legen / device_utils.py
RafaG's picture
Upload 31 files
f0cf837 verified
"""Device detection utilities.
Provides a robust way to detect available accelerators (CUDA, MPS) and returns
useful metadata that can be used to pick the best compute backend.
When possible we rely on PyTorch for accurate information, falling back to
``nvidia-smi`` for a lightweight probe so that we can still inform the user
about available GPUs even when PyTorch is not ready to use them.
"""
from __future__ import annotations
from dataclasses import dataclass, field
from contextlib import contextmanager
import importlib
import math
import shutil
import subprocess
import warnings
import sys
from typing import List, Optional, Tuple
_GIB = 1024 ** 3
_MODEL_VRAM_REQUIREMENTS_GB = {
"tiny": {"int8": 0.4, "int8_float16": 0.6, "float16": 1.0, "float32": 1.8},
"tiny.en": {"int8": 0.4, "int8_float16": 0.6, "float16": 1.0, "float32": 1.8},
"base": {"int8": 0.5, "int8_float16": 0.8, "float16": 1.1, "float32": 2.0},
"base.en": {"int8": 0.5, "int8_float16": 0.8, "float16": 1.1, "float32": 2.0},
"small": {"int8": 0.9, "int8_float16": 1.3, "float16": 2.0, "float32": 3.5},
"small.en": {"int8": 0.9, "int8_float16": 1.3, "float16": 2.0, "float32": 3.5},
"medium": {"int8": 2.2, "int8_float16": 3.0, "float16": 5.0, "float32": 9.0},
"medium.en": {"int8": 2.2, "int8_float16": 3.0, "float16": 5.0, "float32": 9.0},
"large": {"int8": 3.5, "int8_float16": 4.5, "float16": 10.0, "float32": 18.0},
"large-v1": {"int8": 3.5, "int8_float16": 4.5, "float16": 10.0, "float32": 18.0},
"large-v2": {"int8": 3.5, "int8_float16": 4.5, "float16": 10.0, "float32": 18.0},
"large-v3": {"int8": 3.5, "int8_float16": 4.5, "float16": 10.0, "float32": 18.0},
"large-v3-turbo": {"int8": 2.0, "int8_float16": 2.8, "float16": 6.0, "float32": 10.0},
"turbo": {"int8": 2.0, "int8_float16": 2.8, "float16": 6.0, "float32": 10.0},
"distil-large-v2": {"int8": 2.0, "int8_float16": 2.8, "float16": 6.0, "float32": 10.0},
"distil-medium.en": {"int8": 1.1, "int8_float16": 1.6, "float16": 3.0, "float32": 5.5},
"distil-small.en": {"int8": 0.5, "int8_float16": 0.8, "float16": 1.5, "float32": 2.5},
}
_DEFAULT_MODEL_VRAM_GB = {
"int8": 1.0,
"int8_float16": 1.5,
"float16": 6.0,
"float32": 10.0,
}
_FALLBACK_MODEL_VRAM_GB = 6.0
_GPU_ONLY_COMPUTE_TYPES = {"float16", "fp16", "bfloat16", "int8_float16", "int8_bfloat16"}
_FP16_COMPUTE_TYPES = {"float16", "fp16", "bfloat16", "int8_float16", "int8_bfloat16"}
_COMPUTE_CANONICAL = {
"int8": "int8",
"int8_float16": "int8_float16",
"int8_bfloat16": "int8_float16",
"int8_float32": "float32",
"float16": "float16",
"fp16": "float16",
"bfloat16": "float16",
"float32": "float32",
"int16": "float32",
"default": "float16",
"auto": "float16",
}
_TORCH_WARNINGS_CONFIGURED = False
@dataclass
class DeviceInfo:
"""Structured information about the selected compute backend."""
