lea97338 commited on
Commit
9d1c9f7
·
verified ·
1 Parent(s): 3cd1802

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +81 -0
README.md ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ language:
4
+ - fr
5
+ pipeline_tag: text-generation
6
+ ---
7
+ ```!pip -q install torch huggingface_hub gradio
8
+
9
+ import re, torch, gradio as gr
10
+ import torch.nn as nn
11
+ from huggingface_hub import hf_hub_download
12
+
13
+ REPO = "lea97338/LunaLMV2"
14
+ DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
15
+
16
+ # Téléchargement public : aucun token
17
+ path = hf_hub_download(REPO, "luna_model.pth")
18
+ ckpt = torch.load(path, map_location="cpu", weights_only=False)
19
+
20
+ class LunaLM(nn.Module):
21
+ def __init__(self, vocab_size, d_model=128, max_len=64):
22
+ super().__init__()
23
+ self.max_len, self.vocab_size = max_len, vocab_size
24
+ self.wte = nn.Embedding(vocab_size, d_model)
25
+ self.wpe = nn.Embedding(max_len, d_model)
26
+ self.ln1, self.ln2, self.ln_f = (
27
+ nn.LayerNorm(d_model), nn.LayerNorm(d_model),
28
+ nn.LayerNorm(d_model))
29
+ self.q_proj, self.k_proj, self.v_proj, self.out_proj = [
30
+ nn.Linear(d_model, d_model) for _ in range(4)]
31
+ self.mlp = nn.Sequential(
32
+ nn.Linear(d_model, d_model*2), nn.GELU(),
33
+ nn.Linear(d_model*2, d_model))
34
+ self.lm_head = nn.Linear(d_model, vocab_size, bias=False)
35
+
36
+ def forward(self, idx):
37
+ _, T = idx.shape
38
+ x = self.wte(idx) + self.wpe(
39
+ torch.arange(T, device=idx.device).unsqueeze(0))
40
+ n = self.ln1(x)
41
+ q, k, v = self.q_proj(n), self.k_proj(n), self.v_proj(n)
42
+ a = q @ k.transpose(-2, -1) / (k.size(-1) ** 0.5)
43
+ mask = torch.tril(torch.ones(T, T, device=idx.device)).bool()
44
+ x = x + self.out_proj(
45
+ torch.softmax(a.masked_fill(~mask, float("-inf")), -1) @ v)
46
+ return self.lm_head(self.ln_f(x + self.mlp(self.ln2(x))))
47
+
48
+ @staticmethod
49
+ def clean_tok(text):
50
+ return re.findall(r"\w+|[^\w\s]", text)
51
+
52
+ model = LunaLM(len(ckpt["w2i"]))
53
+ model.load_state_dict(ckpt["model"])
54
+ model.w2i, model.i2w = ckpt["w2i"], ckpt["i2w"]
55
+ model = model.to(DEVICE).eval()
56
+
57
+ @torch.inference_mode()
58
+ def generate_stream(prompt):
59
+ ids = [model.w2i["<bos>"]] + [
60
+ model.w2i.get(w, 3)
61
+ for w in ["user"] + model.clean_tok(prompt) + ["luna"]]
62
+ ids = ids[:model.max_len]
63
+ result = ""
64
+
65
+ while len(ids) < model.max_len:
66
+ x = torch.tensor([ids], device=DEVICE)
67
+ token = int(model(x)[0, -1].argmax().item())
68
+ if token in (model.w2i["<eos>"], model.w2i["<pad>"]):
69
+ break
70
+ word = model.i2w.get(token, "")
71
+ result += ("" if word in ".,!?;:" else " ") + word
72
+ ids.append(token)
73
+ yield result.strip()
74
+
75
+ demo = gr.ChatInterface(
76
+ fn=generate_stream,
77
+ title="🌙 LunaLMV2",
78
+ description="Inférence seule — génération en streaming."
79
+ )
80
+ demo.queue().launch(share=True)
81
+ ```