llama_cpp.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

"""
llama-server inference backend for GGUF models.

Manages a llama-server subprocess and proxies chat completions
through its OpenAI-compatible /v1/chat/completions endpoint.
"""

import atexit
import contextlib
import json
import os
import re
import struct
import structlog
from loggers import get_logger
import shutil
import socket
import subprocess
import sys
import threading
import time
from pathlib import Path
from typing import Generator, Iterable, List, Optional
from urllib.parse import urlparse

import httpx

from core.tool_healing import (
    _TC_END_TAG_RE,
    _TC_FUNC_CLOSE_RE,
    _TC_FUNC_START_RE,
    _TC_JSON_START_RE,
    _TC_PARAM_CLOSE_RE,
    _TC_PARAM_START_RE,
    _TOOL_ALL_PATS,
    _TOOL_CLOSED_PATS,
    parse_tool_calls_from_text,
    strip_tool_call_markup,
)
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,
)
from core.inference.tool_call_parser import (
    parse_tool_calls_from_text as _shared_parse_tool_calls_from_text,
)

logger = get_logger(__name__)


# ── Pre-compiled patterns for plan-without-action re-prompt ──
# Forward-looking intent signals that indicate the model is
# describing what it *will* do rather than giving a final answer.
_INTENT_SIGNAL = re.compile(
    r"(?i)("
    # Direct intent: "I'll ...", "I will ...", "Let me ...", "I am going to ..."
    # Handles both straight and curly apostrophes.
    # Excludes "I can", "I should", "I want to", "let's" which
    # appear frequently in direct answers / explanations.
    # Negative lookahead drops negated forms ("I will not", "I'll never")
    # so a refusal doesn't trigger a re-prompt.
    r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
    r"|"
    # Step/plan framing: "First ...", "Step 1:", "Here's my plan"
    r"\b(?:first\b|step \d+:?|here['\u2019]?s (?:my |the |a )?(?:plan|approach))"
    r"|"
    # "Now I" / "Next I" patterns
    r"\b(?:now i|next i)\b"
    r")"
)
_MAX_REPROMPTS = 3

# Without max_tokens, llama-server defaults to n_predict = n_ctx (up to
# 262144 for Qwen3.5), producing many-minute zombie decodes when cancel
# fails. t_max_predict_ms is a wall-clock backstop applied unconditionally,
# but the llama.cpp README notes it ONLY fires after a newline has been
# generated -- a model stuck in a long unbroken non-newline sequence is
# unbounded by it. So we still want a token cap as the front-line limiter.
#
# The cap is the model's effective context length when we know it,
# falling back to a generous floor when metadata is unavailable. 4096 was
# too low: Qwen3 / gpt-oss reasoning traces routinely exceed it, and any
# OpenAI-API caller that omits max_tokens (langchain, llama-index, raw
# curl) sees responses silently truncated mid-sentence.
_DEFAULT_MAX_TOKENS_FLOOR = 32768
_DEFAULT_T_MAX_PREDICT_MS = 600_000  # 10 min
_REPROMPT_MAX_CHARS = 2000

# ── Pre-compiled patterns for GGUF shard detection ───────────
_SHARD_FULL_RE = re.compile(r"^(.*)-(\d{5})-of-(\d{5})\.gguf$")
_SHARD_RE = re.compile(r"^(.*)-\d{5}-of-\d{5}\.gguf$")


# ── Sliding-window-pattern resolver ───────────────────────────
# Resolves the per-layer SWA mask when a GGUF reports a sliding window
# but no `sliding_window_pattern` field. Tier order in
# `_resolve_swa_pattern`: GGUF metadata, on-disk cache, bootstrap dict
# below, transformers introspection, HF Hub config.json, legacy 1/4
# fallback. Period N means layer i is SWA iff `(i + 1) % N != 0`,
# matching transformers. Skipped on purpose: phi3 (no key/val length
# in GGUF, window >= ctx anyway), qwen2 family (converter strips
# sliding_window when use_sliding_window=False), mistral v0.1/v0.2
# (all-SWA can't be expressed as a period).
_BOOTSTRAP_SWA_DEFAULTS: dict[str, int] = {
    "gemma2": 2,  # Gemma2Config.sliding_window_pattern
    "gemma3": 6,  # Gemma3TextConfig.sliding_window_pattern
    "gemma3n": 5,  # text_config.layer_types: SWA*4 + FULL
    "gpt_oss": 2,  # text_config.layer_types: alternating
    "cohere2": 4,  # Cohere2Config.sliding_window_pattern
}

# Process-wide cache backed by JSON on disk. Values are int period or
# list[bool] mask. Lazy-loaded.
_SWA_CACHE: Optional[dict] = None
_SWA_CACHE_LOCK = threading.Lock()


def _probe_dns_dead(host: str = "huggingface.co", timeout: float = 2.0) -> bool:
    """Quick DNS check. Runs on a daemon thread so concurrent sockets
    in the same process are not affected by socket.setdefaulttimeout."""
    result: list[Optional[bool]] = [None]

    def _probe() -> None:
        try:
            socket.gethostbyname(host)
            result[0] = False
        except Exception:
            result[0] = True

    t = threading.Thread(target = _probe, daemon = True)
    t.start()
    t.join(timeout)
    # Thread still running -> resolver wedged -> treat as dead.
    return True if result[0] is None else result[0]


@contextlib.contextmanager
def _hf_offline_if_dns_dead():
    """Set HF_HUB_OFFLINE for the body of this block only when DNS to
    huggingface.co fails. Restores the env on exit so a transient
    resolver hiccup at the start of one load can't quarantine the whole
    process. Respects an explicit user setting (no-op if already set)."""
    if "HF_HUB_OFFLINE" in os.environ:
        yield False
        return
    if not _probe_dns_dead():
        yield False
        return

    transformers_was_set = "TRANSFORMERS_OFFLINE" in os.environ
    os.environ["HF_HUB_OFFLINE"] = "1"
    if not transformers_was_set:
        os.environ["TRANSFORMERS_OFFLINE"] = "1"
    logger.warning("huggingface.co unreachable; using local HF cache for this load.")
    try:
        yield True
    finally:
        os.environ.pop("HF_HUB_OFFLINE", None)
        if not transformers_was_set:
            os.environ.pop("TRANSFORMERS_OFFLINE", None)


def _swa_cache_path() -> Path:
    home = os.environ.get("UNSLOTH_STUDIO_HOME") or os.environ.get("STUDIO_HOME")
    base = Path(home) if home else Path.home() / ".unsloth" / "studio"
    return base / "swa_cache.json"


def _load_swa_cache() -> dict:
    global _SWA_CACHE
    with _SWA_CACHE_LOCK:
        if _SWA_CACHE is not None:
            return _SWA_CACHE
        try:
            with open(_swa_cache_path()) as f:
                _SWA_CACHE = json.load(f)
                if not isinstance(_SWA_CACHE, dict):
                    _SWA_CACHE = {}
        except (FileNotFoundError, json.JSONDecodeError, OSError):
            _SWA_CACHE = {}
        return _SWA_CACHE


def _save_swa_cache(cache: dict) -> None:
    try:
        path = _swa_cache_path()
        path.parent.mkdir(parents = True, exist_ok = True)
        tmp = path.with_suffix(".json.tmp")
        with open(tmp, "w") as f:
            json.dump(cache, f, indent = 2, sort_keys = True)
        tmp.replace(path)
    except OSError:
        pass


def _period_from_layer_types(layer_types: list) -> Optional[int]:
    """Smallest period N where `(i+1) % N != 0` matches the SWA mask,
    or None if no fixed period fits."""
    if not layer_types:
        return None
    is_swa = ["full" not in str(t).lower() for t in layer_types]
    n = len(is_swa)
    for N in range(1, n + 1):
        if all(((i + 1) % N != 0) == is_swa[i] for i in range(n)):
            return N
    return None


def _fetch_swa_entry_from_hf(repo_id: str) -> Optional[object]:
    try:
        from huggingface_hub import hf_hub_download

        cfg_path = hf_hub_download(repo_id, "config.json", repo_type = "model")
        with open(cfg_path) as f:
            cfg = json.load(f)
    except Exception:
        return None

    src = cfg.get("text_config") if isinstance(cfg.get("text_config"), dict) else cfg
    period = src.get("sliding_window_pattern")
    if isinstance(period, int) and period > 0:
        return period
    lt = src.get("layer_types")
    if isinstance(lt, list) and lt:
        return _period_from_layer_types(lt) or [
            "full" not in str(t).lower() for t in lt
        ]
    return None


def _arch_aliases(arch: str) -> tuple:
    # GGUF emits `falcon-h1`; HF model_type is `falcon_h1`. Normalise both ways.
    seen = []
    for a in (arch, arch.replace("-", "_"), arch.replace("_", "-")):
        if a and a not in seen:
            seen.append(a)
    return tuple(seen)


def _swa_entry_from_config_obj(cfg) -> Optional[object]:
    src = getattr(cfg, "text_config", None) or cfg
    period = getattr(src, "sliding_window_pattern", None)
    if isinstance(period, int) and period > 0:
        return period
    lt = getattr(src, "layer_types", None)
    if isinstance(lt, list) and lt:
        return _period_from_layer_types(lt) or [
            "full" not in str(t).lower() for t in lt
        ]
    return None


_SWA_PATTERN_SOURCE_RE = re.compile(
    r"sliding_window_pattern\s*(?::\s*[\w\[\], ]*)?\s*=\s*(\d+)"
)


def _resolve_swa_entry_from_transformers(arch: str) -> Optional[object]:
    """Default-instantiate the matching Config; on failure, regex-parse
    its source for `sliding_window_pattern = N`."""
    try:
        from transformers.models.auto.configuration_auto import (
            CONFIG_MAPPING,
            CONFIG_MAPPING_NAMES,
        )
    except Exception:
        return None

    cfg_class = None
    for alias in _arch_aliases(arch):
        if alias in CONFIG_MAPPING_NAMES:
            try:
                cfg_class = CONFIG_MAPPING[alias]
                break
            except Exception:
                cfg_class = None
    if cfg_class is None:
        return None

    try:
        if (entry := _swa_entry_from_config_obj(cfg_class())) is not None:
            return entry
    except Exception:
        pass

    import inspect

    candidates = [cfg_class]
    text_cfg_class = getattr(cfg_class, "sub_configs", {}).get("text_config")
    if text_cfg_class is not None:
        candidates.append(text_cfg_class)
    for cls in candidates:
        try:
            src = inspect.getsource(cls)
        except (OSError, TypeError):
            continue
        if m := _SWA_PATTERN_SOURCE_RE.search(src):
            period = int(m.group(1))
            if period > 0:
                return period
    return None


def _resolve_swa_pattern(
    arch: Optional[str],
    n_layers: Optional[int],
    source_repo_candidates: tuple = (),
    *,
    allow_network: Optional[bool] = None,
) -> Optional[list]:
    if not arch or not n_layers:
        return None
    if allow_network is None:
        allow_network = os.environ.get("UNSLOTH_STUDIO_OFFLINE", "0") not in (
            "1",
            "true",
            "True",
            "yes",
        )

    cache = _load_swa_cache()

    def _entry_to_mask(entry):
        if isinstance(entry, int) and entry > 0:
            return [(i + 1) % entry != 0 for i in range(n_layers)]
        if isinstance(entry, list) and entry:
            return [bool(entry[i % len(entry)]) for i in range(n_layers)]
        return None

    def _persist(entry):
        with _SWA_CACHE_LOCK:
            cache[arch] = entry
        _save_swa_cache(cache)

    if (entry := cache.get(arch)) is not None:
        if (mask := _entry_to_mask(entry)) is not None:
            return mask

    if (entry := _BOOTSTRAP_SWA_DEFAULTS.get(arch)) is not None:
        return _entry_to_mask(entry)

    entry = _resolve_swa_entry_from_transformers(arch)
    if entry is not None:
        _persist(entry)
        return _entry_to_mask(entry)

    # Tier 3: live HF fetch (with persistent caching of the result)
    if allow_network:
        for repo_id in source_repo_candidates:
            if not repo_id:
                continue
            entry = _fetch_swa_entry_from_hf(repo_id)
            if entry is not None:
                _persist(entry)
                return _entry_to_mask(entry)

    return None


def _hf_repo_from_url(url: Optional[str]) -> Optional[str]:
    """Strip `https://huggingface.co/owner/name(/...)` to `owner/name`."""
    if not url or "huggingface.co/" not in url:
        return None
    tail = url.split("huggingface.co/", 1)[1].rstrip("/")
    parts = tail.split("/")
    if len(parts) < 2:
        return None
    return f"{parts[0]}/{parts[1]}"


# Model size extraction — lazy import to avoid pulling in transformers
# at module level.  See PR description for the full explanation.
def _extract_model_size_b(model_id: str):
    from utils.models import extract_model_size_b

    return extract_model_size_b(model_id)


_TOOL_TEMPLATE_MARKERS = (
    "{%- if tools %}",
    "{%- if tools -%}",
    "{% if tools %}",
    "{% if tools -%}",
    '"role" == "tool"',
    "'role' == 'tool'",
    'message.role == "tool"',
    "message.role == 'tool'",
)


def detect_reasoning_flags(
    chat_template: Optional[str],
    model_identifier: Optional[str] = None,
    *,
    log_source: Optional[str] = None,
) -> dict:
    """Classify a chat template's reasoning and tool-calling capabilities.

