rasbt commited on
Commit
44656d5
·
verified ·
1 Parent(s): e436a25

Upload folder using huggingface_hub

Browse files
.DS_Store ADDED
Binary file (6.15 kB). View file
 
.ipynb_checkpoints/model-checkpoint.py ADDED
@@ -0,0 +1,334 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Sebastian Raschka under Apache License 2.0 (see LICENSE.txt).
2
+ # Source for "Build a Large Language Model From Scratch"
3
+ # https://github.com/rasbt/LLMs-from-scratch/blob/main/ch05/07_gpt_to_llama/standalone-llama32.ipynb
4
+
5
+
6
+ import torch
7
+ import torch.nn as nn
8
+
9
+
10
+ LLAMA32_CONFIG_1B = {
11
+ "vocab_size": 128_256, # Vocabulary size
12
+ "context_length": 8192, # Maximum context length to use (reduced to save memory)
13
+ "orig_context_length": 131_072, # Context length that was used to train the model
14
+ "emb_dim": 2048, # Embedding dimension
15
+ "n_heads": 32, # Number of attention heads
16
+ "n_layers": 16, # Number of layers
17
+ "hidden_dim": 8192, # Size of the intermediate dimension in FeedForward
18
+ "n_kv_groups": 8, # Key-Value groups for grouped-query attention
19
+ "rope_base": 500_000.0, # The base in RoPE's "theta"
20
+ "dtype": torch.bfloat16, # Lower-precision dtype to reduce memory usage
21
+ "rope_freq": { # RoPE frequency scaling
22
+ "factor": 32.0,
23
+ "low_freq_factor": 1.0,
24
+ "high_freq_factor": 4.0,
25
+ "original_context_length": 8192,
26
+ }
27
+ }
28
+
29
+ LLAMA32_CONFIG_3B = {
30
+ "vocab_size": 128_256, # Vocabulary size
31
+ "context_length": 8192, # Maximum context length to use (reduced to save memory)
32
+ "orig_context_length": 131_072, # Context length that was used to train the model
33
+ "emb_dim": 3072, # Embedding dimension
34
+ "n_heads": 24, # Number of attention heads
35
+ "n_layers": 28, # Number of layers
36
+ "hidden_dim": 8192, # Size of the intermediate dimension in FeedForward
37
+ "n_kv_groups": 8, # Key-Value groups for grouped-query attention
38
+ "rope_base": 500_000.0, # The base in RoPE's "theta"
39
+ "dtype": torch.bfloat16, # Lower-precision dtype to reduce memory usage
40
+ "rope_freq": { # RoPE frequency scaling
41
+ "factor": 32.0,
42
+ "low_freq_factor": 1.0,
43
+ "high_freq_factor": 4.0,
44
+ "original_context_length": 8192,
45
+ }
46
+ }
47
+
48
+
49
+ class Llama3Model(nn.Module):
50
+ def __init__(self, cfg):
51
+ super().__init__()
52
+
53
+ # Main model parameters
54
+ self.tok_emb = nn.Embedding(cfg["vocab_size"], cfg["emb_dim"], dtype=cfg["dtype"])
55
+
56
+ self.trf_blocks = nn.ModuleList( # ModuleList since Sequential can only accept one input, and we need `x, mask, cos, sin`
57
+ [TransformerBlock(cfg) for _ in range(cfg["n_layers"])]
58
+ )
59
+
60
+ self.final_norm = nn.RMSNorm(cfg["emb_dim"], eps=1e-5, dtype=cfg["dtype"])
61
+ self.out_head = nn.Linear(cfg["emb_dim"], cfg["vocab_size"], bias=False, dtype=cfg["dtype"])
62
+
63
+ # Reusuable utilities
64
+ self.register_buffer(
65
+ "mask", torch.triu(torch.ones(cfg["context_length"], cfg["context_length"]), diagonal=1).bool(),
66
+ persistent=False
67
+ )
68
+
69
+ if cfg["orig_context_length"] != cfg["context_length"]:
70
+ cfg["rope_base"] = rescale_theta(
71
+ cfg["rope_base"],
72
+ cfg["orig_context_length"],
73
+ cfg["context_length"]
74
+ )
75
+ cos, sin = compute_rope_params(
76
+ head_dim=cfg["emb_dim"] // cfg["n_heads"],
77
+ theta_base=cfg["rope_base"],
78
+ context_length=cfg["context_length"],
79
+ freq_config=cfg["rope_freq"]
80
+ )
81
+ self.register_buffer("cos", cos, persistent=False)
82
+ self.register_buffer("sin", sin, persistent=False)
83
+ self.cfg = cfg
84
+
85
+ def forward(self, in_idx):
86
+ # Forward pass
87
+ tok_embeds = self.tok_emb(in_idx)
88
+ x = tok_embeds
89
+
90
+ for block in self.trf_blocks:
91
+ x = block(x, self.mask, self.cos, self.sin)
92
+ x = self.final_norm(x)
93
+ logits = self.out_head(x.to(self.cfg["dtype"]))
94
+ return logits
95
+
96
+
97
+ class TransformerBlock(nn.Module):
98
+ def __init__(self, cfg):
99
+ super().__init__()
100
+ self.att = GroupedQueryAttention(
101
+ d_in=cfg["emb_dim"],
102
+ d_out=cfg["emb_dim"],
103
+ num_heads=cfg["n_heads"],
104
+ num_kv_groups=cfg["n_kv_groups"],
105
+ dtype=cfg["dtype"]
106
+ )
107
+ self.ff = FeedForward(cfg)
108
+ self.norm1 = nn.RMSNorm(cfg["emb_dim"], eps=1e-5, dtype=cfg["dtype"])
109
+ self.norm2 = nn.RMSNorm(cfg["emb_dim"], eps=1e-5, dtype=cfg["dtype"])
110
+
111
+ def forward(self, x, mask, cos, sin):
112
+ # Shortcut connection for attention block
113
+ shortcut = x
114
+ x = self.norm1(x)
115
+ x = self.att(x, mask, cos, sin) # Shape [batch_size, num_tokens, emb_size]
116
+ x = x + shortcut # Add the original input back
117
+
118
+ # Shortcut connection for feed-forward block
119
+ shortcut = x
120
+ x = self.norm2(x)
121
+ x = self.ff(x)
122
+ x = x + shortcut # Add the original input back
123
+
124
+ return x
125
+
126
+
127
+ class FeedForward(nn.Module):
128
+ def __init__(self, cfg):
129
+ super().__init__()
130
+ self.fc1 = nn.Linear(cfg["emb_dim"], cfg["hidden_dim"], dtype=cfg["dtype"], bias=False)
131
+ self.fc2 = nn.Linear(cfg["emb_dim"], cfg["hidden_dim"], dtype=cfg["dtype"], bias=False)
132
+ self.fc3 = nn.Linear(cfg["hidden_dim"], cfg["emb_dim"], dtype=cfg["dtype"], bias=False)
133
+
134
+ def forward(self, x):
135
+ x_fc1 = self.fc1(x)
136
+ x_fc2 = self.fc2(x)
137
+ x = nn.functional.silu(x_fc1) * x_fc2
138
+ return self.fc3(x)
139
+
140
+
141
+ class GroupedQueryAttention(nn.Module):
142
+ def __init__(
143
+ self, d_in, d_out, num_heads,
144
+ num_kv_groups,
145
+ dtype=None
146
+ ):
147
+ super().__init__()
148
+ assert d_out % num_heads == 0, "d_out must be divisible by num_heads"
149
+ assert num_heads % num_kv_groups == 0, "num_heads must be divisible by num_kv_groups"
150
+
151
+ self.d_out = d_out
152
+ self.num_heads = num_heads
153
+ self.head_dim = d_out // num_heads
154
+
155
+ self.W_key = nn.Linear(d_in, num_kv_groups * self.head_dim, bias=False, dtype=dtype)
156
+ self.W_value = nn.Linear(d_in, num_kv_groups * self.head_dim, bias=False, dtype=dtype)
157
+ self.num_kv_groups = num_kv_groups
158
+ self.group_size = num_heads // num_kv_groups
159
+
160
+ self.W_query = nn.Linear(d_in, d_out, bias=False, dtype=dtype)
161
