
    Vhx1                        d dl Z d dlmZ d dlZd dlmc mZ d dlm	Z	 d dl
mZ ddlmZmZ ddlmZ  e j"                  e      Zdej(                  d	ej*                  d
efdZdej(                  d	ej*                  ded
dfdZdej(                  d
ej4                  fdZ eeeef             Zdej4                  d
dfdZdej4                  d
efdZ dej(                  dedefdZ!dej(                  fdZ"d Z#y)    N)cast)is_symbolic)
OrderedSet   )configir)Vxcomm_buffer_typereturnc                    t        |       }t        |t        j                        ry|j	                         }t        |t        j
                        ryt        |t        j                        rt        |j                               syy)ze
    Check if an input can be realized as a comm buffer of the specified
    `comm_buffer_type`.
    TF)		_get_data
isinstancer   Loopsget_output_specCommBufferLayoutFlexibleLayoutr   	get_numel)r
   r   datalayouts       M/home/dcms/DCMS/lib/python3.12/site-packages/torch/_inductor/comm_lowering.pycan_realize_as_comm_bufferr   8   sc     Q<D$!!!#F&"--.&"++,[AQ5R    
group_namec                    | j                          t        |       }t        |t        j                        sJ |j                         }t        |t        j                        ryt        |t        j                        st        d| d      t        |j                               rt        d| d      t        j                  |||      |_        y)z
    Realize an input as a comm buffer of the specified `comm_buffer_type`.

    Specifically, this realizes the underlying buffer if it's still unrealized
    and changes the layout of the buffer to `ir.CommBufferLayout`.
    NzOA buffer can only be realized as a comm buffer if it has `FlexibleLayout` (got ).zGA buffer with symbolic shape cannot be converted to a comm buffer (got )r   r   r   )realizer   r   r   Bufferr   r   r   AssertionErrorr   r   r   )r
   r   r   bufferr   s        r   realize_as_comm_bufferr!   N   s     IIKq\Ffbii(((##%F&"--.fb//0))/4
 	

 6##%&""(-
 	

 '')FMr   c                 h   t        | j                  t        j                        r$| j                  j	                         j                  S t        | j                  t        j
                        r.t        t        j                  | j                  j                        S t        d| j                   d      )Nz\Expect the data attr of a `TensorBox` to be either an `ir.BaseView` or `ir.StorageBox` (got r   )	r   r   r   BaseViewunwrap_view
StorageBoxr   r   r   r
   s    r   r   r   r   s{    !&&"++&vv!!#(((	AFFBMM	*BIIqvv{{++889xrC
 	
r   c                 |    t         j                  t        t        j                        | j                         f       y)z
    If a non-blocking collective is lowered as a blocking collective, the wait
    node in the original graph becomes useless and we can skip the lowering it.
    N)_bufs_to_skip_waitaddidr	   graphget_namer&   s    r   mark_as_skip_waitr-      s%    
 BqwwK67r   c                 `    t        t        j                        | j                         ft        v S N)r*   r	   r+   r,   r(   r&   s    r   should_skip_waitr0      s!    qwwK&*<<<r   inp	reduce_opc                 B   ddl m} | j                         | j                         j                  z  }t
        j                  j                  xrS  ||      xrI t        | t        j                  j                        xr# |dv xr |t
        j                  j                  k  S )Nr   )is_symm_mem_enabled_for_group)sum)#torch.distributed._symmetric_memoryr4   r   	get_dtypeitemsizer   _collectiveauto_selectr   r   CommBufferTypeSYMM_MEM#one_shot_all_reduce_threshold_bytes)r1   r2   r   r4   inp_sizes        r   $_should_lower_as_one_shot_all_reducer?      s     R}}!9!99H&& 	O)*5	O&sB,=,=,F,FG	O !	O **NNNr   c           	      @   t        | t        j                  j                  |       t	        j
                  t        j                  j                  t        j                  j                  t        j                  j                  j                  j                  | ||            S r/   )r!   r   r;   r<   pytreetree_map	TensorBoxcreateFallbackKerneltorchopssymm_memone_shot_all_reducedefault)r1   r2   r   s      r   _one_shot_all_reducerK      sn    3 1 1 : :JG??

