Spaces:
Sleeping
Sleeping
File size: 48,177 Bytes
36ed1cd e9c64c8 36ed1cd e9c64c8 36ed1cd e9c64c8 36ed1cd e9c64c8 36ed1cd | 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 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 | """FastAPI backend for GPU Memory Calculator web application."""
import hashlib
import json
import logging
from pathlib import Path
from typing import Any
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from fastapi.staticfiles import StaticFiles
from fastapi.templating import Jinja2Templates
from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
from starlette.requests import Request
from gpu_mem_calculator.config.presets import load_presets
from gpu_mem_calculator.core.calculator import GPUMemoryCalculator
from gpu_mem_calculator.core.models import (
EngineConfig,
GPUConfig,
InferenceConfig,
InferenceEngineType,
InterconnectType,
MemoryResult,
ModelConfig,
NodeConfig,
ParallelismConfig,
TrainingConfig,
)
from gpu_mem_calculator.core.multinode import MultiNodeCalculator
from gpu_mem_calculator.exporters.manager import ExportFormat, ExportManager
from gpu_mem_calculator.huggingface import (
HuggingFaceClient,
HuggingFaceConfigMapper,
HuggingFaceError,
InvalidConfigError,
ModelNotFoundError,
PrivateModelAccessError,
)
from gpu_mem_calculator.inference.calculator import InferenceMemoryCalculator
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# Create FastAPI app
app = FastAPI(
title="GPU Memory Calculator",
description="Calculate GPU memory requirements for LLM training",
version="0.1.0",
)
# Configure CORS
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# Setup templates and static files
BASE_DIR = Path(__file__).parent
templates = Jinja2Templates(directory=str(BASE_DIR / "templates"))
# Mount static files
static_dir = BASE_DIR / "static"
if static_dir.exists():
app.mount("/static", StaticFiles(directory=str(static_dir)), name="static")
# Request/Response models
class CalculateRequest(BaseModel):
"""Request model for memory calculation with comprehensive validation."""
model: dict[str, Any] = Field(description="Model configuration")
training: dict[str, Any] = Field(description="Training configuration")
parallelism: dict[str, Any] | None = Field(
default=None,
description="Parallelism configuration",
)
engine: dict[str, Any] | None = Field(default=None, description="Engine configuration")
hardware: dict[str, Any] | None = Field(default=None, description="Hardware configuration")
@field_validator("model")
@classmethod
def validate_moe_settings(cls, v: dict[str, Any]) -> dict[str, Any]:
"""Validate MoE-specific constraints."""
if v.get("moe_enabled"):
num_experts = v.get("num_experts", 1)
top_k = v.get("top_k", 1)
if top_k > num_experts:
raise ValueError(f"MoE top_k ({top_k}) cannot exceed num_experts ({num_experts})")
if num_experts < 1 or num_experts > 256:
raise ValueError(f"num_experts must be between 1 and 256, got {num_experts}")
if top_k < 1 or top_k > 8:
raise ValueError(f"top_k must be between 1 and 8, got {top_k}")
return v
@model_validator(mode="after")
def validate_parallelism_consistency(self) -> "CalculateRequest":
"""Validate parallelism settings consistency."""
if self.parallelism and self.hardware:
tensor_pp = self.parallelism.get("tensor_parallel_size", 1)
pipeline_pp = self.parallelism.get("pipeline_parallel_size", 1)
data_pp = self.parallelism.get("data_parallel_size", 1)
num_gpus = self.hardware.get("num_gpus", 1)
effective_gpus = tensor_pp * pipeline_pp * data_pp
if effective_gpus != num_gpus:
raise ValueError(
f"Parallelism mismatch: tensor_pp ({tensor_pp}) × "
f"pipeline_pp ({pipeline_pp}) × data_pp ({data_pp}) = "
f"{effective_gpus} GPUs, but num_gpus is set to {num_gpus}. "
f"These must match."
)
# Validate sequence parallel requires tensor parallel > 1
if self.parallelism and self.parallelism.get("sequence_parallel"):
tensor_pp = self.parallelism.get("tensor_parallel_size", 1)
if tensor_pp <= 1:
raise ValueError(
f"Sequence parallelism requires tensor_parallel_size > 1, " f"got {tensor_pp}"
)
return self
@model_validator(mode="after")
def validate_engine_settings(self) -> "CalculateRequest":
"""Validate engine-specific settings."""
if not self.engine:
return self
engine_type = self.engine.get("type")
zero_stage = self.engine.get("zero_stage", 0)
# ZeRO stages only valid for DeepSpeed engines
if engine_type not in ["deepspeed", "megatron_deepspeed"] and zero_stage > 0:
raise ValueError(
f"ZeRO stages are only supported for DeepSpeed engines, "
f"got engine_type='{engine_type}' with zero_stage={zero_stage}"
)
# Validate ZeRO stage range
if zero_stage < 0 or zero_stage > 3:
raise ValueError(f"zero_stage must be between 0 and 3, got {zero_stage}")
return self
class PresetInfo(BaseModel):
"""Information about a preset model configuration."""
