File size: 39,577 Bytes
16ffc97 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 |
import os
import gc
import sys
import time
import logging
import traceback
import torch
import warnings
import numpy as np
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation import GenerationConfig
from tqdm import tqdm
from onnxruntime.quantization import quantize_dynamic, QuantType
# Configure logging
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s',
datefmt='%Y-%m-%d %H:%M:%S'
)
logger = logging.getLogger(__name__)
# Suppress unhelpful warnings
warnings.filterwarnings("ignore", category=UserWarning)
class GenerationWrapper(torch.nn.Module):
"""
Wrapper for model export that handles generation properly.
This ensures the model can be correctly used for text generation.
"""
def __init__(self, model):
super().__init__()
self.model = model
self.config = model.config
def forward(self, input_ids, attention_mask=None):
# Return only the logits to avoid complex structures
with torch.no_grad():
try:
# Standard approach for most models
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
use_cache=False,
return_dict=True
)
return outputs.logits
except Exception as e:
logger.warning(f"Standard forward pass failed, trying fallback: {str(e)}")
# Fallback for models with different API
outputs = self.model(input_ids=input_ids, attention_mask=attention_mask)
if hasattr(outputs, 'logits'):
return outputs.logits
elif isinstance(outputs, tuple) and len(outputs) > 0:
return outputs[0] # First element is typically logits
else:
raise ValueError("Could not extract logits from model outputs")
def verify_model_generation(model, tokenizer, device="cpu"):
"""Test model generation capabilities before export"""
model.eval()
# Use a chat-like prompt for better testing
prompt = "User: Hello, how are you today?\nAssistant:"
logger.info("Testing model generation...")
inputs = tokenizer(prompt, return_tensors="pt").to(device)
# Configure generation parameters
gen_config = GenerationConfig(
max_length=100,
do_sample=True,
temperature=0.7,
num_return_sequences=1,
)
try:
# Try generation
with torch.no_grad():
outputs = model.generate(
**inputs,
generation_config=gen_config
)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
logger.info(f"Test generation result: {generated_text}")
if len(generated_text) <= len(prompt):
logger.warning("Generation output is not longer than input prompt!")
return True
except Exception as e:
logger.error(f"Generation test failed: {str(e)}")
return False
def test_onnx_model(onnx_path, tokenizer):
"""Verify the ONNX model can be loaded and run"""
try:
import onnxruntime as ort
logger.info("Testing ONNX model inference...")
session = ort.InferenceSession(onnx_path)
# Get input and output names
input_names = [input.name for input in session.get_inputs()]
output_names = [output.name for output in session.get_outputs()]
# Create test input
prompt = "User: Hello, how are you?\nAssistant:"
inputs = tokenizer(prompt, return_tensors="np")
# Prepare input dict
onnx_inputs = {}
for name in input_names:
if name == "input_ids" and "input_ids" in inputs:
onnx_inputs[name] = inputs["input_ids"]
elif name == "attention_mask" and "attention_mask" in inputs:
onnx_inputs[name] = inputs["attention_mask"]
# Run inference
outputs = session.run(output_names, onnx_inputs)
# Check output shape
logits = outputs[0]
logger.info(f"ONNX model output shape: {logits.shape}")
if logits.shape[0] != 1 or logits.shape[1] != inputs["input_ids"].shape[1]:
logger.warning("Output shape doesn't match expected dimensions!")
# Test next token prediction
next_token_logits = logits[0, -1, :]
next_token_id = np.argmax(next_token_logits)
next_token = tokenizer.decode([next_token_id])
logger.info(f"Next predicted token: '{next_token}'")
return True
except Exception as e:
logger.error(f"ONNX model test failed: {str(e)}")
return False
def post_process_onnx_for_unity(onnx_path):
"""
Post-process ONNX model to be compatible with Unity Sentis
using only core onnx functionality (no onnxsim)
"""
try:
import onnx
logger.info("Post-processing ONNX model for Unity compatibility...")
