# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project # Copyright (c) 2023, Tri Dao. # ruff: noqa: E501 import torch # isort: off # We need to import the CUDA kernels after importing torch # Use relative import to support build-from-source installation in vLLM try: from . import _vllm_fa2_C # type: ignore[attr-defined] # noqa: F401 FA2_UNAVAILABLE_REASON = None FA2_AVAILABLE = True except ImportError as e: FA2_UNAVAILABLE_REASON = str(e) FA2_AVAILABLE = False try: from . import _vllm_fa3_C # type: ignore[attr-defined] # noqa: F401 FA3_UNAVAILABLE_REASON = None FA3_AVAILABLE = True except ImportError as e: FA3_UNAVAILABLE_REASON = str(e) FA3_AVAILABLE = False try: import os _cute_interface_path = os.path.join( os.path.dirname(__file__), "cute", "interface.py" ) if not os.path.exists(_cute_interface_path): raise ImportError("vllm.vllm_flash_attn.cute.interface not found") FA4_UNAVAILABLE_REASON = None FA4_AVAILABLE = True except (ImportError, ModuleNotFoundError) as e: FA4_UNAVAILABLE_REASON = str(e) FA4_AVAILABLE = False # isort: on DEFAULT_FA_VERSION = 2 def _is_fa2_supported() -> tuple[bool, str | None]: if not FA2_AVAILABLE: return False, f"FA2 is unavailable due to: {FA2_UNAVAILABLE_REASON}" from vllm.platforms import current_platform if not current_platform.has_device_capability(80): return False, "FA2 is only supported on devices with compute capability >= 8" return True, None def _is_fa3_supported() -> tuple[bool, str | None]: if not FA3_AVAILABLE: return False, f"FA3 is unavailable due to: {FA3_UNAVAILABLE_REASON}" from vllm.platforms import current_platform if not current_platform.is_device_capability_family(90): return False, "FA3 is only supported on devices with compute capability 9.x" return True, None def _is_fa4_supported() -> tuple[bool, str | None]: if not FA4_AVAILABLE: return False, f"FA4 is unavailable due to: {FA4_UNAVAILABLE_REASON}" from vllm.platforms import current_platform if not ( current_platform.is_device_capability_family(90) or current_platform.is_device_capability_family(100) or current_platform.is_device_capability_family(110) ): return ( False, "FA4 is only supported on devices with compute capability 9.x, 10.x, or 11.x", ) return True, None def is_fa_version_supported(fa_version: int) -> bool: if fa_version == 2: return _is_fa2_supported()[0] elif fa_version == 3: return _is_fa3_supported()[0] elif fa_version == 4: return _is_fa4_supported()[0] else: raise ValueError(f"Unsupported FA version: {fa_version}") def fa_version_unsupported_reason(fa_version: int) -> str | None: if fa_version == 2: return _is_fa2_supported()[1] elif fa_version == 3: return _is_fa3_supported()[1] elif fa_version == 4: return _is_fa4_supported()[1] else: raise ValueError(f"Unsupported FA version: {fa_version}") # # For vLLM we only care about `flash_attn_varlen_func` and # `flash_attn_with_kvcache` so we only maintain wrappers for these two. # def maybe_contiguous(x): return x.contiguous() if x is not None and x.stride(-1) != 1 else x # NOTE only used in FA3 def get_scheduler_metadata( batch_size, max_seqlen_q, max_seqlen_k, num_heads_q, num_heads_kv, headdim, cache_seqlens: torch.Tensor, qkv_dtype=torch.bfloat16, headdim_v=None, cu_seqlens_q: torch.Tensor | None = None, cu_seqlens_k_new: torch.Tensor | None = None, cache_leftpad: torch.Tensor | None = None, page_size: int | None = None, max_seqlen_k_new=0, causal=False, window_size=(-1, -1), # -1 means infinite context window has_softcap=False, num_splits=0, # Can be tuned for speed pack_gqa=None, # Can be tuned for speed sm_margin=0, # Can be tuned if some SMs are used for communication ): cache_seqlens = maybe_contiguous(cache_seqlens) if headdim_v is None: headdim_v = headdim scheduler_metadata = torch.ops._vllm_fa3_C.get_scheduler_metadata( batch_size, max_seqlen_q, max_seqlen_k, num_heads_q, num_heads_kv, headdim, headdim_v, qkv_dtype, cache_seqlens, cu_seqlens_q, None, # cu_seqlens_k cu_seqlens_k_new, None, # seqused_q cache_leftpad, page_size, max_seqlen_k_new, causal, window_size[0], window_size[1], has_softcap, num_splits, pack_gqa, sm_margin, ) return scheduler_metadata def flash_attn_varlen_func( q, k, v, max_seqlen_q, cu_seqlens_q, max_seqlen_k, cu_seqlens_k=None, # only used for non-paged prefill seqused_k=None, q_v=None, dropout_p=0.0, softmax_scale=None, causal=False, window_size: list[int] | None = None, softcap=0.0, # 0.0 means deactivated alibi_slopes=None, deterministic=False, return_attn_probs=False, block_table=None, return_softmax_lse=False, out=None, # FA3 Only scheduler_metadata=None, q_descale=None, k_descale=None, v_descale=None, num_splits: int = 0, # FA4 Only output_scale=None, # Version selector fa_version: int = DEFAULT_FA_VERSION, s_aux=None, cp_world_size=1, cp_rank=0, cp_tot_seqused_k=None, # FA4 only mask_mod=None, block_sparse_tensors=None, aux_tensors=None, aux_tensor_leading_dims=None, dynamic_causal: "torch.Tensor | None" = None, gather_kv_indices=None, gather_kv_valid_length=None, ): """dropout_p should be set to 0.0 during evaluation Supports multi-query and grouped-query attention (MQA/GQA) by passing in K, V with fewer heads than Q. Note that the number of heads in Q must be divisible by the number of heads in KV. For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head 0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V. If causal=True, the causal mask is aligned to the bottom right corner of the attention matrix. For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is: 1 1 1 1 0 1 1 1 1 1 If seqlen_q = 5 and seqlen_k = 2, the causal mask is: 0 0 0 0 0 0 1 0 1 1 If the row of the mask is all zero, the output will be zero. If window_size != (-1, -1), implements sliding window local attention. Query at position i will only attend to keys between [i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive. Arguments: q: (total_q, nheads, headdim), where total_q = total number of query tokens in the batch. k: (total_k, nheads_k, headdim), where total_k = total number of key tokens in the batch. v: (total_k, nheads_k, headdim), where total_k = total number of key tokens in the batch. cu_seqlens_q: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths of the sequences in the batch, used to index into q. cu_seqlens_k: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths of the sequences in the batch, used to index into kv. max_seqlen_q: int. Maximum query sequence length in the batch. max_seqlen_k: int. Maximum key sequence length in the batch. dropout_p: float. Dropout probability. softmax_scale: float. The scaling of QK^T before applying softmax. Default to 1 / sqrt(headdim). causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling). window_size: (left, right). If not (-1, -1), implements sliding window local attention. softcap: float. Anything > 0 activates softcapping attention. alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i + seqlen_k - seqlen_q - j|) is added to the attention score of query i and key j. deterministic: bool. Whether to use the deterministic implementation of the backward pass, which is slightly slower and uses more memory. The forward pass is always deterministic. return_attn_probs: bool. Whether to return the attention probabilities. This option is for testing only. The returned probabilities are not guaranteed to be correct (they might not have the right scaling). seqused_k: (batch_size,), dtype torch.int32. Actual number of key tokens per sequence. Mutually exclusive with cu_seqlens_k except on FA4, which reads cu_seqlens_k as the KV offset and seqused_k as the length. Required when block_table is given. q_v: optional value tensor fused with q. FA3 and