+ self.out_proj = nn.Linear(d_out, d_out, bias=False, dtype=dtype)
162
+
163
+ def forward(self, x, mask, cos, sin):
164
+ b, num_tokens, d_in = x.shape
165
+
166
+ queries = self.W_query(x) # Shape: (b, num_tokens, d_out)
167
+ keys = self.W_key(x) # Shape: (b, num_tokens, num_kv_groups * head_dim)
168
+ values = self.W_value(x) # Shape: (b, num_tokens, num_kv_groups * head_dim)
169
+
170
+ # Reshape queries, keys, and values
171
+ queries = queries.view(b, num_tokens, self.num_heads, self.head_dim)
172
+ keys = keys.view(b, num_tokens, self.num_kv_groups, self.head_dim)
173
+ values = values.view(b, num_tokens, self.num_kv_groups, self.head_dim)
174
+
175
+ # Transpose keys, values, and queries
176
+ keys = keys.transpose(1, 2) # Shape: (b, num_heads, num_tokens, head_dim)
177
+ values = values.transpose(1, 2) # Shape: (b, num_heads, num_tokens, head_dim)
178
+ queries = queries.transpose(1, 2) # Shape: (b, num_query_groups, num_tokens, head_dim)
179
+
180
+ # Apply RoPE
181
+ keys = apply_rope(keys, cos, sin)
182
+ queries = apply_rope(queries, cos, sin)
183
+
184
+ # Expand keys and values to match the number of heads
185
+ # Shape: (b, num_heads, num_tokens, head_dim)
186
+ keys = keys.repeat_interleave(self.group_size, dim=1) # Shape: (b, num_heads, num_tokens, head_dim)
187
+ values = values.repeat_interleave(self.group_size, dim=1) # Shape: (b, num_heads, num_tokens, head_dim)
188
+ # For example, before repeat_interleave along dim=1 (query groups):
189
+ # [K1, K2]
190
+ # After repeat_interleave (each query group is repeated group_size times):
191
+ # [K1, K1, K2, K2]
192
+ # If we used regular repeat instead of repeat_interleave, we'd get:
193
+ # [K1, K2, K1, K2]
194
+
195
+ # Compute scaled dot-product attention (aka self-attention) with a causal mask
196
+ # Shape: (b, num_heads, num_tokens, num_tokens)
197
+ attn_scores = queries @ keys.transpose(2, 3) # Dot product for each head
198
+
199
+ # Use the mask to fill attention scores
200
+ attn_scores = attn_scores.masked_fill(mask[:num_tokens, :num_tokens], -torch.inf)
201
+
202
+ attn_weights = torch.softmax(attn_scores / keys.shape[-1]**0.5, dim=-1)
203
+ assert keys.shape[-1] == self.head_dim
204
+
205
+ # Shape: (b, num_tokens, num_heads, head_dim)
206
+ context_vec = (attn_weights @ values).transpose(1, 2)
207
+
208
+ # Combine heads, where self.d_out = self.num_heads * self.head_dim
209
+ context_vec = context_vec.reshape(b, num_tokens, self.d_out)
210
+ context_vec = self.out_proj(context_vec) # optional projection
211
+
212
+ return context_vec
213
+
214
+
215
+ def compute_rope_params(head_dim, theta_base=10_000, context_length=4096, freq_config=None, dtype=torch.float32):
216
+ assert head_dim % 2 == 0, "Embedding dimension must be even"
217
+
218
+ # Compute the inverse frequencies
219
+ inv_freq = 1.0 / (theta_base ** (torch.arange(0, head_dim, 2, dtype=dtype)[: (head_dim // 2)].float() / head_dim))
220
+
221
+ # Frequency adjustments
222
+ if freq_config is not None:
223
+ low_freq_wavelen = freq_config["original_context_length"] / freq_config["low_freq_factor"]
224