name: str
display_name: str
description: str
config: dict[str, Any]
class HuggingFaceRequest(BaseModel):
"""Request for fetching HuggingFace model metadata."""
model_config = ConfigDict(protected_namespaces=())
model_id: str = Field(description="HuggingFace model ID (e.g., meta-llama/Llama-2-7b-hf)")
token: str | None = Field(default=None, description="HF token for private models")
# Simple in-memory cache for calculation results
# In production, use Redis or similar
_calculation_cache: dict[str, tuple[MemoryResult, float]] = {} # key -> (result, timestamp)
_CACHE_TTL = 3600 # 1 hour
_MAX_CACHE_SIZE = 1000
def _cache_key_from_request(request: CalculateRequest) -> str:
"""Generate cache key from request."""
request_dict = request.model_dump()
# Sort keys for consistent hashing
request_str = json.dumps(request_dict, sort_keys=True)
return hashlib.md5(request_str.encode()).hexdigest()
def _get_cached_result(key: str) -> MemoryResult | None:
"""Get cached result if available and not expired."""
if key in _calculation_cache:
result, timestamp = _calculation_cache[key]
import time
if time.time() - timestamp < _CACHE_TTL:
return result
else:
# Expired, remove from cache
del _calculation_cache[key]
return None
def _cache_result(key: str, result: MemoryResult) -> None:
"""Cache calculation result."""
import time
# Simple cache eviction if too large
if len(_calculation_cache) >= _MAX_CACHE_SIZE:
# Remove oldest entry (first key)
oldest_key = next(iter(_calculation_cache))
del _calculation_cache[oldest_key]
_calculation_cache[key] = (result, time.time())
# Load presets at startup using shared preset loader
# The shared loader reads from web/presets/models.json
def _load_presets_from_shared() -> dict[str, PresetInfo]:
"""Load presets using the shared preset loader."""
all_presets = load_presets()
return {
name: PresetInfo(
name=name,
display_name=preset.get("display_name", name),
description=preset.get("description", ""),
config=preset.get("config", {}),
)
for name, preset in all_presets.items()
}
PRESETS = _load_presets_from_shared()
# API Routes
@app.get("/")
async def index(request: Request) -> Any:
"""Serve the main web page."""
return templates.TemplateResponse("index.html", {"request": request})
@app.get("/api/engines")
async def list_engines() -> dict[str, str]:
"""List supported training engines."""
return {
"pytorch_ddp": "PyTorch DDP (Distributed Data Parallel)",
"deepspeed": "DeepSpeed ZeRO",
"megatron_lm": "Megatron-LM",
"fsdp": "PyTorch FSDP (Fully Sharded Data Parallel)",
"megatron_deepspeed": "Megatron-LM + DeepSpeed",
}
@app.get("/api/optimizers")
async def list_optimizers() -> dict[str, str]:
"""List supported optimizers."""
return {
"adam": "Adam",
"adamw": "AdamW",
"adamw_8bit": "AdamW 8-bit",
"sgd": "SGD",
}
@app.get("/api/dtypes")
async def list_dtypes() -> dict[str, str]:
"""List supported data types."""
return {
"fp32": "FP32 (32-bit floating point)",
"fp16": "FP16 (16-bit floating point)",
"bf16": "BF16 (16-bit bfloat)",
"int8": "INT8 (8-bit integer)",
"int4": "INT4 (4-bit integer)",
}
@app.get("/api/presets")
async def list_presets() -> dict[str, dict[str, str]]:
"""List all preset model configurations."""
return {
name: {
"display_name": preset.display_name,
"description": preset.description,
}
for name, preset in PRESETS.items()
}
@app.get("/api/preset/{preset_name}")
async def get_preset(preset_name: str) -> dict[str, Any]:
"""Get a specific preset configuration."""
if preset_name not in PRESETS:
raise HTTPException(status_code=404, detail=f"Preset '{preset_name}' not found")
return PRESETS[preset_name].config
@app.post("/api/hf/fetch")
async def fetch_huggingface_model(request: HuggingFaceRequest) -> dict[str, Any]:
"""Fetch model metadata from HuggingFace Hub.
