aj\1#LdZddlZddlmZddlmZmZddlZeje Z ddgZ de e dzde e fd Zed d e defd ZGddeZejdi d1dejdejdejdejdejde de dededzde e dzdeejejejffdZej d1dejdejdejdejdejde de dededzde e dzdeejejejffdZdddddejdejdejdejdejde de dedzdedzdee e fdejeejejfzfd Zd!ed"eed#fd$eddfd%Zejd&i d2d'ejdejdejdejd(ejd)ejdejdejde de ded*ejdedzde e dzdeejejejffd+Zej d2d'ejdejdejdejd(ejd)ejdejdejde de ded*ejdedzde e dzdeejejejffd,Zd!ed'ejd-ejd.ejdeejdzd#ff d/Zeee0dS)3z Variable-length attention implementation using Flash Attention. This module provides a high-level Python interface for variable-length attention that calls into the optimized Flash Attention kernels. N) lru_cache)Any NamedTuple varlen_attn AuxRequest window_sizereturncv|ddg}t|dkrtdt||S)Nz$window_size must have length 2, got )len ValueError)rs ^/home/longshao/multi-rider-rag/.venv/lib/python3.11/site-packages/torch/nn/attention/varlen.py_normalize_window_sizersI2h  ;1RK@P@PRRSSS )maxsize device_indexcdS)z;Cache device capability check to avoid repeated CUDA calls.F)rs r_should_use_cudnnrs  5rc"eZdZUdZdZeed<dS)rz Request which auxiliary outputs to compute from varlen_attn. Each field is a boolean indicating whether that auxiliary output should be computed. FlseN)__name__ __module__ __qualname____doc__rbool__annotations__rrrrr#s. Crztorch_attn::_varlen_attn) mutates_argsFquerykeyvaluecu_seq_qcu_seq_kmax_qmax_k is_causalscalec t| } |jot|jj} | rt d| ddks | ddkrtdtj j |||d||||dd|d | } | d| d| d }} } n]t d tj j |||||||d|d || d| d \} } }}}tj dtj|j}| | |fS)z Private custom op for variable-length attention. This is the internal implementation. Users should use the public varlen_attn function instead. #Using cuDNN backend for varlen_attnrr TcuDNN backend does not support window attention. Please use Flash Attention backend.NTFr)-Using Flash Attention backend for varlen_attn)return_debug_maskr)window_size_leftwindow_size_rightr dtypedevice)ris_cudarr8indexloginfo RuntimeErrortorchopsaten_cudnn_attention_forward_flash_attention_forwardzerosuint64)r!r"r#r$r%r&r'r(r)r use_cudnnresultoutput softmax_lse rng_state_ rng_state_s r _varlen_attnrL-sz$)55K G"3EL4F"G"GI(  6777 q>R  ;q>R#7#7f 88            9   *0F1IvayY  @AAA/4y~/V/V         #(^)!n0W0 0 , Y1  ELJ ; **rc t| } tj|} |d} |d} tjjrB|ddz } tj| | |ftj|j}n(tj| | ftj|j}tjdtj |j}| ||fS)z Fake implementation for meta tensor computation and tracing. Based on the 3D varlen path from meta__flash_attention_forward: - query shape: (total, num_heads, head_dim) - logsumexp shape: (num_heads, total_q) rr,r6r5) rr> empty_likesizeversionhipemptyfloatr8rD)r!r"r#r$r%r&r'r(r)rrGtotal_q num_heads batch_size logsumexprIs r_varlen_attn_fakerXss()55K e $ $FjjmmG 1 I }  ]]1%%) K E *%+el   K   EL    D U\JJJI 9i ''r)r r ) return_auxr)rrYc| dk} tjj|||||||| |t |  \} } } | |jr| | fS| S)au Compute variable-length attention using Flash Attention. This function is similar to scaled_dot_product_attention but optimized for variable-length sequences using cumulative sequence position tensors. Args: query (Tensor): Query tensor; shape :math:`(T_q, H, D)` key (Tensor): Key tensor; shape :math:`(T_k, H, D)` value (Tensor): Value tensor; shape :math:`(T_k, H, D)` cu_seq_q (Tensor): Cumulative sequence positions for queries; shape :math:`(N+1,)` cu_seq_k (Tensor): Cumulative sequence positions for keys/values; shape :math:`(N+1,)` max_q (int): Maximum query sequence length in the batch. max_k (int): Maximum key/value sequence length in the batch. return_aux (Optional[AuxRequest]): If not None and ``return_aux.lse`` is True, also returns the logsumexp tensor. scale (float, optional): Scaling factor for attention