# 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 math from collections import OrderedDict import numpy as np import tensorrt as trt from ..._utils import str_dtype_to_trt, trt_dtype_to_str from ...functional import (Tensor, allgather, arange, chunk, concat, constant, cos, exp, expand, shape, silu, sin, slice, split, unsqueeze) from ...layers import MLP, BertAttention, Conv2d, Embedding, LayerNorm, Linear from ...mapping import Mapping from ...module import Module, ModuleList from ...parameter import Parameter from ...plugin import current_all_reduce_helper from ...quantization import QuantMode from ..modeling_utils import PretrainedConfig, PretrainedModel def modulate(x, shift, scale, dtype): ones = 1.0 if dtype is not None: ones = constant(np.ones(1, dtype=np.float32)).cast(dtype) return x * (ones + unsqueeze(scale, 1)) + unsqueeze(shift, 1) class TimestepEmbedder(Module): def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None): super().__init__() self.dtype = dtype self.mlp1 = Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype) self.mlp2 = Linear(hidden_size, hidden_size, bias=True, dtype=dtype) self.frequency_embedding_size = frequency_embedding_size def timestep_embedding(self, t, dim, max_period=10000): half = dim // 2 freqs = exp( -math.log(max_period) * arange(start=0, end=half, dtype=trt_dtype_to_str(trt.float32)) / constant(np.array([half], dtype=np.float32))) args = unsqueeze(t, -1).cast(trt.float32) * unsqueeze(freqs, 0) embedding = concat([cos(args), sin(args)], dim=-1) if self.dtype is not None: embedding = embedding.cast(self.dtype) assert dim % 2 == 0 return embedding def forward(self, t): t_freq = self.timestep_embedding(t, self.frequency_embedding_size) t_emb = self.mlp2(silu(self.mlp1(t_freq))) return t_emb class LabelEmbedder(Module): def __init__(self, num_classes, hidden_size, dropout_prob, dtype=None): super().__init__() use_cfg_embedding = dropout_prob > 0 self.embedding_table = Embedding(num_classes + use_cfg_embedding, hidden_size, dtype=dtype) self.num_classes = num_classes self.dropout_prob = dropout_prob def forward(self, labels, force_drop_ids=None): assert force_drop_ids is None embeddings = self.embedding_table(labels) return embeddings class PatchEmbed(Module): def __init__(self, img_size: int, patch_size: int, input_c: int, output_c: int, bias: bool = True, dtype: trt.DataType = None): super().__init__() self.img_size = img_size self.patch_size = patch_size self.num_patches = (img_size // patch_size)**2 self.proj = Conv2d(input_c, output_c, kernel_size=(patch_size, patch_size), stride=(patch_size, patch_size), bias=bias, dtype=dtype) def forward(self, x): assert x.shape[2] == self.img_size assert x.shape[3] == self.img_size x = self.proj(x) x = x.flatten(2).transpose(1, 2) # NCHW -> NLC return x class DiTBlock(Module): def __init__(self, hidden_size, num_heads, mapping=Mapping(), mlp_ratio=4.0, dtype=None, quant_mode=QuantMode(0)): super().__init__() self.dtype = dtype self.norm1 = LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) self.attn = BertAttention(hidden_size, num_heads, tp_group=mapping.tp_group, tp_size=mapping.tp_size, tp_rank=mapping.tp_rank, cp_group=mapping.cp_group, cp_size=mapping.cp_size, cp_rank=mapping.cp_rank, dtype=dtype, quant_mode=quant_mode) self.norm2 = LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) self.mlp = MLP(hidden_size=hidden_size, ffn_hidden_size=int(hidden_size * mlp_ratio), hidden_act='gelu', tp_group=mapping.tp_group, tp_size=mapping.tp_size, dtype=dtype, quant_mode=quant_mode) self.adaLN_modulation = Linear(hidden_size, 6 * hidden_size, tp_group=mapping.tp_group, tp_size=mapping.tp_size, bias=True, dtype=dtype) def forward(self, x, c, input_lengths): c = self.adaLN_modulation(silu(c)) shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = chunk( c, 6, dim=1) x = x + unsqueeze(gate_msa, 1) * self.attn(modulate( self.norm1(x), shift_msa, scale_msa, self.dtype), input_lengths=input_lengths) x = x + unsqueeze(gate_mlp, 1) * self.mlp( modulate(self.norm2(x), shift_mlp, scale_mlp, self.dtype)) return x class FinalLayer(Module): def __init__(self, hidden_size, patch_size, out_channels, mapping=Mapping(), dtype=None): super().__init__() self.dtype = dtype self.norm_final = LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) self.linear = Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype) self.adaLN_modulation = Linear(hidden_size, 2 * hidden_size, tp_group=mapping.tp_group, tp_size=mapping.tp_size, bias=True, dtype=dtype) def forward(self, x, c): shift, scale = chunk(self.adaLN_modulation(silu(c)), 2, dim=1) x = modulate(self.norm_final(x), shift, scale, self.dtype) x = self.linear(x) return x class DiT(PretrainedModel): def __init__(self, config: PretrainedConfig): self.check_config(config) super().