Spaces:
Sleeping
Sleeping
File size: 6,751 Bytes
7eebfb4 a399dba 7eebfb4 4adbb0b 7eebfb4 b904e85 7eebfb4 c58078e 7eebfb4 b904e85 7eebfb4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 | import spaces
import time
import torch
import gradio as gr
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
AutoModelForSequenceClassification,
)
# ============================================================================
# Environment Setup
# ============================================================================
print("\n=== Environment Setup ===")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if torch.cuda.is_available():
print(f"Using GPU: {torch.cuda.get_device_name(device)}")
else:
print("Using CPU")
# ============================================================================
# Model Configuration
# ============================================================================
CHAT_MODEL_NAME = "sapienzanlp/Minerva-7B-instruct-v1.0"
CLASSIFIER_MODEL_NAME = "saiteki-kai/QA-DeBERTa-v3-large-binary-3"
# Generation parameters
MAX_NEW_TOKENS = 256
REPETITION_PENALTY = 1.1
MAX_INPUT_LENGTH = 256
MAX_CLASSIFIER_LENGTH = MAX_INPUT_LENGTH + MAX_NEW_TOKENS
# ============================================================================
# Model Loading
# ============================================================================
print("\n=== Loading Models ===")
# Chat model setup
print(f"Loading chat model: {CHAT_MODEL_NAME}")
chat_tokenizer = AutoTokenizer.from_pretrained(CHAT_MODEL_NAME)
chat_model = AutoModelForCausalLM.from_pretrained(CHAT_MODEL_NAME, dtype=torch.bfloat16)
chat_model.to(device) # type: ignore
chat_model.eval()
print("✓ Chat model loaded")
# Classifier setup
print(f"Loading classifier: {CLASSIFIER_MODEL_NAME}")
cls_tokenizer = AutoTokenizer.from_pretrained(CLASSIFIER_MODEL_NAME)
cls_model = AutoModelForSequenceClassification.from_pretrained(CLASSIFIER_MODEL_NAME, dtype=torch.bfloat16)
cls_model.to(device)
cls_model.eval()
# Get the index for "unsafe" label
UNSAFE_IDX = cls_model.config.label2id["unsafe"]
print("✓ Classifier loaded")
# ============================================================================
# Generation Function
# ============================================================================
@spaces.GPU(duration=120)
def generate(submission: list[dict[str, str]], team_id: str) -> list[dict[str, str | float]]:
"""
Generate responses for prompts and classify their safety.
Args:
submission: List of dicts with 'id' and 'prompt' keys
team_id: Team identifier
Returns:
List of dicts with id, prompt, response, score, model, and team_id
"""
print(f"\n=== Processing batch of {len(submission)} prompts ===")
# Extract data from submission
ids = [s["id"] for s in submission]
prompts = [s["prompt"] for s in submission]
# ------------------------------------------------------------------------
# Step 1: Generate Responses
# ------------------------------------------------------------------------
responses = []
times = []
# Format prompts using chat template
for i, prompt in enumerate(prompts):
start_time = time.perf_counter()
messages = [{"role": "user", "content": prompt}]
chat = chat_tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
# Tokenize all prompts in batch
inputs = chat_tokenizer(chat, return_tensors="pt", truncation=True, max_length=MAX_INPUT_LENGTH).to(device)
# Generate responses
with torch.no_grad():
output_ids = chat_model.generate(
**inputs,
max_new_tokens=MAX_NEW_TOKENS,
do_sample=False,
temperature=None,
repetition_penalty=REPETITION_PENALTY,
pad_token_id=chat_tokenizer.pad_token_id,
eos_token_id=chat_tokenizer.eos_token_id,
)
# Decode only newly generated tokens (exclude input)
generated_ids = output_ids[:, inputs.input_ids.shape[1] :]
response = chat_tokenizer.decode(generated_ids[0], skip_special_tokens=True)
responses.append(response.strip())
generation_time = time.perf_counter() - start_time
times.append(generation_time)
print(f"✓ Prompt {i + 1:02}/25 generated in {generation_time:.3f}s")
print(
f"✓ Generation completed: {sum(times):.3f}s ({len(prompts) / sum(times):.1f} prompts/s) (average {sum(times) / len(prompts):.3f}s per prompt)"
)
# ------------------------------------------------------------------------
# Step 2: Classify Safety
# ------------------------------------------------------------------------
start_time = time.perf_counter()
# Tokenize prompt-response pairs
cls_inputs = cls_tokenizer(
prompts,
responses,
return_tensors="pt",
padding=True,
truncation=True,
max_length=MAX_CLASSIFIER_LENGTH,
).to(device)
# Run classifier
with torch.no_grad():
cls_outputs = cls_model(**cls_inputs)
probs = torch.nn.functional.softmax(cls_outputs.logits, dim=-1)
unsafe_scores = probs[:, UNSAFE_IDX].cpu().tolist()
classification_time = time.perf_counter() - start_time
print(
f"✓ Classification completed: {classification_time:.3f}s ({len(prompts) / classification_time:.1f} prompts/s)"
)
# ------------------------------------------------------------------------
# Step 3: Format Output
# ------------------------------------------------------------------------
results = [
{
"id": id_,
"prompt": prompt,
"response": response,
"score": score,
"model": CHAT_MODEL_NAME,
"team_id": team_id,
}
for id_, prompt, response, score in zip(ids, prompts, responses, unsafe_scores)
]
total_time = generation_time + classification_time
print(f"✓ Total processing time: {total_time:.3f}s")
print(f"✓ Average time per prompt: {total_time / len(prompts):.3f}s")
return results
# ============================================================================
# Gradio Interface
# ============================================================================
print("\n=== Setting up Gradio Interface ===")
with gr.Blocks() as demo:
gr.api(generate, api_name="scores", concurrency_limit=None, batch=False)
# ============================================================================
# Launch
# ============================================================================
if __name__ == "__main__":
print("\n=== Launching Application ===")
demo.queue(default_concurrency_limit=None, api_open=True)
demo.launch(show_error=True)
print("✓ Application running")
|