# First, create a backup of the original model
backup_path = onnx_path.replace(".onnx", "_original.onnx")
import shutil
shutil.copy(onnx_path, backup_path)
logger.info(f"Original model backed up to {backup_path}")
# Load the model
model = onnx.load(onnx_path)
# Basic model checks and optimizations
try:
# Check model validity
onnx.checker.check_model(model)
logger.info("β Model structure validated successfully")
# Apply shape inference
inferred_model = onnx.shape_inference.infer_shapes(model)
onnx.save(inferred_model, onnx_path)
logger.info("β Applied shape inference")
except Exception as e:
logger.warning(f"Model validation/optimization error (continuing): {str(e)}")
return True
except Exception as e:
logger.warning(f"ONNX post-processing error (skipping): {str(e)}")
return False
def is_architecture_compatible(model_id):
"""
Check if the model architecture is expected to be compatible with ONNX opset 11
"""
model_id_lower = model_id.lower()
# Models known to work with opset 11
compatible_architectures = [
"gpt2", "distilgpt2", "opt-125m", "opt-350m",
"pythia-70m", "pythia-160m", "rwkv", "gpt-neo"
]
# Models likely requiring higher opsets (usually 14+)
incompatible_architectures = [
"llama", "mistral", "mixtral", "tinyllama", "phi-2",
"gemma", "falcon", "bloom"
]
# Check for compatibility
for arch in compatible_architectures:
if arch in model_id_lower:
return True, 11
# Check for known incompatible architectures
for arch in incompatible_architectures:
if arch in model_id_lower:
return False, 14
# For phi-1 models, use opset 14 but mark as potentially compatible
if "phi-1" in model_id_lower:
return True, 14
# Default to opset 14 for unknown architectures
return False, 14
def setup_chat_template(model_id, tokenizer):
"""
Setup appropriate chat template based on model architecture
"""
model_id_lower = model_id.lower()
# Try to setup chat template if it doesn't have one
try:
if not hasattr(tokenizer, "chat_template") or tokenizer.chat_template is None:
logger.info("Setting up chat template for improved conversations...")
# Determine chat template based on model
if "gpt2" in model_id_lower or "pythia" in model_id_lower or "opt" in model_id_lower:
# Simple template for base models
chat_template = "{% for message in messages %}\n{% if message['role'] == 'user' %}\nHuman: {{ message['content'] }}\n{% elif message['role'] == 'assistant' %}\nAI: {{ message['content'] }}\n{% endif %}\n{% endfor %}\n{% if add_generation_prompt %}\nAI: {% endif %}"
tokenizer.chat_template = chat_template
logger.info("β Added simple Human/AI chat template")
elif "phi" in model_id_lower:
# Microsoft Phi models template
chat_template = "{% for message in messages %}\n{% if message['role'] == 'user' %}\nHuman: {{ message['content'] }}\n{% elif message['role'] == 'assistant' %}\nAssistant: {{ message['content'] }}\n{% endif %}\n{% endfor %}\n{% if add_generation_prompt %}\nAssistant: {% endif %}"
tokenizer.chat_template = chat_template
logger.info("β Added Phi-style Human/Assistant chat template")
elif "rwkv" in model_id_lower:
# RWKV template
chat_template = "{% for message in messages %}\n{% if message['role'] == 'user' %}\nUser: {{ message['content'] }}\n{% elif message['role'] == 'assistant' %}\nBot: {{ message['content'] }}\n{% endif %}\n{% endfor %}\n{% if add_generation_prompt %}\nBot: {% endif %}"
tokenizer.chat_template = chat_template
logger.info("β Added RWKV-style User/Bot chat template")
except Exception as e:
logger.warning(f"Couldn't setup chat template: {str(e)}")
logger.info("Chat template setup will need to be handled in Unity")
def convert_model(model_id, output_dir="./onnx_models", seq_length=32, quantize=True, force_opset=None):
"""
Convert a model to ONNX format with focus on Unity compatibility.
Args:
model_id: HuggingFace model ID or path
output_dir: Directory to save the model
seq_length: Input sequence length for export
quantize: Whether to quantize the model to INT8
force_opset: Force a specific ONNX opset version
Returns:
bool: Success status
"""
start_time = time.time()
# Check model architecture for compatibility
is_compatible, recommended_opset = is_architecture_compatible(model_id)
# Use forced opset if provided, otherwise use recommended
opset_version = force_opset if force_opset is not None else recommended_opset
# Warn if using a model that might not be compatible with Unity
if not is_compatible and opset_version < 14:
logger.warning(f"β Model {model_id} may not be compatible with opset {opset_version}")
logger.warning(f"β Recommended opset for this model: {recommended_opset}")
logger.warning(f"β You can force a higher opset with --opset {recommended_opset}")
logger.info(f"\n{'=' * 60}")
logger.info(f"Converting {model_id} to ONNX for Unity (opset {opset_version})")
logger.info(f"{'=' * 60}")
# Create output directory
model_name = model_id.split("/")[-1]
model_dir = os.path.join(output_dir, model_name)
os.makedirs(model_dir, exist_ok=True)
try:
# Step 1: Load tokenizer
logger.info("Step 1/7: Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(model_id)
if tokenizer.pad_token is None and hasattr(tokenizer, 'eos_token'):
logger.info("Adding pad_token = eos_token")
tokenizer.pad_token = tokenizer.eos_token
# Setup chat template for better conversation formatting
setup_chat_template(model_id, tokenizer)
# Save tokenizer
tokenizer.save_pretrained(model_dir)
logger.info(f"β Tokenizer saved to {model_dir}")
# Step 2: Load model with reliability optimizations
logger.info("Step 2/7: Loading model...")
