# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

"""
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 copy
import gc
import glob
import os
import platform
import re
import subprocess
import sys
import types
from importlib.metadata import PackageNotFoundError, version as pkg_version
import structlog
from loggers import get_logger
from enum import Enum
from pathlib import Path
from typing import Optional, Dict, Any

logger = get_logger(__name__)


# ── GPU index ordering ──────────────────────────────────────────────────────
# CUDA defaults to CUDA_DEVICE_ORDER=FASTEST_FIRST, numbering GPUs by compute
# performance. nvidia-smi -- and every free-VRAM probe in Unsloth -- numbers GPUs
# by PCI bus id instead. On a mixed-GPU host (e.g. an RTX 5090 alongside an RTX
# PRO 6000) the two orderings disagree, so an index picked from nvidia-smi data
# ("the emptiest card is GPU 1") gets written into CUDA_VISIBLE_DEVICES and then
# reinterpreted by CUDA against FASTEST_FIRST -- landing the model on a different
# physical GPU than the one selected. Pinning PCI_BUS_ID makes torch, nvidia-smi,
# and CUDA_VISIBLE_DEVICES share a single index space, matching what users see in
# `nvidia-smi -L`. Set at import (before any torch.cuda call latches the order
# at context creation) and inherited by child processes, since the llama-server
# and spawn workers copy os.environ. setdefault so an explicit user override wins.
os.environ.setdefault("CUDA_DEVICE_ORDER", "PCI_BUS_ID")

# Unsloth workers can import MLX without importing unsloth first, so mirror the
# package bootstrap here. Keep an explicit user value authoritative.
if platform.system() == "Darwin" and platform.machine() == "arm64":
    os.environ.setdefault("AGX_RELAX_CDM_CTXSTORE_TIMEOUT", "1")


# ========== Device Enum ==========


class DeviceType(str, Enum):
    """Supported compute backends. str subclass for clean JSON serialization."""

    CUDA = "cuda"
    XPU = "xpu"
    MLX = "mlx"
    CPU = "cpu"


# ========== Global State (set once by detect_hardware) ==========

DEVICE: Optional[DeviceType] = None
CHAT_ONLY: bool = True  # No CUDA GPU -> GGUF chat only (Mac, CPU-only, etc.)
# Why CHAT_ONLY is True (Train/Export disabled). None when training is enabled.
# "mlx_unavailable": Apple Silicon but the MLX stack is missing, too old, or broken
# (the usual cause of "Train/Export greyed out" on Macs after a reinstall dropped MLX);
# "intel_mac": Intel Mac (no PyTorch/MLX); "no_gpu": CPU-only non-Mac host.
CHAT_ONLY_REASON: Optional[str] = None
IS_ROCM: bool = False  # True when running on AMD ROCm (HIP) -- routes GPU monitoring to amd.py


def _backend_label(device: DeviceType) -> str:
    """Return the user-facing backend name for API responses.

    ROCm hosts stay ``DeviceType.CUDA`` internally (ROCm reuses ``torch.cuda.*``),
    but "cuda" is misleading in JSON, so swap to ``"rocm"`` when ``IS_ROCM`` is set.
    """
    if IS_ROCM and device == DeviceType.CUDA:
        return "rocm"
    return device.value


# ========== Detection ==========


def is_apple_silicon() -> bool:
    """True on Apple Silicon (pure platform check, no ML imports)."""
    return platform.system() == "Darwin" and platform.machine() == "arm64"


def _has_torch() -> bool:
    """True if PyTorch is importable."""
    try:
        import torch
        return True
    except ImportError:
        return False


def _has_mlx() -> bool:
    """True if MLX is importable."""
    try:
        import mlx.core
        return True
    except ImportError:
        return False


def _has_usable_mlx_stack() -> bool:
    """True only when the FULL Unsloth MLX training/export stack is usable
    (mlx + mlx-lm + mlx-vlm at the minimum versions unsloth-zoo requires), not
    just a bare ``import mlx.core``. A backtracked/old mlx-vlm still imports but
    breaks VLM Train/Export, so the training gate must match the self-heal's own
    criterion (utils.mlx_repair.mlx_stack_available) -- otherwise detect_hardware
    would enable Train/Export on exactly the inadequate stack the MLX self-heal
    is trying to repair, leaving the user with greyed-in-but-broken buttons."""
    try:
        from utils.mlx_repair import mlx_stack_available
        return mlx_stack_available()
    except Exception as exc:
        # mlx_repair should always import; if it somehow cannot, fall back to the
        # bare import check rather than forcing a working host into chat-only.
        logger.debug("MLX stack availability check failed, using bare import: %s", exc)
        return _has_mlx()


def _print_cuda_device_list(is_rocm: bool) -> None:
    """List every visible CUDA/ROCm GPU with its index at startup.

    The "Hardware detected" banner names only device 0, which hides the other
    cards on a multi-GPU host. This lists the full visible set in CUDA-ordinal
    order, matching `nvidia-smi -L` when no CUDA_VISIBLE_DEVICES mask is set
    (under a mask the indices are visible ordinals, not physical PCI ids).
    CUDA_DEVICE_ORDER governs only CUDA, so it is shown for CUDA but not ROCm.
    No-ops on single-GPU hosts and never raises -- it is purely informational.
    """
    try:
        import torch

        count = torch.cuda.device_count()
        if count <= 1:
            return
        if is_rocm:
            header = f"ROCm devices ({count}):"
        else:
            order = os.environ.get("CUDA_DEVICE_ORDER", "default")
            header = f"CUDA devices ({count}, CUDA_DEVICE_ORDER={order}):"
        lines = [header]
        for i in range(count):
            try:
                name = torch.cuda.get_device_properties(i).name
            except Exception as e:
                logger.debug("CUDA device %d property probe failed: %s", i, e)
                name = "<unavailable>"
            lines.append(f"  [{i}] {name}")
        print("\n".join(lines))
    except Exception:
        return  # purely informational; never disrupt startup


def detect_hardware() -> DeviceType:
    """
    Detect the best compute device and set the module-level DEVICE global.

    Call once at FastAPI lifespan startup; idempotent.

    Detection order:
      1. CUDA  (NVIDIA GPU, requires torch)
      2. MLX   (Apple Silicon via MLX framework)
      3. CPU   (fallback)
    """
    global DEVICE, CHAT_ONLY, CHAT_ONLY_REASON, IS_ROCM
    CHAT_ONLY = True  # reset -- only CUDA/ROCm/XPU/MLX sets it to False
    CHAT_ONLY_REASON = None
    IS_ROCM = False

    # --- CUDA / ROCm: try PyTorch ---
    if _has_torch():
        import torch
        if torch.cuda.is_available():
            DEVICE = DeviceType.CUDA
            CHAT_ONLY = False
            try:
                device_name = torch.cuda.get_device_properties(0).name
            except Exception as e:
                logger.debug("CUDA device 0 property probe failed: %s", e)
                device_name = "<unavailable>"

            # Distinguish ROCm from CUDA for display only (DeviceType stays CUDA).
            # AMD SDK wheels don't set torch.version.hip, so fall back to __version__.
            _hip_ver = getattr(torch.version, "hip", None)
            if _hip_ver is not None or "rocm" in torch.__version__.lower():
                IS_ROCM = True
                _hip_label = _hip_ver or torch.__version__
                print(f"Hardware detected: ROCm (HIP {_hip_label}) -- {device_name}")
            else:
                print(f"Hardware detected: CUDA -- {device_name}")
            _print_cuda_device_list(IS_ROCM)
            return DEVICE

    # --- XPU: Intel GPU ---
    if _has_torch():
        import torch
        if hasattr(torch, "xpu") and torch.xpu.is_available():
            DEVICE = DeviceType.XPU
            CHAT_ONLY = False
            device_name = torch.xpu.get_device_name(0)
            print(f"Hardware detected: XPU — {device_name}")
            return DEVICE

    # --- MLX: Apple Silicon ---
    # Require the full mlx/mlx-lm/mlx-vlm stack (not a bare `import mlx.core`) so
    # the gate matches utils.mlx_repair: a partial/backtracked stack stays
    # chat-only (reason "mlx_unavailable") and the background self-heal repairs it.
    if is_apple_silicon() and _has_usable_mlx_stack():
        DEVICE = DeviceType.MLX
        CHAT_ONLY = False
        # Use platform.machine() ("arm64"); platform.processor() returns "i386"
        # on universal2 / Rosetta builds even on native arm64.
        chip = platform.machine() or "arm64"
        print(f"Hardware detected: MLX — Apple Silicon ({chip})")
        return DEVICE

    # --- Fallback ---
    DEVICE = DeviceType.CPU
    # CHAT_ONLY is still True here (every training-capable branch returned early),
    # so record WHY so the UI can explain the greyed-out Train/Export instead of
    # silently disabling them.
    if is_apple_silicon():
        # Reached the CPU fallback on Apple Silicon, so the MLX stack is missing,
        # too old, or broken. This is usually an environment problem recoverable
        # with `unsloth studio update`.
        CHAT_ONLY_REASON = "mlx_unavailable"
        logger.warning(
            "Apple Silicon detected but the MLX stack is incomplete or too old; "
            "Train/Export disabled (chat-only). Run `unsloth studio update` to "
            "restore MLX training."
        )
    elif platform.system() == "Darwin":
        CHAT_ONLY_REASON = "intel_mac"  # Intel Mac: no PyTorch/MLX -> GGUF-only by design.
    else:
        CHAT_ONLY_REASON = "no_gpu"
    print("Hardware detected: CPU (no GPU backend available)")
    return DEVICE


# ========== Convenience helpers ==========


def get_device() -> DeviceType:
    """
    Return the detected device, auto-detecting if detect_hardware() hasn't run.
    Prefer calling detect_hardware() explicitly at startup.
    """
    global DEVICE
    if DEVICE is None:
        detect_hardware()
    return DEVICE


def export_capability() -> dict:
    """Whether model export can run here, with a torch-aware reason when it cannot.

    Export runs through Unsloth, which hard-requires an accelerator (it calls ``torch.cuda`` at
    import and has no CPU path), so it is supported iff ``get_device() in {CUDA, XPU, MLX}``. The
    reason distinguishes a --no-torch install from a bare-CPU host. Safe to call without torch.