Args:
request: Request with model_id and optional token
Returns:
Model config with fields filled from HF, plus list of missing fields
Raises:
HTTPException: If model not found, access denied, or invalid config
"""
try:
# Initialize HF client
client = HuggingFaceClient(token=request.token)
# Fetch metadata
metadata = await client.fetch_model_metadata(request.model_id)
# Map to ModelConfig
mapper = HuggingFaceConfigMapper()
result = mapper.map_to_model_config(metadata["config"], metadata.get("model_info"))
return {
"model_id": request.model_id,
"config": result["config"],
"missing_fields": result["missing_fields"],
"found_fields": result["found_fields"],
"warnings": [],
}
except PrivateModelAccessError as e:
raise HTTPException(
status_code=401,
detail={
"error": "Authentication required",
"message": str(e),
"type": "auth_error",
},
) from e
except ModelNotFoundError as e:
raise HTTPException(
status_code=404,
detail={
"error": "Model not found",
"message": str(e),
"type": "not_found",
},
) from e
except InvalidConfigError as e:
raise HTTPException(
status_code=422,
detail={
"error": "Invalid model configuration",
"message": str(e),
"type": "invalid_config",
},
) from e
except HuggingFaceError as e:
logger.error(f"HuggingFace error: {str(e)}")
raise HTTPException(
status_code=500,
detail={
"error": "HuggingFace API error",
"message": str(e),
"type": "api_error",
},
) from e
except Exception as e:
logger.error(f"Unexpected error fetching HF model: {str(e)}", exc_info=True)
raise HTTPException(
status_code=500,
detail={
"error": "Internal server error",
"message": "An unexpected error occurred",
"type": "server_error",
},
) from e
@app.post("/api/calculate")
async def calculate_memory(request: CalculateRequest) -> MemoryResult:
"""Calculate GPU memory requirements.
Args:
request: Calculation request with model, training, and hardware configs
Returns:
MemoryResult with complete memory breakdown
"""
# Check cache first
cache_key = _cache_key_from_request(request)
cached_result = _get_cached_result(cache_key)
if cached_result is not None:
logger.info(f"Cache hit for key: {cache_key[:8]}...")
return cached_result
try:
# Parse model configuration
model_data = request.model.copy()
# Parse num_parameters if it's a string (e.g., "7B", "7000M")
if "num_parameters" in model_data and isinstance(
model_data["num_parameters"],
str,
):
from gpu_mem_calculator.config.parser import ConfigParser
model_data["num_parameters"] = ConfigParser._parse_num_params(
model_data["num_parameters"],
)
model_config = ModelConfig(**model_data)
# Parse training configuration
training_config = TrainingConfig(**request.training)
# Parse optional configurations with defaults
parallelism_config = (
ParallelismConfig(**request.parallelism) if request.parallelism else ParallelismConfig()
)
engine_config = EngineConfig(**request.engine) if request.engine else EngineConfig()
gpu_config = GPUConfig(**request.hardware) if request.hardware else GPUConfig()
# Create calculator and compute
calculator = GPUMemoryCalculator(
model_config=model_config,
training_config=training_config,
parallelism_config=parallelism_config,
engine_config=engine_config,
gpu_config=gpu_config,
)
result = calculator.calculate()
# Cache the result
_cache_result(cache_key, result)
logger.info(
f"Calculation successful: {model_config.name}, "
f"{result.total_memory_per_gpu_gb:.2f} GB per GPU"
)
return result
except ValueError as e:
# User input validation error
logger.warning(f"Validation error: {str(e)}")
raise HTTPException(
status_code=400,
detail={"error": "Validation error", "message": str(e), "type": "validation_error"},
) from e
except Exception as e:
# Unexpected system error
logger.error(f"Calculation error: {str(e)}", exc_info=True)
raise HTTPException(
status_code=500,
detail={
"error": "Internal server error",
"message": "An unexpected error occurred during calculation",
},
) from e
@app.post("/api/export/deepspeed")
async def export_deepspeed_config(request: CalculateRequest) -> dict[str, Any]:
"""Export DeepSpeed configuration file.
Args:
request: Calculation request with model, training, and hardware configs
Returns:
DeepSpeed config JSON and memory result
"""
try:
# First calculate memory
calc_result = await calculate_memory(request)
# Generate DeepSpeed config
parallelism = request.parallelism or {}
training = request.training
engine = request.engine or {}
train_batch_size = (
training.get("batch_size", 1)
* training.get("gradient_accumulation_steps", 1)
* parallelism.get("data_parallel_size", 1)
)
zero_stage = engine.get("zero_stage", 0)
offload_optimizer = engine.get("offload_optimizer", "none")
offload_param = engine.get("offload_param", "none")
deepspeed_config = {
"train_batch_size": train_batch_size,
"train_micro_batch_size_per_gpu": training.get("batch_size", 1),
"gradient_accumulation_steps": training.get("gradient_accumulation_steps", 1),
"optimizer": {
"type": training.get("optimizer", "AdamW"),
"params": {"lr": 0.0001, "betas": [0.9, 0.999], "eps": 1e-8, "weight_decay": 0.01},
},
"scheduler": {
"type": "WarmupLR",
"params": {"warmup_min_lr": 0, "warmup_max_lr": 0.0001, "warmup_num_steps": 2000},
},
"fp16": {"enabled": training.get("dtype") in ["fp16", "int4", "int8"]},
"bf16": {"enabled": training.get("dtype") == "bf16"},
"zero_optimization": {"stage": zero_stage},
"gradient_clipping": training.get("gradient_clipping", 1.0),
"steps_per_print": 100,
}
# Add offload config if ZeRO stage >= 1
if zero_stage >= 1:
deepspeed_config["zero_optimization"]["offload_optimizer"] = {
"device": offload_optimizer
}
deepspeed_config["zero_optimization"]["offload_param"] = {"device": offload_param}
return {"config": deepspeed_config, "memory_result": calc_result}
except HTTPException:
raise
except Exception as e:
logger.error(f"DeepSpeed export error: {str(e)}", exc_info=True)
raise HTTPException(
status_code=500, detail=f"Failed to generate DeepSpeed config: {str(e)}"
) from e
@app.post("/api/optimize/batch-size")
async def optimize_batch_size(request: CalculateRequest) -> dict[str, Any]:
"""Find maximum batch size that fits in GPU memory.
