unsloth/studio/backend/utils/transformers_version.py

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

"""Automatic transformers version switching.

Some newer model architectures (Ministral-3, GLM-4.7-Flash, Qwen3-30B-A3B MoE,
tiny_qwen3_moe) require transformers>=5.3.0, while Gemma 4 models require a
newer 5.x sidecar.  Dense NemotronH models (e.g. NVIDIA-Nemotron-3-Nano-4B) use
MLP layers that only transformers>=5.10 can parse natively, so they go on the
5.10 sidecar too.  Everything else needs the default 4.57.x that ships with
Unsloth.

Two separate target directories are maintained:
  - .venv_t5_530/  — transformers 5.3.0 (Ministral-3, GLM, Qwen3 MoE, etc.)
  - .venv_t5_550/  — transformers 5.5.0 (Gemma 4)
  - .venv_t5_510/  — transformers 5.10.2 (Gemma 4 Unified / 12B)

When loading a LoRA adapter with a custom name, we resolve the base model from
``adapter_config.json`` and check *that* against the model list.

Strategy:
  Training and inference run in subprocesses that activate the correct version
  via sys.path (prepending the appropriate .venv_t5_*/ directory). See:
    - core/training/worker.py
    - core/inference/worker.py

  For export (still in-process), ensure_transformers_version() does a lightweight
  sys.path swap using the same directories pre-installed by setup.sh.
"""

import importlib
import json
import structlog
from loggers import get_logger
import os
import shutil
import subprocess
import sys
from pathlib import Path

from utils.native_path_leases import child_env_without_native_path_secret
from utils.subprocess_compat import (
    windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
)

logger = get_logger(__name__)


_OFFLINE_TRUE_VALUES = {"1", "true", "yes", "on"}


def _env_offline() -> bool:
    """True if an HF offline env var is truthy (canonical strip+lower parse); gates the urllib fetches below."""
    return (
        os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _OFFLINE_TRUE_VALUES
        or os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _OFFLINE_TRUE_VALUES
    )


def hf_endpoint_unreachable(timeout: int = 3) -> bool:
    """Bounded reachability probe to the HF endpoint. A HEAD request runs in a daemon thread
    joined with a deadline, so a resolver blackhole cannot block past ~timeout+1s. True if
    unreachable. urllib natively honors *_PROXY / NO_PROXY, so this verifies real egress
    (the proxy can reach HF), not just that the proxy is up. No ML imports, so it is safe to
    call before transformers version activation. Mirrors the probe in export._hf_offline."""
    import ssl
    import threading
    import urllib.error
    import urllib.request

    endpoint = os.environ.get("HF_ENDPOINT", "https://huggingface.co")
    if "://" not in endpoint:
        endpoint = "https://" + endpoint

    result = {"online": False}

    def _probe():
        try:
            req = urllib.request.Request(endpoint, method = "HEAD")
            with urllib.request.urlopen(req, timeout = timeout):
                result["online"] = True
        except urllib.error.HTTPError as exc:
            # The server/proxy answered: reachable unless it is a gateway error.
            result["online"] = exc.code not in (502, 503, 504)
        except urllib.error.URLError as exc:
            # A TLS/cert failure means we DID reach the server; treat as reachable so the real
            # load surfaces it (consistent with _is_offline_related_error not retrying TLS).
            result["online"] = isinstance(exc.reason, ssl.SSLError)
        except ssl.SSLError:
            result["online"] = True
        except Exception:
            result["online"] = False

    t = threading.Thread(target = _probe, daemon = True)
    t.start()
    t.join(timeout + 1)
    return t.is_alive() or not result["online"]


def _safe_is_file(p: Path) -> bool:
    """``p.is_file()`` returning False instead of raising on a bad path."""
    try:
        return p.is_file()
    except (OSError, ValueError):
        return False


def _safe_is_dir(p: Path) -> bool:
    """``p.is_dir()`` returning False instead of raising on a bad path."""
    try:
        return p.is_dir()
    except (OSError, ValueError):
        return False


# ---------------------------------------------------------------------------
# Detection
# ---------------------------------------------------------------------------

# Lowercase substrings — any match in the lowered model name needs transformers 5.3.0.
TRANSFORMERS_5_MODEL_SUBSTRINGS: tuple[str, ...] = (
    "ministral-3-",  # Ministral-3-{3,8,14}B-{Instruct,Reasoning,Base}-2512
    "glm-4.7-flash",  # GLM-4.7-Flash
    "qwen3-30b-a3b",  # Qwen3-30B-A3B-Instruct-2507 and variants
    "qwen3.5",  # Qwen3.5 family (35B-A3B, etc.)
    "qwen3-next",  # Qwen3-Next and variants
    "tiny_qwen3_moe",  # imdatta0/tiny_qwen3_moe_2.8B_0.7B
    "lfm2.5-vl-450m",  # LiquidAI/LFM2.5-VL-450M
)

# Lowercase substrings for models that require transformers 5.10.x (checked first).
TRANSFORMERS_510_MODEL_SUBSTRINGS: tuple[str, ...] = (
    "gemma-4-12b",  # Gemma 4 Unified 12B
    "gemma4-12b",
)

# Lowercase substrings for models that require the Gemma 4 transformers 5.5 sidecar.
TRANSFORMERS_550_MODEL_SUBSTRINGS: tuple[str, ...] = (
    "gemma-4",  # Gemma-4 (E2B-it, E4B-it, 31B-it, 26B-A4B-it)
    "gemma4",  # Gemma-4 alternate naming
    "qwen3.6",
)

# Architecture classes / model_type values that require transformers 5.10.x.
# Checked via config.json (local or HuggingFace).
_TRANSFORMERS_510_ARCHITECTURES: set[str] = {
    "Gemma4UnifiedForConditionalGeneration",
    "Gemma4AssistantForCausalLM",
    "Gemma4UnifiedAssistantForCausalLM",
}
_TRANSFORMERS_510_MODEL_TYPES: set[str] = {
    "gemma4_unified",
    "gemma4_assistant",
    "gemma4_unified_assistant",
}

# Architecture classes / model_type values that require transformers 5.5.0.
# Checked via config.json (local or HuggingFace).
_TRANSFORMERS_550_ARCHITECTURES: set[str] = {
    "Gemma4ForConditionalGeneration",
}
_TRANSFORMERS_550_MODEL_TYPES: set[str] = {
    "gemma4",
}

# Architecture classes / model_type values that require transformers 5.3.0.
# Checked via config.json (local or HuggingFace).
_TRANSFORMERS_530_ARCHITECTURES: set[str] = {
    "Qwen3_5ForCausalLM",
    "Qwen3_5ForConditionalGeneration",
    "Qwen3_5MoeForCausalLM",
    "Qwen3_5MoeForConditionalGeneration",
    "Qwen3MoeForCausalLM",
    "Qwen3NextForCausalLM",
    "Glm4MoeLiteForCausalLM",
    "Lfm2VlForConditionalGeneration",
}
_TRANSFORMERS_530_MODEL_TYPES: set[str] = {
    "qwen3_5",
    "qwen3_5_text",
    "qwen3_5_moe",
    "qwen3_5_moe_text",
    "qwen3_moe",
    "qwen3_next",
    "glm4_moe_lite",
    "lfm2_vl",
}

# Tokenizer classes that only exist in transformers>=5.x.
_TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
    "TokenizersBackend",
}

# Caches keyed on (model_name, token-hash) so authed/unauthed reads stay separate (a
# gated/private repo's unauthenticated miss must not poison a later authenticated lookup).
# Offline negatives are NOT written (see the _env_offline branches) so they cannot poison a
# later online read in this persistent worker.
_tokenizer_class_cache: dict[tuple[str, str | None], bool] = {}
_config_json_cache: dict[tuple[str, str | None], dict | None] = {}
_config_needs_510_cache: dict[tuple[str, str | None], bool] = {}
_config_needs_550_cache: dict[tuple[str, str | None], bool] = {}
_config_needs_530_cache: dict[tuple[str, str | None], bool] = {}

# AutoConfig-probe tier cache for the process lifetime (cleared on restart), keyed by
# model_name plus a local config.json signature (see _probe_cache_key) so an overwritten
# checkpoint re-probes. Not keyed by Hub sha, so the probe never imports huggingface_hub
# before a worker's sidecar venv is activated (which would pin the wrong hub).
_probe_tier_cache: dict[str, str] = {}

