Spaces:
Running
on
Zero
Running
on
Zero
Adding initial eval code
Browse files
app.py
CHANGED
@@ -50,26 +50,31 @@ def fused_sim(a:Image.Image,b:Image.Image,α=.5):
|
|
50 |
lp_sim = 1 - _LP(ta,tb,normalize=True).item()
|
51 |
return α*clip_sim + (1-α)*lp_sim
|
52 |
|
|
|
|
|
53 |
# ---------- load models once at startup ---------------------
|
54 |
@spaces.GPU
|
55 |
def load_models():
|
56 |
from unsloth import FastLanguageModel
|
57 |
global base, tok, lora
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
|
|
|
|
|
|
|
|
|
71 |
|
72 |
-
load_models()
|
73 |
|
74 |
def build_prompt(desc:str):
|
75 |
msgs=[{"role":"system","content":"You are an SVG illustrator."},
|
@@ -80,6 +85,7 @@ def build_prompt(desc:str):
|
|
80 |
@spaces.GPU
|
81 |
@torch.no_grad()
|
82 |
def draw(model, desc:str):
|
|
|
83 |
prompt = build_prompt(desc)
|
84 |
ids = tok(prompt, return_tensors="pt").to(DEVICE)
|
85 |
out = model.generate(**ids, max_new_tokens=MAX_NEW,
|
|
|
50 |
lp_sim = 1 - _LP(ta,tb,normalize=True).item()
|
51 |
return α*clip_sim + (1-α)*lp_sim
|
52 |
|
53 |
+
bnb_cfg = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_use_double_quant=True)
|
54 |
+
|
55 |
# ---------- load models once at startup ---------------------
|
56 |
@spaces.GPU
|
57 |
def load_models():
|
58 |
from unsloth import FastLanguageModel
|
59 |
global base, tok, lora
|
60 |
+
if base is None:
|
61 |
+
print("Loading BASE …")
|
62 |
+
base, tok = FastLanguageModel.from_pretrained(
|
63 |
+
BASE_MODEL, max_seq_length=2048,
|
64 |
+
load_in_4bit=True, quantization_config=bnb_cfg, device_map="auto")
|
65 |
+
tok.pad_token = tok.eos_token
|
66 |
+
|
67 |
+
print("Loading LoRA …")
|
68 |
+
lora, _ = FastLanguageModel.from_pretrained(
|
69 |
+
ADAPTER_DIR, max_seq_length=2048,
|
70 |
+
load_in_4bit=True, quantization_config=bnb_cfg, device_map="auto")
|
71 |
+
print("✔ models loaded")
|
72 |
|
73 |
+
@spaces.GPU
|
74 |
+
def ensure_models():
|
75 |
+
load_models()
|
76 |
+
return True # small, pickle-able sentinel
|
77 |
|
|
|
78 |
|
79 |
def build_prompt(desc:str):
|
80 |
msgs=[{"role":"system","content":"You are an SVG illustrator."},
|
|
|
85 |
@spaces.GPU
|
86 |
@torch.no_grad()
|
87 |
def draw(model, desc:str):
|
88 |
+
ensure_models()
|
89 |
prompt = build_prompt(desc)
|
90 |
ids = tok(prompt, return_tensors="pt").to(DEVICE)
|
91 |
out = model.generate(**ids, max_new_tokens=MAX_NEW,
|