    Returns {export_supported, export_unsupported_reason, export_unsupported_message}.
    """
    if get_device() in (DeviceType.CUDA, DeviceType.XPU, DeviceType.MLX):
        return {
            "export_supported": True,
            "export_unsupported_reason": None,
            "export_unsupported_message": None,
        }
    # No accelerator: name the blocker. Apple Silicon first -- its path is MLX, so "install PyTorch"
    # would be wrong advice on a Mac even when torch is also absent.
    if is_apple_silicon():
        reason = "mlx_unavailable"
        message = (
            "Export on Apple Silicon requires the MLX stack, which is unavailable or too old. Run "
            "`unsloth studio update` to restore MLX and enable export."
        )
    elif not _has_torch():
        reason = "pytorch_not_installed"
        message = (
            "PyTorch is not installed. Model export requires PyTorch with a supported accelerator "
            "(NVIDIA, AMD, or Intel GPU) or Apple Silicon (MLX). Install PyTorch to enable export."
        )
    else:
        reason = "no_accelerator"
        message = (
            "Export requires an NVIDIA, AMD, or Intel GPU, or Apple Silicon (MLX). No supported "
            "accelerator was found on this host. (PyTorch is installed, but Unsloth cannot export "
            "on CPU only.)"
        )
    return {
        "export_supported": False,
        "export_unsupported_reason": reason,
        "export_unsupported_message": message,
    }


def clear_gpu_cache():
    """
    Clear GPU memory cache for the current device.
    Safe on any platform — no-ops gracefully.
    """
    gc.collect()

    device = get_device()

    if device == DeviceType.CUDA:
        import torch

        torch.cuda.synchronize()
        torch.cuda.empty_cache()
        torch.cuda.ipc_collect()
    elif device == DeviceType.XPU:
        import torch
        torch.xpu.synchronize()
        torch.xpu.empty_cache()
    elif device == DeviceType.MLX:
        # MLX manages memory automatically; gc.collect() above is enough.
        pass


def get_gpu_memory_info() -> Dict[str, Any]:
    """
    Get GPU memory info.
    Supports CUDA (NVIDIA), MLX (Apple Silicon), and CPU-only.
    """
    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": _backend_label(device),
                "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": _backend_label(device),
                "error": str(e),
            }

    # ---- XPU path (Intel GPU) ----
    if device == DeviceType.XPU:
        try:
            import torch

            idx = torch.xpu.current_device()
            props = torch.xpu.get_device_properties(idx)

            total = props.total_memory
            allocated = torch.xpu.memory_allocated(idx)
            reserved = torch.xpu.memory_reserved(idx)

            return {
                "available": True,
                "backend": _backend_label(device),
                "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("Error getting XPU GPU info: %s", e)
            return {
                "available": False,
                "backend": _backend_label(device),
                "error": str(e),
            }

    # ---- MLX path (Apple Silicon) ----
    if device == DeviceType.MLX:
        try:
            import mlx.core as mx
            import psutil

            # Unified memory: total = system RAM, GPU used from IORegistry AGX.
            total = psutil.virtual_memory().total
            agx = _read_apple_gpu_stats()
            allocated = agx.get("vram_used_bytes", 0) if agx else 0

            try:
                info = mx.device_info()
                # prefer machine(); processor() can return "i386" on native arm64.
                gpu_name = info.get("device_name") or platform.machine() or "arm64"
            except Exception:
                gpu_name = platform.machine() or "arm64"

            return {
                "available": True,
                "backend": _backend_label(device),
                "device": 0,
                "device_name": f"Apple Silicon ({gpu_name})",
                "total_gb": total / (1024**3),
                "allocated_gb": allocated / (1024**3),
                "reserved_gb": allocated / (1024**3),
                "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": _backend_label(device),
                "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),
            "vram_free_gb": round(mem.get("free_gb", 0), 2),
        }
    return {"gpu_name": None, "vram_total_gb": None, "vram_free_gb": None}


def get_package_versions() -> Dict[str, Optional[str]]:
    """
    Return installed versions of key ML packages.

    Uses importlib.metadata (stdlib), no subprocess. CUDA version from
    torch.version.cuda. Returns dict keyed unsloth/torch/transformers/cuda;
    missing packages yield None.
    """
    packages = ("unsloth", "torch", "transformers")
    versions: Dict[str, Optional[str]] = {}

    for name in packages:
        try:
            versions[name] = pkg_version(name)
        except PackageNotFoundError:
            versions[name] = None

    # GPU runtime version bundled with torch
    try:
        import torch
        versions["cuda"] = getattr(torch.version, "cuda", None)
        versions["rocm"] = getattr(torch.version, "hip", None)
    except Exception:
        versions["cuda"] = None
        versions["rocm"] = None

    return versions


# ========== Torch-based GPU fallbacks (AMD ROCm, Intel XPU, nvidia-smi missing) ==========


def _torch_get_device_module():
    """Return the appropriate torch device module (cuda or xpu) and its name."""
    device = get_device()
    import torch

    if device == DeviceType.CUDA:
        return torch.cuda, "cuda"
    if device == DeviceType.XPU and hasattr(torch, "xpu"):
        return torch.xpu, "xpu"
    return None, None


def _torch_get_physical_gpu_count() -> Optional[int]:
    mod, _ = _torch_get_device_module()
    if mod is None:
        return None
    try:
        return mod.device_count()
    except Exception:
        return None


def _torch_get_per_device_info(device_indices: list[int]) -> list[Dict[str, Any]]:
    """Query torch for per-GPU name, total VRAM, and used VRAM.

    ``used_gb`` is ``None`` on Windows ROCm when ``hipMemGetInfo`` reports
    ``free == total`` (ROCm/ROCm#1909): that 0 means unknown, not empty.
    """
    mod, _ = _torch_get_device_module()
    if mod is None:
        return []

    # free==total is a Windows-ROCm-only quirk.
    _win_rocm = sys.platform == "win32" and IS_ROCM
    devices = []
    for ordinal, phys_idx in enumerate(device_indices):
        try:
            # torch ordinals are 0-based relative to CUDA_VISIBLE_DEVICES.
            props = mod.get_device_properties(ordinal)
            total_bytes = props.total_memory
            used_bytes: Optional[int]
            # Prefer mem_get_info (system-wide) so auto-select sees other consumers.
            if hasattr(mod, "mem_get_info"):
                free_bytes, total_bytes = mod.mem_get_info(ordinal)
                used_bytes = total_bytes - free_bytes
                # free==total is the broken-API sentinel, not an idle GPU.
                if _win_rocm and free_bytes == total_bytes:
                    used_bytes = None
            else:
                used_bytes = mod.memory_allocated(ordinal)
            devices.append(
                {
                    "index": phys_idx,
                    "visible_ordinal": ordinal,
                    "name": props.name,
                    "total_gb": round(total_bytes / (1024**3), 2),
                    "used_gb": round(used_bytes / (1024**3), 2) if used_bytes is not None else None,
                }
            )
        except Exception as e:
            logger.debug("torch device query failed for ordinal %d: %s", ordinal, e)
    return devices


# ========== Live GPU Utilization ==========


def _smi_query(func_name: str, *args, **kwargs) -> Optional[Dict[str, Any]]:
    """Query the appropriate SMI backend (amd-smi or nvidia-smi).

    Returns the result dict if available, else None.
    """
    if IS_ROCM:
        backend_name = "amd-smi"
        try:
            from . import amd as _backend
        except Exception as e:
            logger.warning("%s import failed: %s", backend_name, e)
            return None
    else:
        backend_name = "nvidia-smi"
        try:
            from . import nvidia as _backend
        except Exception as e:
            logger.warning("%s import failed: %s", backend_name, e)
            return None
    try:
        func = getattr(_backend, func_name)
        result = func(*args, **kwargs)
        if isinstance(result, dict) and result.get("available"):
            return result
    except Exception as e:
        logger.warning("%s %s query failed: %s", backend_name, func_name, e)
    return None


def _read_apple_gpu_stats() -> Dict[str, Any]:
    """Query macOS IORegistry for AGX (Apple GPU) live stats. No sudo needed.

    Returns dict with utilization_pct, vram_used_bytes (system-wide GPU
    memory), or empty dict on failure.
    """
    try:
        result = subprocess.run(
            ["ioreg", "-r", "-c", "AGXAccelerator"],
            capture_output = True,
            timeout = 2,
        )
        text = result.stdout.decode("utf-8", errors = "replace")
    except Exception:
        return {}

    # PerformanceStatistics block has GPU utilization and in-use memory
    m = re.search(r'"PerformanceStatistics" = \{([^}]+)\}', text)
    if not m:
        return {}
    stats_str = m.group(1)
    pairs = re.findall(r'"([^"]+)"=(\d+)', stats_str)
    stats = {k: int(v) for k, v in pairs}

    return {
        "utilization_pct": stats.get("Device Utilization %", 0),
        "vram_used_bytes": stats.get("In use system memory", 0),
    }


def _rocm_linux_sysfs_gpu_busy_pct() -> Optional[float]:
    """Query AMD GPU compute utilization via Linux DRM sysfs gpu_busy_percent."""
    if platform.system() != "Linux":
        return None
    try:
        files = glob.glob("/sys/class/drm/card*/device/gpu_busy_percent")
        if not files:
            return None
        values = [int(open(f).read().strip()) for f in files]
        return round(sum(values) / len(values), 1)
    except Exception:
        return None


def _rocm_linux_sysfs_temp_c() -> Optional[float]:
    """Query AMD GPU edge temperature via Linux DRM hwmon sysfs (temp1_input, millidegrees C)."""
    if platform.system() != "Linux":
        return None
    try:
        files = glob.glob("/sys/class/drm/card*/device/hwmon/hwmon*/temp1_input")
        if not files:
            return None
        temps = [int(open(f).read().strip()) / 1000.0 for f in files]
        return round(max(temps), 1)
    except Exception:
        return None


def _rocm_linux_sysfs_power_w() -> Optional[float]:
    """Query AMD GPU average power draw via Linux DRM hwmon sysfs (microwatts)."""
    if platform.system() != "Linux":
        return None
    try:
        for pattern in (
            "/sys/class/drm/card*/device/hwmon/hwmon*/power1_average",
            "/sys/class/drm/card*/device/hwmon/hwmon*/power1_input",
        ):
            files = glob.glob(pattern)
            if files:
                watts = sum(int(open(f).read().strip()) / 1_000_000.0 for f in files)
                return round(watts, 1)
        return None
    except Exception:
        return None


def _rocm_windows_perf_counter_gpu_util_pct() -> Optional[float]:
    """Query AMD GPU compute utilization via Windows Performance Counters (3D engine nodes)."""
    if platform.system() != "Windows":
        return None
    try:
        ps = (
            "$s=(Get-Counter '\\GPU Engine(*engtype_3D*)\\Utilization Percentage'"
            " -ErrorAction SilentlyContinue).CounterSamples;"
            "if($s){[math]::Min(($s|Measure-Object CookedValue -Sum).Sum,100)}else{-1}"
        )
        r = subprocess.run(
            ["powershell", "-NoProfile", "-NonInteractive", "-Command", ps],
            capture_output = True,
            text = True,
            timeout = 5,
        )
        if r.returncode != 0 or not r.stdout.strip():
            return None
        val = float(r.stdout.strip())
        return round(val, 1) if val >= 0 else None
    except Exception:
        return None


def _rocm_linux_sysfs_vram_gb() -> tuple[Optional[float], Optional[float]]:
    """Query system-wide AMD GPU VRAM via Linux DRM sysfs.

    Reads /sys/class/drm/card*/device/mem_info_vram_*, which the kernel
    updates in real-time across all processes. No tools required.
    Returns (used_gb, total_gb) or (None, None) on failure.
    """
    if platform.system() != "Linux":
        return None, None
    try:
        used_files = glob.glob("/sys/class/drm/card*/device/mem_info_vram_used")
        total_files = glob.glob("/sys/class/drm/card*/device/mem_info_vram_total")
        if not used_files or not total_files:
            return None, None
        used_bytes = sum(int(open(f).read().strip()) for f in used_files)
        total_bytes = sum(int(open(f).read().strip()) for f in total_files)
        if total_bytes == 0:
            return None, None
        return round(used_bytes / (1024**3), 2), round(total_bytes / (1024**3), 2)
    except Exception:
        return None, None


# 0x1002. NVIDIA's open kernel module also registers KFD nodes (vendor_id 0x10DE);
# a non-AMD node is not a HIP device and must never take an ordinal.
_AMD_PCI_VENDOR_ID = 4098


def _rocm_kfd_gpu_pci_ids() -> list[str]:
    """PCI addresses of the GPUs ROCm enumerates, in HIP device order.

    Reads /sys/class/kfd/kfd/topology/nodes/<N>/properties, the topology ROCm
    itself enumerates from: AMD GPU nodes (simd_count > 0 excludes CPUs,
    vendor_id == AMD excludes NVIDIA) in node-id order are HIP's device order, so
    position N is ROCm physical device N. Unlike DRM sysfs, an amdgpu adapter HIP
    cannot enumerate has no node here, so it never consumes an ordinal.