backend: str
n_gpus: int = 0
gpu_names: List[str] = field(default_factory=list)
gpu_vram_bytes: List[int] = field(default_factory=list)
gpu_capabilities: List[Tuple[int, int]] = field(default_factory=list)
cuda_version: Optional[str] = None
driver_version: Optional[str] = None
messages: List[str] = field(default_factory=list)
issues: List[str] = field(default_factory=list)
notes: List[str] = field(default_factory=list)
resolved_compute_type: Optional[str] = None
selected_gpu_index: Optional[int] = None
def primary_gpu_name(self) -> Optional[str]:
if self.selected_gpu_index is not None and 0 <= self.selected_gpu_index < len(self.gpu_names):
return self.gpu_names[self.selected_gpu_index]
return self.gpu_names[0] if self.gpu_names else None
def _format_gib(byte_count: int | float | None) -> str:
if byte_count is None:
return "unknown"
return f"{byte_count / _GIB:.1f} GiB"
def _normalize_model_name(model_name: Optional[str]) -> str:
if not model_name:
return ""
return str(model_name).strip().lower()
def _canonical_compute(compute_type: Optional[str]) -> str:
key = (compute_type or "float16").lower()
return _COMPUTE_CANONICAL.get(key, "float16")
def _estimate_required_vram_bytes(model_name: Optional[str], compute_type: Optional[str]) -> int:
normalized = _normalize_model_name(model_name)
canonical = _canonical_compute(compute_type)
model_table = _MODEL_VRAM_REQUIREMENTS_GB.get(normalized)
if model_table is not None:
requirement_gb = model_table.get(canonical)
if requirement_gb is None:
requirement_gb = _DEFAULT_MODEL_VRAM_GB.get(canonical, _FALLBACK_MODEL_VRAM_GB)
else:
requirement_gb = _DEFAULT_MODEL_VRAM_GB.get(canonical, _FALLBACK_MODEL_VRAM_GB)
return int(math.ceil(requirement_gb * _GIB))
def _gpu_supports_fp16(capability: Optional[Tuple[int, int]]) -> bool:
if capability is None:
return False
major, minor = capability
if major is None or minor is None:
return False
return (major > 5) or (major == 5 and minor >= 3)
def _probe_nvidia_smi() -> Optional[DeviceInfo]:
"""Try to query ``nvidia-smi`` for GPU information."""
if shutil.which("nvidia-smi") is None:
return None
try:
out = subprocess.check_output(
[
"nvidia-smi",
"--query-gpu=name,memory.total,driver_version",
"--format=csv,noheader,nounits",
],
text=True,
stderr=subprocess.DEVNULL,
)
except Exception:
return None
lines = [line.strip() for line in out.splitlines() if line.strip()]
if not lines:
return None
gpu_names: List[str] = []
gpu_vram: List[int] = []
driver_version: Optional[str] = None
for line in lines:
parts = [part.strip() for part in line.split(",")]
if not parts:
continue
gpu_names.append(parts[0])
if len(parts) > 1:
try:
gpu_vram.append(int(float(parts[1])) * 1024 * 1024)
except (TypeError, ValueError):
gpu_vram.append(0)
if len(parts) > 2 and driver_version is None:
driver_version = parts[2]
return DeviceInfo(
backend="cuda",
n_gpus=len(gpu_names),
gpu_names=gpu_names,
gpu_vram_bytes=gpu_vram,
driver_version=driver_version,
)
def _suppress_known_torch_warnings() -> None:
"""Silence noisy torch.cuda capability warnings on older GPUs."""
global _TORCH_WARNINGS_CONFIGURED
if _TORCH_WARNINGS_CONFIGURED:
return
patterns = [
r"torch\.cuda",
r"torch\._C",
]
for module_pattern in patterns:
warnings.filterwarnings(
"ignore",
category=UserWarning,
module=module_pattern,
)
message_patterns = [
r"Found GPU\d+ .*cuda capability",
r"Please install PyTorch with a following CUDA",
r"not compatible with the current PyTorch installation",
]
for message_pattern in message_patterns:
warnings.filterwarnings(
"ignore",
category=UserWarning,
message=message_pattern,
)
_TORCH_WARNINGS_CONFIGURED = True
@contextmanager
def _suppress_torch_cuda_calls():
with warnings.catch_warnings():
warnings.filterwarnings(
"ignore",
category=UserWarning,
module=r"torch\.cuda",
)
warnings.filterwarnings(
"ignore",
category=UserWarning,
module=r"torch\._C",
)
yield
def _load_torch_module():
existing = sys.modules.get("torch")
if existing is not None:
return existing
with warnings.catch_warnings(record=True) as caught:
warnings.simplefilter("always")
module = importlib.import_module("torch")
for warning_msg in caught:
filename = getattr(warning_msg, "filename", "") or ""
normalized = filename.replace("\\", "/")
if isinstance(warning_msg.message, UserWarning) and "torch/cuda" in normalized:
continue
warnings.showwarning(
warning_msg.message,
warning_msg.category,
warning_msg.filename,
warning_msg.lineno,
)
return module
def _resolve_compute_type(
backend: str,
requested: Optional[str],
supports_fp16: bool,
auto_mode: bool,
) -> Tuple[str, List[str]]:
req = (requested or "auto").lower()
issues: List[str] = []
if req in {"auto", "default"}:
if backend == "cuda":
if not supports_fp16:
issues.append("FP16 may be unsupported on this GPU; it could run slower or fail. Consider int8_float16 if issues occur.")