    Returns a dict with the same five keys populated by the GGUF sniffer:
    ``supports_reasoning``, ``reasoning_style``
    (``"enable_thinking"`` | ``"reasoning_effort"``),
    ``reasoning_always_on``, ``supports_preserve_thinking``, and
    ``supports_tools``. Used by both the llama-server backend at load
    time and the safetensors/transformers paths in ``routes/inference``
    so the two agree on what the frontend will see.
    """
    flags = {
        "supports_reasoning": False,
        "reasoning_style": "enable_thinking",
        "reasoning_always_on": False,
        "supports_preserve_thinking": False,
        "supports_tools": False,
    }
    if not chat_template:
        return flags
    tpl = chat_template
    prefix = f"{log_source}: " if log_source else ""

    if "enable_thinking" in tpl:
        flags["supports_reasoning"] = True
        flags["reasoning_style"] = "enable_thinking"
        logger.info(f"{prefix}model supports reasoning (enable_thinking)")
    elif "reasoning_effort" in tpl:
        # gpt-oss / Harmony templates use reasoning_effort
        # ("low" | "medium" | "high") instead of a boolean.
        flags["supports_reasoning"] = True
        flags["reasoning_style"] = "reasoning_effort"
        logger.info(f"{prefix}model supports reasoning (reasoning_effort)")
    elif "thinking" in tpl:
        # DeepSeek uses 'thinking' instead of 'enable_thinking'
        normalized_id = (model_identifier or "").lower()
        if "deepseek" in normalized_id:
            flags["supports_reasoning"] = True
            logger.info(f"{prefix}model supports reasoning (DeepSeek thinking)")

    # Hardcoded <think> tags or reasoning_content in the template mean
    # thinking is always on (no toggle to disable it).
    if not flags["supports_reasoning"]:
        if ("<think>" in tpl and "</think>" in tpl) or "reasoning_content" in tpl:
            flags["supports_reasoning"] = True
            flags["reasoning_always_on"] = True
            logger.info(f"{prefix}model always reasons (<think> tags in template)")

    # preserve_thinking is an independent kwarg on some Qwen templates
    # that keeps historical <think> blocks in prior assistant turns.
    if "preserve_thinking" in tpl:
        flags["supports_preserve_thinking"] = True
        logger.info(f"{prefix}model supports preserve_thinking")

    if any(marker in tpl for marker in _TOOL_TEMPLATE_MARKERS):
        flags["supports_tools"] = True
        logger.info(f"{prefix}model supports tool calling")

    return flags


def _is_mtp_model_name(
    model_identifier: Optional[str],
    gguf_path: Optional[str] = None,
) -> bool:
    """Name-based MTP detector. Fallback for the metadata signal."""
    for cand in (model_identifier, Path(gguf_path).name if gguf_path else None):
        if cand and "-mtp" in cand.lower():
            return True
    return False


def _extra_args_set_spec_type(extra_args: Optional[Iterable[str]]) -> bool:
    """User passed --spec-type / --spec-default? llama-server takes a
    single --spec-type (comma-separated to chain), so suppress
    auto-emit when this is true."""
    if not extra_args:
        return False
    for raw in extra_args:
        tok = str(raw)
        if not tok.startswith("--"):
            continue
        flag = tok.split("=", 1)[0]
        if flag in ("--spec-type", "--spec-default"):
            return True
    return False


def _build_ngram_mod_flags(
    caps: Optional[dict],
    n_match: int = 24,
    n_min: int = 48,
    n_max: int = 64,
) -> list[str]:
    """Emit the right ngram-mod knob flags for the running llama-server.

    Post-rename builds expose ``--spec-ngram-mod-n-{match,min,max}``;
    pre-rename builds expose the legacy ``--spec-ngram-size-n`` /
    ``--draft-min`` / ``--draft-max``. ``caps`` comes from
    ``probe_server_capabilities``; ``ngram_mod_flavor`` tells us which
    set is real (vs a removal-stub entry). Returns ``[]`` when neither
    set is available so the caller can drop ngram-mod entirely.
    """
    flavor = caps.get("ngram_mod_flavor") if caps else None
    if flavor == "new":
        return [
            "--spec-ngram-mod-n-match",
            str(n_match),
            "--spec-ngram-mod-n-min",
            str(n_min),
            "--spec-ngram-mod-n-max",
            str(n_max),
        ]
    if flavor == "legacy":
        # Legacy llama.cpp before the spec arg rename: same knobs lived
        # under --spec-ngram-size-n (lookup length) and the generic
        # --draft-min / --draft-max (ngram size N range).
        return [
            "--spec-ngram-size-n",
            str(n_match),
            "--draft-min",
            str(n_min),
            "--draft-max",
            str(n_max),
        ]
    return []


# Canonical Speculative Decoding modes exposed by the Studio chat UI.
# The dropdown renders five options (auto, mtp, ngram, mtp+ngram, off);
# the load API also accepts legacy values that the original Switch and
# external callers emit (default, draft-mtp, ngram-mod, ngram-simple).
_CANONICAL_SPEC_MODES = {"auto", "mtp", "ngram", "mtp+ngram", "off", "ngram-simple"}
_LEGACY_SPEC_MODE_MAP = {
    "default": "auto",
    "draft-mtp": "mtp",
    "ngram-mod": "ngram",
}


def _canonicalize_spec_mode(value):
    """Map any accepted ``speculative_type`` input onto a canonical mode.

    Returns one of ``auto``, ``mtp``, ``ngram``, ``mtp+ngram``, ``off``,
    ``ngram-simple``, or ``None`` (callers treat ``None`` as ``auto``).
    Unknown strings collapse to ``auto`` so a stale UI value or typo
    falls back to the safe platform-aware path.
    """
    if value is None:
        return None
    if not isinstance(value, str):
        return None
    stripped = value.strip().lower()
    if not stripped:
        return None
    if stripped in _CANONICAL_SPEC_MODES:
        return stripped
    if stripped in _LEGACY_SPEC_MODE_MAP:
        return _LEGACY_SPEC_MODE_MAP[stripped]
    # llama.cpp comma-chains are emitted by old persisted state e.g.
    # "ngram-mod,draft-mtp"; collapse the most common one explicitly.
    pieces = [p.strip() for p in stripped.split(",") if p.strip()]
    has_mtp = any(p in ("mtp", "draft-mtp") for p in pieces)
    has_ngram = any(p in ("ngram", "ngram-mod") for p in pieces)
    if has_mtp and has_ngram:
        return "mtp+ngram"
    if has_mtp:
        return "mtp"
    if has_ngram:
        return "ngram"
    return "auto"


def _backfill_usage_from_timings(usage, timings):
    """Synthesize ``usage`` from llama-server's ``timings`` when the
    OpenAI-style usage block is missing or reports zero tokens.

    The Studio chat UI computes generation t/s from
    ``meta.usage.completion_tokens / totalStreamTime``. llama-server
    always populates ``timings.predicted_n`` (true decoded count) and
    ``timings.prompt_n``, but the ``usage`` field on the final SSE chunk
    can be absent or zero on some server builds / streaming
    configurations, which makes the UI fall back to wall-clock t/s and
    dilute speculative-decoding speedups.
    """
    if not timings:
        return usage
    if usage and usage.get("completion_tokens"):
        return usage
    predicted_n = timings.get("predicted_n")
    prompt_n = timings.get("prompt_n")
    if predicted_n is None and prompt_n is None:
        return usage
    out = dict(usage or {})
    if not out.get("completion_tokens") and predicted_n is not None:
        out["completion_tokens"] = predicted_n
    if not out.get("prompt_tokens") and prompt_n is not None:
        out["prompt_tokens"] = prompt_n
    out["total_tokens"] = int(out.get("prompt_tokens") or 0) + int(
        out.get("completion_tokens") or 0
    )
    return out


# Probe script run in a short-lived subprocess so the Vulkan instance never
# lives in the long-running backend process. Loads the bundled ggml Vulkan
# backend and prints "<idx>\t<free_bytes>\t<total_bytes>" per device. The
# indices are ggml's own Vulkan device ordinals -- the space
# GGML_VK_VISIBLE_DEVICES expects -- which need not match nvidia-smi order.
_VULKAN_PROBE_SCRIPT = r"""
import ctypes, os, sys
bindir = sys.argv[1]
if sys.platform == "win32":
    base_name, vk_name = "ggml-base.dll", "ggml-vulkan.dll"
    try:
        os.add_dll_directory(bindir)
    except Exception:
        pass
else:
    base_name, vk_name = "libggml-base.so", "libggml-vulkan.so"
try:
    ctypes.CDLL(os.path.join(bindir, base_name), mode=ctypes.RTLD_GLOBAL)
    lib = ctypes.CDLL(os.path.join(bindir, vk_name), mode=ctypes.RTLD_GLOBAL)
except OSError:
    sys.exit(0)
lib.ggml_backend_vk_get_device_count.restype = ctypes.c_int
lib.ggml_backend_vk_get_device_count.argtypes = []
lib.ggml_backend_vk_get_device_memory.restype = None
lib.ggml_backend_vk_get_device_memory.argtypes = [
    ctypes.c_int,
    ctypes.POINTER(ctypes.c_size_t),
    ctypes.POINTER(ctypes.c_size_t),
]
rows = []
for i in range(lib.ggml_backend_vk_get_device_count()):
    free, total = ctypes.c_size_t(0), ctypes.c_size_t(0)
    lib.ggml_backend_vk_get_device_memory(i, ctypes.byref(free), ctypes.byref(total))
    rows.append("%d\t%d\t%d" % (i, free.value, total.value))
sys.stdout.write("\n".join(rows))
"""


def _vulkan_lib_filename() -> str:
    return "ggml-vulkan.dll" if sys.platform == "win32" else "libggml-vulkan.so"


class LlamaCppBackend:
    """
    Manages a llama-server subprocess for GGUF model inference.

