inference.py

import os
import sys
import site

try:
    cudnn_path = os.path.join(site.getsitepackages()[0], 'nvidia', 'cudnn', 'lib')
    if os.path.exists(cudnn_path):
        if 'LD_LIBRARY_PATH' in os.environ:
            os.environ['LD_LIBRARY_PATH'] = f"{cudnn_path}:{os.environ['LD_LIBRARY_PATH']}"
        else:
            os.environ['LD_LIBRARY_PATH'] = cudnn_path
        if "RESTARTED" not in os.environ:
            os.environ["RESTARTED"] = "1"
            os.execv(sys.executable, [sys.executable] + sys.argv)
except Exception:
    pass

import onnxruntime as ort

import tiktoken
import numpy as np
import time

# --- Configuration ---
MODEL_PATH = "Apex_1.5_DYNAMIC.onnx"
VOCAB_SIZE = 50304 
enc = tiktoken.get_encoding("gpt2")

# Setup ONNX Session with CUDA
options = ort.SessionOptions()
options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL

print(f"🚀 Loading Dynamic ONNX Model: {MODEL_PATH}...")
providers = [
    ('CUDAExecutionProvider', {
        'device_id': 0,
        'arena_extend_strategy': 'kNextPowerOfTwo',
    }),
    'CPUExecutionProvider'
]

try:
    session = ort.InferenceSession(MODEL_PATH, sess_options=options, providers=providers)
    print(f"✅ Active Provider: {session.get_providers()[0]}")
except Exception as e:
    print(f"❌ Error loading model: {e}")
    sys.exit()

def get_param(prompt, default):
    """Reads input and returns default if empty."""
    val = input(f"{prompt} (Default: {default}): ").strip()
    if not val:
        return default
    return type(default)(val)

def apply_sampling(logits, temperature, top_k, repetition_penalty, history):
    """
    Applies Top-K, Temperature and Repetition Penalty to logits.
    """
    # 1. Repetition Penalty
    if repetition_penalty != 1.0 and len(history) > 0:
        unique_tokens = np.unique(history)
        # Apply penalty: divide positive logits, multiply negative ones
        for token in unique_tokens:
            if token < len(logits):
                if logits[token] > 0:
                    logits[token] /= repetition_penalty
                else:
                    logits[token] *= repetition_penalty

    # 2. Temperature Scaling
    logits = logits / max(temperature, 1e-6)

    # 3. Top-K Sampling
    top_k = min(top_k, len(logits))
    indices_to_remove = logits < np.partition(logits, -top_k)[-top_k]
    logits[indices_to_remove] = -float('Inf')

    # 4. Softmax and Random Choice
    exp_logits = np.exp(logits - np.max(logits))
    probs = exp_logits / np.sum(exp_logits)
    
    return int(np.random.choice(len(logits), p=probs))

def run_chat():
    print("\n" + "="*50)
    print("   APEX 1.5 DYNAMIC ONNX INTERACTIVE CHAT")
    print("="*50 + "\n")

    while True:
        user_input = input("You: ")
        if user_input.lower() in ["exit", "quit", "beenden"]:
            break

        # Prompt Parameters
        temp = get_param("  Temperature", 0.55)
        tk = get_param("  Top-K", 40)
        rp = get_param("  Repetition Penalty", 1.2)
        max_tk = get_param("  Max New Tokens", 500)

        # Tokenize and Setup
        prompt = f"Instruction:\n{user_input}\n\nResponse:\n"
        input_ids = enc.encode(prompt)
        history = list(input_ids)

        print("\nApex 1.5: ", end="", flush=True)
        
        start_time = time.time()
        token_count = 0
        last_printed_len = 0
        full_response_ids = []

        # Generation Loop
        for _ in range(max_tk):
            # Dynamic Input Shape (1, Sequence_Length)
            # We take the last 1024 tokens if it grows too long
            current_ctx = input_ids[-1024:]
            input_array = np.array([current_ctx], dtype=np.int64)

            # Run ONNX Inference
            outputs = session.run(None, {'input': input_array})
            
            # Extract Logits for the last token [Batch, Seq, Vocab]
            # Since it's dynamic, we grab index -1
            logits = outputs[0][0, -1, :VOCAB_SIZE].astype(np.float32)

            # Sampling Logic
            next_token = apply_sampling(logits, temp, tk, rp, history)

            if next_token == enc.eot_token or next_token >= 50257:
                break

            # Update state
            input_ids.append(next_token)
            full_response_ids.append(next_token)
            history.append(next_token)
            token_count += 1

            # Decode and Print
            decoded_text = enc.decode(full_response_ids)
            new_text = decoded_text[last_printed_len:]
            
            # Simple Stop Condition
            if "Instruction:" in new_text:
                break
                
            print(new_text, end="", flush=True)
            last_printed_len = len(decoded_text)

        duration = time.time() - start_time
        tps = token_count / duration if duration > 0 else 0
        
        print(f"\n\n[Speed: {tps:.2f} tokens/s | Time: {duration:.2f}s]")
        print("-" * 40 + "\n")

if __name__ == "__main__":
    run_chat()