Upload folder using huggingface_hub

config.json

@@ -4,16 +4,22 @@
     ],
     "model_type": "smdm",
     "vocab_size": 32000,
-    "hidden_size": 2048,
-    "num_hidden_layers": 22,
-    "num_attention_heads": 32,
+    "n_embd": 2048,
+    "n_layer": 22,
+    "n_head": 32,
+    "n_query_groups": 32,
     "intermediate_size": 5632,
     "hidden_dropout_prob": 0.0,
     "attention_probs_dropout_prob": 0.0,
-    "max_position_embeddings": 2048,
+    "block_size": 2048,
     "initializer_range": 0.02,
-    "layer_norm_eps": 1e-5,
+    "norm_eps": 1e-5,
     "use_cache": true,
+    "rotary_percentage": 1.0,
+    "condense_ratio": 1,
+    "parallel_residual": true,
+    "shared_attention_norm": false,
+    "bias": true,
     "bos_token_id": 1,
     "eos_token_id": 2,
     "pad_token_id": 0,

modeling_smdm.py

@@ -1,8 +1,11 @@
+import math
 import torch
 import torch.nn as nn
 from transformers import PreTrainedModel, PretrainedConfig
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from typing import Optional, Tuple, Union
+from typing import Optional, Tuple, Union, List
+from flash_attn import flash_attn_func
+from xformers.ops import SwiGLU
 
 class SMDMConfig(PretrainedConfig):
     model_type = "smdm"
@@ -10,30 +13,44 @@ class SMDMConfig(PretrainedConfig):
     def __init__(
         self,
         vocab_size: int = 32000,
-        hidden_size: int = 2048,
-        num_hidden_layers: int = 22,
-        num_attention_heads: int = 32,
+        n_embd: int = 2048,
+        n_layer: int = 22,
+        n_head: int = 32,
+        n_query_groups: int = 32,
         intermediate_size: int = 5632,
         hidden_dropout_prob: float = 0.0,
         attention_probs_dropout_prob: float = 0.0,
-        max_position_embeddings: int = 2048,
+        block_size: int = 2048,
         initializer_range: float = 0.02,
-        layer_norm_eps: float = 1e-5,
+        norm_eps: float = 1e-5,
         use_cache: bool = True,
+        rotary_percentage: float = 1.0,
+        condense_ratio: int = 1,
+        parallel_residual: bool = True,
+        shared_attention_norm: bool = False,
+        bias: bool = True,
         **kwargs
     ):
         super().__init__(**kwargs)
         self.vocab_size = vocab_size
-        self.hidden_size = hidden_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_query_groups = n_query_groups
         self.intermediate_size = intermediate_size
         self.hidden_dropout_prob = hidden_dropout_prob
         self.attention_probs_dropout_prob = attention_probs_dropout_prob
-        self.max_position_embeddings = max_position_embeddings
+        self.block_size = block_size
         self.initializer_range = initializer_range
-        self.layer_norm_eps = layer_norm_eps
+        self.norm_eps = norm_eps
         self.use_cache = use_cache
+        self.rotary_percentage = rotary_percentage
+        self.condense_ratio = condense_ratio
+        self.parallel_residual = parallel_residual
+        self.shared_attention_norm = shared_attention_norm
+        self.bias = bias
+        self.head_size = n_embd // n_head
+        self.padded_vocab_size = vocab_size
 
 class SMDMForCausalLM(PreTrainedModel):
     config_class = SMDMConfig
@@ -43,31 +60,54 @@ class SMDMForCausalLM(PreTrainedModel):
         super().__init__(config)
         self.config = config
         
-        # Initialize model components here
-        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
-        self.layers = nn.ModuleList([SMDMBlock(config) for _ in range(config.num_hidden_layers)])
-        self.norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
-        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        # Initialize model components
+        self.lm_head = nn.Linear(config.n_embd, config.padded_vocab_size, bias=False)
+        self.transformer = nn.ModuleDict(
+            dict(
+                wte=nn.Embedding(config.padded_vocab_size + 1, config.n_embd),
+                h=nn.ModuleList([SMDMBlock(config) for _ in range(config.n_layer)]),
+                ln_f=nn.LayerNorm(config.n_embd, eps=config.norm_eps),
+            )
+        )
+        self.rope_cache = None
         