    Returns [] (disabling the overlay) when KFD is absent, and FAILS CLOSED the
    same way on any unreadable node or an AMD node with no location_id: dropping
    one would shift every later ordinal and let a similar-capacity GPU pass the
    total-size guard while showing another card's usage.

    location_id is the kernel's (bus << 8) | devfn; domain is separate.
    """
    nodes: list[tuple[int, str]] = []
    try:
        node_dirs = glob.glob("/sys/class/kfd/kfd/topology/nodes/*")
    except Exception:
        return []
    for node_dir in node_dirs:
        m = re.fullmatch(r".*/(\d+)", node_dir)
        if m is None:
            continue
        props: dict[str, int] = {}
        try:
            with open(os.path.join(node_dir, "properties")) as f:
                for line in f:
                    parts = line.split()
                    if len(parts) == 2:
                        try:
                            props[parts[0]] = int(parts[1])
                        except ValueError:
                            continue
        except OSError:
            return []  # unreadable node could be a GPU: fail closed, don't shift
        if props.get("simd_count", 0) <= 0:
            continue  # CPU node, not a GPU
        if props.get("vendor_id") != _AMD_PCI_VENDOR_ID:
            continue  # non-AMD GPU node (NVIDIA open driver): not a HIP device
        location_id = props.get("location_id")
        if location_id is None:
            return []  # an AMD GPU we cannot place: fail closed for the whole map
        domain = props.get("domain", 0)
        bus = (location_id >> 8) & 0xFF
        devfn = location_id & 0xFF
        bdf = f"{domain:04x}:{bus:02x}:{(devfn >> 3) & 0x1F:02x}.{devfn & 0x7}"
        nodes.append((int(m.group(1)), bdf))
    nodes.sort(key = lambda n: n[0])
    return [bdf for _node_id, bdf in nodes]


def _rocm_linux_amdgpu_cards() -> list[tuple[str, int, str]]:
    """The amdgpu-bound DRM cards in PCI order: ``(pci_bdf, card_no, device_dir)``.

    Membership is by the BOUND DRIVER, not the VRAM sysfs files: an AMD device
    with incomplete sysfs support (some APUs expose no mem_info_vram_*) still
    consumes a ROCm ordinal, and dropping it would shift every later card down.
    PCI order is HIP's default enumeration order, so list position is the ROCm
    ordinal; card_no is a stable tiebreak when the BDF cannot be resolved.

    NOTE this is a superset of the ROCm-visible set (a HIP-unsupported amdgpu
    adapter appears too), so callers must check the counts agree before assuming
    a 1:1 mapping onto torch devices.
    """
    if platform.system() != "Linux":
        return []
    amd_cards: list[tuple[str, int, str]] = []
    try:
        for card_path in glob.glob("/sys/class/drm/card*"):
            # Match card<N> exactly so connector nodes (card0-DP-1) are skipped.
            m = re.fullmatch(r".*/card(\d+)", card_path)
            if m is None:
                continue
            dev_dir = os.path.join(card_path, "device")
            try:
                driver = os.path.basename(os.path.realpath(os.path.join(dev_dir, "driver")))
            except OSError:
                continue
            if driver != "amdgpu":
                continue  # foreign adapter: not a ROCm device, takes no ordinal
            try:
                bdf = os.path.basename(os.path.realpath(dev_dir))
            except OSError:
                bdf = ""
            amd_cards.append((bdf, int(m.group(1)), dev_dir))
    except Exception:
        return []
    amd_cards.sort(key = lambda c: (c[0], c[1]))
    return amd_cards


def _rocm_linux_sysfs_vram_by_pci_gb() -> dict[str, tuple[float, float]]:
    """System-wide AMD VRAM via Linux DRM sysfs, keyed by the card's PCI address.

    Reads each card's mem_info_vram_{used,total} (kernel-updated across all
    processes) so every GPU gets its own figure, unlike _rocm_linux_sysfs_vram_gb
    which sums the host. Keyed by PCI address, not an ordinal, so the caller can
    join it to _rocm_kfd_gpu_pci_ids() by identity: DRM card numbers include
    foreign adapters and this set includes cards HIP does not enumerate, so any
    ordinal from this list alone can be shifted relative to ROCm's. A card with
    missing/unreadable/zero-total figures simply has no entry. Empty off Linux.
    """
    if platform.system() != "Linux":
        return {}

    try:
        by_pci: dict[str, tuple[float, float]] = {}
        for bdf, _card_no, dev_dir in _rocm_linux_amdgpu_cards():
            if not bdf:
                continue
            try:
                with open(os.path.join(dev_dir, "mem_info_vram_used")) as f:
                    used_bytes = int(f.read().strip())
                with open(os.path.join(dev_dir, "mem_info_vram_total")) as f:
                    total_bytes = int(f.read().strip())
            except (OSError, ValueError):
                continue
            if total_bytes <= 0:
                continue
            by_pci[bdf.lower()] = (
                round(used_bytes / (1024**3), 2),
                round(total_bytes / (1024**3), 2),
            )
        return by_pci
    except Exception:
        return {}


# ── Windows AMD/ROCm per-adapter VRAM (issue #7072) ──────────────────────────
# amd-smi is disabled and hipMemGetInfo reports free==total, so read used from the
# per-LUID "GPU Adapter Memory" perf counters and take each total from torch, so
# every GPU shows instead of one fake device with GPU 0's total.
# Placeholder adapters (Basic Render Driver / idle iGPU) drop only when they would
# outnumber the real torch devices.
_ROCM_WIN_ADAPTER_MIN_BYTES = 64 * 1024 * 1024  # 64 MiB


def _rocm_windows_perf_counter_vram_by_adapter() -> Optional[list[tuple[str, float]]]:
    """Per-adapter dedicated VRAM usage on Windows via Performance Counters.

    Returns ``[(instance_name, used_bytes)]`` (one per LUID-named adapter), or
    ``None`` when the counter is unavailable/localized/empty so callers fall back.
    """
    if platform.system() != "Windows":
        return None
    try:
        # Emit "<InstanceName>|<CookedValue>" per sample, or a __NONE__ sentinel.
        ps = (
            "$s=(Get-Counter '\\GPU Adapter Memory(*)\\Dedicated Usage'"
            " -ErrorAction SilentlyContinue).CounterSamples;"
            "if($s){$s|ForEach-Object{'{0}|{1}' -f $_.InstanceName,[int64]$_.CookedValue}}"
            "else{'__NONE__'}"
        )
        r = subprocess.run(
            ["powershell", "-NoProfile", "-NonInteractive", "-Command", ps],
            capture_output = True,
            text = True,
            timeout = 5,
        )
        if r.returncode != 0 or not r.stdout.strip():
            return None
        adapters: list[tuple[str, float]] = []
        for line in r.stdout.splitlines():
            line = line.strip()
            if not line or line == "__NONE__" or "|" not in line:
                continue
            instance, _, raw = line.rpartition("|")
            try:
                used = float(raw.strip())
            except (ValueError, TypeError):
                continue
            if used < 0:
                continue
            adapters.append((instance.strip(), used))
        return adapters or None
    except Exception:
        return None


def _match_adapter_used_to_devices(
    adapter_useds: list[float], device_totals: list[float]
) -> list[Optional[float]]:
    """Attribute per-adapter used bytes to torch devices by capacity ranking.

    Windows shares no key between LUID counters and torch ordinals, so usages are
    ranked against device totals and each is trusted only when capacity *forces* it
    (it exceeds every smaller device); an ambiguous ranking reports unknown
    (``None``) rather than fabricate a per-index free.

    Extra counters mean a hidden/display adapter, and the noise filter may have
    dropped a real reading, so values are emitted only when the supra-threshold
    counters number EXACTLY the visible devices AND capacity forces the mapping;
    otherwise every device is unknown. Best-effort but correct for the common
    loaded-card case (#7072). Returns a list aligned to ``device_totals``.
    """
    n = len(device_totals)
    if n == 0:
        return []
    useds = sorted(adapter_useds, reverse = True)
    ranked_positions = sorted(range(n), key = lambda i: -device_totals[i])
    ranked_totals = [device_totals[pos] for pos in ranked_positions]
    assigned: list[Optional[float]]
    # More counters than devices -> a hidden/display adapter (check before noise filter).
    if len(useds) > n:
        non_trivial = [u for u in useds if u >= _ROCM_WIN_ADAPTER_MIN_BYTES]
        if len(non_trivial) != n:
            # Not a clean bijection (a masked GPU is busy or a visible card idle):
            # no counter maps to a specific card, so report unknown.
            return [None] * n
        # Exactly n supra-threshold counters: extras were placeholders, so a
        # capacity-ranked bijection is plausible.
        useds = non_trivial
        ranked_useds = [useds[rank] for rank in range(n)]
        # A usage above its ranked capacity is a hidden larger GPU; clamping onto the
        # smaller card would fabricate a fully-used reading.
        for rank in range(n):
            if ranked_useds[rank] > ranked_totals[rank]:
                return [None] * n
        # Capacity forces the mapping only when the usage exceeds the next-smaller
        # capacity; the smallest card and merely-fitting usages stay unknown.
        # Keeps 40 GiB over 48/8 GiB -> [40, None].
        assigned = [None] * n
        for rank, pos in enumerate(ranked_positions):
            if rank + 1 < n and ranked_useds[rank] > ranked_totals[rank + 1]:
                assigned[pos] = min(ranked_useds[rank], device_totals[pos])
        return assigned
    # No hidden adapters: every counter is a visible card, so ranking is a permutation.
    ranked_useds = [useds[rank] if rank < len(useds) else 0.0 for rank in range(n)]
    # Ambiguous if a strictly larger usage also fits the next smaller card: the two
    # could be swapped without breaking capacity, so ranking can't tell them apart.
    for rank in range(n - 1):
        upper, lower = ranked_useds[rank], ranked_useds[rank + 1]
        if upper > lower and upper <= ranked_totals[rank + 1]:
            return [None] * n
    assigned = [None] * n
    for rank, pos in enumerate(ranked_positions):
        if rank < len(useds):
            assigned[pos] = min(useds[rank], device_totals[pos])
    return assigned


def _rocm_windows_per_device_vram(device_indices: list[int]) -> list[Dict[str, Any]]:
    """Per-GPU VRAM on Windows AMD/ROCm: total from torch properties (reliable),
    used from the per-adapter Dedicated Usage counter.

    Returns ``{index, visible_ordinal, name, used_gb, total_gb}`` per visible GPU
    (``used_gb`` may be ``None`` when the counter is unavailable), or ``[]`` when
    torch can't enumerate devices so callers fall through to the torch last resort.
    """
    if platform.system() != "Windows":
        return []
    mod, _ = _torch_get_device_module()
    if mod is None:
        return []
    # Totals/names from torch properties (mem_get_info's free==total quirk zeroes used).
    dev_meta: list[Dict[str, Any]] = []
    for ordinal, phys_idx in enumerate(device_indices):
        try:
            props = mod.get_device_properties(ordinal)
            dev_meta.append(
                {
                    "index": phys_idx,
                    "visible_ordinal": ordinal,
                    "name": props.name,
                    "total_bytes": int(props.total_memory),
                }
            )
        except Exception as e:
            logger.debug("torch property probe failed for ordinal %d: %s", ordinal, e)
    if not dev_meta:
        return []

    adapters = _rocm_windows_perf_counter_vram_by_adapter()
    if adapters:
        assigned = _match_adapter_used_to_devices(
            [used for _, used in adapters],
            [d["total_bytes"] for d in dev_meta],
        )
    else:
        # Counter unavailable: show every GPU with a correct total, used unknown.
        assigned = [None] * len(dev_meta)

    devices: list[Dict[str, Any]] = []
    for meta, used_bytes in zip(dev_meta, assigned):
        total_gb = round(meta["total_bytes"] / (1024**3), 2)
        used_gb = round(used_bytes / (1024**3), 2) if used_bytes is not None else None
        devices.append(
            {
                "index": meta["index"],
                "visible_ordinal": meta["visible_ordinal"],
                "name": meta["name"],
                "used_gb": used_gb,
                "total_gb": total_gb,
            }
        )
    return devices