# Versions
TRANSFORMERS_510_VERSION = "5.10.2"
TRANSFORMERS_550_VERSION = "5.5.0"
TRANSFORMERS_530_VERSION = "5.3.0"
TRANSFORMERS_DEFAULT_VERSION = "4.57.6"
# Backwards-compat alias — points to the highest 5.x tier.
# Consumers should prefer TRANSFORMERS_510_VERSION / TRANSFORMERS_550_VERSION /
# TRANSFORMERS_530_VERSION.
TRANSFORMERS_5_VERSION = TRANSFORMERS_510_VERSION

# Pre-installed directories — created by setup.sh / setup.ps1.
from utils.paths.storage_roots import studio_root as _studio_root  # noqa: E402

_VENV_T5_530_DIR = str(_studio_root() / ".venv_t5_530")
_VENV_T5_550_DIR = str(_studio_root() / ".venv_t5_550")
_VENV_T5_510_DIR = str(_studio_root() / ".venv_t5_510")
# Backwards-compat alias
_VENV_T5_DIR = _VENV_T5_550_DIR

# llm-compressor-main shadow for FP8/FP4 export of newer-transformers models. Like the .venv_t5_*
# sidecars but also shadows llm-compressor main + compressed-tensors; installed --no-deps so it
# reuses the workspace torch (torch-agnostic).
_VENV_LLMCOMPRESSOR_DIR = str(_studio_root() / ".venv_llmcompressor")

# Tier precedence: higher rank wins in _higher_tier.
_TIER_RANK = {"default": 0, "530": 1, "550": 2, "510": 3}


def _higher_tier(a: str, b: str) -> str:
    return a if _TIER_RANK.get(a, 0) >= _TIER_RANK.get(b, 0) else b


def activate_transformers_for_subprocess(model_name: str, hf_token: str | None = None) -> None:
    """Activate the correct transformers version in a subprocess worker.

    Call BEFORE any ML imports. Resolves LoRA adapters to their base model,
    determines the required tier, prepends the appropriate ``.venv_t5_*`` dir to
    ``sys.path``, and propagates it via ``PYTHONPATH`` for child processes
    (e.g. GGUF converter). Used by training, inference, and export workers.

    ``hf_token`` is forwarded to tier detection so a gated/private model whose only 5.x
    signal is an authenticated config/tokenizer reaches the right sidecar, not the default.
    """
    # Pre-resolve only LoRA adapters; full checkpoints go to get_transformers_tier so their
    # local config.json drives the tier (a full checkpoint with a private/offline
    # _name_or_path must not resolve to an unreachable HF id and skip its own config).
    if _is_lora_adapter_dir(Path(model_name)):
        resolved = _resolve_base_model(model_name)
    else:
        resolved = model_name
    tier = get_transformers_tier(resolved, hf_token)
    if model_name != resolved and _safe_is_file(Path(model_name) / "config.json"):
        # Gate on a real local config.json: a checkpoint carries config the base may not
        # surface, but path names alone must not upgrade a plain adapter.
        tier = _higher_tier(tier, get_transformers_tier(model_name, hf_token))

    if tier == "510":
        if not _ensure_venv_t5_510_exists():
            raise RuntimeError(
                f"Cannot activate transformers {TRANSFORMERS_510_VERSION}: "
                f".venv_t5_510 missing at {_VENV_T5_510_DIR}"
            )
        if _VENV_T5_510_DIR not in sys.path:
            sys.path.insert(0, _VENV_T5_510_DIR)
        logger.info(
            "Prepended transformers %s venv to sys.path from %s "
            "(path only; the loaded version is confirmed later by "
            "'Subprocess loaded transformers ...' on first import)",
            TRANSFORMERS_510_VERSION,
            _VENV_T5_510_DIR,
        )
        _pp = os.environ.get("PYTHONPATH", "")
        os.environ["PYTHONPATH"] = _VENV_T5_510_DIR + (os.pathsep + _pp if _pp else "")
    elif tier == "550":
        if not _ensure_venv_t5_550_exists():
            raise RuntimeError(
                f"Cannot activate transformers {TRANSFORMERS_550_VERSION}: "
                f".venv_t5_550 missing at {_VENV_T5_550_DIR}"
            )
        if _VENV_T5_550_DIR not in sys.path:
            sys.path.insert(0, _VENV_T5_550_DIR)
        logger.info(
            "Prepended transformers %s venv to sys.path from %s "
            "(path only; the loaded version is confirmed later by "
            "'Subprocess loaded transformers ...' on first import)",
            TRANSFORMERS_550_VERSION,
            _VENV_T5_550_DIR,
        )
        _pp = os.environ.get("PYTHONPATH", "")
        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
    elif tier == "530":
        if not _ensure_venv_t5_530_exists():
            raise RuntimeError(
                f"Cannot activate transformers 5.3.0: "
                f".venv_t5_530 missing at {_VENV_T5_530_DIR}"
            )
        if _VENV_T5_530_DIR not in sys.path:
            sys.path.insert(0, _VENV_T5_530_DIR)
        logger.info(
            "Prepended transformers %s venv to sys.path from %s "
            "(path only; the loaded version is confirmed later by "
            "'Subprocess loaded transformers ...' on first import)",
            TRANSFORMERS_530_VERSION,
            _VENV_T5_530_DIR,
        )
        _pp = os.environ.get("PYTHONPATH", "")
        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
    else:
        logger.info("Using default transformers (4.57.x) for %s", model_name)


def _has_adapter_weights(path: Path) -> bool:
    """True if *path* holds LoRA adapter weight files (``adapter_model.*``)."""
    try:
        return any(path.glob("adapter_model*.safetensors")) or any(path.glob("adapter_model*.bin"))
    except OSError:
        return False


def _is_lora_adapter_dir(path: Path) -> bool:
    """True if *path* is a local LoRA dir (adapter_config.json or adapter_model-only
    weights). Import-light so it can run during subprocess activation."""
    try:
        if not path.is_dir():
            return False
        return (path / "adapter_config.json").is_file() or _has_adapter_weights(path)
    except OSError:
        return False


def _is_same_path(value: str, local_path: Path) -> bool:
    """True if *value* resolves to *local_path* (relative/absolute/symlink)."""
    if value == str(local_path):
        return True
    try:
        return os.path.realpath(value) == os.path.realpath(str(local_path))
    except OSError:
        return False


def _resolve_base_model(model_name: str) -> str:
    """If *model_name* points to a LoRA adapter, return its base model.

    Checks ``adapter_config.json`` locally first. Only calls the heavier
    ``get_base_model_from_lora`` for real local directories (avoids noisy
    warnings for plain HF model IDs). Returns *model_name* unchanged if not a
    LoRA adapter.
    """
    # --- Fast local check ---------------------------------------------------
    local_path = Path(model_name)
    adapter_cfg_path = local_path / "adapter_config.json"
    if _safe_is_file(adapter_cfg_path):
        try:
            with open(adapter_cfg_path) as f:
                cfg = json.load(f)
            base = cfg.get("base_model_name_or_path")
            if base:
                logger.info(
                    "Resolved LoRA adapter '%s' → base model '%s'",
                    model_name,
                    base,
                )
                return base
        except Exception as exc:
            logger.debug("Could not read %s: %s", adapter_cfg_path, exc)

    # --- config.json fallback (works for both LoRA and full fine-tune) ------
    config_json_path = local_path / "config.json"
    if _safe_is_file(config_json_path):
        try:
            with open(config_json_path) as f:
                cfg = json.load(f)
            # Unsloth writes model_name, HF writes _name_or_path; skip a self-reference.
            for _key in ("model_name", "_name_or_path"):
                base = cfg.get(_key)
                if isinstance(base, str) and base and not _is_same_path(base, local_path):
                    logger.info(
                        "Resolved checkpoint '%s' → base model '%s' (via config.json)",
                        model_name,
                        base,
                    )
                    return base
        except Exception as exc:
            logger.debug("Could not read %s: %s", config_json_path, exc)

    # Gate the heavy resolver on adapter_config.json: importing utils.models pulls
    # in transformers, which would pin the default into sys.modules before the
    # sidecar venv is prepended during activation.
    if _safe_is_file(adapter_cfg_path):
        try:
            from utils.models import get_base_model_from_lora
            base = get_base_model_from_lora(model_name)
            if base:
                logger.info(
                    "Resolved LoRA adapter '%s' → base model '%s' "
                    "(via get_base_model_from_lora)",
                    model_name,
                    base,
                )
                return base
        except Exception as exc:
            logger.debug(
                "get_base_model_from_lora failed for '%s': %s",
                model_name,
                exc,
            )

    # adapter_model-only LoRA: no config to resolve from, so use the
    # unsloth_<model>_<timestamp> dir-name convention (pure string parse).
    if local_path.name.startswith("unsloth_") and _has_adapter_weights(local_path):
        parts = local_path.name.split("_")
        if len(parts) >= 2:  # unsloth_<model...>_<timestamp>
            base = "unsloth/" + "_".join(parts[1:-1])
            logger.info(
                "Resolved adapter-only LoRA '%s' → base model '%s' (via directory name)",
                model_name,
                base,
            )
            return base

    return model_name


def _token_cache_key(model_name: str, hf_token: str | None) -> tuple[str, str | None]:
    """Cache key that keeps authenticated and unauthenticated reads separate, so an
    unauthenticated miss on a gated/private repo never poisons a later authed lookup."""
    import hashlib

    tok = hashlib.sha256(hf_token.encode()).hexdigest()[:16] if hf_token else None
    return (model_name, tok)


def _is_canonical_repo_id(model_name: str) -> bool:
    """True for a canonical ``owner/repo`` Hub id (not a local or relative path)."""
    return bool(
        model_name
        and model_name.count("/") == 1
        and model_name[0] not in "/.~"
        and "\\" not in model_name
    )


def _adapter_base_from_hf_cache(model_name: str) -> str | None:
    """``base_model_name_or_path`` from a remote adapter's cached ``adapter_config.json``.