    Lifecycle:
        1. load_model()  — starts llama-server with the GGUF file
        2. generate_chat_completion() — proxies to /v1/chat/completions, streams back
        3. unload_model() — terminates llama-server subprocess
    """

    def __init__(self):
        self._process: Optional[subprocess.Popen] = None
        self._port: Optional[int] = None
        self._model_identifier: Optional[str] = None
        self._gguf_path: Optional[str] = None
        self._hf_repo: Optional[str] = None
        self._hf_variant: Optional[str] = None
        self._is_vision: bool = False
        self._healthy = False
        # Set by _classify_gpu_offload after _wait_for_health.
        self._gpu_offload_active: Optional[bool] = None
        self._context_length: Optional[int] = None
        self._effective_context_length: Optional[int] = None
        self._max_context_length: Optional[int] = None
        self._chat_template: Optional[str] = None
        self._chat_template_override: Optional[str] = None
        self._supports_reasoning: bool = False
        self._reasoning_always_on: bool = False
        self._reasoning_style: str = "enable_thinking"
        self._supports_preserve_thinking: bool = False
        self._supports_tools: bool = False
        self._cache_type_kv: Optional[str] = None
        self._reasoning_default: bool = True
        self._speculative_type: Optional[str] = None
        # Canonical UI-facing mode the user requested: one of
        # ``auto``/``mtp``/``ngram``/``mtp+ngram``/``off``/``ngram-simple``.
        # Round-tripped through the status API so the dropdown reflects
        # the picked mode rather than the resolved internal flag set
        # (auto on a 27B MTP GGUF resolves to draft-mtp but the dropdown
        # should still read "Auto").
        self._requested_spec_mode: Optional[str] = None
        # User-supplied --spec-draft-n-max override (None = platform default).
        self._spec_draft_n_max: Optional[int] = None
        # KV-cache estimation fields (populated by _read_gguf_metadata)
        self._n_layers: Optional[int] = None
        self._n_kv_heads: Optional[int] = None
        self._n_kv_heads_by_layer: Optional[list[int]] = None
        self._n_heads: Optional[int] = None
        self._embedding_length: Optional[int] = None
        # Architecture-aware KV fields for 5-path estimation
        self._kv_key_length: Optional[int] = None
        self._kv_value_length: Optional[int] = None
        self._sliding_window: Optional[int] = None
        self._sliding_window_pattern: Optional[list[bool]] = None
        self._full_attention_interval: Optional[int] = None
        self._kv_lora_rank: Optional[int] = None
        self._key_length_mla: Optional[int] = None
        self._kv_key_length_swa: Optional[int] = None
        self._kv_value_length_swa: Optional[int] = None
        self._ssm_inner_size: Optional[int] = None
        self._ssm_state_size: Optional[int] = None
        # Last N layers reuse KV from earlier layers and don't allocate
        # their own cache (Gemma 3n / Gemma 4: <arch>.attention.shared_kv_layers).
        self._shared_kv_layers: Optional[int] = None
        # MTP head count (llama.cpp #22673); >0 enables --spec-type draft-mtp.
        self._nextn_predict_layers: Optional[int] = None
        self._lock = threading.Lock()
        # Wraps load_model() end-to-end so concurrent loads serialise
        # and never coexist as two llama-server processes (#5401).
        self._serial_load_lock = threading.Lock()
        # Last extra_args / requested n_ctx, preserved across unload so
        # the chat UI's /unload+/load Apply path can inherit them (#5401).
        # ``_extra_args_source`` records the (model_identifier, hf_variant)
        # the stored args came from so the route can refuse cross-model
        # inheritance.
        self._extra_args: Optional[List[str]] = None
        self._extra_args_source: Optional[tuple[str, Optional[str]]] = None
        self._requested_n_ctx: int = 0
        self._stdout_lines: list[str] = []
        self._stdout_thread: Optional[threading.Thread] = None
        # llama-server tee log (see _drain_stdout / _kill_process).
        self._llama_log_fh = None
        self._llama_log_path: Optional[Path] = None
        self._cancel_event = threading.Event()
        self._api_key: Optional[str] = None
        # True once a probe has completed; cleared on transient failure.
        self._is_audio: bool = False
        self._audio_type: Optional[str] = None
        self._audio_probed: bool = False
        # Monotonic timestamp set in _kill_process; read by load_model
        # to decide whether to wait for the VRAM reclaim to finish.
        self._last_kill_monotonic: float = 0.0

        self._kill_orphaned_servers()
        atexit.register(self._cleanup)

    # ── Properties ────────────────────────────────────────────────

    @property
    def is_loaded(self) -> bool:
        return self._process is not None and self._healthy

    @property
    def is_active(self) -> bool:
        """True if a llama-server process exists (loading or loaded)."""
        return self._process is not None

    @property
    def base_url(self) -> str:
        return f"http://127.0.0.1:{self._port}"

    @property
    def model_identifier(self) -> Optional[str]:
        return self._model_identifier

    @property
    def is_vision(self) -> bool:
        return self._is_vision

    @property
    def hf_variant(self) -> Optional[str]:
        return self._hf_variant

    @property
    def extra_args(self) -> Optional[List[str]]:
        """Extra llama-server flags from the last load. Copy; None = never
        set, [] = explicitly cleared. Used by the route for inheritance."""
        return list(self._extra_args) if self._extra_args is not None else None

    @property
    def requested_n_ctx(self) -> int:
        """n_ctx the last load was invoked with (not the effective cap).
        0 means Auto. Used by the route to detect Auto-vs-explicit flips."""
        return self._requested_n_ctx

    @property
    def extra_args_source(self) -> Optional[tuple[str, Optional[str]]]:
        """(model_identifier, hf_variant) the stored extra_args came from.
        ``None`` if no extras have ever been recorded. Used by the route
        to refuse cross-model inheritance (#5401)."""
        return self._extra_args_source

    @property
    def context_length(self) -> Optional[int]:
        """Return the effective context length the server is running at."""
        return self._effective_context_length or self._context_length

    @property
    def max_context_length(self) -> Optional[int]:
        """Return the largest context that fits on this hardware at load time.

        This is the "safe zone" threshold the UI renders warnings
        against. For a model whose weights fit on some GPU subset, it
        is the binary-search cap from ``_fit_context_to_vram`` for that
        subset. For a model whose weights exceed 90% of every GPU
        subset, it is the 4096 fallback -- the spec's default when the
        model will not fit. The UI slider ceiling is
        ``native_context_length``; dragging above ``max_context_length``
        triggers the "might be slower" warning.
        """
        return self._max_context_length or self._context_length

    @property
    def native_context_length(self) -> Optional[int]:
        """Return the model's native context length from GGUF metadata."""
        return self._context_length

    def load_progress(self) -> Optional[dict]:
        """Return live model-load progress, or None if not loading.

        While llama-server is warming up, its process is typically in
        kernel state D (disk sleep) mmap'ing the weight shards into
        page cache before pushing layers to VRAM. During that window
        ``/api/inference/status`` only reports ``loading``, which gives
        the UI nothing to display besides a spinner that looks stuck
        for minutes on large MoE models.

        This method samples ``/proc/<pid>/status VmRSS`` against the
        sum of the GGUF shard sizes so the UI can render a real bar
        and compute rate / ETA. Returns ``None`` when no load is in
        flight (no process, or process already healthy).

        Shape::

            {
                "phase": "mmap" | "ready",
                "bytes_loaded": int,   # VmRSS of the llama-server
                "bytes_total":  int,   # sum of shard file sizes
                "fraction": float,     # bytes_loaded / bytes_total, 0..1
            }

        Linux-only in the current implementation. On macOS/Windows the
        equivalent would be a different API; this returns ``None`` on
        platforms where ``/proc/<pid>/status`` is unavailable.
        """
        proc = self._process
        if proc is None:
            return None
        pid = proc.pid
        if pid is None:
            return None

        # Sum up shard sizes (primary + any extras sitting alongside).
        bytes_total = 0
        gguf_path = self._gguf_path
        if gguf_path:
            primary = Path(gguf_path)
            try:
                if primary.is_file():
                    bytes_total += primary.stat().st_size
            except OSError:
                pass
            # Extra shards live alongside the primary with the same prefix
            # before the shard index (e.g. ``-00001-of-00004.gguf``).
            try:
                parent = primary.parent
                stem = primary.name
                m = _SHARD_RE.match(stem)
                prefix = m.group(1) if m else None
                if prefix and parent.is_dir():
                    for sibling in parent.iterdir():
                        if (
                            sibling.is_file()
                            and sibling.name.startswith(prefix)
                            and sibling.name != stem
                            and sibling.suffix == ".gguf"
                        ):
                            try:
                                bytes_total += sibling.stat().st_size
                            except OSError:
                                pass
            except OSError:
                pass

        # Read VmRSS from /proc/<pid>/status. Kilobytes on Linux.
        bytes_loaded = 0
        try:
            with open(f"/proc/{pid}/status", "r", encoding = "utf-8") as f:
                for line in f:
                    if line.startswith("VmRSS:"):
                        kb = int(line.split()[1])
                        bytes_loaded = kb * 1024
                        break
        except (FileNotFoundError, PermissionError, ValueError, OSError):
            return None

        phase = "ready" if self._healthy else "mmap"
        fraction = 0.0
        if bytes_total > 0:
            fraction = min(1.0, bytes_loaded / bytes_total)
        return {
            "phase": phase,
            "bytes_loaded": bytes_loaded,
            "bytes_total": bytes_total,
            "fraction": round(fraction, 4),
        }

    @property
    def chat_template(self) -> Optional[str]:
        return self._chat_template

    @property
    def chat_template_override(self) -> Optional[str]:
        return self._chat_template_override

    @property
    def supports_reasoning(self) -> bool:
        return self._supports_reasoning

    @property
    def reasoning_always_on(self) -> bool:
        return self._reasoning_always_on

    @property
    def reasoning_style(self) -> str:
        return self._reasoning_style

    @property
    def supports_preserve_thinking(self) -> bool:
        return self._supports_preserve_thinking

    @property
    def reasoning_default(self) -> bool:
        return self._reasoning_default

    def _reasoning_kwargs(self, enable_thinking: bool) -> dict:
        if self._reasoning_style == "reasoning_effort":
            return {"reasoning_effort": "high" if enable_thinking else "low"}
        return {"enable_thinking": enable_thinking}

    def _request_reasoning_kwargs(
        self,
        enable_thinking: Optional[bool],
        reasoning_effort: Optional[str] = None,
        preserve_thinking: Optional[bool] = None,
    ) -> Optional[dict]:
        """Build chat_template_kwargs from per-request reasoning fields.

        Produces a merged dict covering the active model's reasoning style
        (``enable_thinking`` or ``reasoning_effort``) plus the independent
        ``preserve_thinking`` kwarg when the template supports it.
        """
        kwargs: dict = {}
        # Always-on reasoning models hardcode <think> tags in their template
        # and do not consume enable_thinking / reasoning_effort -- skip.
        if self._supports_reasoning and not self._reasoning_always_on:
            if self._reasoning_style == "reasoning_effort":
                if reasoning_effort in ("low", "medium", "high"):
                    kwargs["reasoning_effort"] = reasoning_effort
                elif enable_thinking is not None:
                    kwargs["reasoning_effort"] = "high" if enable_thinking else "low"
            else:
                if enable_thinking is not None:
                    kwargs["enable_thinking"] = enable_thinking
        if self._supports_preserve_thinking and preserve_thinking is not None:
            kwargs["preserve_thinking"] = preserve_thinking
        return kwargs or None

    @property
    def supports_tools(self) -> bool:
        return self._supports_tools

    @property
    def cache_type_kv(self) -> Optional[str]:
        return self._cache_type_kv

    @property
    def speculative_type(self) -> Optional[str]:
        return self._speculative_type

    @property
    def requested_spec_mode(self) -> Optional[str]:
        """Canonical UI-facing mode the user requested (see field doc)."""
        return self._requested_spec_mode

    @property
    def spec_draft_n_max(self) -> Optional[int]:
        """User --spec-draft-n-max override active on the load, or None
        when the platform default (6 GPU / 3 CPU) is in effect."""
        return self._spec_draft_n_max

    # ── Binary discovery ──────────────────────────────────────────

    @staticmethod
    def _find_llama_server_binary() -> Optional[str]:
        """
        Locate the llama-server binary.

        Search order:
        1.  LLAMA_SERVER_PATH environment variable (direct path to binary)
        1b. UNSLOTH_LLAMA_CPP_PATH env var (custom llama.cpp install dir)
        2.  ~/.unsloth/llama.cpp/llama-server        (make build, root dir)
        3.  ~/.unsloth/llama.cpp/build/bin/llama-server  (cmake build, Linux)
        4.  ~/.unsloth/llama.cpp/build/bin/Release/llama-server.exe  (cmake build, Windows)
        5.  ./llama.cpp/llama-server                 (legacy: make build, root dir)
        6.  ./llama.cpp/build/bin/llama-server        (legacy: cmake in-tree build)
        7.  llama-server on PATH                     (system install)
        8.  ./bin/llama-server                       (legacy: extracted binary)
        """
        import os
        import sys

        binary_name = "llama-server.exe" if sys.platform == "win32" else "llama-server"

        # 1. Env var — direct path to binary
        env_path = os.environ.get("LLAMA_SERVER_PATH")
        if env_path and Path(env_path).is_file():
            return env_path

        # 1b. UNSLOTH_LLAMA_CPP_PATH — custom llama.cpp install directory
        custom_llama_cpp = os.environ.get("UNSLOTH_LLAMA_CPP_PATH")
        if custom_llama_cpp:
            custom_dir = Path(custom_llama_cpp)
            # Root dir (make builds)
            root_bin = custom_dir / binary_name
            if root_bin.is_file():
                return str(root_bin)
            # build/bin/ (cmake builds on Linux)
            cmake_bin = custom_dir / "build" / "bin" / binary_name
            if cmake_bin.is_file():
                return str(cmake_bin)
            # build/bin/Release/ (cmake builds on Windows)
            if sys.platform == "win32":
                win_bin = custom_dir / "build" / "bin" / "Release" / binary_name
                if win_bin.is_file():
                    return str(win_bin)

