# 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. from typing import Optional import tensorrt as trt from .._common import default_net from ..functional import (ACT2FN, AllReduceParams, cast, concat, gemm_swiglu, is_gated_activation, low_latency_gemm_swiglu) from ..module import Module from ..quantization import QuantMode from ..quantization.functional import quantize from ..quantization.layers import FP8Linear, FP8RowLinear from .linear import ColumnLinear, RowLinear from .lora import LoraRuntimeParams from .normalization import LayerNorm def fc_gate_lora(hidden_states, lora, lora_layer_params): 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) mlp_fc_lora, mlp_gate_lora = lora(hidden_states, mlp_in_lora_params) mlp_in_result = concat([mlp_gate_lora, mlp_fc_lora], dim=mlp_fc_lora.rank() - 1) return mlp_in_result return None 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), inner_layernorm=False, eps=1e-05, is_expert=False, ): 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 in [ 'swiglu', 'gegelu' ] else ffn_hidden_size self.inner_layernorm = LayerNorm(ffn_hidden_size, dtype=dtype, eps=eps) if inner_layernorm else None self.fc = ColumnLinear(hidden_size, fc_output_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, gather_output=False) self.proj = RowLinear(ffn_hidden_size, hidden_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, is_expert=is_expert) self.hidden_size = hidden_size self.ffn_hidden_size = ffn_hidden_size self.hidden_act = hidden_act self.dtype = dtype self.bias = bias self.tp_group = tp_group self.tp_size = tp_size self.quant_mode = quant_mode self.eps = eps self.is_expert = is_expert # see optimize_model's add_lora for LoRA initialization self.lora = None def forward(self, hidden_states, lora_layer_params=None, gegelu_limit=None): if is_gated_activation(self.hidden_act): inter = self.fc(hidden_states) lora_result = fc_gate_lora(hidden_states, self.lora, lora_layer_params) if lora_result is not None: inter = inter + lora_result else: 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") inter = self.fc(hidden_states, mlp_fc_lora_params) 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") if self.hidden_act == 'gegelu': inter = ACT2FN[self.hidden_act](inter, gegelu_limit) else: inter = ACT2FN[self.hidden_act](inter) if self.inner_layernorm is not None: inter = self.inner_layernorm(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), inner_layernorm=False, eps=1e-05, is_expert=False, ): 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, inner_layernorm=inner_layernorm, eps=eps, is_expert=is_expert) self.hidden_size = hidden_size self.ffn_hidden_size = ffn_hidden_size self.tp_group = tp_group self.tp_size = tp_size self.gate = ColumnLinear(hidden_size, ffn_hidden_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, gather_output=False) def forward(self, hidden_states, lora_layer_params=None, all_reduce_params: Optional[AllReduceParams] = 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 if self.inner_layernorm is not None: intermediate = self.inner_layernorm(intermediate) output = self.proj(intermediate, lora_runtime_params=mlp_proj_lora_params, all_reduce_params=all_reduce_params) return output class FusedGatedMLP(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), inner_layernorm=False, eps=1e-05, is_expert=False, ): super().__init__() 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 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, ) self.inner_layernorm = LayerNorm(ffn_hidden_size, dtype=dtype, eps=eps) if inner_layernorm else None self.proj = RowLinear(ffn_hidden_size, hidden_size, bias=bias, dtype=dtype, tp_group=tp_group, tp_size=tp_size, is_expert=is_expert) # see optimize_model's add_lora for LoRA initialization self.lora = None def fc_gate_plugin(self, hidden_states, lora_layer_params=None): # Combine the following pattern # # SiLU(FC(x)) * Gate(x) # # into: # # SwiGLU(FusedFC(x)) if default_net( ).plugin_config.low_latency_gemm_swiglu_plugin is not None: p_dtype = default_net().plugin_config.low_latency_gemm_swiglu_plugin else: p_dtype = default_net().plugin_config.gemm_swiglu_plugin use_fp8 = p_dtype == 'fp8' assert use_fp8, "gemm_swiglu_plugin and low_latency_gemm_swiglu_plugin only supports fp8 now" 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 or mlp_gate_lora_params is not None: raise NotImplementedError( f"LoRA not yet implemented for gemm_swiglu_plugin") if self.hidden_act != 'silu': raise NotImplementedError( f"Activation {self.hidden_act} not yet implemented for gemm_swiglu_plugin" ) if self.bias: raise NotImplementedError( f"bias not yet implemented for gemm_swiglu_plugin fp8") assert isinstance( self.fused_fc, FP8Linear), "fp8 gemm_swiglu only supports fp8 weights" assert isinstance( self.proj, FP8RowLinear), "fp8 gemm_swiglu only supports fp8 weights" assert self.fused_fc.weight.shape == ( self.hidden_size, self.ffn_hidden_size * 2 // self.tp_size), "fp8 gemm_swiglu only supports (k, n) weights" scale_d0 = (self.fused_fc.weights_scaling_factor.raw_value.item() * self.fused_fc.activation_scaling_factor.raw_value.item()) scale_d1 = scale_d0 scale_output = 1.0 / self.proj.activation_scaling_factor.raw_value.item( ) activation_scaling_factor = cast( self.fused_fc.activation_scaling_factor.value, self.dtype) if hidden_states.dtype != trt.fp8: hidden_states = quantize(hidden_states, activation_scaling_factor, 'fp8') if default_net( ).plugin_config.low_latency_gemm_swiglu_plugin is not None: inter = low_latency_gemm_swiglu(hidden_states, self.fused_fc.weight.value, scale_d0, scale_d1, scale_output) else: inter = gemm_swiglu(hidden_states, self.fused_fc.weight.value, None, scale_d0, scale_d1, scale_output) return inter def fc_gate(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 inter = self.fused_fc(hidden_states) lora_result = fc_gate_lora(hidden_states, self.lora, lora_layer_params) if lora_result is not None: inter = inter + lora_result if self.hidden_act == 'silu': inter = ACT2FN['swiglu'](inter) elif self.hidden_act == 'gelu': inter = ACT2FN['geglu'](inter) else: raise NotImplementedError( f"Activation {self.hidden_act} not yet implemented for {self.__class__.__name__}." ) return inter def forward(self, hidden_states, lora_layer_params=None, all_reduce_params: Optional[AllReduceParams] = None): if default_net().plugin_config.gemm_swiglu_plugin or default_net( ).plugin_config.low_latency_gemm_swiglu_plugin: inter = self.fc_gate_plugin(hidden_states, lora_layer_params) else: inter = self.fc_gate(hidden_states, lora_layer_params) if self.inner_layernorm is not None: inter = self.inner_layernorm(inter) 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, all_reduce_params=all_reduce_params) return output