FA4 only. block_table: (batch_size, max_blocks_per_seq), dtype torch.int32. Paged KV cache block table. Requires seqused_k. return_softmax_lse: bool. Whether to also return softmax_lse. out: optional tensor to write the output into. scheduler_metadata: precomputed scheduling metadata. Not supported by FA2. q_descale: descale factor for FP8 q. Not supported by FA2, and ignored by FA4 when the inputs are not FP8. k_descale: descale factor for FP8 k. Same constraints as q_descale. v_descale: descale factor for FP8 v. Same constraints as q_descale. num_splits: int. Number of splits for the split-KV kernel; 0 lets the kernel choose. FA2 only supports 0 or 1. output_scale: scale for fused FP8 output quantization. FA4 only. fa_version: int. FlashAttention major version to dispatch to. s_aux: learnable per-head attention sink logits. Not supported by FA2. cp_world_size: int. Context parallel world size. cp_rank: int. Rank of this shard within the context parallel group. cp_tot_seqused_k: (batch_size,), dtype torch.int32. Key lengths across the whole context parallel group. mask_mod: optional callable applying a custom mask. Not supported by FA2. block_sparse_tensors: block-sparse index tensors. FA4 only, and must materialize full_block_cnt and full_block_idx. aux_tensors: auxiliary tensors consumed by mask_mod. FA4 only. aux_tensor_leading_dims: leading dimensions of each entry in aux_tensors. dynamic_causal: optional per-sequence causal offsets. FA4 only. gather_kv_indices: (total_q, k) int32 KV rows per query token, `-1` padded, k a multiple of 128. Requires q_v, excludes block_table. FA4 only. gather_kv_valid_length: (total_q,) int32, real entries per row. FA4 only. Return: out: (total, nheads, headdim). softmax_lse [optional, if return_softmax_lse=True]: (nheads, total_q_seqlen), or (total_q, nheads) on FA4's q_v path. The logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax normalization factor). """ assert cu_seqlens_k is not None or seqused_k is not None, ( "cu_seqlens_k or seqused_k must be provided" ) assert fa_version == 4 or cu_seqlens_k is None or seqused_k is None, ( "cu_seqlens_k and seqused_k cannot be provided at the same time" ) assert block_table is None or seqused_k is not None, ( "seqused_k must be provided if block_table is provided" ) assert output_scale is None or fa_version == 4, ( f"Fused FP8 output (output_scale) is only supported by FA4, " f"got fa_version={fa_version}" ) assert fa_version == 4 or ( gather_kv_indices is None and gather_kv_valid_length is None ), f"gather_kv_* is only supported by FA4, got fa_version={fa_version}" if softmax_scale is None: softmax_scale = q.shape[-1] ** (-0.5) # custom op does not support non-tuple input real_window_size: tuple[int, int] if window_size is None: real_window_size = (-1, -1) else: assert len(window_size) == 2 real_window_size = (window_size[0], window_size[1]) q, k, v = [maybe_contiguous(x) for x in (q, k, v)] dummy_cu_seqlens_k = torch.empty_like(cu_seqlens_q) if fa_version == 2: if ( scheduler_metadata is not None and q_descale is not None and k_descale is not None and v_descale is not None ): raise NotImplementedError( "FA2 does not support scheduler_metadata, q_descale, " "k_descale, v_descale" ) if s_aux is not None: raise NotImplementedError("FA2 does not support s_aux") if num_splits > 1: raise NotImplementedError("FA2 does not support num_splits > 1") if mask_mod is not None: raise NotImplementedError("FA2 does not support mask_mod") if aux_tensors is not None: raise NotImplementedError("FA2 does not support aux_tensors") out, softmax_lse = torch.ops._vllm_fa2_C.varlen_fwd( q, k, v, out, cu_seqlens_q, # cu_seqlens_k not used since we use seqused_k, but flash_api.cpp # still wants it so we pass all zeros dummy_cu_seqlens_k if cu_seqlens_k is None else