def _rocm_windows_device_payload_entry(
    device: DeviceType, dev: Dict[str, Any], gpu_util_pct: Optional[float]
) -> Dict[str, Any]:
    """Build a ``get_gpu_utilization`` device entry from a per-device VRAM dict."""
    total_gb = dev["total_gb"]
    used_gb = dev["used_gb"]
    return {
        "available": True,
        "backend": _backend_label(device),
        "index": dev["index"],
        "visible_ordinal": dev["visible_ordinal"],
        "name": dev.get("name", "Unknown"),
        "gpu_utilization_pct": gpu_util_pct,
        "temperature_c": None,
        "vram_used_gb": used_gb,
        "vram_total_gb": total_gb,
        "vram_utilization_pct": round((used_gb / total_gb) * 100, 1)
        if total_gb and total_gb > 0 and used_gb is not None
        else None,
        "power_draw_w": None,
        "power_limit_w": None,
        "power_utilization_pct": None,
    }


def _gpu_utilization_payload(
    device: DeviceType, devices: list[Dict[str, Any]], **metadata: Any
) -> Dict[str, Any]:
    """Keep the legacy primary-GPU shape and append all visible devices."""
    backend = _backend_label(device)
    normalized = []
    for ordinal, raw in enumerate(devices):
        dev = dict(raw)
        dev.setdefault("available", True)
        dev.setdefault("backend", backend)
        if dev.get("visible_ordinal") is None:
            dev["visible_ordinal"] = ordinal
        normalized.append(dev)

    normalized.sort(key = lambda dev: dev.get("visible_ordinal", dev.get("index", 0)))
    payload: Dict[str, Any] = {
        "available": bool(normalized),
        "backend": backend,
        "devices": normalized,
    }
    payload.update(metadata)
    if normalized:
        payload.update(normalized[0])
        payload["available"] = True
        payload["backend"] = normalized[0].get("backend", backend)
        payload["devices"] = normalized
    return payload


def get_gpu_utilization() -> Dict[str, Any]:
    """Live utilization snapshot for the primary GPU plus all visible GPUs."""
    device = get_device()

    if device == DeviceType.XPU:
        result = get_visible_gpu_utilization()
        return _gpu_utilization_payload(
            device,
            result.get("devices", []),
            parent_visible_gpu_ids = result.get("parent_visible_gpu_ids", []),
            index_kind = result.get("index_kind"),
        )

    if device == DeviceType.CUDA:
        parent_visible_spec = _get_parent_visible_gpu_spec()
        result = _smi_query(
            "get_visible_gpu_utilization",
            parent_visible_spec["numeric_ids"],
            parent_cuda_visible_devices = parent_visible_spec["raw"],
        )
        if result is not None and "devices" in result:
            devices = result["devices"]
            numeric_ids = parent_visible_spec.get("numeric_ids")
            if IS_ROCM and numeric_ids is not None:
                _reconcile_rocm_unified_memory(result, numeric_ids)

            return _gpu_utilization_payload(
                device,
                devices,
                backend_cuda_visible_devices = result.get("backend_cuda_visible_devices"),
                parent_visible_gpu_ids = result.get("parent_visible_gpu_ids", []),
                index_kind = result.get("index_kind"),
            )

        # Fallback Windows ROCm: per-adapter VRAM attribution (issue #7072), so
        # every visible GPU is shown instead of a sum collapsed onto one device.
        if IS_ROCM and platform.system() == "Windows":
            _win_ids = _get_parent_visible_gpu_spec().get("numeric_ids")
            if not _win_ids:
                _win_ids = list(range(_torch_get_physical_gpu_count() or 0))
            _win_devices = _rocm_windows_per_device_vram(_win_ids)
            if _win_devices:
                # A single visible GPU can own the aggregate 3D-engine utilization;
                # across several GPUs the sum isn't per-device, so leave it unset.
                _win_util = (
                    _rocm_windows_perf_counter_gpu_util_pct() if len(_win_devices) == 1 else None
                )
                return _gpu_utilization_payload(
                    device,
                    [
                        _rocm_windows_device_payload_entry(device, _wd, _win_util)
                        for _wd in _win_devices
                    ],
                )

        # Fallback Linux ROCm
        if IS_ROCM and platform.system() == "Linux":
            _linux_used, _linux_total = _rocm_linux_sysfs_vram_gb()
            if _linux_used is not None and _linux_total is not None:
                _linux_util = _rocm_linux_sysfs_gpu_busy_pct()
                _linux_temp = _rocm_linux_sysfs_temp_c()
                _linux_power = _rocm_linux_sysfs_power_w()
                return _gpu_utilization_payload(
                    device,
                    [
                        {
                            "available": True,
                            "backend": _backend_label(device),
                            "index": 0,
                            "visible_ordinal": 0,
                            "gpu_utilization_pct": _linux_util,
                            "temperature_c": _linux_temp,
                            "vram_used_gb": _linux_used,
                            "vram_total_gb": _linux_total,
                            "vram_utilization_pct": round((_linux_used / _linux_total) * 100, 1)
                            if _linux_total > 0
                            else None,
                            "power_draw_w": _linux_power,
                            "power_limit_w": None,
                            "power_utilization_pct": None,
                        }
                    ],
                )

        # Last resort: torch mem_get_info (process-local) for all visible GPUs
        _visible_spec = _get_parent_visible_gpu_spec()
        _numeric_ids = _visible_spec.get("numeric_ids") or []
        if not _numeric_ids:
            visible_count = _torch_get_physical_gpu_count() or 0
            _numeric_ids = list(range(visible_count))

        _torch_devices = _torch_get_per_device_info(_numeric_ids)
        if _torch_devices:
            gpu_array = []
            for _td in _torch_devices:
                _total = _td["total_gb"]
                _used = _td["used_gb"]
                gpu_array.append(
                    {
                        "available": True,
                        "backend": _backend_label(device),
                        "index": _td["index"],
                        "name": _td.get("name", "Unknown"),
                        "gpu_utilization_pct": None,
                        "temperature_c": None,
                        "vram_used_gb": _used,
                        "vram_total_gb": _total,
                        "vram_utilization_pct": round((_used / _total) * 100, 1)
                        if _total > 0 and _used is not None
                        else None,
                        "power_draw_w": None,
                        "power_limit_w": None,
                        "power_utilization_pct": None,
                    }
                )
            return _gpu_utilization_payload(device, gpu_array)

    # MLX
    if device == DeviceType.MLX:
        try:
            import psutil
            agx = _read_apple_gpu_stats()
            total_bytes = psutil.virtual_memory().total
        except Exception as e:
            logger.error(f"Error getting MLX GPU utilization: {e}")
            return {"available": False, "backend": device.value, "devices": [], "error": str(e)}

        allocated_bytes = agx.get("vram_used_bytes", 0) or 0
        vram_used_gb = allocated_bytes / (1024**3)
        total_gb = total_bytes / (1024**3)

        try:
            from core.training import get_training_backend

            tb = get_training_backend()
            tb_progress = getattr(tb, "_progress", None)
            if tb_progress is not None and getattr(tb_progress, "is_training", False):
                tb_peak = getattr(tb_progress, "peak_memory_gb", None)
                if tb_peak is not None and tb_peak > 0:
                    vram_used_gb = float(tb_peak)
        except Exception:
            pass

        from . import apple

        return _gpu_utilization_payload(
            device,
            [
                {
                    "available": True,
                    "backend": device.value,
                    "index": 0,
                    "visible_ordinal": 0,
                    "gpu_utilization_pct": agx.get("utilization_pct") if agx else None,
                    "temperature_c": apple.read_gpu_temperature_c(),
                    "vram_used_gb": round(vram_used_gb, 2),
                    "vram_total_gb": round(total_gb, 2),
                    "vram_utilization_pct": round((vram_used_gb / total_gb) * 100, 1)
                    if total_gb > 0
                    else None,
                    "power_draw_w": apple.read_gpu_power_w(),
                    "power_limit_w": None,
                    "power_utilization_pct": None,
                }
            ],
        )

    mem = get_gpu_memory_info()
    if device != DeviceType.CPU and mem.get("available"):
        return _gpu_utilization_payload(
            device,
            [
                {
                    "available": True,
                    "backend": _backend_label(device),
                    "index": mem.get("device", 0),
                    "visible_ordinal": 0,
                    "gpu_utilization_pct": None,
                    "temperature_c": None,
                    "vram_used_gb": round(mem.get("allocated_gb", 0), 2),
                    "vram_total_gb": round(mem.get("total_gb", 0), 2),
                    "vram_utilization_pct": round(mem.get("utilization_pct", 0), 1),
                    "power_draw_w": None,
                    "power_limit_w": None,
                    "power_utilization_pct": None,
                }
            ],
        )

    return {"available": False, "backend": _backend_label(device), "devices": []}


def _apply_unified_memory_correction(
    device_metrics: Dict[str, Any], torch_info: Dict[str, Any]
) -> None:
    """Per-device reconciliation: when torch reports a larger memory total
    than amd-smi, overwrite the smi VRAM fields in place.

    Used by both the multi-device and primary-device reconcilers so the two
    endpoints stay in sync on AMD iGPUs with unified memory.
    """
    torch_total_gb = torch_info["total_gb"]
    torch_used_gb = torch_info.get("used_gb")
    smi_total_gb = device_metrics.get("vram_total_gb") or 0.0
    # torch sees the full unified (GTT) pool; amd-smi only the dedicated carve-out.
    # Adopt torch's larger total regardless of used: on Windows ROCm torch_used is
    # None (free==total sentinel) but its total stays authoritative. Overwrite used
    # only when torch's is known, then recompute utilization against whatever remains.
    if torch_total_gb > smi_total_gb:
        device_metrics["vram_total_gb"] = torch_total_gb
        if torch_used_gb is not None:
            device_metrics["vram_used_gb"] = torch_used_gb
        _used_for_pct = device_metrics.get("vram_used_gb")
        device_metrics["vram_utilization_pct"] = (
            round((_used_for_pct / torch_total_gb) * 100, 1)
            if torch_total_gb > 0 and _used_for_pct is not None
            else None
        )
        logger.debug(
            "ROCm unified memory: adopted torch mem_get_info total (%.2f GB) over "
            "amd-smi (%.2f GB) for device %s",
            torch_total_gb,
            smi_total_gb,
            torch_info.get("index"),
        )


def _reconcile_rocm_unified_memory(utilization: Dict[str, Any], device_indices: list[int]) -> None:
    """Fix amd-smi VRAM for ROCm unified-memory GPUs (e.g. Strix Halo).

    amd-smi reports only the dedicated slice; torch sees the full GTT pool. When
    torch total > smi total, overwrite per-device VRAM fields with the real value.
    """
    torch_devices = _torch_get_per_device_info(device_indices)
    if not torch_devices:
        return
    torch_by_index = {td["index"]: td for td in torch_devices}
    for dev in utilization.get("devices", []):
        td = torch_by_index.get(dev.get("index"))
        if td is None:
            continue
        _apply_unified_memory_correction(dev, td)


def _reconcile_primary_rocm_unified_memory(
    utilization: Dict[str, Any], parent_visible_spec: Dict[str, Any]
) -> None:
    """Same fix as _reconcile_rocm_unified_memory for the flat primary-GPU dict."""
    numeric_ids = parent_visible_spec.get("numeric_ids")
    if numeric_ids is None:
        # No visibility env var set: torch ordinal 0 is the primary device.
        primary_idx = [0]
    elif len(numeric_ids) == 0:
        # Empty mask: no GPU visible. Querying torch device 0 would raise or
        # return stale data, so bail rather than write bad values.
        return
    else:
        primary_idx = [int(numeric_ids[0])]
    torch_devices = _torch_get_per_device_info(primary_idx)
    if not torch_devices:
        return
    _apply_unified_memory_correction(utilization, torch_devices[0])


def _rocm_visibility_mask_active() -> bool:
    """True when any ROCm/CUDA visibility variable filters the device set."""
    for var in (
        "HIP_VISIBLE_DEVICES",
        "ROCR_VISIBLE_DEVICES",
        "CUDA_VISIBLE_DEVICES",
        "GPU_DEVICE_ORDINAL",
    ):
        value = os.environ.get(var)
        if value and value.strip():
            return True
    return False


def _overlay_system_wide_vram(devices: list[Dict[str, Any]]) -> None:
    """Replace process-local torch VRAM with system-wide Linux ROCm figures.