        # 2-4. Match installer layout: env-mode -> $STUDIO_HOME/llama.cpp;
        # default/HOME-redirect -> ~/.unsloth/llama.cpp (sibling of studio).
        legacy_llama = Path.home() / ".unsloth" / "llama.cpp"
        try:
            from utils.paths.storage_roots import studio_root as _sr  # noqa: WPS433

            _resolved_sr = _sr()
            _legacy_studio = Path.home() / ".unsloth" / "studio"
            try:
                _is_legacy = _resolved_sr.resolve() == _legacy_studio.resolve()
            except (OSError, ValueError):
                _is_legacy = _resolved_sr == _legacy_studio
            if _is_legacy:
                search_roots = [legacy_llama]
            else:
                # why: _kill_orphaned_servers excludes the legacy root in custom
                # mode; discovery must match so we never spawn a server we then
                # refuse to clean up. UNSLOTH_LLAMA_CPP_PATH (handled earlier)
                # is the explicit way to share a build across roots.
                search_roots = [_resolved_sr / "llama.cpp"]
        except (ImportError, OSError, ValueError):
            search_roots = [legacy_llama]
        _seen_roots: set[str] = set()
        _unique_roots: list[Path] = []
        for r in search_roots:
            k = str(r)
            if k not in _seen_roots:
                _seen_roots.add(k)
                _unique_roots.append(r)
        for unsloth_home in _unique_roots:
            home_root = unsloth_home / binary_name
            if home_root.is_file():
                return str(home_root)
            home_linux = unsloth_home / "build" / "bin" / binary_name
            if home_linux.is_file():
                return str(home_linux)
            if sys.platform == "win32":
                home_win = unsloth_home / "build" / "bin" / "Release" / binary_name
                if home_win.is_file():
                    return str(home_win)

        # 5–6. Legacy: in-tree build (older setup.sh / setup.ps1 versions)
        project_root = Path(__file__).resolve().parents[4]
        # Root dir (make builds)
        root_path = project_root / "llama.cpp" / binary_name
        if root_path.is_file():
            return str(root_path)
        # build/bin/ (cmake builds)
        build_path = project_root / "llama.cpp" / "build" / "bin" / binary_name
        if build_path.is_file():
            return str(build_path)
        if sys.platform == "win32":
            win_path = (
                project_root / "llama.cpp" / "build" / "bin" / "Release" / binary_name
            )
            if win_path.is_file():
                return str(win_path)

        # 7. System PATH
        system_path = shutil.which("llama-server")
        if system_path:
            return system_path

        # 8. Legacy: extracted to bin/
        bin_path = project_root / "bin" / binary_name
        if bin_path.is_file():
            return str(bin_path)

        return None

    # ── llama-server capability probe ─────────────────────────────

    # Cached on (path, mtime); `unsloth studio update` bumps mtime.
    _capability_cache: dict[tuple[str, int], dict[str, object]] = {}

    @classmethod
    def probe_server_capabilities(
        cls, binary: Optional[str] = None
    ) -> dict[str, object]:
        """Parse `llama-server --help` for feature flags. Returns
        {found, mtp_token, supports_mtp, ngram_mod_flavor,
        supports_ngram_mod, spec_draft_n_max_flag}.

        ``ngram_mod_flavor`` is ``"new"`` when the binary exposes the
        post-rename ``--spec-ngram-mod-n-match / -n-min / -n-max`` as
        real args, ``"legacy"`` when only the pre-rename
        ``--spec-ngram-size-n / --draft-min / --draft-max`` are real
        (the rename ships with stub removal entries for the legacy
        names; we tell stubs apart by the "argument has been removed"
        description), or ``None`` if neither set is usable.

        ``spec_draft_n_max_flag`` is the actual flag name the binary
        accepts: ``--spec-draft-n-max`` on post-rename builds, or
        ``--draft-max`` on legacy. ``None`` means n_max cannot be set.
        """
        bin_path = binary or cls._find_llama_server_binary()
        if not bin_path or not Path(bin_path).is_file():
            return {
                "found": False,
                "mtp_token": None,
                "supports_mtp": False,
                "ngram_mod_flavor": None,
                "supports_ngram_mod": False,
                "spec_draft_n_max_flag": None,
            }
        try:
            mtime = int(Path(bin_path).stat().st_mtime)
        except OSError:
            mtime = 0
        cache_key = (bin_path, mtime)
        cached = cls._capability_cache.get(cache_key)
        if cached is not None:
            return cached

        mtp_token: Optional[str] = None
        ngram_mod_flavor: Optional[str] = None
        spec_draft_n_max_flag: Optional[str] = None
        try:
            result = subprocess.run(
                [bin_path, "--help"],
                capture_output = True,
                text = True,
                timeout = 10,
                check = False,
            )
            help_text = (result.stdout or "") + "\n" + (result.stderr or "")
            # Split into per-flag blocks: each --flag line plus its
            # indented continuation lines, so the "argument has been
            # removed" description sits with its flag.
            blocks: dict[str, str] = {}
            current_flags: list[str] = []
            current_desc: list[str] = []
            for line in help_text.splitlines():
                stripped = line.strip()
                if stripped.startswith("-") and not line.startswith(" "):
                    # New flag line; flush previous.
                    if current_flags:
                        desc = " ".join(current_desc)
                        for f in current_flags:
                            blocks[f] = desc
                    current_flags = []
                    current_desc = [stripped]
                    # Extract long-form flag tokens from the DECLARATION
                    # prefix only (comma-separated aliases). Stop at the
                    # first token that isn't itself a flag, so flag
                    # references inside descriptions are ignored.
                    for tok in re.split(r"[,\s]+", stripped):
                        if tok.startswith("--") and re.match(
                            r"--[A-Za-z][A-Za-z0-9_-]*$", tok
                        ):
                            current_flags.append(tok)
                        elif tok.startswith("-") and len(tok) > 1:
                            # short alias like -fa; keep scanning aliases.
                            continue
                        else:
                            # First non-flag token marks end of decl.
                            break
                else:
                    current_desc.append(stripped)
            if current_flags:
                desc = " ".join(current_desc)
                for f in current_flags:
                    blocks[f] = desc

            def _is_real(flag: str) -> bool:
                """True if the flag exists AND is not a removal stub."""
                desc = blocks.get(flag)
                if desc is None:
                    return False
                return "argument has been removed" not in desc

            # MTP token detection from --spec-type line.
            spec_line = ""
            for line in help_text.splitlines():
                if "--spec-type" in line:
                    spec_line = line
                    break
            # PR #22673 used draft-mtp; later renamed to mtp.
            if "draft-mtp" in spec_line:
                mtp_token = "draft-mtp"
            elif re.search(r"[|,\[]mtp[|,\]]", spec_line):
                mtp_token = "mtp"

            # ngram-mod flag flavor. Post-rename builds advertise both
            # the new args (real) and the legacy ones (stubs); pre-rename
            # builds only have the legacy ones as real.
            new_ngram_real = (
                _is_real("--spec-ngram-mod-n-match")
                and _is_real("--spec-ngram-mod-n-min")
                and _is_real("--spec-ngram-mod-n-max")
            )
            legacy_ngram_real = (
                _is_real("--spec-ngram-size-n")
                and _is_real("--draft-max")
                and _is_real("--draft-min")
            )
            if new_ngram_real:
                ngram_mod_flavor = "new"
            elif legacy_ngram_real:
                ngram_mod_flavor = "legacy"

            # n_max flag: prefer post-rename, fall back to legacy.
            if _is_real("--spec-draft-n-max"):
                spec_draft_n_max_flag = "--spec-draft-n-max"
            elif _is_real("--draft-max"):
                spec_draft_n_max_flag = "--draft-max"
        except (OSError, subprocess.SubprocessError) as exc:
            logger.debug(f"llama-server --help probe failed: {exc}")

        info = {
            "found": True,
            "mtp_token": mtp_token,
            "supports_mtp": mtp_token is not None,
            "ngram_mod_flavor": ngram_mod_flavor,
            "supports_ngram_mod": ngram_mod_flavor is not None,
            "spec_draft_n_max_flag": spec_draft_n_max_flag,
        }
        cls._capability_cache[cache_key] = info
        return info

    # ── GPU allocation ────────────────────────────────────────────

    @staticmethod
    def _get_gguf_size_bytes(model_path: str) -> int:
        """Get total GGUF size in bytes, including split shards."""
        main = Path(model_path)
        total = main.stat().st_size

        # Check for split shards (e.g., model-00001-of-00003.gguf)
        m = _SHARD_FULL_RE.match(main.name)
        if m:
            prefix, _, num_total = m.group(1), m.group(2), m.group(3)
            sibling_pat = re.compile(
                r"^"
                + re.escape(prefix)
                + r"-\d{5}-of-"
                + re.escape(num_total)
                + r"\.gguf$"
            )
            for sibling in main.parent.iterdir():
                if sibling != main and sibling_pat.match(sibling.name):
                    total += sibling.stat().st_size

        return total

    @staticmethod
    def _is_vulkan_backend(binary: Optional[str] = None) -> bool:
        """True if the installed llama.cpp build is the Vulkan one.

        Builds are single-backend, so the presence of the Vulkan ggml
        backend library next to llama-server is sufficient. Used to keep
        the free-memory probe and the GPU pin in the same device-index
        space (ggml's Vulkan ordinals, not nvidia-smi order).
        """
        binary = binary or LlamaCppBackend._find_llama_server_binary()
        if not binary:
            return False
        return (Path(binary).parent / _vulkan_lib_filename()).is_file()

    @staticmethod
    def _get_gpu_free_memory(binary: Optional[str] = None) -> list[tuple[int, int]]:
        """Query free memory per GPU across all supported backends.

        On a Vulkan build, the ggml Vulkan probe is authoritative so the
        returned indices are Vulkan ordinals (the space the GPU pin writes
        to ``GGML_VK_VISIBLE_DEVICES``). Otherwise ``nvidia-smi`` / torch
        cover NVIDIA + AMD ROCm, with the Vulkan probe as a last resort.

        Returns list of (gpu_index, free_mib) sorted by index. Empty
        list if no supported GPU is reachable.
        """
        binary = binary or LlamaCppBackend._find_llama_server_binary()
        if LlamaCppBackend._is_vulkan_backend(binary):
            return LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
        gpus = LlamaCppBackend._get_gpu_free_memory_nvidia_torch()
        if gpus:
            return gpus
        return LlamaCppBackend._get_gpu_free_memory_vulkan(binary)

    @staticmethod
    def _get_gpu_free_memory_nvidia_torch() -> list[tuple[int, int]]:
        """Query free memory per GPU.

        Order:
          1. ``nvidia-smi`` (NVIDIA CUDA hosts) -- respects
             ``CUDA_VISIBLE_DEVICES``.
          2. ``torch.cuda.mem_get_info`` -- universal fallback that
             works on AMD ROCm too because the HIP runtime
             reuses the entire ``torch.cuda.*`` namespace. Covers the
             AMD case for issue #5106 (nvidia-smi-only probe silently
             returned [] on AMD hosts) and also rescues NVIDIA hosts
             where ``nvidia-smi`` is missing from PATH.

