loocorez commited on
Commit
c29ec17
·
verified ·
1 Parent(s): 9799787

Upload folder using huggingface_hub

Browse files
Files changed (4) hide show
  1. README.md +6 -7
  2. app.py +78 -0
  3. gpt_infer.py +190 -0
  4. requirements.txt +5 -0
README.md CHANGED
@@ -1,12 +1,11 @@
1
  ---
2
- title: Nanochat Zerogpu Demo
3
- emoji: 🐢
4
- colorFrom: blue
5
- colorTo: gray
6
  sdk: gradio
7
- sdk_version: 5.49.1
8
  app_file: app.py
9
- pinned: false
10
  ---
11
 
12
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
1
  ---
2
+ title: nanochat ZeroGPU Demo
3
+ emoji: 🤖
4
+ colorFrom: indigo
5
+ colorTo: blue
6
  sdk: gradio
7
+ sdk_version: 4.41.0
8
  app_file: app.py
 
9
  ---
10
 
11
+ ZeroGPU demo Space serving nanochat SFT/MID/BASE models (CPU inference).
app.py ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import os
3
+ import json
4
+ import gradio as gr
5
+ import torch
6
+ from huggingface_hub import hf_hub_download
7
+ from tokenizers import Tokenizer as HFTokenizer
8
+ from gpt_infer import GPT, GPTConfig
9
+
10
+ DEFAULT_MODEL = os.environ.get('NANOCHAT_DEFAULT_MODEL', 'loocorez/nanochat-sft-d20-step650')
11
+
12
+ def load_model(repo_id: str):
13
+ cfg_path = hf_hub_download(repo_id, 'config.json')
14
+ with open(cfg_path, 'r') as f:
15
+ cfg = json.load(f)
16
+ model_config = GPTConfig(
17
+ sequence_len=cfg.get('n_ctx', 2048),
18
+ vocab_size=cfg['vocab_size'],
19
+ n_layer=cfg['n_layer'],
20
+ n_head=cfg['n_head'],
21
+ n_kv_head=cfg.get('n_kv_head', cfg['n_head']),
22
+ n_embd=cfg['n_embd'],
23
+ )
24
+ model = GPT(model_config)
25
+ model.eval()
26
+ try:
27
+ from safetensors.torch import load_file
28
+ weights_path = hf_hub_download(repo_id, 'model.safetensors')
29
+ sd = load_file(weights_path)
30
+ except Exception:
31
+ weights_path = hf_hub_download(repo_id, 'pytorch_model.bin')
32
+ sd = torch.load(weights_path, map_location='cpu')
33
+ model.load_state_dict(sd, strict=True, assign=True)
34
+ tok_path = hf_hub_download(repo_id, 'tokenizer.json')
35
+ tok = HFTokenizer.from_file(tok_path)
36
+ return model, tok
37
+
38
+ model_cache = {}
39
+
40
+ def get_model(repo_id: str):
41
+ if repo_id not in model_cache:
42
+ model_cache[repo_id] = load_model(repo_id)
43
+ return model_cache[repo_id]
44
+
45
+ @torch.inference_mode()
46
+ def generate(repo_id: str, prompt: str, max_tokens: int, temperature: float, top_k: int|None):
47
+ model, tok = get_model(repo_id)
48
+ bos_id = tok.token_to_id('<|bos|>')
49
+ ids = tok.encode(prompt).ids
50
+ if bos_id is not None:
51
+ ids = [bos_id] + ids
52
+ out_tokens = []
53
+ for token in model.generate(ids, max_tokens=max_tokens, temperature=temperature, top_k=top_k or None):
54
+ out_tokens.append(token)
55
+ text = tok.decode(out_tokens, skip_special_tokens=False)
56
+ return text
57
+
58
+ with gr.Blocks() as demo:
59
+ gr.Markdown('# nanochat (ZeroGPU)')
60
+ gr.Markdown('Select a model and generate text.')
61
+ repo = gr.Dropdown(choices=[
62
+ 'loocorez/nanochat-sft-d20-step650',
63
+ 'loocorez/nanochat-mid-d20-step765',
64
+ 'loocorez/nanochat-base-d20-step21400',
65
+ ], value=DEFAULT_MODEL, label='Model Repo')
66
+ prompt = gr.Textbox(label='Prompt', lines=6)
67
+ with gr.Row():
68
+ max_tokens = gr.Slider(1, 256, value=128, step=1, label='Max tokens')
69
+ temperature = gr.Slider(0.0, 1.5, value=0.8, step=0.05, label='Temperature')
70
+ top_k = gr.Slider(0, 100, value=40, step=1, label='Top-k (0=disabled)')
71
+ btn = gr.Button('Generate')
72
+ output = gr.Textbox(label='Output', lines=10)
73
+ btn.click(fn=lambda r,p,m,t,k: generate(r,p,int(m),float(t),int(k) if int(k)>0 else None),
74
+ inputs=[repo, prompt, max_tokens, temperature, top_k],
75
+ outputs=output)
76
+
77
+ if __name__ == '__main__':
