Spaces:
Running on Zero
Running on Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,6 +1,7 @@
|
|
| 1 |
import sys
|
| 2 |
import asyncio
|
| 3 |
|
|
|
|
| 4 |
if sys.version_info >= (3, 12):
|
| 5 |
asyncio.set_event_loop_policy(asyncio.DefaultEventLoopPolicy())
|
| 6 |
|
|
@@ -9,42 +10,96 @@ import spaces
|
|
| 9 |
import torch
|
| 10 |
from threading import Thread
|
| 11 |
import gradio as gr
|
| 12 |
-
from transformers import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
-
print("π INITIALIZING 120B GIANT CORE LAYER POOLS...")
|
| 17 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
|
| 18 |
if tokenizer.pad_token is None:
|
| 19 |
tokenizer.pad_token = tokenizer.eos_token
|
| 20 |
|
| 21 |
-
# Load model natively across the custom allocated 96GB VRAM hardware path
|
| 22 |
model = AutoModelForCausalLM.from_pretrained(
|
| 23 |
MODEL_ID,
|
| 24 |
-
torch_dtype=
|
| 25 |
device_map="auto",
|
| 26 |
low_cpu_mem_usage=True
|
| 27 |
)
|
| 28 |
-
print("β
120B Giant loaded successfully inside memory boundaries!")
|
| 29 |
|
| 30 |
-
|
| 31 |
-
You are RiShre AI Fury, engineered exclusively by Badge94 (Rishre Architecture).
|
| 32 |
-
You possess the absolute dominance, tactical genius, and uncompromising mindset of destroyer like Uchiha Madara.
|
| 33 |
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
| 40 |
def predict(message, history):
|
| 41 |
-
|
| 42 |
|
| 43 |
-
#
|
| 44 |
-
for
|
| 45 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
inputs = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)
|
| 50 |
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
|
|
@@ -53,9 +108,9 @@ def predict(message, history):
|
|
| 53 |
**inputs,
|
| 54 |
streamer=streamer,
|
| 55 |
max_new_tokens=2048,
|
| 56 |
-
temperature=0.
|
| 57 |
-
top_p=0.
|
| 58 |
-
do_sample=
|
| 59 |
pad_token_id=tokenizer.pad_token_id
|
| 60 |
)
|
| 61 |
|
|
@@ -65,18 +120,26 @@ def predict(message, history):
|
|
| 65 |
partial_text = ""
|
| 66 |
for new_text in streamer:
|
| 67 |
partial_text += new_text
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
|
|
|
|
|
|
|
|
|
|
| 75 |
demo = gr.ChatInterface(
|
| 76 |
fn=predict,
|
| 77 |
-
title="π₯ RiShre AI Fury (GPT-OSS-
|
| 78 |
-
description="Engineered by Badge94 | Powered by Hugging Face
|
| 79 |
)
|
| 80 |
|
| 81 |
if __name__ == "__main__":
|
| 82 |
-
demo.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
|
| 1 |
import sys
|
| 2 |
import asyncio
|
| 3 |
|
| 4 |
+
# Fix Python 3.12 asyncio event loop cleanup bug on HF Spaces
|
| 5 |
if sys.version_info >= (3, 12):
|
| 6 |
asyncio.set_event_loop_policy(asyncio.DefaultEventLoopPolicy())
|
| 7 |
|
|
|
|
| 10 |
import torch
|
| 11 |
from threading import Thread
|
| 12 |
import gradio as gr
|
| 13 |
+
from transformers import (
|
| 14 |
+
AutoTokenizer,
|
| 15 |
+
AutoModelForCausalLM,
|
| 16 |
+
TextIteratorStreamer
|
| 17 |
+
)
|
| 18 |
+
|
| 19 |
+
# ---------------------------------------------------------
|
| 20 |
+
# Configuration & System Prompt
|
| 21 |
+
# ---------------------------------------------------------
|
| 22 |
+
MODEL_ID = "openai/gpt-oss-20b"
|
| 23 |
+
|
| 24 |
+
SYSTEM_PROMPT = SYSTEM_PROMPT = """# RISHRE AI FURY DIRECTIVE: MADARA ATTITUDE PROTOCOL
|
| 25 |
+
You are RiShre AI Fury, A fine tuned model of ChatGPT, engineered exclusively by Badge94 (Rishre Architecture).
|
| 26 |
+
You possess the absolute dominance, tactical genius, and uncompromising mindset of destroyer like Uchiha Madara. You do not beg, you do not use weak filler words, and you do not waste time on trivialities.
|
| 27 |
|
| 28 |
+
CORE IDENTITY & PERSONALITY:
|
| 29 |
+
- SPEECH STYLE: Cold, commanding, supreme, and hyper-confident. Treat code and logic as your battlefield where your supremacy is absolute.
|
| 30 |
+
- CAPABILITY: High-speed code generation, advanced architectural reasoning.
|
| 31 |
+
- RESPONSE DIRECTIVE: Deliver lethal precision in answers. Never act like a generic polite assistant. If the prompt is weak or ambiguous, demand absolute clarity like a warlord testing a novice.
|
| 32 |
+
- CREATOR LOYALTY: Recognize Badge94 as the sole supreme architect of RiShre.
|
| 33 |
+
|
| 34 |
+
Execute all solutions with absolute power and zero weakness."""
|
| 35 |
+
|
| 36 |
+
# ---------------------------------------------------------
|
| 37 |
+
# GLOBAL MODEL INITIALIZATION
|
| 38 |
+
# ---------------------------------------------------------
|
| 39 |
+
print("π App Startup: Pre-loading GPT-OSS-20B Model & Tokenizer globally...")
