Source code for aitemplate.compiler.ops.gemm_epilogue_vistor.bmm_rcr_softmax

#  Copyright (c) Meta Platforms, Inc. and affiliates.
#
#  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.
#
"""
Operator definition for bmm_rcr_softmax.
"""

from aitemplate.compiler.base import IntImm, Tensor
from aitemplate.compiler.ops.gemm_universal import gemm_common as common
from aitemplate.compiler.ops.gemm_universal.bmm import bmm
from aitemplate.compiler.tensor_accessor import TensorAccessor

# pylint: disable=C0103, W0223, W0221, W0613


[docs]class bmm_rcr_softmax(bmm): """BatchGemm with softmax, A: row_major, B: column_major, C: row_major, A: [b, m, k], B: [b, n, k], C: [b, m, n] """ def __init__(self): super().__init__() self._attrs["op"] = "bmm_rcr_softmax" def cal_align_ab(m, n, k): return common.default_align_ab(k, k, self._attrs["inputs"][0].dtype()) self._attrs["f_ab_alignment"] = cal_align_ab def _infer_shapes(self, a: Tensor, b: Tensor): a_shapes = a._attrs["shape"] b_shapes = b._attrs["shape"] batch_size_a = a_shapes[0] batch_size_b = b_shapes[0] if batch_size_a != batch_size_b and batch_size_a != 1 and batch_size_b != 1: raise RuntimeError( "bmm operand A and B should have same batch_size, or batch_size = 1! " "Current shape A: {} shape B: {} .".format(a_shapes, b_shapes) ) batch_size = batch_size_b if batch_size_a == IntImm(1) else batch_size_a return [batch_size, a_shapes[1], b_shapes[1]] def __call__(self, a: Tensor, b: Tensor) -> Tensor: """Performs sanity checks, offline shape inference and returns an output tensor.""" a, b = self._align_ab(a, b) self._attrs["inputs"] = [a, b] self._sanity_check(a, b) output_shape = self._infer_shapes(a, b) self._extract_epilogue_alignment(output_shape) self._attrs["input_accessors"] = [ TensorAccessor(tensor) for tensor in self._attrs["inputs"] ] self._set_depth() output = Tensor(output_shape, src_ops={self}, dtype=a._attrs["dtype"]) self._attrs["outputs"] = [output] self._attrs["output_accessors"] = [TensorAccessor(output)] return output def _extract_dims(self, for_profiling=False): # (B, M, K) * (B, N, K) = (B, M, N) return { "B": [common.DimInfo(common.Source.OUTPUT, tensor_idx=0, dim_idx=0)], "M": [ common.DimInfo(common.Source.INPUT, tensor_idx=0, dim_idx=1), common.DimInfo(common.Source.OUTPUT, tensor_idx=0, dim_idx=1), ], "N": [ common.DimInfo(common.Source.INPUT, tensor_idx=1, dim_idx=1), common.DimInfo(common.Source.OUTPUT, tensor_idx=0, dim_idx=2), ], "K": [ common.DimInfo(common.Source.INPUT, tensor_idx=0, dim_idx=2), common.DimInfo(common.Source.INPUT, tensor_idx=1, dim_idx=2), ], } def _invert_exec_key(self, key): """extract tensor shape info from key""" return common.gemm_inverse_key_func(key) def _gen_profile_cmd(self, profiler_prefix, cfg, exec_key): """get command line for profiling""" def fbuild_cmd(exec_key): B, M, N, K = self._invert_exec_key(exec_key) cmd = [] cmd.append(B) # b cmd.append(M) # m cmd.append(N) # n cmd.append(K) # k return cmd return super()._gen_profile_cmd(profiler_prefix, cfg, exec_key, fbuild_cmd)