    Stdlib path resolution of the HF hub cache (no ``huggingface_hub`` import); the newest
    snapshot wins. Lets an offline cached LoRA still resolve its base.
    """
    if not _is_canonical_repo_id(model_name):
        return None
    hub = (
        os.environ.get("HF_HUB_CACHE")
        or os.environ.get("HUGGINGFACE_HUB_CACHE")
        or os.path.join(
            os.environ.get("HF_HOME") or os.path.expanduser("~/.cache/huggingface"), "hub"
        )
    )
    repo_dir = Path(hub) / ("models--" + model_name.replace("/", "--"))
    candidates = []
    ref_main = repo_dir / "refs" / "main"

    def _mtime(p: Path) -> float:
        try:
            return p.stat().st_mtime
        except OSError:
            return 0.0

    try:
        if ref_main.is_file():
            candidates.append(
                repo_dir / "snapshots" / ref_main.read_text().strip() / "adapter_config.json"
            )
        candidates += sorted(
            repo_dir.glob("snapshots/*/adapter_config.json"), key = _mtime, reverse = True
        )
        for cfg_path in candidates:
            if cfg_path.is_file():
                base = json.loads(cfg_path.read_text()).get("base_model_name_or_path")
                return base or None
    except Exception as exc:
        logger.debug("HF cache adapter_config.json lookup failed for '%s': %s", model_name, exc)
    return None


def _remote_lora_base(model_name: str, hf_token: str | None = None) -> str | None:
    """``base_model_name_or_path`` from a remote adapter's ``adapter_config.json``, or None.

    Raw HTTP (no huggingface_hub / transformers import), so a remote LoRA's base is known
    before any ML import. Offline (or on a transient failure) it reads the local hub cache,
    since a cached adapter is still loadable; a definitive 404 returns None (the repo is not
    a LoRA) rather than a stale cached base. Skipped for local/non-canonical ids.
    """
    if not _is_canonical_repo_id(model_name):
        return None
    try:
        from utils.paths import is_local_path
        if is_local_path(model_name):
            return None  # an existing relative path is a local checkpoint, not a Hub repo
    except Exception:
        pass
    if _env_offline():
        return _adapter_base_from_hf_cache(model_name)

    import urllib.error
    import urllib.request

    endpoint = (os.environ.get("HF_ENDPOINT") or "https://huggingface.co").rstrip("/")
    url = f"{endpoint}/{model_name}/raw/main/adapter_config.json"
    headers = {"User-Agent": "unsloth-studio"}
    if hf_token:
        headers["Authorization"] = f"Bearer {hf_token}"
    try:
        req = urllib.request.Request(url, headers = headers)
        with urllib.request.urlopen(req, timeout = 10) as resp:
            cfg = json.loads(resp.read().decode())
        base = cfg.get("base_model_name_or_path")
        if base:
            logger.info("Resolved remote LoRA adapter '%s' → base model '%s'", model_name, base)
        return base or None
    except urllib.error.HTTPError as exc:
        if exc.code == 404:
            return None  # definitively not a LoRA; do not serve a stale cached base
        logger.debug("adapter_config.json fetch failed for '%s': %s", model_name, exc)
        return _adapter_base_from_hf_cache(model_name)
    except Exception as exc:
        logger.debug("No remote adapter_config.json for '%s': %s", model_name, exc)
        return _adapter_base_from_hf_cache(model_name)


def _check_tokenizer_config_needs_v5(model_name: str, hf_token: str | None = None) -> bool:
    """True if the model's tokenizer_class requires transformers 5.x.

    Checks local tokenizer_config.json, else fetches from HuggingFace (authenticated
    with ``hf_token`` so gated/private repos resolve). Cached in
    ``_tokenizer_class_cache``, keyed by (model, token) so an unauthenticated miss does
    not poison a later authed read. Returns False on any network/parse error
    (fail-open to default version).
    """
    cache_key = _token_cache_key(model_name, hf_token)
    if cache_key in _tokenizer_class_cache:
        return _tokenizer_class_cache[cache_key]

    # --- Check local tokenizer_config.json first ---------------------------
    local_path = Path(model_name)
    local_tc = local_path / "tokenizer_config.json"
    if _safe_is_file(local_tc):
        try:
            with open(local_tc) as f:
                data = json.load(f)
            tokenizer_class = data.get("tokenizer_class", "")
            result = tokenizer_class in _TRANSFORMERS_5_TOKENIZER_CLASSES
            if result:
                logger.info(
                    "Local check: %s uses tokenizer_class=%s (requires transformers 5.x)",
                    model_name,
                    tokenizer_class,
                )
            _tokenizer_class_cache[cache_key] = result
            return result
        except Exception as exc:
            logger.debug("Could not read %s: %s", local_tc, exc)

    # Local checkpoint without the file yet: don't fetch it as a Hub id or cache the miss,
    # so a file written later this process (in-progress checkpoint) is read next call.
    if _safe_is_dir(local_path):
        return False

    # Offline: skip the 10s urllib fetch (fail-open to lower tier). Do NOT cache this
    # assumed negative, so a later online read of the same id re-fetches the real value.
    if _env_offline():
        return False

    # --- Fall back to fetching from HuggingFace ----------------------------
    import urllib.request

    url = f"https://huggingface.co/{model_name}/raw/main/tokenizer_config.json"
    headers = {"User-Agent": "unsloth-studio"}
    if hf_token:
        headers["Authorization"] = f"Bearer {hf_token}"
    try:
        req = urllib.request.Request(url, headers = headers)
        with urllib.request.urlopen(req, timeout = 10) as resp:
            data = json.loads(resp.read().decode())
        tokenizer_class = data.get("tokenizer_class", "")
        result = tokenizer_class in _TRANSFORMERS_5_TOKENIZER_CLASSES
        if result:
            logger.info(
                "Dynamic check: %s uses tokenizer_class=%s (requires transformers 5.x)",
                model_name,
                tokenizer_class,
            )
        _tokenizer_class_cache[cache_key] = result
        return result
    except Exception as exc:
        logger.debug("Could not fetch tokenizer_config.json for '%s': %s", model_name, exc)
        _tokenizer_class_cache[cache_key] = False
        return False


def _safe_mtime(path: Path) -> float:
    try:
        return path.stat().st_mtime
    except OSError:
        return 0.0


def _config_json_from_hf_cache(model_name: str) -> dict | None:
    """Parsed ``config.json`` from the local HF hub cache, or None.

    Stdlib-only path resolution (no ``huggingface_hub`` import) so tier detection never
    loads the default-env hub before a sidecar venv is activated.
    """
    # Only a canonical ``owner/repo`` Hub id maps to a cache dir; reject local paths.
    if not model_name or model_name.count("/") != 1 or model_name[0] in "/.~" or "\\" in model_name:
        return None
    hub = (
        os.environ.get("HF_HUB_CACHE")
        or os.environ.get("HUGGINGFACE_HUB_CACHE")
        or os.path.join(
            os.environ.get("HF_HOME") or os.path.expanduser("~/.cache/huggingface"), "hub"
        )
    )
    repo_dir = Path(hub) / ("models--" + model_name.replace("/", "--"))
    candidates = []
    ref_main = repo_dir / "refs" / "main"
    try:
        if ref_main.is_file():
            candidates.append(repo_dir / "snapshots" / ref_main.read_text().strip() / "config.json")
        # No refs/main (e.g. commit-pinned downloads): newest snapshot by mtime, not a stale
        # lexicographically-first SHA, matching what the Hub cache would actually load.
        candidates += sorted(
            repo_dir.glob("snapshots/*/config.json"), key = _safe_mtime, reverse = True
        )
        for cfg_path in candidates:
            if cfg_path.is_file():
                with open(cfg_path) as f:
                    return json.load(f)
    except Exception as exc:
        logger.debug("HF cache config.json lookup failed for '%s': %s", model_name, exc)
    return None


def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | None:
    """Return parsed ``config.json`` for *model_name*, checking local files first.

