modeling_heron.py

from dataclasses import dataclass
from typing import Unpack, cast

import torch
import torch.nn as nn
from transformers.cache_utils import Cache
from transformers.generation import GenerationMixin
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_outputs import ModelOutput
from transformers.modeling_utils import PreTrainedModel
from transformers.models.auto import AutoModel
from transformers.utils import TransformersKwargs

from .configuration_heron import HeronConfig


# Copied from https://github.com/NVlabs/VILA/blob/36f6adcd11a10be1580caeb7e647e1b6f8517f89/llava/model/multimodal_projector/base_projector.py#L83
def _flat_square_2x2(x):
    n, w, h, c = x.size()
    if w % 2 == 1:
        x = torch.concat([x, torch.zeros((n, 1, h, c), dtype=x.dtype).to(x.device)], dim=1).contiguous()
        n, w, h, c = x.size()
    x = x.contiguous()
    if h % 2 == 1:
        x = torch.concat([x, torch.zeros((n, w, 1, c), dtype=x.dtype).to(x.device)], dim=2).contiguous()
        n, w, h, c = x.size()
    x = x.view(n, w, int(h / 2), int(c * 2))
    x = x.permute(0, 2, 1, 3).contiguous()
    x = x.view(n, int(h / 2), int(w / 2), int(c * 4))
    x = x.permute(0, 2, 1, 3).contiguous()
    return x


class DownSample2x2BlockFix(nn.Module):
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        vit_embeds = x
        h = w = int(vit_embeds.shape[1] ** 0.5)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
        vit_embeds = _flat_square_2x2(vit_embeds)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
        return vit_embeds


# Copied from https://github.com/NVlabs/VILA/blob/36f6adcd11a10be1580caeb7e647e1b6f8517f89/llava/model/multimodal_projector/base_projector.py#L109
def _flat_square_3x3(x):
    n, w, h, c = x.size()
    if w % 3 != 0:
        x = torch.concat([x, torch.zeros((n, 3 - (w % 3), h, c), dtype=x.dtype).to(x.device)], dim=1).contiguous()
        n, w, h, c = x.size()
    x = x.contiguous()
    if h % 3 != 0:
        x = torch.concat([x, torch.zeros((n, w, 3 - (h % 3), c), dtype=x.dtype).to(x.device)], dim=2).contiguous()
        n, w, h, c = x.size()
    x = x.view(n, w, int(h / 3), int(c * 3))
    x = x.permute(0, 2, 1, 3).contiguous()
    x = x.view(n, int(h / 3), int(w / 3), int(c * 9))
    x = x.permute(0, 2, 1, 3).contiguous()
    return x


class DownSample3x3BlockFix(nn.Module):
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        vit_embeds = x
        h = w = int(vit_embeds.shape[1] ** 0.5)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
        vit_embeds = _flat_square_3x3(vit_embeds)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
        return vit_embeds


class HeronMultiModalProjector(nn.Module):
    def __init__(self, config: HeronConfig):
        super().__init__()

        mm_projector_type = config.mm_projector_type
        mm_hidden_size = config.vision_config.hidden_size
        output_hidden_size = config.text_config.hidden_size

        if mm_projector_type == "mlp_downsample_2x2_fix":
            self.layers = nn.Sequential(
                DownSample2x2BlockFix(),
                nn.LayerNorm(mm_hidden_size * 4),
                nn.Linear(mm_hidden_size * 4, output_hidden_size),
                nn.GELU(),
                nn.Linear(output_hidden_size, output_hidden_size),
            )
            self.downsample_rate = 2
        elif mm_projector_type == "mlp_downsample_3x3_fix":
            self.layers = nn.Sequential(
                DownSample3x3BlockFix(),
                nn.LayerNorm(mm_hidden_size * 9),
                nn.Linear(mm_hidden_size * 9, mm_hidden_size * 3),
                nn.GELU(),
                nn.LayerNorm(mm_hidden_size * 3),
                nn.Linear(mm_hidden_size * 3, output_hidden_size),
                nn.GELU(),
                nn.Linear(output_hidden_size, output_hidden_size),
            )
            self.downsample_rate = 3
        else:
            raise ValueError(f"Unknown projector type: {mm_projector_type}")

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.layers(x)