  II22::		
 r   c            	         	 t         j                  j                  j                   ddlmm	m
mm fd} t         j                  j                   | j                        dt        j                  dt        dt        dt        j                  ffd	       } | j                         dt        j                  dt        dt        dt        j                  ffd
       } | j"                        fd       } | j$                        fd       } | j&                        fd       } | j(                        fd       } | j*                        fd       } | j,                        fd       } | j.                        fd       }	 | j0                        fd       }
 | j2                        fd       } | j4                        fd       } | t         j                  j6                  j8                        d        } | j:                        fd       }y # t        $ r t
        j                  d       Y y w xY w)NzRInductor support for distributed collectives depends on building torch.distributedr   )add_layout_constraintcloneconstrain_to_fx_stridescopy_register_loweringc                 &     |         |       S r/    )fnrM   rO   rQ   s    r   register_comm_loweringz7register_comm_lowerings.<locals>.register_comm_lowering   s    b"9: $$r   r1   r2   r   r   c                    t        | ||      rt        | ||      S  |       } t        j                  rG| j	                          t
        j                  j                  j                  | j                                t        j                  j                  |       } t        j                  j                  j                  j                   | ||       | S r/   )r?   rK   r    reorder_for_compute_comm_overlapr   r	   r+   no_fuse_buffer_namesr)   r,   r   ExternKernelrequire_contiguous_CollectiveKernelcreate_inplaceall_reduce_rJ   )r1   r2   r   c10drN   s      r   _all_reducez,register_comm_lowerings.<locals>._all_reduce   s    /Y
K'Y
CC Cj22 KKMGG((,,S\\^<oo005
++$$c9j	
 
r   c                    t        | ||      r! | t        | ||            }t        |       | S t        j                  j                  |       } t        j                  j                  j                  j                  | ||       | S r/   )
r?   rK   r-   r   rY   rZ   r[   r\   r]   rJ   )r1   r2   r   retr^   rP   s       r   _all_reduce_z-register_comm_lowerings.<locals>._all_reduce_   s     0Y
K$S)Z@C c"J oo005
++$$c9j	
 
r   c                     | D cg c]
  } |       } }t         j                  j                  j                  j                  | ||       | S c c}w r/   r   r[   r\   all_reduce_coalesced_rJ   )inputsr2   r   r1   r^   rN   s       r   _all_reduce_coalescedz6register_comm_lowerings.<locals>._all_reduce_coalesced   sR    (./%*//
++&&..		
  0s   Ac                 t    t         j                  j                  j                  j                  | ||       | S r/   rd   )rf   r2   r   r^   s      r   _all_reduce_coalesced_z7register_comm_lowerings.<locals>._all_reduce_coalesced_   s6    
++&&..		
 r   c                     t         j                  j                  t         j                  j	                  j
                  j                  | ||            S r/   )r   rC   rD   r[   create_out_of_placeall_gather_into_tensorrJ   )r1   
group_sizer   r^   s      r   _all_gather_into_tensorz8register_comm_lowerings.<locals>._all_gather_into_tensor   sF    ||""  44++33	
 	
r   c           	          t        j                  t        j                  j                  t        j
                  j                  j                  j                  | ||            S r/   )	rA   rB   r   rC   rD   r[   rk    all_gather_into_tensor_coalescedrJ   )rf   rm   r   r^   s      r   !_all_gather_into_tensor_coalescedzBregister_comm_lowerings.<locals>._all_gather_into_tensor_coalesced  sM    LL  4455==	
 	
r   c                x    t         j                  j                  j                  j                  | |||       |S )N)out)r   r[   r\   all_gather_into_tensor_outrJ   )r1   rm   r   rs   r^   s       r   _all_gather_into_tensor_outz<register_comm_lowerings.<locals>._all_gather_into_tensor_out  s>    
++++33 	, 	
 
