optimum/habana/transformers/models/siglip/modeling_siglip.py

from typing import Optional, Tuple, Union

import torch
from torch import nn
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
from transformers.models.siglip.configuration_siglip import SiglipConfig
from transformers.models.siglip.modeling_siglip import (
    SiglipAttention,
    SiglipEncoder,
    SiglipEncoderLayer,
    SiglipMLP,
    SiglipVisionEmbeddings,
    SiglipVisionModel,
    SiglipVisionTransformer,
)

from ..modeling_all_models import Matmul


try:
    from habana_frameworks.torch.hpex.kernels import FusedSDPA
except ImportError:
    print("Not using HPU fused scaled dot-product attention kernel.")
    FusedSDPA = None


class GaudiSiglipVisionEmbeddings(SiglipVisionEmbeddings):
    def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
        _, _, height, width = pixel_values.shape
        target_dtype = self.patch_embedding.weight.dtype
        # if HQT quantization enabled, remove the explicit cast to float8 to avoid HQT casting error
        if "float8" in str(target_dtype) and pixel_values.device.type == "hpu":
            target_dtype = torch.bfloat16
        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
        embeddings = patch_embeds.flatten(2).transpose(1, 2)

        if interpolate_pos_encoding:
            embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
        else:
            embeddings = embeddings + self.position_embedding(self.position_ids)
        return embeddings


class ModuleFusedSDPA(torch.nn.Module):
    def __init__(self, fusedSDPA):
        super().__init__()
        self._hpu_kernel_fsdpa = fusedSDPA

    def forward(self, query, key, value, attn_mask, dropout_p, is_casual, scale, softmax_mode):
        return self._hpu_kernel_fsdpa.apply(query, key, value, attn_mask, dropout_p, is_casual, scale, softmax_mode)


class Softmax(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x, dim=None, invAttnHead=None):
        return torch.nn.functional.softmax(x, dim)


class GaudiSiglipAttention(SiglipAttention):
    def __init__(self, config):
        super().__init__(config=config)
        self.fused_scaled_dot_product_attention = ModuleFusedSDPA(FusedSDPA) if FusedSDPA else None
        self.bmm1 = Matmul()
        self.bmm2 = Matmul()
        self.softmax = Softmax()

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = False,
        use_flash_attention: Optional[bool] = False,
        flash_attention_recompute: Optional[bool] = False,
        flash_attention_fast_softmax: Optional[bool] = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        """
        Copied from CLIPAttention.forward: https://github.com/huggingface/transformers/blob/ab0f050b42d903f34d6eb97f3f8c0c07f0517ad2/src/transformers/models/clip/modeling_clip.py
        The only differences are:
        - add new args use_flash_attention to enable FusedSDPA
        - add new args flash_attention_recompute
        - add new args flash_attention_fast_softmax
        """
        """Input shape: Batch x Time x Channel"""

        batch_size, q_len, _ = hidden_states.size()
        attn_weights = None

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)

        k_v_seq_len = key_states.shape[-2]

        if FusedSDPA and use_flash_attention:
            import habana_frameworks.torch.hpu as ht

            softmax_mode = "fast" if flash_attention_fast_softmax else "None"
            with ht.sdp_kernel(enable_recompute=flash_attention_recompute):
                attn_output = self.fused_scaled_dot_product_attention(
                    query_states,
                    key_states,
                    value_states,
                    attention_mask,
                    self.dropout,
                    False,
                    1,
                    softmax_mode,
                )
        else:
            attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale

            if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
                raise ValueError(
                    f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
                    f" {attn_weights.size()}"
                )

            if attention_mask is not None:
                if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
                    raise ValueError(
                        f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
                    )
                attn_weights = attn_weights + attention_mask

            # upcast attention to fp32
            attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
            attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
            attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights


class GaudiSiglipEncoderLayer(SiglipEncoderLayer):
    def __init__(self, config: SiglipConfig):
        super(SiglipEncoderLayer, self).__init__()
        self.embed_dim = config.hidden_size
        self.self_attn = GaudiSiglipAttention(config)
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = SiglipMLP(config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        output_attentions: Optional[bool] = False,
        use_flash_attention: Optional[bool] = False,
        flash_attention_recompute: Optional[bool] = False,
    ) -> Tuple[torch.FloatTensor]:
        """
        Args:
            hidden_states (`torch.FloatTensor`):
                Input to the layer of shape `(batch, seq_len, embed_dim)`.
            attention_mask (`torch.FloatTensor`):
                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states, attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            use_flash_attention=use_flash_attention,
            flash_attention_recompute=flash_attention_recompute,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.layer_norm2(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


class GaudiSiglipEncoder(SiglipEncoder):
    def forward(
        self,
        inputs_embeds,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        use_flash_attention: Optional[bool] = False,
        flash_attention_recompute: Optional[bool] = False,
    ) -> Union[Tuple, BaseModelOutput]:
        r"""
        Args:
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        hidden_states = inputs_embeds
        for encoder_layer in self.layers:
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    encoder_layer.__call__,
                    hidden_states,
                    attention_mask,
                    output_attentions,
                )
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    attention_mask,
                    output_attentions=output_attentions,
                    use_flash_attention=use_flash_attention,
                    flash_attention_recompute=flash_attention_recompute,
                )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


class GaudiSiglipVisionTransformer(SiglipVisionTransformer):
    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        interpolate_pos_encoding: Optional[bool] = False,
        use_flash_attention: Optional[bool] = False,
        flash_attention_recompute: Optional[bool] = False,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        r"""
        Returns:

        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)

        encoder_outputs = self.encoder(
            inputs_embeds=hidden_states,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            use_flash_attention=use_flash_attention,
            flash_attention_recompute=flash_attention_recompute,
        )

        last_hidden_state = encoder_outputs[0]
        last_hidden_state = self.post_layernorm(last_hidden_state)

        pooler_output = self.head(last_hidden_state) if self.use_head else None
        if not return_dict:
            return (last_hidden_state, pooler_output) + encoder_outputs[1:]

        return BaseModelOutputWithPooling(
            last_hidden_state=last_hidden_state,
            pooler_output=pooler_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


class GaudiSiglipVisionModel(SiglipVisionModel):
    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        use_flash_attention: Optional[bool] = False,
        flash_attention_recompute: Optional[bool] = False,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        """
        Copied from CLIPVisionModel.forward: https://github.com/huggingface/transformers/blob/ab0f050b42d903f34d6eb97f3f8c0c07f0517ad2/src/transformers/models/clip/modeling_clip.py
        The only differences are:
        - add new args use_flash_attention
        - add new args flash_attention_recompute
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        return self.vision_model(
            pixel_values=pixel_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            use_flash_attention=use_flash_attention,
            flash_attention_recompute=flash_attention_recompute,
        )