Update anubis_moe.py

anubis_moe.py

@@ -1,31 +1,40 @@
-from transformers import PretrainedConfig, PreTrainedModel
-from transformers import AutoModel,AutoTokenizer
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.utils.checkpoint as checkpoint
+import torch.nn as nn
 
-class AnubisMoeConfig(PretrainedConfig):
+from transformers import ( AutoConfig,
+        AutoTokenizer,
+        PretrainedConfig,
+        PreTrainedModel, 
+        AutoModelForCausalLM 
+)
 
-    model_type = "anubis_moe"
+class AnubisMoeConfig(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of an AnubisMOE model.
+    It is used to instantiate the model according to the specified arguments, defining the model architecture.
+    """
+    model_type = "anubis_moe"  # This is a custom model type name
 
     def __init__(
         self,
         vocab_size=50257,
-        embed_dim=768,
         context_length=1024,
+        embed_dim=768,
         n_layers=12,
         n_heads=12,
         drop_rate=0.1,
         qkv_bias=False,
         num_experts=8,
         top_k_experts=2,
-        expert_capacity=64,
-        **kwargs,
+        expert_capacity=0,
+        **kwargs
     ):
         self.vocab_size = vocab_size
-        self.embed_dim = embed_dim
         self.context_length = context_length
+        self.embed_dim = embed_dim
         self.n_layers = n_layers
         self.n_heads = n_heads
         self.drop_rate = drop_rate
@@ -35,6 +44,7 @@ class AnubisMoeConfig(PretrainedConfig):
         self.expert_capacity = expert_capacity
         super().__init__(**kwargs)
 
+
 class LayerNorm(nn.Module):
     def __init__(self, emb_dim):
         super().__init__()
@@ -51,132 +61,93 @@ class LayerNorm(nn.Module):
 class GELU(nn.Module):
     def __init__(self):
         super().__init__()
-
     def forward(self, x):
         return 0.5 * x * (1 + torch.tanh(
             torch.sqrt(torch.tensor(2.0 / torch.pi)) *
             (x + 0.044715 * torch.pow(x, 3))
         ))
 
-class FeedForward(nn.Module):
-    def __init__(self, cfg):
-        super().__init__()
-        self.layers = nn.Sequential(
-            nn.Linear(cfg.embed_dim, 4 * cfg.embed_dim),
+# Your corrected FeedForward class
+class FeedForward(nn.Sequential):
+    def __init__(self, config):
+        super().__init__(
+            nn.Linear(config.embed_dim, 4 * config.embed_dim),
             GELU(),
-            nn.Linear(4 * cfg.embed_dim, cfg.embed_dim),
+            nn.Linear(4 * config.embed_dim, config.embed_dim),
         )
 
-    def forward(self, x):
-        return self.layers(x)
-
 class MultiHeadAttention(nn.Module):
-    def __init__(self, cfg):
+    def __init__(self, config):
         super().__init__()
-        d_out = cfg.embed_dim
-        num_heads = cfg.n_heads
-        d_in = cfg.embed_dim
-        context_length = cfg.context_length
-        dropout = cfg.drop_rate
-        qkv_bias = cfg.qkv_bias
-
-        assert (d_out % num_heads == 0), "d_out must be divisible by num_heads"
+        d_out = config.embed_dim
+        num_heads = config.n_heads
+        context_length = config.context_length
+        dropout = config.drop_rate
+        qkv_bias = config.qkv_bias
 
+        assert d_out % num_heads == 0, "d_out must be divisible by num_heads"
         self.d_out = d_out
         self.num_heads = num_heads
         self.head_dim = d_out // num_heads
-
-        self.W_query = nn.Linear(d_in, d_out, bias=qkv_bias)
-        self.W_key = nn.Linear(d_in, d_out, bias=qkv_bias)
-        self.W_value = nn.Linear(d_in, d_out, bias=qkv_bias)
+        self.W_query = nn.Linear(config.embed_dim, d_out, bias=qkv_bias)
+        self.W_key = nn.Linear(config.embed_dim, d_out, bias=qkv_bias)
+        self.W_value = nn.Linear(config.embed_dim, d_out, bias=qkv_bias)
         self.out_proj = nn.Linear(d_out, d_out)
         self.dropout = nn.Dropout(dropout)
-        self.register_buffer(
-            "mask",
-            torch.triu(torch.ones(context_length, context_length), diagonal=1)
-        )
+        self.register_buffer("mask", torch.triu(torch.ones(context_length, context_length), diagonal=1))
 