__init__(config) self.learn_sigma = config.learn_sigma self.in_channels = config.in_channels self.out_channels = config.in_channels * 2 if config.learn_sigma else config.in_channels self.input_size = config.input_size self.patch_size = config.patch_size self.num_heads = config.num_attention_heads self.dtype = str_dtype_to_trt(config.dtype) self.cfg_scale = config.cfg_scale self.mapping = config.mapping self.x_embedder = PatchEmbed(config.input_size, config.patch_size, config.in_channels, config.hidden_size, bias=True, dtype=self.dtype) self.t_embedder = TimestepEmbedder(config.hidden_size, dtype=self.dtype) self.y_embedder = LabelEmbedder(config.num_classes, config.hidden_size, config.class_dropout_prob, dtype=self.dtype) num_patches = self.x_embedder.num_patches self.pos_embed = Parameter(shape=(1, num_patches, config.hidden_size), dtype=self.dtype) self.blocks = ModuleList([ DiTBlock(config.hidden_size, config.num_attention_heads, mlp_ratio=config.mlp_ratio, mapping=config.mapping, dtype=self.dtype, quant_mode=config.quant_mode) for _ in range(config.num_hidden_layers) ]) self.final_layer = FinalLayer(config.hidden_size, config.patch_size, self.out_channels, mapping=config.mapping, dtype=self.dtype) # We need to invoke default `__post_init__()` for quantized layers. # def __post_init__(self): # return def check_config(self, config: PretrainedConfig): config.set_if_not_exist('input_size', 32) config.set_if_not_exist('patch_size', 2) config.set_if_not_exist('in_channels', 4) config.set_if_not_exist('mlp_ratio', 4.0) config.set_if_not_exist('class_dropout_prob', 0.1) config.set_if_not_exist('num_classes', 1000) config.set_if_not_exist('learn_sigma', True) config.set_if_not_exist('dtype', None) config.set_if_not_exist('cfg_scale', None) def unpatchify(self, x: Tensor): c = self.out_channels p = self.x_embedder.patch_size h = w = int(x.shape[1]**0.5) assert h * w == x.shape[1] x = x.view(shape=(x.shape[0], h, w, p, p, c)) x = x.permute((0, 5, 1, 3, 2, 4)) imgs = x.view(shape=(x.shape[0], c, h * p, h * p)) return imgs def forward(self, latent, timestep, label): """ Forward pass of DiT. latent: (N, C, H, W) timestep: (N,) label: (N,) """ if self.cfg_scale is not None: output = self.forward_with_cfg(latent, timestep, label) else: output = self.forward_without_cfg(latent, timestep, label) output.mark_output('output', self.dtype) return output def forward_without_cfg(self, x, t, y): """ Forward pass without classifier-free guidance. """ x = self.x_embedder(x) + self.pos_embed.value t = self.t_embedder(t) y = self.y_embedder(y) self.register_network_output('t_embedder', t) self.register_network_output('x_embedder', x) self.register_network_output('y_embedder', y) c = t + y input_length = constant(np.array([x.shape[1]], dtype=np.int32)) input_lengths = expand(input_length, unsqueeze(shape(x, 0), 0)) # Split squeence for CP here if self.mapping.cp_size > 1: assert x.shape[1] % self.mapping.cp_size == 0 x = chunk(x, self.mapping.cp_size, dim=1)[self.mapping.cp_rank] input_lengths = input_lengths // self.mapping.cp_size for block in self.blocks: x = block(x, c, input_lengths) # (N, T, D) self.register_network_output('before_final_layer', x) x = self.final_layer(x, c) # (N, T, patch_size ** 2 * out_channels) self.register_network_output('final_layer', x) # All gather after CP if self.mapping.cp_size > 1: x = allgather(x, self.mapping.cp_group, gather_dim=1) x = self.unpatchify(x) # (N, out_channels, H, W) self.register_network_output('unpatchify', x) return x def forward_with_cfg(self, x, t, y): """ Forward pass with classifier-free guidance. """ batch_size = shape(x, 0) half = slice( x, [0, 0, 0, 0], concat([batch_size / 2, x.shape[1], x.shape[2], x.shape[3]])) combined = concat([half, half], dim=0) self.register_network_output('combined', combined) model_out = self.forward_without_cfg(combined, t, y) _, d, h, w = model_out.shape eps, rest = split(model_out, [3, d - 3], dim=1) cond_eps = slice(eps, [0, 0, 0, 0], concat([batch_size / 2, 3, h, w])) uncond_eps = slice(eps, concat([batch_size / 2, 0, 0, 0]), concat([batch_size / 2, 3, h, w])) self.register_network_output('cond_eps', cond_eps) self.register_network_output('uncond_eps', uncond_eps) half_eps = uncond_eps + self.cfg_scale * (cond_eps - uncond_eps) eps = concat([half_eps, half_eps], dim=0) self.register_network_output('eps', eps) return concat([eps, rest], dim=1) def prepare_inputs(self, max_batch_size, **kwargs): '''@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes. @return: a list contains values which can be fed into the self.forward() ''' mapping = self.config.mapping if mapping.tp_size > 1: current_all_reduce_helper().set_workspace_tensor(mapping, 1) def dit_default_range(max_batch_size): return [2, max(2, (max_batch_size + 1) // 2), max_batch_size] default_range = dit_default_range if self.cfg_scale is not None: max_batch_size *= 2 latent = Tensor( name='latent', dtype=self.dtype, shape=[-1, self.in_channels, self.input_size, self.input_size], dim_range=OrderedDict([ ('batch_size', [default_range(max_batch_size)]), ('in_channels', [[self.in_channels] * 3]), ('latent_height', [[self.input_size] * 3]), ('latent_width', [[self.input_size] * 3]), ])) timestep = Tensor(name='timestep', dtype=trt.int32, shape=[-1], dim_range=OrderedDict([ ('batch_size', [default_range(max_batch_size)]), ])) label = Tensor(name='label', dtype=trt.int32, shape=[-1], dim_range=OrderedDict([ ('batch_size', [default_range(max_batch_size)]), ])) return {'latent': latent, 'timestep': timestep, 'label': label}