model_16_float.py

from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions
import torch
import torch.nn as nn
from torch.nn import functional as F
from transformers.modeling_outputs import CausalLMOutput


class BVVConfig(PretrainedConfig):
    model_type = "model_16_float"
    
    def __init__(
        self,
        vocab_size = 65536,
        n_embed = 16,
        d_model = 1024,
        n_head = 32,  
        n_layer = 16, 
        block_size = 1024,
        pad_id = 57344,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.block_size  = block_size 
        self.n_embed = n_embed
        self.d_model = d_model
        self.n_layer = n_layer
        self.n_head = n_head
        self.pad_id = pad_id 
        self.scale = d_model // n_embed 


class RotaryEmbedding(nn.Module):
    def __init__(self, dim):  # dim = head_dim (?? d_model!)
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def forward(self, seq_len, device):
        t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
        freqs = torch.einsum('i,j->ij', t, self.inv_freq)
        emb = torch.cat([freqs, freqs], dim=-1)    # (seq_len, dim)
        return emb

def apply_rotary_emb(x, rot_emb):
    # x: (B, n_head, seq_len, head_dim)
    # rot_emb: (seq_len, head_dim)
    seq_len = x.shape[-2]
    rot_emb = rot_emb[:seq_len]
    
    cos = torch.cos(rot_emb).unsqueeze(0).unsqueeze(0)  # (1, 1, seq_len, head_dim)
    sin = torch.sin(rot_emb).unsqueeze(0).unsqueeze(0)
    
    x_shape = x.shape
    x = x.reshape(*x_shape[:-1], -1, 2)  # (..., head_dim/2, 2)
    x1 = x[..., 0]  
    x2 = x[..., 1] 
    
    cos = cos.reshape(*cos.shape[:-1], -1, 2)[..., 0]
    sin = sin.reshape(*sin.shape[:-1], -1, 2)[..., 0]
    
    x1_rot = x1 * cos - x2 * sin
    x2_rot = x1 * sin + x2 * cos
    
    x_rot = torch.stack([x1_rot, x2_rot], dim=-1)
    return x_rot.reshape(x_shape)

class MultiHeadSelfAttention(nn.Module):
    def __init__(self, d_model, n_head, block_size):
        super().__init__()
        assert d_model % n_head == 0
        self.d_model = d_model
        self.n_head = n_head
        self.head_dim = d_model // n_head

        self.q_proj = nn.Linear(d_model, d_model, bias=False)
        self.k_proj = nn.Linear(d_model, d_model, bias=False)
        self.v_proj = nn.Linear(d_model, d_model, bias=False)
        self.o_proj = nn.Linear(d_model, d_model, bias=False)

        self.rotary_emb = RotaryEmbedding(self.head_dim)
        self.dropout = nn.Dropout(0.0)

        self.register_buffer(
            "tril", torch.tril(torch.ones(block_size, block_size)), persistent=False
        )

    def forward(self, x):
        # x: (B, T, d_model)
        B, T, C = x.shape

        q = self.q_proj(x)    # (B, T, d_model)
        k = self.k_proj(x)
        v = self.v_proj(x)

        q = q.view(B, T, self.n_head, self.head_dim).transpose(1, 2)  # (B, n_head, T, head_dim)
        k = k.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
        v = v.view(B, T, self.n_head, self.head_dim).transpose(1, 2)

        # Rotary embeddings
        rot_emb = self.rotary_emb(seq_len=T, device=x.device)   # (T, head_dim)
        q = apply_rotary_emb(q, rot_emb)
        k = apply_rotary_emb(k, rot_emb)