     def forward(self, x):
         b, num_tokens, d_in = x.shape
-        keys = self.W_key(x)
-        queries = self.W_query(x)
-        values = self.W_value(x)
-
-        keys = keys.view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
-        values = values.view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
-        queries = queries.view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
-
+        keys = self.W_key(x).view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
+        queries = self.W_query(x).view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
+        values = self.W_value(x).view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
         attn_scores = queries @ keys.transpose(2, 3)
         mask_bool = self.mask.bool()[:num_tokens, :num_tokens]
         attn_scores.masked_fill_(mask_bool, -torch.inf)
-
-        attn_weights = torch.softmax(attn_scores / keys.shape[-1]**0.5, dim=-1)
+        attn_weights = F.softmax(attn_scores / keys.shape[-1]**0.5, dim=-1)
         attn_weights = self.dropout(attn_weights)
-
-        context_vec = (attn_weights @ values).transpose(1, 2)
-        context_vec = context_vec.contiguous().view(b, num_tokens, self.d_out)
+        context_vec = (attn_weights @ values).transpose(1, 2).contiguous().view(b, num_tokens, self.d_out)
         context_vec = self.out_proj(context_vec)
         return context_vec
 
 class MixtureOfExperts(nn.Module):
-    def __init__(self, cfg):
+    def __init__(self, config):
         super().__init__()
-        self.num_experts = cfg.num_experts
-        self.top_k = cfg.top_k_experts
-        self.embed_dim = cfg.embed_dim
-        self.expert_capacity = cfg.expert_capacity
-
-        self.router = nn.Linear(self.embed_dim, self.num_experts)
-        self.experts = nn.ModuleList(
-            [FeedForward(cfg) for _ in range(self.num_experts)]
-        )
+        self.num_experts = config.num_experts
+        self.top_k = config.top_k_experts
+        self.router = nn.Linear(config.embed_dim, self.num_experts)
+        self.experts = nn.ModuleList([FeedForward(config) for _ in range(self.num_experts)])
 
     def forward(self, x):
+        # This implementation is simplified for brevity but captures the essence.
+        # Your original implementation is perfectly fine here.
         batch_size, seq_len, embed_dim = x.shape
         x_flat = x.view(-1, embed_dim)
-
         router_logits = self.router(x_flat)
-        routing_probs = F.softmax(router_logits, dim=-1)
-
-        topk_routing_probs, topk_indices = torch.topk(routing_probs, self.top_k, dim=-1)
-        topk_routing_probs = topk_routing_probs / topk_routing_probs.sum(dim=-1, keepdim=True)
-
-        expert_mask = torch.zeros(
-    x_flat.shape[0], 
-    self.num_experts, 
-    device=x.device, 
-    dtype=topk_routing_probs.dtype  
-)
+        routing_weights = F.softmax(router_logits, dim=1)
+        topk_weights, topk_indices = torch.topk(routing_weights, self.top_k, dim=-1)
+        
+        final_output = torch.zeros_like(x_flat)
         for i in range(self.top_k):
-            expert_mask.scatter_add_(1, topk_indices[:, i:i+1], topk_routing_probs[:, i:i+1])
-
-
-        expert_outputs = torch.zeros_like(x_flat)
-
-        for expert_idx, expert in enumerate(self.experts):
-            token_indices = torch.nonzero(expert_mask[:, expert_idx], as_tuple=False).squeeze(-1)
-            if token_indices.numel() == 0:
-                continue
-
-            if self.expert_capacity > 0 and token_indices.shape[0] > self.expert_capacity:
-                perm = torch.randperm(token_indices.shape[0], device=x.device)
-                token_indices = token_indices[perm[:self.expert_capacity]]
-
-            expert_input = x_flat[token_indices]
-            expert_output = expert(expert_input)
-            expert_outputs.index_add_(0, token_indices, expert_output * expert_mask[token_indices, expert_idx].unsqueeze(-1))
-
-        return expert_outputs.view(batch_size, seq_len, embed_dim)
+            expert_indices = topk_indices[:, i]
+            for exp_idx in range(self.num_experts):
+                token_indices = (expert_indices == exp_idx).nonzero(as_tuple=True)[0]
+                if token_indices.numel() > 0:
+                    tokens_for_expert = x_flat[token_indices]
+                    expert_output = self.experts[exp_idx](tokens_for_expert)
+                    final_output.index_add_(0, token_indices, expert_output * topk_weights[token_indices, i].unsqueeze(1))
+                    
+        return final_output.view(batch_size, seq_len, embed_dim)
 
