}t ||qS rA   )r   rE   rA   rA   rH   
<dictcomp>|      zunique.<locals>.<dictcomp>)valuesr  rA   rA   rH   unique{     r  number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: , )rn   ro   ExprrV   sympifyrh   r   runtime_ceildiv)r  r  rA   rA   rH   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%d&}|d'd( t| D  t| t r`| S d)||  S )*Nz*i8.r   rm   i1
float8e4nvfp8e4nvfloat8e5fp8e5float8e4b15fp8e4b15float8e4b15x4
fp8e4b15x4float8_e4m3fnfloat8_e5m2float8_e8m0fnuu8float4_e2m1fn_x2r   fp16bfloat16bf16float32fp32float64fp64int8i8int16i16int32i32int64i64u16u32u64)uint8uint16uint32uint64c                 S  s   i | ]}||qS rA   rA   )rF   rk   rA   rA   rH   r    s    z_type_of.<locals>.<dictcomp>*)r@   splitupdatelistr  rn   )r  	dtype_strtysrA   rA   rH   _type_of  sZ   
rL  lst"Iterable[Union[int, torch.SymInt]]list[sympy.Expr]c                 C     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 rA   )ro   r  rF   r   rA   rA   rH   rI     r  z-convert_shape_to_inductor.<locals>.<listcomp>rA   rM  rA   rA   rH   convert_shape_to_inductor  s   rS  r   Union[int, torch.SymInt]c                 C  sB   ddl m} t| tr| S t| tjrt| S |jjjj	| ddS )zL
    Like convert_shape_to_symint, but operates on a single expression.
    r*   VN)hint)
virtualizedrV  rn   rh   ro   r|   graphsizevars	shape_envcreate_symintnode)r   rV  rA   rA   rH   convert_to_symint  s   
r]   Iterable[Union[int, sympy.Expr]]list[Union[int, torch.SymInt]]c                 C  rP  )zz
    Takes a list of shapes from Inductor and converts them into symints (or just
    ints if all shapes are static).
    c                 S  s   g | ]}t |qS rA   )r]  rQ  rA   rA   rH   rI         z+convert_shape_to_symint.<locals>.<listcomp>rA   rR  rA   rA   rH   convert_shape_to_symint  s   ra  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 r{   )
alias_inforF   r  rA   rA   rH   r         zis_view.<locals>.<genexpr>)any_schema	argumentsrb  rA   rA   rH   is_view  s   rk  c                 C     dS NFrA   )r   rA   rA   rH   <lambda>      rn  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 r{   )is_pointwise_use)rF   urq  rA   rH   r     rf  z#is_pointwise_use.<locals>.<genexpr>)rb  rn   targetrC   _ops
OpOverloadr  getitemr   rk  rr   usersTag	pointwisetags)rp  rq  rw  rA   rv  rH   rt    s   

rt  rw  r
   ru   	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 )Nr  )appendplaceholderrK   )r  g
graph_argsrA   rH   add_tensor_arg  s   
z)gen_gm_and_inputs.<locals>.add_tensor_argr*   r   Tensor)r  r  r?   r)   )rC   fxGraphrs  r   r  rK   rh  returnsr@   r   outputr'   )rw  ru   r  r  nodegmrA   r  rH   gen_gm_and_inputs  s   

r  r;   Nonec                 C  s,   | dkrd S t | }| r|  d S d S Nr   )rP   rD   r   r   rA   rA   rH   r     s   r   modelCallable[..., Any]example_inputsSequence[Any]r   c                 C  sT   t | td t }t|D ]
}| | }t | qt }|d us&J || S )Ni9  )r   rC   manual_seedtimeperf_counterr   )r  r  r   r   t0r   resultt1rA   rA   rH   timed  s   

r  rA   
         ?repeatbaselinec                   sH   t  fddt|D }t | }t|| d | S )Nc                   s   g | ]	}t  qS rA   )r  r   r   r  r  r   rA   rH   rI   3  r   z%print_performance.<locals>.<listcomp>z.6f)rC   r   r   medianprintr   )r  r  r   r  r  r   timingstookrA   r  rH   print_performance*  s   r  objmethodc                   s$   t | |  t| | fdd dS )zKReplace obj.method() with a new method that returns a precomputed constant.c                     s    S r{   rA   rA   r  rA   rH   rn  =  ro  z#precompute_method.<locals>.<lambda>N)rB   setattr)r  r  rA   r  rH   precompute_method:  s   r  methodsr   c                 C  s   |D ]}t | | qdS )zFReplace methods with new methods that returns a precomputed constants.N)r  )r  r  r  rA   rA   rH   precompute_methods@  s   r  r  r  c                 C  s   t | |kt | |k  S r{   )rh   )r  r  rA   rA   rH   cmpF     r  rG   Union[int, Sequence[int]]sizeSequence[int]c                 C  s:   t | tr
| g| S t| dkrt| | d g| S | S )Nr*   r   )rn   rh   rK   r   )rG   r  rA   rA   rH   pad_listlikeJ  s
   

r  tuple[_T, ...]list[_T]c                 C  s&   t | dkrg S d	dd}t| |dS )
Nr   elemrb   r?   r@   c                 S  s0   t | tr| S ddlm} t | |sJ |  S )Nr*   )r9   )rn   r@   	schedulerr9   get_name)r  r9   rA   rA   rH   	sort_funcW  s
   
ztuple_sorted.<locals>.sort_funcr  )r  rb   r?   r@   )rK   sorted)rG   r  rA   rA   rH   tuple_sortedS  s   
	r  PRV)	covariantc                   @  s$   e Zd ZedddZdddZdS )CachedMethodr   r
   r?   r  c                 C     d S r{   rA   )r   rA   rA   rH   clear_cacheh     zCachedMethod.clear_cacheru   P.argsr  P.kwargsr  c                 O  r  r{   rA   selfru   r  rA   rA   rH   __call__k  ro  zCachedMethod.__call__N)r   r
   r?   r  )ru   r  r  r  r?   r  )r   r   r   staticmethodr  r  rA   rA   rA   rH   r  g  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?   r  c                   s   t |  rt|   d S d S r{   )r   delattrr  r  rA   rH   r    s   
z"cache_on_self.<locals>.clear_cache)r  r
   r?   r  )r   execlstripr  wrapsr  )r   r   ctxwrapperr  rA   r  rH   cache_on_selfo  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 )r  )r   r  originsrF   r  rA   rA   rH   rI     s    z%aggregate_origins.<locals>.<listcomp>) r  rn   rI  r  r  r  or_r   r1   r  )r  r  rA   rA   rH   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 )rs  r  N)rb  r   _overloadpacketr   rF   originrA   rA   rH   rI     s    

z)get_fused_kernel_name.<locals>.<listcomp>rC   rs  source_fn_stackr   r*   inductor_nodec                 S  s   g | ]
}|j d kr|jqS rs  )rb  r   r  rA   rA   rH   rI     s    r   fused)r  r  r   rb  r   rn   r@   r  r   NotImplementedErrorjoin)r  r  all_originssourcesr  	source_fnrA   rA   rH   get_fused_kernel_name  s.   r  r  r-   tuple[str, str]c                   s  t | }dd |D }tt}tt}d  t|rKtdd |D }t|dkrK|d j t dsAdd	 t j	D }| _
|j fd
dd |D ]3}d|jv rk|jd d urkt|jd j}	||	 |j d|jv r|jd d j}	||	 |j qM 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  rj  r  rA   rA   rH   rI     r   z'get_kernel_metadata.<locals>.<listcomp>c                 s  r   r{   )rY  )rF   nrA   rA   rH   r     r   z&get_kernel_metadata.<locals>.<genexpr>r*   r   )_inductor_kernel_metadata_node_to_idx_mapc                 S     i | ]\}}||qS rA   rA   )rF   idxr  rA   rA   rH   r    r  z'get_kernel_metadata.<locals>.<dictcomp>c                   s
    j |  S r{   )r  r  single_graphrA   rH   rn    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defaultdictrI  rK   r   rY  r   r   nodesr  sortr   r@   r  r  r   commentr  keysr  itemsformat_node)r  r  r  inductor_nodesfrom_node_dictoriginal_aten_dictunique_graphsnode_to_idx_mapr  r  sort_strmetadatadetailed_metadataoriginal_noder  r  rA   r  rH   get_kernel_metadata  sL   





