File size: 50,633 Bytes
8a520fb 501d5b0 19f193d 501d5b0 8a520fb 4984ed2 1eb637f 8a520fb 0ed0bda 8a520fb 5819251 8a520fb 73df3b2 19f193d 73df3b2 8a520fb dd87592 73df3b2 1eb637f 73df3b2 0ed0bda 1eb637f 0ed0bda 1eb637f 0ed0bda 1eb637f 0ed0bda 1eb637f 0ed0bda 73df3b2 0ed0bda 73df3b2 0ed0bda 73df3b2 1eb637f 73df3b2 972a07c 8257213 972a07c 73df3b2 972a07c 73df3b2 0ed0bda 73df3b2 0ed0bda 73df3b2 0ed0bda 73df3b2 0ed0bda 73df3b2 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 972a07c 8257213 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 8a520fb 19f193d 0ed0bda 19f193d 0ed0bda 19f193d 0ed0bda 19f193d f3a8698 19f193d f3a8698 19f193d f3a8698 19f193d 8a520fb cc80389 8a520fb cc80389 8a520fb cc80389 8a520fb cc80389 8a520fb cc80389 8a520fb cc80389 dd87592 cc80389 dd87592 cc80389 8a520fb 0ed0bda 8a520fb d3e88f3 8a520fb d3e88f3 8a520fb 5819251 8a520fb 5819251 8a520fb d3e88f3 8a520fb d3e88f3 8a520fb 501d5b0 8a520fb d3e88f3 501d5b0 19f193d 501d5b0 1a079db 501d5b0 1a079db 972a07c 501d5b0 19f193d 501d5b0 19f193d d3e88f3 8a520fb 501d5b0 8a520fb cc80389 8a520fb 4d4fb80 |
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 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 |
import os
import json
from datetime import datetime
import asyncio
import aiohttp
from typing import Dict, List, Optional
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel, HttpUrl
import uvicorn
from git_clone import clone_repository
# ===== CONFIG =====
class Settings:
# Server URLs and Ports
CONTROLLER_HOST = "0.0.0.0" # Listen on all interfaces
CONTROLLER_PORT = 8000
# This should be the actual IP or hostname where controller is accessible
CONTROLLER_BASE_URL = os.getenv("CONTROLLER_BASE_URL", "http://192.168.1.100:8000")
# List of tensor server URLs - should be actual IP addresses or hostnames
TENSOR_SERVER_URLS = [
url for url in os.getenv("TENSOR_SERVER_URLS", "").split(",") if url
] or [
"https://fred808-ilob.hf.space",
"https://fred808-tserv.hf.space",
"https://fred808-tserve2.hf.space",
]
AGGREGATOR_URL = os.getenv("AGGREGATOR_URL", "http://192.168.1.104:8002")
# Model settings
MODEL_REPO = "https://huggingface.co/inference-net/Schematron-8B"
# Server settings
TENSOR_SERVER_TIMEOUT = 30 # seconds
MAX_ERROR_THRESHOLD = 5 # maximum number of errors
SERVER_TIMEOUT = 60 # seconds before marking as error
MONITORING_INTERVAL = 15 # seconds between health checks
# Dynamic distribution settings
@classmethod
def get_optimal_chunk_size(cls, total_params: int, num_servers: int) -> int:
"""Calculate optimal chunk size based on number of servers"""
# Aim for 2-3 chunks per server for better parallelism
target_chunks = num_servers * 2
return max(1, total_params // target_chunks)
@classmethod
def get_min_servers_required(cls) -> int:
"""Dynamically calculate minimum servers needed based on registered servers"""
return max(2, len(cls.TENSOR_SERVER_URLS) // 3) # At least 1/3 of registered servers
@classmethod
def get_min_replica_count(cls, num_servers: int) -> int:
"""Calculate minimum replicas based on server count"""
return max(2, num_servers // 4) # At least 25% of servers should have each chunk
# Tokenizer settings
MAX_SEQUENCE_LENGTH = 2048
VOCAB_SIZE = 50257
@classmethod
def from_env(cls):
"""Load settings from environment variables"""
cls.CONTROLLER_HOST = os.getenv("CONTROLLER_HOST", cls.CONTROLLER_HOST)
cls.CONTROLLER_PORT = int(os.getenv("CONTROLLER_PORT", cls.CONTROLLER_PORT))
cls.CONTROLLER_BASE_URL = os.getenv("CONTROLLER_BASE_URL", cls.CONTROLLER_BASE_URL)
# Load tensor server URLs from environment
tensor_urls = os.getenv("TENSOR_SERVER_URLS")
if tensor_urls:
cls.TENSOR_SERVER_URLS = tensor_urls.split(",")
cls.AGGREGATOR_HOST = os.getenv("AGGREGATOR_HOST", cls.AGGREGATOR_HOST)
cls.AGGREGATOR_PORT = int(os.getenv("AGGREGATOR_PORT", cls.AGGREGATOR_PORT))
cls.AGGREGATOR_URL = os.getenv("AGGREGATOR_URL",
f"http://{cls.AGGREGATOR_HOST}:{cls.AGGREGATOR_PORT}")
return cls
# ===== State Models =====
class ServerMetrics(BaseModel):
"""Metrics for tensor server performance and load"""
cpu_usage: float = 0.0
memory_usage: float = 0.0
gpu_usage: Optional[float] = None
active_requests: int = 0
total_requests: int = 0
average_response_time: float = 0.0
last_error: Optional[str] = None
error_count: int = 0
class TensorServer(BaseModel):
"""Represents a registered tensor server"""
url: HttpUrl
status: str = "initializing" # initializing, ready, busy, error, degraded
last_heartbeat: datetime = datetime.now()
model_chunks: List[int] = [] # List of chunk IDs assigned to this server
metrics: ServerMetrics = ServerMetrics()
version: str = "1.0.0"
capabilities: Dict[str, bool] = {
"gpu_available": False,
"quantization_support": False,
