unsloth/studio/backend/utils/hardware/hardware.py

258 lines
7.7 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

"""
Hardware detection — run once at startup, read everywhere.
Usage:
# At FastAPI lifespan startup:
from utils.hardware import detect_hardware
detect_hardware()
# Anywhere else:
from utils.hardware import DEVICE, DeviceType, is_apple_silicon
if DEVICE == DeviceType.CUDA:
import torch
...
"""
import platform
import logging
from enum import Enum
from typing import Optional, Dict, Any
logger = logging.getLogger(__name__)
# ========== Device Enum ==========
class DeviceType(str, Enum):
"""Supported compute backends. Inherits from str so it serializes cleanly in JSON."""
CUDA = "cuda"
MLX = "mlx"
CPU = "cpu"
# ========== Global State (set once by detect_hardware) ==========
DEVICE: Optional[DeviceType] = None
# ========== Detection ==========
def is_apple_silicon() -> bool:
"""Check if running on Apple Silicon hardware (pure platform check, no ML imports)."""
return platform.system() == "Darwin" and platform.machine() == "arm64"
def _has_torch() -> bool:
"""Check if PyTorch is importable."""
try:
import torch
return True
except ImportError:
return False
def _has_mlx() -> bool:
"""Check if MLX is importable."""
try:
import mlx.core
return True
except ImportError:
return False
def detect_hardware() -> DeviceType:
"""
Detect the best available compute device and set the module-level DEVICE global.
Should be called exactly once during FastAPI lifespan startup.
Safe to call multiple times (idempotent).
Detection order:
1. CUDA (NVIDIA GPU, requires torch)
2. MLX (Apple Silicon via MLX framework)
3. CPU (fallback)
"""
global DEVICE
# --- CUDA: try PyTorch ---
if _has_torch():
import torch
if torch.cuda.is_available():
DEVICE = DeviceType.CUDA
device_name = torch.cuda.get_device_properties(0).name
logger.info(f"Hardware detected: CUDA — {device_name}")
return DEVICE
# --- MLX: Apple Silicon ---
if is_apple_silicon() and _has_mlx():
DEVICE = DeviceType.MLX
chip = platform.processor() or platform.machine()
logger.info(f"Hardware detected: MLX — Apple Silicon ({chip})")
return DEVICE
# --- Fallback ---
DEVICE = DeviceType.CPU
logger.info("Hardware detected: CPU (no GPU backend available)")
return DEVICE
# ========== Convenience helpers ==========
def get_device() -> DeviceType:
"""
Return the detected device. Auto-detects if detect_hardware() hasn't been called yet.
Prefer calling detect_hardware() explicitly at startup instead.
"""
global DEVICE
if DEVICE is None:
detect_hardware()
return DEVICE
def clear_gpu_cache():
"""
Clear GPU memory cache for the current device.
Safe to call on any platform — no-ops gracefully.
"""
import gc
gc.collect()
device = get_device()
if device == DeviceType.CUDA:
import torch
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
elif device == DeviceType.MLX:
# MLX manages memory automatically; no explicit cache clear needed.
# mlx.core has no empty_cache equivalent — gc.collect() above is enough.
pass
def get_gpu_memory_info() -> Dict[str, Any]:
"""
Get GPU memory information.
Supports CUDA (NVIDIA), MLX (Apple Silicon), and CPU-only environments.
"""
device = get_device()
# ---- CUDA path ----
if device == DeviceType.CUDA:
try:
import torch
idx = torch.cuda.current_device()
props = torch.cuda.get_device_properties(idx)
total = props.total_memory
allocated = torch.cuda.memory_allocated(idx)
reserved = torch.cuda.memory_reserved(idx)
return {
"available": True,
"backend": device.value,
"device": idx,
"device_name": props.name,
"total_gb": total / (1024**3),
"allocated_gb": allocated / (1024**3),
"reserved_gb": reserved / (1024**3),
"free_gb": (total - allocated) / (1024**3),
"utilization_pct": (allocated / total) * 100,
}
except Exception as e:
logger.error(f"Error getting CUDA GPU info: {e}")
return {"available": False, "backend": device.value, "error": str(e)}
# ---- MLX path (Apple Silicon) ----
if device == DeviceType.MLX:
try:
import mlx.core as mx
import psutil
# MLX uses unified memory — report system memory as the pool
total = psutil.virtual_memory().total
# MLX doesn't expose per-process GPU allocation; report 0 as allocated
allocated = 0
return {
"available": True,
"backend": device.value,
"device": 0,
"device_name": f"Apple Silicon ({platform.processor() or platform.machine()})",
"total_gb": total / (1024**3),
"allocated_gb": allocated / (1024**3),
"reserved_gb": 0,
"free_gb": (total - allocated) / (1024**3),
"utilization_pct": (allocated / total) * 100 if total else 0,
}
except Exception as e:
logger.error(f"Error getting MLX GPU info: {e}")
return {"available": False, "backend": device.value, "error": str(e)}
# ---- CPU-only ----
return {"available": False, "backend": "cpu"}
def log_gpu_memory(context: str):
"""Log GPU memory usage with context."""
memory_info = get_gpu_memory_info()
if memory_info.get("available"):
backend = memory_info.get("backend", "unknown").upper()
device_name = memory_info.get("device_name", "")
label = f"{backend}" + (f" ({device_name})" if device_name else "")
logger.info(
f"GPU Memory [{context}] {label}: "
f"{memory_info['allocated_gb']:.2f}GB/{memory_info['total_gb']:.2f}GB "
f"({memory_info['utilization_pct']:.1f}% used, "
f"{memory_info['free_gb']:.2f}GB free)"
)
else:
logger.info(f"GPU Memory [{context}]: No GPU available (CPU-only)")
# ========== GPU Summary & Package Versions ==========
def get_gpu_summary() -> Dict[str, Any]:
"""
Return a compact summary of the primary GPU.
Returns dict with keys:
gpu_name e.g. "NVIDIA L4" (or None)
vram_total_gb e.g. 22.17 (or None)
"""
mem = get_gpu_memory_info()
if mem.get("available"):
return {
"gpu_name": mem.get("device_name"),
"vram_total_gb": round(mem.get("total_gb", 0), 2),
}
return {"gpu_name": None, "vram_total_gb": None}
def get_package_versions() -> Dict[str, Optional[str]]:
"""
Return the installed versions of key ML packages.
Uses importlib.metadata (stdlib) so no subprocess is needed.
CUDA version comes from torch.version.cuda.
Returns dict with keys: unsloth, torch, transformers, cuda.
Missing packages yield None.
"""
from importlib.metadata import version as pkg_version, PackageNotFoundError
packages = ("unsloth", "torch", "transformers")
versions: Dict[str, Optional[str]] = {}
for name in packages:
try:
versions[name] = pkg_version(name)
except PackageNotFoundError:
versions[name] = None
# CUDA toolkit version bundled with torch
try:
import torch
versions["cuda"] = getattr(torch.version, "cuda", None)
except Exception:
versions["cuda"] = None
return versions