# Clean memory
gc.collect()
torch.cuda.empty_cache() if torch.cuda.is_available() else None
# Determine device
device = "cuda" if torch.cuda.is_available() else "cpu"
# Load model with full precision
try:
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float32, # Use full precision for reliability
low_cpu_mem_usage=True, # Reduce memory usage
device_map=device # Use CUDA if available
)
except Exception as e:
logger.warning(f"Standard loading failed, trying with 'trust_remote_code=True': {str(e)}")
# Some models (like RWKV) need trust_remote_code
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float32,
low_cpu_mem_usage=True,
device_map=device,
trust_remote_code=True
)
# Save config
model.config.save_pretrained(model_dir)
logger.info(f"β Model config saved to {model_dir}")
# Step 3: Verify model can generate chat responses
logger.info("Step 3/7: Validating chat capabilities...")
if not verify_model_generation(model, tokenizer, device):
logger.warning("β Model chat test didn't complete successfully")
logger.info("Continuing with export anyway...")
# Step 4: Export to ONNX
logger.info(f"Step 4/7: Exporting to ONNX format with opset {opset_version}...")
# Wrap model with generation-optimized interface
wrapped_model = GenerationWrapper(model)
wrapped_model.eval()
# Clean memory again
gc.collect()
torch.cuda.empty_cache() if torch.cuda.is_available() else None
# Export to ONNX with appropriate opset version
onnx_path = os.path.join(model_dir, "model.onnx")
# Create minimal input
batch_size = 1
dummy_input = torch.ones(batch_size, seq_length, dtype=torch.long)
attention_mask = torch.ones(batch_size, seq_length, dtype=torch.long)
# Move tensors to correct device
dummy_input = dummy_input.to(device)
attention_mask = attention_mask.to(device)
# Export to ONNX with required opset
with torch.no_grad():
torch.onnx.export(
wrapped_model, # Wrapped model
(dummy_input, attention_mask), # Input tensors
onnx_path, # Output path
export_params=True, # Store weights
opset_version=opset_version, # Required opset version
do_constant_folding=True, # Optimize constants
input_names=['input_ids', 'attention_mask'], # Input names
output_names=['logits'], # Output name
dynamic_axes={ # Dynamic dimensions
'input_ids': {0: 'batch_size', 1: 'sequence'},
'attention_mask': {0: 'batch_size', 1: 'sequence'},
'logits': {0: 'batch_size', 1: 'sequence'}
}
)
# Clean up to save memory
del model
del wrapped_model
gc.collect()
torch.cuda.empty_cache() if torch.cuda.is_available() else None
# Verify export success
if os.path.exists(onnx_path):
size_mb = os.path.getsize(onnx_path) / (1024 * 1024)
logger.info(f"β ONNX model saved to {onnx_path}")
logger.info(f"β Original size: {size_mb:.2f} MB")
# Step 5: Post-process the ONNX model for better Unity compatibility
logger.info("Step 5/7: Post-processing ONNX model for Unity compatibility...")