         # Initialize weights
-        self.apply(self._init_weights)
+        self.apply(lambda module: self._init_weights(module, config.n_layer))
         
-    def _init_weights(self, module):
-        if isinstance(module, nn.Linear):
-            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+    def _init_weights(self, module, n_layer):
+        if isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=math.sqrt(2.0 / 5 / self.config.n_embd))
+        elif isinstance(module, nn.Linear):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=math.sqrt(2.0 / 5 / self.config.n_embd))
             if module.bias is not None:
-                module.bias.data.zero_()
-        elif isinstance(module, nn.Embedding):
-            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
-            if module.padding_idx is not None:
-                module.weight.data[module.padding_idx].zero_()
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, (SMDMMLP, SMDMAttention)):
+            if hasattr(module, 'proj'):
+                nn.init.normal_(module.proj.weight, mean=0.0, std=1 / math.sqrt(self.config.n_embd) / n_layer)
+                
+    def build_rope_cache(self, idx: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        seq_len = self.config.block_size
+        n_elem = int(self.config.rotary_percentage * self.config.head_size)
+        base = 10000
+        
+        theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, device=idx.device) / n_elem))
+        seq_idx = torch.arange(seq_len, device=idx.device) / self.config.condense_ratio
+        idx_theta = torch.outer(seq_idx, theta)
+        
+        cos, sin = torch.cos(idx_theta), torch.sin(idx_theta)
+        
+        if idx.dtype == torch.bfloat16:
+            return cos.bfloat16(), sin.bfloat16()
+        if idx.dtype in (torch.float16, torch.bfloat16, torch.int8):
+            return cos.half(), sin.half()
+        return cos, sin
                 
     def forward(
         self,
         input_ids: torch.LongTensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[List[Tuple[torch.FloatTensor]]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         labels: Optional[torch.LongTensor] = None,
         use_cache: Optional[bool] = None,
@@ -75,63 +115,123 @@ class SMDMForCausalLM(PreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
     ) -> Union[Tuple, CausalLMOutputWithPast]:
-        # Implementation of forward pass
-        # This is a placeholder - you'll need to implement the actual forward logic
-        # based on your model's architecture
+        B, T = input_ids.size()
+        
+        if self.rope_cache is None:
+            self.rope_cache = self.build_rope_cache(input_ids)
+            
+        cos, sin = self.rope_cache
+        cos = cos[:T]
+        sin = sin[:T]
+        
+        x = self.transformer.wte(input_ids)
         
+        for block in self.transformer.h:
+            x = block(x, (cos, sin))
+            
+        x = self.transformer.ln_f(x)
+        logits = self.lm_head(x)
+        
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+            
         return CausalLMOutputWithPast(
-            loss=None,  # Implement loss calculation if labels are provided
-            logits=None,  # Implement logits calculation
-            past_key_values=None,  # Implement past key values if using cache
-            hidden_states=None,  # Implement hidden states if requested
-            attentions=None,  # Implement attention weights if requested
+            loss=loss,
+            logits=logits,
+            past_key_values=None,
+            hidden_states=None,
+            attentions=None,
         )
 
 class SMDMBlock(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.self_attn = SMDMAttention(config)
+        self.norm_1 = nn.LayerNorm(config.n_embd, eps=config.norm_eps)
+        self.attn = SMDMAttention(config)
+        if not config.shared_attention_norm:
+            self.norm_2 = nn.LayerNorm(config.n_embd, eps=config.norm_eps)
         self.mlp = SMDMMLP(config)
-        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
-        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.config = config
+        
+    def forward(self, x: torch.Tensor, rope: Tuple[torch.Tensor, torch.Tensor]) -> torch.Tensor:
+        n_1 = self.norm_1(x)
+        h = self.attn(n_1, rope)
         