    The torch fallback is process-local, so a model served by the separate
    llama-server process reads as ~0 used even with the GPU full (#7072). DRM
    sysfs gives per-card figures the kernel updates across all processes. Sources
    are matched by the device's PHYSICAL index (never list position), and only
    when NO visibility mask is active and the device count equals the host GPU
    count; under any mask the index is not a verifiable host ordinal, so torch's
    figures are kept. Best-effort, in place: a device with no matching card, or a
    unified-memory APU whose sysfs total is below torch's GTT-backed total, keeps
    torch's (mirrors _apply_unified_memory_correction).

    Windows is intentionally not overlaid: its per-adapter perf counters cannot be
    mapped to ROCm ordinals and miss WDDM shared memory, so the multi-GPU view
    keeps torch there rather than risk misattributing another adapter's usage.
    """
    if not devices or platform.system() != "Linux":
        return
    # Match by PCI identity, never list position: index N in KFD topology is ROCm
    # physical device N and carries its PCI address, which DRM sysfs keys on too.
    # The two gates below verify ``index`` really is a host-physical ordinal
    # (torch exposes no PCI id to check directly):
    #   * No visibility mask -- any mask makes ``index`` container/ROCR-relative
    #     rather than a host ordinal.
    #   * Device count == host GPU count -- rules out a device-cgroup container
    #     that sets no env var yet compacts torch's indices from zero.
    pci_by_ordinal = _rocm_kfd_gpu_pci_ids()
    if not pci_by_ordinal:
        return
    if _rocm_visibility_mask_active() or len(devices) != len(pci_by_ordinal):
        return
    vram_by_pci = _rocm_linux_sysfs_vram_by_pci_gb()
    for dev in devices:
        index = dev.get("index")
        if not isinstance(index, int) or not (0 <= index < len(pci_by_ordinal)):
            continue
        entry = vram_by_pci.get(pci_by_ordinal[index].lower())
        if entry is None:
            continue
        used, total = entry
        dev_total = dev.get("vram_total_gb") or 0.0
        # Overlay only a device that maps 1:1 to the whole card: torch total must
        # match sysfs total within ~10%. A mismatch either way means a different
        # memory scope -- a unified-memory APU (sysfs sees only the dedicated
        # slice, torch the GTT pool) or a partitioned MI300 (sysfs reports the
        # whole card, dwarfing a partition) -- and overlaying would misstate free
        # VRAM (a partition would look like it has the whole card free).
        if dev_total <= 0 or abs(total - dev_total) > 0.1 * dev_total:
            continue
        dev["vram_used_gb"] = used
        dev["vram_total_gb"] = total
        dev["vram_utilization_pct"] = round((used / total) * 100, 1) if total > 0 else None


def get_visible_gpu_utilization() -> Dict[str, Any]:
    device = get_device()

    if device == DeviceType.CUDA:
        parent_visible_spec = _get_parent_visible_gpu_spec()
        result = _smi_query(
            "get_visible_gpu_utilization",
            parent_visible_spec["numeric_ids"],
            parent_cuda_visible_devices = parent_visible_spec["raw"],
        )
        if result is not None:
            result["backend"] = _backend_label(device)
            numeric_ids = parent_visible_spec.get("numeric_ids")
            if IS_ROCM and numeric_ids is not None:
                # Fix unified-memory VRAM on AMD iGPUs (Strix Halo etc.).
                _reconcile_rocm_unified_memory(result, numeric_ids)
            return result

        # Windows AMD/ROCm (issue #7072): the System tab's VRAM source. The torch
        # fallback below would report used==0 (free==total), so read per-adapter
        # Dedicated Usage instead; total from torch properties.
        if IS_ROCM and platform.system() == "Windows":
            win_numeric_ids = parent_visible_spec.get("numeric_ids")
            if win_numeric_ids:
                win_ids = win_numeric_ids
                win_index_kind = "physical"
            else:
                win_ids = list(range(_torch_get_physical_gpu_count() or 0))
                win_index_kind = "relative"
            win_devices = _rocm_windows_per_device_vram(win_ids)
            if win_devices:
                devices = []
                for wd in win_devices:
                    total = wd["total_gb"]
                    used = wd["used_gb"]
                    devices.append(
                        {
                            "index": wd["index"],
                            "index_kind": win_index_kind,
                            "visible_ordinal": wd["visible_ordinal"],
                            "name": wd.get("name"),
                            "gpu_utilization_pct": None,
                            "temperature_c": None,
                            "vram_used_gb": used,
                            "vram_total_gb": total,
                            "vram_utilization_pct": round((used / total) * 100, 1)
                            if total and total > 0 and used is not None
                            else None,
                            "power_draw_w": None,
                            "power_limit_w": None,
                            "power_utilization_pct": None,
                        }
                    )
                return {
                    "available": True,
                    "backend": _backend_label(device),
                    "parent_visible_gpu_ids": win_numeric_ids or [],
                    "devices": devices,
                    "index_kind": win_index_kind,
                }

    # Torch-based fallback for CUDA (nvidia-smi unavailable, AMD ROCm) and XPU (Intel)
    if device in (DeviceType.CUDA, DeviceType.XPU):
        parent_ids = get_parent_visible_gpu_ids()
        # Empty parent_ids (UUID/MIG mask or no CVD): enumerate torch ordinals.
        if parent_ids:
            torch_indices = parent_ids
            index_kind = "physical"
        else:
            visible_count = _torch_get_physical_gpu_count() or 0
            torch_indices = list(range(visible_count))
            index_kind = "relative"
        torch_devices = _torch_get_per_device_info(torch_indices)
        if torch_devices:
            devices = []
            for td in torch_devices:
                total = td["total_gb"]
                used = td["used_gb"]
                devices.append(
                    {
                        "index": td["index"],
                        "index_kind": index_kind,
                        "visible_ordinal": td["visible_ordinal"],
                        "gpu_utilization_pct": None,
                        "temperature_c": None,
                        "vram_used_gb": used,
                        "vram_total_gb": total,
                        "vram_utilization_pct": round((used / total) * 100, 1)
                        if total > 0 and used is not None
                        else None,
                        "power_draw_w": None,
                        "power_limit_w": None,
                        "power_utilization_pct": None,
                    }
                )
            if IS_ROCM and index_kind == "physical":
                # Swap process-local torch VRAM for system-wide sysfs so a model
                # held by the separate llama-server process shows up (#7072).
                # Physical-index only: a relative index (UUID/MIG mask) is not a
                # host GPU id. The overlay verifies the rest itself.
                _overlay_system_wide_vram(devices)
            return {
                "available": True,
                "backend": _backend_label(device),
                "parent_visible_gpu_ids": parent_ids,
                "devices": devices,
                "index_kind": index_kind,
            }

    if device == DeviceType.MLX:
        mem = get_gpu_memory_info()
        if not mem.get("available"):
            return {
                "available": False,
                "backend": _backend_label(device),
                "parent_visible_gpu_ids": [],
                "devices": [],
                "index_kind": "relative",
            }
        return {
            "available": True,
            "backend": _backend_label(device),
            "parent_visible_gpu_ids": [0],
            "devices": [
                {
                    "index": 0,
                    "index_kind": "relative",
                    "visible_ordinal": 0,
                    "gpu_utilization_pct": None,
                    "temperature_c": None,
                    "vram_used_gb": round(mem.get("allocated_gb", 0), 2),
                    "vram_total_gb": round(mem.get("total_gb", 0), 2),
                    "vram_utilization_pct": round(mem.get("utilization_pct", 0), 1),
                    "power_draw_w": None,
                    "power_limit_w": None,
                    "power_utilization_pct": None,
                }
            ],
            "index_kind": "relative",
        }

    return {
        "available": False,
        "backend": _backend_label(device),
        "parent_visible_gpu_ids": [],
        "devices": [],
        "index_kind": "relative",
    }


# ========== Multi-GPU Detection & Safe num_proc ==========

_physical_gpu_count: Optional[int] = None
_visible_gpu_count: Optional[int] = None


def _get_parent_visible_gpu_spec() -> Dict[str, Any]:
    # ROCm uses HIP/ROCR_VISIBLE_DEVICES on top of CUDA_VISIBLE_DEVICES; check
    # them first. Explicit None checks (not `or`) so "" reads as "no visible GPUs".
    cuda_visible = None
    # Prefer ROCm masks only on a ROCm host or when no CUDA mask is set, so a
    # stale HIP_VISIBLE_DEVICES on NVIDIA can't override CUDA_VISIBLE_DEVICES.
    _is_rocm_spec = IS_ROCM or (
        "CUDA_VISIBLE_DEVICES" not in os.environ
        and ("HIP_VISIBLE_DEVICES" in os.environ or "ROCR_VISIBLE_DEVICES" in os.environ)
    )
    if _is_rocm_spec:
        hip_vis = os.environ.get("HIP_VISIBLE_DEVICES")
        rocr_vis = os.environ.get("ROCR_VISIBLE_DEVICES")
        if hip_vis is not None:
            cuda_visible = hip_vis
        elif rocr_vis is not None:
            cuda_visible = rocr_vis
    if cuda_visible is None:
        cuda_visible = os.environ.get("CUDA_VISIBLE_DEVICES")

    if cuda_visible is None:
        return {
            "raw": None,
            "numeric_ids": list(range(get_physical_gpu_count())),
            "supports_explicit_gpu_ids": True,
        }

    cuda_visible = cuda_visible.strip()
    if cuda_visible == "" or cuda_visible == "-1":
        return {
            "raw": cuda_visible,
            "numeric_ids": [],
            "supports_explicit_gpu_ids": True,
        }

    tokens = [value.strip() for value in cuda_visible.split(",") if value.strip()]
    try:
        numeric_ids = [int(value) for value in tokens]
    except ValueError:
        return {
            "raw": cuda_visible,
            "numeric_ids": None,
            "supports_explicit_gpu_ids": False,
        }

    return {
        "raw": cuda_visible,
        "numeric_ids": numeric_ids,
        "supports_explicit_gpu_ids": True,
    }


def get_parent_visible_gpu_ids() -> list[int]:
    parent_visible_ids = _get_parent_visible_gpu_spec()["numeric_ids"]
    return list(parent_visible_ids) if parent_visible_ids is not None else []


def resolve_requested_gpu_ids(gpu_ids: Optional[list[int]]) -> list[int]:
    parent_visible_spec = _get_parent_visible_gpu_spec()
    parent_visible_ids = get_parent_visible_gpu_ids()
    physical_gpu_count = get_physical_gpu_count()

    if gpu_ids is None:
        return parent_visible_ids

    requested_ids = list(gpu_ids)
    if len(requested_ids) == 0:
        return parent_visible_ids

    if not parent_visible_spec["supports_explicit_gpu_ids"]:
        raise ValueError(
            f"Invalid gpu_ids {requested_ids}: explicit physical GPU IDs are "
            f"unsupported when CUDA_VISIBLE_DEVICES uses UUID/MIG entries "
            f"({parent_visible_spec['raw']!r}). Omit gpu_ids to use the "
            "parent-visible devices."
        )

    if len(set(requested_ids)) != len(requested_ids):
        raise ValueError(
            f"Invalid gpu_ids {requested_ids}: duplicate GPU IDs are not allowed. "
            f"Parent-visible GPUs: {parent_visible_ids}"
        )

    # Reject negative IDs.
    negative_ids = [gpu_id for gpu_id in requested_ids if gpu_id < 0]
    if negative_ids:
        raise ValueError(
            f"Invalid gpu_ids {requested_ids}: GPU IDs must be non-negative. "
            f"Rejected IDs: {negative_ids}. Parent-visible GPUs: {parent_visible_ids}"
        )