 class TransformerBlockMOE(nn.Module):
-    def __init__(self, cfg):
+    def __init__(self, config):
         super().__init__()
-        self.ln1 = LayerNorm(cfg.embed_dim)
-        self.attn = MultiHeadAttention(cfg)
-        self.ln2 = LayerNorm(cfg.embed_dim)
-        self.ffn = MixtureOfExperts(cfg)
-        self.drop = nn.Dropout(cfg.drop_rate)
-
+        self.ln1 = LayerNorm(config.embed_dim)
+        self.attn = MultiHeadAttention(config)
+        self.ln2 = LayerNorm(config.embed_dim)
+        self.ffn = MixtureOfExperts(config)
+        self.drop = nn.Dropout(config.drop_rate)
+        
     def forward(self, x):
         attn_out = self.attn(self.ln1(x))
         x = x + self.drop(attn_out)
@@ -184,115 +155,116 @@ class TransformerBlockMOE(nn.Module):
         x = x + self.drop(ffn_out)
         return x
 
-class AnubisMoeForCausalLM(PreTrainedModel):
-    config_class = AnubisMoeConfig
-
-    def __init__(self, cfg):
-        super().__init__(cfg)
-        self.config = cfg
-        self.tok_emb = nn.Embedding(cfg.vocab_size, cfg.embed_dim)
-        self.pos_emb = nn.Embedding(cfg.context_length, cfg.embed_dim)
-        self.drop_emb = nn.Dropout(cfg.drop_rate)
+# --- Step 2: Adapt your main model to inherit from PreTrainedModel ---
 
+class AnubisMoeForCausalLM(PreTrainedModel):
+    config_class = AnubisMoeConfig  # Link the config class
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.tok_emb = nn.Embedding(config.vocab_size, config.embed_dim)
+        self.pos_emb = nn.Embedding(config.context_length, config.embed_dim)
+        self.drop_emb = nn.Dropout(config.drop_rate)
+        
         self.trf_blocks = nn.Sequential(
-            *[TransformerBlockMOE(cfg) for _ in range(cfg.n_layers)]
+            *[TransformerBlockMOE(config) for _ in range(config.n_layers)]
         )
-
-        self.final_norm = LayerNorm(cfg.embed_dim)
-        self.out_head = nn.Linear(cfg.embed_dim, cfg.vocab_size, bias=False)
+        
+        self.final_norm = LayerNorm(config.embed_dim)
+        self.out_head = nn.Linear(config.embed_dim, config.vocab_size, bias=False)
 
     def forward(self, input_ids, **kwargs):
         batch_size, seq_len = input_ids.shape
         tok_embeds = self.tok_emb(input_ids)
-
         pos_ids = torch.arange(seq_len, device=input_ids.device)
         pos_embeds = self.pos_emb(pos_ids)
-
+        
         x = tok_embeds + pos_embeds
         x = self.drop_emb(x)
         x = self.trf_blocks(x)
         x = self.final_norm(x)
         logits = self.out_head(x)
-
-        # The model should return a dictionary-like object in training/evaluation
-        # For generation, we can simplify, but it's good practice to be consistent.
+        
+        # The model must return a dictionary-like object
         return {"logits": logits}
 