        Returns list of (gpu_index, free_mib) sorted by index. Empty
        list if no supported GPU is reachable.
        """
        import os

        # ── NVIDIA via nvidia-smi ────────────────────────────────────
        try:
            result = subprocess.run(
                [
                    "nvidia-smi",
                    "--query-gpu=index,memory.free",
                    "--format=csv,noheader,nounits",
                ],
                capture_output = True,
                text = True,
                timeout = 10,
                env = child_env_without_native_path_secret(),
                **_windows_hidden_subprocess_kwargs(),
            )
            if result.returncode == 0:
                allowed: Optional[set[int]] = None
                cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
                if cvd is not None:
                    try:
                        # `if x.strip()` filters trailing-comma masks like
                        # "0,1," which would otherwise raise ValueError on
                        # an empty token. An explicitly empty mask (CVD="")
                        # yields an empty `allowed` set so all GPUs are
                        # filtered out, matching the codebase convention.
                        allowed = set(
                            int(x.strip()) for x in cvd.split(",") if x.strip()
                        )
                    except ValueError:
                        pass
                gpus: list[tuple[int, int]] = []
                for line in result.stdout.strip().splitlines():
                    parts = line.split(",")
                    if len(parts) == 2:
                        idx = int(parts[0].strip())
                        free_mib = int(parts[1].strip())
                        if allowed is not None and idx not in allowed:
                            continue
                        gpus.append((idx, free_mib))
                # Match the docstring's sort-by-id guarantee. nvidia-smi
                # almost always returns sorted output, but driver order
                # is not formally guaranteed.
                gpus.sort(key = lambda g: g[0])
                if gpus:
                    return gpus
        except Exception as e:
            logger.debug(f"nvidia-smi probe failed: {e}")

        # ── Torch fallback (covers AMD ROCm and missing nvidia-smi) ──
        try:
            import torch

            if not hasattr(torch, "cuda") or not torch.cuda.is_available():
                return []
            if not hasattr(torch.cuda, "mem_get_info"):
                return []
            # torch.cuda enumerates GPUs RELATIVE to the visibility mask.
            # On NVIDIA builds the mask is CUDA_VISIBLE_DEVICES; on AMD
            # ROCm builds it is HIP_VISIBLE_DEVICES (or ROCR_VISIBLE_DEVICES
            # if HIP is unset). Downstream we feed these IDs back into the
            # llama-server subprocess as CVD, so we must translate visible
            # ordinals back to physical indices first; otherwise launching
            # with ``CUDA_VISIBLE_DEVICES=2,3`` would get rewritten to
            # ``CUDA_VISIBLE_DEVICES=0,1`` and target the wrong GPUs.
            physical_ids: Optional[list[int]] = None
            # Match the codebase convention in
            # ``utils/hardware/hardware.py::_get_parent_visible_gpu_spec``:
            # treat an explicitly empty mask (``HIP_VISIBLE_DEVICES=""``)
            # as "set to no GPUs" rather than falling through to the next
            # var. ``or`` would coerce empty string to falsy and silently
            # promote the wrong source.
            if getattr(torch.version, "hip", None) is not None:
                hip_v = os.environ.get("HIP_VISIBLE_DEVICES")
                rocr_v = os.environ.get("ROCR_VISIBLE_DEVICES")
                cvd = (
                    hip_v
                    if hip_v is not None
                    else rocr_v
                    if rocr_v is not None
                    else os.environ.get("CUDA_VISIBLE_DEVICES")
                )
            else:
                cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
            if cvd is not None:
                try:
                    # Empty mask (CVD="") yields an empty list so the
                    # below loop produces no GPUs, consistent with the
                    # nvidia-smi path and utils/hardware/hardware.py.
                    physical_ids = [int(x.strip()) for x in cvd.split(",") if x.strip()]
                except ValueError:
                    physical_ids = None
            gpus = []
            for ordinal in range(torch.cuda.device_count()):
                free_bytes, _total_bytes = torch.cuda.mem_get_info(ordinal)
                idx = (
                    physical_ids[ordinal]
                    if physical_ids is not None and ordinal < len(physical_ids)
                    else ordinal
                )
                gpus.append((idx, free_bytes // (1024 * 1024)))
            # Match the nvidia-smi path's docstring guarantee of sorted-by-id.
            return sorted(gpus, key = lambda g: g[0])
        except Exception as e:
            logger.debug(f"torch GPU probe failed: {e}")
            return []

    @staticmethod
    def _get_gpu_free_memory_vulkan(
        binary: Optional[str] = None,
    ) -> list[tuple[int, int]]:
        """Query free VRAM per device via the bundled ggml Vulkan backend.

        Loads ``libggml-vulkan`` in a short-lived subprocess and calls
        ``ggml_backend_vk_get_device_memory`` for each device, so no Vulkan
        instance is created in this process. Returns list of
        (device_index, free_mib) sorted by index, where the index is ggml's
        own Vulkan device ordinal (the space ``GGML_VK_VISIBLE_DEVICES``
        expects). Returns [] when no Vulkan build is installed or no device
        is reachable.
        """
        binary = binary or LlamaCppBackend._find_llama_server_binary()
        if not binary:
            return []
        binary_dir = Path(binary).parent
        if not (binary_dir / _vulkan_lib_filename()).is_file():
            return []

        env = child_env_without_native_path_secret()
        if sys.platform != "win32":
            # Let the loader resolve sibling ggml libs next to the binary.
            existing_ld = env.get("LD_LIBRARY_PATH", "")
            env["LD_LIBRARY_PATH"] = (
                f"{binary_dir}:{existing_ld}" if existing_ld else str(binary_dir)
            )
        try:
            result = subprocess.run(
                [sys.executable, "-c", _VULKAN_PROBE_SCRIPT, str(binary_dir)],
                capture_output = True,
                text = True,
                timeout = 15,
                env = env,
                **_windows_hidden_subprocess_kwargs(),
            )
        except Exception as e:
            logger.debug(f"vulkan GPU probe failed: {e}")
            return []

        gpus: list[tuple[int, int]] = []
        for line in result.stdout.strip().splitlines():
            parts = line.split("\t")
            if len(parts) != 3:
                continue
            try:
                idx = int(parts[0])
                free_mib = int(parts[1]) // (1024 * 1024)
            except ValueError:
                continue
            gpus.append((idx, free_mib))
        gpus.sort(key = lambda g: g[0])
        if gpus:
            logger.info(
                "Vulkan GPU memory detected: "
                + ", ".join(f"VK{idx}={free}MiB" for idx, free in gpus)
            )
        return gpus

    # Skip the wait when the last kill is older than this; the GPU
    # driver has already reclaimed the prior process's allocations.
    _VRAM_SETTLE_WINDOW_S: float = 15.0

    @staticmethod
    def _wait_for_vram_settle(
        max_wait: float = 2.0,
        interval: float = 0.25,
        tolerance_mib: int = 256,
        since_kill: float = 0.0,
    ) -> None:
        """Poll ``_get_gpu_free_memory`` until free VRAM stabilises.

        The GPU driver reclaims a dead process's allocations
        asynchronously, so sampling free memory in the kill-to-spawn
        window reads artificially low and pushes ``_select_gpus`` /
        ``_fit_context_to_vram`` toward needless CPU offload -- on a
        tight VRAM card this is the Apply-reload OOM that bare-shell
        launches with the same flags never see.

        Short-circuits on cold start (``since_kill`` zero) or stale
        kill (older than ``_VRAM_SETTLE_WINDOW_S``); also on CPU-only
        hosts (empty probe), probe exceptions, and GPU-set changes.
        ``max_wait`` is a wall-clock bound that includes probe time,
        so a wedged ``nvidia-smi`` cannot extend the reload.
        """
        now = time.monotonic()
        if since_kill <= 0.0:
            return
        if now - since_kill > LlamaCppBackend._VRAM_SETTLE_WINDOW_S:
            return
        deadline = now + max_wait

        def _probe_or_none():
            if time.monotonic() >= deadline:
                return None
            try:
                return LlamaCppBackend._get_gpu_free_memory()
            except Exception:
                return None

        prev = _probe_or_none()
        if prev is None or not prev:
            return
        while time.monotonic() < deadline:
            remaining = deadline - time.monotonic()
            if remaining <= 0:
                return
            # Clip the nap so a near-zero ``max_wait`` is respected.
            time.sleep(min(interval, remaining))
            curr = _probe_or_none()
            if curr is None or not curr or len(curr) != len(prev):
                return
            prev_map = dict(prev)
            stable = True
            for idx, free in curr:
                if idx not in prev_map:
                    stable = False
                    break
                prev_free = prev_map[idx]
                # Adaptive: 2 % of the larger sample dominates the
                # 256 MiB floor on large-VRAM cards.
                per_gpu_tol = max(tolerance_mib, int(max(free, prev_free) * 0.02))
                if abs(free - prev_free) >= per_gpu_tol:
                    stable = False
                    break
            if stable:
                return
            prev = curr

    # Free-VRAM fraction at which Studio pins the GPU directly instead
    # of deferring to ``--fit on``. 5% headroom covers CUDA context +
    # compute buffers; 0.90 was too conservative and dropped 91-94%
    # fits to CPU offload (#5106). The fork's --fit on still catches
    # the truly-too-large case.
    _GPU_PIN_VRAM_FRACTION = 0.95

    @staticmethod
    def _windows_pip_nvidia_dll_dirs(prefix: str) -> list[str]:
        """Return DLL dirs from pip-installed CUDA wheels under
        ``<prefix>/Lib/site-packages/`` so llama-server.exe can load
        ``cudart64_X.dll`` / ``cublas64_X.dll`` without a system CUDA
        toolkit. Mirrors the Linux ``nvidia/cu*/lib`` LD_LIBRARY_PATH
        block, with parity for the Windows-specific wheel layouts seen
        in the wild. Covered patterns:
          * ``nvidia/<pkg>/bin`` -- legacy modular wheels
            (``nvidia-cuda-runtime-cu12``, ``nvidia-cublas-cu12``, etc.).
          * ``nvidia/<pkg>/bin/x86_64`` and ``.../bin/x64`` -- current
            CUDA 13 wheel layout used by the unsuffixed
            ``nvidia-cuda-runtime`` / ``nvidia-cublas`` packages, which
            ship under ``nvidia/cu13/bin/x86_64/`` (#5106).
          * ``nvidia/<pkg>/Library/bin`` (and arch subdirs) -- conda-
            style wheel repacks.
          * ``torch/lib`` -- PyTorch's own CUDA-bundled Windows wheel,
            which can ship ``cudart64_*.dll`` directly here instead of
            as separate ``nvidia-*`` wheels. The install-side helper
            ``python_runtime_dirs`` in ``install_llama_prebuilt.py``
            covers this path for the same reason.

        Walks the tree with ``Path.iterdir`` rather than ``glob.glob``
        so the resolver is safe against Windows paths containing
        ``[`` or ``]`` (valid in usernames; would otherwise be
        interpreted as a glob character class and silently miss
        existing dirs)."""
        site_packages = Path(prefix) / "Lib" / "site-packages"
        out: list[str] = []
        seen: set[str] = set()

        def _add(path: Path) -> None:
            if not path.is_dir():
                return
            key = os.path.normcase(os.path.abspath(str(path)))
            if key in seen:
                return
            seen.add(key)
            out.append(str(path))

        nvidia_root = site_packages / "nvidia"
        if nvidia_root.is_dir():
            for pkg_dir in nvidia_root.iterdir():
                if not pkg_dir.is_dir():
                    continue
                # Order matters for PATH search: arch-specific subdirs
                # first so the explicit cudart64_X.dll location wins
                # over a sibling ``bin`` that might be empty.
                for sub in (
                    pkg_dir / "bin" / "x86_64",
                    pkg_dir / "bin" / "x64",
                    pkg_dir / "bin",
                    pkg_dir / "Library" / "bin" / "x86_64",
                    pkg_dir / "Library" / "bin" / "x64",
                    pkg_dir / "Library" / "bin",
                ):
                    _add(sub)
        _add(site_packages / "torch" / "lib")
        return out

    @staticmethod
    def _build_windows_path_dirs(
        binary_dir: str, prefix: str, cuda_path: str
    ) -> list[str]:
        """Ordered PATH entries the win32 branch of start_llama_server
        prepends so llama-server.exe resolves cudart / cublas DLLs:
        binary_dir, pip nvidia wheels, CUDA_PATH/bin, CUDA_PATH/bin/x64.
        Extracted so test_windows_gpu_detection_mock asserts against
        production logic, not a hand-copy. #5106."""
        path_dirs = [binary_dir]
        path_dirs.extend(LlamaCppBackend._windows_pip_nvidia_dll_dirs(prefix))
        if cuda_path:
            cuda_bin = os.path.join(cuda_path, "bin")
            if os.path.isdir(cuda_bin):
                path_dirs.append(cuda_bin)
            cuda_bin_x64 = os.path.join(cuda_path, "bin", "x64")
            if os.path.isdir(cuda_bin_x64):
                path_dirs.append(cuda_bin_x64)
        return path_dirs

    @staticmethod
    def _select_gpus(
        model_size_bytes: int,
        gpus: list[tuple[int, int]],
    ) -> tuple[Optional[list[int]], bool]:
        """Pick GPU(s) for a model based on estimated VRAM and free memory.

        ``model_size_bytes`` should include both model weights and estimated
        KV cache.  The ``_GPU_PIN_VRAM_FRACTION`` threshold provides headroom
        for compute buffers, CUDA context, and other runtime overhead.