cu_seqlens_k, seqused_k, None, block_table, alibi_slopes, max_seqlen_q, max_seqlen_k, dropout_p, softmax_scale, False, causal, real_window_size[0], real_window_size[1], softcap, return_softmax_lse and dropout_p > 0, num_splits, None, ) elif fa_version == 3: assert alibi_slopes is None, "Alibi is not supported in FA3" if mask_mod is not None: raise NotImplementedError("FA3 does not support mask_mod") if aux_tensors is not None: raise NotImplementedError("FA3 does not support aux_tensors") out, softmax_lse, _, _ = torch.ops._vllm_fa3_C.fwd( q, k, v, None, None, # k_new, v_new q_v, out, cu_seqlens_q, cu_seqlens_k, # cu_seqlens_k None, # cu_seqlens_k_new None, seqused_k, # seqused_q, seqused_k max_seqlen_q, max_seqlen_k, block_table, None, # kv_batch_idx None, # leftpad_k None, None, None, # rotary_cos, rotary_sin, seqlens_rotary q_descale, k_descale, v_descale, softmax_scale, causal, real_window_size[0], real_window_size[1], softcap, True, # rotary_interleaved scheduler_metadata, num_splits, None, # pack_gqa 0, # sm_margin s_aux, # s_aux cp_world_size, cp_rank, cp_tot_seqused_k, ) elif fa_version == 4: assert alibi_slopes is None, "Alibi is not supported in FA4" if block_sparse_tensors is not None: assert block_sparse_tensors.full_block_cnt is not None, ( "FA4 block_sparse_tensors must materialize empty full_block_cnt " "instead of passing None" ) assert block_sparse_tensors.full_block_idx is not None, ( "FA4 block_sparse_tensors must materialize empty full_block_idx " "instead of passing None" ) # FA4 only accepts descales for FP8 inputs. The attention backends # initialize these tensors even when the inputs are not quantized. if v.dtype not in (torch.float8_e4m3fn, torch.float8_e5m2): q_descale = k_descale = v_descale = None from vllm.vllm_flash_attn.cute.interface import _flash_attn_fwd # SM90 FA4 fp8-KV path: fp8 e4m3 paged K/V dequantized and the K/V descale folded # in-kernel; accepts bf16/fp16 Q and writes O in its native dtype (no Q cast, no # output copy). Only the (batch, num_kv_heads) f32 K/V descales are forwarded. fa4_fp8_kv_dequant = ( k.dtype == torch.float8_e4m3fn and torch.cuda.get_device_capability()[0] == 9 ) out, softmax_lse, _, _ = _flash_attn_fwd( q, k, v, qv=q_v, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, seqused_k=seqused_k, max_seqlen_q=max_seqlen_q, max_seqlen_k=max_seqlen_k, page_table=block_table, gather_kv_indices=gather_kv_indices, gather_kv_valid_length=gather_kv_valid_length, softmax_scale=softmax_scale, causal=causal, dynamic_causal=dynamic_causal, softcap=softcap, window_size_left=real_window_size[0] if real_window_size[0] >= 0 else None, window_size_right=real_window_size[1] if real_window_size[1] >= 0 else None, num_splits=num_splits, return_lse=return_softmax_lse, out=out, learnable_sink=s_aux, mask_mod=mask_mod, block_sparse_tensors=block_sparse_tensors, aux_tensors=aux_tensors, aux_tensor_leading_dims=aux_tensor_leading_dims, q_descale=None if fa4_fp8_kv_dequant else q_descale, k_descale=k_descale, v_descale=v_descale, output_scale=output_scale, fp8_kv_dequant=fa4_fp8_kv_dequant, ) else: raise ValueError(f"Unsupported FA version: {fa_version}") return (out, softmax_lse) if return_softmax_lse else out def compile_flash_attn_varlen_func_from_specs( *, q_shape: tuple[int, ...], k_shape: tuple[int, ...], v_shape: tuple[int, ...], q_dtype: torch.dtype, v_stride: tuple[int, ...] | None = None, k_stride: tuple[int, ...] | None = None, cu_seqlens_q_shape: tuple[int, ...] | None = None, cu_seqlens_k_shape: tuple[int, ...] | None = None, max_seqlen_q: int | None = None, max_seqlen_k: int | None = None, dropout_p: float = 0.0, softmax_scale=None, causal=False, window_size: list[int] | None = None, deterministic=False, return_softmax_lse=False, num_splits: int = 0, fa_version: int = DEFAULT_FA_VERSION, seqused_k_shape: tuple[int, ...] | None = None, page_table_shape: tuple[int, ...] | None = None, softcap: float | None = None, ) -> None: if fa_version != 4: raise ValueError( f"Compile-only FlashAttention is only supported for FA4, got FA{fa_version}" ) if dropout_p != 0.0: raise NotImplementedError("FA4 compile-only wrapper does not support dropout") del deterministic from vllm.vllm_flash_attn.cute.interface import ( _flash_attn_fwd, _make_compile_only_tensor_spec, ) real_window_size: tuple[int, int] if window_size is None: real_window_size = (-1, -1) else: assert len(window_size) == 2 real_window_size = (window_size[0], window_size[1]) if softmax_scale is None: softmax_scale = q_shape[-1] ** (-0.5) q = _make_compile_only_tensor_spec(q_shape, q_dtype) k = _make_compile_only_tensor_spec(k_shape, q_dtype, stride=k_stride) v = _make_compile_only_tensor_spec(v_shape, q_dtype, stride=v_stride) cu_seqlens_q = _make_compile_only_tensor_spec(cu_seqlens_q_shape, torch.int32, 4) out = _make_compile_only_tensor_spec( (*q_shape[:-1], v_shape[-1]), q_dtype, ) lse = None if return_softmax_lse: assert q is not None if cu_seqlens_q_shape is None: lse_shape = (*q.shape[:-3], q.shape[-2], q.shape[-3]) lse_stride = None else: lse_shape = (q.shape[-2], q.shape[0]) lse_stride = (None, 1) lse = _make_compile_only_tensor_spec( lse_shape, torch.float32, 4, stride=lse_stride, ) _flash_attn_fwd( q=q, k=k, v=v, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=_make_compile_only_tensor_spec(cu_seqlens_k_shape, torch.int32, 4), seqused_k=_make_compile_only_tensor_spec(seqused_k_shape, torch.int32, 4), page_table=_make_compile_only_tensor_spec(page_table_shape, torch.int32, 4), max_seqlen_q=max_seqlen_q, max_seqlen_k=max_seqlen_k, softmax_scale=softmax_scale, causal=causal, softcap=softcap, window_size_left=real_window_size[0], window_size_right=real_window_size[1], num_splits=num_splits, return_lse=return_softmax_lse, out=out, lse=lse, compile_only=True, ) def sparse_attn_func( q, k, v, block_count, block_offset, column_count, column_index, dropout_p=0.0, softmax_scale=None, causal=False, softcap=0.0, # 0.0 means deactivated alibi_slopes=None, deterministic=False, return_attn_probs=False, *, return_softmax_lse=False, out=None, ): """Compute attention with vertical and slash sparsity patterns. Most Arguments are the same with the flash_attn_func interface, except for 4 extra args: block_count and block_offset for slash sparsity patterns, and column_count and column_index for vertical sparsity patterns. For more details please refer to Appendix C.4.2 of paper https://arxiv.org/abs/2407.02490. Arguments: q: (batch_size, seqlen, nheads, headdim) k: (batch_size, seqlen, nheads_k, headdim) v: (batch_size, seqlen, nheads_k, headdim) block_count: (batch_size, nheads, cdiv(seqlen, BLOCK_M)) block_offset: (batch_size, nheads, cdiv(seqlen, BLOCK_M), NNZ_S) column_count: (batch_size, nheads, cdiv(seqlen, BLOCK_M)) column_index: (batch_size, nheads, cdiv(seqlen, BLOCK_M), NNZ_V) dropout_p: float. Dropout probability. softmax_scale: float. The scaling of QK^T before applying softmax. Default to 1 / sqrt(headdim). causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling). alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i + seqlen_k - seqlen_q - j|) is added to the attention score of query i and key j. deterministic: bool. Whether to use the deterministic implementation of the backward pass, which is slightly slower and uses more memory. The forward pass is always deterministic. return_attn_probs: bool. Whether to return the attention probabilities. This option is for testing only. The returned probabilities are not guaranteed to be correct (they might not have the right scaling). softcap: float. Anything > 0 activates softcapping attention. return_softmax_lse: bool. Whether to also return softmax_lse. out: optional tensor to