Uses binary search to find the maximum batch size that doesn't OOM.
Args:
request: Calculation request with model, training, and hardware configs
Returns:
Maximum batch size that fits and corresponding memory result
"""
try:
# Create a mutable copy for testing
from copy import deepcopy
min_batch = 1
max_batch = 512 # Reasonable upper bound
best_batch = 1
while min_batch <= max_batch:
mid = (min_batch + max_batch) // 2
# Create modified request with test batch size
test_request = deepcopy(request)
test_request.training["batch_size"] = mid
try:
# Validate and calculate
CalculateRequest.model_validate(test_request)
result = await calculate_memory(test_request)
if result.fits_on_gpu:
best_batch = mid
min_batch = mid + 1
else:
max_batch = mid - 1
except (ValueError, HTTPException):
# Invalid config or doesn't fit
max_batch = mid - 1
# Get final result for best batch size
final_request = deepcopy(request)
final_request.training["batch_size"] = best_batch
final_result = await calculate_memory(final_request)
return {"max_batch_size": best_batch, "memory_result": final_result}
except Exception as e:
logger.error(f"Batch size optimization error: {str(e)}", exc_info=True)
raise HTTPException(
status_code=500, detail=f"Failed to optimize batch size: {str(e)}"
) from e
@app.post("/api/validate")
async def validate_config(request: CalculateRequest) -> dict[str, Any]:
"""Validate a configuration without calculating memory.
Args:
request: Configuration to validate
Returns:
Validation result with valid flag and any errors
"""
try:
# Pydantic validation happens automatically when creating CalculateRequest
# If we get here, the request is valid
return {"valid": True, "errors": []}
except ValueError as e:
# Validation error
return {"valid": False, "errors": [str(e)]}
except Exception as e:
# Unexpected error
logger.error(f"Validation error: {str(e)}", exc_info=True)
return {"valid": False, "errors": [str(e)]}
@app.post("/api/explain-formula")
async def explain_formula(request: CalculateRequest) -> dict[str, Any]:
"""Explain the memory formula used for calculation.
Returns detailed information about which formula is being used,
with the user's values plugged in, and links to documentation.
Args:
request: Calculation request with model, training, and hardware configs
Returns:
Formula explanation with formula type, breakdown, and references
"""
try:
# Get configuration details
engine_type = request.engine.get("type", "pytorch_ddp") if request.engine else "pytorch_ddp"
num_params = request.model.get("num_parameters", 0)
# Parse num_parameters if it's a string (e.g., "7B", "7000M")
if isinstance(num_params, str):
from gpu_mem_calculator.config.parser import ConfigParser
num_params = ConfigParser._parse_num_params(num_params)
optimizer = request.training.get("optimizer", "adamw")
num_gpus = request.hardware.get("num_gpus", 1) if request.hardware else 1
batch_size = request.training.get("batch_size", 1)
# Calculate memory to get the breakdown
result = await calculate_memory(request)
# Determine formula description based on engine type
formula_info = {
"engine_type": engine_type,
"engine_name": _get_engine_name(engine_type),
"formula_components": [],
"total_memory_gb": round(result.total_memory_per_gpu_gb, 2),
"breakdown": {
"model_params_gb": round(result.breakdown.model_params_gb, 2),
"gradients_gb": round(result.breakdown.gradients_gb, 2),
"optimizer_states_gb": round(result.breakdown.optimizer_states_gb, 2),
"activations_gb": round(result.breakdown.activations_gb, 2),
"overhead_gb": round(result.breakdown.overhead_gb, 2),
},
"references": _get_formula_references(engine_type),
}
# Add engine-specific formula details
if engine_type == "pytorch_ddp":
formula_info["formula_description"] = (
"PyTorch DDP stores complete copies of model parameters, gradients, "
"and optimizer states on each GPU."
)
formula_info["formula_components"] = [
{
"name": "Model Parameters",
"formula": f"{num_params:,} × 2 bytes (FP16/BF16)",
"result": f"{result.breakdown.model_params_gb:.2f} GB",
"description": "Full model stored on each GPU",
},
{
"name": "Gradients",
"formula": f"{num_params:,} × 2 bytes (FP16)",
"result": f"{result.breakdown.gradients_gb:.2f} GB",
"description": "Full gradients during backward pass",
},
{
"name": "Optimizer States",
"formula": _get_optimizer_formula(optimizer, num_params)["formula"],
"result": f"{result.breakdown.optimizer_states_gb:.2f} GB",
"description": _get_optimizer_formula(optimizer, num_params)["description"],
},
]
elif engine_type in ["deepspeed", "megatron_deepspeed"]:
zero_stage = request.engine.get("zero_stage", 0) if request.engine else 0
offload_optimizer = (
request.engine.get("offload_optimizer", "none") if request.engine else "none"
)
offload_param = (
request.engine.get("offload_param", "none") if request.engine else "none"
)
if zero_stage == 0:
stage_name = "ZeRO-0 (Baseline)"
formula_info["formula_description"] = (
f"{stage_name}: No memory optimization. Same as PyTorch DDP."