+ high_freq_wavelen = freq_config["original_context_length"] / freq_config["high_freq_factor"]
225
+
226
+ wavelen = 2 * torch.pi / inv_freq
227
+
228
+ inv_freq_llama = torch.where(
229
+ wavelen > low_freq_wavelen, inv_freq / freq_config["factor"], inv_freq
230
+ )
231
+
232
+ smooth_factor = (freq_config["original_context_length"] / wavelen - freq_config["low_freq_factor"]) / (
233
+ freq_config["high_freq_factor"] - freq_config["low_freq_factor"]
234
+ )
235
+
236
+ smoothed_inv_freq = (
237
+ (1 - smooth_factor) * (inv_freq / freq_config["factor"]) + smooth_factor * inv_freq
238
+ )
239
+
240
+ is_medium_freq = (wavelen <= low_freq_wavelen) & (wavelen >= high_freq_wavelen)
241
+ inv_freq_llama = torch.where(is_medium_freq, smoothed_inv_freq, inv_freq_llama)
242
+ inv_freq = inv_freq_llama
243
+
244
+ # Generate position indices
245
+ positions = torch.arange(context_length, dtype=dtype)
246
+
247
+ # Compute the angles
248
+ angles = positions[:, None] * inv_freq[None, :] # Shape: (context_length, head_dim // 2)
249
+
250
+ # Expand angles to match the head_dim
251
+ angles = torch.cat([angles, angles], dim=1) # Shape: (context_length, head_dim)
252
+
253
+ # Precompute sine and cosine
254
+ cos = torch.cos(angles)
255
+ sin = torch.sin(angles)
256
+
257
+ return cos, sin
258
+
259
+
260
+ def apply_rope(x, cos, sin):
261
+ # x: (batch_size, num_heads, seq_len, head_dim)
262
+ batch_size, num_heads, seq_len, head_dim = x.shape
263
+ assert head_dim % 2 == 0, "Head dimension must be even"
264
+
265
+ # Split x into first half and second half
266
+ x1 = x[..., : head_dim // 2] # First half
267
+ x2 = x[..., head_dim // 2:] # Second half
268
+
269
+ # Adjust sin and cos shapes
270
+ cos = cos[:seq_len, :].unsqueeze(0).unsqueeze(0) # Shape: (1, 1, seq_len, head_dim)
271
+ sin = sin[:seq_len, :].unsqueeze(0).unsqueeze(0)
272
+
273
+ # Apply the rotary transformation
274
+ rotated = torch.cat((-x2, x1), dim=-1)
275
+ x_rotated = (x * cos) + (rotated * sin)
276
+
277
+ # It's ok to use lower-precision after applying cos and sin rotation
278
+ return x_rotated.to(dtype=x.dtype)
279
+
280
+
281
+ def rescale_theta(theta_old, context_length_old, context_length_new):
282
+ scaling_factor = context_length_new / context_length_old
283
+ theta_new = theta_old * scaling_factor
284
+ return theta_new
285
+
286
+
287
+ def text_to_token_ids(text, tokenizer):
288
+ encoded = tokenizer.encode(text)
289
+ encoded_tensor = torch.tensor(encoded).unsqueeze(0) # add batch dimension
290
+ return encoded_tensor
291
+
292
+
293
+ def token_ids_to_text(token_ids, tokenizer):
294
+ flat = token_ids.squeeze(0) # remove batch dimension
295
+ return tokenizer.decode(flat.tolist())
296
+
297
+
298
+ def generate(model, idx, max_new_tokens, context_size, temperature=0.0, top_k=None, eos_id=None):
299
+
300
+ # For-loop is the same as before: Get logits, and only focus on last time step
301
+ for _ in range(max_new_tokens):
302
+ idx_cond = idx[:, -context_size:]
303
+ with torch.no_grad():
304
+ logits = model(idx_cond)
305
+ logits = logits[:, -1, :]
306
+
307
+ # Filter logits with top_k sampling
