3i\1#dZddlZddlmZddlmZmZddlZeje Z ddgZ de e dzde e fd Zed d e defd ZGddeZej$j'di d0dej(dej(dej(dej(dej(de de dededzde e dzdeej(ej(ej(ffdZej0 d0dej(dej(dej(dej(dej(de de dededzde e dzdeej(ej(ej(ffdZdddddej(dej(dej(dej(dej(de de dedzdedzdee e fdej(eej(ej(fzfdZd ed!eed"fd#eddfd$Zej$j'd%i d1d&ej(dej(dej(dej(d'ej(d(ej(dej(dej(de de ded)ej(dedzde e dzdeej(ej(ej(ffd*Zej0 d1d&ej(dej(dej(dej(d'ej(d(ej(dej(dej(de de ded)ej(dedzde e dzdeej(ej(ej(ffd+Zd ed&ej(d,ej(d-ej(deej(dzd"ff d.Zej?ee/y)2z Variable-length attention implementation using Flash Attention. This module provides a high-level Python interface for variable-length attention that calls into the optimized Flash Attention kernels. N) lru_cache)Any NamedTuple varlen_attn AuxRequest window_sizereturnc\|ddg}t|dk7rtdt||S)Nz$window_size must have length 2, got )len ValueError)rs [/var/www/pod-logistic/pod-ai/venv/lib/python3.12/site-packages/torch/nn/attention/varlen.py_normalize_window_sizers=2h  ;1?K@P?QRSS )maxsize device_indexcy)z;Cache device capability check to avoid repeated CUDA calls.F)rs r_should_use_cudnnrs rc eZdZUdZdZeed<y)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_argsquerykeyvaluecu_seq_qcu_seq_kmax_qmax_k is_causalscalec zt| } |jxrt|jj} | rvt j d| ddk7s| ddk7r tdtjjj|||d||||dd|d | } | d| d| d }} } nWt j d tjjj|||||||d|d || d| d \} } }}}tjdtj|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$r%r&r'r(r)r use_cudnnresultoutput softmax_lse rng_state_ rng_state_s r _varlen_attnrL-sW$)5K G"3ELL4F4F"GI 67 q>R ;q>R#7f 88            9  *0F1IvayY  @A/4yy~~/V/V         #(^)!n0W0 , Y1  ELLJ ; **rc t| } tj|} |jd} |jd} tjj rH|jddz } tj | | |ftj|j}n2tj | | ftj|j}tj dtj|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()5K  e $FjjmG 1 I }}]]1%) KK E *%++ell KK   ELL  D U\\JI 9i ''r)r r ) return_auxr)rrYc| dk(} tjjj|||||||| |t |  \} } } ||j r| | 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 rrrsn\w&I))&&33       [ KCa*..Cx Jrctxinputs.rGc |\ }}}}}}} } } } |\} }}|j|||||| ||||_| |_| |_| |_| |_yN)save_for_backwardr&r'r(r)r)r^r_rGr!r"r#r$r%r&r'r(r)rr]rrIs r_setup_contextrcsu         Ci%eXxc9UCICICMCI!COrz!torch_attn::_varlen_attn_backwardgrad_outr]rrIcNt| } tjd|j}|jxrt |jj }|rmtjd| ddk7s| ddk7r tdtjjj|||||||||| d| | || \}}}nYtjd tjjj|||||||||| d| | || | d| d \}}}|||fS) Nr)r8r+r r,r-r.r/r1)r)r3r4)rr>rRr8r9rr:r;r<r=r?r@_cudnn_attention_backward_flash_attention_backward)rdr!r"r#r]rr$r%r&r'r(rIr)runusedrEdqdkdvs r_varlen_attn_backwardrlsD")5K [[5<< 0F G"3ELL4F4F"GI 67 q>R ;q>R#7f YY^^==              > B$ @AYY^^==              (^)!n#> B& r2:rct| } tj|}tj|}tj|}|||fS)zF Fake implementation for meta tensor computation and tracing. )rr>rN)rdr!r"r#r]rr$r%r&r'r(rIr)r grad_querygrad_key grad_values r_varlen_attn_backward_fakerq\sK()5K!!%(J$H!!%(J x ++rgrad_lsegrad_rngc0|j\}}}}}} } } |j} |j} |j}|j}|j }t jjj||||| | ||| | || ||\}}}|||dddddddf Sra) saved_tensorsr&r'r(r)rr>r?r[rl)r^rdrrrsr!r"r#r$r%r]rrIr&r'r(r)rrirjrks r _backwardrvysBEARAR>E3x3Y IIE IIE I IIE//K%%;;        JBB r2tT4tT4 ??r) setup_context)FNN)NN) rlogging functoolsrtypingrrr> getLoggerrr;__all__r\intrrrrlibrary custom_opTensorrStuplerL register_fakerXrrcrlrqrvregister_autogradrrrrs" g! , 'S D(8T#Y 1CD 3"E$(B+ <<B+ B+ <<B+ll B+ ll B+  B+ B+B+ 4<B+cT!B+ 5<<u|| 34B+FB+J$((( <<(( (( <<((ll (( ll ((  (( (((( 4<((cT!(( 5<<u|| 34((((h%)#+] <<] ] <<]ll ] ll ]  ] ]T!] 4<]sCx] \\E%,, 455]@""U38_"c"d"0<2N$(AllA <<A A << A  A  AllAllA A AA||A 4<AcT!A 5<<u|| 34AOAH$$$(,ll, <<, , << ,  ,  ,ll,ll, , ,,||, 4<,cT!, 5<<u|| 34,%,8@ @ @05 @HM @ 5<<$  #$@< yGr