    # Only enforce the physical upper bound when the count is reliable (nvidia-smi).
    # A torch count reflects only visible devices, so it could falsely reject valid
    # physical indices. The parent-visible check below is always authoritative.
    if physical_gpu_count > 0 and parent_visible_ids:
        max_parent_id = max(parent_visible_ids)
        if physical_gpu_count > max_parent_id:
            # Count is plausibly physical, so enforce it.
            out_of_range = [gpu_id for gpu_id in requested_ids if gpu_id >= physical_gpu_count]
            if out_of_range:
                raise ValueError(
                    f"Invalid gpu_ids {requested_ids}: IDs must be physical GPU IDs "
                    f"between 0 and {physical_gpu_count - 1}. "
                    f"Rejected IDs: {out_of_range}. Parent-visible GPUs: {parent_visible_ids}"
                )

    disallowed_ids = [gpu_id for gpu_id in requested_ids if gpu_id not in parent_visible_ids]
    if disallowed_ids:
        raise ValueError(
            f"Invalid gpu_ids {requested_ids}: requested GPUs {disallowed_ids} are "
            f"outside the parent-visible set {parent_visible_ids}"
        )

    return requested_ids


def _resolve_model_identifier_for_gpu_estimate(
    model_name: str, hf_token: Optional[str] = None
) -> str:
    try:
        from utils.models.model_config import ModelConfig

        config = ModelConfig.from_identifier(model_name, hf_token = hf_token)
        if config and config.is_lora and config.base_model:
            return config.base_model
        return config.identifier if config else model_name
    except Exception as e:
        logger.debug("Could not resolve base model for GPU estimate '%s': %s", model_name, e)
        return model_name


def _get_local_weight_size_bytes(model_name: str) -> Optional[int]:
    model_path = Path(model_name)
    if not model_path.exists():
        return None

    weight_exts = (".safetensors", ".bin", ".pt", ".pth")
    # Skip intermediate training checkpoints: a run dir can hold several
    # checkpoint-*/global_step* snapshots, but export loads only the model at
    # the root, so counting them would multiply the estimate.
    skip_prefixes = ("checkpoint-", "global_step")
    total = 0
    for file in model_path.rglob("*"):
        if not file.is_file() or file.suffix not in weight_exts:
            continue
        rel = file.relative_to(model_path)
        if any(part.startswith(skip_prefixes) for part in rel.parts):
            continue
        total += file.stat().st_size
    return total if total > 0 else None


def _get_hf_safetensors_total_params(
    model_name: str, hf_token: Optional[str] = None
) -> Optional[int]:
    try:
        from huggingface_hub import model_info as hf_model_info

        info = hf_model_info(model_name, token = hf_token)
        safetensors = getattr(info, "safetensors", None)
        if isinstance(safetensors, dict):
            total = safetensors.get("total")
            if total:
                return int(total)
    except Exception as e:
        logger.warning("Could not get safetensors metadata for '%s': %s", model_name, e)
    return None


def _load_config_for_gpu_estimate(model_name: str, hf_token: Optional[str] = None):
    # Estimation needs only declarative config.json fields, and this probe runs
    # on model selection, so read raw config.json (never run auto_map Python) and
    # expose it as an attribute namespace for downstream getattr access.
    try:
        from utils.transformers_version import _load_config_json

        cfg = _load_config_json(model_name, hf_token = hf_token)
        if cfg is None:
            return None

        def _to_ns(d):
            if isinstance(d, dict):
                return types.SimpleNamespace(**{k: _to_ns(v) for k, v in d.items()})
            return d

        return _to_ns(cfg)
    except Exception as e:
        # A 5.x-only config can't be parsed by the default transformers; that is
        # expected (the worker reloads under the sidecar), so only warn for default tier.
        tier = "default"
        try:
            from utils.transformers_version import get_transformers_tier
            tier = get_transformers_tier(model_name)
        except Exception:
            pass
        if tier != "default":
            _tier_version = {"510": "5.10.x", "530": "5.3.0", "550": "5.5.0"}.get(tier, "5.x")
            logger.info(
                "Config for '%s' not parseable by the default transformers; "
                "needs transformers %s and will be loaded with that sidecar in the worker",
                model_name,
                _tier_version,
            )
        else:
            logger.warning("Could not load config for '%s': %s", model_name, e)
        return None


def _determine_attention_impl_for_gpu_estimate(config) -> str:
    # torch.distributed is incomplete on Windows ROCm (torch._C._distributed_c10d
    # can't be imported). Inject stubs into sys.modules before importing
    # torch.distributed, then patch the missing process-group helpers.
    if sys.platform == "win32" and IS_ROCM:
        # Dummy for any name torch.distributed imports from these stubs.
        class _Dummy:
            pass

        for _c10d_name in (
            "torch._C._distributed_c10d",
            "torch._C._distributed_autograd",
            "torch._C._distributed_rpc",
        ):
            if _c10d_name not in sys.modules:
                _stub = types.ModuleType(_c10d_name)
                # No-op dummies for names torch.distributed imports from _distributed_c10d.
                for _sym in (
                    "FakeProcessGroup",
                    "ProcessGroup",
                    "Work",
                    "Store",
                    "PrefixStore",
                    "FileStore",
                    "TCPStore",
                    "HashStore",
                    "Reducer",
                    "Logger",
                    "DistributedDebugLevel",
                    "GradBucket",
                    "BuiltinCommHookType",
                ):
                    setattr(_stub, _sym, _Dummy)
                sys.modules[_c10d_name] = _stub

    try:
        import torch.distributed as _td
        for _attr, _stub in (
            ("is_initialized", lambda: False),
            ("is_available", lambda: False),
            ("get_rank", lambda: 0),
            ("get_world_size", lambda: 1),
            ("is_torchelastic_launched", lambda: False),
        ):
            if not hasattr(_td, _attr):
                setattr(_td, _attr, _stub)
    except ImportError:
        pass

    from unsloth.models._utils import resolve_attention_implementation
    from transformers import AutoModel, AutoModelForCausalLM

    # why: resolve_attention_implementation writes _attn_implementation onto the
    # config and propagates to nested sub-configs; a shallow copy would still
    # mutate the cached config's shared inner objects. Deepcopy isolates them.
    config_copy = copy.deepcopy(config)

    model_class = None
    for auto_model in (AutoModelForCausalLM, AutoModel):
        mapping = getattr(auto_model, "_model_mapping", None)
        if mapping is None:
            continue
        try:
            if config_copy.__class__ in mapping:
                model_class = mapping[config_copy.__class__]
                break
        except Exception:
            continue

    return resolve_attention_implementation(model_class, config_copy)


def _estimate_fp16_model_size_bytes_from_config(config) -> Optional[int]:
    from .vram_estimation import extract_arch_config, compute_total_params

    arch = extract_arch_config(config)
    if arch is None:
        return None
    return compute_total_params(arch) * 2


def _estimate_fp16_model_size_bytes_from_vllm_utils(config) -> Optional[int]:
    if config is None:
        return None

    previous_unsloth_present = os.environ.get("UNSLOTH_IS_PRESENT")
    os.environ["UNSLOTH_IS_PRESENT"] = "1"
    try:
        from unsloth_zoo import vllm_utils as _vllm_utils

        synthetic_total_bytes = 1024 * (1024**3)
        original_get_mem_info = _vllm_utils.get_mem_info
        try:
            _vllm_utils.get_mem_info = lambda: (
                synthetic_total_bytes,
                synthetic_total_bytes,
            )
            _, _, _, memory_left_for_kv_cache_gb = _vllm_utils.approximate_vllm_memory_usage(
                config,
                load_in_4bit = False,
                load_in_8bit = False,
                max_seq_length = 1,
                gpu_memory_utilization = 1.0,
                enable_lora = False,
                account_for_gradients = False,
                cuda_graph_overhead = False,
            )
        finally:
            _vllm_utils.get_mem_info = original_get_mem_info
    except Exception as e:
        logger.debug("Could not estimate model size via vllm_utils: %s", e)
        return None
    finally:
        if previous_unsloth_present is None:
            os.environ.pop("UNSLOTH_IS_PRESENT", None)
        else:
            os.environ["UNSLOTH_IS_PRESENT"] = previous_unsloth_present

    model_size_gb = 1024.0 - memory_left_for_kv_cache_gb
    if model_size_gb <= 0:
        return None
    return int(round(model_size_gb * (1024**3)))


def estimate_fp16_model_size_bytes(
    model_name: str, hf_token: Optional[str] = None
) -> tuple[Optional[int], str]:
    estimate_model = _resolve_model_identifier_for_gpu_estimate(model_name, hf_token = hf_token)

    total_params = None
    if "/" in estimate_model and not Path(estimate_model).exists():
        total_params = _get_hf_safetensors_total_params(estimate_model, hf_token = hf_token)
    if total_params:
        return int(total_params * 2), "safetensors"

    config = _load_config_for_gpu_estimate(estimate_model, hf_token = hf_token)
    config_bytes: Optional[int] = None
    if config is not None:
        config_bytes = _estimate_fp16_model_size_bytes_from_config(config)

    local_bytes = _get_local_weight_size_bytes(estimate_model)

    # why: config-derived bytes cover only the text tower; local safetensors
    # include vision/audio towers. Take the larger so the multimodal
    # extra_bytes correction can fire.
    if config_bytes is not None and local_bytes is not None:
        if local_bytes > config_bytes:
            return local_bytes, "weight_bytes"
        return config_bytes, "config"
    if config_bytes is not None:
        return config_bytes, "config"
    if local_bytes is not None:
        return local_bytes, "weight_bytes"

    vllm_bytes = _estimate_fp16_model_size_bytes_from_vllm_utils(config)
    if vllm_bytes is not None:
        return vllm_bytes, "vllm_utils"

    return None, "unavailable"


def estimate_required_model_memory_gb(
    model_name: str,
    *,
    hf_token: Optional[str] = None,
    training_type: Optional[str] = None,
    load_in_4bit: bool = True,
    batch_size: int = 4,
    max_seq_length: int = 2048,
    lora_rank: int = 16,
    target_modules: Optional[list] = None,
    gradient_checkpointing: str = "unsloth",
    optimizer: str = "adamw_8bit",
) -> tuple[Optional[float], Dict[str, Any]]:
    from .vram_estimation import (
        TrainingVramConfig,
        extract_arch_config,
        estimate_training_vram,
        compute_total_params,
        compute_optimizer_bytes,
        compute_gradient_bytes,
        CUDA_OVERHEAD_BYTES,
        QUANT_4BIT_FACTOR,
        DEFAULT_TARGET_MODULES,
    )

    model_size_bytes, source = estimate_fp16_model_size_bytes(model_name, hf_token = hf_token)
    metadata: Dict[str, Any] = {
        "mode": "inference" if training_type is None else "training",
        "model_size_source": source,
    }
    if model_size_bytes is None:
        metadata["required_gb"] = None
        return None, metadata

    model_size_gb = model_size_bytes / (1024**3)
    metadata["model_size_gb"] = round(model_size_gb, 3)
    min_buffer_gb = 2.0

    if training_type is None:
        if load_in_4bit:
            base_4bit_gb = model_size_gb / QUANT_4BIT_FACTOR
            required_gb = base_4bit_gb + max(base_4bit_gb * 0.3, min_buffer_gb)
        else:
            required_gb = model_size_gb * 1.3
        metadata["required_gb"] = round(required_gb, 3)
        return required_gb, metadata

    training_method = (
        "full" if training_type == "Full Finetuning" else ("qlora" if load_in_4bit else "lora")
    )
    vram_config = TrainingVramConfig(
        training_method = training_method,
        batch_size = batch_size,
        max_seq_length = max_seq_length,
        lora_rank = lora_rank,
        target_modules = target_modules or list(DEFAULT_TARGET_MODULES),
        gradient_checkpointing = gradient_checkpointing,
        optimizer = optimizer,
        load_in_4bit = load_in_4bit,
    )

