PaliGemma 구현 Part 2

PaliGemma 구현 Part 2

(Reference: https://www.youtube.com/watch?v=vAmKB7iPkWw&t=19426s)

• [Vision] modeling_siglip.py
• [Vision2Text] processing_paligemma.py
• [Total] modeling_gemma.py

[Vision] modeling_siglip.py

SiglipVisionModel (configuration: SiglipVisionConfig)

• SiglipVisionTransformer
  • SiglipVisionEmbeddings: ViT의 (첫 번째) patch embedding layer
  • SiglipEncoder: ViT의 (main) Encoder
    • SiglipEncoderLayer: ViT Encoder의 layer
      • SiglipAttention: MHA
      • SiglipMLP: MLP

[Vision2Text] processing_paligemma.py

PaliGemmaProcessor: Gemma의 입력을 위한 image & text 토큰 전처리

• add_image_tokens_to_prompt: image token을 text token 앞에 이어서 붙인다.
• process_images: numpy로 된 image를 전처리하는 함수
  • rescale, resize, normalize

[Total] modeling_gemma.py

PaliGemmaForConditionalGeneration (configuration: PaliGemmaConfig - GemmaConfig)

• SiglipVisionModel: Vision encoder
• PaliGemmaMultiModalProjector: Vision encoder에서 나온 image token을 LLM space로 projection
• GemmaForCausalLM: Multimodal decoder + LM head
  • GemmaModel: Multimodal decoder
    • GemmaDecoderLayer: Multimodal decoder layer
      • GemmaAttention
        • GemmaRotaryEmbedding: RoPE
      • GemmaMLP: MLP layer (gating + gelu)
      • GemmaRMSNorm: RMS norm
    • GemmaRMSNorm: RMS norm

3. Total (modeling_gemma.py)

import torch
from torch import nn
from typing import Optional, Tuple, List
from torch.nn import CrossEntropyLoss
import math
from modeling_siglip import SiglipVisionConfig, SiglipVisionModel

PaliGemmaConfig: Multimodal Decoder의 configuration

• GemmaConfig : 그 중 text 관련의 configuration

class PaliGemmaConfig():

    def __init__(
        self,
        vision_config=None,
        text_config=None,
        ignore_index=-100,
        image_token_index=256000,
        vocab_size=257152,
        projection_dim=2048,
        hidden_size=2048,
        pad_token_id=None,
        **kwargs,
    ):
        super().__init__()
        self.ignore_index = ignore_index
        self.image_token_index = image_token_index
        self.vocab_size = vocab_size
        self.projection_dim = projection_dim
        self.hidden_size = hidden_size
        self.vision_config = vision_config
        self.is_encoder_decoder = False
        self.pad_token_id = pad_token_id

        self.vision_config = SiglipVisionConfig(**vision_config)
        self.text_config = GemmaConfig(**text_config, pad_token_id=pad_token_id)
        self.vocab_size = self.text_config.vocab_size
        self.text_config.num_image_tokens = (self.vision_config.image_size // self.vision_config.patch_size) ** 2
        self.vision_config.projection_dim = projection_dim

class GemmaConfig():
    def __init__(
        self,
        vocab_size,
        hidden_size,
        intermediate_size,
        num_hidden_layers,
        num_attention_heads,
        num_key_value_heads,
        head_dim=256,
        max_position_embeddings=8192,
        rms_norm_eps=1e-6,
        rope_theta=10000.0,
        attention_bias=False,
        attention_dropout=0.0,
        pad_token_id=None,
        **kwargs,
    ):
        super().__init__()
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.head_dim = head_dim
        self.num_key_value_heads = num_key_value_heads
        self.rms_norm_eps = rms_norm_eps
        self.rope_theta = rope_theta
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.pad_token_id = pad_token_id

PaliGemmaForConditionalGeneration :