"tensor_parallelism": False
}
class ModelChunk(BaseModel):
"""Represents a chunk of the model to be sent to a tensor server"""
chunk_id: int
files: List[str] # files included in this chunk
config: Dict # configuration for this chunk
size_bytes: int = 0
server_assignments: List[str] = [] # URLs of servers holding this chunk
status: str = "unassigned" # unassigned, assigned, loaded, error
metrics: Dict[str, float] = {
"load_time": 0.0,
"memory_usage": 0.0,
"average_inference_time": 0.0
}
# ===== FastAPI App =====
app = FastAPI(
title="Florence-2 Model Controller",
description="Controls model distribution across tensor servers",
version="1.0.0"
)
# ===== Global State =====
class ControllerState:
def __init__(self):
self.model_files: Dict[str, str] = {} # Mapping of filename to file path
self.model_config: Dict = {} # Model configuration
self.tensor_servers: Dict[str, TensorServer] = {}
self.model_chunks: Dict[int, ModelChunk] = {}
self.is_model_loaded = False
self.model_path: str = "" # Base path where model files are stored
self.chunks_dir: str = "" # Directory containing chunk files
self.operation_results: Dict[str, Dict] = {} # Track operation results from tensor servers
self.pending_operations: Dict[str, asyncio.Task] = {} # Track ongoing operations
state = ControllerState()
# ===== Helper Functions =====
async def split_model_weights():
"""Split model files into chunks based on available servers without loading into memory"""
try:
import os
import math
import shutil
from pathlib import Path
# Find model weight files (safetensors or pytorch)
weight_files = [f for f in state.model_files.values() if f.endswith(('.safetensors', '.bin'))]
if not weight_files:
raise Exception("No model weight files found")
# The current splitting logic only supports splitting a single file.
# If there are multiple files, we assume they are sharded and need a different approach.
# For now, we will select the largest file to split, or the first one if all are small.
model_file = max(weight_files, key=os.path.getsize) if len(weight_files) > 1 else weight_files[0]
if len(weight_files) > 1:
print(f"[WARN] Found multiple weight files. Selecting the largest one for splitting: {model_file}")
else:
print(f"[INFO] Found model weight file: {model_file}")
# Get file size and calculate chunks
try:
with open(model_file, 'rb') as f:
# Get actual file size by seeking to end
f.seek(0, 2) # Seek to end
file_size = f.tell() # Get position (total size)
f.seek(0) # Reset to beginning
# Read first few bytes to verify file isn't corrupted
header = f.read(8)
if len(header) == 0:
raise ValueError(f"File is empty: {model_file}")
except Exception as e:
raise Exception(f"Failed to read model file {model_file}: {str(e)}")
# Verify file size is reasonable
if file_size < 1024: # Less than 1KB
raise ValueError(f"Model file suspiciously small ({file_size} bytes). Possible corruption or incomplete download.")
num_servers = len(state.tensor_servers) or len(Settings.TENSOR_SERVER_URLS)
num_chunks = num_servers # One chunk per server initially
chunk_size = math.ceil(file_size / num_chunks)
# Format sizes for display
def format_size(size_bytes):
if size_bytes >= 1024*1024*1024: # GB
return f"{size_bytes / (1024*1024*1024):.2f} GB ({size_bytes:,} bytes)"
elif size_bytes >= 1024*1024: # MB
return f"{size_bytes / (1024*1024):.2f} MB ({size_bytes:,} bytes)"
elif size_bytes >= 1024: # KB
return f"{size_bytes / 1024:.2f} KB ({size_bytes:,} bytes)"
else:
return f"{size_bytes:,} bytes"
print(f"[INFO] Model file size: {format_size(file_size)}")
print(f"[INFO] Creating {num_chunks} chunks of approximately {format_size(chunk_size)} each")
# Use the chunks directory from state
os.makedirs(state.chunks_dir, exist_ok=True)
# Split the file into chunks
with open(model_file, 'rb') as f:
chunk_sizes = [] # Track actual chunk sizes
for chunk_id in range(num_chunks):
chunk_path = os.path.join(state.chunks_dir, f"chunk_{chunk_id}.bin")
# Calculate chunk boundaries
start_pos = chunk_id * chunk_size
remaining = file_size - start_pos
current_chunk_size = min(chunk_size, remaining)
if current_chunk_size <= 0:
break
# Read and write chunk
try:
f.seek(start_pos)
chunk_data = f.read(current_chunk_size)
actual_chunk_size = len(chunk_data)
if actual_chunk_size != current_chunk_size:
print(f"[WARN] Chunk {chunk_id} size mismatch. Expected: {current_chunk_size}, Got: {actual_chunk_size}")
with open(chunk_path, 'wb') as chunk_file:
chunk_file.write(chunk_data)
chunk_sizes.append(actual_chunk_size)
print(f"[DEBUG] Chunk {chunk_id} data: First few bytes: {chunk_data[:20].hex()}")
except Exception as e:
raise Exception(f"Failed to process chunk {chunk_id} at offset {start_pos}: {str(e)}")
