QPromaQ's picture
Upload folder using huggingface_hub
c289d87 verified
Raw
History Blame Contribute Delete
2.19 kB
from __future__ import annotations
import os
from dataclasses import dataclass
from typing import Any, Dict
AUTO_MINUS_FOUR = "auto-minus-4"
@dataclass(frozen=True)
class ThreadAllocation:
policy: str
system_threads: int
reserve_threads: int
threads_used: int
def detect_system_threads() -> int:
return max(1, int(os.cpu_count() or 1))
def resolve_threads_used(
configured_value: Any,
*,
reserve_threads: int = 4,
system_threads: int | None = None,
) -> ThreadAllocation:
total = max(1, int(system_threads or detect_system_threads()))
reserve = max(0, int(reserve_threads))
policy_raw = configured_value if configured_value is not None else AUTO_MINUS_FOUR
policy = str(policy_raw).strip().lower()
if policy in {"", "auto", AUTO_MINUS_FOUR}:
used = max(1, total - reserve)
return ThreadAllocation(
policy=AUTO_MINUS_FOUR,
system_threads=total,
reserve_threads=reserve,
threads_used=used,
)
try:
used = max(1, int(policy_raw))
except (TypeError, ValueError) as exc:
raise ValueError(
f"Invalid backend.parallel_jobs value `{configured_value}`. Use integer or `{AUTO_MINUS_FOUR}`."
) from exc
return ThreadAllocation(
policy="fixed",
system_threads=total,
reserve_threads=reserve,
threads_used=min(total, used),
)
def enforce_thread_fairness(config: Dict[str, Any], *, reserve_threads: int = 4) -> ThreadAllocation:
backend = config.setdefault("backend", {})
configured = backend.get("parallel_jobs", AUTO_MINUS_FOUR)
alloc = resolve_threads_used(configured, reserve_threads=reserve_threads)
backend["parallel_jobs"] = int(alloc.threads_used)
backend["thread_policy"] = alloc.policy
runtime = config.setdefault("runtime", {})
runtime["system_threads"] = int(alloc.system_threads)
runtime["reserve_threads"] = int(alloc.reserve_threads)
runtime["threads_used"] = int(alloc.threads_used)
runtime["thread_policy"] = alloc.policy
runtime["thread_formula"] = "threads_used = max(1, system_threads - 4)"
return alloc