write the output into. Return: out: (batch_size, seqlen, nheads, headdim). softmax_lse [optional, if return_softmax_lse=True]: (batch_size, nheads, seqlen). The logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax normalization factor). """ if softmax_scale is None: softmax_scale = q.shape[-1] ** (-0.5) q, k, v = [maybe_contiguous(x) for x in (q, k, v)] out, softmax_lse = torch.ops._vllm_fa2_C.fwd_sparse( q, k, v, block_count, block_offset, column_count, column_index, out, alibi_slopes, dropout_p, softmax_scale, causal, softcap, return_attn_probs and dropout_p > 0, None, ) return (out, softmax_lse) if return_softmax_lse else out def sparse_attn_varlen_func( q, k, v, block_count, block_offset, column_count, column_index, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, dropout_p=0.0, softmax_scale=None, causal=False, softcap=0.0, # 0.0 means deactivated alibi_slopes=None, deterministic=False, return_attn_probs=False, *, return_softmax_lse=False, out=None, ): """Compute attention with vertical and slash sparsity patterns. Most Arguments are the same with the flash_attn_varlen_func interface, except for 4 extra args: block_count and block_offset for slash sparsity patterns, and column_count and column_index for vertical sparsity patterns. For more details please refer to Appendix C.4.2 of paper https://arxiv.org/abs/2407.02490. Arguments: q: (total_q, nheads, headdim), where total_q = total number of query tokens in the batch. k: (total_k, nheads_k, headdim), where total_k = total number of key tokens in the batch. v: (total_k, nheads_k, headdim), where total_k = total number of key tokens in the batch. block_count: (batch_size, nheads, cdiv(seqlen, BLOCK_M)) block_offset: (batch_size, nheads, cdiv(seqlen, BLOCK_M), NNZ_S) column_count: (batch_size, nheads, cdiv(seqlen, BLOCK_M)) column_index: (batch_size, nheads, cdiv(seqlen, BLOCK_M), NNZ_V) cu_seqlens_q: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths of the sequences in the batch, used to index into q. cu_seqlens_k: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths of the sequences in the batch, used to index into kv. max_seqlen_q: int. Maximum query sequence length in the batch. max_seqlen_k: int. Maximum key sequence length in the batch. dropout_p: float. Dropout probability. softmax_scale: float. The scaling of QK^T before applying softmax. Default to 1 / sqrt(headdim). causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling). softcap: float. Anything > 0 activates softcapping attention. alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i + seqlen_k - seqlen_q - j|) is added to the attention score of query i and key j. deterministic: bool. Whether to use the deterministic implementation of the backward pass, which is slightly slower and uses more memory. The forward pass is always deterministic. return_attn_probs: bool. Whether to return the attention probabilities. This option is for testing only. The returned probabilities are not guaranteed to be correct (they might not have the right scaling). return_softmax_lse: bool. Whether to also return softmax_lse. out: optional tensor to write the output into. Return: out: (total, nheads, headdim). softmax_lse [optional, if return_softmax_lse=True]: (nheads, total_q_seqlen). The logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax normalization factor). """ if softmax_scale is None: softmax_scale = q.shape[-1] ** (-0.5) q, k, v = [maybe_contiguous(x) for x in (q, k, v)] out, softmax_lse = torch.ops._vllm_fa2_C.varlen_fwd_sparse( q, k, v, block_count, block_offset, column_count, column_index, out, cu_seqlens_q, cu_seqlens_k, None, alibi_slopes, max_seqlen_q, max_seqlen_k, dropout_p, softmax_scale, False, causal, softcap, return_attn_probs and dropout_p > 0, None, ) return (out, softmax_lse) if return_softmax_lse else out