Upload app.py
Browse files
app.py
CHANGED
@@ -1,7 +1,7 @@
|
|
1 |
-
# app.py โโ Math-solution classifier
|
2 |
-
#
|
3 |
-
#
|
4 |
-
# gradio torch transformers peft accelerate spaces
|
5 |
|
6 |
import os
|
7 |
import json
|
@@ -9,107 +9,83 @@ import logging
|
|
9 |
from typing import Tuple
|
10 |
|
11 |
import gradio as gr
|
12 |
-
import
|
13 |
-
import spaces
|
14 |
-
|
15 |
-
from transformers import (
|
16 |
-
AutoTokenizer,
|
17 |
-
AutoModelForSequenceClassification,
|
18 |
-
)
|
19 |
-
|
20 |
-
# PEFT imports (optional)
|
21 |
-
try:
|
22 |
-
from peft.auto import (
|
23 |
-
AutoPeftModelForSequenceClassification,
|
24 |
-
AutoPeftModelForCausalLM,
|
25 |
-
)
|
26 |
-
PEFT_AVAILABLE = True
|
27 |
-
except ImportError: # PEFT not installed
|
28 |
-
PEFT_AVAILABLE = False
|
29 |
|
30 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
31 |
-
#
|
32 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
33 |
-
|
34 |
-
logger = logging.getLogger(__name__)
|
35 |
-
|
36 |
-
ADAPTER_PATH = os.getenv("ADAPTER_PATH", "./lora_adapter") # local dir or Hub ID
|
37 |
FALLBACK_MODEL = "distilbert-base-uncased"
|
38 |
LABELS = {0: "โ
Correct",
|
39 |
1: "๐ค Conceptual Error",
|
40 |
2: "๐ข Computational Error"}
|
41 |
|
42 |
-
|
|
|
43 |
|
|
|
44 |
model = None
|
45 |
tokenizer = None
|
46 |
model_ty = None # "classification" | "causal_lm" | "baseline"
|
47 |
|
48 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
49 |
-
#
|
50 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
51 |
-
def
|
52 |
-
"""
|
|
|
|
|
|
|
53 |
global model, tokenizer, model_ty
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
|
55 |
-
dtype = torch.float16
|
56 |
-
|
57 |
-
|
58 |
-
logger.info(f"Found adapter at {ADAPTER_PATH}")
|
59 |
|
60 |
-
# 1)
|
61 |
-
try:
|
62 |
model = AutoPeftModelForSequenceClassification.from_pretrained(
|
63 |
-
ADAPTER_PATH,
|
64 |
-
torch_dtype=dtype,
|
65 |
-
device_map="auto" if device == "cuda" else None,
|
66 |
)
|
67 |
model_ty = "classification"
|
68 |
-
logger.info("Loaded adapter as sequence-classifier")
|
69 |
except ValueError:
|
70 |
-
|
71 |
-
logger.info("Adapter is not a classifier โ trying causal-LM")
|
72 |
model = AutoPeftModelForCausalLM.from_pretrained(
|
73 |
-
ADAPTER_PATH,
|
74 |
-
torch_dtype=dtype,
|
75 |
-
device_map="auto" if device == "cuda" else None,
|
76 |
)
|
77 |
model_ty = "causal_lm"
|
78 |
-
logger.info("Loaded adapter as causal-LM")
|
79 |
|
80 |
tokenizer = AutoTokenizer.from_pretrained(ADAPTER_PATH)
|
81 |
|
82 |
else:
|
83 |
-
logger.warning("No adapter found โ using baseline DistilBERT
|
84 |
tokenizer = AutoTokenizer.from_pretrained(FALLBACK_MODEL)
|
85 |
model = AutoModelForSequenceClassification.from_pretrained(
|
86 |
-
FALLBACK_MODEL,
|
87 |
-
num_labels=3,
|
88 |
-
ignore_mismatched_sizes=True,
|
89 |
-
torch_dtype=dtype,
|
90 |
)
|
91 |
model_ty = "baseline"
|
92 |
|
93 |
-
# Make sure we have a pad token
|
94 |
if tokenizer.pad_token is None:
|
95 |
tokenizer.pad_token = tokenizer.eos_token or tokenizer.sep_token
|
96 |
|
97 |
-
model.to(device)
|
98 |
model.eval()
|
99 |
-
logger.info(f"Model ready
|
|
|
100 |
|
101 |
-
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
102 |
-
# Inference helpers
|
103 |
-
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
104 |