    estimate_model = _resolve_model_identifier_for_gpu_estimate(model_name, hf_token = hf_token)
    config = _load_config_for_gpu_estimate(estimate_model, hf_token = hf_token)
    if config is not None:
        try:
            vram_config.attention_implementation = _determine_attention_impl_for_gpu_estimate(
                config
            )
        except Exception as e:
            # Debug-level: fires every estimate on Windows ROCm (stub lacks Store);
            # expected and non-actionable -- eager is the safe fallback.
            logger.debug(
                "Could not resolve attention implementation for '%s': %s",
                estimate_model,
                e,
            )
            # why: charge the quadratic non-flash activation path so GPU
            # selection stays conservative when flash attn isn't proven usable.
            vram_config.attention_implementation = "eager"
    arch = extract_arch_config(config) if config is not None else None

    if arch is not None:
        breakdown = estimate_training_vram(arch, vram_config)
        # why: extract_arch_config only sees text_config; add the vision/audio
        # tower bytes that the text-arch fp16 total misses.
        arch_fp16_bytes = compute_total_params(arch) * 2
        extra_bytes = max(0, int(model_size_bytes) - arch_fp16_bytes)
        if extra_bytes > 0:
            breakdown.model_weights += extra_bytes
            if training_method == "full":
                # why: full fine-tuning makes extra params trainable; optimizer +
                # gradient bytes scale with them.
                extra_params = extra_bytes // 2
                breakdown.optimizer_states += compute_optimizer_bytes(
                    extra_params,
                    vram_config.optimizer,
                )
                breakdown.gradients += compute_gradient_bytes(extra_params)
        required_gb = breakdown.total / (1024**3)
        metadata["required_gb"] = round(required_gb, 3)
        metadata["estimation_mode"] = "detailed"
        metadata["attention_implementation"] = vram_config.attention_implementation
        metadata["vram_breakdown"] = breakdown.to_gb_dict()
        max_gpus = max(1, get_visible_gpu_count())
        for n_gpus in range(1, max_gpus + 1):
            metadata["vram_breakdown"][f"min_per_gpu_{n_gpus}"] = round(
                breakdown.min_gpu_vram(n_gpus) / (1024**3), 3
            )
        return required_gb, metadata

    # Fallback when model config is unavailable.
    overhead_gb = CUDA_OVERHEAD_BYTES / (1024**3)
    if training_method == "full":
        required_gb = model_size_gb * 3.5 + overhead_gb
    elif training_method == "qlora":
        base_4bit_gb = model_size_gb / QUANT_4BIT_FACTOR
        lora_overhead_gb = model_size_gb * 0.04
        act_gb = model_size_gb * 0.15 * (batch_size / 4) * (max_seq_length / 2048)
        required_gb = base_4bit_gb + lora_overhead_gb + act_gb + overhead_gb
    else:
        lora_overhead_gb = model_size_gb * 0.04
        act_gb = model_size_gb * 0.15 * (batch_size / 4) * (max_seq_length / 2048)
        required_gb = model_size_gb + lora_overhead_gb + act_gb + overhead_gb

    metadata["required_gb"] = round(required_gb, 3)
    metadata["estimation_mode"] = "fallback"
    return required_gb, metadata


def auto_select_gpu_ids(
    model_name: str,
    *,
    hf_token: Optional[str] = None,
    training_type: Optional[str] = None,
    load_in_4bit: bool = True,
    batch_size: int = 4,
    max_seq_length: int = 2048,
    lora_rank: int = 16,
    target_modules: Optional[list] = None,
    gradient_checkpointing: str = "unsloth",
    optimizer: str = "adamw_8bit",
) -> tuple[Optional[list[int]], Dict[str, Any]]:
    metadata: Dict[str, Any] = {"selection_mode": "auto"}

    if get_device() != DeviceType.CUDA:
        metadata["selection_mode"] = "non_cuda"
        return None, metadata

    required_gb, estimate_metadata = estimate_required_model_memory_gb(
        model_name,
        hf_token = hf_token,
        training_type = training_type,
        load_in_4bit = load_in_4bit,
        batch_size = batch_size,
        max_seq_length = max_seq_length,
        lora_rank = lora_rank,
        target_modules = target_modules,
        gradient_checkpointing = gradient_checkpointing,
        optimizer = optimizer,
    )
    metadata.update(estimate_metadata)
    parent_visible_spec = _get_parent_visible_gpu_spec()
    metadata["parent_cuda_visible_devices"] = parent_visible_spec["raw"]

    if not parent_visible_spec["supports_explicit_gpu_ids"]:
        metadata["selection_mode"] = "inherit_parent_visible"
        metadata["selected_gpu_ids"] = None
        return None, metadata

    if required_gb is None:
        # Can't estimate size -- use all visible GPUs rather than risk one too small.
        parent_ids = get_parent_visible_gpu_ids()
        metadata["selection_mode"] = "fallback_all"
        metadata["selected_gpu_ids"] = parent_ids
        return parent_ids, metadata

    utilization = get_visible_gpu_utilization()
    devices = utilization.get("devices", [])
    parent_ids = get_parent_visible_gpu_ids()

    if not devices:
        metadata["selection_mode"] = "fallback_all"
        metadata["selected_gpu_ids"] = parent_ids
        return parent_ids, metadata

    gpu_candidates = []
    for device in devices:
        total_gb = device.get("vram_total_gb")
        used_gb = device.get("vram_used_gb")
        if total_gb is None or used_gb is None:
            continue
        free_gb = max(total_gb - used_gb, 0.0)
        gpu_candidates.append(
            {
                "index": device["index"],
                "free_gb": free_gb,
            }
        )

    if not gpu_candidates:
        metadata["selection_mode"] = "fallback_all"
        metadata["selected_gpu_ids"] = parent_ids
        return parent_ids, metadata

    ranked = sorted(gpu_candidates, key = lambda item: (-item["free_gb"], item["index"]))
    free_by_index = {item["index"]: item["free_gb"] for item in ranked}
    selected: list[int] = []
    usable_gb = 0.0
    # Sharding has inter-GPU overhead, so each extra GPU contributes less than
    # its raw free memory (first GPU keeps full capacity). 0.85 is empirical on
    # 2-8 GPU setups: covers NCCL buffers, pipeline bubbles, fragmentation.
    multi_gpu_overhead = 0.85

    # Per-GPU check: activations don't shard, so each GPU needs its weight shard
    # + full activation cost. Uses precomputed min_per_gpu_N values.
    vram_breakdown = estimate_metadata.get("vram_breakdown", {})

    for candidate in ranked:
        selected.append(candidate["index"])
        if len(selected) == 1:
            usable_gb = candidate["free_gb"]
        else:
            first_gpu_id = selected[0]
            usable_gb = free_by_index[first_gpu_id] + sum(
                free_by_index[gpu_id] * multi_gpu_overhead for gpu_id in selected[1:]
            )

        total_fits = usable_gb >= required_gb

        per_gpu_fits = True
        if total_fits and len(selected) > 1:
            min_key = f"min_per_gpu_{len(selected)}"
            min_per_gpu_gb = vram_breakdown.get(min_key)
            if min_per_gpu_gb is not None:
                smallest_free = min(free_by_index[gpu_id] for gpu_id in selected)
                per_gpu_fits = smallest_free >= min_per_gpu_gb

        if total_fits and per_gpu_fits:
            metadata["usable_gb"] = round(usable_gb, 3)
            metadata["selection_mode"] = "auto"
            metadata["selected_gpu_ids"] = selected
            logger.debug(
                "Selected GPUs automatically",
                model_name = model_name,
                selected_gpu_ids = selected,
                usable_gb = metadata["usable_gb"],
                required_gb = metadata.get("required_gb"),
                multi_gpu_overhead = multi_gpu_overhead,
            )
            return selected, metadata

    # Use only GPUs with verified VRAM data.
    fallback_all = [c["index"] for c in gpu_candidates] if gpu_candidates else parent_ids
    metadata["selection_mode"] = "fallback_all"
    if ranked:
        fallback_usable = ranked[0]["free_gb"] + sum(
            c["free_gb"] * multi_gpu_overhead for c in ranked[1:]
        )
    else:
        fallback_usable = 0.0
    metadata["usable_gb"] = round(fallback_usable, 3)
    metadata["selected_gpu_ids"] = fallback_all
    logger.warning(
        "Falling back to all visible GPUs -- model may not fit",
        model_name = model_name,
        selected_gpu_ids = fallback_all,
        usable_gb = metadata["usable_gb"],
        required_gb = metadata.get("required_gb"),
        multi_gpu_overhead = multi_gpu_overhead,
    )
    return fallback_all, metadata


def prepare_gpu_selection(
    gpu_ids: Optional[list[int]],
    *,
    model_name: str,
    hf_token: Optional[str] = None,
    training_type: Optional[str] = None,
    load_in_4bit: bool = True,
    batch_size: int = 4,
    max_seq_length: int = 2048,
    lora_rank: int = 16,
    target_modules: Optional[list] = None,
    gradient_checkpointing: str = "unsloth",
    optimizer: str = "adamw_8bit",
) -> tuple[Optional[list[int]], Dict[str, Any]]:
    """Resolve which physical GPUs to use for a model load.

    GPU selection modes:
      - **Explicit** (``gpu_ids=[5, 6, 7]``): caller chooses exact GPUs.
        All listed GPUs are used and the model is sharded via
        ``device_map="balanced"``, even if it would fit on fewer. IDs are
        validated against the parent-visible set.
      - **Auto** (``gpu_ids=None`` or ``[]``): ``auto_select_gpu_ids``
        estimates VRAM needs and picks the *minimum* GPUs needed,
        preferring those with the most free memory.

    The returned ``gpu_ids`` is later passed to ``get_device_map()`` (maps it
    to a Hugging Face ``device_map`` string) and to ``apply_gpu_ids()`` in the
    worker subprocess (narrows ``CUDA_VISIBLE_DEVICES`` before torch/CUDA init).
    """
    if gpu_ids and get_device() != DeviceType.CUDA:
        raise ValueError(
            f"gpu_ids {list(gpu_ids)} is only supported on CUDA devices, "
            f"but the current backend is '{get_device().value}'."
        )

    if gpu_ids:
        resolved = resolve_requested_gpu_ids(gpu_ids)
        metadata = {
            "selection_mode": "explicit",
            "selected_gpu_ids": resolved,
        }
        return resolved, metadata

    selected_gpu_ids, metadata = auto_select_gpu_ids(
        model_name,
        hf_token = hf_token,
        training_type = training_type,
        load_in_4bit = load_in_4bit,
        batch_size = batch_size,
        max_seq_length = max_seq_length,
        lora_rank = lora_rank,
        target_modules = target_modules,
        gradient_checkpointing = gradient_checkpointing,
        optimizer = optimizer,
    )
    return selected_gpu_ids, metadata


def get_physical_gpu_count() -> int:
    """
    Return the number of physical GPUs on the machine.

    Uses ``nvidia-smi -L`` on NVIDIA (unaffected by CUDA_VISIBLE_DEVICES),
    with a torch fallback for AMD ROCm and Intel XPU. Cached after first call.
    """
    global _physical_gpu_count
    if _physical_gpu_count is not None:
        return _physical_gpu_count

    device = get_device()

    if device == DeviceType.CUDA:
        try:
            if IS_ROCM:
                from . import amd as _smi_mod
            else:
                from . import nvidia as _smi_mod
            count = _smi_mod.get_physical_gpu_count()
            if count is not None:
                _physical_gpu_count = count
                return _physical_gpu_count
        except Exception:
            pass
        # SMI unavailable -- fall back to torch.
        count = _torch_get_physical_gpu_count()
        _physical_gpu_count = count if count is not None else 1
        return _physical_gpu_count

    if device == DeviceType.XPU:
        count = _torch_get_physical_gpu_count()
        _physical_gpu_count = count if count is not None else 1
        return _physical_gpu_count

    if device == DeviceType.MLX:
        _physical_gpu_count = 1
        return _physical_gpu_count

    _physical_gpu_count = 0

    return _physical_gpu_count


def _backend_visible_devices_env() -> Optional[str]:
    """Return the raw visibility env string that applies to this backend.