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)rI  r   rL   r{  addr  )r  r  dominated_setr  userrA   rA   rH   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  r3   r?   rm   c                   sD   t |  jr| jS t |  jr| jS t |  jo!t |  jS r{   )rn   	TensorBoxdata
StorageBoxr3   	Pointwiser  r  is_unrealized_noderA   rH   r%    s
   

z*gather_origins.<locals>.is_unrealized_nodec                      g | ]	} |r|j qS rA   r  )rF   valr%  rA   rH   rI      r   z"gather_origins.<locals>.<listcomp>c                   r&  rA   r'  rF   r  r)  rA   rH   rI   !  r   )r  r3   r?   rm   )	itertoolsr  r  r  r   chain)ru   r  r+  kwarg_originsarg_originsrA   r$  rH   gather_origins  s   r/  exprc                   s@   ddd d fdd	d fd
ddfdd| 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.
    r0  rl   r?   rm   c                 S  s(   t | tjot| jdko| jd dkS )N   r   r   )rn   ro   MulrK   ru   r0  rA   rA   rH   is_neg_lead,  s   &zsympy_str.<locals>.is_neg_leadr@   c                   sj   t | tjr1t| jdkr( | jd r(| jd  d| jd jd  S dt| jS | S )Nr1  r*   r   z - z + )rn   ro   rp   rK   ru   r  rs   r3  )r4  sympy_str_mulrA   rH   sympy_str_add1  s
   (z sympy_str.<locals>.sympy_str_addc                   sB   t | tjr | rd| jd  S dt| jS | S )N-r*   z * )rn   ro   r2  ru   r  rs   r3  )r4  sympy_str_atomrA   rH   r5  <  s
   z sympy_str.<locals>.sympy_str_mulc                   sp   t | tjr	| jS t | tjtjfrd |  dS t | tttt	fr4| j
j ddtt| j dS t| S )N()r  )rn   ro   Symbolr   rp   r2  rZ   rW   rX   rY   funcr   r  rs   	sympy_strru   r@   r3  )r6  rA   rH   r8  G  s   "z!sympy_str.<locals>.sympy_str_atomN)r0  rl   r?   rm   r0  rl   r?   r@   rA   r3  rA   )r4  r6  r8  r5  rH   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*   rU  current_node
index_expr)
rX  rV  r_   compute_all_boundsrB   interpreterrw  r]   r^   unknown)r   rV  fx_noderA   rA   rH   get_bounds_index_exprT  s   
rF  prefixc                 C  s   | d dkS )Nr   rrA   )rG  rA   rA   rH   prefix_is_reductionb     rI  r\   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)r\   SIZEr[   )rG  r  rA   rA   rH   sympy_index_symbol_with_prefixf  s   rQ  checkc                 C  s   | st jot jS r{   )r_   debug_index_assertsassert_indirect_indexing)rR  rA   rA   rH   generate_assertr     rU  r   c                 C  s    | d dksJ t j| dddS )rL  r   r   TrM  )ro   r;  r   rA   rA   rH   sympy_index_symbolv  s   rX  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.
    replacedrl   replacementUnion[sympy.Expr, str]r?   rK  c                 S  s2   t | tjsJ t |trtj|| j| jdS |S )NrM  )rn   ro   r  r@   r;  r   is_nonnegative)r[  r\  rA   rA   rH   	to_symbol  s   
zsympy_subs.<locals>.to_symbolc                   s   i | ]
\}}| ||qS rA   rA   rF   krk   r_  rA   rH   r        zsympy_subs.<locals>.<dictcomp>N)r[  rl   r\  r]  r?   rK  )ro   r  xreplacer  )r0  rY  rA   rb  rH   
sympy_subs  s   

re  ,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 r{   is_symbolicrE   rA   rA   rH   r         zis_symbolic.<locals>.<genexpr>)	rn   rC   r%   r  rg  r+  r,  r  stride)r  rA   rA   rH   ri    s    ri  c                  G     t dd | D S )Nc                 s  rg  r{   rh  re  rA   rA   rH   r     rj  z"any_is_symbolic.<locals>.<genexpr>rg  )ru   rA   rA   rH   any_is_symbolic  r  rn  r  torch.fx.GraphModuleOptional[torch.fx.Node]c                 C  s   ddl m} tg d}t r|d | jjD ]9}t|j	|v r&|  S tj
jjs@t|j	tjjr@tjjj|j	jv r@|  S |jd }d urR||rR|  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_shapesrq  r   rC   $are_deterministic_algorithms_enabledrH  rY  r  r@   rw  	_inductorr_   graph_partitionrn   rx  ry  r   r|  cudagraph_unsafer~  r   get)r  rq  forbidden_setr  r(  rA   rA   rH   %get_first_incompatible_cudagraph_node  s*   r{  c                 C  s&   t tt| jj}|jdksJ |S )z$Get the output node from an FX graphr  )nextiterreversedrY  r  rb  )r  	last_noderA   rA   rH   output_node  s   r  OrderedSet[torch.device]c                 C  s\   | j jdd}tdd |D }t| jd }t|tr|n|f}tdd |D }||B S )Nr  rj  c                 s  s0    | ]}t |jd tjr|jd  jV  qdS r(  N)rn   r   ry  rC   r  r   r  rA   rA   rH   r     s    

z"get_all_devices.<locals>.<genexpr>r   c                 s  s>    | ]}t |tjjrt |jd tjr|jd  jV  qdS r  )rn   rC   r  r)   r   ry  r  r   r*  rA   rA   rH   r     s    

)rY  
find_nodesr   r  ru   rn   tuple)r  placeholder_nodesinput_devicesout_argout_argsout_devicesrA   rA   rH   get_all_devices  s   r  c                  C  s   t tj D ]B} | dsqtj|  }|j D ]+}|drDt||}t|tj	j
jjrD|jD ]}t|tj	j
jjrC|jjj  q1qtj| = qdtjv r_tjd }t|jjj`|jj`t  d S )Nz&torch._inductor.runtime.compile_tasks.triton_ztriton.runtime.driver)rI  sysmodulesr  
startswith__dict__rB   rn   rC   rv  runtimetriton_heuristicsCachingAutotunercompile_resultsTritonCompileResultkernelrunmod__del__r   driveractiveutilsinstancegccollect)module_namem	attr_namer  r  r  rA   rA   rH   unload_xpu_triton_pyds  s.   








r  _registered_cachesc                 C  s0   t | dr
t| jst|  dt|  | S )zh
    Use this decorator to register any caches that should be cache_clear'd
    with fresh_cache().
    cache_clearz# does not have a cache_clear method)r   callabler  AttributeErrorr  r  r  rA   rA   rH   clear_on_fresh_cache  s   
r  c                  C  s   t D ]} |   qdS )z&
    Clear all registered caches.
    N)r  r  r  rA   rA   rH   clear_caches*  s   
r  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 pt2 caches.

    Optionally, pass a dict as 'cache_entries' to get a list of filenames and sizes
    generated with this cache instance.
    )r  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  )rF   f)triton_cache_dirrA   rH   r  N  s
    zfresh_cache.<locals>.<dictcomp>c                   s   t jd |dS )Nz*Failed to remove temporary cache dir at %s)exc_info)r   warning)r<  r  r  )inductor_cache_dirrA   rH   rn  ]  s
    zfresh_cache.<locals>.<lambda>)onerrorz(on error, temporary cache dir kept at %s)r  tempfilemkdtempr   patchdictr  environr   r   r  r  rn   rK   existslistdirrH  
is_windowsrC   r=   rD   r  shutilrmtree	Exceptionr  )r  r  r  filesrA   )r  r  rH   fresh_cache2  sL   




