PlanetDOGE
commited on
Commit
·
160bd5c
1
Parent(s):
be55ff8
Delete KAI-1B_Demo.py
Browse files- KAI-1B_Demo.py +0 -143
KAI-1B_Demo.py
DELETED
@@ -1,143 +0,0 @@
|
|
1 |
-
from mistral.cache import RotatingBufferCache
|
2 |
-
import torch
|
3 |
-
import fire
|
4 |
-
from typing import List
|
5 |
-
from pathlib import Path
|
6 |
-
|
7 |
-
from mistral.model import Transformer
|
8 |
-
from mistral.tokenizer import Tokenizer
|
9 |
-
|
10 |
-
|
11 |
-
def sample_top_p(probs: torch.Tensor, p: float):
|
12 |
-
assert 0 <= p <= 1
|
13 |
-
|
14 |
-
probs_sort, probs_idx = torch.sort(probs, dim=-1, descending=True)
|
15 |
-
probs_sum = torch.cumsum(probs_sort, dim=-1)
|
16 |
-
mask = probs_sum - probs_sort > p
|
17 |
-
probs_sort[mask] = 0.0
|
18 |
-
probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
|
19 |
-
next_token = torch.multinomial(probs_sort, num_samples=1)
|
20 |
-
return torch.gather(probs_idx, -1, next_token)
|
21 |
-
|
22 |
-
|
23 |
-
def sample(logits: torch.Tensor, temperature: float, top_p: float):
|
24 |
-
if temperature > 0:
|
25 |
-
probs = torch.softmax(logits / temperature, dim=-1)
|
26 |
-
next_token = sample_top_p(probs, top_p)
|
27 |
-
else:
|
28 |
-
next_token = torch.argmax(logits, dim=-1).unsqueeze(0)
|
29 |
-
|
30 |
-
return next_token.reshape(-1)
|
31 |
-
|
32 |
-
|
33 |
-
@torch.inference_mode()
|
34 |
-
def generate(prompts: List[str], model: Transformer, tokenizer: Tokenizer, *, max_tokens: int, chunk_size: int = None, temperature: float = 0.7):
|
35 |
-
model = model.eval()
|
36 |
-
B, V = len(prompts), model.args.vocab_size
|
37 |
-
|
38 |
-
# Tokenize
|
39 |
-
encoded_prompts = [tokenizer.encode(prompt, bos=True) for prompt in prompts]
|
40 |
-
seqlens = [len(x) for x in encoded_prompts]
|
41 |
-
|
42 |
-
# Cache
|
43 |
-
cache_window = min(model.args.sliding_window, max(seqlens) + max_tokens)
|
44 |
-
cache = RotatingBufferCache(model.args.n_layers, model.args.max_batch_size, cache_window, model.args.n_kv_heads, model.args.head_dim)
|
45 |
-
cache.to(device=model.device, dtype=model.dtype)
|
46 |
-
cache.reset()
|
47 |
-
|
48 |
-
# Bookkeeping
|
49 |
-
logprobs = [[] for _ in range(B)]
|
50 |
-
last_token_prelogits = None
|
51 |
-
|
52 |
-
# One chunk if size not specified
|
53 |
-
max_prompt_len = max(seqlens)
|
54 |
-
if chunk_size is None:
|
55 |
-
chunk_size = max_prompt_len
|
56 |
-
|
57 |
-
# Encode prompt by chunks
|
58 |
-
for s in range(0, max_prompt_len, chunk_size):
|
59 |
-
prompt_chunks = [p[s:s+chunk_size] for p in encoded_prompts]
|
60 |
-
assert all(len(p) > 0 for p in prompt_chunks)
|
61 |
-
prelogits = model.forward(
|
62 |
-
torch.tensor(sum(prompt_chunks, []), device=model.device, dtype=torch.long),
|
63 |
-
seqlens=[len(p) for p in prompt_chunks],
|
64 |
-
cache=cache
|
65 |
-
)
|
66 |
-
logits = torch.log_softmax(prelogits, dim=-1)
|
67 |
-
|
68 |
-
if last_token_prelogits is not None:
|
69 |
-
# Pass > 1
|
70 |
-