        Returns (gpu_indices, use_fit):
          - ([1], False)       model fits on 1 GPU at the headroom threshold
          - ([1, 2], False)    model needs 2 GPUs
          - (None, True)       model too large, let --fit handle it
        """
        if not gpus:
            return None, True

        model_size_mib = model_size_bytes / (1024 * 1024)
        usable_fraction = LlamaCppBackend._GPU_PIN_VRAM_FRACTION

        # Sort GPUs by free memory descending
        ranked = sorted(gpus, key = lambda g: g[1], reverse = True)

        # Try fitting on 1 GPU at the usable-VRAM threshold.
        if ranked[0][1] * usable_fraction >= model_size_mib:
            return [ranked[0][0]], False

        # Try fitting on N GPUs (accumulate free memory from most-free)
        cumulative = 0
        selected = []
        for idx, free_mib in ranked:
            selected.append(idx)
            cumulative += free_mib * usable_fraction
            if cumulative >= model_size_mib:
                return sorted(selected), False

        # Model is too large even for all GPUs, let --fit handle it
        logger.debug(
            "Model does not fit in available GPU memory, falling back to --fit",
            model_size_mib = round(model_size_mib, 2),
            ranked_gpus = ranked,
        )
        return None, True

    # ── KV cache VRAM estimation ─────────────────────────────────────

    def _can_estimate_kv(self) -> bool:
        """True if we have enough GGUF metadata to estimate KV cache size."""
        if self._n_layers is None:
            return False
        # MLA: kv_lora_rank is sufficient (K-only cache)
        if self._kv_lora_rank is not None:
            return True
        # New-style: need both explicit key AND value dimensions
        if self._kv_key_length is not None and self._kv_value_length is not None:
            return True
        # Legacy: need embedding_length + a head count (scalar or per-layer).
        return self._embedding_length is not None and (
            self._n_kv_heads is not None
            or self._n_heads is not None
            or self._n_kv_heads_by_layer is not None
        )

    def _kv_heads_for_layer(self, layer_idx: int, fallback: int) -> int:
        if self._n_kv_heads_by_layer is not None and layer_idx < len(
            self._n_kv_heads_by_layer
        ):
            return self._n_kv_heads_by_layer[layer_idx]
        return fallback

    def _estimate_kv_cache_bytes(
        self,
        n_ctx: int,
        cache_type_kv: Optional[str] = None,
        *,
        swa_full: bool = False,
        n_parallel: int = 1,
        kv_unified: bool = True,
        ctx_checkpoints: int = 0,
    ) -> int:
        """Estimate KV cache VRAM for a given context length.

        Uses 5-path architecture-aware estimation:
          1. MLA      -- compressed KV latent + RoPE, K-only (no separate V)
          2. Hybrid   -- only attention layers need KV (Mamba layers don't)
          3. SWA      -- sliding-window layers cache min(ctx, window) tokens
          4. GQA      -- standard full KV with explicit key/value dimensions
          5. Legacy   -- fallback using embed // n_heads

        Server-flag knobs (mirror llama-server's CLI):
          swa_full        -- ``--swa-full``: force SWA layers to cache the
                             full ``n_ctx`` (collapses path 3 to path 4
                             sizing for the SWA layers).
          n_parallel      -- ``--parallel``: number of server slots.
                             Verified empirically against llama-server:
                             non-SWA layers stay constant (cells split
                             across slots), SWA layers scale linearly
                             (per-slot window).
          kv_unified      -- ``--kv-unified`` (default on): retained for
                             API forward-compat. Currently a no-op for
                             memory math because the unified buffer total
                             matches per-slot buffers in measured cases.
          ctx_checkpoints -- ``--ctx-checkpoints``: SWA snapshot count per
                             slot (PR #15293). Each snapshot stores one
                             sliding-window of state per SWA layer.

        Returns 0 if metadata is insufficient for estimation.
        """
        if not self._can_estimate_kv() or n_ctx <= 0:
            return 0

        n_layers = self._n_layers  # type: ignore[assignment]
        # Gemma 3n / Gemma 4 reuse KV from earlier layers in the last
        # ``shared_kv_layers`` blocks -- those don't allocate their own
        # cache.  Floor at 1 so a misconfigured GGUF can't zero out KV.
        shared = self._shared_kv_layers or 0
        n_layers_kv = max(1, n_layers - shared)
        n_kv = self._n_kv_heads or self._n_heads or 1  # type: ignore[assignment]

        # Bytes per element depends on KV cache quantization
        bpe = {
            "f32": 4.0,
            "f16": 2.0,
            "bf16": 2.0,
            "q8_0": 34 / 32,
            "q5_1": 0.75,
            "q5_0": 0.6875,
            "q4_1": 0.625,
            "q4_0": 0.5625,
            "iq4_nl": 0.5625,
        }.get(cache_type_kv or "f16", 2.0)

        slots = max(1, n_parallel)

        # Path 1: MLA (DeepSeek-V2/V3, GLM-4.7, GLM-5, Kimi-K2.5)
        # MLA stores one compressed KV latent per token/layer (shared across heads).
        # V is reconstructed from the latent on the fly -- no separate V cache.
        # key_length = kv_lora_rank + rope_dim (the full compressed representation).
        # MLA GGUFs set head_count_kv=1; default to 1 if absent to avoid
        # falling back to n_heads (e.g., 128 for DeepSeek-V3) which would 128x.
        if self._kv_lora_rank is not None:
            n_kv_mla = self._n_kv_heads or 1
            rope_dim = self._key_length_mla or 64
            key_len = self._kv_key_length or (self._kv_lora_rank + rope_dim)
            return int(n_layers_kv * n_ctx * n_kv_mla * key_len * bpe)

        key_len = self._kv_key_length
        val_len = self._kv_value_length

        # Path 2: Hybrid Mamba/Attention (Qwen3.5-27B, Qwen3.5-35B-A3B)
        # Only 1 in N layers is attention; the rest are Mamba (no KV cache).
        if (
            self._ssm_inner_size is not None
            and self._full_attention_interval is not None
        ):
            fai = self._full_attention_interval
            n_attn = -(-n_layers // fai) if fai > 0 else n_layers  # ceiling division
            if key_len is not None and val_len is not None:
                return int(n_attn * n_ctx * n_kv * (key_len + val_len) * bpe)
            head_dim = self._embedding_length // self._n_heads if self._n_heads else 128  # type: ignore[operator]
            return int(n_attn * n_ctx * n_kv * 2 * head_dim * bpe)

        # Path 3: Sliding window (Gemma 2/3/3n/4, gpt-oss, Cohere2 ...).
        # Pattern is filled in by the resolver at parse time; if absent,
        # falls through to the legacy 1/4-global heuristic below.
        # Per-layer-type ``--parallel N`` accounting (verified empirically
        # against ``llama-server``):
        #   * non-SWA layers:    total cells = n_ctx, partitioned across
        #                         slots -> total memory CONSTANT in slots.
        #   * SWA layers:         per-slot cells = 2 * sliding_window
        #                         (capped at n_ctx and at per_slot_ctx
        #                         when ctx is split among many slots) ->
        #                         total memory grows LINEARLY in slots.
        # ``--swa-full`` forces full n_ctx for SWA layers instead.
        # ``--ctx-checkpoints N`` adds N snapshots per SWA layer per slot.
        if (
            self._sliding_window is not None
            and self._sliding_window > 0
            and key_len is not None
            and val_len is not None
        ):
            swa = self._sliding_window
            per_slot_ctx = max(1, n_ctx // slots)
            # ``--swa-full`` makes SWA layers cache the full context just
            # like non-SWA: cells get partitioned across slots, so per-slot
            # cells = per_slot_ctx and the slots*per-slot product collapses
            # back to the constant ``n_ctx`` total.  Otherwise SWA caches
            # 2*sliding_window per slot, clamped at the per-slot ctx.
            swa_cells_per_slot = (
                per_slot_ctx if swa_full else min(n_ctx, 2 * swa, per_slot_ctx)
            )
            key_len_swa = self._kv_key_length_swa or key_len
            val_len_swa = self._kv_value_length_swa or val_len
            if self._sliding_window_pattern is not None:
                global_bytes = 0.0  # constant across slots
                swa_bytes_per_slot = 0.0  # multiplied by slots
                checkpoint_extra_per_slot = 0.0
                # Iterate only over layers that allocate their own KV;
                # the trailing ``shared`` layers reuse earlier caches.
                for layer_idx in range(n_layers_kv):
                    layer_n_kv = self._kv_heads_for_layer(layer_idx, n_kv)
                    is_swa = (
                        layer_idx < len(self._sliding_window_pattern)
                        and self._sliding_window_pattern[layer_idx]
                    )
                    if is_swa:
                        swa_bytes_per_slot += (
                            swa_cells_per_slot
                            * layer_n_kv
                            * (key_len_swa + val_len_swa)
                            * bpe
                        )
                        if ctx_checkpoints > 0 and not swa_full:
                            checkpoint_extra_per_slot += (
                                ctx_checkpoints
                                * swa
                                * layer_n_kv
                                * (key_len_swa + val_len_swa)
                                * bpe
                            )
                    else:
                        global_bytes += n_ctx * layer_n_kv * (key_len + val_len) * bpe
                return int(
                    global_bytes
                    + slots * (swa_bytes_per_slot + checkpoint_extra_per_slot)
                )
            n_global = max(1, n_layers_kv // 4)
            n_swa = n_layers_kv - n_global
            kv_per_token = n_kv * (key_len + val_len) * bpe
            kv_per_token_swa = n_kv * (key_len_swa + val_len_swa) * bpe
            global_bytes = n_global * n_ctx * kv_per_token
            swa_bytes_per_slot = n_swa * swa_cells_per_slot * kv_per_token_swa
            checkpoint_extra_per_slot = (
                ctx_checkpoints * n_swa * swa * kv_per_token_swa
                if ctx_checkpoints > 0 and not swa_full
                else 0.0
            )
            return int(
                global_bytes + slots * (swa_bytes_per_slot + checkpoint_extra_per_slot)
            )

        # Path 4: Standard GQA with explicit key/value dimensions
        if key_len is not None and val_len is not None:
            return int(n_layers_kv * n_ctx * n_kv * (key_len + val_len) * bpe)

        # Path 5: Legacy fallback (old GGUFs without explicit dimensions)
        head_dim = self._embedding_length // self._n_heads if self._n_heads else 128  # type: ignore[operator]
        return int(2 * n_kv * head_dim * n_layers_kv * n_ctx * bpe)

    def _fit_context_to_vram(
        self,
        requested_ctx: int,
        available_mib: int,
        model_size_bytes: int,
        cache_type_kv: Optional[str] = None,
        min_ctx: int = 4096,
        *,
        swa_full: bool = False,
        n_parallel: int = 1,
        kv_unified: bool = True,
        ctx_checkpoints: int = 0,
        kv_on_gpu: bool = True,
        mtp_engaged: bool = False,
    ) -> int:
        """Return the largest context length that fits in GPU VRAM.

        Uses 90% of available VRAM as the ctx-fit budget. Tighter than
        ``_GPU_PIN_VRAM_FRACTION`` on purpose: over-promising context
        OOMs at runtime, while pinning conservatively just defers to
        --fit on. If the weights alone don't fit, returns
        ``requested_ctx`` unchanged.

        ``kv_on_gpu`` mirrors ``--kv-offload`` (default on). When False
        the KV cache lives in CPU RAM and doesn't compete with weights
        for VRAM; the requested context is honored verbatim. The other
        keyword args mirror ``_estimate_kv_cache_bytes``.

        ``mtp_engaged`` reserves extra VRAM for the MTP draft model's
        KV cache + compute graph buffers. llama.cpp's MTP path keeps a
        secondary cache sized off the target's KV; on tight VRAM tiers
        (e.g. 32 GB) auto-fit at native context would otherwise spill
        and force llama-server into a slower partial-offload path.
        """
        if not self._can_estimate_kv():
            logger.debug(
                "Skipping context fit because KV cache metadata is unavailable",
                requested_ctx = requested_ctx,
                available_mib = available_mib,
            )
            return requested_ctx

        # KV lives off-GPU: no VRAM accounting needed for the cache itself.
        if not kv_on_gpu:
            return requested_ctx

        kv_kwargs = dict(
            swa_full = swa_full,
            n_parallel = n_parallel,
            kv_unified = kv_unified,
            ctx_checkpoints = ctx_checkpoints,
        )

        # MTP needs a tighter budget; drop from 0.90 to 0.85.
        budget_frac = 0.85 if mtp_engaged else 0.90
        budget_bytes = available_mib * 1024 * 1024 * budget_frac
        model_footprint = model_size_bytes

        # Check if requested context already fits
        kv = self._estimate_kv_cache_bytes(requested_ctx, cache_type_kv, **kv_kwargs)
        if model_footprint + kv <= budget_bytes:
            return requested_ctx

        # Model weights alone exceed budget -- can't help by reducing ctx.
        # Return requested_ctx unchanged; --fit will handle VRAM management.
        if model_footprint >= budget_bytes:
            logger.debug(
                "Model footprint exceeds GPU budget before KV cache",
                requested_ctx = requested_ctx,
                available_mib = available_mib,
                model_size_gb = round(model_footprint / (1024**3), 2),
            )
            return requested_ctx

        # Binary search for max context that fits
        remaining = budget_bytes - model_footprint
        effective_min = min(min_ctx, requested_ctx)
        lo, hi = effective_min, requested_ctx
        best = effective_min
        while lo <= hi:
            mid = (lo + hi) // 2
            kv = self._estimate_kv_cache_bytes(mid, cache_type_kv, **kv_kwargs)
            if kv <= remaining:
                best = mid
                lo = mid + 1
            else:
                hi = mid - 1

        # Round down to nearest 256 for alignment, but never exceed requested_ctx
        best = (best // 256) * 256
        best = max(effective_min, best)
        best = min(best, requested_ctx)
        return best

    # ── Variant fallback ────────────────────────────────────────────

    @staticmethod
    def _find_smallest_fitting_variant(
        hf_repo: str,
        free_bytes: int,
        hf_token: Optional[str] = None,
    ) -> Optional[tuple[str, int]]:
        """Find the smallest GGUF variant (including all shards) that fits.