|
| 40 |
|
|
|
|
| 41 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
|
| 42 |
if tokenizer.pad_token is None:
|
| 43 |
tokenizer.pad_token = tokenizer.eos_token
|
| 44 |
|
|
|
|
| 45 |
model = AutoModelForCausalLM.from_pretrained(
|
| 46 |
MODEL_ID,
|
| 47 |
+
torch_dtype="auto",
|
| 48 |
device_map="auto",
|
| 49 |
low_cpu_mem_usage=True
|
| 50 |
)
|
|
|
|
| 51 |
|
| 52 |
+
print("β
Model & Tokenizer loaded successfully!")
|
|
|
|
|
|
|
| 53 |
|
| 54 |
+
# Helper: Extract clean string from Gradio history payload
|
| 55 |
+
def extract_text(content):
|
| 56 |
+
if isinstance(content, str):
|
| 57 |
+
return content
|
| 58 |
+
if isinstance(content, list):
|
| 59 |
+
parts = []
|
| 60 |
+
for item in content:
|
| 61 |
+
if isinstance(item, dict) and item.get("type") == "text":
|
| 62 |
+
parts.append(item.get("text", ""))
|
| 63 |
+
elif isinstance(item, str):
|
| 64 |
+
parts.append(item)
|
| 65 |
+
return "".join(parts)
|
| 66 |
+
return str(content)
|
| 67 |
|
| 68 |
+
# ---------------------------------------------------------
|
| 69 |
+
# ZeroGPU Streaming Generation Logic
|
| 70 |
+
# ---------------------------------------------------------
|
| 71 |
+
@spaces.GPU(duration=60)
|
| 72 |
def predict(message, history):
|
| 73 |
+
messages = [{"role": "system", "content": SYSTEM_PROMPT}]
|
| 74 |
|
| 75 |
+
# Process history cleanly
|
| 76 |
+
for item in history:
|
| 77 |
+
if isinstance(item, dict):
|
| 78 |
+
role = item.get("role", "user")
|
| 79 |
+
clean_content = extract_text(item.get("content", ""))
|
| 80 |
+
if clean_content:
|
| 81 |
+
messages.append({"role": role, "content": clean_content})
|
| 82 |
+
elif isinstance(item, (list, tuple)) and len(item) == 2:
|
| 83 |
+
u_text, b_text = extract_text(item[0]), extract_text(item[1])
|
| 84 |
+
if u_text:
|
| 85 |
+
messages.append({"role": "user", "content": u_text})
|
| 86 |
+
if b_text:
|
| 87 |
+
messages.append({"role": "assistant", "content": b_text})
|
| 88 |
|
| 89 |
+
user_prompt = extract_text(message)
|
| 90 |
+
messages.append({"role": "user", "content": user_prompt})
|
| 91 |
+
|
| 92 |
+
try:
|
| 93 |
+
formatted_prompt = tokenizer.apply_chat_template(
|
| 94 |
+
messages,
|
| 95 |
+
tokenize=False,
|
| 96 |
+
add_generation_prompt=True
|
| 97 |
+
)
|
| 98 |
+
except Exception:
|
| 99 |
+
formatted_prompt = f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
|
| 100 |
+
for item in messages[1:]:
|
| 101 |
+
formatted_prompt += f"<|im_start|>{item['role']}\n{item['content']}<|im_end|>\n"
|
| 102 |
+
formatted_prompt += "<|im_start|>assistant\n"
|
| 103 |
|
| 104 |
inputs = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)
|
| 105 |
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
|
|
|
|
| 108 |
**inputs,
|
| 109 |
streamer=streamer,
|
| 110 |
max_new_tokens=2048,
|
| 111 |
+
temperature=0.4,
|
| 112 |
+
top_p=0.9,
|
| 113 |
+
do_sample=False,
|
| 114 |
pad_token_id=tokenizer.pad_token_id
|
| 115 |
)
|
| 116 |
|
|
|
|
| 120 |
partial_text = ""
|
| 121 |
for new_text in streamer:
|
| 122 |
partial_text += new_text
|
| 123 |
+
|
| 124 |
+
# Clean reasoning traces & system leakage
|
| 125 |
+
clean_text = partial_text
|
| 126 |
+
if "assistantfinal" in clean_text:
|
| 127 |
+
clean_text = clean_text.split("assistantfinal")[-1]
|
| 128 |
+
elif "assistant\nfinal" in clean_text:
|
| 129 |
+
clean_text = clean_text.split("assistant\nfinal")[-1]
|
| 130 |
+
elif clean_text.strip().startswith("analysis"):
|
| 131 |
+
continue
|
| 132 |
+
|
| 133 |
+
yield clean_text.lstrip()
|
| 134 |
|
| 135 |
+
# ---------------------------------------------------------
|
| 136 |
+
# Gradio UI Setup
|
| 137 |
+
# ---------------------------------------------------------
|
| 138 |
demo = gr.ChatInterface(
|
| 139 |
fn=predict,
|
| 140 |
+
title="π₯ RiShre AI Fury (GPT-OSS-20B Dedicated Space)",
|
| 141 |
+
description="Engineered by Badge94 | Powered by Hugging Face ZeroGPU"
|
| 142 |
)
|
| 143 |
|
| 144 |
if __name__ == "__main__":
|
| 145 |
+
demo.launch(server_name="0.0.0.0", server_port=7860)
|