comfy/text_encoders/t5.py

import torch
import math
from comfy.ldm.modules.attention import optimized_attention_for_device
import comfy.ops

class T5LayerNorm(torch.nn.Module):
    def __init__(self, hidden_size, eps=1e-6, dtype=None, device=None, operations=None):
        super().__init__()
        self.weight = torch.nn.Parameter(torch.empty(hidden_size, dtype=dtype, device=device))
        self.variance_epsilon = eps

    def forward(self, x):
        variance = x.pow(2).mean(-1, keepdim=True)
        x = x * torch.rsqrt(variance + self.variance_epsilon)
        return comfy.ops.cast_to_input(self.weight, x) * x

activations = {
    "gelu_pytorch_tanh": lambda a: torch.nn.functional.gelu(a, approximate="tanh"),
    "relu": torch.nn.functional.relu,
}

class T5DenseActDense(torch.nn.Module):
    def __init__(self, model_dim, ff_dim, ff_activation, dtype, device, operations):
        super().__init__()
        self.wi = operations.Linear(model_dim, ff_dim, bias=False, dtype=dtype, device=device)
        self.wo = operations.Linear(ff_dim, model_dim, bias=False, dtype=dtype, device=device)
        # self.dropout = nn.Dropout(config.dropout_rate)
        self.act = activations[ff_activation]

    def forward(self, x):
        x = self.act(self.wi(x))
        # x = self.dropout(x)
        x = self.wo(x)
        return x

class T5DenseGatedActDense(torch.nn.Module):
    def __init__(self, model_dim, ff_dim, ff_activation, dtype, device, operations):
        super().__init__()
        self.wi_0 = operations.Linear(model_dim, ff_dim, bias=False, dtype=dtype, device=device)
        self.wi_1 = operations.Linear(model_dim, ff_dim, bias=False, dtype=dtype, device=device)
        self.wo = operations.Linear(ff_dim, model_dim, bias=False, dtype=dtype, device=device)
        # self.dropout = nn.Dropout(config.dropout_rate)
        self.act = activations[ff_activation]

    def forward(self, x):
        hidden_gelu = self.act(self.wi_0(x))
        hidden_linear = self.wi_1(x)
        x = hidden_gelu * hidden_linear
        # x = self.dropout(x)
        x = self.wo(x)
        return x

class T5LayerFF(torch.nn.Module):
    def __init__(self, model_dim, ff_dim, ff_activation, gated_act, dtype, device, operations):
        super().__init__()
        if gated_act:
            self.DenseReluDense = T5DenseGatedActDense(model_dim, ff_dim, ff_activation, dtype, device, operations)
        else:
            self.DenseReluDense = T5DenseActDense(model_dim, ff_dim, ff_activation, dtype, device, operations)

        self.layer_norm = T5LayerNorm(model_dim, dtype=dtype, device=device, operations=operations)
        # self.dropout = nn.Dropout(config.dropout_rate)

    def forward(self, x):
        forwarded_states = self.layer_norm(x)
        forwarded_states = self.DenseReluDense(forwarded_states)
        # x = x + self.dropout(forwarded_states)
        x += forwarded_states
        return x

class T5Attention(torch.nn.Module):
    def __init__(self, model_dim, inner_dim, num_heads, relative_attention_bias, dtype, device, operations):
        super().__init__()

        # Mesh TensorFlow initialization to avoid scaling before softmax
        self.q = operations.Linear(model_dim, inner_dim, bias=False, dtype=dtype, device=device)
        self.k = operations.Linear(model_dim, inner_dim, bias=False, dtype=dtype, device=device)
        self.v = operations.Linear(model_dim, inner_dim, bias=False, dtype=dtype, device=device)
        self.o = operations.Linear(inner_dim, model_dim, bias=False, dtype=dtype, device=device)
        self.num_heads = num_heads

        self.relative_attention_bias = None
        if relative_attention_bias:
            self.relative_attention_num_buckets = 32
            self.relative_attention_max_distance = 128
            self.relative_attention_bias = operations.Embedding(self.relative_attention_num_buckets, self.num_heads, device=device, dtype=dtype)

    @staticmethod
    def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
        """
        Adapted from Mesh Tensorflow:
        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593

        Translate relative position to a bucket number for relative attention. The relative position is defined as
        memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
        position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
        small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
        positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
        This should allow for more graceful generalization to longer sequences than the model has been trained on

        Args:
            relative_position: an int32 Tensor
            bidirectional: a boolean - whether the attention is bidirectional
            num_buckets: an integer
            max_distance: an integer

        Returns:
            a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
        """
        relative_buckets = 0
        if bidirectional:
            num_buckets //= 2
            relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
            relative_position = torch.abs(relative_position)
        else:
            relative_position = -torch.min(relative_position, torch.zeros_like(relative_position))
        # now relative_position is in the range [0, inf)

        # half of the buckets are for exact increments in positions
        max_exact = num_buckets // 2
        is_small = relative_position < max_exact

        # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
        relative_position_if_large = max_exact + (
            torch.log(relative_position.float() / max_exact)
            / math.log(max_distance / max_exact)
            * (num_buckets - max_exact)
        ).to(torch.long)
        relative_position_if_large = torch.min(
            relative_position_if_large, torch.full_like(relative_position_if_large, num_buckets - 1)
        )

        relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
        return relative_buckets

    def compute_bias(self, query_length, key_length, device, dtype):
        """Compute binned relative position bias"""
        context_position = torch.arange(query_length, dtype=torch.long, device=device)[:, None]
        memory_position = torch.arange(key_length, dtype=torch.long, device=device)[None, :]
        relative_position = memory_position - context_position  # shape (query_length, key_length)
        relative_position_bucket = self._relative_position_bucket(
            relative_position,  # shape (query_length, key_length)
            bidirectional=True,
            num_buckets=self.relative_attention_num_buckets,
            max_distance=self.relative_attention_max_distance,
        )
        values = self.relative_attention_bias(relative_position_bucket, out_dtype=dtype)  # shape (query_length, key_length, num_heads)
        values = values.permute([2, 0, 1]).unsqueeze(0)  # shape (1, num_heads, query_length, key_length)
        return values

    def forward(self, x, mask=None, past_bias=None, optimized_attention=None):
        q = self.q(x)
        k = self.k(x)
        v = self.v(x)
        if self.relative_attention_bias is not None:
            past_bias = self.compute_bias(x.shape[1], x.shape[1], x.device, x.dtype)

        if past_bias is not None:
            if mask is not None:
                mask = mask + past_bias
            else:
                mask = past_bias

        out = optimized_attention(q, k * ((k.shape[-1] / self.num_heads) ** 0.5), v, self.num_heads, mask)
        return self.o(out), past_bias

class T5LayerSelfAttention(torch.nn.Module):
    def __init__(self, model_dim, inner_dim, ff_dim, num_heads, relative_attention_bias, dtype, device, operations):
        super().__init__()
        self.SelfAttention = T5Attention(model_dim, inner_dim, num_heads, relative_attention_bias, dtype, device, operations)
        self.layer_norm = T5LayerNorm(model_dim, dtype=dtype, device=device, operations=operations)
        # self.dropout = nn.Dropout(config.dropout_rate)

    def forward(self, x, mask=None, past_bias=None, optimized_attention=None):
        normed_hidden_states = self.layer_norm(x)
        output, past_bias = self.SelfAttention(self.layer_norm(x), mask=mask, past_bias=past_bias, optimized_attention=optimized_attention)
        # x = x + self.dropout(attention_output)
        x += output
        return x, past_bias

class T5Block(torch.nn.Module):
    def __init__(self, model_dim, inner_dim, ff_dim, ff_activation, gated_act, num_heads, relative_attention_bias, dtype, device, operations):
        super().__init__()
        self.layer = torch.nn.ModuleList()
        self.layer.append(T5LayerSelfAttention(model_dim, inner_dim, ff_dim, num_heads, relative_attention_bias, dtype, device, operations))
        self.layer.append(T5LayerFF(model_dim, ff_dim, ff_activation, gated_act, dtype, device, operations))

    def forward(self, x, mask=None, past_bias=None, optimized_attention=None):
        x, past_bias = self.layer[0](x, mask, past_bias, optimized_attention)
        x = self.layer[-1](x)
        return x, past_bias

class T5Stack(torch.nn.Module):
    def __init__(self, num_layers, model_dim, inner_dim, ff_dim, ff_activation, gated_act, num_heads, relative_attention, dtype, device, operations):
        super().__init__()

        self.block = torch.nn.ModuleList(
            [T5Block(model_dim, inner_dim, ff_dim, ff_activation, gated_act, num_heads, relative_attention_bias=((not relative_attention) or (i == 0)), dtype=dtype, device=device, operations=operations) for i in range(num_layers)]
        )
        self.final_layer_norm = T5LayerNorm(model_dim, dtype=dtype, device=device, operations=operations)
        # self.dropout = nn.Dropout(config.dropout_rate)

    def forward(self, x, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
        mask = None
        if attention_mask is not None:
            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))

        intermediate = None
        optimized_attention = optimized_attention_for_device(x.device, mask=attention_mask is not None, small_input=True)
        past_bias = None
        for i, l in enumerate(self.block):
            x, past_bias = l(x, mask, past_bias, optimized_attention)
            if i == intermediate_output:
                intermediate = x.clone()
        x = self.final_layer_norm(x)
        if intermediate is not None and final_layer_norm_intermediate:
            intermediate = self.final_layer_norm(intermediate)
        return x, intermediate

class T5(torch.nn.Module):
    def __init__(self, config_dict, dtype, device, operations):
        super().__init__()
        self.num_layers = config_dict["num_layers"]
        model_dim = config_dict["d_model"]

        self.encoder = T5Stack(self.num_layers, model_dim, model_dim, config_dict["d_ff"], config_dict["dense_act_fn"], config_dict["is_gated_act"], config_dict["num_heads"], config_dict["model_type"] != "umt5", dtype, device, operations)
        self.dtype = dtype
        self.shared = operations.Embedding(config_dict["vocab_size"], model_dim, device=device, dtype=dtype)

    def get_input_embeddings(self):
        return self.shared

    def set_input_embeddings(self, embeddings):
        self.shared = embeddings

    def forward(self, input_ids, *args, **kwargs):
        x = self.shared(input_ids, out_dtype=kwargs.get("dtype", torch.float32))
        if self.dtype not in [torch.float32, torch.float16, torch.bfloat16]:
            x = torch.nan_to_num(x) #Fix for fp8 T5 base
        return self.encoder(x, *args, **kwargs)