experimental "lin" blocks insted of attention sparely

.gitignore

@@ -0,0 +1,2 @@
+*.safetensors
+*.csv

data.py

@@ -18,27 +18,26 @@ def dataWorker(q, dataset, encoding, batch_size, block_size,chat):
     batch_x, batch_y = [], []
     while True:
         for text in dataset:
-            tokens = None
+            tokens = []
             if(chat):
-                txt=""
                 for msg in text['messages']:
                     role = msg['role']
                     content = msg['content']
-                    txt = txt + f"<|{role}|>{content}<|end|>"
+                    txt = f"<|{role}|>{content}<|end|> "
-                tokens = [encoding.bos_token_id]+encoding.encode(txt)
+                    tokens += encoding.encode(txt) + [encoding.eos_token_id]
             else:
-                tokens = [encoding.bos_token_id]+encoding.encode(text["text"])
+                tokens = encoding.encode(text["text"])
             for i in range(0, len(tokens)-block_size+1,block_size):
                 x = tokens[i:i+block_size]
                 y = tokens[i+1:i+block_size+1]
 
                 if len(x) < block_size:
-                    pad = len(x)-(block_size-1)
+                    pad = len(x)-(block_size)
-                    x = x + [encoding.eos_token_id] + [encoding.pad_token_id] * pad
+                    x = x + [encoding.eos_token_id] * pad
 
                 if len(y) < block_size:
-                    pad = len(y)-(block_size-1)
+                    pad = len(y)-(block_size)
-                    y = y + [encoding.eos_token_id] + [encoding.pad_token_id] * pad
+                    y = y + [encoding.eos_token_id] * pad
 
                 batch_x.append(x)
                 batch_y.append(y)

model.py

@@ -1,12 +1,13 @@
 from tinygrad import Tensor,nn,TinyJit
 
 class MultiHeadAttention:
-    def __init__(self,embed_size,n_heads):
+    def __init__(self,embed_size,n_heads,lin):
         assert embed_size % n_heads == 0
         self.head_size = embed_size//n_heads
         self.n_heads = n_heads
         self.qkv = nn.Linear(embed_size, embed_size*3,bias=False)
         self.projection = nn.Linear(embed_size, embed_size,bias=False)
+        self.lin = lin
     def __call__(self,x):
         B,T,C=x.shape
 
@@ -15,10 +16,16 @@ class MultiHeadAttention:
         k = k.view(B, T, self.n_heads, self.head_size).transpose(1, 2)
         v = v.view(B, T, self.n_heads, self.head_size).transpose(1, 2)
         #B H T S
-        #TODO attention free transformer
 
-        out = q.scaled_dot_product_attention(k,v,is_causal=True,dropout_p=0.01)
+        out = None
+        if self.lin:
+            q = q.sigmoid()
+            k = k.sigmoid()
+            out = ((q*k).exp()/(q*k)) * v
+        else:
+            out = q.scaled_dot_product_attention(k,v,is_causal=True)
         out = out.transpose(1,2).view(B,T,C)
+
         return self.projection(out)
     def cast(self,dtype):
         self.qkv.weight = self.qkv.weight.cast(dtype)
@@ -43,8 +50,8 @@ class FeedForwardNetwork:
         return self
 
 class Block:
-    def __init__(self,embed_size,n_heads):
+    def __init__(self,embed_size,n_heads,lin):
-        self.mha = MultiHeadAttention(embed_size,n_heads)
+        self.mha = MultiHeadAttention(embed_size,n_heads,lin)
         self.ffn = FeedForwardNetwork(embed_size)
         self.mhaNorm = nn.RMSNorm(embed_size)
         self.ffnNorm = nn.RMSNorm(embed_size)
@@ -61,9 +68,9 @@ class Transformer():
     def __init__(self,vocab_size,embed_size,n_heads,n_blocks,block_size):
         self.tok_embed = nn.Embedding(vocab_size,embed_size)
         self.pos_embed = nn.Embedding(block_size,embed_size)
-        self.pos_idx = Tensor.arange(block_size, requires_grad=False)
+        self.pos_idx = Tensor.arange(block_size, requires_grad=False).sin()
 
-        self.blocks = [Block(embed_size,n_heads) for _ in range(n_blocks)]
+        self.blocks = [Block(embed_size,n_heads,i%4==0) for i in range(n_blocks)]
         self.norm = nn.RMSNorm(embed_size)
         self.output = nn.Linear(embed_size,vocab_size,bias=False)
     def __call__(self,x):

train.py

@@ -12,20 +12,20 @@ import sys
 
 hypr = {
     "embed_size":  768,
-    "n_heads":     12,
+    "n_heads":     8,
     "n_blocks":    12,
     "block_size":  512,
     "batch_size":  8,
     "starting_lr": 6e-4,
     "minimum_lr":  6e-5,
-    "warmup":      1_000,
+    "warmup":      5_000,
-    "steps":       20_000,
+    "steps":       535_000,
-    "encoding":    "gpt2",
+    "encoding":    "TinyLlama/TinyLlama_v1.1",
     "dataset":     "HuggingFaceTB/smollm-corpus",
     "subset":      "cosmopedia-v2",
     "chat_dataset": "HuggingFaceTB/smoltalk",
     "chat_subset":  "all",
-    "half":         False,
+    "half":         True,
 }
 
 print(Device.DEFAULT)