def _classify_logits(question: str, solution: str) -> Tuple[str, str, str]:
|
|
|
105 |
text = f"Question: {question}\n\nSolution:\n{solution}"
|
106 |
inputs = tokenizer(
|
107 |
-
text,
|
108 |
-
|
109 |
-
padding=True,
|
110 |
-
truncation=True,
|
111 |
-
max_length=512,
|
112 |
-
).to(device)
|
113 |
|
114 |
with torch.no_grad():
|
115 |
logits = model(**inputs).logits
|
@@ -117,14 +93,14 @@ def _classify_logits(question: str, solution: str) -> Tuple[str, str, str]:
|
|
117 |
pred = int(torch.argmax(probs))
|
118 |
conf = f"{probs[pred].item():.3f}"
|
119 |
|
120 |
-
return LABELS
|
|
|
121 |
|
122 |
def _classify_generate(question: str, solution: str) -> Tuple[str, str, str]:
|
123 |
-
|
124 |
prompt = (
|
125 |
"You are a mathematics tutor.\n"
|
126 |
-
"You are given a math word problem and a student's solution. "
|
127 |
-
"Decide whether the solution is correct.\n\n"
|
128 |
"- Correct = all reasoning and calculations are correct.\n"
|
129 |
"- Conceptual Error = reasoning is wrong.\n"
|
130 |
"- Computational Error= reasoning okay but arithmetic off.\n\n"
|
@@ -135,8 +111,7 @@ def _classify_generate(question: str, solution: str) -> Tuple[str, str, str]:
|
|
135 |
f"Question: {question}\n\nSolution:\n{solution}\n\nAnswer:"
|
136 |
)
|
137 |
|
138 |
-
|
139 |
-
inputs = tokenizer(prompt, return_tensors="pt").to(device)
|
140 |
with torch.no_grad():
|
141 |
out_ids = model.generate(
|
142 |
**inputs,
|
@@ -148,12 +123,9 @@ def _classify_generate(question: str, solution: str) -> Tuple[str, str, str]:
|
|
148 |
skip_special_tokens=True,
|
149 |
).strip()
|
150 |
|
151 |
-
|
152 |
-
# Try to parse last JSON line
|
153 |
verdict = "Unparsed"
|
154 |
try:
|
155 |
-
|
156 |
-
data = json.loads(line)
|
157 |
v = data.get("verdict", "").lower()
|
158 |
if v.startswith("corr"):
|
159 |
verdict = LABELS[0]
|
@@ -166,21 +138,30 @@ def _classify_generate(question: str, solution: str) -> Tuple[str, str, str]:
|
|
166 |
|
167 |
return verdict, "", generated
|
168 |
|
169 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
170 |
if not question.strip() or not solution.strip():
|
171 |
-
return "Please
|
172 |
|
173 |
if model_ty in ("classification", "baseline"):
|
174 |
return _classify_logits(question, solution)
|
175 |
-
|
176 |
return _classify_generate(question, solution)
|
177 |
-
else:
|
178 |
-
return "Model not loaded.", "", ""
|
179 |
|
180 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
181 |
-
#
|
182 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
183 |
-
|
|
|
|
|
184 |
|
185 |
with gr.Blocks(title="Math Solution Classifier") as demo:
|
186 |
gr.Markdown("# ๐งฎ Math Solution Classifier")
|
@@ -191,15 +172,15 @@ with gr.Blocks(title="Math Solution Classifier") as demo:
|
|
191 |
|
192 |
with gr.Row():
|
193 |
with gr.Column():
|
194 |
-
q_in
|
195 |
-
s_in
|
196 |
-
btn
|
197 |
with gr.Column():
|
198 |
verdict = gr.Textbox(label="Verdict", interactive=False)
|
199 |
conf = gr.Textbox(label="Confidence", interactive=False)
|
200 |
raw = gr.Textbox(label="Model Output", interactive=False)
|
201 |
|
202 |
-
btn.click(
|
203 |
|
204 |
gr.Examples(
|
205 |
[
|
@@ -210,11 +191,6 @@ with gr.Blocks(title="Math Solution Classifier") as demo:
|
|
210 |
inputs=[q_in, s_in],
|
211 |
)
|
212 |
|
213 |
-
|
214 |
-
@spaces.GPU # or @spaces.CPU if you deploy on CPU