# Create chunk metadata
# Assign vocab_offset based on cumulative sizes of earlier chunks
# so that chunks map to disjoint vocab ranges for aggregation.
cumulative = 0
for cid, c in state.model_chunks.items():
try:
cumulative += int(c.config.get('shard_dim', c.config.get('size_bytes', 1)))
except Exception:
cumulative += 1
cfg = {
"start_offset": start_pos,
"size_bytes": current_chunk_size,
"is_last_chunk": chunk_id == num_chunks - 1,
"total_chunks": num_chunks,
"original_file": os.path.basename(model_file),
# minimal shard mapping; users should adjust shard_dim to real local vocab size
"vocab_offset": cumulative,
# shard_dim should reflect how many vocab ids this chunk covers.
# Default to 1 when unknown; prefer explicitly setting this in chunk metadata.
"shard_dim": int(cfg.get('shard_dim', 1)) if isinstance(cfg := {} , dict) else 1
}
state.model_chunks[chunk_id] = ModelChunk(
chunk_id=chunk_id,
files=[f"chunk_{chunk_id}.bin"],
config=cfg,
size_bytes=current_chunk_size,
status="ready"
)
print(f"[INFO] Created chunk {chunk_id}: {format_size(current_chunk_size)} ({current_chunk_size:,} bytes)")
# Verify distribution
total_size_actual = sum(chunk_sizes)
if total_size_actual != file_size:
print(f"[WARN] Total chunk size ({format_size(total_size_actual)}) differs from original file size ({format_size(file_size)})")
print(f"[WARN] Difference: {format_size(abs(total_size_actual - file_size))}")
# Calculate statistics
avg_chunk_size = sum(chunk_sizes) / len(chunk_sizes) if chunk_sizes else 0
min_chunk_size = min(chunk_sizes) if chunk_sizes else 0
max_chunk_size = max(chunk_sizes) if chunk_sizes else 0
print(f"\n[INFO] Distribution Summary:")
print(f"- Original file: {os.path.basename(model_file)}")
print(f"- Total size: {format_size(file_size)} ({file_size:,} bytes)")
print(f"- Number of chunks: {len(state.model_chunks)}")
print(f"- Chunks directory: {state.chunks_dir}")
print(f"- Average chunk size: {format_size(avg_chunk_size)}")
print(f"- Smallest chunk: {format_size(min_chunk_size)}")
print(f"- Largest chunk: {format_size(max_chunk_size)}")
print(f"- Size variance: {((max_chunk_size - min_chunk_size) / avg_chunk_size * 100):.1f}%")
return True
except Exception as e:
print(f"[ERROR] Failed to split model weights: {str(e)}")
return False
# Calculate total model size and chunks
total_size_bytes = sum(p.nelement() * p.element_size() for p in weights.values())
num_servers = len(state.tensor_servers) or len(Settings.TENSOR_SERVER_URLS)
# Determine optimal number of chunks based on server count
# If 2 servers -> 2 chunks (500MB each for 1GB)
# If 3 servers -> 3 chunks (333MB each for 1GB)
num_chunks = num_servers
bytes_per_chunk = math.ceil(total_size_bytes / num_chunks)
print(f"[INFO] Total model size: {total_size_bytes / (1024*1024*1024):.2f} GB")
print(f"[INFO] Available servers: {num_servers}")
print(f"[INFO] Creating {num_chunks} chunks")
print(f"[INFO] Target chunk size: {bytes_per_chunk / (1024*1024):.2f} MB")
current_chunk = []
current_chunk_size = 0
chunk_id = 0
chunk_sizes = [] # Track actual chunk sizes for verification
# Sort weights by size for better distribution
sorted_weights = sorted(
weights.items(),
key=lambda x: x[1].nelement() * x[1].element_size(),
reverse=True
)
for key, tensor in weights.items():
tensor_size = tensor.numel()
# Calculate tensor size in bytes
tensor_size = tensor.nelement() * tensor.element_size()
# If adding this tensor would exceed chunk size and we have tensors in current chunk
if (current_chunk_size + tensor_size > bytes_per_chunk and current_chunk) or \
(chunk_id == num_chunks - 1): # Last chunk gets remaining tensors
# Save current chunk
chunk_path = os.path.join(state.model_path, f"chunk_{chunk_id}.safetensors")
chunk_weights = {k: weights[k] for k in current_chunk}
torch.save(chunk_weights, chunk_path)
# Calculate chunk stats
chunk_total_size = sum(weights[k].nelement() * weights[k].element_size()
for k in current_chunk)
chunk_sizes.append(chunk_total_size)
# Create chunk metadata
state.model_chunks[chunk_id] = ModelChunk(
chunk_id=chunk_id,
files=[f"chunk_{chunk_id}.safetensors"],
config={
"weight_keys": current_chunk,
"size_bytes": chunk_total_size,
"num_parameters": sum(weights[k].nelement() for k in current_chunk),
"input_size": weights[current_chunk[0]].size(1) if len(current_chunk) > 0 else 0,
"output_size": weights[current_chunk[-1]].size(0) if len(current_chunk) > 0 else 0,
# assign a vocab_offset cumulatively
"vocab_offset": sum(int(c.config.get('shard_dim', 1)) for c in state.model_chunks.values()),
