# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # SPDX-License-Identifier: Apache-2.0 # # 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 numpy as np from .._utils import trt_dtype_to_np from ..functional import ACT2FN, concat from ..module import Module from ..quantization import QuantMode from ..quantization.layers import FP8Linear, FP8RowLinear from .linear import ColumnLinear, RowLinear from .lora import Lora, LoraRuntimeParams class MLP(Module): def __init__( self, hidden_size, ffn_hidden_size, hidden_act, bias=True, dtype=None, tp_group=None, tp_size=1, quant_mode=QuantMode(0), max_lora_rank=None, ): super().__init__() if hidden_act not in ACT2FN: raise ValueError( 'unsupported activation function: {}'.format(hidden_act)) fc_output_size = 2 * ffn_hidden_size if hidden_act == 'swiglu' else ffn_hidden_size self.use_fp8_qdq = quant_mode.has_fp8_qdq() if self.use_fp8_qdq: self.fc = FP8Linear(hidden_size, fc_output_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, gather_output=False, max_lora_rank=max_lora_rank) self.proj = FP8RowLinear(ffn_hidden_size, hidden_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, max_lora_rank=max_lora_rank) else: self.fc = ColumnLinear(hidden_size, fc_output_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, gather_output=False, max_lora_rank=max_lora_rank) self.proj = RowLinear(ffn_hidden_size, hidden_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, max_lora_rank=max_lora_rank) self.hidden_act = hidden_act self.dtype = dtype self.bias = bias def forward(self, hidden_states, lora_layer_params=None): mlp_fc_lora_params = None if lora_layer_params is not None: mlp_fc_lora_params = lora_layer_params.get_runtime_params( 0, "mlp_h_to_4h") mlp_proj_lora_params = None if lora_layer_params is not None: mlp_proj_lora_params = lora_layer_params.get_runtime_params( 0, "mlp_4h_to_h") inter = self.fc(hidden_states, mlp_fc_lora_params) inter = ACT2FN[self.hidden_act](inter) output = self.proj(inter, lora_runtime_params=mlp_proj_lora_params) return output class GatedMLP(MLP): def __init__( self, hidden_size, ffn_hidden_size, hidden_act, bias=True, dtype=None, tp_group=None, tp_size=1, quant_mode=QuantMode(0), max_lora_rank=None, ): super().__init__(hidden_size, ffn_hidden_size, hidden_act, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, quant_mode=quant_mode) self.hidden_size = hidden_size self.ffn_hidden_size = ffn_hidden_size self.hidden_act = hidden_act self.bias = bias self.dtype = dtype self.tp_group = tp_group self.tp_size = tp_size self.quant_mode = quant_mode if self.use_fp8_qdq: self.gate = FP8Linear(hidden_size, ffn_hidden_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, gather_output=False, max_lora_rank=max_lora_rank) else: self.gate = ColumnLinear(hidden_size, ffn_hidden_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, gather_output=False, max_lora_rank=max_lora_rank) def forward(self, hidden_states, lora_layer_params=None): mlp_fc_lora_params = None if lora_layer_params is not None: mlp_fc_lora_params = lora_layer_params.get_runtime_params( 0, "mlp_h_to_4h") mlp_gate_lora_params = None if lora_layer_params is not None: mlp_gate_lora_params = lora_layer_params.get_runtime_params( 0, "mlp_gate") mlp_proj_lora_params = None if lora_layer_params is not None: mlp_proj_lora_params = lora_layer_params.get_runtime_params( 0, "mlp_4h_to_h") inter = self.fc(hidden_states, mlp_fc_lora_params) inter = ACT2FN[self.hidden_act](inter) gate = self.gate(hidden_states, mlp_gate_lora_params) intermediate = inter * gate output = self.proj(intermediate, lora_runtime_params=mlp_proj_lora_params) return output class FusedGatedMLP(GatedMLP): def __init__( self, hidden_size, ffn_hidden_size, hidden_act, bias=True, dtype=None, tp_group=None, tp_size=1, quant_mode=QuantMode(0), max_lora_rank=None, ): super().