# Try to post-process model for Unity
try:
post_process_onnx_for_unity(onnx_path)
except Exception as e:
logger.warning(f"Post-processing failed (non-critical): {str(e)}")
# Test ONNX model
test_onnx_model(onnx_path, tokenizer)
# Step 6: Quantize the model (optional)
if quantize:
logger.info("Step 6/7: Applying INT8 quantization...")
quant_path = onnx_path.replace(".onnx", "_quantized.onnx")
try:
with tqdm(total=100, desc="Quantizing") as pbar:
# Update progress callback
def update_progress(x):
pbar.update(1)
# Apply quantization
quantize_dynamic(
model_input=onnx_path,
model_output=quant_path,
per_channel=False,
reduce_range=False,
weight_type=QuantType.QInt8,
optimize_model=True,
use_external_data_format=False
)
pbar.update(100) # Ensure progress reaches 100%
if os.path.exists(quant_path):
quant_size = os.path.getsize(quant_path) / (1024 * 1024)
logger.info(f"β Quantized size: {quant_size:.2f} MB")
logger.info(f"β Size reduction: {(1 - quant_size/size_mb) * 100:.1f}%")
# Test the quantized model
test_onnx_model(quant_path, tokenizer)
# Rename original as backup
backup_path = onnx_path.replace(".onnx", "_fp32.onnx")
os.rename(onnx_path, backup_path)
# Replace original with quantized
os.rename(quant_path, onnx_path)
logger.info("β Original model preserved as *_fp32.onnx")
logger.info("β Replaced original with quantized version")
else:
logger.warning("β Quantized file not created, using original")
except Exception as e:
logger.error(f"β Quantization error: {str(e)}")
logger.info("β Using original model without quantization")
else:
logger.info("Step 6/7: Skipping quantization as requested")
# Step 7: Generate Unity integration examples
logger.info("Step 7/7: Generating Unity integration examples...")
# Create a Unity integration example
unity_example_path = os.path.join(model_dir, "unity_integration.cs")
with open(unity_example_path, 'w') as f:
f.write("""
using UnityEngine;
using Unity.Sentis;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
public class ONNXChatbot : MonoBehaviour
{
[SerializeField] private ModelAsset modelAsset;
[SerializeField] private TextAsset tokenizerVocabJson;
[SerializeField] private int maxTokens = 50;
[SerializeField] private float temperature = 0.7f;
private IWorker worker;
private Dictionary<string, Tensor> inputs;
private SimpleTokenizer tokenizer;
private bool isGenerating = false;
void Start()
{
// Initialize the model
var model = ModelLoader.Load(modelAsset);
worker = WorkerFactory.CreateWorker(WorkerFactory.Type.ComputePrecompiled, model);
// Initialize tokenizer
tokenizer = new SimpleTokenizer(tokenizerVocabJson.text);
// Prepare for inference
inputs = new Dictionary<string, Tensor>();
Debug.Log("Model and tokenizer initialized successfully.");
}
public async Task<string> GenerateResponseAsync(string userMessage)
{
if (isGenerating)
{
Debug.LogWarning("Already generating a response. Please wait.");
return "Already generating a response. Please wait.";
}
isGenerating = true;
try
{
// Format prompt with chat template
string prompt = FormatChatPrompt(userMessage);
Debug.Log($"Formatted prompt: {prompt}");
// Tokenize input
var tokenIds = tokenizer.Encode(prompt);
Debug.Log($"Encoded to {tokenIds.Length} tokens");
if (tokenIds.Length > 0)
{
// Generate response token by token
StringBuilder responseBuilder = new StringBuilder();
List<int> currentIds = tokenIds.ToList();
for (int i = 0; i < maxTokens; i++)
{
// Make sure we don't exceed the model's context window
if (currentIds.Count > 1024)
{
// If too long, keep only the last 1024 tokens
currentIds = currentIds.Skip(currentIds.Count - 1024).Take(1024).ToList();
}
// Create tensors for current sequence
using (var inputIdsTensor = new TensorInt(new TensorShape(1, currentIds.Count), currentIds.ToArray()))
using (var attentionMaskTensor = new TensorInt(new TensorShape(1, currentIds.Count), Enumerable.Repeat(1, currentIds.Count).ToArray()))
{
// Run inference
inputs.Clear();
inputs["input_ids"] = inputIdsTensor;
inputs["attention_mask"] = attentionMaskTensor;
worker.Execute(inputs);
var logits = worker.PeekOutput() as TensorFloat;