last_token_logits = torch.log_softmax(last_token_prelogits, dim=-1)
|
71 |
-
for i_seq in range(B):
|
72 |
-
logprobs[i_seq].append(last_token_logits[i_seq, prompt_chunks[i_seq][0]].item())
|
73 |
-
|
74 |
-
offset = 0
|
75 |
-
for i_seq, sequence in enumerate(prompt_chunks):
|
76 |
-
logprobs[i_seq].extend([logits[offset + i, sequence[i + 1]].item() for i in range(len(sequence) - 1)])
|
77 |
-
offset += len(sequence)
|
78 |
-
|
79 |
-
last_token_prelogits = prelogits.index_select(0, torch.tensor([len(p) for p in prompt_chunks], device=prelogits.device).cumsum(dim=0) - 1)
|
80 |
-
assert last_token_prelogits.shape == (B, V)
|
81 |
-
|
82 |
-
# decode
|
83 |
-
generated_tokens = []
|
84 |
-
for i_token in range(max_tokens):
|
85 |
-
next_token = sample(last_token_prelogits, temperature=temperature, top_p=0.8)
|
86 |
-
|
87 |
-
last_token_logits = torch.log_softmax(last_token_prelogits, dim=-1)
|
88 |
-
for i in range(B):
|
89 |
-
logprobs[i].append(last_token_logits[i, next_token[i]].item())
|
90 |
-
|
91 |
-
generated_tokens.append(next_token[:, None])
|
92 |
-
last_token_prelogits = model.forward(next_token, seqlens=[1] * len(prompts), cache=cache)
|
93 |
-
assert last_token_prelogits.shape == (B, V)
|
94 |
-
|
95 |
-
generated_words = []
|
96 |
-
if generated_tokens:
|
97 |
-
generated_tokens = torch.cat(generated_tokens, 1)
|
98 |
-
for i, x in enumerate(encoded_prompts):
|
99 |
-
generated_words.append(tokenizer.decode(x + generated_tokens[i].tolist()))
|
100 |
-
|
101 |
-
return generated_words, logprobs
|
102 |
-
|
103 |
-
|
104 |
-
def interactive(model_path: str, max_tokens: int = 35, temperature: float = 0.7):
|
105 |
-
tokenizer = Tokenizer(str(Path(model_path) / "tokenizer.model"))
|
106 |
-
transformer = Transformer.from_folder(Path(model_path), max_batch_size=3)
|
107 |
-
|
108 |
-
while True:
|
109 |
-
prompt = input("Prompt: ")
|
110 |
-
res, _logprobs = generate(
|
111 |
-
[prompt],
|
112 |
-
transformer,
|
113 |
-
tokenizer,
|
114 |
-
max_tokens=max_tokens,
|
115 |
-
temperature=temperature,
|
116 |
-
)
|
117 |
-
print(res[0])
|
118 |
-
print("=====================")
|
119 |
-
|
120 |
-
def demo(model_path: str, max_tokens: int = 35, temperature: float = 0):
|
121 |
-
tokenizer = Tokenizer(str(Path(model_path) / "tokenizer.model"))
|
122 |
-
transformer = Transformer.from_folder(Path(model_path), max_batch_size=3)
|
123 |
-
|
124 |
-
res, _logprobs = generate(
|
125 |
-
[
|
126 |
-
"This is a test",
|
127 |
-
"This is another test",
|
128 |
-
"This is a third test, KAI is very good at testing. ",
|
129 |
-
],
|
130 |
-
transformer,
|
131 |
-
tokenizer,
|
132 |
-
max_tokens=max_tokens,
|
133 |
-
temperature=temperature,
|
134 |
-
)
|
135 |
-
for x in res:
|
136 |
-
print(x)
|
137 |
-
print("=====================")
|
138 |
-
|
139 |
-
if __name__ == "__main__":
|
140 |
-
fire.Fire({
|
141 |
-
"interactive": interactive,
|
142 |
-
"demo": demo,
|
143 |
-
})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|