Updated model architecture

llama_diffusion_model.py

@@ -1,73 +1,172 @@
-import torch
 import torch.nn as nn
-import torch.nn.functional as F
+from transformers import AutoModelForCausalLM,  PreTrainedModel, PretrainedConfig
+from transformers.models.llama.modeling_llama import LlamaAttention
 from torch.amp import autocast
-from transformers import AutoModelForCausalLM, PreTrainedModel, PretrainedConfig
 from peft import LoraConfig, get_peft_model
+from typing import  Optional, Tuple
+import torch
 import os
 
+
+
 hf_token = os.getenv("HF_TOKEN")
 
-class BidirectionalLlamaAttention(nn.Module):
-    def __init__(self, original_layer, masking='unidirectional'):
-        super().__init__()
-        self.original = original_layer
+class BidirectionalLlamaAttention(LlamaAttention):
+    def __init__(self, original_layer, masking = 'unidirectional'):
+        super().__init__(original_layer.config, layer_idx=original_layer.layer_idx)
         self.masking = masking
 
-        self.q_proj = original_layer.q_proj
-        self.k_proj = original_layer.k_proj
-        self.v_proj = original_layer.v_proj
-        self.o_proj = original_layer.o_proj
+        # Copy weights from original layer
+        self.q_proj.weight = original_layer.q_proj.weight
+        self.k_proj.weight = original_layer.k_proj.weight
+        self.v_proj.weight = original_layer.v_proj.weight
+        self.o_proj.weight = original_layer.o_proj.weight
+
+    def repeat_kv(self, hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+        """
+        This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+        num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+        """
+        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)
+
+    def eager_attention_forward(self,
+        module: nn.Module,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        scaling: float,
+        dropout: float = 0.0,
+        **kwargs,
+        ):
+        key_states = self.repeat_kv(key, module.num_key_value_groups)
+        value_states = self.repeat_kv(value, module.num_key_value_groups)
+
+        attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1).to(query.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+        attn_output = attn_output.transpose(1, 2).contiguous()
+
+        return attn_output, attn_weights
 
-        self.head_dim = self.q_proj.out_features // original_layer.num_heads
-        self.num_heads = original_layer.num_heads
-        self.num_key_value_groups = original_layer.num_key_value_groups
-        self.attention_dropout = original_layer.attention_dropout
-        self.layer_idx = original_layer.layer_idx
-        self.scaling = original_layer.scaling
+    def rotate_half(self, x):
+        """Rotates half the hidden dims of the input."""
+        x1 = x[..., : x.shape[-1] // 2]
+        x2 = x[..., x.shape[-1] // 2 :]
 
-    def forward(self, hidden_states, position_embeddings, attention_mask=None, past_key_value=None, cache_position=None, **kwargs):
-        bsz, seq_len, _ = hidden_states.size()
+        return torch.cat((-x2, x1), dim=-1)
 
-        query_states = self._split_heads(self.q_proj(hidden_states))
-        key_states = self._split_heads(self.k_proj(hidden_states))
-        value_states = self._split_heads(self.v_proj(hidden_states))
 
+    def apply_rotary_pos_emb(self, q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+        """Applies Rotary Position Embedding to the query and key tensors.
+
+        Args:
+            q (`torch.Tensor`): The query tensor.
+            k (`torch.Tensor`): The key tensor.
+            cos (`torch.Tensor`): The cosine part of the rotary embedding.
+            sin (`torch.Tensor`): The sine part of the rotary embedding.
+            position_ids (`torch.Tensor`, *optional*):
+                Deprecated and unused.
+            unsqueeze_dim (`int`, *optional*, defaults to 1):
+                The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+                sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+                that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+                k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+                cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+                the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+        Returns:
+            `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+        """
+        cos = cos.unsqueeze(unsqueeze_dim)
+        sin = sin.unsqueeze(unsqueeze_dim)
+        q_embed = (q * cos) + (self.rotate_half(q) * sin)
+        k_embed = (k * cos) + (self.rotate_half(k) * sin)
+
+        return q_embed, k_embed
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[torch.Tensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ):
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        # Apply rotary embeddings
         cos, sin = position_embeddings
-        query_states, key_states = self._apply_rotary(query_states, key_states, cos, sin)
+        query_states, key_states = self.apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # 🔄 **Modify the Attention Mask**
+        seq_len = hidden_states.shape[1]
+        batch_size = hidden_states.shape[0]
         if self.masking == 'bidirectional':
-            attn_mask = torch.ones((bsz, 1, seq_len, seq_len), device=hidden_states.device)
-        else:
-            attn_mask = torch.tril(torch.ones(seq_len, seq_len, device=hidden_states.device)).unsqueeze(0).unsqueeze(0)
+            base_mask = torch.ones((seq_len, seq_len), device=hidden_states.device, dtype=torch.bool)
+            attn_mask = base_mask.unsqueeze(0).unsqueeze(1).expand(batch_size, 1, seq_len, seq_len).clone()  # ✅ Copy for each batch
+        elif self.masking == 'bidirectional_masked':
+            base_mask = torch.ones((seq_len, seq_len), device=hidden_states.device, dtype=torch.bool)
+            base_mask[:, 1:].fill_diagonal_(False)  # ✅ Apply diagonal masking only in 2D
+            attn_mask = base_mask.unsqueeze(0).unsqueeze(1).expand(batch_size, 1, seq_len, seq_len).clone()  # ✅ Copy for each batch
+        else: # unidirectional
+            # 🚀 Standard autoregressive (causal) mask
+            attn_mask = torch.tril(torch.ones(seq_len, seq_len, device=hidden_states.device, dtype=torch.bool))
+            attn_mask = base_mask.unsqueeze(0).unsqueeze(1).expand(batch_size, 1, seq_len, seq_len).clone()  # ✅ Copy for each batch
+
+
+        # Call the default attention function
+        attn_output, attn_weights = self.eager_attention_forward(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attn_mask,  # ✅ Custom mask is applied here
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
 