|
215 |
-
|
216 |
def launch_app():
|
217 |
-
return demo
|
218 |
-
|
219 |
-
if __name__ == "__main__":
|
220 |
-
demo.launch()
|
|
|
1 |
+
# app.py โโ Math-solution classifier on HF Spaces (Zero-GPU-safe)
|
2 |
+
#
|
3 |
+
# Pin in requirements.txt:
|
4 |
+
# gradio==4.44.0 torch==2.1.0 transformers==4.35.0 peft==0.7.1 accelerate==0.25.0 spaces
|
5 |
|
6 |
import os
|
7 |
import json
|
|
|
9 |
from typing import Tuple
|
10 |
|
11 |
import gradio as gr
|
12 |
+
import spaces # <- Hugging Face Spaces SDK (Zero)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
15 |
+
# CONSTANTS (no CUDA use here)
|
16 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
17 |
+
ADAPTER_PATH = os.getenv("ADAPTER_PATH", "./lora_adapter") # dir or Hub repo
|
|
|
|
|
|
|
18 |
FALLBACK_MODEL = "distilbert-base-uncased"
|
19 |
LABELS = {0: "โ
Correct",
|
20 |
1: "๐ค Conceptual Error",
|
21 |
2: "๐ข Computational Error"}
|
22 |
|
23 |
+
logging.basicConfig(level=logging.INFO)
|
24 |
+
logger = logging.getLogger(__name__)
|
25 |
|
26 |
+
# Globals that will live **inside the GPU worker**
|
27 |
model = None
|
28 |
tokenizer = None
|
29 |
model_ty = None # "classification" | "causal_lm" | "baseline"
|
30 |
|
31 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
32 |
+
# GPU-SIDE INITIALISATION & INFERENCE
|
33 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
34 |
+
def _load_model_gpu():
|
35 |
+
"""
|
36 |
+
Runs **inside the GPU worker**.
|
37 |
+
Tries LoRA classification adapter โ LoRA causal-LM adapter โ plain baseline.
|
38 |
+
"""
|
39 |
global model, tokenizer, model_ty
|
40 |
+
import torch
|
41 |
+
from transformers import (
|
42 |
+
AutoTokenizer,
|
43 |
+
AutoModelForSequenceClassification,
|
44 |
+
)
|
45 |
+
from peft.auto import (
|
46 |
+
AutoPeftModelForSequenceClassification,
|
47 |
+
AutoPeftModelForCausalLM,
|
48 |
+
)
|
49 |
|
50 |
+
dtype = torch.float16
|
51 |
+
if os.path.isdir(ADAPTER_PATH):
|
52 |
+
logger.info(f"[GPU] Loading adapter from {ADAPTER_PATH}")
|
|
|
53 |
|
54 |
+
try: # 1) classification adapter
|
|
|
55 |
model = AutoPeftModelForSequenceClassification.from_pretrained(
|
56 |
+
ADAPTER_PATH, torch_dtype=dtype, device_map="auto"
|
|
|
|
|
57 |
)
|
58 |
model_ty = "classification"
|
|
|
59 |
except ValueError:
|
60 |
+
logger.info("[GPU] Not a classifier, trying causal-LM")
|
|
|
61 |
model = AutoPeftModelForCausalLM.from_pretrained(
|
62 |
+
ADAPTER_PATH, torch_dtype=dtype, device_map="auto"
|
|
|
|
|
63 |
)
|
64 |
model_ty = "causal_lm"
|
|
|
65 |
|
66 |
tokenizer = AutoTokenizer.from_pretrained(ADAPTER_PATH)
|
67 |
|
68 |
else:
|
69 |
+
logger.warning("[GPU] No adapter found โ using baseline DistilBERT")
|
70 |
tokenizer = AutoTokenizer.from_pretrained(FALLBACK_MODEL)
|
71 |
model = AutoModelForSequenceClassification.from_pretrained(
|
72 |
+
FALLBACK_MODEL, num_labels=3, ignore_mismatched_sizes=True
|
|
|
|
|
|
|
73 |
)
|
74 |
model_ty = "baseline"
|
75 |
|
|
|
76 |
if tokenizer.pad_token is None:
|
77 |
tokenizer.pad_token = tokenizer.eos_token or tokenizer.sep_token
|
78 |
|
|
|
79 |
model.eval()
|
80 |
+
logger.info(f"[GPU] Model ready ({model_ty})")
|
81 |
+
|
82 |
|
|
|
|
|
|
|
83 |
def _classify_logits(question: str, solution: str) -> Tuple[str, str, str]:
|
84 |
+
import torch
|
85 |
text = f"Question: {question}\n\nSolution:\n{solution}"