• SiglipVisionModel: Vision encoder
• PaliGemmaMultiModalProjector: Vision encoder에서 나온 image token을 LLM space로 projection
• GemmaForCausalLM: Multimodal decoder + LM head

class PaliGemmaForConditionalGeneration(nn.Module):
    def __init__(self, config: PaliGemmaConfig):
        super().__init__()
        self.config = config
        self.vision_tower = SiglipVisionModel(config.vision_config)
        self.multi_modal_projector = PaliGemmaMultiModalProjector(config)
        self.vocab_size = config.vocab_size

        language_model = GemmaForCausalLM(config.text_config)
        self.language_model = language_model

        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
		
    # (1) Input embedding = Output embedding
    def tie_weights(self):
        return self.language_model.tie_weights()
		
    # (2) Text & Image 합치기
    def _merge_input_ids_with_image_features(
        self, image_features: torch.Tensor, inputs_embeds: torch.Tensor, input_ids: torch.Tensor, attention_mask: torch.Tensor, kv_cache: Optional[KVCache] = None
    ):
        _, _, embed_dim = image_features.shape
        batch_size, sequence_length = input_ids.shape
        dtype, device = inputs_embeds.dtype, inputs_embeds.device
        scaled_image_features = image_features / (self.config.hidden_size**0.5)
    
        # 최종으로 사용할 embedding틀 ([B,L,D])
        # [B,L,D], where L = "모든" token의 길이(개수)
        final_embedding = torch.zeros(batch_size, sequence_length, embed_dim, 
                                      dtype=inputs_embeds.dtype, 
                                      device=inputs_embeds.device)

        # Text, Image, Pad 토큰을 식별해주는 mask(크기:[B,L])
        text_mask = (input_ids != self.config.image_token_index) & (input_ids != self.pad_token_id)
        image_mask = input_ids == self.config.image_token_index
        pad_mask = input_ids == self.pad_token_id

        # 위의 mask의 크기 맞춰주기 (복제)
        text_mask_expanded = text_mask.unsqueeze(-1).expand(-1, -1, embed_dim)
        pad_mask_expanded = pad_mask.unsqueeze(-1).expand(-1, -1, embed_dim)
        image_mask_expanded = image_mask.unsqueeze(-1).expand(-1, -1, embed_dim)

        # Final embedding에 차례로 Text-Image-Pad 토큰 추가하기 (덮어쓰기)
        final_embedding = torch.where(text_mask_expanded, inputs_embeds, final_embedding)
        final_embedding = final_embedding.masked_scatter(image_mask_expanded, scaled_image_features)
        final_embedding = torch.where(pad_mask_expanded, torch.zeros_like(final_embedding), final_embedding)

        # Attention mask 만들기
        dtype, device = inputs_embeds.dtype, inputs_embeds.device
        min_dtype = torch.finfo(dtype).min
        q_len = inputs_embeds.shape[1]
    
        if kv_cache is None or kv_cache.num_items() == 0:
            # Training (Prefill, KV cache X)
						# (LLaMA와는 달리, PaliGemma에서는) Prefix 부분은 전부 참고 가능! 따라서 causal mask 없음.
            causal_mask = torch.full(
                (batch_size, q_len, q_len), fill_value=0, dtype=dtype, device=device
            )
        else:
            # Inference (KV cache O)
            # KV caching을 하므로, matrix의 last row (마지막 단어)에 대해서만 계산을 함.
            # 따라서, (q_len, q_len)이 아니라 (q_len, kv_len)
            assert q_len == 1
            kv_len = kv_cache.num_items() + q_len
            causal_mask = torch.full(
                (batch_size, q_len, kv_len), fill_value=0, dtype=dtype, device=device
            )

        # [B, Q_Len, KV_Len] -> [B, 1, Q_Len, KV_Len] (head dimension 추가)
        causal_mask = causal_mask.unsqueeze(1)

        if kv_cache is not None and kv_cache.num_items() > 0:
            position_ids = attention_mask.cumsum(-1)[:, -1]
            if position_ids.dim() == 1:
                position_ids = position_ids.unsqueeze(0)
        else:
            position_ids = (attention_mask.cumsum(-1)).masked_fill_((attention_mask == 0), 1).to(device)

        return final_embedding, causal_mask, position_ids

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        kv_cache: Optional[KVCache] = None,
    ) -> Tuple:

        assert torch.all(attention_mask == 1), "The input cannot be padded"