# Default shard_dim to 1; set correct value in chunk metadata if known
"shard_dim": int(1)
}
)
print(f"[INFO] Created chunk {chunk_id}: {chunk_total_size / (1024*1024):.2f} MB, "
f"{len(current_chunk)} tensors")
# Reset for next chunk
current_chunk = []
current_chunk_size = 0
chunk_id += 1
# If we've created all chunks except last one, put remaining tensors in last chunk
if chunk_id == num_chunks - 1:
remaining_tensors = [k for k, _ in sorted_weights if k not in sum([c.config["weight_keys"]
for c in state.model_chunks.values()], [])]
current_chunk.extend(remaining_tensors)
continue
# Add tensor to current chunk
current_chunk.append(key)
current_chunk_size += tensor_size
# Save last chunk if not empty
if current_chunk:
chunk_path = os.path.join(state.model_path, f"chunk_{chunk_id}.safetensors")
chunk_weights = {k: weights[k] for k in current_chunk}
torch.save(chunk_weights, chunk_path)
# Calculate final chunk stats
chunk_total_size = sum(weights[k].nelement() * weights[k].element_size()
for k in current_chunk)
chunk_sizes.append(chunk_total_size)
state.model_chunks[chunk_id] = ModelChunk(
chunk_id=chunk_id,
files=[f"chunk_{chunk_id}.safetensors"],
config={
"weight_keys": current_chunk,
"size_bytes": chunk_total_size,
"num_parameters": sum(weights[k].nelement() for k in current_chunk),
"input_size": weights[current_chunk[0]].size(1),
"output_size": weights[current_chunk[-1]].size(0)
}
)
print(f"[INFO] Created final chunk {chunk_id}: {chunk_total_size / (1024*1024):.2f} MB, "
f"{len(current_chunk)} tensors")
# Verify distribution
total_size_actual = sum(chunk_sizes)
size_std_dev = torch.tensor(chunk_sizes).std().item() / (1024*1024) # MB
size_mean = torch.tensor(chunk_sizes).mean().item() / (1024*1024) # MB
print(f"\n[INFO] Distribution Summary:")
print(f"- Total model size: {total_size_actual / (1024*1024*1024):.2f} GB")
print(f"- Number of chunks: {len(state.model_chunks)}")
print(f"- Average chunk size: {size_mean:.2f} MB")
print(f"- Chunk size std dev: {size_std_dev:.2f} MB")
print(f"- Size variation: {(size_std_dev/size_mean*100):.1f}%")
# Verify all weights were distributed
all_distributed = set(sum([c.config["weight_keys"] for c in state.model_chunks.values()], []))
if len(all_distributed) != len(weights):
missing = set(weights.keys()) - all_distributed
print(f"[WARN] Some weights were not distributed: {missing}")
return True
except Exception as e:
print(f"[ERROR] Failed to split model weights: {str(e)}")
return False
async def send_chunk_to_server(server_url: str, chunk_id: int, chunk_info: Dict):
"""Send a model chunk to a tensor server"""
try:
print(f"[INFO] Sending chunk {chunk_id} to server {server_url}")
chunk_path = os.path.join(state.chunks_dir, f"chunk_{chunk_id}.bin")
if not os.path.exists(chunk_path):
raise Exception(f"Chunk file not found: {chunk_path}")
# Get chunk metadata
chunk = state.model_chunks[chunk_id]
chunk_data = {
'chunk_id': chunk_id,
'files': [os.path.basename(chunk_path)],
'config': chunk.config
}
async with aiohttp.ClientSession() as session:
# Step 1: Send chunk configuration
async with session.post(
f"{server_url}/load_chunk",
json=chunk_data,
timeout=Settings.TENSOR_SERVER_TIMEOUT
) as response:
if response.status != 200:
error_msg = await response.text()
raise Exception(f"Failed to register chunk: {error_msg}")
result = await response.json()
if not result.get("ready_for_data", False):
raise Exception("Server not ready for chunk data")
# Step 2: Upload chunk data
with open(chunk_path, 'rb') as f:
chunk_file = f.read()
form = aiohttp.FormData()
form.add_field('file',
chunk_file,
filename=os.path.basename(chunk_path),
content_type='application/octet-stream')
async with session.post(
f"{server_url}/upload_chunk_data/{chunk_id}",
data=form,
timeout=Settings.TENSOR_SERVER_TIMEOUT
) as upload_response:
if upload_response.status != 200:
error_msg = await upload_response.text()
raise Exception(f"Failed to upload chunk data: {error_msg}")
upload_result = await upload_response.json()
print(f"[INFO] Successfully uploaded chunk {chunk_id} to {server_url} ({upload_result.get('size_bytes', 0)} bytes)")
return True
except Exception as e:
print(f"[ERROR] Failed to send chunk {chunk_id} to {server_url}: {str(e)}")
return False
async def distribute_model_chunks():
"""Distribute model chunks across available tensor servers"""
try:
available_servers = [
server for server in state.tensor_servers.values()
if server.status in ["ready", "busy"] and server.metrics.error_count < Settings.MAX_ERROR_THRESHOLD
]
min_required = Settings.get_min_servers_required()