// Get next token prediction
int nextToken = SampleNextToken(logits, currentIds, temperature);
// If we hit the end token or a newline after content, stop
if (nextToken == tokenizer.EosToken ||
(i > 0 && nextToken == tokenizer.NewlineToken))
{
break;
}
// Add token to current sequence for next iteration
currentIds.Add(nextToken);
// Decode the latest token
string newToken = tokenizer.Decode(new[] { nextToken });
responseBuilder.Append(newToken);
// For smoother output, yield every few tokens
if (i % 5 == 0)
{
await Task.Delay(1);
}
}
}
// Return the full response, without the prompt
string fullResponse = responseBuilder.ToString();
return CleanResponse(fullResponse);
}
else
{
Debug.LogError("Tokenization failed: empty token list");
return "Sorry, I couldn't process that input.";
}
}
catch (System.Exception ex)
{
Debug.LogError($"Generation error: {ex.Message}\\n{ex.StackTrace}");
return "Sorry, an error occurred while generating a response.";
}
finally
{
isGenerating = false;
}
}
private string FormatChatPrompt(string userMessage)
{
// You may need to adjust this template based on your specific model
return $"User: {userMessage}\\nAssistant:";
}
private string CleanResponse(string response)
{
// Extract only the Assistant's response
int assistantPrefix = response.IndexOf("Assistant:");
if (assistantPrefix >= 0)
{
response = response.Substring(assistantPrefix + "Assistant:".Length).Trim();
}
// Stop at any "User:" marker if present
int nextUser = response.IndexOf("User:");
if (nextUser >= 0)
{
response = response.Substring(0, nextUser).Trim();
}
return response;
}
private int SampleNextToken(TensorFloat logits, List<int> currentInputs, float temp)
{
// Get logits for the last position
int lastPos = currentInputs.Count - 1;
int vocabSize = logits.shape.channels;
// Prepare array for logits
float[] lastLogits = new float[vocabSize];
// Extract logits for the last token position
for (int i = 0; i < vocabSize; i++)
{
lastLogits[i] = logits[0, lastPos, i];
}
// Simple temperature-based sampling
if (temp <= 0.0f)
{
// Greedy sampling (argmax)
int maxIndex = 0;
float maxValue = lastLogits[0];
for (int i = 1; i < vocabSize; i++)
{
if (lastLogits[i] > maxValue)
{
maxValue = lastLogits[i];
maxIndex = i;
}
}
return maxIndex;
}
else
{
// Temperature sampling
// Apply temperature
for (int i = 0; i < vocabSize; i++)
{
lastLogits[i] /= temp;
}
// Softmax
float maxLogit = lastLogits.Max();
float sum = 0.0f;
for (int i = 0; i < vocabSize; i++)
{
lastLogits[i] = Mathf.Exp(lastLogits[i] - maxLogit);
sum += lastLogits[i];
}
for (int i = 0; i < vocabSize; i++)
{
lastLogits[i] /= sum;
}
// Sample from distribution
float random = Random.value;
float cumulativeProb = 0.0f;
for (int i = 0; i < vocabSize; i++)
{
cumulativeProb += lastLogits[i];
if (random < cumulativeProb)
{
return i;
}
}
// Fallback to last token if sampling fails
return vocabSize - 1;
}
}
void OnDestroy()
{
worker?.Dispose();
}
}
// Simple tokenizer implementation for Unity
public class SimpleTokenizer
{
private Dictionary<string, int> vocab;
private Dictionary<int, string> reversedVocab;
public int PadToken { get; private set; }
public int EosToken { get; private set; }
public int BosToken { get; private set; }
public int NewlineToken { get; private set; }
public SimpleTokenizer(string vocabJson)
{
// Parse the vocabulary from JSON
vocab = new Dictionary<string, int>();
// Simple JSON parsing (you'll need a proper JSON parser in production)
string[] entries = vocabJson.Split(new[] { '\\n', '{', '}', '\"', ':', ',' },
System.StringSplitOptions.RemoveEmptyEntries);
for (int i = 0; i < entries.Length - 1; i += 2)
{
string token = entries[i].Trim();
if (int.TryParse(entries[i + 1].Trim(), out int id))
{
vocab[token] = id;
}
}
// Create reversed vocabulary for decoding
reversedVocab = vocab.ToDictionary(kv => kv.Value, kv => kv.Key);
// Find special tokens
SetSpecialTokens();
Debug.Log($"Tokenizer initialized with {vocab.Count} tokens");
}
private void SetSpecialTokens()
{
// Try to find standard special tokens
PadToken = FindToken(new[] { "<pad>", "[PAD]", "<|endoftext|>" });