        # (1) Text embedding
        # Token ID -> Token Embedding ... [B,L,D]
        inputs_embeds = self.language_model.get_input_embeddings()(input_ids)

        # (2) Image embedding
        # Image pixel -> Image Embedding .... [B,C,H,W] -> [B,N,D]
        selected_image_feature = self.vision_tower(pixel_values.to(inputs_embeds.dtype))
        # Projection ... [B,N,D] -> [B,N,D2]
        image_features = self.multi_modal_projector(selected_image_feature)

        # (3) Text + Image embedding
        inputs_embeds, attention_mask, position_ids = self._merge_input_ids_with_image_features(image_features, inputs_embeds, input_ids, attention_mask, kv_cache)
        
        # (4) LM으로 generation
        outputs = self.language_model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            kv_cache=kv_cache,
        )

        return outputs

PaliGemmaMultiModalProjector: image embedding을 LLM space로 매핑

class PaliGemmaMultiModalProjector(nn.Module):
    def __init__(self, config: PaliGemmaConfig):
        super().__init__()
        self.linear = nn.Linear(config.vision_config.hidden_size, 
                                config.vision_config.projection_dim, 
                                bias=True)

    def forward(self, image_features):
        # [B,N,D] -> [B,N,D2]
        hidden_states = self.linear(image_features)
        return hidden_states

GemmaForCausalLM: Multimodal decoder + LM head

• GemmaModel: Multimodal decoder

class GemmaForCausalLM(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.vocab_size = config.vocab_size
        
        # (1) Multimodal decoder
        self.model = GemmaModel(config)
        # (2) LM head
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

    def get_input_embeddings(self):
        return self.model.embed_tokens
    
    def tie_weights(self):
        self.lm_head.weight = self.model.embed_tokens.weight

    def forward(
        self,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        kv_cache: Optional[KVCache] = None,
    ) -> Tuple:

        # (1) Last token에 대한 embedding vector
        # (input_embeds: [B,L,D] -> outputs: [B,L,D])
        outputs = self.model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            kv_cache=kv_cache,
        )
        hidden_states = outputs
        
        # (2) Logit 계산하기
        logits = self.lm_head(hidden_states)
        logits = logits.float()

        return_data = {"logits": logits,}
				
        # (3) KV cache 업데이트해주기
        if kv_cache is not None:
            return_data["kv_cache"] = kv_cache

        return return_data

GemmaModel: Multimodal decoder

• GemmaDecoderLayer: Multimodal decoder layer
• GemmaRMSNorm: RMS norm

class GemmaModel(nn.Module):
    def __init__(self, config: GemmaConfig):
        super().__init__()
        self.config = config
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [GemmaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def get_input_embeddings(self):
        return self.embed_tokens

    def forward(
        self,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        kv_cache: Optional[KVCache] = None,
    ) -> torch.FloatTensor:
      
        # inputs_embeds: [B,L,D] -> shape은 계속 유지됨
        hidden_states = inputs_embeds
        normalizer = torch.tensor(self.config.hidden_size**0.5, dtype=hidden_states.dtype)
        hidden_states = hidden_states * normalizer

        for decoder_layer in self.layers:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                kv_cache=kv_cache,
            )
				
        # RMSNorm
        hidden_states = self.norm(hidden_states)
        return hidden_states

GemmaDecoderLayer: Multimodal decoder layer

• GemmaAttention
  • GemmaRotaryEmbedding: RoPE
• GemmaMLP: MLP layer (gating + gelu)
• GemmaRMSNorm: RMS norm

class GemmaDecoderLayer(nn.Module):

    def __init__(self, config: GemmaConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = GemmaAttention(config=config, layer_idx=layer_idx)

        self.mlp = GemmaMLP(config)
        self.input_layernorm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        kv_cache: Optional[KVCache] = None,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        residual = hidden_states
        
        # inputs_embeds: [B,L,D] -> shape은 계속 유지됨
        # (Pre) RMS Norm
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, _, = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            kv_cache=kv_cache,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        # (Post) RMS Norm
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states