    On ROCm, HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES take precedence over
    CUDA_VISIBLE_DEVICES; this mirrors ``_get_parent_visible_gpu_spec`` so
    ``backend_cuda_visible_devices`` reports the value actually narrowing the
    visible device set.
    """
    if IS_ROCM:
        return _get_parent_visible_gpu_spec().get("raw")
    return os.environ.get("CUDA_VISIBLE_DEVICES")


def get_backend_visible_gpu_info() -> Dict[str, Any]:
    device = get_device()
    if device in (DeviceType.CUDA, DeviceType.XPU):
        parent_visible_ids = get_parent_visible_gpu_ids()
        # Try native SMI first (nvidia-smi; skipped for ROCm).
        if device == DeviceType.CUDA and not IS_ROCM:
            try:
                from . import nvidia

                parent_visible_spec = _get_parent_visible_gpu_spec()
                result = nvidia.get_backend_visible_gpu_info(
                    parent_visible_spec["numeric_ids"],
                    parent_visible_spec["raw"],
                )
                if result.get("available"):
                    result["backend"] = _backend_label(device)
                    return result
            except Exception as e:
                logger.warning("Backend GPU visibility query failed: %s", e)

        # Torch fallback (ROCm, XPU, nvidia-smi missing). Empty parent_visible_ids
        # (UUID/MIG mask) -> enumerate by torch ordinal so the UI shows devices.
        if parent_visible_ids:
            torch_indices = parent_visible_ids
            index_kind = "physical"
        else:
            visible_count = _torch_get_physical_gpu_count() or 0
            torch_indices = list(range(visible_count))
            index_kind = "relative"
        torch_devices = _torch_get_per_device_info(torch_indices)
        if torch_devices:
            devices = [
                {
                    "index": td["index"],
                    "index_kind": index_kind,
                    "visible_ordinal": td["visible_ordinal"],
                    "name": td["name"],
                    "memory_total_gb": td["total_gb"],
                }
                for td in torch_devices
            ]
            return {
                "available": True,
                "backend": _backend_label(device),
                "backend_cuda_visible_devices": _backend_visible_devices_env(),
                "parent_visible_gpu_ids": parent_visible_ids,
                "devices": devices,
                "index_kind": index_kind,
            }

        return {
            "available": False,
            "backend": _backend_label(device),
            "backend_cuda_visible_devices": _backend_visible_devices_env(),
            "parent_visible_gpu_ids": parent_visible_ids,
            "devices": [],
            "index_kind": "physical",
        }

    if device == DeviceType.MLX:
        mem = get_gpu_memory_info()
        if not mem.get("available"):
            return {
                "available": False,
                "backend": _backend_label(device),
                "backend_cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES"),
                "parent_visible_gpu_ids": [],
                "devices": [],
                "index_kind": "relative",
            }
        return {
            "available": True,
            "backend": _backend_label(device),
            "backend_cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES"),
            "parent_visible_gpu_ids": [0],
            "devices": [
                {
                    "index": 0,
                    "index_kind": "relative",
                    "visible_ordinal": 0,
                    "name": mem.get("device_name", "MLX"),
                    "memory_total_gb": round(mem.get("total_gb", 0), 2),
                }
            ],
            "index_kind": "relative",
        }

    return {
        "available": False,
        "backend": _backend_label(device),
        "backend_cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES"),
        "parent_visible_gpu_ids": [],
        "devices": [],
        "index_kind": "relative",
    }


def get_visible_gpu_count() -> int:
    """
    Return the number of GPUs visible to this process.

    Respects ``CUDA_VISIBLE_DEVICES`` -- if set, only those GPUs count.
    Falls back to physical count if unset or torch is unavailable.
    Cached after the first call.
    """
    global _visible_gpu_count
    if _visible_gpu_count is not None:
        return _visible_gpu_count

    # _get_parent_visible_gpu_spec() already handles HIP_VISIBLE_DEVICES /
    # ROCR_VISIBLE_DEVICES on ROCm.
    visible_spec = _get_parent_visible_gpu_spec()
    if visible_spec["raw"] is not None:
        raw = visible_spec["raw"].strip()
        if raw == "" or raw == "-1":
            _visible_gpu_count = 0
        elif visible_spec["numeric_ids"] is not None:
            _visible_gpu_count = len(visible_spec["numeric_ids"])
        else:
            _visible_gpu_count = len([x for x in raw.split(",") if x.strip()])
        return _visible_gpu_count

    # No visibility env var set -- try torch, else physical count
    try:
        import torch
        if get_device() == DeviceType.XPU and hasattr(torch, "xpu"):
            _visible_gpu_count = torch.xpu.device_count()
        else:
            _visible_gpu_count = torch.cuda.device_count()
    except Exception:
        _visible_gpu_count = get_physical_gpu_count()

    return _visible_gpu_count


def apply_gpu_ids(gpu_ids) -> None:
    if gpu_ids is None:
        return

    # Empty list -> treat like None (inherit parent); setting CUDA_VISIBLE_DEVICES=""
    # disables CUDA entirely and crashes downstream torch calls.
    if isinstance(gpu_ids, (list, tuple)) and len(gpu_ids) == 0:
        return

    global _visible_gpu_count

    if isinstance(gpu_ids, (list, tuple)):
        value = ",".join(str(g) for g in gpu_ids)
    else:
        value = str(gpu_ids)

    os.environ["CUDA_VISIBLE_DEVICES"] = value
    # Keep ROCm visibility env vars in sync. Workers may call apply_gpu_ids()
    # before detect_hardware() (IS_ROCM still False), so also mirror when the
    # parent set a ROCm visibility var, with a torch.version.hip probe fallback.
    _inherits_rocm_visibility = (
        "HIP_VISIBLE_DEVICES" in os.environ or "ROCR_VISIBLE_DEVICES" in os.environ
    )
    _is_rocm = IS_ROCM or _inherits_rocm_visibility
    if not _is_rocm:
        # torch.version.hip is set on ROCm, None on CUDA; AMD SDK wheels may leave
        # it unset but encode "rocm" in __version__. Broad except: never crash a worker.
        try:
            import torch as _torch
            _is_rocm = (
                getattr(_torch.version, "hip", None) is not None
                or "rocm" in getattr(_torch, "__version__", "").lower()
            )
        except Exception as e:
            logger.debug(
                "apply_gpu_ids: torch ROCm probe skipped (%s: %s)",
                type(e).__name__,
                e,
            )
    if _is_rocm:
        os.environ["HIP_VISIBLE_DEVICES"] = value
        # ROCR_VISIBLE_DEVICES operates at the HSA agent level and uses
        # different indexing semantics to HIP_VISIBLE_DEVICES. Setting it
        # to a physical GPU index breaks multi-GPU ROCm systems where the
        # parent already set ROCR_VISIBLE_DEVICES (e.g. "0,1"): narrowing
        # to "1" causes torch.cuda.is_available() to return False in the
        # worker subprocess. HIP_VISIBLE_DEVICES is sufficient for GPU
        # selection on ROCm -- leave ROCR_VISIBLE_DEVICES inherited.
    _visible_gpu_count = None
    if _is_rocm:
        logger.info("Applied gpu_ids: CUDA_VISIBLE_DEVICES='%s' (rocm)", value)
    else:
        logger.info("Applied gpu_ids: CUDA_VISIBLE_DEVICES='%s'", value)


def get_device_map(gpu_ids: Optional[list[int]] = None) -> str:
    """Return the Hugging Face ``device_map`` string for model loading.

    Returns ``"balanced"`` (shard evenly across GPUs) when:
      - ``gpu_ids`` explicitly lists >1 GPU, **or**
      - ``CUDA_VISIBLE_DEVICES`` uses UUID/MIG identifiers (non-numeric) and
        >1 GPU is visible (fallback: numeric IDs unresolvable, so assume
        multi-GPU is intended).

    Returns ``"sequential"`` (single device) otherwise, including non-CUDA
    backends (CPU, MLX).

    Use ``prepare_gpu_selection()`` upstream to determine ``gpu_ids`` -- it
    handles auto-selecting the minimum GPUs needed for a model.
    """
    device = get_device()
    if device == DeviceType.CUDA:
        multi_gpu = gpu_ids is not None and len(gpu_ids) > 1

        if not multi_gpu:
            # UUID/MIG masks can't be split into numeric IDs; >1 visible GPU
            # means multi-GPU sharding is intended.
            parent_visible_spec = _get_parent_visible_gpu_spec()
            if parent_visible_spec["numeric_ids"] is None and get_visible_gpu_count() > 1:
                multi_gpu = True

        if multi_gpu:
            return "balanced"

    return "sequential"


def get_offloaded_device_map_entries(model) -> dict[str, str]:
    hf_device_map = getattr(model, "hf_device_map", None)
    if not isinstance(hf_device_map, dict):
        return {}
    return {
        module_name: placement
        for module_name, placement in hf_device_map.items()
        if placement in ("cpu", "disk")
    }


def raise_if_offloaded(
    model,
    device_map: str,
    context: str = "Loading",
) -> None:
    """Raise ``ValueError`` if *model* has modules offloaded to CPU or disk."""
    offloaded = get_offloaded_device_map_entries(model)
    if not offloaded:
        return
    example = ", ".join(f"{name}={placement}" for name, placement in list(offloaded.items())[:5])
    raise ValueError(
        f"{context} does not support models loaded with CPU or disk offload. "
        f"device_map='{device_map}' produced offloaded modules: {example}"
    )


def safe_num_proc(desired: Optional[int] = None) -> int:
    """
    Return a safe ``num_proc`` for ``dataset.map()`` calls.

    On Windows always returns 1: Python uses ``spawn`` not ``fork``, so
    re-importing torch/transformers/unsloth per worker is typically slower
    than single-process for normal dataset sizes.

    On multi-GPU machines (multiple GPUs *visible* to this process) the
    NVIDIA driver spawns extra background threads, making ``os.fork()``
    deadlock-prone with many workers, so this caps ``num_proc`` to 4.
    The cap does not apply when ``CUDA_VISIBLE_DEVICES`` restricts to one GPU.

    Args:
        desired: The num_proc you *want*. If None, auto-computes from
                 ``os.cpu_count()``.

    Returns:
        A safe integer ≥ 1.
    """
    # Windows/macOS use 'spawn'; re-importing torch/transformers/unsloth per
    # worker is typically slower than single-process.
    if sys.platform in ("win32", "darwin"):
        return 1

    if desired is None or not isinstance(desired, int):
        desired = max(1, (os.cpu_count() or 1) // 3)

    visible = get_visible_gpu_count()
    if visible > 1:
        capped = max(1, min(4, desired))
        logger.info(
            f"Multi-GPU detected ({visible} visible GPUs) -- "
            f"capping num_proc {desired} -> {capped} to avoid fork deadlocks"
        )
        return capped

    return max(1, desired)


def safe_thread_num_proc(desired: Optional[int] = None) -> int:
    """
    Return a safe worker count for ``ThreadPoolExecutor`` calls.

    Unlike ``safe_num_proc()``, does NOT cap to 1 on macOS/Windows: threads
    share the parent address space, unaffected by ``spawn`` vs ``fork``.

    Args:
        desired: The thread count you *want*. If None, auto-computes
                 from ``os.cpu_count()``.

    Returns:
        A safe integer >= 1.
    """
    if desired is None or not isinstance(desired, int):
        desired = max(1, (os.cpu_count() or 1) // 3)

    return max(1, desired)


def dataset_map_num_proc(desired: Optional[int] = None) -> Optional[int]:
    """
    Return a safe ``num_proc`` for ``Dataset.map()`` and ``Dataset.filter()``.

    Returns ``None`` on spawn platforms (Windows, macOS) because ``datasets``
    treats ``num_proc=1`` as multiprocessing (creates ``Pool(1)``); only
    ``num_proc=None`` guarantees in-process execution.
    """
    if sys.platform in ("win32", "darwin"):
        return None
    return safe_num_proc(desired)