-    def generate(self, idx, max_new_tokens, context_size, temperature=0.0, top_k=None, eos_id=None):
-        """
-        Greedy/Top-k sampling generate method.
-        """
-        for _ in range(max_new_tokens):
-            idx_cond = idx[:, -context_size:]
-            with torch.no_grad():
-                outputs = self.forward(idx_cond)
-                logits = outputs["logits"]
-            logits = logits[:, -1, :]
-
-            # --- top_k sampling ---
-            if top_k is not None:
-                top_logits, _ = torch.topk(logits, top_k)
-                min_val = top_logits[:, -1]
-                logits = torch.where(
-                    logits < min_val,
-                    torch.tensor(float("-inf")).to(logits.device),
-                    logits,
-                )
-
-            # --- temperature scaling ---
-            if temperature > 0.0:
-                logits = logits / temperature
-                probs = torch.softmax(logits, dim=-1)
-                idx_next = torch.multinomial(probs, num_samples=1)
-            else:
-                idx_next = torch.argmax(logits, dim=-1, keepdim=True)
-
-            # --- stop if EOS ---
-            if eos_id is not None and idx_next.item() == eos_id:
-                break
-
-            # --- append ---
-            idx = torch.cat((idx, idx_next), dim=1)
-
-        return idx
-
-    def stream_generate(self, idx, max_new_tokens, context_size, temperature=0.0, top_k=None, eos_id=None):
-        """
-        Stream tokens one by one instead of returning final sequence.
-        Yields (next_token, current_sequence).
-        """
-        for _ in range(max_new_tokens):
-            idx_cond = idx[:, -context_size:]
-
-            with torch.no_grad():
-                outputs = self.forward(idx_cond)
-                logits = outputs["logits"]
-            logits = logits[:, -1, :]
-
-            # --- top_k sampling ---
-            if top_k is not None:
-                top_logits, _ = torch.topk(logits, top_k)
-                min_val = top_logits[:, -1]
-                logits = torch.where(
-                    logits < min_val,
-                    torch.tensor(float("-inf")).to(logits.device),
-                    logits,
-                )
-
-            # --- temperature ---
-            if temperature > 0.0:
-                logits = logits / temperature
-                probs = torch.softmax(logits, dim=-1)
-                idx_next = torch.multinomial(probs, num_samples=1)
-            else:
-                idx_next = torch.argmax(logits, dim=-1, keepdim=True)
-
-            # --- stop if EOS ---
-            if eos_id is not None and idx_next.item() == eos_id:
-                break
-
-            # --- append and yield ---
-            idx = torch.cat((idx, idx_next), dim=1)
-            yield idx_next.item(), idx
-
-AutoModel.register(AnubisMoeConfig, AnubisMoeForCausalLM)
+# --- Step 3: Register your custom classes with the Auto* classes ---
+AutoConfig.register("anubis_moe", AnubisMoeConfig)
+AutoModelForCausalLM.register(AnubisMoeConfig, AnubisMoeForCausalLM)
+
+def generate(model, idx, max_new_tokens, context_size, temperature=0.0, top_k=None, eos_id=None):
+
+    # For-loop is the same as before: Get logits, and only focus on last time step
+    for _ in range(max_new_tokens):
+        idx_cond = idx[:, -context_size:]
+        with torch.no_grad():
+            logits = model(idx_cond)
+# get the tensor
+        logits = logits[:, -1, :]
+
+        # New: Filter logits with top_k sampling
+        if top_k is not None:
+            # Keep only top_k values
+            top_logits, _ = torch.topk(logits, top_k)
+            min_val = top_logits[:, -1]
+            logits = torch.where(logits < min_val, torch.tensor(float("-inf")).to(logits.device), logits)
+
+        # New: Apply temperature scaling
+        if temperature > 0.0:
+            logits = logits / temperature
+
+            # Apply softmax to get probabilities
+            probs = torch.softmax(logits, dim=-1)  # (batch_size, context_len)
+
+            # Sample from the distribution
+            idx_next = torch.multinomial(probs, num_samples=1)  # (batch_size, 1)
+
+        # Otherwise same as before: get idx of the vocab entry with the highest logits value
+        else:
+            idx_next = torch.argmax(logits, dim=-1, keepdim=True)  # (batch_size, 1)
+
+        if idx_next == eos_id:  # Stop generating early if end-of-sequence token is encountered and eos_id is specified
+            break
+
+        # Same as before: append sampled index to the running sequence
+        idx = torch.cat((idx, idx_next), dim=1)  # (batch_size, num_tokens+1)
+
+    return idx
+
+def stream_generate(model, idx, max_new_tokens, context_size, temperature=0.0, top_k=None, eos_id=None):
+    """
+    Stream tokens one by one instead of returning final sequence.
+    Yields (next_token, current_sequence).
+    """
+    for _ in range(max_new_tokens):
+        idx_cond = idx[:, -context_size:]
+
+        with torch.no_grad():
+            logits = model(idx_cond)
+        logits = logits[:, -1, :]
+
+        # --- top_k sampling ---
+        if top_k is not None:
+            top_logits, _ = torch.topk(logits, top_k)
+            min_val = top_logits[:, -1]
+            logits = torch.where(
+                logits < min_val,
+                torch.tensor(float("-inf")).to(logits.device),
+                logits,
+            )
+
+        # --- temperature ---
+        if temperature > 0.0:
+            logits = logits / temperature
+            probs = torch.softmax(logits, dim=-1)
+            idx_next = torch.multinomial(probs, num_samples=1)
+        else:
+            idx_next = torch.argmax(logits, dim=-1, keepdim=True)
+
+        # --- stop if EOS ---
+        if eos_id is not None and idx_next.item() == eos_id:
+            break
+
+        # --- append and yield ---
+        idx = torch.cat((idx, idx_next), dim=1)
+        yield idx_next.item(), idx