)
elif zero_stage == 1:
stage_name = "ZeRO-1"
formula_info["formula_description"] = (
f"{stage_name}: Shards optimizer states across {num_gpus} GPUs. "
f"Reduces optimizer memory by {num_gpus}x."
)
elif zero_stage == 2:
stage_name = "ZeRO-2"
formula_info["formula_description"] = (
f"{stage_name}: Shards optimizer states AND gradients across {num_gpus} GPUs. "
f"Reduces memory by {num_gpus}x for both components."
)
elif zero_stage == 3:
stage_name = "ZeRO-3"
formula_info["formula_description"] = (
f"{stage_name}: Shards parameters, gradients, AND optimizer states. "
f"Only largest layer stored intact. Linear memory reduction with GPU count."
)
formula_info["zero_stage"] = zero_stage
formula_info["offload_optimizer"] = offload_optimizer
formula_info["offload_param"] = offload_param
# Add ZeRO-specific components
if zero_stage == 3:
# Estimate largest layer (approx 10% of params for typical models)
largest_params = num_params // 10
formula_info["formula_components"] = [
{
"name": "Largest Layer",
"formula": f"{largest_params:,} × 4 bytes (FP16 params + grads)",
"result": f"{result.breakdown.model_params_gb:.2f} GB",
"description": "Gathered during compute, largest layer kept intact",
},
{
"name": "Sharded Parameters",
"formula": f"({num_params:,} × 2 bytes) / {num_gpus} GPUs",
"result": "Included in model params",
"description": "Remaining parameters sharded across GPUs",
},
{
"name": "Sharded Optimizer States",
"formula": (
(
f"({_get_optimizer_formula(optimizer, num_params)['formula']}) "
f"/ {num_gpus} GPUs"
)
if offload_optimizer == "none"
else f"Offloaded to {offload_optimizer}"
),
"result": f"{result.breakdown.optimizer_states_gb:.2f} GB",
"description": (
_get_optimizer_formula(optimizer, num_params)["description"]
+ " (sharded or offloaded)"
),
},
]
else:
# ZeRO-1 or ZeRO-2
formula_info["formula_components"] = [
{
"name": "Model Parameters",
"formula": f"{num_params:,} × 2 bytes (FP16)",
"result": f"{result.breakdown.model_params_gb:.2f} GB",
"description": "Full model on each GPU",
},
{
"name": "Gradients",
"formula": (
f"{num_params:,} × 2 bytes"
if zero_stage < 2
else f"({num_params:,} × 2 bytes) / {num_gpus} GPUs"
),
"result": f"{result.breakdown.gradients_gb:.2f} GB",
"description": (
"Sharded across GPUs" if zero_stage >= 2 else "Full gradients"
),
},
{
"name": "Optimizer States",
"formula": (
(
f"({_get_optimizer_formula(optimizer, num_params)['formula']}) "
f"/ {num_gpus} GPUs"
)
if offload_optimizer == "none"
else f"Offloaded to {offload_optimizer}"
),
"result": f"{result.breakdown.optimizer_states_gb:.2f} GB",
"description": (
_get_optimizer_formula(optimizer, num_params)["description"]
+ " (sharded or offloaded)"
),
},
]
elif engine_type == "fsdp":
sharding_strategy = (
request.engine.get("sharding_strategy", "full_shard")
if request.engine
else "full_shard"
)
if sharding_strategy == "no_shard":
strategy_name = "No Sharding (like DDP)"
elif sharding_strategy == "shard_grad_op":
strategy_name = "Shard Gradients + Optimizer (like ZeRO-2)"
else:
strategy_name = "Full Shard (like ZeRO-3)"
formula_info["sharding_strategy"] = sharding_strategy
formula_info["strategy_name"] = strategy_name
formula_info["formula_description"] = f"FSDP with {strategy_name.lower()} strategy."
elif engine_type == "megatron_lm":
formula_info["formula_description"] = (
"Megatron-LM uses tensor and/or pipeline parallelism to "
"split the model across GPUs, reducing memory per GPU."