r  seq	list[int]c                 C  s(   | j }tt| }ttt||ddS )NT)r  reverse)__getitem__r   rK   rI  r~  r  )r  gettera_rrA   rA   rH   argsortp  s   r  r[  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?   rh   c                   sZ   | \}}|\}}d
 fdd}|||k rdS |||krdS ||k r%dS ||kr+dS d	S )Nr0  %Union[bool, torch.SymInt, sympy.Expr]r?   rm   c                   s   t | tr| S  j| ddS )NT)size_oblivious)rn   rm   evaluate_exprr3  r[  rA   rH   evaluate~  s   
z*argsort_sym.<locals>.cmp.<locals>.evaluater   r*   r   )r0  r  r?   rm   rA   )r  r  a_idxa_valb_idxb_valr  r  rA   rH   r  z  s   zargsort_sym.<locals>.cmpc                 S  s,   g | ]\}}|t |tjr|jjn|fqS rA   )rn   rC   r%   r  r0  )rF   r  r   rA   rA   rH   rI     s    zargsort_sym.<locals>.<listcomp>r  c                 S  s   g | ]\}}|qS rA   rA   )rF   r  r   rA   rA   rH   rI     r`  )r  r  r  r  r?   rh   )r   r  r  
cmp_to_key)r[  r  r  exprsr  rA   r  rH   argsort_symw  s   r  r   torch.dtypec                 C  s    | t jkrdS t jd| d S )Nrf   rA   r   )rC   rE  r   element_sizer  rA   rA   rH   get_dtype_size  s   
r  c                   @  s   e Zd ZU ded< dS )LineContextr
   contextNr   r   r   r   rA   rA   rA   rH   r    s   
 r  c                   @     e Zd ZU ded< ded< dS )ValueWithLineMapr@   ry   zlist[tuple[int, LineContext]]line_mapNr  rA   rA   rA   rH   r       
 r  c                   @  s   e Zd ZdZd@dAddZejdBddZdCddZdDddZ	dDddZ
dEddZdFddZdDddZdEddZdGd d!ZdHd$d%ZdIdJd)d*ZdIdKd+d,ZdIdKd-d.Z	/dLdMd3d4ZdNd7d8ZdDd9d:ZdOd=d>Zd?S )PIndentedBuffer   r   initial_indentrh   r?   r  c                 C  s   g | _ || _d S r{   )_lines_indent)r  r  rA   rA   rH   __init__     
zIndentedBuffer.__init__tabwidthr  c                 c  s*    | j }z|| _ d V  W || _ d S || _ w r{   )r  )r  r  prevrA   rA   rH   set_tabwidth  s   zIndentedBuffer.set_tabwidthr  c                 C  s   t  }d}g }| jD ]:}t|tr| }|d u rq
nt|tr(|||jf q
|}t|ts1J || |d |d|	d 7 }q
t
| |S )Nr*   r  )r	   r  rn   DeferredLineBaser  r  r  r@   writecountr  getvalue)r  bufr   linemaplilinerA   rA   rH   getvaluewithlinemap  s$   




z"IndentedBuffer.getvaluewithlinemapr@   c                 C  s
   |   jS r{   )r  ry   r  rA   rA   rH   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  rn   r  r  r@   endswithr  r  )r  r  r  r  rA   rA   rH   getrawvalue  s    




zIndentedBuffer.getrawvaluec                 C  s   | j   d S r{   )r  clearr  rA   rA   rH   r        zIndentedBuffer.clearrm   c                 C  
   t | jS r{   )rm   r  r  rA   rA   rH   __bool__  r  zIndentedBuffer.__bool__c                 C  s   d| j | j  S )Nr  )r  r  r  rA   rA   rH   rG    rV  zIndentedBuffer.prefixc                 C  s   |  d d S )Nr  	writeliner  rA   rA   rH   newline  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  )rn   r  r  r  r  with_prefixrG  stripr  r  rA   rA   rH   r    s   

zIndentedBuffer.writelinelines3Sequence[Union[LineContext, DeferredLineBase, str]]c                 C  s   |D ]}|  | qd S r{   r  )r  r  r  rA   rA   rH   
writelines  s   zIndentedBuffer.writelinesr*   offset'contextlib.AbstractContextManager[None]c                   s   t jd fdd}| S )Nr?   r  c                	   3  s<     j  7  _ zd V  W  j  8  _ d S  j  8  _ w r{   r  rA   r  r  rA   rH   r    
   "z"IndentedBuffer.indent.<locals>.ctxr?   r  )
contextlibcontextmanager)r  r  r  rA   r  rH   indent  s   zIndentedBuffer.indentc                 C  s   |  j |7  _ d S r{   r  r  r  rA   rA   rH   	do_indent  r  zIndentedBuffer.do_indentc                 C  s   |  j |8  _ d S r{   r  r  rA   rA   rH   do_unindent  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  )rn   r  r   r  r  minrK   r  mathisinfr  r  rh   textwrapdedentrstriprG  )r  r  r
  r"  r  r   rA   rA   rH   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 rA   rA   )rF   r  r<  rA   rH   rI   1  r`  z&IndentedBuffer.map.<locals>.<listcomp>)r  r  r  )r  r<  r   rA   r'  rH   rs   /  s   zIndentedBuffer.mapc                 C  s   t |  d|   dS )Nr9  r:  )r   r  r  rA   rA   rH   __repr__4  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   rA   rA   rH   __add__7  s
   zIndentedBuffer.__add__Nr   )r  rh   r?   r  )r  rh   r?   r  )r?   r  r?   r@   r?   r  r?   rm   )r  r  r?   r  )r  r  r?   r  rx   )r  rh   r?   r  )r  rh   r?   r  F)r  r  r
  rm   r?   r  )r<  r%  r?   r  )r)  r   r?   r  )r   r   r   r  r  r  r  r  r  r  r  r   r  rG  r  r  r  r  r  r  r$  rs   r(  r*  rA   rA   rA   rH   r    s,    











r  c                      s(   e Zd Zd
 fddZddd	Z  ZS )FakeIndentedBufferr?   r  c                   s   t    d S r{   )superr  r  	__class__rA   rH   r  A  r  zFakeIndentedBuffer.__init__r   r@   r
   c                 C  s$   |dkr
t | |S td| d)Nr2  zTried to call self.z on FakeIndentedBuffer. This bufferis currently used on TritonTemplateKernel to prevent actualwrites to the body without explicitly specifying the body with`TritonTemplateKernel.set_subgraph_body(name)`)object__getattribute__r   )r  r   rA   rA   rH   r4  D  s
   
z#FakeIndentedBuffer.__getattribute__r,  )r   r@   r?   r
   )r   r   r   r  r4  __classcell__rA   rA   r1  rH   r/  @  s    r/  c               	   c  s<    t jt j} }zd V  W | |t _t _d S | |t _t _w r{   )r  stdoutstderr)initial_stdoutinitial_stderrrA   rA   rH   restore_stdout_stderrO  r  r:  c                   @  s`   e Zd ZdZdddZddd	ZdddZd ddZd!ddZd"ddZ	d#ddZ
d$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  rA   rA   rH   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  rA   rA   rH   r  `     zDeferredLineBase.__call__r   c                 C  r<  )z3Returns a new deferred line with the same conditionr=  r  rA   rA   rH   	_new_lined  r>  zDeferredLineBase._new_linerG  c                 C  s   |  | | j S r{   r?  r  )r  rG  rA   rA   rH   r	  h  r	  zDeferredLineBase.with_prefixc                 C  s   |  | j S r{   )r?  r  r  r  rA   rA   rH   r  k  rV  zDeferredLineBase.lstripr   Union[int, slice]c                 C  s   |  | j| S r{   r@  )r  r   rA   rA   rH   r  n  rV  zDeferredLineBase.__getitem__rm   c                 C  r  r{   )rm   r  r  rA   rA   rH   r  q  r  zDeferredLineBase.__bool__rh   c                 C  r  r{   )rK   r  r  rA   rA   rH   __len__t  r  zDeferredLineBase.__len__N)r  r@   )r?   r;  )r  r@   r?   r   )rG  r@   r?   r   )r?   r   )r   rA  r?   r   r-  r?   rh   )r   r   r   r   r  r  r?  r	  r  r  r  rB  rA   rA   rA   rH   r  X  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 r{   )r0  r  r  rE  )r  r  rE  r  r1  rA   rH   r  {  s   
zDelayReplaceLine.__init__r?   c                 C  s   | j | j|  S r{   )r  replacer  rE  r  rA   rA   rH   r    r	  zDelayReplaceLine.__call__c                 C  s   t | j| j|S r{   )rD  r  rE  r  rA   rA   rH   r?    rV  zDelayReplaceLine._new_line)r  r@   rE  rF  r  r@   r+  )r  r@   r?   rD  )r   r   r   r   r  r  r?  r5  rA   rA   r1  rH   rD  x  s
    