if len(available_servers) < min_required:
raise Exception(f"Not enough healthy servers. Need {min_required}, got {len(available_servers)}")
# Create or update weight chunks based on current server count
if not state.model_chunks or len(state.model_chunks) > len(available_servers) * 3:
if not await split_model_weights():
raise Exception("Failed to split model weights")
# Prepare for parallel distribution
tasks = []
min_replicas = Settings.get_min_replica_count(len(available_servers))
chunks_per_server = len(state.model_chunks) / len(available_servers)
print(f"[INFO] Distributing chunks with min {min_replicas} replicas per chunk")
print(f"[INFO] Target chunks per server: {chunks_per_server:.1f}")
# Distribute chunks
for chunk_id, chunk in state.model_chunks.items():
# Calculate optimal number of replicas based on chunk size and server capacity
target_replicas = max(min_replicas,
int(chunks_per_server * len(available_servers) / len(state.model_chunks)))
current_assignments = set(chunk.server_assignments)
current_healthy = [url for url in current_assignments
if state.tensor_servers[url].status in ["ready", "busy"]]
# Remove unhealthy assignments
chunk.server_assignments = current_healthy
# Add new assignments if needed
while len(chunk.server_assignments) < target_replicas:
# Find least loaded eligible server
eligible_servers = [
server for server in available_servers
if str(server.url) not in chunk.server_assignments
and len(server.model_chunks) < (len(state.model_chunks) / len(available_servers) * 1.5)
]
if not eligible_servers:
break
# Sort by load and error count
eligible_servers.sort(key=lambda s: (
len(s.model_chunks),
s.metrics.error_count,
s.metrics.cpu_usage
))
# Assign to best server
best_server = eligible_servers[0]
chunk.server_assignments.append(str(best_server.url))
best_server.model_chunks.append(chunk_id)
print(f"[INFO] Assigned chunk {chunk_id} to server {best_server.url}")
return True
except Exception as e:
print(f"[ERROR] Failed to distribute model chunks: {str(e)}")
return False
async def monitor_tensor_servers():
"""Periodically check health and update metrics of all tensor servers"""
while True:
for server_url, server in state.tensor_servers.items():
try:
# Check basic health
is_healthy = await check_tensor_server_health(server_url)
if not is_healthy:
server.status = "error"
server.metrics.error_count += 1
print(f"[WARN] Server {server_url} is unhealthy")
continue
# Get detailed metrics
async with aiohttp.ClientSession() as session:
async with session.get(f"{server_url}/metrics", timeout=Settings.TENSOR_SERVER_TIMEOUT) as response:
if response.status == 200:
metrics = await response.json()
server.metrics = ServerMetrics(**metrics)
# Update server status based on metrics
if server.metrics.error_count > Settings.MAX_ERROR_THRESHOLD:
server.status = "degraded"
elif server.metrics.cpu_usage > 90 or server.metrics.memory_usage > 90:
server.status = "busy"
else:
server.status = "ready"
server.last_heartbeat = datetime.now()
except Exception as e:
print(f"[ERROR] Failed to monitor server {server_url}: {str(e)}")
server.status = "error"
server.metrics.last_error = str(e)
server.metrics.error_count += 1
# Check for servers that haven't responded in a while
current_time = datetime.now()
for server_url, server in state.tensor_servers.items():
if (current_time - server.last_heartbeat).seconds > Settings.SERVER_TIMEOUT:
print(f"[WARN] Server {server_url} hasn't responded in {Settings.SERVER_TIMEOUT} seconds")
server.status = "error"
await asyncio.sleep(Settings.MONITORING_INTERVAL)
def get_next_model_version(base_dir: str, model_name: str) -> int:
"""Get the next available version number for the model"""
existing_versions = []
model_base_dir = os.path.join(base_dir, model_name)
if os.path.exists(model_base_dir):
for d in os.listdir(model_base_dir):
if d.startswith('v') and d[1:].isdigit():
existing_versions.append(int(d[1:]))
return max(existing_versions + [0]) + 1
def check_existing_model(model_path: str) -> bool:
"""Check if a model exists and has required files"""
if not os.path.exists(model_path):
return False
# Check for essential files
required_files = ['config.json']
model_files = os.listdir(model_path)
# Check for any weight files
has_weights = any(f.endswith(('.bin', '.safetensors')) for f in model_files)
return all(f in model_files for f in required_files) and has_weights
async def download_model_files():
"""Downloads the model files using Hugging Face Hub API"""
try:
print(f"[INFO] Processing model from {Settings.MODEL_REPO}...")