return "float16", issues
if backend == "mps":
return "float16", issues
return "float32", issues
if backend == "cpu" and req in _GPU_ONLY_COMPUTE_TYPES:
replacement = "float32" if auto_mode else req
issues.append(f"{req} requires a GPU; using {replacement}.")
return replacement, issues
if backend in {"cuda", "mps"} and req in _FP16_COMPUTE_TYPES and not supports_fp16:
issues.append(f"{req} may be unsupported on detected GPU; it could run slower or fail. Consider int8_float16 or float32 if problems occur.")
return req, issues
return req, issues
def select_torch_device(
preferred: str = "auto",
*,
model_name: Optional[str] = None,
compute_type: Optional[str] = None,
) -> DeviceInfo:
"""Select and validate the best compute device."""
pref = (preferred or "auto").lower()
auto_mode = pref == "auto"
requested_compute = (compute_type or "auto").lower()
info = DeviceInfo(backend="cpu")
torch_module = None
torch_import_error = None
try:
_suppress_known_torch_warnings()
torch_module = _load_torch_module()
except Exception as exc: # pragma: no cover - depends on environment
torch_import_error = exc
cuda_available = False
cuda_device_count = 0
cuda_names: List[str] = []
cuda_vram: List[int] = []
cuda_capabilities: List[Tuple[int, int]] = []
if torch_module is not None and getattr(torch_module, "cuda", None) is not None:
try:
with _suppress_torch_cuda_calls():
cuda_available = bool(torch_module.cuda.is_available())
except Exception:
cuda_available = False
try:
with _suppress_torch_cuda_calls():
cuda_device_count = int(torch_module.cuda.device_count())
except Exception:
cuda_device_count = 0
if cuda_available and cuda_device_count:
for idx in range(cuda_device_count):
name = None
total_mem = None
capability = None
try:
with _suppress_torch_cuda_calls():
props = torch_module.cuda.get_device_properties(idx)
except Exception:
props = None
if props is not None:
name = getattr(props, "name", None)
total_mem = getattr(props, "total_memory", None)
capability = (
getattr(props, "major", None),
getattr(props, "minor", None),
)
if name is None:
try:
name = torch_module.cuda.get_device_name(idx)
except Exception:
name = f"CUDA GPU {idx}"
cuda_names.append(str(name))
cuda_vram.append(int(total_mem) if total_mem is not None else 0)
if capability is not None and capability[0] is not None and capability[1] is not None:
cuda_capabilities.append((int(capability[0]), int(capability[1])))
else:
cuda_capabilities.append((0, 0))
mps_available = False
if torch_module is not None:
mps_backend = getattr(getattr(torch_module, "backends", None), "mps", None)
if mps_backend is not None and hasattr(mps_backend, "is_available"):
try:
mps_available = bool(mps_backend.is_available())
except Exception:
mps_available = False
smi_info = _probe_nvidia_smi()
if not cuda_names and smi_info is not None:
cuda_names = smi_info.gpu_names
cuda_vram = smi_info.gpu_vram_bytes
info.driver_version = smi_info.driver_version
info.n_gpus = smi_info.n_gpus
if torch_module is not None:
info.cuda_version = getattr(getattr(torch_module, "version", None), "cuda", None)
if cuda_available and cuda_device_count:
info.backend = "cuda"
info.n_gpus = cuda_device_count
info.gpu_names = cuda_names
info.gpu_vram_bytes = cuda_vram
info.gpu_capabilities = cuda_capabilities
info.selected_gpu_index = 0 if cuda_device_count else None
elif pref == "cuda":
info.backend = "cpu"
info.issues.append("CUDA backend unavailable in PyTorch; using CPU.")