rD  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	   r  z6GPU arch does not support max_autotune_gemm mode usageFTr=   rc   D   z,Not enough SMs to use max_autotune_gemm mode)min_sms	avail_sms)extra)rn   rC   r   rO   r   createversionhipmajorr   r  r   multi_processor_count)rH  r   proprL  rM  rA   rA   rH   
is_big_gpu  s&   

rU  c                   C  s   t jdjS )Nr;   )rC   r;   get_device_propertiesrS  rA   rA   rA   rH   get_max_num_sms     rW  c                  C  s"   t j } t | dur|  S d S )zFHandle experimental carveout if set otherwise return hardware SM countNr   )rC   r   _get_sm_carveout_experimentalrW  )carveoutrA   rA   rH   get_num_sms  s   
r[  num_tma_descriptorsnum_programsOptional[int]r+   c                 C  sH   ddl m}m} |du rt }|d}||  t }||||| dS )zKBuilds and returns a WorkspaceArg for the device side TMA workspace buffer.r*   )r+   WorkspaceZeroModeNF)r  	zero_moder   
outer_name)codegen.commonr+   r_  r[  	from_boolTMA_DESCRIPTOR_SIZEunique_name)r\  r   r]  r+   r_  r`  r  rA   rA   rH   get_tma_workspace_arg  s   
rf  layoutr4   allowed_layout_dtypeslist[torch.dtype]c                 C  s:   | j |vrtd| j | t| jjo| j |v ot| jS )NzDNot using template since dtype %s is not in allowed layout dtypes %s)r   r   r   is_gpur   r   rU  )rg  rh  rA   rA   rH   _use_template_for_gpu  s   
rk  backendc                 C  "   |   dd tj  dD v S )Nc                 S     g | ]}|  qS rA   r
  rE   rA   rA   rH   rI         z)_use_autotune_backend.<locals>.<listcomp>,)upperr_   max_autotune_gemm_backendsrG  rl  rA   rA   rH   _use_autotune_backend     ru  c                 C  rm  )Nc                 S  rn  rA   ro  rE   rA   rA   rH   rI     rp  z._use_conv_autotune_backend.<locals>.<listcomp>rq  )rr  r_   max_autotune_conv_backendsrG  rt  rA   rA   rH   _use_conv_autotune_backend  rv  rx  F)enable_int32enable_float8ry  rz  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 oMtjpBtjoMtdoM|| j|jS )Nr*   )BackendFeaturehas_backend_featurer   TRITON)rb  r{  r|  rC   r   r1  r3  r;  extendr+  r,  rj  r   r   rk  r   r_   max_autotunemax_autotune_gemmru  TRITON_TEMPLATES)rg  ry  rz  r{  r|  layout_dtypesrA   rA   rH   use_triton_template  s"   
	r  matricesr3   c                    s^   ddl m}m} ddlm  d fd	d
| rtjrdS tjjo.| o.t	fdd| D S )Nr   )has_triton_stable_tma_apihas_triton_tma_devicer*   rU  rG   r3   r?   rm   c                   s   t |  dkr
dS |  }|tjtjtjfvrdS |  }| }|	 s*|s*dS |j
d }|r6|j
d }|tjkrE jj|drEdS ||j } jj|tS )Nr1  Fr*   r       )rK   get_size	get_dtyperC   r   r1  r+  
get_layoutis_transposedis_contiguousr  rY  rZ  statically_known_ltitemsizestatically_known_multiple_ofTMA_ALIGNMENT)rG   r   rg  
transposed	inner_diminner_bytesrU  rA   rH   _is_tma_compatible  s$   


z3use_triton_tma_template.<locals>._is_tma_compatibleFc                 3      | ]} |V  qd S r{   rA   )rF   r  )r  rA   rH   r     rj  z*use_triton_tma_template.<locals>.<genexpr>rG   r3   r?   rm   )
torch.utils._tritonr  r  rX  rV  r_   cpp_wrapperr  enable_persistent_tma_matmulrr   )r  r  r  rA   )rV  r  rH   use_triton_tma_template  s   r  r  r  ra  c           	      C  s   ddl m} |jjj|| | dd}|dks|tjjk rdS ddlm	} t
jjr+dS t
jt
jt
jg}t| |oAtjp=tjoAtd}|rN| sNtd	 dS |S )
Nr*   rU  r   fallbackr   F)try_import_cutlassCUTLASSzFailed to import CUTLASS lib. Please check whether _inductor.config.cuda.cutlass_dir is set correctly. Skipping CUTLASS backend for now.)rX  rV  rY  rZ  	size_hintr_   r;   cutlass_backend_min_gemm_sizecodegen.cuda.cutlass_utilsr  rC   rP  rQ  r   r1  r;  rk  r  r  ru  r   r  )	rg  r  r  ra  rV  	gemm_sizer  r  r   rA   rA   rH   use_cutlass_template!  s(   

r  op_namec                 C  s4   t jj }|dkrdS |  dd |dD v S )z8Check if CUTLASS should be used for the given operation.ALLTc                 S  rn  rA   ro  rE   rA   rA   rH   rI   F  r`  z'_use_cutlass_for_op.<locals>.<listcomp>rq  )r_   r;   cutlass_enabled_opsrr  rG  )r  enabled_opsrA   rA   rH   _use_cutlass_for_opA  s   r  r  r   )rc   r  re   rd      r   _IntLikec              
   C  sV   ddl m} |jjtt|t|  t|t| o*|jj	 o*|jj
 o*tj S )Nr   rU  )torch._inductor.virtualizedrV  rY  rZ  statically_known_truero   AndGedecompose_k_thresholdaot_moder  r_   disable_decompose_k)r  r  ra  rV  rA   rA   rH   use_decompose_k_choiceT  s   r  c           
        s$  t |tjr|jstS t | tjr| jrt |tjr |js d n	t||  ||  dt|} fdd|D }g g g }}}|D ].}|| }|dk rOqD||d @ dkra|dkra|| qD|d dkrm|| qD|| qDtj	d	kr~|| | S t
|tkr|S || | }	|	d t S )
Nr  r1  c                   s    g | ]}| kr|kr|qS rA   rA   )rF   divisormax_k_splitmin_k_splitrA   rH   rI   u  s
    z get_k_splits.<locals>.<listcomp>rd   r*   r   r  
EXHAUSTIVE)rn   ro   r  	is_numberdefault_k_splitsr  divisorsr  r_   max_autotune_gemm_search_spacerK   k_splits_limit)
r  r  ra  r  pow_of_2_divisorsmul_of_32_divisorsrest_of_splitsdkPartbest_splitsrA   r  rH   get_k_splitsd  s<   


r  c                 C  s   t j| jS r{   )rC   r;   rV  gcnArchNamer   rA   rA   rH   _rocm_native_device_arch_name  rX  r  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?   r  c                   S     g S r{   rA   rA   rA   rA   rH   r    r  z*try_import_ck_lib.<locals>.gen_ops_libraryc                   S  r  r{   rA   rA   rA   rA   rH   r    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   rA   rA   rA   rH   r    s    r  )r?   r  )ck4inductor(ck4inductor.universal_gemm.gen_instancesr  r  ck4inductor.universal_gemm.opr  r  r  dirname__file__r   )r  r  r  r  package_dirnamerA   rA   rH   try_import_ck_lib  s   

r  c                   s   t jst j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rMdS t \}}}}|s]td	 dS t  re|t j	_t j	jsptd
 dS |t j	jkr}td dS dS )NFr;   c                 S  s   i | ]
}| d d |qS ):r   )rG  rF   ra  rA   rA   rH   r    rc  z#use_ck_template.<locals>.<dictcomp>r  r   c                   s   g | ]} | qS rA   rA   r  requested_archsrA   rH   rI     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  rC   rP  rQ  r   r   r  rocmarchrG  r  ck_supported_archr   r   r1  r3  r  r   r  	is_fbcodeck_dir)rg  native_archrequested_supported_archsck_package_dirnamer   rA   r  rH   use_ck_template  s<   




r  c                 C  :   ddl m} tdot| o|jjj|| | dddkS )Nr*   rU  CKr   r  r   rX  rV  ru  r  rY  rZ  r  rg  r  r  ra  rV  rA   rA   rH   use_ck_gemm_template     r  c                 C  r  )Nr*   rU  CKTILEr   r  r   r  r  rA   rA   rH   use_ck_tile_gemm_template  r  r  c                 C  s   t dot| S )Nr  )rx  r  rg  rA   rA   rH   use_ck_conv_template   rV  r  c                 C  s   t jpt jo| jjdkS r  )r_   r  r  r   r   r  rA   rA   rH   _use_template_for_cpu  s   