        Groups split shards by variant prefix and sums their sizes.
        For example, UD-Q4_K_XL with 9 shards of 50 GB each = 450 GB total.

        Returns (first_shard_filename, total_size_bytes) or None if nothing fits.
        """
        try:
            from huggingface_hub import get_paths_info, list_repo_files

            files = list_repo_files(hf_repo, token = hf_token)
            gguf_files = [
                f for f in files if f.endswith(".gguf") and "mmproj" not in f.lower()
            ]
            if not gguf_files:
                return None

            # Get sizes for all GGUF files
            path_infos = list(get_paths_info(hf_repo, gguf_files, token = hf_token))
            size_map = {p.path: (p.size or 0) for p in path_infos}

            # Group files by variant: shards share a prefix before -NNNNN-of-NNNNN
            variants: dict[str, list[str]] = {}
            for f in gguf_files:
                m = _SHARD_RE.match(f)
                key = m.group(1) if m else f
                variants.setdefault(key, []).append(f)

            # Sum shard sizes per variant, track the first shard (for download)
            variant_sizes: list[tuple[str, int, list[str]]] = []
            for key, shard_files in variants.items():
                total = sum(size_map.get(f, 0) for f in shard_files)
                first = sorted(shard_files)[0]
                variant_sizes.append((first, total, shard_files))

            # Sort by total size ascending and pick the smallest that fits
            variant_sizes.sort(key = lambda x: x[1])
            for first_file, total_size, _ in variant_sizes:
                if total_size > 0 and total_size <= free_bytes:
                    return first_file, total_size

            return None
        except Exception:
            return None

    # ── Port allocation ───────────────────────────────────────────

    @staticmethod
    def _find_free_port() -> int:
        """Find an available TCP port."""
        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
            s.bind(("127.0.0.1", 0))
            return s.getsockname()[1]

    # ── Stdout drain (prevents pipe deadlock on Windows) ─────────

    def _drain_stdout(self):
        """
        Read lines from the subprocess stdout in a background thread.

        This prevents a pipe-buffer deadlock on Windows where the default
        pipe buffer is only ~4 KB.  Without draining, llama-server blocks
        on writes and never becomes healthy.

        Each line is also teed to ``self._llama_log_fh`` when set so a
        post-mortem (especially in CI) has the full subprocess output
        even if the crash predates the drain-thread join in
        ``_wait_for_health``.
        """
        try:
            for line in self._process.stdout:
                line = line.rstrip()
                if line:
                    self._stdout_lines.append(line)
                    logger.debug(f"[llama-server] {line}")
                    fh = getattr(self, "_llama_log_fh", None)
                    if fh is not None:
                        try:
                            fh.write(line + "\n")
                            fh.flush()
                        except (ValueError, OSError):
                            # Log file closed under us; tee silently.
                            pass
        except (ValueError, OSError):
            # Pipe closed — process is terminating
            pass

    # GGUF KV type sizes for fast skipping
    _GGUF_TYPE_SIZE = {
        0: 1,
        1: 1,
        2: 2,
        3: 2,
        4: 4,
        5: 4,
        6: 4,
        7: 1,
        10: 8,
        11: 8,
        12: 8,
    }

    @staticmethod
    def _gguf_skip_value(f, vtype: int) -> None:
        """Skip a GGUF KV value without reading it."""
        sz = LlamaCppBackend._GGUF_TYPE_SIZE.get(vtype)
        if sz is not None:
            f.seek(sz, 1)
        elif vtype == 8:  # STRING
            slen = struct.unpack("<Q", f.read(8))[0]
            f.seek(slen, 1)
        elif vtype == 9:  # ARRAY
            atype = struct.unpack("<I", f.read(4))[0]
            alen = struct.unpack("<Q", f.read(8))[0]
            elem_sz = LlamaCppBackend._GGUF_TYPE_SIZE.get(atype)
            if elem_sz is not None:
                f.seek(elem_sz * alen, 1)
            elif atype == 8:
                for _ in range(alen):
                    slen = struct.unpack("<Q", f.read(8))[0]
                    f.seek(slen, 1)
            else:
                for _ in range(alen):
                    LlamaCppBackend._gguf_skip_value(f, atype)

    @staticmethod
    def _gguf_read_array_value(f, atype: int, alen: int) -> Optional[list]:
        if atype == 4:  # UINT32
            return [struct.unpack("<I", f.read(4))[0] for _ in range(alen)]
        if atype == 5:  # INT32
            return [struct.unpack("<i", f.read(4))[0] for _ in range(alen)]
        if atype == 7:  # BOOL
            return [struct.unpack("<?", f.read(1))[0] for _ in range(alen)]

        for _ in range(alen):
            LlamaCppBackend._gguf_skip_value(f, atype)
        return None

    def _read_gguf_metadata(self, gguf_path: str) -> None:
        """Read context_length, architecture params, and chat_template from a GGUF header.

        Parses only the KV pairs we need (~30ms even for multi-GB files).
        For split GGUFs, metadata is always in shard 1.
        """
        # Reset metadata from any previously loaded model so stale flags
        # (eg _supports_reasoning) do not carry over when switching models.
        self._context_length = None
        self._chat_template = None
        self._supports_reasoning = False
        self._reasoning_always_on = False
        self._reasoning_style = "enable_thinking"
        self._reasoning_default = True
        self._supports_preserve_thinking = False
        self._supports_tools = False
        self._n_layers = None
        self._n_kv_heads = None
        self._n_kv_heads_by_layer = None
        self._n_heads = None
        self._embedding_length = None
        self._kv_key_length = None
        self._kv_value_length = None
        self._sliding_window = None
        self._sliding_window_pattern = None
        self._full_attention_interval = None
        self._kv_lora_rank = None
        self._key_length_mla = None
        self._kv_key_length_swa = None
        self._kv_value_length_swa = None
        self._ssm_inner_size = None
        self._ssm_state_size = None
        self._shared_kv_layers = None
        self._nextn_predict_layers = None

        try:
            WANTED = {
                "general.architecture",
                "tokenizer.chat_template",
                # Source-repo hints for the SWA resolver's HF fallback.
                "general.source.huggingface.repository",
                "general.source.url",
                "general.source.repo_url",
                "general.base_model.0.repo_url",
                "general.base_model.0.organization",
                "general.base_model.0.name",
                "general.basename",
                "general.organization",
                "general.size_label",
                "general.finetune",
            }
            # Additional arch-specific keys are added dynamically once
            # we know the architecture name.
            arch_keys: dict[str, str] = {}  # gguf_key -> attribute name
            arch = None
            sliding_window_pattern_period: Optional[int] = None
            general: dict[str, str] = {}

            with open(gguf_path, "rb") as f:
                magic = struct.unpack("<I", f.read(4))[0]
                if magic != 0x46554747:  # b"GGUF" as little-endian u32
                    return
                _version = struct.unpack("<I", f.read(4))[0]
                _tensor_count, kv_count = struct.unpack("<QQ", f.read(16))

                for _ in range(kv_count):
                    # Tolerate truncated input (e.g., a partial header
                    # fetched via HTTP byte-range): bail out gracefully
                    # so the resolver fallback still runs on whatever
                    # we did manage to parse.
                    try:
                        key_len_bytes = f.read(8)
                        if len(key_len_bytes) < 8:
                            break
                        key_len = struct.unpack("<Q", key_len_bytes)[0]
                        key_bytes = f.read(key_len)
                        if len(key_bytes) < key_len:
                            break
                        key = key_bytes.decode("utf-8")
                        vtype_bytes = f.read(4)
                        if len(vtype_bytes) < 4:
                            break
                        vtype = struct.unpack("<I", vtype_bytes)[0]
                    except (struct.error, UnicodeDecodeError):
                        break

                    try:
                        if key in WANTED or key in arch_keys:
                            if vtype == 8:  # STRING
                                slen = struct.unpack("<Q", f.read(8))[0]
                                val_s = f.read(slen).decode("utf-8")
                                if (
                                    key.startswith("general.")
                                    and key != "general.architecture"
                                ):
                                    general[key] = val_s
                                if key == "general.architecture":
                                    arch = val_s
                                    arch_keys = {
                                        f"{arch}.context_length": "context_length",
                                        f"{arch}.block_count": "n_layers",
                                        f"{arch}.attention.head_count_kv": "n_kv_heads",
                                        f"{arch}.attention.head_count": "n_heads",
                                        f"{arch}.embedding_length": "embedding_length",
                                        f"{arch}.attention.key_length": "kv_key_length",
                                        f"{arch}.attention.value_length": "kv_value_length",
                                        f"{arch}.attention.sliding_window": "sliding_window",
                                        f"{arch}.attention.sliding_window_pattern": "sliding_window_pattern",
                                        f"{arch}.full_attention_interval": "full_attention_interval",
                                        f"{arch}.attention.kv_lora_rank": "kv_lora_rank",
                                        f"{arch}.attention.key_length_mla": "key_length_mla",
                                        f"{arch}.attention.key_length_swa": "kv_key_length_swa",
                                        f"{arch}.attention.value_length_swa": "kv_value_length_swa",
                                        f"{arch}.attention.shared_kv_layers": "shared_kv_layers",
                                        f"{arch}.ssm.inner_size": "ssm_inner_size",
                                        f"{arch}.ssm.state_size": "ssm_state_size",
                                        f"{arch}.nextn_predict_layers": "nextn_predict_layers",
                                    }
                                elif key == "tokenizer.chat_template":
                                    self._chat_template = val_s
                            elif vtype in (4, 10):  # UINT32 or UINT64
                                val_i = (
                                    struct.unpack("<I", f.read(4))[0]
                                    if vtype == 4
                                    else struct.unpack("<Q", f.read(8))[0]
                                )
                                attr = arch_keys.get(key)
                                if attr:
                                    if attr == "sliding_window_pattern":
                                        sliding_window_pattern_period = val_i
                                    else:
                                        setattr(self, f"_{attr}", val_i)
                            elif vtype == 9:  # ARRAY
                                atype = struct.unpack("<I", f.read(4))[0]
                                alen = struct.unpack("<Q", f.read(8))[0]
                                val_a = self._gguf_read_array_value(f, atype, alen)
                                attr = arch_keys.get(key)
                                if attr == "n_kv_heads" and val_a is not None:
                                    self._n_kv_heads_by_layer = [int(x) for x in val_a]
                                    if self._n_kv_heads is None and val_a:
                                        self._n_kv_heads = max(int(x) for x in val_a)
                                elif (
                                    attr == "sliding_window_pattern"
                                    and val_a is not None
                                ):
                                    self._sliding_window_pattern = [
                                        bool(x) for x in val_a
                                    ]
                                    sliding_window_pattern_period = None
                            else:
                                self._gguf_skip_value(f, vtype)
                        else:
                            self._gguf_skip_value(f, vtype)
                    except (struct.error, UnicodeDecodeError):
                        # Truncated input (e.g., HTTP byte-range fetch
                        # of just the GGUF header); break so the
                        # resolver fallback still runs on what we have.
                        break

            # Expand a scalar period straight from the GGUF first.
            if (
                self._sliding_window_pattern is None
                and sliding_window_pattern_period
                and self._n_layers
            ):
                self._sliding_window_pattern = [
                    (i + 1) % sliding_window_pattern_period != 0
                    for i in range(self._n_layers)
                ]