)
# Add parallelism info
if request.parallelism:
tp_size = request.parallelism.get("tensor_parallel_size", 1)
pp_size = request.parallelism.get("pipeline_parallel_size", 1)
formula_info["parallelism"] = {
"tensor_parallel_size": tp_size,
"pipeline_parallel_size": pp_size,
}
# Add activation memory explanation
components: list[dict[str, Any]] = formula_info["formula_components"] # type: ignore[assignment]
components.append(
{
"name": "Activations",
"formula": (
f"batch_size({batch_size}) × seq_len × hidden_size × "
f"layers × ~16 bytes/token/layer"
),
"result": f"{result.breakdown.activations_gb:.2f} GB",
"description": "Memory from intermediate activations during forward/backward pass",
}
)
return formula_info
except Exception as e:
logger.error(f"Formula explanation error: {str(e)}", exc_info=True)
raise HTTPException(
status_code=500, detail=f"Failed to generate formula explanation: {str(e)}"
) from e
def _get_engine_name(engine_type: str) -> str:
"""Get human-readable engine name."""
names = {
"pytorch_ddp": "PyTorch DDP (Distributed Data Parallel)",
"deepspeed": "DeepSpeed ZeRO",
"megatron_lm": "Megatron-LM",
"fsdp": "PyTorch FSDP (Fully Sharded Data Parallel)",
"megatron_deepspeed": "Megatron-LM + DeepSpeed",
}
return names.get(engine_type, engine_type)
def _get_optimizer_formula(optimizer: str, num_params: int) -> dict[str, str]:
"""Get optimizer memory formula based on optimizer type.
Args:
optimizer: Optimizer type (adam, adamw, sgd, adamw_8bit)
num_params: Number of model parameters
Returns:
Dictionary with 'formula' and 'description' keys
"""
num_params_formatted = f"{num_params:,}"
if optimizer in ["adam", "adamw"]:
return {
"formula": f"{num_params_formatted} × 12 bytes (Adam/AdamW FP32)",
"description": "4 bytes FP32 params + 4 bytes momentum + 4 bytes variance",
}
elif optimizer == "adamw_8bit":
return {
"formula": f"{num_params_formatted} × 2 bytes (AdamW 8-bit)",
"description": "8-bit quantized optimizer states (2 bytes per parameter)",
}
elif optimizer == "sgd":
return {
"formula": f"{num_params_formatted} × 4 bytes (SGD)",
"description": "4 bytes FP32 params (no momentum for SGD)",
}
else:
# Default to AdamW
return {
"formula": f"{num_params_formatted} × 12 bytes (Adam/AdamW FP32)",
"description": "4 bytes FP32 params + 4 bytes momentum + 4 bytes variance",
}
def _get_formula_references(engine_type: str) -> list[dict[str, str]]:
"""Get authoritative references for the formula."""
references = [
{
"title": "EleutherAI Transformer Math 101",
"url": "https://blog.eleuther.ai/transformer-math/",
"description": "Comprehensive transformer memory breakdown with formulas",
},
{
"title": "Microsoft Research ZeRO Blog",
"url": "https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/",
"description": "ZeRO optimization techniques and memory formulas",
},
]
if engine_type in ["deepspeed", "megatron_deepspeed"]:
references.append(
{
"title": "DeepSpeed Memory Documentation",
"url": "https://deepspeed.readthedocs.io/en/latest/memory.html",
"description": "Official DeepSpeed memory requirements and formulas",
}
)
elif engine_type == "megatron_lm" or engine_type == "megatron_deepspeed":
references.append(
{
"title": "NVIDIA Megatron-LM",
"url": "https://github.com/NVIDIA/Megatron-LM",
"description": "Megatron-LM tensor and pipeline parallelism",
}
)
elif engine_type == "fsdp":
references.append(
{
"title": "PyTorch FSDP Documentation",
"url": "https://pytorch.org/docs/stable/fsdp.html",
"description": "PyTorch Fully Sharded Data Parallel documentation",
}
)
return references
@app.post("/api/inference/calculate")
async def calculate_inference_memory(request: dict[str, Any]) -> dict[str, Any]:
"""Calculate GPU memory requirements for inference.