r  mat1Union[ReinterpretView, Buffer]mat2c                 C  s6   ddl m} t|j|sJ t| ||ddo|j S )Nr*   )r4   F)require_constant_mat2)r  r4   rn   rg  use_cpp_gemm_templater  )rg  r  r  r4   rA   rA   rH   use_cpp_bmm_template
  s
   r  mat2_transposedr  is_woq_int4q_group_sizec                 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  rG   r3   r?   rm   c                 S  s   |    |  d dkS )Nr   r*   )freeze_layout
get_striderG   rA   rA   rH   is_last_dim_stride1J  s   z2use_cpp_gemm_template.<locals>.is_last_dim_stride1r  )r  r  codegen.cpp_micro_gemmr  codegen.cpp_utilsr  kernel.mm_commonr  r  ru  r_   cppweight_prepackr  rC   rB  r7  r3  r1  halfr   has_free_symbolsrn   BaseViewunwrap_viewparallel_num_threadsr"  is_module_buffer)rg  r  r  r  r  r  r  r  r  r  r  	int8_gemmr  r  r  ra  r  r   r  r  rA   rA   rH   r    sX   		


r  c                   C  s   t jpt j p
tdS )NATEN)r_   r  r  ru  rA   rA   rA   rH   use_aten_gemm_kernelsW  s   
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?   r  c                 C  s   t tj| _d S r{   )r|  r  counterr   r  rA   rA   rH   r  a  rV  zDebugDirManager.__init__c                 C  s0   t jjj| _| j d| j | _| jt jj_d S )N_tmp_)rC   _dynamor_   debug_dir_rootr  r   new_namer  rA   rA   rH   	__enter__d  s   zDebugDirManager.__enter__ru   r
   c                 G  s   t | j | jtjj_d S r{   )r  r  r  r  rC   r  r_   r  )r  ru   rA   rA   rH   __exit__i  s   zDebugDirManager.__exit__Nr,  )ru   r
   r?   r  )
r   r   r   r+  r  r  r   r  r  r  rA   rA   rA   rH   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?   r  c                        |  d S r{   r  r  source_codesrA   rH   save_output_codew  r  z*run_and_get_code.<locals>.save_output_coder#  r  r@   r?   r  rY  r/   r   r  r3  rC   r  reset)r   ru   r  r/   r#  r  rA   r!  rH   run_and_get_coden  s   

r'  c                 O  sF   t | g|R i |\}}g }|D ]}|td|tj q||fS )Nz	'''.*?''')r'  r~  refindallDOTALL)r   ru   r  r  r"  kernelsr  rA   rA   rH   run_and_get_kernels  s
   r,  tuple[Any, list[str]]c                   s   d fdd}t |S )Nr?   r
   c                    s     } |     | S r{   )r   backwardr  r   rA   rH   run_with_backward  s   z1run_fw_bw_and_get_code.<locals>.run_with_backward)r?   r
   )r'  )r   r0  rA   r/  rH   run_fw_bw_and_get_code  s   r1  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?   r  c                   r  r{   r  r   r!  rA   rH   r#    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$   e Zd ZdZdddZdd	d
ZdS )z@get_code.<locals>.patched_compile_to_module.<locals>.DummyModulez4This is empty to replace the generated triton moduler?   r  c                 S  r  r{   rA   r  rA   rA   rH   r    r  zIget_code.<locals>.patched_compile_to_module.<locals>.DummyModule.__init__ru   r
   r  c                 _  r  r{   rA   r  rA   rA   rH   call  r>  zEget_code.<locals>.patched_compile_to_module.<locals>.DummyModule.callNr,  ru   r
   r  r
   r?   r  )r   r   r   r   r  r2  rA   rA   rA   rH   DummyModule  s    
r4  )r  codegen_with_cpp_wrappercodegenry   )r  r4  wrapper_codekernel_code)r#  rA   rH   patched_compile_to_module  s   

z+get_code.<locals>.patched_compile_to_modulecompile_to_moduler#  Nr$  )r  r/   r?   r
   r%  )r   ru   r  r/   r9  r   rA   )r#  r"  rH   get_code  s$   
(


r;  c                 O  sJ   t | g|R i |}dt|  krdks!n J dt| |d S Nr*   r1  z%expected one or two code outputs got r   )r;  rK   )r   ru   r  r"  rA   rA   rH   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'  rK   )r   ru   r  r   r"  rA   rA   rH   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.   r7   ru   r
   r  r?   r  c                    s2   | i | | d }t | sJ | d S )Nr1  )rn   r  )ru   r  rY  r/   graph_lowerings	real_initrA   rH   	fake_init  s   z-run_and_get_graph_lowering.<locals>.fake_initr  r3  )torch._inductor.graphr/   torch._inductor.output_coder8   r  r   r  r3  )r   ru   r  r8   rC  r  rA   r@  rH   run_and_get_graph_lowering  s   
rF  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._inductorrI  	loweringsr  partial)rG  rH  rI  orig_fnrA   rA   rH   override_lowering  s   
rN  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 r{   rA   )r  r  outrM  rP  rO  rA   rH   r    s
   


z(add_scheduler_init_hook.<locals>.wrapperr  N)r  r
   r  r
   r?   r
   )torch._inductor.schedulerrR  r  unittestr   r  r3  )rO  rP  rR  r  rA   rT  rH   add_scheduler_init_hook  s   rW  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)rX  rA   rA   rH   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   r7  z--only=N)r  argvr   rK   
ValueErrorr  )r  r  rA   rA   rH   get_benchmark_name  s   

r^  r  c                 C  rl  )Nc                 s      | ]}|d kV  qdS r*   NrA   rE   rA   rA   rH   r   =  rj  zis_ones.<locals>.<genexpr>rr   r  rA   rA   rH   is_ones<  r  rc  c                 C  rl  )Nc                 s  r_  )r   NrA   rE   rA   rA   rH   r   A  rj  zis_zeros.<locals>.<genexpr>ra  rb  rA   rA   rH   is_zeros@  r  rd  inputsSequence[torch.Tensor]c                 C  rl  )Nc                 s  s,    | ]}t |tjr|jtd kV  qdS )r   N)rn   rC   r  r   )rF   r   rA   rA   rH   r   E  s    

z is_cpu_device.<locals>.<genexpr>ra  )re  rA   rA   rH   is_cpu_deviceD  s   rg  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)rn   ro   r  r   rC   r=  r5  )r(  rA   rA   rH   get_sympy_Expr_dtypeL  s   rh  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 r{   )rC   r   r   )ri  ru   r  r   rA   rA   rH   maybe_profileV  s   "
rk  c                  C  s   t jj} | dk rt } | S Nr*   )r_   r  threadsrC   get_num_threads)rm  rA   rA   rH   r  _  s   r  c                  C  s,   ddl m}  |  }|dtjjrdS dS )Nr*   )get_backend_options
num_stagesr1     )runtime.triton_helpersro  ry  rC   rP  rQ  )ro  optionsrA   rA   rH   get_backend_num_stagesf  s   rt  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.testingru  rv  rC   r   r1  r3  inspect	signature
parametersry  torch._utils_internalrx  backendsr;   matmul
allow_tf32)r   ru  rv  rx  sm_clockrA   rA   rH   get_device_tflopsn  s   


r  c                  C  s   ddl m}  |  S )Nr   get_dram_gbps)ry  r  r  rA   rA   rH   get_gpu_dram_gbps  s   r  c                  C  s"   ddl m}  | jjdddS )Nr   r  max_shared_mem)triton.runtimer  r  r  rV  ry  r  rA   rA   rH   get_gpu_shared_memory  s   r  reduction_typec                 C  s
   |  dS )Nwelford)r  r  rA   rA   rH   is_welford_reduction  r  r  c                 C  s   t | rdS | dkrdS dS )Nrq  online_softmax_reducer1  r*   )r  r  rA   rA   rH   reduction_num_outputs  s
   r  c                   C  s   t  dkS )NLinux)platformsystemrA   rA   rA   rH   is_linux  rJ  r  c                   C  s
   t jdkS )Nra   )r  r  rA   rA   rA   rH   r    r  r  itrIterable[Any]c                 C  rl  )Nc                 s  s$    | ]}t |tjo|j V  qd S r{   )rn   ro   r  r  rE   rA   rA   rH   r     s   " z#has_free_symbols.<locals>.<genexpr>rm  )r  rA   rA   rH   r
    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  rA   Tzunexpected type for is_dynamic F)r  r  rn   r   r"  r  ComputedBufferr0   r
  maybe_get_sizemaybe_get_strider3   	TypeErrorr   )ru   r  trA   rA   rH   
is_dynamic  s   
r  c                   @  s   e Zd ZdZdZdS )PlaceholderKERNEL_NAMEDESCRIPTIVE_NAMEN)r   r   r   r  r  rA   rA   rA   rH   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encodingr  )r  	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	   rU   rQ   	propagater  rY  r   nowrT   lint	recompiler  r   rZ  r   )r<  r  r  rX  r  r  	before_ioafter_io
start_timetime_elapsedr  rA   rA   rH   pass_execution_and_save  s>   

