File size: 7,571 Bytes
9382e3f |
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 |
import json
import os
import time
import torch
import gradio as gr
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
import random
# Environment variables
os.environ["TOKENIZERS_PARALLELISM"] = "0"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
# Global variables to store the model and tokenizer
model = None
tokenizer = None
# Load model and tokenizer
def load_model_and_tokenizer(model_name, dtype, kv_bits):
global model, tokenizer
if model is None or tokenizer is None:
print("Loading model and tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(model_name)
special_tokens = {"pad_token": "<PAD>"}
tokenizer.add_special_tokens(special_tokens)
config = AutoConfig.from_pretrained(model_name)
if kv_bits != "unquantized":
quantizer_path = f"codebooks/{model_name.split('/')[-1]}_{kv_bits}bit.xmad"
setattr(config, "quantizer_path", quantizer_path)
if dtype == "bf16":
dtype = torch.bfloat16
elif dtype == "fp16":
dtype = torch.float16
elif dtype == "fp32":
dtype = torch.float32
model = AutoModelForCausalLM.from_pretrained(model_name, config=config, torch_dtype=dtype, device_map="auto")
if len(tokenizer) > model.get_input_embeddings().weight.shape[0]:
model.resize_token_embeddings(len(tokenizer))
tokenizer.padding_side = "left"
model.config.pad_token_id = tokenizer.pad_token_id
return model, tokenizer
# Format response
def format_response(dialog, response):
question = next((turn['content'] for turn in dialog if turn['role'] == 'user'), 'No question found')
answer = response.split("assistant")[-1].strip()
return {"question": question, "answer": answer}
# Load questions
def load_questions(prompts_path, custom_questions):
with open(prompts_path, "r") as file:
dialogs = json.load(file)
selected_dialogs = []
if custom_questions:
for question in custom_questions:
if question.strip():
custom_dialog = [{"role": "user", "content": question}]
selected_dialogs.append(custom_dialog)
num_questions = 30 - len(selected_dialogs)
random.shuffle(dialogs)
selected_dialogs.extend(dialogs[:num_questions])
return selected_dialogs[:30]
# Inference
def infer(model_name, dialogs, num_new_tokens, temperature, dtype, kv_bits, progress=gr.Progress()):
print("Starting inference...")
model, tokenizer = load_model_and_tokenizer(model_name, dtype, kv_bits)
batch_inputs = [
tokenizer.apply_chat_template(dialog, tokenize=False, add_generation_prompt=True)
for dialog in dialogs
]
responses = []
start_time = time.time()
batch_size = 30 # Set batch size for processing, this can be adjusted
num_dialogs = len(dialogs)
total_time = 0
total_tokens = 0
num_batches = (num_dialogs + batch_size - 1) // batch_size
for batch_idx in range(num_batches):
start_idx = batch_idx * batch_size
end_idx = min(start_idx + batch_size, num_dialogs)
batch = batch_inputs[start_idx:end_idx]
encoded_inputs = tokenizer(batch, padding=True, truncation=False, return_tensors="pt")
input_ids = encoded_inputs["input_ids"].to(model.device)
attention_mask = encoded_inputs["attention_mask"].to(model.device)
with torch.no_grad():
torch.cuda.synchronize()
batch_start_time = time.perf_counter()
output_tokens = model.generate(
input_ids,
attention_mask=attention_mask,
max_new_tokens=num_new_tokens,
do_sample=True,
temperature=temperature,
pad_token_id=tokenizer.pad_token_id,
eos_token_id=tokenizer.eos_token_id
)
torch.cuda.synchronize()
batch_end_time = time.perf_counter()
batch_time = batch_end_time - batch_start_time
total_time += batch_time
total_tokens += output_tokens.numel()
decoded_outputs = tokenizer.batch_decode(output_tokens, skip_special_tokens=True)
for i, response in enumerate(decoded_outputs):
original_dialog = dialogs[start_idx + i]
formatted_response = format_response(original_dialog, response)
responses.append(formatted_response)
yield {
"Time Taken (seconds)": time.time() - start_time,
"Tokens per Second": total_tokens / total_time if total_time > 0 else 0,
"Formatted Responses": f"**Question**: {formatted_response['question']}\n\n**Answer**: {formatted_response['answer']}\n\n---\n\n"
}
progress((batch_idx + 1) / num_batches, desc="Processing batches")
elapsed_time = time.time() - start_time
print(f"Inference completed in {elapsed_time:.2f} seconds.")
# Demo function
def demo(num_new_tokens, temperature, custom_questions_text, kv_bits, progress=gr.Progress()):
custom_questions = custom_questions_text.split("\n")
print("Loading questions...")
dialogs = load_questions("chats_sys_none.json", custom_questions)
print(f"{len(dialogs)} questions loaded. Starting inference...")
result_gen = infer("NousResearch/Meta-Llama-3-8B-Instruct", dialogs, num_new_tokens, temperature, "fp16", kv_bits, progress=progress)
time_taken, tokens_per_second, formatted_responses = None, None, ""
for result in result_gen:
time_taken = result["Time Taken (seconds)"]
tokens_per_second = result["Tokens per Second"]
formatted_responses += result["Formatted Responses"]
yield time_taken, tokens_per_second, formatted_responses
# Load JSON data
with open("chats_sys_none.json", "r") as file:
json_data = json.load(file)
json_data_str = json.dumps(json_data, indent=2)
# Show JSON function
def show_json():
return json_data_str
# Gradio interface
app = gr.Blocks()
with app:
with gr.Tab("LLM Inference Demo"):
num_new_tokens = gr.Slider(label="Number of New Tokens", minimum=128, maximum=1024, step=128, value=512)
temperature = gr.Slider(label="Temperature", minimum=0.0, maximum=1.0, step=0.1, value=0.4)
custom_questions_text = gr.Textbox(label="Custom Questions", placeholder="Type your custom questions here, one per line...", lines=5)
kv_bits = gr.Dropdown(label="KV Bits", choices=["1", "2", "4", "unquantized"], value="1")
time_taken = gr.Number(label="Time Taken (seconds)")
tokens_per_second = gr.Number(label="Tokens per Second")
formatted_responses = gr.Markdown(label="Formatted Responses")
demo_btn = gr.Button("Run Inference")
demo_btn.click(demo, inputs=[num_new_tokens, temperature, custom_questions_text, kv_bits], outputs=[time_taken, tokens_per_second, formatted_responses])
with gr.Tab("Show JSON"):
json_output = gr.HTML("<pre>{}</pre>".format(json_data_str))
json_interface = gr.Interface(fn=show_json, inputs=[], outputs=[json_output], live=False)
json_interface.render()
if __name__ == "__main__":
print("Loading model and tokenizer on startup...")
load_model_and_tokenizer("NousResearch/Meta-Llama-3-8B-Instruct", "fp16", "1")
print("Model and tokenizer loaded. Starting Gradio interface...")
app.queue(default_concurrency_limit=5).launch()
|