    ``hf_token`` authenticates the raw fetch so gated/private repos resolve. The
    cache is keyed on the token so an unauthenticated miss never poisons a later
    authenticated read. The HF hub cache is consulted only offline or after a failed
    network fetch, so an online read never serves stale metadata.
    """
    import hashlib

    tok = hashlib.sha256(hf_token.encode()).hexdigest()[:16] if hf_token else None
    cache_key = (model_name, tok)
    if cache_key in _config_json_cache:
        return _config_json_cache[cache_key]

    local_cfg = Path(model_name) / "config.json"
    if _safe_is_file(local_cfg):
        try:
            with open(local_cfg) as f:
                cfg = json.load(f)
            _config_json_cache[cache_key] = cfg
            return cfg
        except Exception as exc:
            logger.debug("Could not read %s: %s", local_cfg, exc)
            _config_json_cache[cache_key] = None
            return None

    # Local checkpoint without the file yet: don't fetch it as a Hub id or cache the miss,
    # so a file written later this process (in-progress checkpoint) is read next call.
    if _safe_is_dir(Path(model_name)):
        return None

    if _env_offline():
        # No network: a previously downloaded repo can still tier from the hub cache. Cache a
        # real hit, but never the miss (None) so a later online read still fetches the config.
        cfg = _config_json_from_hf_cache(model_name)
        if cfg is not None:
            _config_json_cache[cache_key] = cfg
        return cfg

    import urllib.error
    import urllib.request

    url = f"https://huggingface.co/{model_name}/raw/main/config.json"
    headers = {"User-Agent": "unsloth-studio"}
    if hf_token:
        headers["Authorization"] = f"Bearer {hf_token}"
    try:
        req = urllib.request.Request(url, headers = headers)
        with urllib.request.urlopen(req, timeout = 10) as resp:
            cfg = json.loads(resp.read().decode())
        _config_json_cache[cache_key] = cfg
        return cfg
    except urllib.error.HTTPError as exc:
        # 401/403/404 is a definitive access answer: never serve another caller's cached
        # private metadata to an unauthenticated/wrong-token request.
        if exc.code in (401, 403, 404):
            logger.debug("config.json access denied for '%s': %s", model_name, exc)
            return None
        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
        return _config_json_from_hf_cache(model_name)
    except Exception as exc:
        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
        # Transient: serve the hub cache uncached so the next call retries the network.
        return _config_json_from_hf_cache(model_name)


def _config_json_is_definitive(model_name: str, hf_token: str | None = None) -> bool:
    """True if the last ``_load_config_json`` read for this model+token was cached
    (definitive), not a transient fallback (not stored, so callers re-check next call)."""
    return _token_cache_key(model_name, hf_token) in _config_json_cache


def _config_matches_tier(cfg: dict, architectures: set[str], model_types: set[str]) -> bool:
    # Defensive: a malformed config may carry non-string values (e.g. list model_type).
    archs = cfg.get("architectures")
    if isinstance(archs, (list, tuple)) and any(a in architectures for a in archs):
        return True
    mt = cfg.get("model_type")
    return isinstance(mt, str) and mt in model_types


def _config_needs_550(cfg: dict) -> bool:
    return _config_matches_tier(
        cfg,
        _TRANSFORMERS_550_ARCHITECTURES,
        _TRANSFORMERS_550_MODEL_TYPES,
    )


_NESTED_CONFIG_KEYS = ("llm_config", "text_config", "language_config", "thinker_config")


def _nemotron_h_needs_mlp_support(cfg: dict) -> bool:
    """True for a dense NemotronH config using MLP (``-``) layers.

    transformers only gained ``-`` -> ``mlp`` in 5.10; 5.3/5.5 raise ``KeyError: '-'``.
    Read from ``hybrid_override_pattern`` or ``layers_block_type``, recursing into nested
    language configs (VL wrappers hold the dense LM under ``llm_config``/``text_config``).
    """
    if not isinstance(cfg, dict):
        return False
    if cfg.get("model_type") == "nemotron_h":
        pattern = cfg.get("hybrid_override_pattern")
        if isinstance(pattern, str) and "-" in pattern:
            return True
        block_types = cfg.get("layers_block_type")
        if isinstance(block_types, (list, tuple)) and "mlp" in block_types:
            return True
    return any(_nemotron_h_needs_mlp_support(cfg.get(key)) for key in _NESTED_CONFIG_KEYS)


def _config_needs_510(cfg: dict) -> bool:
    if _config_matches_tier(
        cfg,
        _TRANSFORMERS_510_ARCHITECTURES,
        _TRANSFORMERS_510_MODEL_TYPES,
    ):
        return True
    return _nemotron_h_needs_mlp_support(cfg)


def _config_needs_530(cfg: dict) -> bool:
    return _config_matches_tier(
        cfg,
        _TRANSFORMERS_530_ARCHITECTURES,
        _TRANSFORMERS_530_MODEL_TYPES,
    )


def _check_config_needs_550(model_name: str, hf_token: str | None = None) -> bool:
    """True if ``config.json`` needs transformers 5.5.0 (e.g. Gemma 4). Local first, else
    fetched (authenticated with ``hf_token``); cached by (model, token) only for a definitive
    read so a transient miss retries. False on error.
    """
    cache_key = _token_cache_key(model_name, hf_token)
    if cache_key in _config_needs_550_cache:
        return _config_needs_550_cache[cache_key]

    cfg = _load_config_json(model_name, hf_token)
    result = bool(cfg) and _config_needs_550(cfg)
    if result:
        logger.info(
            "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
            model_name,
            TRANSFORMERS_550_VERSION,
            cfg.get("architectures", []),
            cfg.get("model_type"),
        )
    if _config_json_is_definitive(model_name, hf_token):
        _config_needs_550_cache[cache_key] = result
    return result


def _check_config_needs_530(model_name: str, hf_token: str | None = None) -> bool:
    """True if ``config.json`` needs transformers 5.3.0 (Qwen3.5, Qwen3 MoE, GLM-4.7, LFM2.5-VL).
    Local first, else fetched (authenticated with ``hf_token``); cached by (model, token) only
    for a definitive read so a transient miss retries. False on error.
    """
    cache_key = _token_cache_key(model_name, hf_token)
    if cache_key in _config_needs_530_cache:
        return _config_needs_530_cache[cache_key]

    cfg = _load_config_json(model_name, hf_token)
    result = bool(cfg) and _config_needs_530(cfg)
    if result:
        logger.info(
            "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
            model_name,
            TRANSFORMERS_530_VERSION,
            cfg.get("architectures", []),
            cfg.get("model_type"),
        )
    if _config_json_is_definitive(model_name, hf_token):
        _config_needs_530_cache[cache_key] = result
    return result


def _check_config_needs_510(model_name: str, hf_token: str | None = None) -> bool:
    """Check ``config.json`` for Gemma 4 Unified / 12B architectures (authenticated with
    ``hf_token``; cached by (model, token) only for a definitive read)."""
    cache_key = _token_cache_key(model_name, hf_token)
    if cache_key in _config_needs_510_cache:
        return _config_needs_510_cache[cache_key]

    cfg = _load_config_json(model_name, hf_token)
    result = bool(cfg) and _config_needs_510(cfg)
    if result:
        logger.info(
            "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
            model_name,
            TRANSFORMERS_510_VERSION,
            cfg.get("architectures", []),
            cfg.get("model_type"),
        )
    if _config_json_is_definitive(model_name, hf_token):
        _config_needs_510_cache[cache_key] = result
    return result


def _config_saved_by_transformers_5(cfg: dict | None) -> bool:
    """True if ``config.json``'s ``transformers_version`` is >= 5. Only a cheap "worth
    probing" hint (the saving version, not the minimum to load); the default-first probe
    decides the actual tier."""
    if not isinstance(cfg, dict):
        return False
    ver = cfg.get("transformers_version")
    if not isinstance(ver, str):
        return False
    try:
        return int(ver.strip().split(".", 1)[0]) >= 5
    except ValueError:
        return False


def _cached_config_json(model_name: str, hf_token: str | None) -> dict | None:
    """Already-fetched config.json from the in-process cache (no new fetch); the tier checks
    above populate it, and a miss just skips the version-field probe."""
    return _config_json_cache.get(_token_cache_key(model_name, hf_token))