"r  	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  rn   CppTemplateBufferrg  MultiOutputLayoutr  r  rA   rA   rH   is_multi_outputs_template  s   r  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  rn   MultiOutputrK   re  r  r  rA   rA   rH   #is_output_of_multi_outputs_template  s   r  r   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   rC   r   torchrecr  defaultr  r  r  rb  r  rA   rA   rH   is_collective	  s"   

r  "Optional[Union[IRNode, Operation]]c                 C  s   ddl m} t| |jkS Nr*   r  )r  r  r   _WaitKernelr  r  rA   rA   rH   is_wait0	  s   r  snoder9   c                 C  4   ddl m} t| |rtdd | jD S t| jS )Nr   GroupedSchedulerNodec                 s  rg  r{   )contains_collectiverE   rA   rA   rH   r   :	  rj  z&contains_collective.<locals>.<genexpr>)rU  r  rn   rg  snodesr  r  r  r  rA   rA   rH   r  6	     

r  c                 C  r  )Nr   r  c                 s  rg  r{   )contains_waitrE   rA   rA   rH   r   C	  rj  z contains_wait.<locals>.<genexpr>)rU  r  rn   rg  r  r  r  r  rA   rA   rH   r  ?	  r  r  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 r  )r  r  rn   rC   rx  ry  r  r  r  rA   rA   rH   is_fallback_opH	  s   r  buf_namename_to_bufname_to_fused_nodec                 C  s   |||  j   S r{   )defining_opr  )r  r  r  rA   rA   rH   buf_name_to_fused_snodeS	  s   r  c                 C  rl  rm  rA   r  rA   rA   rH   rn  ^	  ro  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_dependenciesr  r   find_recursive_deps_of_node)r  r  r  r  r  depdefining_op_for_deprA   rA   rH   r  Y	  s"   

r  c                 C  rl  rm  rA   r  rA   rA   rH   rn  w	  ro  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_outputsr{  r  r  find_recursive_users_of_node)r  r  r  r  r  or  user_oprA   rA   rH   r  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)r1  r   )rC   
_functorchr_   functionalize_rng_ops)r  r  num_rng_seed_offset_inputsrA   rA   rH   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
    rG   r)   r?   rm   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_staterW  r  rA   rA   rH   is_saved_tensor	  s   
z'count_tangents.<locals>.is_saved_tensorr   r  r*   N)rG   r)   r?   rm   )rY  r  rb  r  rI  r   rK   )r  r  	arg_countstatic_arg_idxsr  rA   rA   rH   count_tangents	  s   


r  c                   @  s.   e Zd ZU ded< dddZedd	d
ZdS )	BoxedBoolrm   ry   r?   c                 C  s   | j S r{   )ry   r  rA   rA   rH   r  	  s   zBoxedBool.__bool__r  r
   Union[BoxedBool, bool]c                 C  s   t | tr
d| _| S dS rm  )rn   r  ry   r  rA   rA   rH   disable	  s   
zBoxedBool.disableNr-  )r  r
   r?   r  )r   r   r   r   r  r  r  rA   rA   rA   rH   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@   r8  r  r  gpurm   cpp_definitionr?   r
   c                   s     | | |||||S r{   r  )r  r   r8  r  r  r  r  orig_define_kernelrA   rH   define_kernel	  s   
z.collect_defined_kernels.<locals>.define_kernelr  )NTN)r  r-   r   r@   r8  r@   r  r  r  rm   r  r  r?   r
   )codegen.wrapperr-   r  r   r  r3  )r  r-   r  rA   r  rH   collect_defined_kernels	  s   "r  c                 C  s   | d S )N__original__rA   rW  rA   rA   rH    get_cloned_parameter_buffer_name	     r	  c                 C  s   | t v S r{   )rJ   r  rA   rA   rH   rj  	  r
  rj  c                 C  s   | dkot | S )Nr<   )rj  r  rA   rA   rH   device_need_guard	  rV  r  c                 C  sL   t  r| tjkrtj rtj dkrt jrdS | ttj	tj
tjgv S )N)rJ  r   F)r_   r  rC   r1  r;   rD   get_device_capabilitybfloat16_atomic_adds_enabledr   r=  rm   r  rA   rA   rH   ,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*   )overloadpacketrC   r   atenscatter_reduce_scatter_reducescatter_rj  r  r_   r  fallback_scatter_reduce_sumdynamic_threadsr  rm   r=  ru  )r  r  r  r  r  r  	reduce_tyrA   rA   rH   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  3r  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  rU  r  r  rK   r   rn   is_reductionr  r!  reduction_hintread_writesreadswritesr   r   )r  r  r  r  r  r  is_redr  rA   rA   rH   dump_node_schedule
  s0   




r*  r   r  c                 C  s*   ddl m} ||  t| j t dkS )Nr   )r  )rt  r  storage_offsetr  r   GPU_ALIGN_BYTES)r   r  rA   rA   rH   tensor_is_aligned;
  s   r-  example_inputc                 C  s   t | jjsdS tjpt| S rm  )rj  r   r   r_   assume_aligned_inputsr-  )r.  rA   rA   rH   should_assume_input_alignedI
  s   r0  r  c                  C  s4   t jj } | st S | jj}|st S | S r{   )	rC   _guardsTracingContexttry_getr  nullcontextr  r[  suppress_guards)tracing_contextr[  rA   rA   rH   #maybe_get_suppress_shape_guards_ctxR
  s   r7  tuple[_T, 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)rV  r   r  r3  r_   rC   r  r&  r  loggingr	   StreamHandlertorch._inductor.codecacher9  
addHandlerlevelsetLevelDEBUGr  removeHandler)r   ru   r  r  r:  log_capture_stringchr9  
prev_levelr  r   rA   rA   rH   run_and_get_cpp_codec
  s$   




rE  Sequence[InputType]Optional[ShapeEnv]c                 C  s<   t | }|d ur|jS | D ]}t|tjr|jj  S qd S r{   )rQ   r[  rn   rC   r%   r  )re  r  inputrA   rA   rH   shape_env_from_inputs|
  s   rI  Callable[[list[InputType]], _T]inputs_to_checkmutated_input_idxsOrderedSet[int]c                   s&   t  dkrS d fdd}|S )	Nr   
new_inputslist[InputType]r?   r
   c                   s0   t |  \}}| }t|rt|| |S r{   )copy_misaligned_inputsrK   rC   _foreach_copy_)rN  old_tensorsnew_tensorsrS  rK  r  rL  rA   rH   r  
  s   z)align_inputs_from_check_idxs.<locals>.run)rN  rO  r?   r
   )rK   )r  rK  rL  r  rA   rT  rH   align_inputs_from_check_idxs
  s   rU  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`  rA   )rF   shaperk  rA   rA   rH   r   
  s    z)clone_preserve_strides.<locals>.<genexpr>r*   rx   )r  r   r   rk  rC   
as_stridedclone)rG   needed_sizebufferrA   rA   rH   clone_preserve_strides
  s   "r[  rN  rO  check_inputs_idxsreturn_pair_idxsOptional[OrderedSet[int]]-tuple[list[torch.Tensor], list[torch.Tensor]]c                 C  s   g }g }|du}|D ]3}| | }t |tjsJ dt| | t r=t|| |< |r=||v r=|| || |  q
||fS )z
    Clones misaligned tensors which we inferred were aligned. Returns a tuple of [old_tensors], [new_tensors] for every
    cloned tensor which is in `return_pair_idxs`.
    Nz Expected tensors only, but got: )rn   rC   r  r   data_ptr	ALIGNMENTr[  r  )rN  r\  r]  rR  rS  ret_pair_definedr   _inprA   rA   rH   rP  
  s   