r   c           	          t         j                  j                  t         j                  j	                  j
                  j                  | |||            S r/   )r   rC   rD   r[   rk   reduce_scatter_tensorrJ   )r1   r2   rm   r   r^   s       r   _reduce_scatter_tensorz7register_comm_lowerings.<locals>._reduce_scatter_tensor"  sI    ||""  44**22
 	
r   c           
          t        j                  t        j                  j                  t        j
                  j                  j                  j                  | |||            S r/   )	rA   rB   r   rC   rD   r[   rk   reduce_scatter_tensor_coalescedrJ   )rf   r2   rm   r   r^   s       r    _reduce_scatter_tensor_coalescedzAregister_comm_lowerings.<locals>._reduce_scatter_tensor_coalesced.  sP    LL  4444<<	
 		
r   c           	          t         j                  j                  t         j                  j	                  j
                  j                  | |||            S r/   )r   rC   rD   r[   rk   all_to_all_singlerJ   )r1   output_split_sizesinput_split_sizesr   r^   s       r   _all_to_all_singlez3register_comm_lowerings.<locals>._all_to_all_single;  sI    ||""  44&&.."!
 	
r   c                      |       } t         j                  j                  j                  j                  | ||       | S r/   r   r[   r\   
broadcast_rJ   )r1   srcr   r^   rN   s      r   
_broadcastz+register_comm_lowerings.<locals>._broadcastG  s:    Cj
++OO##S#z	
 
r   c                 t    t         j                  j                  j                  j                  | ||       | S r/   r   )r1   r   r   r^   s      r   _broadcast_z,register_comm_lowerings.<locals>._broadcast_O  s1    
++OO##S#z	
 
r   c           	          t         j                  j                  t         j                  j	                  t
        j                  j                  j                  j                  | |||            S r/   )
r   rC   rD   r[   rk   rF   rG   _dtensorshard_dim_alltoallrJ   )r1   
gather_dim	shard_dimr   s       r   _shard_dim_alltoallz4register_comm_lowerings.<locals>._shard_dim_alltoallV  sR    ||""  44		""55==
 	
r   c                     t        |       r| S t        j                  j                  j                  j
                  |        | S r/   )r0   r   _WaitKernelcreate_waitwait_tensorrJ   )r1   r^   s    r   _wait_tensorz-register_comm_lowerings.<locals>._wait_tensorb  s5    C J
""4#3#3#;#;SA
r   )rF   rG   _c10d_functional
all_reduceAttributeErrorloginfoloweringrM   rN   rO   rP   rQ   r   rC   strr]   all_reduce_coalescedre   rl   rp   rt   rw   rz   r}   	broadcastr   r   r   r   )rU   r_   rb   rg   ri   rn   rq   ru   rx   r{   r   r   r   r   r   rM   r^   rN   rO   rP   rQ   s                  @@@@@@r   register_comm_loweringsr      sf   		""-- % 99%%DDOO, # 3 2<<  -( D,,-\\&)7:	 .$ D556 7 D667 8 D778
 9
 DAAB	
 C	
 D;;< = D667	
 8	
 D@@A

 B

 D223	
 4	
 DNN+ , DOO, - EII..AAB	
 C	
 D,,- .k   	
 	s   $H< <II)$loggingtypingr   rF   torch.utils._pytreeutils_pytreerA   torch._inductor.utilsr   torch.utils._ordered_setr    r   r   virtualizedr	   	getLogger__name__r   rC   r;   boolr   r   r!   IRNoder   tupleintr(   r-   r0   r?   rK   r   rS   r   r   <module>r      s&      $ $ - /   g!T	||')'8'8	,!	||!')'8'8!FI!	!H
 
")) 
 1Zc3h02 8 8t 8=		 =d =	"%36
bll 
}r   