# --- AutoConfig probe: general tier resolution for ambiguous models ----------
# When the cheap signals only say "needs some 5.x", parse config.json with the built-in
# parser in each candidate sidecar (lowest first) instead of guessing. Generalizes beyond
# the hardcoded lists, e.g. dense NemotronH whose '-' (MLP) layer only 5.10 can parse.
_PROBE_TIER_ORDER = ("530", "550", "510")
_PROBE_TIMEOUT_SECS = 60

# config.json-only parse in a sidecar (--target dir on sys.path, no per-venv python).
# Built-in parser only, no repo code, no weights. Exit 0 = parses; token via env, not argv.
_PROBE_CONFIG_SCRIPT = r"""
import sys, os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
target_dir, model_name = sys.argv[1], sys.argv[2]
if target_dir:  # empty = probe the ambient (default 4.57.x) transformers, no sidecar prepend
    sys.path.insert(0, target_dir)
try:
    from transformers import AutoConfig
    AutoConfig.from_pretrained(model_name, trust_remote_code=False)
    sys.exit(0)
except Exception as exc:
    # stderr encoding may not be UTF-8 (e.g. cp1252 on Windows); write bytes so a
    # non-ASCII error message cannot itself raise UnicodeEncodeError.
    sys.stderr.buffer.write((type(exc).__name__ + ": " + str(exc)).encode("utf-8", "replace"))
    sys.exit(1)
"""

# stderr fragments meaning "couldn't fetch/auth", NOT "needs a newer parser".
_PROBE_TRANSIENT_MARKERS = (
    "ConnectionError",
    "HTTPError",
    "Timeout",
    "Max retries",
    "Temporary failure",
    "GatedRepoError",
    "RepositoryNotFoundError",
    "LocalEntryNotFoundError",
    "OfflineModeIsEnabled",
    "401",
    "403",
    "404",
)


def _stderr_is_transient(err: str) -> bool:
    return any(marker in err for marker in _PROBE_TRANSIENT_MARKERS)


def _probe_tier_venvs():
    """tier -> (target_dir, ensure_fn), a function so the later _ensure_* defs resolve. The
    ``default`` entry (empty target_dir = ambient 4.57.x) is only probed with include_default."""
    return {
        "default": ("", lambda: True),
        "530": (_VENV_T5_530_DIR, _ensure_venv_t5_530_exists),
        "550": (_VENV_T5_550_DIR, _ensure_venv_t5_550_exists),
        "510": (_VENV_T5_510_DIR, _ensure_venv_t5_510_exists),
    }


def _probe_autoconfig(target_dir: str, model_name: str, hf_token: str | None) -> bool | None:
    """Parse config.json with the built-in parser inside *target_dir*'s sidecar.
    True = parses, False = parse/version failure (escalate), None = transient
    (auth/network/offline/spawn) so the caller fails safe and does not cache.
    """
    env = child_env_without_native_path_secret()
    if hf_token:
        env["HF_TOKEN"] = hf_token
        # The probe relies on the implicit HF_TOKEN env (no token= arg). Clear any inherited
        # HF_HUB_DISABLE_IMPLICIT_TOKEN=1 so a gated repo authenticates instead of 401ing
        # into the 530 fail-safe.
        env["HF_HUB_DISABLE_IMPLICIT_TOKEN"] = "0"
    if _env_offline():
        env["HF_HUB_OFFLINE"] = "1"
        env["TRANSFORMERS_OFFLINE"] = "1"
    try:
        result = subprocess.run(
            [sys.executable, "-c", _PROBE_CONFIG_SCRIPT, target_dir, model_name],
            capture_output = True,
            text = True,
            errors = "replace",
            timeout = _PROBE_TIMEOUT_SECS,
            env = env,
            **_windows_hidden_subprocess_kwargs(),
        )
    except subprocess.TimeoutExpired:
        logger.warning("AutoConfig probe timed out for '%s' in %s", model_name, target_dir)
        return None
    except Exception as exc:
        logger.warning("AutoConfig probe could not spawn for '%s': %s", model_name, exc)
        return None
    if result.returncode == 0:
        return True
    err = (result.stderr or "").strip()
    if _stderr_is_transient(err):
        logger.warning("AutoConfig probe transient failure for '%s': %s", model_name, err)
        return None
    logger.info("AutoConfig probe parse failure for '%s' in %s: %s", model_name, target_dir, err)
    return False


def _probe_cache_key(model_name: str) -> str:
    """Cache key for the probe result. A local checkpoint can be overwritten in place, so
    fold in a cheap config.json signature (size + mtime) and re-probe when it changes.
    Remote ids key by name alone (resolving a Hub revision would need a pre-activation hub
    import that pins the wrong env)."""
    try:
        config_path = (Path(model_name) / "config.json").resolve()
        st = config_path.stat()
    except OSError:
        return model_name
    return f"{config_path}\0{st.st_size}:{st.st_mtime_ns}"


def _probe_tier(
    model_name: str,
    hf_token: str | None,
    reason: str,
    *,
    include_default: bool = False,
    floor: str = "530",
) -> str:
    """Lowest tier whose built-in parser loads the config; *floor* is the fail-safe.

    Escalates ``_PROBE_TIER_ORDER`` (prefixed with the ambient ``default`` tier when
    ``include_default``), returning the first that parses; never raises or escalates on
    uncertainty:
      - first success wins (cached unless a lower tier was skipped);
      - transient failure (auth/network/offline) -> *floor*, uncached;
      - a skipped/uninstallable sidecar -> uncached (a lower tier may yet be the answer);
      - all tiers probed, none parse -> remote-code/custom model_type; keep *floor*.

    Known-5.x callers use ``floor='530'``; weak-signal callers (config saved by transformers
    5.x) use ``include_default=True, floor='default'`` so a model that still parses on 4.57.x
    stays on the default. Cached per _probe_cache_key (process lifetime). No Hub sha is
    resolved: that would import huggingface_hub before the sidecar is on sys.path.
    """
    if os.environ.get("UNSLOTH_DISABLE_TIER_PROBE", "").lower() in ("1", "true", "yes"):
        return floor
    key = _probe_cache_key(model_name)
    # Key by probe mode: the default-first path can return 'default', which must not be
    # reused for a tokenizer/known-5.x caller (floor='530'). Legacy 530 keeps the bare key.
    if include_default or floor != "530":
        key = f"{key}\0floor={floor}:def={int(include_default)}"
    if key in _probe_tier_cache:
        return _probe_tier_cache[key]

    def _cache(tier: str, *, skipped: bool) -> str:
        # Do not pin a result that depended on a skipped lower tier: once that sidecar is
        # available the lowest valid tier may differ, so re-probe next call.
        if not skipped:
            _probe_tier_cache[key] = tier
        return tier

    venvs = _probe_tier_venvs()
    order = (("default",) + _PROBE_TIER_ORDER) if include_default else _PROBE_TIER_ORDER
    probed_count = 0
    skipped_any = False
    for tier in order:
        target_dir, ensure_fn = venvs[tier]
        try:
            available = ensure_fn()
        except Exception:
            available = False
        if not available:
            skipped_any = True
            continue
        probed_count += 1
        ok = _probe_autoconfig(target_dir, model_name, hf_token)
        if ok is True:
            logger.info(
                "Transformers tier %s selected for %s (AutoConfig probe; %s)",
                tier,
                model_name,
                reason,
            )
            return _cache(tier, skipped = skipped_any)
        if ok is None:
            logger.info("Tier probe inconclusive for %s (%s); using %s", model_name, reason, floor)
            return floor  # transient: retry next load

    # Nothing parsed. Only treat it as conclusive (and cache) when every tier was actually
    # probed; a skipped sidecar means the environment is incomplete, so retry uncached.
    if skipped_any or probed_count == 0:
        logger.info(
            "Tier probe incomplete for %s (%s); using %s (uncached)", model_name, reason, floor
        )
        return floor
    logger.info(
        "Transformers tier %s selected for %s (AutoConfig probe found no higher tier; %s)",
        floor,
        model_name,
        reason,
    )
    return _cache(floor, skipped = False)


def _norm_separators(s: str) -> str:
    """Collapse ``_``/whitespace to ``-`` (underscore aliases) but keep ``.`` so a
    version dot (``qwen3.5``) isn't conflated with a size separator (``Qwen3-5B``)."""
    return "".join("-" if ch in "_ \t" else ch for ch in s)


def _looks_like_hf_id(value: str) -> bool:
    """True if *value* looks like a Hub id (``org/name``), not a local path. An
    existing path is treated as a path, mirroring transformers' own resolution."""
    if not value or not value.strip():
        return False
    if os.path.isabs(value) or value.startswith((".", "~")) or "\\" in value:
        return False
    if os.path.exists(value):
        return False
    return value.count("/") <= 1


def _tier_from_name(name: str) -> tuple[str, str] | None:
    """``(tier, reason)`` from name substrings (order 510 > 550 > 530), or ``None``.