rP  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   )rn   rC   r  r`  ra  r  rK   )re  rd  aligned_static_input_idxsr  rH  rA   rA   rH   remove_unaligned_input_idxs
  s   
rf  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*   rU  T)rX  rV  rC   iinfor;  r   rY  rZ  r  r[  has_hintr  )r   rV  int_maxr  rh  rA   rA   rH   expr_fits_within_32bit
  s   
rj  compiled_graphr8   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 r{   )rh   deserialize_symexprevaluate_symexpr)r   )fakify_first_callr[  rA   rH   map_expr  s
   

z4set_tracing_context_output_strides.<locals>.map_exprc                 3  r  r{   rA   )rF   r   )rp  rA   rH   r     rj  z5set_tracing_context_output_strides.<locals>.<genexpr>)r   r
   r?   rl  )
rC   r1  r2  r3  output_stridesrK   rI  r  ro  r  )r  rk  r  r  r  rA   )ro  rp  r[  rH   "set_tracing_context_output_strides
  s"   
rr  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_cacher  rC   _utils_internalis_fb_unit_testtorch._inductor.fb.remote_cachert  ModuleNotFoundErrorjustknobs_getval_intrs  rA   rA   rH    should_use_remote_fx_graph_cache  s   

r{  c                 C  s   t dd| S )Nz[^a-zA-Z0-9_]r   )r(  subrW  rA   rA   rH   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_e8m0fnuztl.float4_e2m1fn_x2c                 C  r  rA   rA   r`  rA   rA   rH   r  3  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_mappingry  )r   triton_type_namerA   rA   rH   triton_type9  s   r  c                 C  s6   t | | }|dd}tt|}t|tjsJ |S )Nr~  r  )_torch_triton_mappingry  rG  rB   rC   rn   r   )r   adjusted_type	type_namer  rA   rA   rH   triton_type_to_torch?  s
   
r  r!  ry   c                 C  sh   | j  o3|  | ko3|  | ko3| j|jko3| j|jko3|   |  ko3|  | kS r{   )	is_mkldnnr  rk  r   r   untyped_storager`  r+  r!  ry   rA   rA   rH   is_same_tensorG  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 r{   )r  r  r   r   rC   r   mkldnnr`  r  rA   rA   rH   is_same_mkldnn_tensorS  s   

r  tuple[str, ...]c                   C  rl  )N)r   isnanlogical_notlogical_andsignbitand_leltgegteqner  xorrA   rA   rA   rA   rH   boolean_ops]  r>  r  c                   @  r  )OpDtypeRuler&   type_promotion_kindr   override_return_dtypeNr  rA   rA   rA   rH   r  q  r  r  zdict[str, OpDtypeRule]op_dtype_propagation_rulesr  r&   r  c                 C  s   t ||t| < d S r{   )r  r  )r   r  r  rA   rA   rH   #register_op_dtype_propagation_rulesz  s   r  zOrderedSet[str]op_requires_libdevice_fp64c                 C  s   t |  d S r{   )r  r  rW  rA   rA   rH   #register_op_requires_libdevice_fp64  r  r  c                  C  s8   ddl m}  | j j}|dkrtjS |dkrdS tjS )Nr   rU  r   r<   )r  rV  rY  get_current_device_or_throwr   r_   cpu_backendcuda_backend)rV  
device_strrA   rA   rH   get_current_backend  s   r  c                 C  s,   | t jt jfv rtjjrt dkrt jS | S )z"Maybe upcast [b]float16 to float32r  )rC   r   r1  r_   r  codegen_upcast_to_fp32r  r3  r  rA   rA   rH   upcast_compute_type  s   
r  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 r{   r  	new_items)r  r  rA   rA   rH   r    r  zScopedDict.__init__r  r  r?   r  c                 C  s   || j v r
| j | S | j| S r{   r  r  r  r  rA   rA   rH   r    s   


zScopedDict.__getitem__ry   r  c                 C  s   || j |< d S r{   )r  )r  r  ry   rA   rA   rH   __setitem__  r  zScopedDict.__setitem__r3  rm   c                 C  s   || j v p	|| jv S r{   r  r  rA   rA   rH   __contains__  r	  zScopedDict.__contains__Nr  Optional[ValType]c                 C  s"   || j v r
| j | S | j||S r{   )r  r  ry  )r  r  r  rA   rA   rH   ry    s   

zScopedDict.getrh   c                 C  s,   t | j}| jD ]}|| jvr|d7 }q|S rl  )rK   r  r  )r  r  ra  rA   rA   rH   rB    s   


zScopedDict.__len__Iterator[KeyType]c                 c  s.    | j E d H  | jD ]
}|| j vr|V  q
d S r{   r  )r  ra  rA   rA   rH   __iter__  s   

zScopedDict.__iter__c                 C  s   t | jp| jS r{   )rm   r  r  r  rA   rA   rH   r    rV  zScopedDict.__bool__c                 C  r<  r{   r=  r  rA   rA   rH   __delitem__  r  zScopedDict.__delitem__)r  r  )r  r  r?   r  )r  r  ry   r  r?   r  )r  r3  r?   rm   r{   )r  r  r  r  r?   r  rC  )r?   r  r-  )r  r  r?   r  )r   r   r   r   r  r  r  r  ry  rB  r  r  r  rA   rA   rA   rH   r    s    






