""" Copyright (c) 2025 by FlashInfer team. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. """ import functools import math from typing import Optional, Tuple, Union import torch from .api_logging import flashinfer_api from .jit import gen_batch_attention_module from .utils import ( MaskMode, PosEncodingMode, TensorLayout, _check_kv_layout, _unpack_paged_kv_cache, determine_attention_backend, ) from .prefill import BatchPrefillWithPagedKVCacheWrapper from .jit.attention.variants import attention_sink_decl from .jit.utils import filename_safe_dtype_map @functools.cache def get_holistic_attention_module(*args): return gen_batch_attention_module(*args).build_and_load() class BatchAttention: @flashinfer_api def __init__( self, kv_layout: str = "NHD", device: str = "cuda", ): _check_kv_layout(kv_layout) self._kv_layout = kv_layout self.float_workspace_buffer = torch.empty( 384 * 1024 * 1024, dtype=torch.uint8, device=torch.device(device), ) self.int_workspace_buffer = torch.empty( 8 * 1024 * 1024, dtype=torch.uint8, device=torch.device(device), ) self.page_locked_int_workspace_buffer = torch.empty( 8 * 1024 * 1024, dtype=torch.uint8, device=torch.device("cpu"), pin_memory=True, ) @flashinfer_api def plan( self, qo_indptr: torch.Tensor, kv_indptr: torch.Tensor, kv_indices: torch.Tensor, kv_len_arr: torch.Tensor, num_qo_heads: int, num_kv_heads: int, head_dim_qk: int, head_dim_vo: int, page_size: int, causal: bool = False, sm_scale: float = None, logits_soft_cap: Optional[float] = None, q_data_type: torch.dtype = torch.bfloat16, kv_data_type: torch.dtype = torch.bfloat16, use_profiler: bool = False, ) -> None: if logits_soft_cap is None: logits_soft_cap = 0.0 self._logits_soft_cap = logits_soft_cap # get jit module get_module_args = ( q_data_type, kv_data_type, q_data_type, kv_indptr.dtype, head_dim_qk, head_dim_vo, PosEncodingMode["NONE"].value, logits_soft_cap > 0.0, use_profiler, # different compiler path ) self.module = get_holistic_attention_module(*get_module_args) qo_indptr_host = qo_indptr.to(torch.device("cpu"), non_blocking=True) kv_indptr_host = kv_indptr.to(torch.device("cpu"), non_blocking=True) kv_len_arr_host = kv_len_arr.to(torch.device("cpu"), non_blocking=True) torch.cuda.synchronize() batch_size = kv_len_arr.shape[0] self._page_size = page_size self._sm_scale = sm_scale self._mask_mode = MaskMode.CAUSAL.value if causal else MaskMode.NON_CAUSAL.value self._num_qo_heads = num_qo_heads self._num_kv_heads = num_kv_heads self._page_size = page_size self._use_profiler = use_profiler # No addtional buf allocated for CUDA graph tensor # Allocate outside FlashInfer self._kv_indices = kv_indices self._plan_info = self.module.plan( self.float_workspace_buffer, self.int_workspace_buffer, self.page_locked_int_workspace_buffer, qo_indptr_host, kv_indptr_host, kv_len_arr_host, batch_size, num_qo_heads, num_kv_heads, head_dim_vo, causal, ) @flashinfer_api def run( self, q: torch.Tensor, kv_cache: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]], out: Optional[torch.Tensor] = None, lse: Optional[torch.Tensor] = None, k_scale: Optional[torch.Tensor] = None, v_scale: Optional[torch.Tensor] = None, logits_soft_cap: float = 0.0, profiler_buffer: Optional[torch.Tensor] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: if profiler_buffer is None: if self._use_profiler: raise ValueError( "Profiler is enabled, profiler_buffer must be provided" ) if logits_soft_cap > 0.0 and self._logits_soft_cap <= 0.0: raise