@dataclass
class HeronModelOutputWithPast(ModelOutput):
    last_hidden_state: torch.FloatTensor | None = None
    past_key_values: list[torch.FloatTensor] | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    image_hidden_states: tuple[torch.FloatTensor] | None = None


class HeronPreTrainedModel(PreTrainedModel):
    config: HeronConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _skip_keys_device_placement = "past_key_values"  # type: ignore[assignment]
    _supports_flash_attn = True
    _supports_sdpa = True
    _can_compile_fullgraph = True
    _supports_flex_attn = True
    _supports_attention_backend = True


class HeronModel(HeronPreTrainedModel):
    def __init__(self, config: HeronConfig):
        super().__init__(config)

        self.vision_tower: PreTrainedModel = AutoModel.from_config(config.vision_config)
        self.multi_modal_projector = HeronMultiModalProjector(config)
        self.language_model: PreTrainedModel = AutoModel.from_config(config.text_config)

        self.post_init()

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    def get_decoder(self):
        return self.language_model

    def set_decoder(self, decoder):
        self.language_model = decoder

    def get_image_features(self, pixel_values: torch.FloatTensor, vision_feature_layer: int | None = None):
        ndim = pixel_values.ndim
        assert ndim == 5 or ndim == 4, "pixel_values should be of shape (batch_size, num_tiles, n_channels, height, width) or (batch_size, n_channels, height, width)"  # fmt: skip
        if ndim == 4:
            pixel_values = cast(torch.FloatTensor, pixel_values.unsqueeze(1))

        shape = pixel_values.shape
        pixel_values = cast(torch.FloatTensor, pixel_values.view(shape[0] * shape[1], *shape[2:]))

        if vision_feature_layer is None:
            vision_feature_layer = self.config.vision_feature_layer
        outputs = self.vision_tower(pixel_values=pixel_values, output_hidden_states=True)
        image_features = outputs.hidden_states[vision_feature_layer]
        image_features = self.multi_modal_projector(image_features)

        image_features = image_features.view(shape[0], shape[1], *image_features.shape[1:])
        return image_features

    def get_placeholder_mask(
        self, input_ids: torch.LongTensor | None, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor
    ):
        if input_ids is None:
            special_image_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_image_mask = special_image_mask.all(-1)
        else:
            special_image_mask = input_ids == self.config.image_token_id

        n_image_tokens = special_image_mask.sum()
        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if not torch.compiler.is_compiling():
            if inputs_embeds[special_image_mask].numel() != image_features.numel():
                n_image_features = image_features.shape[0] * image_features.shape[1]
                raise ValueError(
                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
                )
        return special_image_mask

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple | HeronModelOutputWithPast:
        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

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        if pixel_values is not None:
            image_features = self.get_image_features(pixel_values)
            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
            special_image_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_features
            )
            inputs_embeds = cast(torch.FloatTensor, inputs_embeds.masked_scatter(special_image_mask, image_features))

        outputs = self.language_model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            **kwargs,
        )

        output = HeronModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
        )
        return output if return_dict else output.to_tuple()


@dataclass
class HeronCausalLMOutputWithPast(ModelOutput):
    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    past_key_values: list[torch.FloatTensor] | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    image_hidden_states: tuple[torch.FloatTensor] | None = None


class HeronForConditionalGeneration(HeronPreTrainedModel, GenerationMixin):  # type: ignore[misc]
    _tied_weights_keys = ["lm_head.weight"]  # type: ignore[assignment]

    def __init__(self, config: HeronConfig):
        super().__init__(config)
        self.model = HeronModel(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.post_init()

    def get_input_embeddings(self):
        return self.model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.model.set_input_embeddings(value)

    def get_output_embeddings(self) -> nn.Module:
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def get_decoder(self):
        return self.model.get_decoder()

    def set_decoder(self, decoder):
        self.model.set_decoder(decoder)

    def forward(  # type: ignore[misc]
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | HeronCausalLMOutputWithPast:
        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

        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs[0]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(
                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
            )

        return HeronCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=outputs.image_hidden_states,
        )

    def prepare_inputs_for_generation(  # type: ignore[override]
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        attention_mask=None,
        cache_position=None,
        logits_to_keep=None,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            logits_to_keep=logits_to_keep,
            **kwargs,
        )

        if cache_position is not None and cache_position[0] == 0:
            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
            # Otherwise we need pixel values to be passed to model
            model_inputs["pixel_values"] = pixel_values

        return model_inputs