# Install required packages if not present
required_packages = ["huggingface_hub", "requests", "tqdm"]
for package in required_packages:
try:
__import__(package)
except ImportError:
print(f"[INFO] Installing {package}...")
import subprocess
subprocess.check_call(["pip", "install", package])
from huggingface_hub import hf_hub_download, snapshot_download, HfFolder
import requests
from tqdm import tqdm
# Create models directory
models_dir = os.path.join(os.getcwd(), "models")
os.makedirs(models_dir, exist_ok=True)
print(f"[INFO] Models directory: {models_dir}")
# Get the model name from the repository URL
repo_id = "/".join(Settings.MODEL_REPO.split('/')[-2:]) # e.g., "facebook/opt-125m"
model_name = repo_id.split('/')[-1]
# Create versioned model directory
version = get_next_model_version(models_dir, model_name)
model_base_dir = os.path.join(models_dir, model_name)
model_version_dir = os.path.join(model_base_dir, f"v{version}")
# Function to download file with progress bar
def download_file(url, filename):
response = requests.get(url, stream=True)
total_size = int(response.headers.get('content-length', 0))
with open(filename, 'wb') as f, tqdm(
desc=os.path.basename(filename),
total=total_size,
unit='iB',
unit_scale=True,
unit_divisor=1024,
) as pbar:
for data in response.iter_content(chunk_size=1024):
size = f.write(data)
pbar.update(size)
# Check if previous version exists and is valid
if version > 1:
prev_version_dir = os.path.join(model_base_dir, f"v{version-1}")
if check_existing_model(prev_version_dir):
print(f"[INFO] Using existing model from {prev_version_dir}")
model_path = prev_version_dir
state.is_model_loaded = True
else:
# Download new version
os.makedirs(model_version_dir, exist_ok=True)
model_path = model_version_dir
else:
# First time download
os.makedirs(model_version_dir, exist_ok=True)
model_path = model_version_dir
if not state.is_model_loaded:
try:
print(f"[INFO] Downloading model files from {repo_id}...")
# Use snapshot_download to get all necessary files at once, which supports all weight file names
print("[INFO] Downloading all model files (this may take a while)...")
# snapshot_download is the most robust way to get all files matching patterns
# This addresses the user's request to download model files that are not just "pytorch.bin"
model_path = snapshot_download(
repo_id=repo_id,
local_dir=model_path,
allow_patterns=["*.bin", "*.safetensors", "*.json", "*.txt", "tokenizer.model"],
ignore_patterns=["*.msgpack", "*.onnx"], # Ignore non-PyTorch/safetensors formats
force_download=True
)
print(f"[INFO] All files downloaded to {model_path}")
state.is_model_loaded = True
except Exception as e:
raise Exception(f"Failed to download model files: {str(e)}")
# Set model paths in state
state.model_path = model_path
state.chunks_dir = os.path.join(model_path, "chunks")
os.makedirs(state.chunks_dir, exist_ok=True)
# Load and parse the config
config_path = os.path.join(model_path, "config.json")
if os.path.exists(config_path):
with open(config_path, 'r') as f:
state.model_config = json.load(f)
print("[INFO] Loaded model configuration")
print(f"[INFO] Model type: {state.model_config.get('model_type', 'unknown')}")
print(f"[INFO] Architecture: {state.model_config.get('architectures', ['unknown'])[0]}")
else:
print("[WARN] No config.json found in model directory")
# Scan for model files
print("[INFO] Scanning for model files...")
for root, _, files in os.walk(model_path):
for file in files:
if file.endswith(('.bin', '.json', '.safetensors')):
file_path = os.path.join(root, file)
state.model_files[file] = file_path
print(f"[INFO] Found model file: {file}")
if state.model_files:
state.is_model_loaded = True
print(f"[INFO] Model files found successfully! Total files: {len(state.model_files)}")
print(f"[INFO] Model location: {model_path}")
return True
else:
raise ValueError("No model files were found in the repository")
except Exception as e:
print(f"[ERROR] Failed to process model files: {e}")
state.is_model_loaded = False
raise
async def check_tensor_server_health(url: HttpUrl) -> bool:
"""Checks if a tensor server is healthy"""
try:
async with aiohttp.ClientSession() as session:
async with session.get(f"{url}/health", timeout=Settings.TENSOR_SERVER_TIMEOUT) as response:
return response.status == 200
except:
return False
# ===== API Endpoints =====
async def execute_tensor_operation(operation_id: str, server_url: HttpUrl, operation: str, data: Dict):
"""Execute an operation on a tensor server and wait for results"""
try:
async with aiohttp.ClientSession() as session:
# Start the operation
async with session.post(
f"{server_url}/{operation}",
json=data,
timeout=Settings.TENSOR_SERVER_TIMEOUT
) as response:
if response.status != 200:
error_msg = await response.text()
raise HTTPException(
status_code=response.status,
detail=f"Operation failed on server {server_url}: {error_msg}"
)
initial_response = await response.json()
if initial_response.get("status") == "completed":
# Operation completed immediately
state.operation_results[operation_id] = initial_response
return initial_response
# Operation is async, poll for results
while True:
await asyncio.sleep(1) # Poll interval
async with session.get(
f"{server_url}/operation/{initial_response['operation_id']}",
timeout=Settings.TENSOR_SERVER_TIMEOUT
) as status_response:
if status_response.status != 200:
raise HTTPException(
status_code=status_response.status,