r  )frozen_defaultr}   Optional[type[Any]]r   c                 s"   d fdd}| d u r|S || S )Nr}   rb   r?   c                   s(   t jdkrtj| d dS tj|  dS )N)rq  r  T)kw_onlyr   r   )r  version_infodataclasses	dataclass)r}   r   rA   rH   wrap  s   
zir_dataclass.<locals>.wrap)r}   rb   r?   rb   rA   )r}   r   r  rA   r   rH   ir_dataclass  s   r  Optional[list[int]]c                  C  s&   t jj } | d ur| jr| jjS d S r{   )rC   r1  r2  r3  fw_metadatabw_donated_idxs)r6  rA   rA   rH   get_donated_idxs  s   r  3Union[Sequence[BaseSchedulerNode], ExternKernelOut]r   	is_externc                   s   ddl m}m} ddlm} ddlm} |r4t| |sJ |jj	
|g    fdd| jD  d S t| ts;J | D ]#}|||fvr`|jd ur`|jj	
|g    fdd|jjD  q=d S )Nr*   r  )r2   rU  c                 3       | ]}|j  vr|j V  qd S r{   rW  r  curr_node_inforA   rH   r         
z:set_kernel_post_grad_provenance_tracing.<locals>.<genexpr>c                 3  r  r{   rW  r  r  rA   rH   r   	  r  )codegen.simd_kernel_featuresr  r  r  r2   rX  rV  rn   r   ._inductor_triton_kernel_to_post_grad_node_info
setdefaultr~  r  rI  r  )r  r   r  r  r  r2   rV  r  rA   r  rH   'set_kernel_post_grad_provenance_tracing  s0   
r  c                   @  s    e Zd ZdZdZdZdZdZdS )TritonAttrsDescriptorVersionr   r*   r1  rq  r  N)r   r   r   V0_NO_TRITONV1_COMPILERV2_BACKENDSV3_BACKENDS_TUPLEV4_DICTrA   rA   rA   rH   r    s    r  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 )Nr  r   AttrsDescriptor)	importlibutil	find_specr  r  triton.backends.compilertriton.compiler.compilerr   r~  compilerr  r  r  )r  rA   rA   rH   #get_triton_attrs_descriptor_version  s   r  c                   C  s   t  tjkS r{   )r  r  r  rA   rA   rA   rH   triton_version_uses_attrs_dict4  rJ  r  r5   c                 C  sF   ddl m} t| |jsdS t| jtjjr!tjj	j
| jjv r!dS dS )zq
    Returns True if the node is an op that is not cudagraphable.
    Usually only custom ops have this tag.
    r*   r  FT)r  r  rn   r  r  rC   rx  ry  r   r|  rx  r~  r  rA   rA   rH   is_cudagraph_unsafe_op8  s   r  c                  C  sX   t jdd} t r*ddlm} | }|r*t j|dd}| r(t j	|| gn|} | S )NLD_LIBRARY_PATHr  r   )get_runtime_pathr  lib)
r  r  ry  r_   r  libfb.py.parutilr  r  r  pathsep)r  r  runtime_pathlib_pathrA   rA   rH   get_ld_library_pathK  s   r  c                 C  s    ddl m} t| |o| jd uS )Nr   )SubgraphPythonWrapperCodegen)torch._inductor.codegen.wrapperr  rn   partition_signatures)r  r  rA   rA   rH   #is_codegen_graph_partition_subgraphX  s   
r  c                 C  s8   ddl m} |jj| dr|jj| drtjS tjS )Nr*   rU  l        i   )	rX  rV  rY  rZ  r  statically_known_geqrC   r;  r=  )r  rV  rA   rA   rH   dtype_from_sizea  s   r  )r   r=   r   c                 C  $   | dkr
t jj S d| v rdS dS )z;
    Returns True if the device supports MKL-DNN BF16.
    r   r=   TF)rC   r   r  _is_mkldnn_bf16_supportedr   rA   rA   rH   is_mkldnn_bf16_supportedo  
   r  c                 C  r  )z;
    Returns True if the device supports MKL-DNN FP16.
    r   r=   TF)rC   r   r  _is_mkldnn_fp16_supportedr  rA   rA   rH   is_mkldnn_fp16_supported{  r  r  r+  )rg   rh   r?   rh   )rk   rl   r?   rm   )r   r   )r   r   r   rh   r   rh   r?   r   r-  )r   r   r?   r   )r   r   r?   rl   )r
  r  r  r  r?   rl   )r   r  r?   r  )r  r  r  r  r?   r  )r  r   r?   r@   )rM  rN  r?   rO  )r   r  r?   rT  )rM  r^  r?   r_  )rb  rc  r?   rm   )rp  r)   rq  rr  r?   rm   )rw  r
   ru   r  r  r  r?   r  )r;   )r   r@   r?   r  )r*   r;   )
r  r  r  r  r   rh   r   r@   r?   r   )rA   r  r  r  r;   )r  r  r  r  r   rh   r  rh   r  r   r   r@   r?   r   )r  r
   r  r@   r?   r  )r  r
   r  r   r?   r  )r  rh   r  rh   r?   rh   )rG   r  r  rh   r?   r  )rG   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  )ru   r  r  r  r?   r  r>  )r   rl   r?   r?  )rG  r@   r?   rm   )rG  r\   r  rh   r?   rK  )rR  rm   r?   rm   )r   r@   r?   rK  )r0  rl   rY  rZ  r?   rl   )r  r
   r?   rf  )ru   r
   r?   rm   )r  ro  r?   rp  )r  ro  r?   r)   )r  ro  r?   r  r,  )r  r
   r?   r
   )NNT)r  r  r  r  r  rm   r?   r  )r  r  r?   r  )r[  r(   r  r  r?   r  )r   r  r?   rh   r  r   )rH  rI  r?   rm   rC  )r\  rh   r   r   r]  r^  r?   r+   )rg  r4   rh  ri  r?   rm   )rl  r@   r?   rm   )rg  r4   ry  rm   rz  rm   r?   rm   )r  r3   r?   rm   )
rg  r4   r  rh   r  rh   ra  rh   r?   rm   )r  r@   r?   rm   )r  r  r  r  ra  r  r?   rm   )r  r  r  r  ra  r  r?   r  )r   r@   r?   r@   )r?   r  )rg  r4   r?   rm   )rg  r4   r  r  r  r3   r?   rm   )FTFN)rg  r4   r  r3   r  r3   r  rm   r  rm   r  rm   r  r^  r?   rm   )r   r  ru   r  r  r  r?   r  )r   r  r?   r-  )r   r  ru   r  r  r  r?   r   )r   r  ru   r  r  r  r?   r@   )r   r  ru   r  r  r  r?   r?  )rG  r  rH  r  r?   r  )rO  r  rP  rQ  r?   r
   )rX  r@   r?   r  )r?   r  )r  r  r?   rm   )re  rf  r?   rm   )r(  rl   r?   r  )ri  rm   ru   r
   r  r
   r?   rj  )r  r@   r?   rm   )r  r@   r?   rh   )r  r  r?   rm   )
r<  r  r  r'   r  r  rX  r@   r?   r  )r  r  r?   rm   )r  r  rb  r  r?   rm   )r  r  r?   rm   )r  r9   r?   rm   )r  r  rb  r  r?   rm   )r  r@   r  r  r  r  r?   r
   )r  r9   r  r  r  r  r  r  r  r  r?   r  )r  rh   r  rh   r?   rh   )r  ro  r?   rh   )r  r   r?   r  )r   r@   r?   r@   )r   r  r?   rm   )r   r@   r?   rm   )r   r  r?   rm   )r  rc  r  r  r  r  r  r  r  r@   r  rm   r?   rm   )r  r  r?   r  )r   r  r?   rm   )r.  r  r?   rm   )r?   r  )r   r  ru   r  r  r  r?   r8  )re  rF  r?   rG  )r  rJ  rK  r  rL  rM  r?   rJ  )rG   r  r?   r  )rN  rO  r\  r  r]  r^  r?   r_  )re  rF  rd  r  r?   r  )r   rl   r?   rm   )r  r  rk  r8   r?   r  )r   r  r?   r@   )r   r@   r?   r  )r!  r  ry   r  r?   rm   )r?   r  )r   r@   r  r&   r  r   r?   r  )r   r@   r?   r  )r   r  r?   r  )r}   r  r   rm   r?   r
   )r?   r  r.  )r  r  r   r@   r  rm   r?   r  )r?   r  )r  r5   r?   rm   )r  r-   r?   rm   )r  rh   r?   r  )r   r@   r?   rm   (S  
__future__r   r  r  r  enumr  r  rz  r  r+  r:  r  r  r  r  r(  r  r   r  r  r!  r  rV  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   r   ro   rC   torch._inductor.runtime.hintsr   torch.utils._ordered_setr   torch.utils._pytreer   OPTIMUS_EXCLUDE_POST_GRADr    r!   r"   r#   r$   r%   torch._prims_commonr&   torch.fxr'   rt  r(   torch.fx.noder)   rb  r+   r  r-   rY  r/   r  r0   r1   r2   r3   r4   r5   r6   output_coder8   r  r9   r:   rJ   r>   r   rO   torch._dynamo.device_interfacerP   torch._dynamo.utilsrQ   torch.autogradrR   torch.autograd.profiler_utilrS   (torch.fx.passes.graph_transform_observerrT   torch.fx.passes.shape_proprU   torch.utils._sympy.functionsrV   rW   rX   rY   rZ   torch.utils._sympy.symbolr[   r\   torch.utils._sympy.value_rangesr]   r^   r  r_   runtime.runtime_utilsr`   r  _IS_WINDOWS	getLoggerr   r   rb   r  r  	VarRangesr  rh   	InputTypeGPU_KERNEL_BIN_EXTSr,  ra  r  rd  ri   rj   rt   Functionrv   r  r   r   r   r   r   r  r  r  rL  rS  r]  ra  rk  rt  r  r   r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r/  r=  rF  rI  rQ  rU  rX  re  ri  rn  r{  r  r  r  r  r  r   r  r  r  r  clear_on_fresh_inductor_cacheclear_inductor_cachesfresh_inductor_cacher  r  	lru_cacher  r  r  r  r/  r:  r  rD  rU  rW  r[  rf  rk  ru  rx  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r'  r,  r1  r;  r=  r>  rF  rN  rW  r[  r^  rc  rd  rg  rh  rk  r  rt  r  r  r  r  r  r  r  r
  r  Enumr  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r	  rj  r  r  r  r*  r-  r0  r7  rE  rI  rU  r[  rP  rf  rj  rr  r{  r}  r  r  r  compiler  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  r  SUPPORTED_MKLDNN_DEVICESr  r  rA   rA   rA   rH   <module>   s~   4 $	


$
HV&
		$;/;8$  
) 
6.

@
	,	!
	
$$		'	


$
0
$
	