Upload 3 files

config.json

@@ -1,9 +1,20 @@
 {
-  "model_type": "PreTrainedTokenizerFast",
-  "model_type": "gpt2",
+  "architectures": ["BVVProForCausalLM"],
+  "auto_map": {
+    "AutoConfig": "modeling_bvv_pro.BVVProConfig",
+    "AutoModel": "modeling_bvv_pro.BVVProForCausalLM",
+    "AutoModelForCausalLM": "modeling_bvv_pro.BVVProForCausalLM"
+  },
+  "model_type": "bvv_pro",
+  "vocab_size": 65536,
+  "block_size ": 1024,
+  "n_embd": 1024,
+  "n_layer": 8,
+  "n_head": 8,
+  "pad_id": 57344,
   "bos_token": "<s>",
   "eos_token": "</s>",
   "unk_token": "<unk>",
   "pad_token": "<pad>",
-  "vocab_size": 65536
+  "torch_dtype": "float32"
 }

modeling_bvv_pro.py

@@ -0,0 +1,156 @@
+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 BVVProConfig(PretrainedConfig):
+    model_type = "bvv_pro"
+    
+    def __init__(
+        self,
+        vocab_size = 65536,
+        n_embd = 1024,
+        n_head = 8,  
+        n_layer = 8, 
+        block_size = 1024,
+        pad_id = 57344,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.block_size  = block_size 
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.pad_id = pad_id 
+
+class SimpleSelfAttentionHead(nn.Module):
+    def __init__(self, head_size, n_embd, block_size):
+        super().__init__()
+        self.q_proj = nn.Linear(n_embd, head_size, bias=False)
+        self.k_proj = nn.Linear(n_embd, head_size, bias=False)
+        self.v_proj = nn.Linear(n_embd, head_size, bias=False)
+        self.o_proj = nn.Linear(head_size, head_size, bias=False)
+
+        self.dropout = nn.Dropout(0.0)
+        self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))
+
+    def forward(self, x):
+        B, T, C = x.shape
+        q = self.q_proj(x)     # (B,T,head_size)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        attn_scores = q @ k.transpose(-2, -1) * (k.shape[-1] ** -0.5)  # (B,T,T)
+        
+        attn_scores = attn_scores.masked_fill(self.tril[:T, :T] == 0, torch.finfo(attn_scores.dtype).min) #float('-inf'))
+
+        attn_probs = F.softmax(attn_scores, dim=-1)
+        attn_probs = self.dropout(attn_probs)
+
+        out = attn_probs @ v   # (B,T,head_size)
+        out = self.o_proj(out) # (B,T,head_size)
+
+        return out
+
+class SimpleMultiHeadSelfAttention(nn.Module):
+    def __init__(self, n_embd, n_head, block_size):
+        super().__init__()
+        self.head_size = n_embd // n_head
+        self.heads = nn.ModuleList([SimpleSelfAttentionHead(self.head_size, n_embd, block_size) for _ in range(n_head)])
+        self.out_proj = nn.Linear(n_head * self.head_size, n_embd)
+        self.dropout = nn.Dropout(0.0)
+
+    def forward(self, x):
+        out = torch.cat([head(x) for head in self.heads], dim=-1) 
+        out = self.dropout(self.out_proj(out))
+        return out
+
+class TransformerMLP(nn.Module):
+    def __init__(self, n_embd):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(n_embd, 4 * n_embd),
+            nn.GELU(),
+            nn.Linear(4 * n_embd, n_embd),
+            nn.Dropout(0.0),
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+class TransformerBlock(nn.Module):
+    def __init__(self, n_embd, n_head, block_size):
+        super().__init__()
+        self.self_attn = SimpleMultiHeadSelfAttention(n_embd, n_head, block_size)
+        self.mlp = TransformerMLP(n_embd)
+        self.input_layernorm = nn.LayerNorm(n_embd)
+        self.post_attention_layernorm = nn.LayerNorm(n_embd)
+
+    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 BVVProForCausalLM(PreTrainedModel):
+    config_class = BVVProConfig
+    
+    def __init__(self, config):
+        super().__init__(config)
+        self.token_embeddings = nn.Embedding(config.vocab_size, config.n_embd)
+        self.position_embeddings = nn.Embedding(config.block_size, config.n_embd)
+        self.transformer_layers = nn.Sequential(*[
+            TransformerBlock(config.n_embd, n_head=config.n_head, block_size=config.block_size) for _ in range(config.n_layer)
+        ])
+        self.final_layernorm = nn.LayerNorm(config.n_embd)
+        self.lm_head = nn.Linear(config.n_embd, 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
+
+        positions = torch.arange(0, T, device=idx.device).unsqueeze(0).expand(B, T)
+
+        x = self.token_embeddings(idx) + self.position_embeddings(positions)
+        
+        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, idx, max_new_tokens):
+        with torch.no_grad():
+            for _ in range(max_new_tokens):
+                idx_cond = idx[:, -self.config.block_size:]
+                outputs = self(idx_cond)
+                logits = outputs.logits
+                logits = logits[:, -1, :]  
+                probs = F.softmax(logits, dim=-1)
+                idx_next = torch.multinomial(probs, num_samples=1)
+                idx = torch.cat((idx, idx_next), dim=1)
+
+        return idx  

tokenizer_config.json

@@ -1,6 +1,9 @@
 {
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "model_type": "gpt2",
+  "bos_token": "<s>",
+  "eos_token": "</s>",
   "unk_token": "<unk>",
   "pad_token": "<pad>",
-  "bos_token": "<s>",
-  "eos_token": "</s>"
+  "vocab_size": 65536
 }