ValueError( "logits_soft_cap used in kernel run but not provided in plan(). This will cause template deduction error." ) k_cache, v_cache = _unpack_paged_kv_cache(kv_cache, self._kv_layout) if out is None: out = torch.empty_like(q) if lse is None: # lse shape: [batch_size, num_qo_heads] lse = torch.empty( q.shape[0], q.shape[1], device=q.device, dtype=torch.float32 ) head_dim_qk = q.shape[2] sm_scale = self._sm_scale if sm_scale is None: sm_scale = 1.0 / math.sqrt(head_dim_qk) if k_scale is not None: sm_scale *= k_scale if v_scale is None: v_scale = 1.0 # profiler_buffer is optional profiler_args = (profiler_buffer,) if self._use_profiler else () self.module.run( self.float_workspace_buffer, self.int_workspace_buffer, self._plan_info, q, k_cache, v_cache, self._kv_indices, out, lse, self._mask_mode, TensorLayout[self._kv_layout].value, self._num_qo_heads, self._num_kv_heads, self._page_size, v_scale, sm_scale, logits_soft_cap, # ADDITIONAL_FUNC_PARAMS # PROFILER_FUNC_PARAMS *profiler_args, ) return out, lse class BatchAttentionWithAttentionSinkWrapper(BatchPrefillWithPagedKVCacheWrapper): r""" Wrapper for prefill and decode attention with paged KV-cache that adds support for attention sinks. This class extends `BatchPrefillWithPagedKVCacheWrapper`, providing a convenient interface for using attention sinks during prefill or decode attention. """ def __init__( self, float_workspace_buffer: torch.Tensor, kv_layout: str = "NHD", use_cuda_graph: bool = False, qo_indptr_buf: Optional[torch.Tensor] = None, paged_kv_indptr_buf: Optional[torch.Tensor] = None, paged_kv_indices_buf: Optional[torch.Tensor] = None, paged_kv_last_page_len_buf: Optional[torch.Tensor] = None, custom_mask_buf: Optional[torch.Tensor] = None, mask_indptr_buf: Optional[torch.Tensor] = None, backend: str = "auto", pos_encoding_mode: str = "NONE", use_fp16_qk_reduction: bool = False, q_data_type: torch.dtype = torch.bfloat16, kv_data_type: torch.dtype = torch.bfloat16, head_dim_qk: int = 128, head_dim_vo: int = 128, window_left: int = -1, ) -> None: # trtllm is separate code path assert backend in ["fa2", "fa3", "auto"] if backend == "auto": # dispatch backend before init jit module backend = determine_attention_backend( float_workspace_buffer.device, PosEncodingMode[pos_encoding_mode].value, use_fp16_qk_reduction, # use_fp16_qk_reduction custom_mask_buf is not None, # use_custom_mask q_data_type, kv_data_type, ) jit_args = [ f"batch_prefill_attention_sink_{filename_safe_dtype_map[q_data_type]}_swa_{window_left >= 0}_{backend}", # uri q_data_type, # dtype_q kv_data_type, # dtype_kv q_data_type, # dtype_o torch.int32, # idtype head_dim_qk, # hidden_dim_qk head_dim_vo, # hidden_dim_vo ["sink"], # additional_tensor_names ["float"], # additional_tensor_dtypes ["sm_scale"], # additional_scalar_names ["double"], # additional_scalar_dtypes "AttentionSink", attention_sink_decl[backend], ] jit_kwargs = { "use_sliding_window": window_left >= 0, "use_fp16_qk_reduction": use_fp16_qk_reduction, "pos_encoding_mode": PosEncodingMode[pos_encoding_mode].value, } super().__init__( float_workspace_buffer=float_workspace_buffer, kv_layout=kv_layout, use_cuda_graph=use_cuda_graph, qo_indptr_buf=qo_indptr_buf, paged_kv_indptr_buf=paged_kv_indptr_buf, paged_kv_indices_buf=paged_kv_indices_buf, paged_kv_last_page_len_buf=paged_kv_last_page_len_buf, custom_mask_buf=custom_mask_buf, mask_indptr_buf=mask_indptr_buf, backend=backend, jit_args=jit_args, jit_kwargs=jit_kwargs, )