__init__(hidden_size, ffn_hidden_size, hidden_act, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, quant_mode=quant_mode) self.is_weight_rewritten = False if self.use_fp8_qdq: self.fused_fc = FP8Linear( self.hidden_size, self.ffn_hidden_size * 2, bias=self.bias, dtype=self.dtype, tp_group=self.tp_group, tp_size=self.tp_size, gather_output=False, ) else: self.fused_fc = ColumnLinear( self.hidden_size, self.ffn_hidden_size * 2, bias=self.bias, dtype=self.dtype, tp_group=self.tp_group, tp_size=self.tp_size, gather_output=False, ) if max_lora_rank is None: max_lora_rank = min(hidden_size, ffn_hidden_size // tp_size) self.mlp_in_lora = Lora( in_hidden_size=hidden_size, out_hidden_sizes=[ ffn_hidden_size // tp_size, ffn_hidden_size // tp_size ], max_low_rank=max_lora_rank, ) def forward(self, hidden_states, lora_layer_params=None): # Combine the following pattern # # SiLU(FC(x)) + Gate(x) # # into: # # SwiGLU(FusedFC(x)) # # Upside is we don't need to modify 4 different weight loading paths just to concat weights if not self.is_weight_rewritten: _np_dtype = trt_dtype_to_np(self.dtype) if self.gate.weight.is_inited() or self.fc.weight.is_inited(): self.fused_fc.weight.value = np.concatenate( [self.gate.weight.raw_value, self.fc.weight.raw_value], axis=0).astype(_np_dtype) if self.bias and (self.gate.bias.is_inited() or self.fc.bias.is_inited()): self.fused_fc.bias.value = np.concatenate( [self.gate.bias.raw_value, self.fc.bias.raw_value], axis=0).astype(_np_dtype) if self.use_fp8_qdq: if self.gate.weights_scaling_factor.is_inited( ) or self.fc.weights_scaling_factor.is_inited(): # TODO: need to align with quantization toolkit; preferably put a constraint to equalize # fc/gate weight scaling factor to allow horizontal fusion without accuracy loss self.fused_fc.weights_scaling_factor.value = max( self.gate.weights_scaling_factor.raw_value, self.fc.weights_scaling_factor.raw_value, ) if self.gate.activation_scaling_factor.is_inited( ) or self.fc.activation_scaling_factor.is_inited(): assert self.fc.activation_scaling_factor.raw_value == self.gate.activation_scaling_factor.raw_value, "Activation scales should be identical" self.fused_fc.activation_scaling_factor.value = self.fc.activation_scaling_factor.raw_value del self._modules["gate"] del self._modules["fc"] self.is_weight_rewritten = True inter = self.fused_fc(hidden_states) if lora_layer_params is not None: mlp_fc_lora_params = lora_layer_params.get_runtime_params( 0, "mlp_h_to_4h") mlp_gate_lora_params = lora_layer_params.get_runtime_params( 0, "mlp_gate") if mlp_fc_lora_params is not None and mlp_gate_lora_params is not None: mlp_in_lora_params = LoraRuntimeParams( lora_ranks=[ mlp_fc_lora_params.lora_ranks[0], mlp_gate_lora_params.lora_ranks[0] ], lora_weights_pointers=[ mlp_fc_lora_params.lora_weights_pointers[0], mlp_gate_lora_params.lora_weights_pointers[0] ], host_request_types=mlp_fc_lora_params.host_request_types, host_context_lengths=mlp_fc_lora_params. host_context_lengths, max_context_length=mlp_fc_lora_params.max_context_length) mlp_fc_lora, mlp_gate_lora = self.mlp_in_lora( hidden_states, mlp_in_lora_params) mlp_in_result = concat([mlp_gate_lora, mlp_fc_lora], dim=mlp_fc_lora.rank() - 1) inter = inter + mlp_in_result if self.hidden_act == 'silu': inter = ACT2FN['swiglu'](inter) else: raise NotImplementedError( f"Activation {self.hidden_act} not yet implemented for FusedGatedMLP" ) mlp_proj_lora_params = None if lora_layer_params is not None: mlp_proj_lora_params = lora_layer_params.get_runtime_params( 0, "mlp_4h_to_h") output = self.proj(inter, lora_runtime_params=mlp_proj_lora_params) return output