detail=f"Failed to get operation status from {server_url}"
)
status_data = await status_response.json()
if status_data["status"] in ["completed", "failed"]:
state.operation_results[operation_id] = status_data
if status_data["status"] == "failed":
raise HTTPException(
status_code=500,
detail=f"Operation failed on server {server_url}: {status_data.get('error')}"
)
return status_data
except asyncio.TimeoutError:
raise HTTPException(
status_code=504,
detail=f"Operation timed out on server {server_url}"
)
except Exception as e:
raise HTTPException(
status_code=500,
detail=f"Error executing operation on {server_url}: {str(e)}"
)
@app.post("/execute/{operation}")
async def execute_operation(operation: str, data: Dict):
"""Execute an operation across tensor servers and collect results"""
operation_id = f"{operation}_{datetime.now().strftime('%Y%m%d_%H%M%S')}_{len(state.operation_results)}"
# Get available servers with required chunks
available_servers = [
server for server in state.tensor_servers.values()
if server.status in ["ready", "busy"]
and server.metrics.error_count < Settings.MAX_ERROR_THRESHOLD
]
if not available_servers:
raise HTTPException(
status_code=503,
detail="No available tensor servers"
)
# Start operations on all relevant servers in parallel
tasks = []
for server in available_servers:
if operation in ["compute", "forward"]:
# For compute operations, only use servers with required chunks
required_chunks = data.get("required_chunks", [])
if not all(chunk_id in server.model_chunks for chunk_id in required_chunks):
continue
task = asyncio.create_task(
execute_tensor_operation(
f"{operation_id}_{server.url}",
server.url,
operation,
data
)
)
tasks.append(task)
state.pending_operations[f"{operation_id}_{server.url}"] = task
if not tasks:
raise HTTPException(
status_code=400,
detail="No servers available with required model chunks"
)
try:
# Wait for all operations to complete
results = await asyncio.gather(*tasks)
# Process and aggregate results
aggregated_result = {
"operation_id": operation_id,
"status": "completed",
"server_results": results,
"timestamp": datetime.now().isoformat()
}
# Clean up
for task_id in list(state.pending_operations.keys()):
if task_id.startswith(operation_id):
del state.pending_operations[task_id]
return aggregated_result
except Exception as e:
# Cancel any remaining tasks
for task in tasks:
if not task.done():
task.cancel()
# Clean up
for task_id in list(state.pending_operations.keys()):
if task_id.startswith(operation_id):
del state.pending_operations[task_id]
raise HTTPException(
status_code=500,
detail=f"Operation failed: {str(e)}"
)
@app.get("/operation/{operation_id}")
async def get_operation_status(operation_id: str):
"""Get the status of an operation"""
# Check completed operations
results = {
k: v for k, v in state.operation_results.items()
if k.startswith(operation_id)
}
if results:
return {
"operation_id": operation_id,
"status": "completed",
"results": results
}
# Check pending operations
pending = {
k: "running" for k in state.pending_operations.keys()
if k.startswith(operation_id)
}
if pending:
return {
"operation_id": operation_id,
"status": "running",
"pending_servers": list(pending.keys())
}
raise HTTPException(
status_code=404,
detail=f"Operation {operation_id} not found"
)
@app.get("/")
async def root():
"""Health check endpoint"""
return {
"status": "running",
"model_loaded": state.is_model_loaded,
"registered_servers": len(state.tensor_servers),
"downloaded_files": len(state.model_files),
"config_loaded": bool(state.model_config)
}
@app.get("/health")
async def health_check():
"""Detailed health check"""
return {
"status": "healthy",
"model_loaded": state.is_model_loaded,
"registered_servers": len(state.tensor_servers),
"downloaded_files": list(state.model_files.keys()),
"config_loaded": bool(state.model_config),
"model_type": state.model_config.get("model_type", "unknown")
}
@app.post("/register_tensor_server")
async def register_tensor_server(server_url: HttpUrl):
"""Register a new tensor server"""
if not await check_tensor_server_health(server_url):
raise HTTPException(status_code=400, detail="Tensor server is not healthy")
state.tensor_servers[str(server_url)] = TensorServer(url=server_url)
print(f"[INFO] Registered new tensor server at {server_url}")
# If model is loaded, automatically distribute chunks
if state.is_model_loaded:
print(f"[INFO] Model is loaded, starting distribution for new server {server_url}")
try:
# Create chunks if they don't exist
if not state.model_chunks:
if await split_model_weights():
print(f"[INFO] Successfully split model into {len(state.model_chunks)} chunks")
else:
print("[ERROR] Failed to split model weights")
# Distribute chunks
if await distribute_model_chunks():
print("[INFO] Successfully distributed chunks to tensor servers")
else:
print("[ERROR] Failed to distribute chunks")
except Exception as e:
print(f"[ERROR] Distribution error during server registration: {str(e)}")
return {
"status": "registered",
"registered_servers": len(state.tensor_servers),
"server_id": str(server_url),
"model_loaded": state.is_model_loaded,
"chunks_distributed": len(state.model_chunks) if state.model_chunks else 0
}
@app.delete("/unregister_tensor_server")
async def unregister_tensor_server(server_url: HttpUrl):
"""Unregister a tensor server"""
if str(server_url) in state.tensor_servers:
# Remove server assignments from chunks
for chunk in state.model_chunks.values():