|
86 |
inputs = tokenizer(
|
87 |
+
text, return_tensors="pt", padding=True, truncation=True, max_length=512
|
88 |
+
).to("cuda")
|
|
|
|
|
|
|
|
|
89 |
|
90 |
with torch.no_grad():
|
91 |
logits = model(**inputs).logits
|
|
|
93 |
pred = int(torch.argmax(probs))
|
94 |
conf = f"{probs[pred].item():.3f}"
|
95 |
|
96 |
+
return LABELS[pred], conf, "โ"
|
97 |
+
|
98 |
|
99 |
def _classify_generate(question: str, solution: str) -> Tuple[str, str, str]:
|
100 |
+
import torch
|
101 |
prompt = (
|
102 |
"You are a mathematics tutor.\n"
|
103 |
+
"You are given a math word problem and a student's solution. Decide whether the solution is correct.\n\n"
|
|
|
104 |
"- Correct = all reasoning and calculations are correct.\n"
|
105 |
"- Conceptual Error = reasoning is wrong.\n"
|
106 |
"- Computational Error= reasoning okay but arithmetic off.\n\n"
|
|
|
111 |
f"Question: {question}\n\nSolution:\n{solution}\n\nAnswer:"
|
112 |
)
|
113 |
|
114 |
+
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
|
|
|
115 |
with torch.no_grad():
|
116 |
out_ids = model.generate(
|
117 |
**inputs,
|
|
|
123 |
skip_special_tokens=True,
|
124 |
).strip()
|
125 |
|
|
|
|
|
126 |
verdict = "Unparsed"
|
127 |
try:
|
128 |
+
data = json.loads(generated.splitlines()[-1])
|
|
|
129 |
v = data.get("verdict", "").lower()
|
130 |
if v.startswith("corr"):
|
131 |
verdict = LABELS[0]
|
|
|
138 |
|
139 |
return verdict, "", generated
|
140 |
|
141 |
+
|
142 |
+
@spaces.GPU # <-- every CUDA op happens inside here
|
143 |
+
def gpu_classify(question: str, solution: str):
|
144 |
+
"""
|
145 |
+
Proxy target for Gradio. Executed in the GPU worker so CUDA is allowed.
|
146 |
+
Returns (verdict, confidence, raw_output)
|
147 |
+
"""
|
148 |
+
if model is None:
|
149 |
+
_load_model_gpu()
|
150 |
+
|
151 |
if not question.strip() or not solution.strip():
|
152 |
+
return "Please fill both fields.", "", ""
|
153 |
|
154 |
if model_ty in ("classification", "baseline"):
|
155 |
return _classify_logits(question, solution)
|
156 |
+
else: # causal_lm
|
157 |
return _classify_generate(question, solution)
|
|
|
|
|
158 |
|
159 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
160 |
+
# CPU-SIDE UI (no torch.cuda here)
|
161 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
162 |
+
def classify_proxy(q, s):
|
163 |
+
"""Simple wrapper so Gradio can call the GPU function."""
|
164 |
+
return gpu_classify(q, s)
|
165 |
|
166 |
with gr.Blocks(title="Math Solution Classifier") as demo:
|
167 |
gr.Markdown("# ๐งฎ Math Solution Classifier")
|
|
|
172 |
|
173 |
with gr.Row():
|
174 |
with gr.Column():
|
175 |
+
q_in = gr.Textbox(label="Math Question", lines=3)
|
176 |
+
s_in = gr.Textbox(label="Proposed Solution", lines=6)
|
177 |
+
btn = gr.Button("Classify", variant="primary")
|
178 |
with gr.Column():
|
179 |
verdict = gr.Textbox(label="Verdict", interactive=False)
|
180 |
conf = gr.Textbox(label="Confidence", interactive=False)
|
181 |
raw = gr.Textbox(label="Model Output", interactive=False)
|
182 |
|
183 |
+
btn.click(classify_proxy, [q_in, s_in], [verdict, conf, raw])
|
184 |
|
185 |
gr.Examples(
|
186 |
[
|
|
|
191 |
inputs=[q_in, s_in],
|
192 |
)
|
193 |
|
194 |
+
@spaces.CPU # UI served from the CPU worker
|
|
|
|
|
195 |
def launch_app():
|
196 |
+
return demo
|
|
|
|
|
|