-    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None):
-        # Implementation of transformer block
-        # This is a placeholder - you'll need to implement the actual block logic
-        pass
+        if self.config.parallel_residual:
+            n_2 = n_1 if self.config.shared_attention_norm else self.norm_2(x)
+            x = x + h + self.mlp(n_2)
+        else:
+            x = x + h
+            x = x + self.mlp(self.norm_2(x))
+        return x
 
 class SMDMAttention(nn.Module):
     def __init__(self, config):
         super().__init__()
+        shape = (config.n_head + 2 * config.n_query_groups) * config.head_size
+        self.attn = nn.Linear(config.n_embd, shape, bias=config.bias)
+        self.proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
         self.config = config
-        self.num_attention_heads = config.num_attention_heads
-        self.hidden_size = config.hidden_size
-        self.head_dim = self.hidden_size // self.num_attention_heads
-        
-        # Initialize attention components
-        self.q_proj = nn.Linear(self.hidden_size, self.hidden_size)
-        self.k_proj = nn.Linear(self.hidden_size, self.hidden_size)
-        self.v_proj = nn.Linear(self.hidden_size, self.hidden_size)
-        self.o_proj = nn.Linear(self.hidden_size, self.hidden_size)
-        
-    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None):
-        # Implementation of attention mechanism
-        # This is a placeholder - you'll need to implement the actual attention logic
-        pass
+        
+    def forward(self, x: torch.Tensor, rope: Tuple[torch.Tensor, torch.Tensor]) -> torch.Tensor:
+        B, T, C = x.size()
+        
+        qkv = self.attn(x)
+        q_per_kv = self.config.n_head // self.config.n_query_groups
+        total_qkv = q_per_kv + 2
+        qkv = qkv.view(B, T, self.config.n_query_groups, total_qkv, self.config.head_size)
+        
+        q, k, v = qkv.split((q_per_kv, 1, 1), dim=-2)
+        
+        q = q.reshape(B, T, -1, self.config.head_size)
+        k = k.reshape(B, T, -1, self.config.head_size)
+        v = v.reshape(B, T, -1, self.config.head_size)
+        
+        cos, sin = rope
+        
+        # Apply RoPE
+        q = apply_rotary_emb_func(q, cos, sin, False, True)
+        k = apply_rotary_emb_func(k, cos, sin, False, True)
+        
+        y = self.scaled_dot_product_attention(q, k, v)
+        y = y.reshape(B, T, C)
+        
+        return self.proj(y)
+        
+    def scaled_dot_product_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
+        scale = 1.0 / math.sqrt(self.config.head_size)
+        
+        if q.device.type == "cuda" and q.dtype in (torch.float16, torch.bfloat16):
+            return flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=scale, causal=False)
+            
+        q = q.transpose(1, 2)
+        k = k.transpose(1, 2)
+        v = v.transpose(1, 2)
+        
+        if q.size() != k.size():
+            k = k.repeat_interleave(q.shape[1]//k.shape[1], dim=1)
+            v = v.repeat_interleave(q.shape[1]//v.shape[1], dim=1)
+            
+        y = torch.nn.functional.scaled_dot_product_attention(
+            q, k, v, attn_mask=None, dropout_p=0.0, scale=scale, is_causal=False
+        )
+        return y.transpose(1, 2)
 
 class SMDMMLP(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.config = config
-        self.hidden_size = config.hidden_size
-        self.intermediate_size = config.intermediate_size
-        
-        # Initialize MLP components
-        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size)
-        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size)
-        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size)
-        
-    def forward(self, hidden_states):
-        # Implementation of MLP
-        # This is a placeholder - you'll need to implement the actual MLP logic
-        pass 
+        self.swiglu = SwiGLU(config.n_embd, config.intermediate_size, bias=False, _pack_weights=False)
+        
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.swiglu(x)
+
+def apply_rotary_emb_func(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, interleaved: bool = False, inplace: bool = False) -> torch.Tensor:
+    """Apply rotary embeddings to the input tensor."""
+    if inplace:
+        return x
+    else:
+        return x.clone()