if str(server_url) in chunk.server_assignments:
chunk.server_assignments.remove(str(server_url))
del state.tensor_servers[str(server_url)]
print(f"[INFO] Unregistered tensor server at {server_url}")
# Trigger redistribution of chunks
await distribute_model_chunks()
return {"status": "unregistered"}
raise HTTPException(status_code=404, detail="Server not found")
@app.get("/server/{server_url}/chunks")
async def get_server_chunks(server_url: HttpUrl):
"""Get the chunks assigned to a specific server"""
if str(server_url) not in state.tensor_servers:
raise HTTPException(status_code=404, detail="Server not found")
server = state.tensor_servers[str(server_url)]
assigned_chunks = [
state.model_chunks[chunk_id]
for chunk_id in server.model_chunks
]
return {
"server_status": server.status,
"assigned_chunks": assigned_chunks,
"metrics": server.metrics.dict()
}
@app.post("/redistribute")
async def redistribute_chunks():
"""Manually trigger redistribution of model chunks"""
success = await distribute_model_chunks()
if not success:
raise HTTPException(status_code=500, detail="Failed to redistribute chunks")
return {
"status": "redistributed",
"chunk_assignments": {
chunk_id: chunk.server_assignments
for chunk_id, chunk in state.model_chunks.items()
}
}
@app.get("/chunks/{chunk_id}/status")
async def get_chunk_status(chunk_id: int):
"""Get the status and assignments of a specific chunk"""
if chunk_id not in state.model_chunks:
raise HTTPException(status_code=404, detail="Chunk not found")
chunk = state.model_chunks[chunk_id]
return {
"chunk_id": chunk_id,
"status": chunk.status,
"server_assignments": chunk.server_assignments,
"metrics": chunk.metrics
}
@app.post("/initialize")
async def initialize_system():
"""Download model files and prepare for distribution"""
await download_model_files()
# Verify downloaded files
files_status = {}
total_size = 0
for filename, filepath in state.model_files.items():
exists = os.path.exists(filepath)
if exists:
size = os.path.getsize(filepath)
total_size += size
files_status[filename] = {"exists": exists, "size_bytes": size}
else:
files_status[filename] = {"exists": exists, "size_bytes": 0}
# Start model distribution if we have tensor servers
distribution_status = "not_started"
if state.tensor_servers:
print("[INFO] Starting automatic model distribution...")
try:
# Split model into chunks
if await split_model_weights():
print(f"[INFO] Successfully split model into {len(state.model_chunks)} chunks")
# Distribute chunks to servers
if await distribute_model_chunks():
print("[INFO] Successfully distributed chunks to tensor servers")
distribution_status = "completed"
else:
print("[ERROR] Failed to distribute chunks")
distribution_status = "distribution_failed"
else:
print("[ERROR] Failed to split model weights")
distribution_status = "split_failed"
except Exception as e:
print(f"[ERROR] Distribution error: {str(e)}")
distribution_status = f"error: {str(e)}"
else:
print("[INFO] No tensor servers registered yet. Will distribute when servers register.")
return {
"status": "initialized",
"model_loaded": state.is_model_loaded,
"files_status": files_status,
"total_size_bytes": total_size,
"config_loaded": bool(state.model_config),
"model_type": state.model_config.get("model_type", "unknown"),
"architecture": state.model_config.get("architectures", ["unknown"])[0],
"distribution_status": distribution_status,
"registered_servers": len(state.tensor_servers),
"chunks_created": len(state.model_chunks) if state.model_chunks else 0
}
# ===== Main Execution =====
@app.on_event("startup")
async def startup_event():
"""Initialize the server and start distribution"""
print("[INFO] Initializing system...")
try:
# Initialize system and download model
await initialize_system()
print("[INFO] Model initialization complete")
# Split model into chunks
if await split_model_weights():
print(f"[INFO] Successfully split model into {len(state.model_chunks)} chunks")
# Distribute chunks to tensor servers
print("[INFO] Starting chunk distribution...")
distribution_tasks = []
# One chunk per server distribution
for chunk_id, chunk in state.model_chunks.items():
# Assign each chunk to exactly one server
server_index = chunk_id % len(Settings.TENSOR_SERVER_URLS)
server_url = Settings.TENSOR_SERVER_URLS[server_index]
# Create task for distributing this chunk to its assigned server
task = asyncio.create_task(
send_chunk_to_server(server_url, chunk_id, {"chunk_id": chunk_id})
)
distribution_tasks.append(task)
print(f"[INFO] Sending chunk {chunk_id} to {server_url}")
# Track assignments for future reference
try:
chunk.server_assignments.append(server_url)
except Exception:
pass
if distribution_tasks:
print(f"[INFO] Distributing {len(distribution_tasks)} chunks...")
results = await asyncio.gather(*distribution_tasks, return_exceptions=True)
success_count = sum(1 for r in results if r is True)
print(f"[INFO] Successfully distributed {success_count} chunks out of {len(distribution_tasks)} attempts")
else:
print("[ERROR] Failed to split model weights")
except Exception as e:
print(f"[ERROR] Startup error: {str(e)}")
print("[INFO] Startup complete")
if __name__ == "__main__":
port = int(os.getenv("PORT", 8000))
print(f"[INFO] Starting controller server on port {port}")
print(f"[INFO] API Documentation available at http://localhost:{port}/docs")
uvicorn.run(
"controller_server_new:app",
host="0.0.0.0",
port=port,
reload=False
) |