# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved. # # This source code is licensed under the BSD license found in the # LICENSE file in the root directory of this source tree. from dataclasses import replace from enum import Enum from functools import partial from typing import Any, Iterable, List, Optional, Set, Tuple, Union import torch from ..common import get_operator, register_operator from . import attn_bias from .attn_bias import ( AttentionBias, AttentionBiasSubTensor, BlockDiagonalCausalLocalAttentionFromBottomRightMask, BlockDiagonalCausalLocalAttentionMask, BlockDiagonalCausalMask, BlockDiagonalCausalWithOffsetPaddedKeysMask, BlockDiagonalMask, LowerTriangularFromBottomRightLocalAttentionMask, LowerTriangularFromBottomRightMask, LowerTriangularMask, LowerTriangularMaskWithTensorBias, ) from .common import ( AttentionBwOpBase, AttentionFwOpBase, Context, Gradients, Inputs, _attn_bias_apply, check_lastdim_alignment_stride1, ) from .torch_attention_compat import is_pt_cutlass_compatible def _uses_tensorcores(sm: int, is_half: bool) -> bool: if sm >= 80: return True if sm >= 70: return is_half return False def _minimum_gemm_alignment(inp: Inputs) -> int: if inp.device.type != "cuda": return 1 cap = torch.cuda.get_device_capability(inp.device) sm = cap[0] * 10 + cap[1] bits_per_scalar = {torch.float: 32, torch.half: 16, torch.bfloat16: 16}[ inp.query.dtype ] uses_tensorcores = _uses_tensorcores(sm, bits_per_scalar == 16) matmul_alignment_mn = 1 if sm >= 80: matmul_alignment_mn = 4 if uses_tensorcores: matmul_alignment_mn = max(matmul_alignment_mn, 128 // bits_per_scalar) return matmul_alignment_mn def _get_seqlen_info( inp: Inputs, ) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor], int, int]: attn_bias = inp.attn_bias if isinstance( attn_bias, (BlockDiagonalMask, BlockDiagonalCausalWithOffsetPaddedKeysMask) ): assert attn_bias.k_seqinfo.seqstart.device == inp.query.device seqstart_k = attn_bias.k_seqinfo.seqstart seqstart_q = attn_bias.q_seqinfo.seqstart max_seqlen_q = attn_bias.q_seqinfo.max_seqlen max_seqlen_k = attn_bias.k_seqinfo.max_seqlen else: seqstart_k = None seqstart_q = None max_seqlen_q = -1 max_seqlen_k = -1 return seqstart_k, seqstart_q, max_seqlen_q, max_seqlen_k def _get_tensor_bias( attn_bias: Optional[Union[torch.Tensor, AttentionBias]] ) -> Optional[torch.Tensor]: if isinstance(attn_bias, AttentionBiasSubTensor): if isinstance(attn_bias, LowerTriangularMaskWithTensorBias): return attn_bias._subtensor elif isinstance(attn_bias, torch.Tensor): return attn_bias return None def _check_bias_alignment( reasons: List[str], attn_bias: Optional[Union[torch.Tensor, AttentionBias]] ) -> None: attn_bias_tensor = _get_tensor_bias(attn_bias) if attn_bias_tensor is not None: alignment = 128 // torch.finfo(attn_bias_tensor.dtype).bits show_padding_hint = False for d in range(attn_bias_tensor.ndim - 1): if attn_bias_tensor.stride(d) % alignment != 0: reasons.append( f"attn_bias.stride(-2) % {alignment} != 0 (attn_bias.stride() = {attn_bias_tensor.stride()})" ) show_padding_hint = True if show_padding_hint: reasons.append( """\ HINT: To use an `attn_bias` with a sequence length that is not a multiple of 8, \ you need to ensure memory is aligned by slicing a bigger tensor. \ Example: use `attn_bias = torch.zeros([1, 1, 5, 8])[:,:,:,:5]` instead of `torch.zeros([1, 1, 5, 5])`""" ) # We can have stride=0 sometimes if dimension=1 if attn_bias_tensor.stride(-1) > 1: reasons.append( f"attn_bias.stride(-1) > 1 (attn_bias.stride() = {attn_bias_tensor.stride()}) - " "you should call `.contiguous()` on the bias" ) class _CustomMaskType(int, Enum): """ (Matches CustomMaskType in C++.) """ NoCustomMask = 0 CausalFromTopLeft = 1 CausalFromBottomRight = 2 def _custom_mask_type(bias: Optional[Union[torch.Tensor, AttentionBias]]) -> int: if isinstance( bias, ( LowerTriangularMask, BlockDiagonalCausalMask, BlockDiagonalCausalLocalAttentionMask, ), ): return int(_CustomMaskType.CausalFromTopLeft) if isinstance( bias, ( LowerTriangularFromBottomRightMask, LowerTriangularFromBottomRightLocalAttentionMask, attn_bias.BlockDiagonalCausalFromBottomRightMask, BlockDiagonalCausalWithOffsetPaddedKeysMask, BlockDiagonalCausalLocalAttentionFromBottomRightMask, ), ): return int(_CustomMaskType.CausalFromBottomRight) return int(_CustomMaskType.NoCustomMask) @register_operator class FwOp(AttentionFwOpBase): """xFormers' MHA kernel based on CUTLASS. Supports a large number of settings (including without TensorCores, f32 ...) and GPUs as old as P100 (Sm60) """ OPERATOR = ( get_operator("aten", "_efficient_attention_forward") if is_pt_cutlass_compatible() else None ) SUPPORTED_DEVICES: Set[str] = {"cuda"} SUPPORTED_DTYPES: Set[torch.dtype] = {torch.float, torch.half, torch.bfloat16} SUPPORTED_MAX_K = 65536 SUPPORTED_ATTN_BIAS_TYPES: Iterable[Any] = ( type(None), torch.Tensor, LowerTriangularMask, LowerTriangularFromBottomRightMask, LowerTriangularFromBottomRightLocalAttentionMask, LowerTriangularMaskWithTensorBias, BlockDiagonalMask, BlockDiagonalCausalMask, BlockDiagonalCausalWithOffsetPaddedKeysMask, attn_bias.BlockDiagonalCausalFromBottomRightMask, attn_bias.BlockDiagonalCausalLocalAttentionMask, BlockDiagonalCausalLocalAttentionFromBottomRightMask, ) SUPPORTS_DROPOUT = True SUPPORTS_CUSTOM_SCALE = True SUPPORTS_DIFFERENT_VALUE_EMBED = True SUPPORTS_BMGHK = True VARLEN_LSE_PACKED = False NAME = "cutlassF-pt" _TEST_K: List[int] = [ 32, # 64x64 kernel 128, # 64x128 kernel 256, # 64x128 with accumulation in gmem ] @classmethod def apply( cls, inp: Inputs, needs_gradient: bool ) -> Tuple[torch.Tensor, Optional[Context]]: if type(inp.attn_bias) not in FwOp.SUPPORTED_ATTN_BIAS_TYPES: raise NotImplementedError("Unsupported attn_bias type") if inp.query.ndim in [3, 4]: return cls.apply_bmhk(inp, needs_gradient=needs_gradient) assert inp.query.ndim == 5, f"query has shape {inp.query.shape}" ctx: Optional[Context] = None # XXX: Hackfix for BMGHK with H=1 # In that case we don't want to run G different streams because it adds # some overhead if inp.query.ndim == 5 and inp.query.shape[3] == 1: slice_op = partial(torch.squeeze, dim=3) inp = replace( inp, query=slice_op(inp.query), key=slice_op(inp.key), value=slice_op(inp.value), attn_bias=_attn_bias_apply( inp.attn_bias, partial(torch.squeeze, dim=2) ), ) out, ctx = cls.apply_bmhk(inp, needs_gradient=needs_gradient) out = out.unsqueeze(3) if ctx is not None: ctx = replace(ctx, lse=ctx.lse.unsqueeze(1), out=out) return out, ctx # Workaround until this is properly implemented in C++ # run each head group in a different stream n_groups = inp.key.shape[2] main_stream = torch.cuda.current_stream() streams = [main_stream] + [ torch.cuda.Stream(device=inp.query.device) for _ in range(n_groups - 1) ] outs = [] for group, stream in enumerate(streams): stream.wait_stream(main_stream) with torch.cuda.stream(stream): query = inp.query[:, :, group] key = inp.key[:, :, group] value = inp.value[:, :, group] bias = _attn_bias_apply( inp.attn_bias, partial(torch.select, dim=1, index=group) ) outs.append( cls.apply_bmhk( replace(inp, query=query, key=key, value=value, attn_bias=bias), needs_gradient=needs_gradient, ) ) for s in streams[1:]: main_stream.wait_stream(s) out = torch.stack([o[0] for o in outs], dim=2) if needs_gradient: ctx = Context( out=out, lse=torch.stack([o[1].lse for o in outs], dim=1), # type: ignore op_bw=outs[0][1].op_bw, # type: ignore ) return out, ctx @classmethod def apply_bmhk( cls, inp: Inputs, needs_gradient: bool ) -> Tuple[torch.Tensor, Optional[Context]]: if type(inp.attn_bias) not in FwOp.SUPPORTED_ATTN_BIAS_TYPES: raise NotImplementedError("Unsupported attn_bias type") seqstart_k, seqstart_q, max_seqlen_q, max_seqlen_k = _get_seqlen_info(inp) out, lse, rng_seed, rng_offset, _, _ = cls.OPERATOR( query=inp.query, key=inp.key, value=inp.value, bias=_get_tensor_bias(inp.attn_bias), cu_seqlens_q=seqstart_q, cu_seqlens_k=seqstart_k, max_seqlen_q=max_seqlen_q, max_seqlen_k=max_seqlen_k, dropout_p=inp.p, compute_log_sumexp=needs_gradient, custom_mask_type=_custom_mask_type(inp.attn_bias), scale=inp.scale, seqlen_k=( inp.attn_bias.k_seqinfo.seqlen if isinstance( inp.attn_bias, BlockDiagonalCausalWithOffsetPaddedKeysMask ) else None ), window_size=( inp.attn_bias._window_size if isinstance( inp.attn_bias, ( BlockDiagonalCausalLocalAttentionMask, BlockDiagonalCausalLocalAttentionFromBottomRightMask, LowerTriangularFromBottomRightLocalAttentionMask, ), ) else None ), ) ctx: Optional[Context] = None if needs_gradient: ctx = Context(out=out, lse=lse) if inp.p != 0: # cutlass forward is only compatible with cutlass backward if # dropout is used (because of the way RNG states are passed and the # way random numbers are generated during backward) ctx.rng_state = (rng_seed, rng_offset) ctx.op_bw = BwOp return out, ctx @classmethod def not_supported_reasons(cls, d: Inputs) -> List[str]: reasons = super(FwOp, cls).not_supported_reasons(d) matmul_alignment_mn = _minimum_gemm_alignment(d) check_lastdim_alignment_stride1(reasons, "query", d.query, matmul_alignment_mn) check_lastdim_alignment_stride1(reasons, "value", d.value, matmul_alignment_mn) _check_bias_alignment(reasons, d.attn_bias) return reasons @register_operator class BwOp(AttentionBwOpBase): __doc__ = FwOp.__doc__ OPERATOR = ( get_operator("aten", "_efficient_attention_backward") if is_pt_cutlass_compatible() else None ) SUPPORTED_DEVICES = FwOp.SUPPORTED_DEVICES SUPPORTED_DTYPES = FwOp.SUPPORTED_DTYPES SUPPORTED_MAX_K = FwOp.SUPPORTED_MAX_K SUPPORTED_ATTN_BIAS_TYPES: Iterable[Any] = ( type(None), torch.Tensor, LowerTriangularMask, LowerTriangularFromBottomRightMask, # TODO: Still some infs/nans in the BW pass for # local + causal # LowerTriangularFromBottomRightLocalAttentionMask, # TODO: Fix handling of gradient through the fMHA autograd function # LowerTriangularMaskWithTensorBias, BlockDiagonalMask, BlockDiagonalCausalMask, attn_bias.BlockDiagonalCausalFromBottomRightMask, attn_bias.BlockDiagonalCausalLocalAttentionMask, ) SUPPORTS_ATTN_BIAS_GRAD = True SUPPORTS_DROPOUT = FwOp.SUPPORTS_DROPOUT SUPPORTS_CUSTOM_SCALE = FwOp.SUPPORTS_CUSTOM_SCALE SUPPORTS_DIFFERENT_VALUE_EMBED = FwOp.SUPPORTS_DIFFERENT_VALUE_EMBED VARLEN_LSE_PACKED = False NAME = "cutlassB-pt" _TEST_K: List[int] = [ 32, # 64x64 kernel 128, # 64x128/128x128 kernel 256, # 64x128 with accumulation in gmem ] @classmethod def not_supported_reasons(cls, d: Inputs) -> List[str]: reasons = super(BwOp, cls).not_supported_reasons(d) matmul_alignment_mn = _minimum_gemm_alignment(d) check_lastdim_alignment_stride1(reasons, "query", d.query, matmul_alignment_mn) check_lastdim_alignment_stride1(reasons, "key", d.key, matmul_alignment_mn) check_lastdim_alignment_stride1(reasons, "value", d.value, matmul_alignment_mn) _check_bias_alignment(reasons, d.attn_bias) attn_bias_tensor = _get_tensor_bias(d.attn_bias) # Backprop of gradient through broadcasted bias is not supported if attn_bias_tensor is not None and attn_bias_tensor.requires_grad: # Don't forget that inputs are either in BMK or BMHK! if d.query.ndim == 3 and attn_bias_tensor.ndim == 3: expected_bias_shape = (*d.query.shape[:2], d.key.shape[1]) else: # bias is B H Mq Mk expected_bias_shape = ( d.query.shape[0], d.query.shape[2] if d.query.ndim == 4 else 1, d.query.shape[1], d.key.shape[1], ) if tuple(attn_bias_tensor.shape) != expected_bias_shape: reasons.append( "Broadcasting the `attn_bias` tensor is not supported " f"(shape: {tuple(attn_bias_tensor.shape)}" f"/ expected: {expected_bias_shape})" ) return reasons @classmethod def apply(cls, ctx: Context, inp: Inputs, grad: torch.Tensor) -> Gradients: if type(inp.attn_bias) not in BwOp.SUPPORTED_ATTN_BIAS_TYPES: raise NotImplementedError("Unsupported attn_bias type") seqstart_k, seqstart_q, max_seqlen_q, max_seqlen_k = _get_seqlen_info(inp) dtype = inp.query.dtype rng_seed = rng_offset = torch.Tensor() if inp.p != 0.0: assert ctx.rng_state is not None rng_seed, rng_offset = ctx.rng_state tensor_bias = _get_tensor_bias(inp.attn_bias) force_pad_inf = torch.cuda.get_device_capability(inp.query.device) == (7, 5) (grad_q, grad_k, grad_v, grad_bias) = cls.OPERATOR( grad.to(dtype), inp.query, inp.key, inp.value, bias=tensor_bias, bias_requires_grad=( tensor_bias.requires_grad if tensor_bias is not None else False ), cu_seqlens_q=seqstart_q, cu_seqlens_k=seqstart_k, max_seqlen_q=max_seqlen_q, max_seqlen_k=max_seqlen_k, logsumexp=ctx.get_padded_lse(32, force_pad_inf=force_pad_inf), out=ctx.out.to(dtype), dropout_p=inp.p, # if not using dropout, seed and offset are irrelevant but still expected # in function signature so just pass 0 # seed and offset could be None if a different FW op other than cutlass # was used. philox_seed=rng_seed, philox_offset=rng_offset, custom_mask_type=_custom_mask_type(inp.attn_bias), scale=inp.scale, num_splits_key=None, # Let C++ determine it window_size=( inp.attn_bias._window_size if isinstance( inp.attn_bias, ( BlockDiagonalCausalLocalAttentionMask, BlockDiagonalCausalLocalAttentionFromBottomRightMask, LowerTriangularFromBottomRightLocalAttentionMask, ), ) else None ), ) # c++/CUDA implementation returns an uninitialized tensor if bias doesn't # require grad if not ( isinstance(inp.attn_bias, torch.Tensor) and inp.attn_bias.requires_grad ): grad_bias = None return Gradients(dq=grad_q, dk=grad_k, dv=grad_v, db=grad_bias)