Args:
request: Dictionary with model, inference, and hardware configs
Returns:
Inference memory result with breakdown
"""
try:
model_data = request.get("model", {})
inference_data = request.get("inference", {})
hardware_data = request.get("hardware", {})
# Parse num_parameters if it's a string
if "num_parameters" in model_data and isinstance(model_data["num_parameters"], str):
from gpu_mem_calculator.config.parser import ConfigParser
model_data["num_parameters"] = ConfigParser._parse_num_params(
model_data["num_parameters"]
)
# Create model config
model_config = ModelConfig(**model_data)
# Create inference config
kv_cache_quantization = inference_data.get("kv_cache_quantization", "none")
if isinstance(kv_cache_quantization, str):
from gpu_mem_calculator.core.models import KVCacheQuantization
kv_cache_quantization = KVCacheQuantization(kv_cache_quantization)
inference_config = InferenceConfig(
batch_size=inference_data.get("batch_size", 1),
kv_cache_quantization=kv_cache_quantization,
use_kv_cache=inference_data.get("use_kv_cache", True),
tensor_parallel_size=inference_data.get("tensor_parallel_size", 1),
gpu_memory_utilization=inference_data.get("gpu_memory_utilization", 0.9),
enable_streaming=inference_data.get("enable_streaming", False),
# TGI-specific parameters
max_total_tokens=inference_data.get("max_total_tokens"),
max_input_tokens=inference_data.get("max_input_tokens"),
max_batch_total_tokens=inference_data.get("max_batch_total_tokens"),
tgi_quantize=inference_data.get("tgi_quantize", "none"),
tgi_dtype=inference_data.get("tgi_dtype", "bfloat16"),
sharded=inference_data.get("sharded", False),
num_shard=inference_data.get("num_shard"),
# vLLM-specific parameters
block_size=inference_data.get("block_size"),
swap_space_gb=inference_data.get("swap_space_gb", 0.0),
enable_prefix_caching=inference_data.get("enable_prefix_caching", False),
enforce_eager=inference_data.get("enforce_eager", False),
max_num_batched_tokens=inference_data.get("max_num_batched_tokens"),
max_num_seqs=inference_data.get("max_num_seqs"),
vllm_quantization=inference_data.get("vllm_quantization", "none"),
# TensorRT-LLM-specific parameters
trt_max_batch_size=inference_data.get("trt_max_batch_size"),
trt_max_input_len=inference_data.get("trt_max_input_len"),
trt_max_seq_len=inference_data.get("trt_max_seq_len"),
trt_max_beam_width=inference_data.get("trt_max_beam_width"),
# SGLang-specific parameters
chunk_size=inference_data.get("chunk_size"),
max_running_requests=inference_data.get("max_running_requests"),
disable_radix_cache=inference_data.get("disable_radix_cache", False),
enable_p2p=inference_data.get("enable_p2p", False),
disable_custom_all_reduce=inference_data.get("disable_custom_all_reduce", False),
attention_backend=inference_data.get("attention_backend", "flashinfer"),
enable_torch_compile=inference_data.get("enable_torch_compile", False),
radix_cache_max_seq_len=inference_data.get("radix_cache_max_seq_len"),
speculative_algo=inference_data.get("speculative_algo", "default"),
multi_lora_enabled=inference_data.get("multi_lora_enabled", False),
)
# Create GPU config
gpu_config = GPUConfig(
num_gpus=hardware_data.get("num_gpus", 1),
gpu_memory_gb=hardware_data.get("gpu_memory_gb", 80),
)
# Get engine type
engine_type_str = inference_data.get("engine_type", "huggingface")
engine_type_map = {
"huggingface": InferenceEngineType.HUGGINGFACE,
"vllm": InferenceEngineType.VLLM,
"tgi": InferenceEngineType.TGI,
"tensorrt_llm": InferenceEngineType.TENSORRT_LLM,
"sglang": InferenceEngineType.SGLANG,
}
engine_type = engine_type_map.get(engine_type_str, InferenceEngineType.HUGGINGFACE)
# Calculate inference memory
calculator = InferenceMemoryCalculator(model_config, inference_config, gpu_config)
result = calculator.calculate(engine_type)
return {
"total_memory_per_gpu_gb": result.total_memory_per_gpu_gb,
"total_memory_all_gpus_gb": result.total_memory_all_gpus_gb,
"breakdown": {
"model_params_gb": result.breakdown.model_params_gb,
"kv_cache_gb": result.breakdown.kv_cache_gb,
"activations_gb": result.breakdown.activations_gb,
"overhead_gb": result.breakdown.overhead_gb,
},
"max_supported_batch_size": result.max_supported_batch_size,
"estimated_throughput_tokens_per_sec": result.estimated_throughput_tokens_per_sec,
"fits_on_gpu": result.fits_on_gpu,
"memory_utilization_percent": result.memory_utilization_percent,
}
except Exception as e:
logger.error(f"Inference calculation error: {str(e)}", exc_info=True)
raise HTTPException(
status_code=500, detail=f"Failed to calculate inference memory: {str(e)}"
) from e
@app.post("/api/multinode/calculate")
async def calculate_multinode(request: dict[str, Any]) -> dict[str, Any]:
"""Calculate network overhead for multi-node training.