-        attn_weights = torch.matmul(query_states, key_states.transpose(-2, -1)) * self.scaling
-        attn_weights = attn_weights + attn_mask.log()
-        attn_weights = F.softmax(attn_weights, dim=-1)
-        attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
 
-        attn_output = torch.matmul(attn_weights, value_states)
-        attn_output = self._merge_heads(attn_output)
-        return self.o_proj(attn_output), attn_weights
+        return attn_output, attn_weights
 
-    def _split_heads(self, x):
-        return x.view(x.size(0), x.size(1), self.num_heads, self.head_dim).transpose(1, 2)
 
-    def _merge_heads(self, x):
-        return x.transpose(1, 2).contiguous().view(x.size(0), -1, self.num_heads * self.head_dim)
-
-    def _apply_rotary(self, q, k, cos, sin):
-        cos = cos.unsqueeze(1)
-        sin = sin.unsqueeze(1)
-        q_rot = (q * cos) + (self._rotate_half(q) * sin)
-        k_rot = (k * cos) + (self._rotate_half(k) * sin)
-        return q_rot, k_rot
-
-    def _rotate_half(self, x):
-        x1 = x[..., : x.shape[-1] // 2]
-        x2 = x[..., x.shape[-1] // 2 :]
-        return torch.cat((-x2, x1), dim=-1)
+    def _split_heads(self, tensor, num_heads, attn_head_size):
+        """
+        Splits hidden_size dim into attn_head_size and num_heads
+        """
+        new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
+        tensor = tensor.view(*new_shape)
+        return tensor.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)
 
+    def _merge_heads(self, tensor, num_heads, attn_head_size):
+        """
+        Merges attn_head_size dim and num_attn_heads dim into hidden_size
+        """
+        tensor = tensor.permute(0, 2, 1, 3).contiguous()
+        new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
+        return tensor.view(new_shape)
 
 class CustomTransformerConfig(PretrainedConfig):
     def __init__(self, vocab_size=128256, hidden_size=4096, num_layers=32, num_heads=32, prediction_chunk=256, dropout=0, max_position_embeddings=4096, **kwargs):
@@ -79,7 +178,7 @@ class CustomTransformerConfig(PretrainedConfig):
         self.dropout = dropout
         self.prediction_chunk = prediction_chunk
         self.max_position_embeddings = max_position_embeddings
-
+        self.input_size = prediction_chunk
 
 class CustomTransformerModel(PreTrainedModel):
     config_class = CustomTransformerConfig
@@ -87,12 +186,15 @@ class CustomTransformerModel(PreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
 
-        self.llama = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-3B", torch_dtype=torch.float16, token=hf_token)
+        # Load pre-trained Llama model (excluding its original lm_head)
+        self.llama = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-3B", torch_dtype=torch.float16, device_map="auto", token = hf_token)
+
         self.llama.resize_token_embeddings(config.vocab_size)
 
         for i, layer in enumerate(self.llama.model.layers):
             layer.self_attn = BidirectionalLlamaAttention(layer.self_attn, masking='bidirectional')
 
+        # Freeze Llama to retain pre-trained knowledge
         for param in self.llama.parameters():
             param.requires_grad = False
 
@@ -103,27 +205,40 @@ class CustomTransformerModel(PreTrainedModel):
             r=256,
             lora_alpha=256,
             lora_dropout=0.0,
-            target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
+            target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],  # Llama-3 uses these attention modules
             bias="none",
             task_type=None
         )
 
         self.llama = get_peft_model(self.llama, lora_config)
+        self.llama.print_trainable_parameters()  # Print number of trainable parameters
         self.llama = self.llama.to(torch.float16)
 
     def forward(self, input_ids, labels=None, **kwargs):
         batch_size, seq_length = input_ids.shape
-        assert seq_length == self.config.prediction_chunk
+        assert seq_length == self.input_size, f"Expected input length input_size, got {seq_length}"
+
+        with autocast("cuda", dtype=torch.float16):  # ✅ Correct future-proof usage
+
 
-        with autocast("cuda", dtype=torch.float16):
-            outputs = self.llama(input_ids=input_ids, output_hidden_states=True, **kwargs)
-            logits = outputs.logits[:, :, :self.config.vocab_size].view(batch_size, self.config.prediction_chunk, self.config.vocab_size)
+            outputs = self.llama(input_ids, output_hidden_states=True, **kwargs)
+
+            logits = outputs.logits[:,:,:self.config.vocab_size]
+
+            # Reshape logits to (batch, input_size, vocab_size)
+            logits = logits.view(batch_size, self.config.prediction_chunk, self.config.vocab_size)
+
+            loss = None
 
-        loss = None
         if labels is not None:
-            loss_fct = nn.CrossEntropyLoss()
+            assert labels.shape == (batch_size, self.input_size), f"Labels shape mismatch: expected (batch, input_size), got {labels.shape}"
+
+            # Compute loss
+            loss_fct = torch.nn.CrossEntropyLoss()
+
             loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
 
+
         return {"loss": loss, "logits": logits} if loss is not None else {"logits": logits}