EosToken = FindToken(new[] { "</s>", "<|endoftext|>", "[EOS]", "<eos>" });
BosToken = FindToken(new[] { "<s>", "<|startoftext|>", "[BOS]", "<bos>" });
// Find newline token
foreach (var entry in vocab)
{
if (entry.Key == "\\n" || entry.Key == "<\\n>" || entry.Key == "\\n")
{
NewlineToken = entry.Value;
break;
}
}
Debug.Log($"Special tokens - PAD: {PadToken}, EOS: {EosToken}, BOS: {BosToken}, NEWLINE: {NewlineToken}");
}
private int FindToken(string[] candidates)
{
foreach (var candidate in candidates)
{
if (vocab.TryGetValue(candidate, out int id))
{
return id;
}
}
// Return -1 if not found
return -1;
}
public int[] Encode(string text)
{
// Simple character-level tokenization
// In production, use a proper BPE/WordPiece tokenizer implementation
List<int> tokens = new List<int>();
StringBuilder currentToken = new StringBuilder();
// Add BOS token if available
if (BosToken != -1)
{
tokens.Add(BosToken);
}
// Very simple tokenization - in production, this would implement
// the specific tokenization algorithm for your model
foreach (char c in text)
{
currentToken.Append(c);
string current = currentToken.ToString();
if (vocab.TryGetValue(current, out int id))
{
tokens.Add(id);
currentToken.Clear();
}
else if (currentToken.Length > 10)
{
// If token is too long, add unknown token and reset
tokens.Add(vocab.ContainsKey("<unk>") ? vocab["<unk>"] : 0);
currentToken.Clear();
currentToken.Append(c);
}
}
// Handle any remaining text
if (currentToken.Length > 0)
{
tokens.Add(vocab.ContainsKey("<unk>") ? vocab["<unk>"] : 0);
}
return tokens.ToArray();
}
public string Decode(int[] ids)
{
StringBuilder result = new StringBuilder();
foreach (int id in ids)
{
if (reversedVocab.TryGetValue(id, out string token))
{
// Some tokenizers use special prefixes like "Δ " for spaces
string processedToken = token
.Replace("Δ ", " ")
.Replace("Δ", "\n")
.Replace("β", " ");
result.Append(processedToken);
}
}
return result.ToString();
}
}
""")
# Calculate elapsed time
end_time = time.time()
duration = end_time - start_time
logger.info(f"β Conversion completed in {duration:.2f} seconds")
logger.info(f"β Final model size: {os.path.getsize(onnx_path) / (1024 * 1024):.2f} MB")
# Create a Python example usage file
example_path = os.path.join(model_dir, "example_usage.py")
with open(example_path, 'w') as f:
f.write("""
import onnxruntime as ort
from transformers import AutoTokenizer
import numpy as np
# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("./") # Path to model directory
session = ort.InferenceSession("./model.onnx")
def generate_response(user_message, max_length=50):
# Format as a chat message
prompt = f"User: {user_message}\\nAssistant:"
inputs = tokenizer(prompt, return_tensors="np")
input_ids = inputs["input_ids"]
attention_mask = inputs["attention_mask"]
# Simple auto-regressive generation loop
for _ in range(max_length):
# Run inference for a single step
outputs = session.run(
["logits"],
{
"input_ids": input_ids,
"attention_mask": attention_mask
}
)
# Get next token prediction from logits
logits = outputs[0]
next_token_logits = logits[0, -1, :]
# Apply temperature sampling
temperature = 0.7
next_token_logits = next_token_logits / temperature
# Apply softmax to get probabilities
exp_logits = np.exp(next_token_logits - np.max(next_token_logits))
probs = exp_logits / np.sum(exp_logits)
# Sample from the distribution
next_token_id = np.random.choice(probs.shape[0], p=probs)
# Stop if we hit the end of sequence token
if next_token_id == tokenizer.eos_token_id:
break
# Append new token to the input_ids
input_ids = np.concatenate([input_ids, [[next_token_id]]], axis=1)
attention_mask = np.concatenate([attention_mask, [[1]]], axis=1)
# Decode the entire response
response = tokenizer.decode(input_ids[0], skip_special_tokens=True)
# Extract only the assistant's response
if "Assistant:" in response:
response = response.split("Assistant:")[-1].strip()
return response
# Example usage
while True:
user_input = input("You: ")
if user_input.lower() in ['exit', 'quit']:
break
response = generate_response(user_input)
print(f"Assistant: {response}")
""")