308
+ if top_k is not None:
309
+ # Keep only top_k values
310
+ top_logits, _ = torch.topk(logits, top_k)
311
+ min_val = top_logits[:, -1]
312
+ logits = torch.where(logits < min_val, torch.tensor(float('-inf')).to(logits.device), logits)
313
+
314
+ # Apply temperature scaling
315
+ if temperature > 0.0:
316
+ logits = logits / temperature
317
+
318
+ # Apply softmax to get probabilities
319
+ probs = torch.softmax(logits, dim=-1) # (batch_size, context_len)
320
+
321
+ # Sample from the distribution
322
+ idx_next = torch.multinomial(probs, num_samples=1) # (batch_size, 1)
323
+
324
+ # Otherwise same as before: get idx of the vocab entry with the highest logits value
325
+ else:
326
+ idx_next = torch.argmax(logits, dim=-1, keepdim=True) # (batch_size, 1)
327
+
328
+ if idx_next == eos_id: # Stop generating early if end-of-sequence token is encountered and eos_id is specified
329
+ break
330
+
331
+ # Same as before: append sampled index to the running sequence
332
+ idx = torch.cat((idx, idx_next), dim=1) # (batch_size, num_tokens+1)
333
+
334
+ return idx
llama3.2-1B-base.pth CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:af0de8aca49b5e6138c53a58bc72eb3a7dc6c37d9bceb9d0778b341cd12f1082
3
- size 3064129762
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a91cbbc417b663619e49cc1f12faa0329ae9d7cec233c146f648ee230d5d78ca
3
+ size 2997020635
llama3.2-1B-instruct.pth CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4a1a0ff1c6e4d1aff871e50ab75684c4704e0cacabe6adf4cdbb5da3d4fcc077
3
- size 3064130434
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6718811b79ea055d51cf6a3c0f18faaaca921b9cd33ecd917a42cc2572756142
3
+ size 2997021239
llama3.2-3B-base.pth CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a41b5e512faeec28da58f144b4bb1739756f16294f105dfc7ee54849645d4253
3
- size 7280710838
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4510deedd5e99d7bca73397e9041333361d7de68c76da94b2508e0a4894f3395
3
+ size 7213601647
llama3.2-3B-instruct.pth CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:459af5c3c14d19f72a48ddeadd01e53c1a521b7b3df517a6c4f9187d2757df0a
3
- size 7280711878
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bf907cd45420d57832d26972b0d824713f5b3bd9c0c961a4e73368709ca5e609
3
+ size 7213602683
model.py CHANGED
@@ -61,7 +61,10 @@ class Llama3Model(nn.Module):
61
  self.out_head = nn.Linear(cfg["emb_dim"], cfg["vocab_size"], bias=False, dtype=cfg["dtype"])
62
 
63
  # Reusuable utilities
64
- self.register_buffer("mask", torch.triu(torch.ones(cfg["context_length"], cfg["context_length"]), diagonal=1).bool())
 
 
 
65
 
66
  if cfg["orig_context_length"] != cfg["context_length"]:
67
  cfg["rope_base"] = rescale_theta(
 
61
  self.out_head = nn.Linear(cfg["emb_dim"], cfg["vocab_size"], bias=False, dtype=cfg["dtype"])
62
 
63
  # Reusuable utilities
64
+ self.register_buffer(
65
+ "mask", torch.triu(torch.ones(cfg["context_length"], cfg["context_length"]), diagonal=1).bool(),
66
+ persistent=False
67
+ )
68
 
69
  if cfg["orig_context_length"] != cfg["context_length"]:
70
  cfg["rope_base"] = rescale_theta(