Args:
request: Dictionary with model, training, parallelism, engine, and node configs
Returns:
Network overhead result with suggestions
"""
try:
model_data = request.get("model", {})
training_data = request.get("training", {})
parallelism_data = request.get("parallelism", {})
engine_data = request.get("engine", {})
node_data = request.get("node_config", {})
# Parse num_parameters if it's a string
if "num_parameters" in model_data and isinstance(model_data["num_parameters"], str):
from gpu_mem_calculator.config.parser import ConfigParser
model_data["num_parameters"] = ConfigParser._parse_num_params(
model_data["num_parameters"]
)
# Create minimal configs for multi-node calculation
model_config = ModelConfig(
name="multinode-model",
num_parameters=model_data.get("num_parameters", 7_000_000_000),
num_layers=32,
hidden_size=4096,
num_attention_heads=32,
)
training_config = TrainingConfig(
dtype=training_data.get("dtype", "bf16"),
batch_size=training_data.get("batch_size", 4),
)
parallelism_config = ParallelismConfig(
tensor_parallel_size=parallelism_data.get("tensor_parallel_size", 1),
pipeline_parallel_size=parallelism_data.get("pipeline_parallel_size", 1),
sequence_parallel=parallelism_data.get("sequence_parallel", False),
)
engine_config = EngineConfig(
type=engine_data.get("type", "deepspeed"),
zero_stage=engine_data.get("zero_stage", 3),
)
interconnect_type_str = node_data.get("interconnect_type", "infiniband")
interconnect_map = {
"infiniband": InterconnectType.INFINIBAND,
"nvlink": InterconnectType.NVLINK,
"ethernet_200g": InterconnectType.ETHERNET_200G,
"ethernet_100g": InterconnectType.ETHERNET_100G,
"ethernet_25g": InterconnectType.ETHERNET_25G,
"ethernet_10g": InterconnectType.ETHERNET_10G,
}
interconnect_type = interconnect_map.get(interconnect_type_str, InterconnectType.INFINIBAND)
node_config = NodeConfig(
num_nodes=node_data.get("num_nodes", 2),
gpus_per_node=node_data.get("gpus_per_node", 8),
interconnect_type=interconnect_type,
)
# Calculate network overhead
calculator = MultiNodeCalculator(
model_config=model_config,
training_config=training_config,
parallelism_config=parallelism_config,
node_config=node_config,
engine_config=engine_config,
)
overhead = calculator.calculate_network_overhead()
# Generate optimization suggestions
suggestions: list[str] = []
if overhead.total_overhead_gb > 10:
suggestions.append("Consider reducing tensor parallelism to lower AllGather overhead")
if overhead.estimated_overhead_ms_per_step and overhead.estimated_overhead_ms_per_step > 50:
overhead_val = overhead.estimated_overhead_ms_per_step
suggestions.append(
f"High communication overhead ({overhead_val:.1f}ms/step). "
"Consider upgrading interconnect or reducing model size."
)
if interconnect_type_str.startswith("ethernet") and node_config.num_nodes > 2:
suggestions.append(
"Ethernet interconnect detected. For multi-node training, "
"consider InfiniBand for better performance."
)
return {
"network_overhead": {
"total_overhead_gb": overhead.total_overhead_gb,
"allreduce_gb": overhead.allreduce_gb,
"allgather_gb": overhead.allgather_gb,
"reducescatter_gb": overhead.reducescatter_gb,
"pipeline_gb": overhead.point_to_point_gb,
"estimated_overhead_ms_per_step": overhead.estimated_overhead_ms_per_step,
"communication_time_ms_per_step": None,
"latency_overhead_ms": None,
},
"suggestions": suggestions,
}
except Exception as e:
logger.error(f"Multi-node calculation error: {str(e)}", exc_info=True)
raise HTTPException(
status_code=500, detail=f"Failed to calculate multi-node overhead: {str(e)}"
) from e
@app.post("/api/export/{format}")
async def export_framework_config(format: str, request: CalculateRequest) -> dict[str, Any]:
"""Export configuration to framework-specific format.
Args:
format: Export format (accelerate, lightning, axolotl, deepspeed, yaml, json)
request: Calculation request with all configurations
Returns:
Exported configuration file content
"""
try:
# Parse configurations
model_data = request.model.copy()
if "num_parameters" in model_data and isinstance(model_data["num_parameters"], str):
from gpu_mem_calculator.config.parser import ConfigParser
model_data["num_parameters"] = ConfigParser._parse_num_params(
model_data["num_parameters"]
)
model_config = ModelConfig(**model_data)
training_config = TrainingConfig(**request.training)
parallelism_config = (
ParallelismConfig(**request.parallelism) if request.parallelism else ParallelismConfig()
)
engine_config = EngineConfig(**request.engine) if request.engine else EngineConfig()
# Create minimal node config (not used for single-node export)
node_config = NodeConfig(num_nodes=1, gpus_per_node=8)
# Map format string to ExportFormat enum
format_map = {
"accelerate": ExportFormat.ACCELERATE,
"lightning": ExportFormat.LIGHTNING,
"axolotl": ExportFormat.AXOLOTL,
"deepspeed": ExportFormat.DEEPSPEED,
"yaml": ExportFormat.YAML,
"json": ExportFormat.JSON,
}
export_format = format_map.get(format.lower())
if not export_format:
raise HTTPException(
status_code=400,
detail=f"Unsupported export format: {format}. Supported: {list(format_map.keys())}",
)
# Export configuration
manager = ExportManager(
model_config=model_config,
training_config=training_config,
parallelism_config=parallelism_config,
engine_config=engine_config,
node_config=node_config,
)
result = manager.export(export_format)
# Generate filename
if isinstance(result, dict):
filename = f"config_{format}.{result.get('extension', 'txt')}"
else:
filename = f"config.{format}"
return {
"format": format,
"content": result,
"filename": filename,
}
except HTTPException:
raise
except Exception as e:
logger.error(f"Export error ({format}): {str(e)}", exc_info=True)
raise HTTPException(
status_code=500, detail=f"Failed to export {format} config: {str(e)}"
) from e
def main() -> None:
"""Run the development server."""
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=8000, reload=True)
if __name__ == "__main__":
main()
|