    Underscore aliases match (``Qwen3_5`` == ``Qwen3.5``); a dot-version substring
    matches only the dot/underscore form, never a hyphen, so ``Qwen3-6B`` size names
    aren't promoted.
    """
    lowered = name.lower()
    norm = _norm_separators(lowered)
    dotted = lowered.replace("_", ".")
    if "assistant" in lowered and ("gemma-4" in norm or "gemma4" in norm):
        return "510", "gemma-4 assistant variant"
    for substrings, tier in (
        (TRANSFORMERS_510_MODEL_SUBSTRINGS, "510"),
        (TRANSFORMERS_550_MODEL_SUBSTRINGS, "550"),
        (TRANSFORMERS_5_MODEL_SUBSTRINGS, "530"),
    ):
        for s in substrings:
            if "." in s:
                if s in lowered or s in dotted:
                    return tier, s
            elif s in lowered or _norm_separators(s) in norm:
                return tier, s
    return None


def _higher_tier_name_override(name_hint: str | None) -> str | None:
    """510/550 tier if *name_hint* names a higher-tier model, else ``None``. Qwen3.6
    reuses Qwen3.5 config ids but needs the 5.5 sidecar, so a name hint overrides 530."""
    if not name_hint:
        return None
    hint = _tier_from_name(name_hint)
    return hint[0] if hint is not None and hint[0] in ("510", "550") else None


def get_transformers_tier(
    model_name: str,
    hf_token: str | None = None,
    probe: bool = True,
) -> str:
    """Return the transformers tier required for *model_name*.

    Returns ``"510"`` for models needing transformers 5.10.x (Gemma 4 Unified),
    ``"550"`` for models needing transformers 5.5.0 (Gemma 4),
    ``"530"`` for models needing transformers 5.3.0 (e.g. Ministral-3, Qwen3 MoE),
    or ``"default"`` for everything else (4.57.x).

    Strong signals (architecture/model_type, name substrings) are fast paths. For local paths,
    ``config.json`` is checked before name heuristics to avoid false-positives from directory
    name fragments. When the only signal is the 5.x tokenizer class, the exact tier is resolved
    by probing AutoConfig in each sidecar; a config saved by transformers 5.x with no fast-path
    match is probed default-first, catching a new 5.x-only arch while 4.57.x-loadable models
    stay on default.

    ``probe=False`` skips the sidecar subprocesses (used by the cheap
    :func:`needs_transformers_5`); it still classifies via cheap signals (a 5.x-saved config
    returns ``"530"``). ``probe=True`` (the activation path) resolves the exact tier.

    Higher 5.x tiers run first.
    """
    # Local path: trust config.json. If its arch matches a known sidecar, return;
    # else fall back to the HF id in the config (not the folder name) for renamed dirs.
    local_cfg = Path(model_name) / "config.json"
    if _safe_is_file(local_cfg):
        cfg = _load_config_json(model_name, hf_token)
        if cfg is not None:
            if _config_needs_510(cfg):
                logger.info(
                    "Transformers tier 510 selected for %s (local config.json check)",
                    model_name,
                )
                return "510"
            if _config_needs_550(cfg):
                logger.info(
                    "Transformers tier 550 selected for %s (local config.json check)",
                    model_name,
                )
                return "550"
            if _config_needs_530(cfg):
                # Qwen3.6 reuses Qwen3.5 config ids but needs 5.5 by name. Only a real
                # Hub id (or the folder basename) may override 530, so a stale local
                # path in _name_or_path can't flip a correct 530 config to 550.
                base = _resolve_base_model(model_name)
                hint_src = (
                    base
                    if (base != model_name and _looks_like_hf_id(base))
                    else Path(model_name).name
                )
                override = _higher_tier_name_override(hint_src)
                if override is not None:
                    logger.info(
                        "Transformers tier %s selected for %s (name overrides 530 config)",
                        override,
                        model_name,
                    )
                    return override
                logger.info(
                    "Transformers tier 530 selected for %s (local config.json check)",
                    model_name,
                )
                return "530"
            # Unknown arch: resolve the base id from config. A resolved local dir
            # recurses (config check); a Hub id uses name rules only (no network).
            resolved = _resolve_base_model(model_name)
            if resolved != model_name:
                if _safe_is_dir(Path(resolved)):
                    tier = get_transformers_tier(resolved, hf_token, probe = probe)
                    if tier != "default":
                        logger.info(
                            "Transformers tier %s selected for %s (resolved local path: %s)",
                            tier,
                            model_name,
                            resolved,
                        )
                        return tier
                elif _looks_like_hf_id(resolved):
                    result = _tier_from_name(resolved)
                    if result is not None:
                        tier, match = result
                        logger.info(
                            "Transformers tier %s selected for %s (resolved HF ID: %s, match: %s)",
                            tier,
                            model_name,
                            resolved,
                            match,
                        )
                        return tier
            local_tc = Path(model_name) / "tokenizer_config.json"
            if _safe_is_file(local_tc) and _check_tokenizer_config_needs_v5(model_name, hf_token):
                if not probe:
                    return "530"
                return _probe_tier(model_name, hf_token, "local tokenizer needs 5.x")
            if _config_saved_by_transformers_5(cfg):
                if not probe:
                    return "530"  # cheap 5.x hint; the real path resolves the exact tier
                tier = _probe_tier(
                    model_name,
                    hf_token,
                    "local config saved by transformers 5.x",
                    include_default = True,
                    floor = "default",
                )
                if tier != "default":
                    return tier
            logger.info(
                "Transformers tier default (4.57.x) selected for %s (local config.json no match)",
                model_name,
            )
            return "default"

    # --- Fast substring checks (no I/O) ------------------------------------
    result = _tier_from_name(model_name)
    if result is not None:
        tier, match = result
        logger.info(
            "Transformers tier %s selected for %s (substring match: %s)",
            tier,
            model_name,
            match,
        )
        return tier

    # --- Slow config fallbacks (network for HF IDs; authenticated with hf_token) --------
    if _check_config_needs_510(model_name, hf_token):
        logger.info("Transformers tier 510 selected for %s (config.json check)", model_name)
        return "510"
    if _check_config_needs_550(model_name, hf_token):
        logger.info("Transformers tier 550 selected for %s (config.json check)", model_name)
        return "550"
    if _check_config_needs_530(model_name, hf_token):
        # Qwen3.6 reuses Qwen3.5 config ids but needs 5.5 by name; honor a real Hub-id name
        # hint from _name_or_path before selecting 530.
        remote_cfg = _load_config_json(model_name, hf_token) or {}
        base = remote_cfg.get("_name_or_path") or remote_cfg.get("model_name")
        override = _higher_tier_name_override(
            base if isinstance(base, str) and base != model_name else None
        )
        if override is not None:
            logger.info(
                "Transformers tier %s selected for %s (name overrides 530 config)",
                override,
                model_name,
            )
            return override
        logger.info("Transformers tier 530 selected for %s (config.json check)", model_name)
        return "530"
    if _check_tokenizer_config_needs_v5(model_name, hf_token):
        if not probe:
            return "530"
        return _probe_tier(model_name, hf_token, "tokenizer needs 5.x")

    if _config_saved_by_transformers_5(_cached_config_json(model_name, hf_token)):
        if not probe:
            return "530"  # cheap 5.x hint; the real path resolves the exact tier
        tier = _probe_tier(
            model_name,
            hf_token,
            "config saved by transformers 5.x",
            include_default = True,
            floor = "default",
        )
        if tier != "default":
            return tier

    logger.info("Transformers tier default (4.57.x) selected for %s (no match)", model_name)
    return "default"


def needs_transformers_5(model_name: str) -> bool:
    """Return True if *model_name* requires any transformers 5.x version.