GemmaAttention: Attention layer

class GemmaAttention(nn.Module):

    def __init__(self, config: GemmaConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx

        self.attention_dropout = config.attention_dropout
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = config.head_dim
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True

        assert self.hidden_size % self.num_heads == 0            

        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
        self.rotary_emb = GemmaRotaryEmbedding(
            self.head_dim,
            max_position_embeddings=self.max_position_embeddings,
            base=self.rope_theta,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        kv_cache: Optional[KVCache] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        # [B,L,D]
        bsz, q_len, _ = hidden_states.size() 
        
        # [B,L,H*d]
        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)
        
        # [B,H,L,d]
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        # [B,L,d], [B,L,d]
        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=None)
        # [Batch_Size, Num_Heads_Q, Seq_Len, Head_Dim], [Batch_Size, Num_Heads_KV, Seq_Len, Head_Dim]
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if kv_cache is not None:
            key_states, value_states = kv_cache.update(key_states, value_states, self.layer_idx)

        # GQA (Grouped Query Attention): 개수만큼 복제하기
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

        assert attention_mask is not None
        attn_weights = attn_weights + attention_mask

        # [B,H,L_q,L_kv]
        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.attention_dropout, training=self.training)
        
        # [B,H,L_q, L_kv] x [B,H,L_kv,d] -> [B,H,L_q,d]
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )
            
        # [B,H,L_q,d] -> [B,L_q,H,d] 
        attn_output = attn_output.transpose(1, 2).contiguous()
        # [B,L_q,H,d]  -> [B,L_q,D] 
        attn_output = attn_output.view(bsz, q_len, -1)
        # SAME
        attn_output = self.o_proj(attn_output)

        return attn_output, attn_weights

GemmaMLP: MLP layer (gating + gelu)

class GemmaMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)

    def forward(self, x):
        return self.down_proj(nn.functional.gelu(self.gate_proj(x), approximate="tanh") * self.up_proj(x))

GemmaRMSNorm: RMS Normalization

class GemmaRMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.zeros(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        output = self._norm(x.float())
        output = output * (1.0 + self.weight.float())
        return output.type_as(x)

KVCache: KV caching

class KVCache():
    def __init__(self) -> None:
        self.key_cache: List[torch.Tensor] = []
        self.value_cache: List[torch.Tensor] = []
    
    def num_items(self) -> int:
        if len(self.key_cache) == 0:
            return 0
        else:
            # [B,H,L,D]의 "L"
            return self.key_cache[0].shape[-2]

    def update(
        self,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
        layer_idx: int,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        if len(self.key_cache) <= layer_idx:
            # KV cache 공간 생성
            self.key_cache.append(key_states)
            self.value_cache.append(value_states)
        else:
            # KV cache에 K,V 추가하기
            # 각 tensor: [B,H,L,d]
            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2)
            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2)

        # 기존 token + 새로운 token에 해당하는 K,V 반환
        return self.key_cache[layer_idx], self.value_cache[layer_idx]

GemmaRotaryEmbedding: RoPE

class GemmaRotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
        super().__init__()

        self.dim = dim # d
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim))
        self.register_buffer("inv_freq", tensor=inv_freq, persistent=False)

    @torch.no_grad()
    def forward(self, x, position_ids, seq_len=None):
        # x: [B,H,L,d]
        self.inv_freq.to(x.device)
        # inv_freq_expanded: [B,d//2, 1]
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        # position_ids_expanded: [B,1,L]
        position_ids_expanded = position_ids[:, None, :].float()
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)

def rotate_half(x):
    x1 = x[..., : x.shape[-1] // 2] 
    x2 = x[..., x.shape[-1] // 2 :] 
    return torch.cat((-x2, x1), dim=-1)

def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
    cos = cos.unsqueeze(unsqueeze_dim) # Add the head dimension
    sin = sin.unsqueeze(unsqueeze_dim) # Add the head dimension
    # Apply the formula (34) of the Rotary Positional Encoding paper.
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)