o iu'@sdZddlZddlmZddlmZmZddlZee Z ddgZ eddd e d e fd d ZGd ddeZejjdid d3dejdejdejdejdejde de de d eejejejffddZej d3dejdejdejdejdejde de de d eejejejffddZ  d4dejdejdejdejdejde de de dedBd ejeejejfBfddZded eed!fd"ed dfd#d$Zejjd%idd&ejdejdejdejd'ejd(ejdejdejde de de d)ejd eejejejffd*d+Zejd&ejdejdejdejd'ejd(ejdejdejde de de d)ejd eejejejffd,d-Zded&ejd.ejd/ejd eejdBd!ff d0d1Zejeed2dS)5z Variable-length attention implementation using Flash Attention. This module provides a high-level Python interface for variable-length attention that calls into the optimized Flash Attention kernels. N) lru_cache)Any NamedTuple varlen_attn AuxRequest)maxsize device_indexreturncCsdS)z;Cache device capability check to avoid repeated CUDA calls.F)r r r [/sda-disk/www/egybert/egybert_env/lib/python3.10/site-packages/torch/nn/attention/varlen.py_should_use_cudnnsr c@seZdZUdZdZeed<dS)rz Request which auxiliary outputs to compute from varlen_attn. Each field is a boolean indicating whether that auxiliary output should be computed. FlseN)__name__ __module__ __qualname____doc__rbool__annotations__r r r r rs ztorch_attn::_varlen_attn) mutates_argsFquerykeyvaluecu_seq_qcu_seq_kmax_qmax_k is_causalcCs|jot|jj}|r1tdtjj |||d||||dd|d } | d| d| d} } } ntd tjjj |||||||d|dd \} } } } } tj d tj |jd }| | |fS) z Private custom op for variable-length attention. This is the internal implementation. Users should use the public varlen_attn function instead. #Using cuDNN backend for varlen_attnNTFr-Using Flash Attention backend for varlen_attn)return_debug_maskdtypedevice) is_cudar r(indexloginfotorchopsaten_cudnn_attention_forward_flash_attention_forwardzerosuint64)rrrrrrrr use_cudnnresultoutput softmax_lse rng_state_ rng_state_r r r _varlen_attn$sF    r;c CsTt|}|d} |d} tj| | ftj|jd} tjdtj|jd} || | 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 r&r$)r- empty_likesizeemptyfloatr(r3) rrrrrrrrr6total_q num_heads logsumexpr8r r r _varlen_attn_fake^s    rC return_auxc Cs<tjj||||||||\} } } |dur|jr| | fS| S)a9 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. - is_causal (bool, optional): If set to True, applies causal masking (default: False). - return_aux (Optional[AuxRequest]): If not None and ``return_aux.lse`` is True, also returns the logsumexp tensor. 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 Returns: - Tensor: Output tensor from attention computation - If ``return_aux`` is not None and ``return_aux.lse`` is True, returns a tuple of Tensors: (output, lse), where lse is the logsumexp 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, is_causal=False ... ) N)r-r. torch_attnr;r) rrrrrrrrrDoutrr9r r r rs G ctxinputs.r6c CsL|\}}}}}}} } |\} } } ||||||| | | ||_| |_| |_dSN)save_for_backwardrrr)rGrHr6rrrrrrrrrFrr8r r r _setup_contexts   rKz!torch_attn::_varlen_attn_backwardgrad_outrFrr8c Cstjd|jd} |jot|jj} | r0tdtjj |||||||||| d| | | \}}}ntdtjj |||||||||| d| | | \}}}|||fS)Nr)r(rrr") r-r>r(r)r r*r+r,r.r/_cudnn_attention_backward_flash_attention_backward)rLrrrrFrrrrrrr8unusedr4dqdkdvr r r _varlen_attn_backwardsL   rSc Cs(t|} t|} t|}| | |fS)zF Fake implementation for meta tensor computation and tracing. )r-r<)rLrrrrFrrrrrrr8 grad_querygrad_key grad_valuer r r _varlen_attn_backward_fakes    rWgrad_lsegrad_rngcCsh|j\}}}}}} } } |j} |j} |j}tjj||||| | ||| | || \}}}|||ddddddf SrI) saved_tensorsrrrr-r.rErS)rGrLrXrYrrrrrrFrr8rrrrPrQrRr r r _backward,s& r[) setup_context)F)FN)rlogging functoolsrtypingrrr- getLoggerrr+__all__intrr rlibrary custom_opTensortupler; register_fakerCrrKrSrWr[register_autogradr r r r s      9  (  "O      8