Spaces:
Sleeping
Sleeping
afnotes commited on
Commit ·
3caefac
1
Parent(s): 037ad30
Added app.py
Browse files- app.py +97 -0
- requirements.txt +10 -0
app.py
ADDED
|
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import torch
|
| 3 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 4 |
+
from peft import PeftModel
|
| 5 |
+
import os
|
| 6 |
+
|
| 7 |
+
# --- Configuration ---
|
| 8 |
+
# The base Qwen model
|
| 9 |
+
BASE_MODEL_ID = "Qwen/Qwen3-4B"
|
| 10 |
+
# Your uploaded adapter model ID on Hugging Face Hub
|
| 11 |
+
ADAPTER_MODEL_ID = "afnotes/Qwen3-4B-YeI-LoRA" # IMPORTANT: Replace with your actual repo ID
|
| 12 |
+
|
| 13 |
+
# --- Model Loading (Optimized for Space) ---
|
| 14 |
+
# This function will be called once when the app starts
|
| 15 |
+
def load_model():
|
| 16 |
+
print(f"Loading tokenizer from: {ADAPTER_MODEL_ID}")
|
| 17 |
+
tokenizer = AutoTokenizer.from_pretrained(ADAPTER_MODEL_ID)
|
| 18 |
+
|
| 19 |
+
if tokenizer.eos_token is None:
|
| 20 |
+
tokenizer.eos_token = "<|endoftext|>"
|
| 21 |
+
print(f"Setting tokenizer.eos_token to: {tokenizer.eos_token}")
|
| 22 |
+
if tokenizer.pad_token is None:
|
| 23 |
+
tokenizer.pad_token = tokenizer.eos_token
|
| 24 |
+
print(f"Setting tokenizer.pad_token to: {tokenizer.pad_token}")
|
| 25 |
+
|
| 26 |
+
print(f"Loading base model: {BASE_MODEL_ID}")
|
| 27 |
+
base_model = AutoModelForCausalLM.from_pretrained(
|
| 28 |
+
BASE_MODEL_ID,
|
| 29 |
+
torch_dtype=torch.float16, # Use float16 for memory efficiency on GPU
|
| 30 |
+
device_map="auto",
|
| 31 |
+
trust_remote_code=True
|
| 32 |
+
)
|
| 33 |
+
|
| 34 |
+
print(f"Loading LoRA adapter from {ADAPTER_MODEL_ID} on top of base model...")
|
| 35 |
+
model = PeftModel.from_pretrained(base_model, ADAPTER_MODEL_ID)
|
| 36 |
+
# Merge LoRA weights for faster inference
|
| 37 |
+
model = model.merge_and_unload()
|
| 38 |
+
model.eval() # Set model to evaluation mode
|
| 39 |
+
print("Model loaded and LoRA merged successfully!")
|
| 40 |
+
return tokenizer, model.to("cuda" if torch.cuda.is_available() else "cpu") # Ensure model is on GPU if available
|
| 41 |
+
|
| 42 |
+
tokenizer, model = load_model() # Load model globally when app starts
|
| 43 |
+
|
| 44 |
+
# --- Generation Function for Gradio ---
|
| 45 |
+
def generate_ye_tweet(prompt_text, max_new_tokens, temperature, top_p, repetition_penalty):
|
| 46 |
+
print(f"--- Generating for prompt ---\n'{prompt_text}'")
|
| 47 |
+
|
| 48 |
+
inputs = tokenizer(prompt_text, return_tensors="pt").to(model.device)
|
| 49 |
+
input_token_len = inputs.input_ids.shape[1]
|
| 50 |
+
|
| 51 |
+
with torch.no_grad():
|
| 52 |
+
outputs = model.generate(
|
| 53 |
+
**inputs,
|
| 54 |
+
max_new_tokens=max_new_tokens,
|
| 55 |
+
temperature=temperature,
|
| 56 |
+
top_p=top_p,
|
| 57 |
+
do_sample=True,
|
| 58 |
+
repetition_penalty=repetition_penalty,
|
| 59 |
+
pad_token_id=tokenizer.eos_token_id,
|
| 60 |
+
eos_token_id=tokenizer.eos_token_id,
|
| 61 |
+
)
|
| 62 |
+
|
| 63 |
+
full_generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 64 |
+
|
| 65 |
+
if full_generated_text.startswith(prompt_text):
|
| 66 |
+
completion_text = full_generated_text[len(prompt_text):].strip()
|
| 67 |
+
else:
|
| 68 |
+
print("Warning: Decoded text did not start exactly with the prompt. Showing full output.")
|
| 69 |
+
completion_text = full_generated_text
|
| 70 |
+
|
| 71 |
+
print(f"\n--- Generated Completion (Kanye-style) ---")
|
| 72 |
+
print(completion_text)
|
| 73 |
+
print("------------------------------------------")
|
| 74 |
+
return completion_text
|
| 75 |
+
|
| 76 |
+
# --- Gradio Interface ---
|
| 77 |
+
iface = gr.Interface(
|
| 78 |
+
fn=generate_ye_tweet,
|
| 79 |
+
inputs=[
|
| 80 |
+
gr.Textbox(lines=2, label="Enter your prompt for Ye-style tweet:", placeholder="My next album will be called..."),
|
| 81 |
+
gr.Slider(minimum=10, maximum=100, value=50, step=5, label="Max New Tokens"),
|
| 82 |
+
gr.Slider(minimum=0.1, maximum=1.0, value=0.9, step=0.05, label="Temperature"),
|
| 83 |
+
gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top P"),
|
| 84 |
+
gr.Slider(minimum=1.0, maximum=2.0, value=1.0, step=0.05, label="Repetition Penalty")
|
| 85 |
+
],
|
| 86 |
+
outputs=gr.Textbox(label="Generated Ye Tweet"),
|
| 87 |
+
title="🎙️ Ye-I: Ye Tweet Generator 🎤",
|
| 88 |
+
description="Generate Ye West-style tweet snippets using a fine-tuned Qwen3-4B model. Remember the ZeroGPU limitations!",
|
| 89 |
+
examples=[
|
| 90 |
+
["I HATE"],
|
| 91 |
+
["My next album will be called"],
|
| 92 |
+
["Fashion is dead without"],
|
| 93 |
+
],
|
| 94 |
+
allow_flagging="never" # You can change this to "auto" if you want to allow flagging of bad outputs
|
| 95 |
+
)
|
| 96 |
+
|
| 97 |
+
iface.launch()
|
requirements.txt
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
bitsandbytes
|
| 2 |
+
hf-xet
|
| 3 |
+
peft
|
| 4 |
+
requests
|
| 5 |
+
safetensors
|
| 6 |
+
tokenizers
|
| 7 |
+
torch
|
| 8 |
+
transformers
|
| 9 |
+
gradio
|
| 10 |
+
accelerate
|