        # Attention
        attn_scores = torch.matmul(q, k.transpose(-2, -1)) * (self.head_dim ** -0.5)  # (B, n_head, T, T)
        attn_scores = attn_scores.masked_fill(self.tril[:T, :T] == 0, float('-inf'))
        attn_probs = F.softmax(attn_scores, dim=-1)
        attn_probs = self.dropout(attn_probs)

        out = torch.matmul(attn_probs, v)  # (B, n_head, T, head_dim)
        out = out.transpose(1, 2).contiguous().view(B, T, C)  # (B, T, d_model)

        return self.o_proj(out)


class TransformerMLP(nn.Module):
    def __init__(self, d_model):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(d_model, 4 * d_model),
            nn.GELU(),
            nn.Linear(4 * d_model, d_model),
            nn.Dropout(0.0),
        )

    def forward(self, x):
        return self.net(x)

class TransformerBlock(nn.Module):
    def __init__(self, d_model, n_head, block_size):
        super().__init__()
        self.self_attn = MultiHeadSelfAttention(d_model, n_head, block_size)
        self.mlp = TransformerMLP(d_model)
        self.input_layernorm = nn.LayerNorm(d_model)
        self.post_attention_layernorm = nn.LayerNorm(d_model)

    def forward(self, x):
        x = x + self.self_attn(self.input_layernorm(x))
        x = x + self.mlp(self.post_attention_layernorm(x))
        return x

class BVVForCausalLM(PreTrainedModel):
    config_class = BVVConfig
    
    def __init__(self, config):
        super().__init__(config)
        self.token_embeddings = nn.Embedding(config.vocab_size, config.n_embed)
        self.scale = config.scale
        self.transformer_layers = nn.Sequential(*[
            TransformerBlock(config.d_model, n_head=config.n_head, block_size=config.block_size) for _ in range(config.n_layer)
        ])
        self.final_layernorm = nn.LayerNorm(config.d_model)
        self.lm_head = nn.Linear(config.d_model, config.vocab_size)

        self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)


    def forward(self, idx, targets=None):
        B, T = idx.shape

        token_emb = self.token_embeddings(idx)  # (B, T, 16)
        
        token_emb = token_emb.repeat_interleave(self.scale, dim=-1)  # (B, T, 1024)

        x = token_emb
        
        x = self.transformer_layers(x)
        x = self.final_layernorm(x)
        logits = self.lm_head(x)

        loss = None
        if targets is not None:
            #logits_flat = logits.view(-1, logits.size(-1))
            #targets_flat = targets.view(-1)
            logits_flat = logits.reshape(-1, logits.size(-1))
            targets_flat = targets.reshape(-1)
            loss = F.cross_entropy(logits_flat, targets_flat, ignore_index = 57344)

        return CausalLMOutput(
            logits=logits,
            loss=loss,
        )


    def generate(self, 
                input_ids=None,
                max_new_tokens=None,
                max_length=None,
                temperature=1.0,
                top_k=None,
                top_p=None,
                do_sample=True,
                pad_token_id=None,
                eos_token_id=None,
                **kwargs):
        
        if input_ids is None:
            raise ValueError("Input_ids must be provided")
        
        idx = input_ids
        
        if max_new_tokens is None:
            if max_length is not None:
                max_new_tokens = max_length - idx.shape[1]
            else:
                max_new_tokens = 50  
        
        with torch.no_grad():
            for _ in range(max_new_tokens):
                idx_cond = idx[:, -self.config.block_size:]
                
                outputs = self(idx_cond)
                logits = outputs.logits[:, -1, :] / temperature
                
                if top_k is not None:
                    v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                    logits[logits < v[:, [-1]]] = float('-inf')
                
                if top_p is not None:
                    sorted_logits, sorted_indices = torch.sort(logits, descending=True)
                    cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
                    
                    sorted_indices_to_remove = cumulative_probs > top_p
                    sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
                    sorted_indices_to_remove[..., 0] = 0
                    
                    indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
                    logits[indices_to_remove] = float('-inf')
                
                probs = F.softmax(logits, dim=-1)
                
                if do_sample:
                    idx_next = torch.multinomial(probs, num_samples=1)
                else:
                    idx_next = torch.argmax(logits, dim=-1, keepdim=True)
                
                idx = torch.cat((idx, idx_next), dim=1)
                
    
                if eos_token_id is not None and (idx_next == eos_token_id).any():
                    break
        
        return idx