78
+ demo.launch()
gpt_infer.py ADDED
@@ -0,0 +1,190 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import math
3
+ import torch
4
+ import torch.nn as nn
5
+ import torch.nn.functional as F
6
+ from dataclasses import dataclass
7
+
8
+
9
+ def norm(x: torch.Tensor) -> torch.Tensor:
10
+ return F.rms_norm(x, (x.size(-1),))
11
+
12
+
13
+ def apply_rotary_emb(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
14
+ assert x.ndim == 4
15
+ d = x.shape[3] // 2
16
+ x1, x2 = x[..., :d], x[..., d:]
17
+ y1 = x1 * cos + x2 * sin
18
+ y2 = x1 * (-sin) + x2 * cos
19
+ out = torch.cat([y1, y2], 3)
20
+ out = out.to(x.dtype)
21
+ return out
22
+
23
+
24
+ def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
25
+ if n_rep == 1:
26
+ return x
27
+ bs, n_kv_heads, slen, head_dim = x.shape
28
+ return (
29
+ x[:, :, None, :, :]
30
+ .expand(bs, n_kv_heads, n_rep, slen, head_dim)
31
+ .reshape(bs, n_kv_heads * n_rep, slen, head_dim)
32
+ )
33
+
34
+
35
+ @dataclass
36
+ class GPTConfig:
37
+ sequence_len: int = 1024
38
+ vocab_size: int = 50304
39
+ n_layer: int = 12
40
+ n_head: int = 6
41
+ n_kv_head: int = 6
42
+ n_embd: int = 768
43
+
44
+
45
+ class CausalSelfAttention(nn.Module):
46
+ def __init__(self, config: GPTConfig, layer_idx: int):
47
+ super().__init__()
48
+ self.layer_idx = layer_idx
49
+ self.n_head = config.n_head
50
+ self.n_kv_head = config.n_kv_head
51
+ self.n_embd = config.n_embd
52
+ self.head_dim = self.n_embd // self.n_head
53
+ assert self.n_embd % self.n_head == 0
54
+ assert self.n_kv_head <= self.n_head and self.n_head % self.n_kv_head == 0
55
+ self.c_q = nn.Linear(self.n_embd, self.n_head * self.head_dim, bias=False)
56
+ self.c_k = nn.Linear(self.n_embd, self.n_kv_head * self.head_dim, bias=False)
57
+ self.c_v = nn.Linear(self.n_embd, self.n_kv_head * self.head_dim, bias=False)
58
+ self.c_proj = nn.Linear(self.n_embd, self.n_embd, bias=False)
59
+
60
+ def forward(self, x: torch.Tensor, cos_sin, kv_cache=None):
61
+ B, T, C = x.size()
62
+ q = self.c_q(x).view(B, T, self.n_head, self.head_dim)
63
+ k = self.c_k(x).view(B, T, self.n_kv_head, self.head_dim)
64
+ v = self.c_v(x).view(B, T, self.n_kv_head, self.head_dim)
65
+ cos, sin = cos_sin
66
+ q, k = apply_rotary_emb(q, cos, sin), apply_rotary_emb(k, cos, sin)
67
+ q, k = norm(q), norm(k)
68
+ q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
69
+ if kv_cache is not None:
70
+ k, v = kv_cache.insert_kv(self.layer_idx, k, v)
71
+ Tq = q.size(2)
72
+ Tk = k.size(2)
73
+ nrep = self.n_head // self.n_kv_head
74
+ k, v = repeat_kv(k, nrep), repeat_kv(v, nrep)
75
+ if kv_cache is None or Tq == Tk:
76
+ y = F.scaled_dot_product_attention(q, k, v, is_causal=True)
77
+ elif Tq == 1:
78
+ y = F.scaled_dot_product_attention(q, k, v, is_causal=False)
79
+ else:
80
+ attn_mask = torch.zeros((Tq, Tk), dtype=torch.bool, device=q.device)
81
+ prefix_len = Tk - Tq
82
+ if prefix_len > 0:
83
+ attn_mask[:, :prefix_len] = True
84
+ attn_mask[:, prefix_len:] = torch.tril(torch.ones((Tq, Tq), dtype=torch.bool, device=q.device))
85
+ y = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
86
+ y = y.transpose(1, 2).contiguous().view(B, T, -1)
87
+ y = self.c_proj(y)
88
+ return y
89
+
90
+
91
+ class MLP(nn.Module):
92
+ def __init__(self, config: GPTConfig):
93
+ super().__init__()
94
+ self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=False)
95
+ self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=False)
96
+
97
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
98
+ x = self.c_fc(x)
99
+ x = F.relu(x).square()
100
+ x = self.c_proj(x)
101
+ return x
102
+
103
+
104
+ class Block(nn.Module):
105
+ def __init__(self, config: GPTConfig, layer_idx: int):
106
+ super().__init__()
107