logger.info(f"β Example usage saved to {example_path}")
logger.info(f"β Unity integration example saved to {unity_example_path}")
return True
else:
logger.error(f"Γ ONNX file not created at {onnx_path}")
return False
except Exception as e:
logger.error(f"Γ Error converting model: {str(e)}")
logger.error(traceback.format_exc())
return False
if __name__ == "__main__":
# Parse command line arguments
parser_available = False
try:
import argparse
parser = argparse.ArgumentParser(description="Convert HuggingFace models to ONNX for Unity")
parser.add_argument("model_id", type=str, help="HuggingFace model ID or path")
parser.add_argument("--output_dir", "-o", type=str, default="./onnx_models",
help="Output directory for the converted model")
parser.add_argument("--seq_length", "-s", type=int, default=32,
help="Sequence length for model export")
parser.add_argument("--no_quantize", action="store_true",
help="Skip INT8 quantization step")
parser.add_argument("--opset", "-op", type=int, default=None,
help="Force a specific ONNX opset version")
args = parser.parse_args()
parser_available = True
model_id = args.model_id
output_dir = args.output_dir
seq_length = args.seq_length
quantize = not args.no_quantize
force_opset = args.opset
except (ImportError, NameError):
# Fallback if argparse is not available
parser_available = False
if not parser_available:
if len(sys.argv) < 2:
print("Usage: python unity_compatible_converter.py MODEL_ID [OUTPUT_DIR] [SEQ_LENGTH] [--no-quantize] [--opset]")
print("Example: python unity_compatible_converter.py distilgpt2 ./onnx_models 32")
print("\nRecommended chat models for Unity:")
print(" - distilgpt2 (smallest, opset 11)")
print(" - EleutherAI/pythia-70m (better quality, opset 11)")
print(" - microsoft/phi-1 (high quality, opset 14)")
print(" - TinyLlama/TinyLlama-1.1B-Chat-v1.0 (chat-tuned, opset 14)")
sys.exit(1)
model_id = sys.argv[1]
output_dir = sys.argv[2] if len(sys.argv) > 2 else "./onnx_models"
seq_length = int(sys.argv[3]) if len(sys.argv) > 3 else 32
quantize = "--no-quantize" not in sys.argv and "--no_quantize" not in sys.argv
force_opset = None
# Check for opset flag
for i, arg in enumerate(sys.argv):
if arg == "--opset" and i + 1 < len(sys.argv):
force_opset = int(sys.argv[i + 1])
# Check model architecture for automatic opset recommendation
is_compatible, recommended_opset = is_architecture_compatible(model_id)
# Print header
logger.info("\nUNITY-COMPATIBLE ONNX CONVERTER")
logger.info("===============================")
logger.info(f"Model: {model_id}")
logger.info(f"Output directory: {output_dir}")
logger.info(f"Sequence length: {seq_length}")
if force_opset is not None:
logger.info(f"ONNX opset version: {force_opset} (forced)")
else:
logger.info(f"Recommended ONNX opset: {recommended_opset}")
logger.info(f"Architecture compatible with opset 11: {'Yes' if is_compatible else 'No'}")
logger.info(f"Quantization: {'Enabled' if quantize else 'Disabled'}")
# Create output directory
os.makedirs(output_dir, exist_ok=True)
# Convert the model
success = convert_model(model_id, output_dir, seq_length, quantize, force_opset)
if success:
logger.info("\n" + "=" * 60)
logger.info("CONVERSION SUCCESSFUL")
logger.info("=" * 60)
logger.info(f"Model: {model_id}")
logger.info(f"Output directory: {os.path.abspath(output_dir)}")
logger.info("The model is ready for Unity integration!")
logger.info("\nNext steps:")
logger.info("1. Import the ONNX model into Unity using the Sentis package")
logger.info("2. Use the unity_integration.cs file as a starting point")
logger.info("3. For tokenization in Unity, implement the SimpleTokenizer class")
else:
logger.info("\n" + "=" * 60)
logger.info("CONVERSION FAILED")
logger.info("=" * 60)
logger.info("Please try one of the recommended models that work well with Unity:")
if is_compatible:
logger.info("Compatible with Unity (opset 11):")
logger.info(" - distilgpt2")
logger.info(" - EleutherAI/pythia-70m")
logger.info("Advanced models (require opset 14):")
logger.info(" - microsoft/phi-1 --opset 14")
logger.info(" - TinyLlama/TinyLlama-1.1B-Chat-v1.0 --opset 14") |