scores window_size (tuple[int, int], optional): Window size for sliding window attention as (left, right). Use (-1, -1) for full attention (default), (-1, 0) for causal attention, or (W, 0) for causal attention with sliding window of size W. Returns: output (Tensor): Output tensor from attention computation; shape :math:`(T_q, H, D)`. If ``return_aux`` is not None and ``return_aux.lse`` is True: lse (Tensor): Log-sum-exp of attention scores; shape :math:`(T_q, H)`. Shape legend: - :math:`N`: Batch size - :math:`T_q`: Total number of query tokens in the batch (sum of all query sequence lengths) - :math:`T_k`: Total number of key/value tokens in the batch (sum of all key/value sequence lengths) - :math:`H`: Number of attention heads - :math:`D`: Head dimension Example:: >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CUDA) >>> batch_size, max_seq_len, embed_dim, num_heads = 2, 512, 1024, 16 >>> head_dim = embed_dim // num_heads >>> seq_lengths = [] >>> for _ in range(batch_size): ... length = torch.randint(1, max_seq_len // 64 + 1, (1,)).item() * 64 ... seq_lengths.append(min(length, max_seq_len)) >>> seq_lengths = torch.tensor(seq_lengths, device="cuda") >>> total_tokens = seq_lengths.sum().item() >>> >>> # Create packed query, key, value tensors >>> query = torch.randn( ... total_tokens, num_heads, head_dim, dtype=torch.float16, device="cuda" ... ) >>> key = torch.randn( ... total_tokens, num_heads, head_dim, dtype=torch.float16, device="cuda" ... ) >>> value = torch.randn( ... total_tokens, num_heads, head_dim, dtype=torch.float16, device="cuda" ... ) >>> >>> # Build cumulative sequence tensor >>> cu_seq = torch.zeros(batch_size + 1, device="cuda", dtype=torch.int32) >>> cu_seq[1:] = seq_lengths.cumsum(0) >>> max_len = seq_lengths.max().item() >>> >>> # Call varlen_attn >>> output = varlen_attn( ... query, key, value, cu_seq, cu_seq, max_len, max_len ... ) )r r)r>r? torch_attnrLlistr)r!r"r#r$r%r&r'rYr)rr(outrrJs rrrsr\w&I)&33       [  KCa*.Cx Jrctxinputs.rGc |\ }}}}}}} } } } |\} }}||||||| ||||_| |_| |_| |_| |_dSN)save_for_backwardr&r'r(r)r)r^r_rGr!r"r#r$r%r&r'r(r)rr]rrIs r_setup_contextrcs         Ci%eXxc9UUUCICICMCI!COOOrz!torch_attn::_varlen_attn_backwardgrad_outr]rrIcTt| } tjd|j}|jot |jj}|rytd| ddks | ddkrtdtj j |||||||||| d| | || \}}}n_td tj j |||||||||| d| | || | d| d \}}}|||fS) Nr)r8r+r r,r-r.r/r1)r)r3r4)rr>rRr8r9rr:r;r<r=r?r@_cudnn_attention_backward_flash_attention_backward)rdr!r"r#r]rr$r%r&r'r(rIr)runusedrEdqdkdvs r_varlen_attn_backwardrlse")55K [5< 0 0 0F G"3EL4F"G"GI+  6777 q>R  ;q>R#7#7f Y^==              >  B$ @AAAY^==              (^)!n#>  B& r2:rct| } tj|}tj|}tj|}|||fS)zF Fake implementation for meta tensor computation and tracing. )rr>rN)rdr!r"r#r]rr$r%r&r'r(rIr)r grad_querygrad_key grad_values r_varlen_attn_backward_fakerq\sN()55K!%((J$$H!%((J x ++rgrad_lsegrad_rngc|j\}}}}}} } } |j} |j} |j}|j}|j}t jj ||||| | ||| | || ||\}}}|||dddddddf Sra) saved_tensorsr&r'r(r)rr>r?r[rl)r^rdrrrsr!r"r#r$r%r]rrIr&r'r(r)rrirjrks r _backwardrvysBEAR>E3x3Y IE IE I IE/K%;;        JBB r2tT4tT4 ??r) setup_context)FNN)NN) rlogging functoolsrtypingrrr> getLoggerrr;__all__r\intrrrrlibrary custom_opTensorrStuplerL register_fakerXrrcrlrqrvregister_autogradrrrrsO"""""""" g!! , 'S D(8T#Y 1CD 3"EE$(B+B+ <B+ B+ <B+l B+ l B+  B+ B+B+ 4<B+cT!B+ 5<u| 34B+B+B+FEB+J$((((( <(( (( <((l (( l ((  (( (((( 4<((cT!(( 5<u| 34((((((((h%)#+]]] <] ] <]l ] l ]  ] ]T!] 4<]sCx] \E%, 455]]]]@""U38_"c"d""""0<2NN$(AAlA <A A < A  A  AlAlA A AA|A 4<AcT!A 5<u| 34AAAONAH$$(,,l, <, , < ,  ,  ,l,l, , ,,|, 4<,cT!, 5<u| 34,,,%$,8@ @ @05 @HM @ 5<$  #$@@@@< yGGGGGr