    Convenience wrapper around :func:`get_transformers_tier`. Passes ``probe=False`` so a
    log-only parent caller never spawns sidecar probes (the worker re-resolves the exact
    tier with ``probe=True`` on the real activation path).
    """
    return get_transformers_tier(model_name, probe = False) != "default"


# ---------------------------------------------------------------------------
# Version switching (in-process — used only by export)
# ---------------------------------------------------------------------------


def _get_in_memory_version() -> str | None:
    """Return the transformers version currently loaded in this process."""
    tf = sys.modules.get("transformers")
    if tf is not None:
        return getattr(tf, "__version__", None)
    return None


# All top-level prefixes that hold references to transformers internals.
_PURGE_PREFIXES = (
    "transformers",
    "huggingface_hub",
    "unsloth",
    "unsloth_zoo",
    "peft",
    "trl",
    "accelerate",
    "auto_gptq",
    # NOTE: bitsandbytes is intentionally EXCLUDED -- it registers torch custom
    # operators via torch.library.define() into torch's global registry, which
    # survives module purge; re-importing after purge -> duplicate registration
    # -> crash.
    # Our own modules that import from transformers at module level.
    "utils.models",
    "core.training",
    "core.inference",
    "core.export",
)


def _purge_modules() -> int:
    """Remove all cached modules for transformers and its dependents.

    Returns the number of modules purged.
    """
    importlib.invalidate_caches()
    to_remove = [
        k
        for k in list(sys.modules.keys())
        if any(k == p or k.startswith(p + ".") for p in _PURGE_PREFIXES)
    ]
    for key in to_remove:
        del sys.modules[key]
    return len(to_remove)


_VENV_T5_530_PACKAGES = (
    f"transformers=={TRANSFORMERS_530_VERSION}",
    "huggingface_hub==1.8.0",
    "hf_xet==1.4.2",
    "tiktoken",
)

_VENV_T5_510_PACKAGES = (
    f"transformers=={TRANSFORMERS_510_VERSION}",
    "huggingface_hub==1.8.0",
    "hf_xet==1.4.2",
    "tiktoken",
)

_VENV_T5_550_PACKAGES = (
    f"transformers=={TRANSFORMERS_550_VERSION}",
    "huggingface_hub==1.8.0",
    "hf_xet==1.4.2",
    "tiktoken",
)

# Backwards-compat alias
_VENV_T5_PACKAGES = _VENV_T5_550_PACKAGES


def _venv_dir_is_valid(venv_dir: str, packages: tuple[str, ...]) -> bool:
    """Return True if *venv_dir* has all *packages* at the correct versions."""
    if not os.path.isdir(venv_dir) or not os.listdir(venv_dir):
        return False
    for pkg_spec in packages:
        parts = pkg_spec.split("==")
        pkg_name = parts[0]
        pkg_version = parts[1] if len(parts) > 1 else None
        pkg_name_norm = pkg_name.replace("-", "_")
        # Directory must exist.
        if not any(
            (Path(venv_dir) / d).is_dir() for d in (pkg_name_norm, pkg_name_norm.replace("_", "-"))
        ):
            return False
        # Unpinned packages: existence is enough.
        if pkg_version is None:
            continue
        # Check version via .dist-info metadata.
        dist_info_found = False
        for di in Path(venv_dir).glob(f"{pkg_name_norm}-*.dist-info"):
            metadata = di / "METADATA"
            if not metadata.is_file():
                continue
            for line in metadata.read_text(errors = "replace").splitlines():
                if line.startswith("Version:"):
                    installed_ver = line.split(":", 1)[1].strip()
                    if installed_ver != pkg_version:
                        logger.warning(
                            "%s has %s==%s but need %s -- venv will be wiped and reinstalled",
                            venv_dir,
                            pkg_name,
                            installed_ver,
                            pkg_version,
                        )
                        return False
                    dist_info_found = True
                    break
            if dist_info_found:
                break
        if not dist_info_found:
            return False
    return True


def _venv_t5_is_valid() -> bool:
    """Backwards-compat: check the Gemma 4 sidecar venv."""
    return _venv_dir_is_valid(_VENV_T5_550_DIR, _VENV_T5_550_PACKAGES)


def _install_to_dir(pkg: str, target_dir: str) -> bool:
    """Install a single package into *target_dir*, preferring uv then pip."""
    # Try uv first (faster) if on PATH -- do NOT install uv at runtime.
    if shutil.which("uv"):
        result = subprocess.run(
            [
                "uv",
                "pip",
                "install",
                "--python",
                sys.executable,
                "--target",
                target_dir,
                "--no-deps",
                "--upgrade",
                pkg,
            ],
            stdout = subprocess.PIPE,
            stderr = subprocess.STDOUT,
            text = True,
            env = child_env_without_native_path_secret(),
            **_windows_hidden_subprocess_kwargs(),
        )
        if result.returncode == 0:
            return True
        logger.warning("uv install of %s failed, falling back to pip", pkg)

    # Fallback to pip.
    result = subprocess.run(
        [
            sys.executable,
            "-m",
            "pip",
            "install",
            "--target",
            target_dir,
            "--no-deps",
            "--upgrade",
            pkg,
        ],
        stdout = subprocess.PIPE,
        stderr = subprocess.STDOUT,
        text = True,
        env = child_env_without_native_path_secret(),
        **_windows_hidden_subprocess_kwargs(),
    )
    if result.returncode != 0:
        logger.error("install failed:\n%s", result.stdout)
        return False
    return True


def _ensure_venv_dir(venv_dir: str, packages: tuple[str, ...], label: str) -> bool:
    """Ensure *venv_dir* exists with all *packages*. Install if missing."""
    if _venv_dir_is_valid(venv_dir, packages):
        return True

    logger.warning("%s not found or incomplete at %s -- installing at runtime", label, venv_dir)
    shutil.rmtree(venv_dir, ignore_errors = True)
    os.makedirs(venv_dir, exist_ok = True)
    total = len(packages)
    for idx, pkg in enumerate(packages, start = 1):
        logger.info("Installing %s (%d/%d) into %s ...", pkg, idx, total, venv_dir)
        if not _install_to_dir(pkg, venv_dir):
            return False
    logger.info("Installed %s to %s", label, venv_dir)
    return True


def _ensure_venv_t5_530_exists() -> bool:
    """Ensure .venv_t5_530/ exists with transformers 5.3.0."""
    return _ensure_venv_dir(_VENV_T5_530_DIR, _VENV_T5_530_PACKAGES, "transformers 5.3.0")


def _ensure_venv_t5_550_exists() -> bool:
    """Ensure .venv_t5_550/ exists with transformers 5.5.0."""
    return _ensure_venv_dir(
        _VENV_T5_550_DIR,
        _VENV_T5_550_PACKAGES,
        f"transformers {TRANSFORMERS_550_VERSION}",
    )


def _ensure_venv_t5_510_exists() -> bool:
    """Ensure .venv_t5_510/ exists with transformers 5.10.x."""
    return _ensure_venv_dir(
        _VENV_T5_510_DIR,
        _VENV_T5_510_PACKAGES,
        f"transformers {TRANSFORMERS_510_VERSION}",
    )


def _ensure_venv_t5_exists() -> bool:
    """Backwards-compat: ensure the Gemma 4 5.5 sidecar venv exists."""
    return _ensure_venv_t5_550_exists()


# --- llm-compressor-main shadow (FP8/FP4 export of newer-transformers models) ---------------------
# Exact, reproducible pins (bump deliberately in review). Full 40-char SHA validated to FP8-quantize
# Qwen3.5 / Gemma-4 / Llama.
_LLMC_MAIN_TRANSFORMERS = "5.10.2"
_LLMC_MAIN_SHA = "973c9c539a84dd9efaf74e115ede5ca419704c18"
_LLMC_MAIN_COMPRESSED_TENSORS = "0.17.2a20260702"
# Installed --no-deps (torch untouched); the full runtime set llm-compressor main needs, pinned.
_VENV_LLMCOMPRESSOR_SPECS = (
    f"transformers=={_LLMC_MAIN_TRANSFORMERS}",
    f"llmcompressor @ git+https://github.com/vllm-project/llm-compressor@{_LLMC_MAIN_SHA}",
    f"compressed-tensors=={_LLMC_MAIN_COMPRESSED_TENSORS}",
    "huggingface-hub==1.21.0",
    "hf-xet==1.5.1",
    "tokenizers==0.22.2",
    "safetensors==0.8.0",
    "accelerate==1.14.0",
    "datasets==5.0.0",
    "pydantic==2.13.4",
    "pydantic-core==2.46.4",
    "typing-inspection==0.4.2",
    "loguru==0.7.3",
    "pyyaml==6.0.3",
    "nvidia-ml-py==13.610.43",
    "pillow==12.3.0",
    "auto-round==0.13.1",
    "regex==2026.6.28",
)
# Fingerprint of the pin set; bump the trailing schema version to force a rebuild on layout changes.
_LLMC_SHADOW_FINGERPRINT = (
    f"{_LLMC_MAIN_SHA}|{_LLMC_MAIN_TRANSFORMERS}|{_LLMC_MAIN_COMPRESSED_TENSORS}|schema=1"
)
_LLMC_SHADOW_MARKER = ".unsloth_llmc_fingerprint"


def _llmcompressor_main_disabled() -> bool:
    """True if the operator forbids the llm-compressor-main shadow (air-gapped / locked-down)."""
    return os.environ.get("UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN", "").strip().lower() in {
        "1",
        "true",
        "yes",
        "on",
    }


def _llmcompressor_shadow_is_valid() -> bool:
    """True if the shadow dir exists with a marker matching the current pin fingerprint."""
    marker = Path(_VENV_LLMCOMPRESSOR_DIR) / _LLMC_SHADOW_MARKER
    try:
        return marker.is_file() and marker.read_text().strip() == _LLMC_SHADOW_FINGERPRINT
    except Exception:
        return False


def _ensure_venv_llmcompressor_exists() -> bool:
    """Ensure .venv_llmcompressor/ has the pinned llm-compressor-main stack. Install if missing.

