tt
Browse files
app.py
CHANGED
@@ -11,40 +11,28 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
11 |
)
|
12 |
model.eval()
|
13 |
|
14 |
-
def blend_generate(
|
15 |
-
|
16 |
-
promptB = f"<|system|>{sysB}\n<|user|>{user_message}\n<|assistant|>"
|
17 |
|
18 |
-
|
19 |
-
|
|
|
20 |
|
21 |
-
|
22 |
-
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
logitsA = model(input_ids=outA).logits[:, -1, :]
|
27 |
-
logitsB = model(input_ids=outB).logits[:, -1, :]
|
28 |
|
29 |
-
|
30 |
-
|
31 |
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
sorted_probs = sorted_probs / sorted_probs.sum(dim=-1, keepdim=True)
|
37 |
|
38 |
-
|
39 |
-
outA = torch.cat([outA, token.unsqueeze(0).unsqueeze(0)], dim=1)
|
40 |
-
outB = torch.cat([outB, token.unsqueeze(0).unsqueeze(0)], dim=1)
|
41 |
-
|
42 |
-
token_str = tokenizer.decode(token)
|
43 |
-
response += token_str
|
44 |
-
yield response
|
45 |
-
|
46 |
-
if token.item() == tokenizer.eos_token_id:
|
47 |
-
break
|
48 |
|
49 |
with gr.Blocks() as demo:
|
50 |
gr.Markdown("## Blended Prompt Chat (TinyLlama)")
|
|
|
11 |
)
|
12 |
model.eval()
|
13 |
|
14 |
+
def blend_generate(prompt, wa, wb):
|
15 |
+
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
|
|
|
16 |
|
17 |
+
with torch.no_grad():
|
18 |
+
output_a = model_a(input_ids)
|
19 |
+
output_b = model_b(input_ids)
|
20 |
|
21 |
+
logits_a = output_a.logits[:, -1, :]
|
22 |
+
logits_b = output_b.logits[:, -1, :]
|
23 |
|
24 |
+
# Weighted sum of raw logits (before softmax)
|
25 |
+
blended_logits = wa * logits_a + wb * logits_b
|
|
|
|
|
26 |
|
27 |
+
# Apply softmax safely to get valid probability distribution
|
28 |
+
probs = torch.softmax(blended_logits, dim=-1)
|
29 |
|
30 |
+
# Sample token from valid probability distribution
|
31 |
+
token = torch.multinomial(probs, 1)
|
32 |
+
next_token_id = token.item()
|
33 |
+
next_token = tokenizer.decode([next_token_id])
|
|
|
34 |
|
35 |
+
return next_token
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
with gr.Blocks() as demo:
|
38 |
gr.Markdown("## Blended Prompt Chat (TinyLlama)")
|