SykoSLM commited on
Commit
cde7cc2
·
verified ·
1 Parent(s): a391906

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +54 -22
README.md CHANGED
@@ -60,34 +60,66 @@ Input → [Global Mem | Local Mem | Chunk Tokens] → Transformer → Output
60
 
61
  ```python
62
  import torch
63
- from transformers import AutoModel, PreTrainedTokenizerFast
 
64
 
65
- # Model ve tokenizer yükle
66
- tokenizer = PreTrainedTokenizerFast.from_pretrained("SykoSLM/SykoLLM-CMN-V5.2-Beta")
67
- model = AutoModel.from_pretrained("SykoSLM/SykoLLM-CMN-V5.2-Beta", trust_remote_code=True)
68
- model.eval()
 
69
 
70
- # Bellek başlat
71
- device = "cuda" if torch.cuda.is_available() else "cpu"
72
- model = model.to(device)
73
 
74
- prev_memory = torch.zeros(1, 16, 768, device=device)
75
- global_memory = torch.zeros(1, 32, 768, device=device)
 
 
 
76
 
77
- # Tokenize et
78
- text = "Kullanıcı: Merhaba, nasılsın?\nAsistan:"
79
- inputs = tokenizer(text, return_tensors="pt").to(device)
 
 
 
 
 
 
 
 
 
 
 
 
 
80
 
81
- # Inference
82
  with torch.no_grad():
83
- logits, prev_memory, global_memory = model(
84
- input_ids=inputs["input_ids"],
85
- prev_memory=prev_memory,
86
- global_memory=global_memory
87
- )
88
-
89
- next_token = logits[0, -1, :].argmax(dim=-1)
90
- print(tokenizer.decode(next_token))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
91
  ```
92
 
93
  ---
 
60
 
61
  ```python
62
  import torch
63
+ import torch.nn.functional as F
64
+ import re
65
 
66
+ def sample_next_token(logits, temperature=0.1, top_p=0.9, repetition_penalty=1.2, input_ids=None):
67
+ # Tekrar cezası
68
+ if input_ids is not None:
69
+ for token_id in set(input_ids[0].tolist()):
70
+ logits[token_id] /= repetition_penalty
71
 
72
+ logits = logits / temperature
73
+ probs = F.softmax(logits, dim=-1)
 
74
 
75
+ sorted_probs, sorted_indices = torch.sort(probs, descending=True)
76
+ cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
77
+ sorted_indices_to_remove = cumulative_probs - sorted_probs > top_p
78
+ sorted_probs[sorted_indices_to_remove] = 0
79
+ sorted_probs /= sorted_probs.sum()
80
 
81
+ next_token = torch.multinomial(sorted_probs, num_samples=1)
82
+ return sorted_indices[next_token]
83
+
84
+ def fix_spacing(text):
85
+ # Noktalama önündeki boşlukları sil
86
+ text = re.sub(r'\s+([.,!?;:)])', r'\1', text)
87
+ # Fazla boşlukları tek boşluğa indir
88
+ text = re.sub(r'\s+', ' ', text)
89
+ return text.strip()
90
+
91
+ prev_memory = torch.zeros(1, config.num_memory_tokens, config.d_model, device=device)
92
+ global_memory = torch.zeros(1, config.num_global_memory_tokens, config.d_model, device=device)
93
+
94
+ text = "Kullanıcı: Merhaba, nasılsın?\nAsistan:"
95
+ inputs = tokenizer(text, return_tensors="pt").to(device)
96
+ generated = inputs["input_ids"].clone()
97
 
 
98
  with torch.no_grad():
99
+ for _ in range(100):
100
+ input_chunk = generated[:, -128:]
101
+
102
+ logits, prev_memory, global_memory = model(
103
+ input_ids=input_chunk,
104
+ prev_memory=prev_memory,
105
+ global_memory=global_memory
106
+ )
107
+
108
+ next_token = sample_next_token(
109
+ logits[0, -1, :],
110
+ temperature=0.1,
111
+ top_p=0.9,
112
+ repetition_penalty=1.3,
113
+ input_ids=generated
114
+ ).unsqueeze(0)
115
+
116
+ generated = torch.cat([generated, next_token], dim=1)
117
+
118
+ if next_token.item() == tokenizer.eos_token_id:
119
+ break
120
+
121
+ output = tokenizer.decode(generated[0], skip_special_tokens=True, clean_up_tokenization_spaces=True)
122
+ print(fix_spacing(output))
123
  ```
124
 
125
  ---