    All specs are installed with --no-deps into a --target dir (mirrors the transformers sidecars),
    so the workspace torch is never touched. Returns True on success.
    """
    if _llmcompressor_shadow_is_valid():
        return True
    if _llmcompressor_main_disabled():
        logger.warning(
            "llm-compressor-main shadow needed but UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN is set; "
            "compressed export of newer-transformers models will fail fast."
        )
        return False
    if _env_offline():
        logger.warning(
            "llm-compressor-main shadow missing and HF/offline mode is set; cannot provision it."
        )
        return False

    logger.warning(
        "Provisioning llm-compressor-main shadow at %s (one-time, ~a few hundred MB, no torch) ...",
        _VENV_LLMCOMPRESSOR_DIR,
    )
    shutil.rmtree(_VENV_LLMCOMPRESSOR_DIR, ignore_errors = True)
    os.makedirs(_VENV_LLMCOMPRESSOR_DIR, exist_ok = True)

    # Prefer uv (faster) then pip; install every spec at once, --no-deps, prereleases allowed
    # (compressed-tensors ships as a pre-release).
    base = [
        "--target",
        _VENV_LLMCOMPRESSOR_DIR,
        "--no-deps",
        "--prerelease=allow",
        *_VENV_LLMCOMPRESSOR_SPECS,
    ]
    cmds = []
    if shutil.which("uv"):
        cmds.append(["uv", "pip", "install", "--python", sys.executable, *base])
    cmds.append(
        [
            sys.executable,
            "-m",
            "pip",
            "install",
            *[a for a in base if a != "--prerelease=allow"],
            "--pre",
        ]
    )

    last_out = ""
    for cmd in cmds:
        result = subprocess.run(
            cmd,
            stdout = subprocess.PIPE,
            stderr = subprocess.STDOUT,
            text = True,
            env = child_env_without_native_path_secret(),
            **_windows_hidden_subprocess_kwargs(),
        )
        last_out = result.stdout or ""
        if result.returncode == 0:
            try:
                (Path(_VENV_LLMCOMPRESSOR_DIR) / _LLMC_SHADOW_MARKER).write_text(
                    _LLMC_SHADOW_FINGERPRINT
                )
            except Exception:
                pass
            logger.info("Provisioned llm-compressor-main shadow at %s", _VENV_LLMCOMPRESSOR_DIR)
            return True
        logger.warning("llm-compressor-main shadow install failed with %s; trying next", cmd[0])

    logger.error(
        "Failed to provision llm-compressor-main shadow (spec: llmcompressor@%s). Output:\n%s",
        _LLMC_MAIN_SHA,
        last_out[-4000:],
    )
    return False


def llmcompressor_shadow_pythonpath() -> str | None:
    """Provision (lazily) the llm-compressor-main shadow and return its sys.path entry, or None.

    Returns None when the shadow is disabled (UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN), offline, or
    provisioning failed - callers then fall back to the fail-fast path.
    """
    if _llmcompressor_main_disabled():
        return None
    if _ensure_venv_llmcompressor_exists():
        return _VENV_LLMCOMPRESSOR_DIR
    return None


def _activate_venv(venv_dir: str, label: str) -> None:
    """Prepend *venv_dir* to sys.path, purge stale modules, reimport."""
    if venv_dir not in sys.path:
        sys.path.insert(0, venv_dir)
        logger.info("Prepended %s to sys.path", venv_dir)

    count = _purge_modules()
    logger.info("Purged %d cached modules", count)

    import transformers

    logger.info("Loaded transformers %s (%s)", transformers.__version__, label)


def _deactivate_5x() -> None:
    """Remove all .venv_t5_*/ dirs from sys.path, purge stale modules, reimport."""
    for d in (_VENV_T5_530_DIR, _VENV_T5_550_DIR, _VENV_T5_510_DIR):
        while d in sys.path:
            sys.path.remove(d)
    logger.info("Removed venv_t5 dirs from sys.path")

    count = _purge_modules()
    logger.info("Purged %d cached modules", count)

    import transformers

    logger.info("Reverted to transformers %s", transformers.__version__)


def ensure_transformers_version(model_name: str) -> None:
    """Ensure the correct ``transformers`` version is active for *model_name*.

    Uses sys.path with .venv_t5_510/, .venv_t5_550/, or .venv_t5_530/
    (pre-installed by setup.sh):
      • Need 5.10.x → prepend .venv_t5_510/ to sys.path, purge modules.
      • Need 5.5.0 → prepend .venv_t5_550/ to sys.path, purge modules.
      • Need 5.3.0 → prepend .venv_t5_530/ to sys.path, purge modules.
      • Need 4.x  → remove all .venv_t5_*/ from sys.path, purge modules.

    For custom-named LoRA adapters, the base model is resolved before checking
    (from ``adapter_config.json`` or, for adapter_model-only LoRAs, the directory
    name).

    NOTE: Training and inference use subprocess isolation instead. Used only by
    the export path (routes/export.py).
    """
    # Only pre-resolve for LoRA adapter dirs; see activate_transformers_for_subprocess.
    if _is_lora_adapter_dir(Path(model_name)):
        resolved = _resolve_base_model(model_name)
    else:
        resolved = model_name
    tier = get_transformers_tier(resolved)
    if model_name != resolved and _safe_is_file(Path(model_name) / "config.json"):
        # Gate on a real local config.json: a checkpoint carries config the base may not
        # surface, but path names alone must not upgrade a plain adapter.
        tier = _higher_tier(tier, get_transformers_tier(model_name))

    if tier == "510":
        target_version = TRANSFORMERS_510_VERSION
        venv_dir = _VENV_T5_510_DIR
        ensure_fn = _ensure_venv_t5_510_exists
    elif tier == "550":
        target_version = TRANSFORMERS_550_VERSION
        venv_dir = _VENV_T5_550_DIR
        ensure_fn = _ensure_venv_t5_550_exists
    elif tier == "530":
        target_version = TRANSFORMERS_530_VERSION
        venv_dir = _VENV_T5_530_DIR
        ensure_fn = _ensure_venv_t5_530_exists
    else:
        target_version = TRANSFORMERS_DEFAULT_VERSION
        venv_dir = None
        ensure_fn = None

    target_major = int(target_version.split(".")[0])

    # Check what's actually loaded in memory
    in_memory = _get_in_memory_version()

    logger.info(
        "Version check for '%s' (resolved: '%s'): need=%s, in_memory=%s",
        model_name,
        resolved,
        target_version,
        in_memory,
    )

    # --- Already correct? ---------------------------------------------------
    if in_memory is not None:
        if in_memory == target_version:
            logger.info(
                "transformers %s already loaded — correct for '%s'",
                in_memory,
                model_name,
            )
            return
        # Different 5.x -> need to switch (e.g. 5.3.0 loaded but need 5.10.x).
        in_memory_major = int(in_memory.split(".")[0])
        if in_memory_major == target_major and venv_dir is None:
            # Both are default (4.x) — close enough.
            logger.info(
                "transformers %s already loaded — correct for '%s'",
                in_memory,
                model_name,
            )
            return

    # --- Switch version -----------------------------------------------------
    if venv_dir is not None:
        # First remove any other 5.x venv from sys.path.
        _deactivate_5x()
        if not ensure_fn():
            raise RuntimeError(
                f"Cannot activate transformers {target_version}: " f"venv missing at {venv_dir}"
            )
        logger.info("Activating transformers %s…", target_version)
        _activate_venv(venv_dir, f"transformers {target_version}")
    else:
        logger.info("Reverting to default transformers %s…", TRANSFORMERS_DEFAULT_VERSION)
        _deactivate_5x()

    final = _get_in_memory_version()
    logger.info("✓ transformers version is now %s", final)