+ self.attn = CausalSelfAttention(config, layer_idx)
108
+ self.mlp = MLP(config)
109
+
110
+ def forward(self, x: torch.Tensor, cos_sin, kv_cache=None) -> torch.Tensor:
111
+ x = x + self.attn(norm(x), cos_sin, kv_cache)
112
+ x = x + self.mlp(norm(x))
113
+ return x
114
+
115
+
116
+ class GPT(nn.Module):
117
+ def __init__(self, config: GPTConfig):
118
+ super().__init__()
119
+ self.config = config
120
+ self.transformer = nn.ModuleDict({
121
+ 'wte': nn.Embedding(config.vocab_size, config.n_embd),
122
+ 'h': nn.ModuleList([Block(config, layer_idx) for layer_idx in range(config.n_layer)]),
123
+ })
124
+ self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
125
+ self.rotary_seq_len = config.sequence_len * 10
126
+ head_dim = config.n_embd // config.n_head
127
+ cos, sin = self._precompute_rotary_embeddings(self.rotary_seq_len, head_dim)
128
+ self.register_buffer('cos', cos, persistent=False)
129
+ self.register_buffer('sin', sin, persistent=False)
130
+ self.transformer.wte.to(dtype=torch.bfloat16)
131
+
132
+ def _precompute_rotary_embeddings(self, seq_len: int, head_dim: int, base: int = 10000, device=None):
133
+ if device is None:
134
+ device = self.transformer.wte.weight.device
135
+ channel_range = torch.arange(0, head_dim, 2, dtype=torch.float32, device=device)
136
+ inv_freq = 1.0 / (base ** (channel_range / head_dim))
137
+ t = torch.arange(seq_len, dtype=torch.float32, device=device)
138
+ freqs = torch.outer(t, inv_freq)
139
+ cos, sin = freqs.cos(), freqs.sin()
140
+ cos, sin = cos.bfloat16(), sin.bfloat16()
141
+ cos, sin = cos[None, :, None, :], sin[None, :, None, :]
142
+ return cos, sin
143
+
144
+ def forward(self, idx: torch.Tensor, targets: torch.Tensor | None = None, kv_cache=None, loss_reduction: str = 'mean'):
145
+ B, T = idx.size()
146
+ assert T <= self.cos.size(1)
147
+ assert idx.device == self.cos.device
148
+ assert self.cos.dtype == torch.bfloat16
149
+ T0 = 0 if kv_cache is None else kv_cache.get_pos()
150
+ cos_sin = self.cos[:, T0:T0+T], self.sin[:, T0:T0+T]
151
+ x = self.transformer.wte(idx)
152
+ x = norm(x)
153
+ for block in self.transformer.h:
154
+ x = block(x, cos_sin, kv_cache)
155
+ x = norm(x)
156
+ softcap = 15
157
+ if targets is not None:
158
+ logits = self.lm_head(x)
159
+ logits = softcap * torch.tanh(logits / softcap)
160
+ logits = logits.float()
161
+ loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1, reduction=loss_reduction)
162
+ return loss
163
+ else:
164
+ logits = self.lm_head(x)
165
+ logits = softcap * torch.tanh(logits / softcap)
166
+ return logits
167
+
168
+ @torch.inference_mode()
169
+ def generate(self, tokens: list[int], max_tokens: int, temperature: float = 1.0, top_k: int | None = None, seed: int = 42):
170
+ assert isinstance(tokens, list)
171
+ device = self.transformer.wte.weight.device
172
+ rng = None
173
+ if temperature > 0:
174
+ rng = torch.Generator(device=device)
175
+ rng.manual_seed(seed)
176
+ ids = torch.tensor([tokens], dtype=torch.long, device=device)
177
+ for _ in range(max_tokens):
178
+ logits = self.forward(ids)
179
+ logits = logits[:, -1, :]
180
+ if top_k is not None:
181
+ v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
182
+ logits[logits < v[:, [-1]]] = -float('Inf')
183
+ if temperature > 0:
184
+ logits = logits / max(temperature, 1e-6)
185
+ probs = F.softmax(logits, dim=-1)
186
+ next_ids = torch.multinomial(probs, num_samples=1, generator=rng)
187
+ else:
188
+ next_ids = torch.argmax(logits, dim=-1, keepdim=True)
189
+ ids = torch.cat((ids, next_ids), dim=1)
190
+ yield next_ids.item()
requirements.txt ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ gradio==4.41.0
2
+ huggingface_hub>=0.24.0
3
+ tokenizers>=0.14.0
4
+ safetensors>=0.4.3
5
+ torch