if torch_import_error is not None:
info.notes.append(f"PyTorch import failed: {torch_import_error}")
elif pref == "mps":
if mps_available:
info.backend = "mps"
else:
info.backend = "cpu"
info.issues.append("MPS backend unavailable; using CPU.")
elif pref == "rocm":
info.backend = "cpu"
info.issues.append("ROCm backend not implemented; using CPU.")
elif pref == "cpu":
info.backend = "cpu"
else: # auto mode
if cuda_available and cuda_device_count:
info.backend = "cuda"
info.n_gpus = cuda_device_count
info.gpu_names = cuda_names
info.gpu_vram_bytes = cuda_vram
info.gpu_capabilities = cuda_capabilities
info.selected_gpu_index = 0 if cuda_device_count else None
elif mps_available:
info.backend = "mps"
else:
info.backend = "cpu"
primary_gpu = info.primary_gpu_name()
if primary_gpu is not None:
info.messages.append(f"Detected {primary_gpu} GPU")
if info.backend == "cuda" and primary_gpu is None and cuda_names:
info.messages.append(f"Detected {cuda_names[0]} GPU")
initial_backend_for_compute = info.backend
available_vram = None
if initial_backend_for_compute == "cuda" and info.selected_gpu_index is not None:
if 0 <= info.selected_gpu_index < len(info.gpu_vram_bytes):
available_vram = info.gpu_vram_bytes[info.selected_gpu_index]
if available_vram is None and initial_backend_for_compute == "cuda" and info.gpu_vram_bytes:
available_vram = info.gpu_vram_bytes[0]
supports_fp16 = False
if initial_backend_for_compute == "cuda" and info.selected_gpu_index is not None:
idx = info.selected_gpu_index
capability = None
if 0 <= idx < len(info.gpu_capabilities):
capability = info.gpu_capabilities[idx]
supports_fp16 = _gpu_supports_fp16(capability)
elif initial_backend_for_compute == "mps":
supports_fp16 = True
resolved_compute_candidate, compute_issues = _resolve_compute_type(
initial_backend_for_compute,
requested_compute,
supports_fp16,
auto_mode,
)
requirement_bytes = None
compute_label = _canonical_compute(resolved_compute_candidate)
if initial_backend_for_compute == "cuda":
requirement_bytes = _estimate_required_vram_bytes(model_name, resolved_compute_candidate)
if (
initial_backend_for_compute == "cuda"
and available_vram is not None
and requirement_bytes is not None
and available_vram < requirement_bytes
):
message = (
f"VRAM too low for model {model_name or 'selected model'} using compute type {compute_label} (~{_format_gib(requirement_bytes)} required, found {_format_gib(available_vram)})."
)
if auto_mode:
message += " Falling back to CPU."
else:
message += " GPU execution may fail."
message += " Consider lowering the compute type (e.g. --transcription_compute_type=int8_float16) or selecting a smaller model via --transcription_model."
if auto_mode:
message += " To force GPU usage, rerun with --transcription_device=cuda."
info.issues.append(message)
if auto_mode:
info.backend = "cpu"
info.selected_gpu_index = None
if info.backend != "cuda" and auto_mode and not primary_gpu and smi_info is not None and smi_info.gpu_names:
info.messages.append("No compatible GPU ready; using CPU.")
if info.backend in {"cuda", "mps"}:
info.resolved_compute_type = resolved_compute_candidate
info.issues.extend(compute_issues)
else:
info.resolved_compute_type = None
if info.backend == "cpu" and primary_gpu is None and not info.messages:
info.messages.append("Using CPU for transcription.")
if (
info.backend == "cuda"
and not supports_fp16
and info.resolved_compute_type == "float32"
):
info.notes.append("Consider reinstalling PyTorch with newer CUDA support for FP16 acceleration.")
return info
def select_torch_device_str(
preferred: str = "auto",
*,
model_name: Optional[str] = None,
compute_type: Optional[str] = None,
) -> str:
"""Compatibility helper returning just the backend string."""
info = select_torch_device(preferred=preferred, model_name=model_name, compute_type=compute_type)
return info.backend