            # Otherwise hand off to the resolver (cache / bootstrap /
            # transformers / HF). See `_resolve_swa_pattern`.
            if (
                self._sliding_window_pattern is None
                and self._sliding_window
                and self._n_layers
            ):
                hf_repo_candidates = (
                    general.get("general.source.huggingface.repository"),
                    _hf_repo_from_url(general.get("general.source.url")),
                    _hf_repo_from_url(general.get("general.source.repo_url")),
                    _hf_repo_from_url(general.get("general.base_model.0.repo_url")),
                    (
                        f"{general['general.base_model.0.organization']}/"
                        f"{general['general.base_model.0.name']}".replace(" ", "-")
                        if general.get("general.base_model.0.organization")
                        and general.get("general.base_model.0.name")
                        else None
                    ),
                    (
                        f"{general['general.organization']}/"
                        f"{general['general.basename']}".replace(" ", "-")
                        if general.get("general.organization")
                        and general.get("general.basename")
                        else None
                    ),
                )
                self._sliding_window_pattern = _resolve_swa_pattern(
                    arch,
                    self._n_layers,
                    hf_repo_candidates,
                )

            if self._context_length:
                logger.info(f"GGUF metadata: context_length={self._context_length}")
            if self._chat_template:
                logger.info(
                    f"GGUF metadata: chat_template={len(self._chat_template)} chars"
                )
                # Detect thinking/reasoning support from chat template
                flags = detect_reasoning_flags(
                    self._chat_template,
                    self._model_identifier,
                    log_source = "GGUF metadata",
                )
                self._supports_reasoning = flags["supports_reasoning"]
                self._reasoning_style = flags["reasoning_style"]
                self._reasoning_always_on = flags["reasoning_always_on"]
                self._supports_preserve_thinking = flags["supports_preserve_thinking"]
                self._supports_tools = flags["supports_tools"]
        except Exception as e:
            logger.warning(f"Failed to read GGUF metadata: {e}")

    # ── HF download (no lock held) ───────────────────────────────

    def _download_gguf(
        self,
        *,
        hf_repo: str,
        hf_variant: Optional[str] = None,
        hf_token: Optional[str] = None,
    ) -> str:
        """Download GGUF file(s) from HuggingFace. Returns local path.

        Runs WITHOUT self._lock so that unload_model() can set
        _cancel_event at any time. Checks _cancel_event between
        each shard download.
        """
        try:
            from huggingface_hub import hf_hub_download
        except ImportError:
            raise RuntimeError(
                "huggingface_hub is required for HF model loading. "
                "Install it with: pip install huggingface_hub"
            )

        # Determine the filename from the variant
        gguf_filename = None
        gguf_extra_shards: list[str] = []
        if hf_variant:
            try:
                from huggingface_hub import list_repo_files

                files = list_repo_files(hf_repo, token = hf_token)
                variant_lower = hf_variant.lower()
                boundary = re.compile(
                    r"(?<![a-zA-Z0-9])" + re.escape(variant_lower) + r"(?![a-zA-Z0-9])"
                )
                gguf_files = sorted(
                    f
                    for f in files
                    if f.endswith(".gguf") and boundary.search(f.lower())
                )
                if gguf_files:
                    gguf_filename = gguf_files[0]
                    m = _SHARD_FULL_RE.match(gguf_filename)
                    if m:
                        prefix = m.group(1)
                        total = m.group(3)
                        sibling_pat = re.compile(
                            r"^"
                            + re.escape(prefix)
                            + r"-\d{5}-of-"
                            + re.escape(total)
                            + r"\.gguf$"
                        )
                        gguf_extra_shards = [
                            f for f in gguf_files[1:] if sibling_pat.match(f)
                        ]
            except Exception as e:
                logger.warning(f"Could not list repo files: {e}")

            # Offline: resolve variant -> filename from the local HF cache.
            # The heuristic below assumes filenames echo the repo name,
            # which breaks for e.g. Qwen3.6-27B-MTP-GGUF (no "MTP" in file).
            # Match against the rel path (not just basename) so subdir
            # layouts like ``BF16/foo.gguf`` are findable.
            if not gguf_filename:
                try:
                    from utils.models.model_config import _iter_hf_cache_snapshots

                    boundary = re.compile(
                        r"(?<![a-zA-Z0-9])"
                        + re.escape(hf_variant.lower())
                        + r"(?![a-zA-Z0-9])"
                    )
                    for snap in _iter_hf_cache_snapshots(hf_repo):
                        matches = sorted(
                            p.relative_to(snap).as_posix()
                            for p in snap.rglob("*.gguf")
                            if "mmproj" not in p.name.lower()
                            and boundary.search(p.relative_to(snap).as_posix().lower())
                        )
                        if not matches:
                            continue
                        gguf_filename = matches[0]
                        m = _SHARD_FULL_RE.match(Path(gguf_filename).name)
                        if m:
                            prefix = m.group(1)
                            total = m.group(3)
                            sibling_pat = re.compile(
                                r"^"
                                + re.escape(prefix)
                                + r"-\d{5}-of-"
                                + re.escape(total)
                                + r"\.gguf$"
                            )
                            gguf_extra_shards = [
                                f
                                for f in matches[1:]
                                if sibling_pat.match(Path(f).name)
                            ]
                        logger.info(
                            "Resolved variant %s -> %s from local HF cache",
                            hf_variant,
                            gguf_filename,
                        )
                        break
                except Exception as e:
                    logger.debug(f"Offline cache lookup for variant failed: {e}")

            if not gguf_filename:
                repo_name = hf_repo.split("/")[-1].replace("-GGUF", "")
                gguf_filename = f"{repo_name}-{hf_variant}.gguf"

        # Check disk space and fall back to a smaller variant if needed
        all_gguf_files = [gguf_filename] + gguf_extra_shards
        try:
            from huggingface_hub import get_paths_info, try_to_load_from_cache

            path_infos = list(get_paths_info(hf_repo, all_gguf_files, token = hf_token))
            total_bytes = sum((p.size or 0) for p in path_infos)

            # Subtract bytes already present in the HF cache so we only
            # preflight against what we actually have to download. Without
            # this, re-loading a cached large model (e.g. MiniMax-M2.7-GGUF
            # at 131 GB) fails cold whenever free disk is below the full
            # weight footprint, even though nothing needs downloading.
            already_cached_bytes = 0
            for p in path_infos:
                if not p.size:
                    continue
                try:
                    cached_path = try_to_load_from_cache(hf_repo, p.path)
                except Exception:
                    cached_path = None
                if isinstance(cached_path, str) and os.path.exists(cached_path):
                    try:
                        on_disk = os.path.getsize(cached_path)
                    except OSError:
                        on_disk = 0
                    # Count as satisfied only when the full blob is present.
                    if on_disk >= p.size:
                        already_cached_bytes += p.size

            total_download_bytes = max(0, total_bytes - already_cached_bytes)

            if total_download_bytes > 0:
                cache_dir = os.environ.get(
                    "HF_HUB_CACHE",
                    str(Path.home() / ".cache" / "huggingface" / "hub"),
                )
                Path(cache_dir).mkdir(parents = True, exist_ok = True)
                free_bytes = shutil.disk_usage(cache_dir).free

                total_gb = total_download_bytes / (1024**3)
                free_gb = free_bytes / (1024**3)
                cached_gb = already_cached_bytes / (1024**3)

                logger.info(
                    f"GGUF download: {total_gb:.1f} GB needed "
                    f"({cached_gb:.1f} GB already cached), "
                    f"{free_gb:.1f} GB free on disk"
                )

                if total_download_bytes > free_bytes:
                    smaller = self._find_smallest_fitting_variant(
                        hf_repo,
                        free_bytes,
                        hf_token,
                    )
                    if smaller:
                        fallback_file, fallback_size = smaller
                        logger.info(
                            f"Selected variant too large ({total_gb:.1f} GB), "
                            f"falling back to {fallback_file} ({fallback_size / (1024**3):.1f} GB)"
                        )
                        gguf_filename = fallback_file
                        _m = _SHARD_RE.match(gguf_filename)
                        _prefix = _m.group(1) if _m else None
                        if _prefix:
                            gguf_extra_shards = sorted(
                                f
                                for f in all_gguf_files
                                if f.startswith(_prefix)
                                and f != gguf_filename
                                and "mmproj" not in f.lower()
                            )
                        else:
                            gguf_extra_shards = []
                    else:
                        raise RuntimeError(
                            f"Not enough disk space to download any variant. "
                            f"Only {free_gb:.1f} GB free in {cache_dir}"
                        )
        except RuntimeError:
            raise
        except Exception as e:
            logger.warning(f"Could not check disk space: {e}")

        gguf_label = f"{hf_repo}/{gguf_filename}" + (
            f" (+{len(gguf_extra_shards)} shards)" if gguf_extra_shards else ""
        )
        logger.info(f"Resolving GGUF: {gguf_label}")
        try:
            if self._cancel_event.is_set():
                raise RuntimeError("Cancelled")
            dl_start = time.monotonic()
            local_path = hf_hub_download(
                repo_id = hf_repo,
                filename = gguf_filename,
                token = hf_token,
            )
            for shard in gguf_extra_shards:
                if self._cancel_event.is_set():
                    raise RuntimeError("Cancelled")
                logger.info(f"Resolving GGUF shard: {shard}")
                hf_hub_download(
                    repo_id = hf_repo,
                    filename = shard,
                    token = hf_token,
                )
        except RuntimeError as e:
            if "Cancelled" in str(e):
                raise
            raise RuntimeError(
                f"Failed to download GGUF file '{gguf_filename}' from {hf_repo}: {e}"
            )
        except Exception as e:
            raise RuntimeError(
                f"Failed to download GGUF file '{gguf_filename}' from {hf_repo}: {e}"
            )

        dl_elapsed = time.monotonic() - dl_start
        if dl_elapsed < 2.0:
            logger.info(f"GGUF resolved from cache: {local_path}")
        else:
            logger.info(f"GGUF downloaded in {dl_elapsed:.1f}s: {local_path}")
        return local_path

    def _download_mmproj(
        self,
        *,
        hf_repo: str,
        hf_token: Optional[str] = None,
    ) -> Optional[str]:
        """Download the mmproj (vision projection) file from a GGUF repo.

        Prefers mmproj-F16.gguf, falls back to any mmproj*.gguf file.
        Returns the local path, or None if no mmproj file exists.
        """

        def _pick_mmproj(candidates: list[str]) -> Optional[str]:
            mmproj_files = sorted(
                f
                for f in candidates
                if f.lower().endswith(".gguf") and "mmproj" in Path(f).name.lower()
            )
            if not mmproj_files:
                return None
            for f in mmproj_files:
                if f.lower().endswith("-f16.gguf"):
                    return f
            return mmproj_files[0]

        target: Optional[str] = None
        try:
            from huggingface_hub import list_repo_files

            target = _pick_mmproj(list_repo_files(hf_repo, token = hf_token))
        except Exception as e:
            logger.debug(f"Could not list repo files for mmproj: {e}")

        # Offline: resolve mmproj from the local HF cache snapshot, same
        # shape as _download_gguf's offline fallback above.
        if target is None:
            try:
                from utils.models.model_config import _iter_hf_cache_snapshots

                for snap in _iter_hf_cache_snapshots(hf_repo):
                    rel_files = [
                        p.relative_to(snap).as_posix() for p in snap.rglob("*.gguf")
                    ]
                    target = _pick_mmproj(rel_files)
                    if target is not None:
                        logger.info("Resolved mmproj %s from local HF cache", target)
                        break
            except Exception as e:
                logger.debug(f"Offline cache lookup for mmproj failed: {e}")

        if target is None:
            return None

        try:
            from huggingface_hub import hf_hub_download

            logger.info(f"Downloading mmproj: {hf_repo}/{target}")
            local_path = hf_hub_download(
                repo_id = hf_repo,
                filename = target,
                token = hf_token,
            )
            return local_path
        except Exception as e:
            logger.warning(f"Could not download mmproj: {e}")
            return None

    def _resolve_launch_mmproj_path(
        self,
        *,
        model_path: str,
        mmproj_path: Optional[str],
    ) -> Optional[str]:
        """Return mmproj_path iff it exists on disk AND matches the model family.

        Returns None if mmproj_path is None, missing on disk, or family-mismatched.
        """
        if not mmproj_path:
            return None

        mmproj = Path(mmproj_path)
        if not mmproj.is_file():
            logger.warning(f"mmproj file not found: {mmproj_path}")
            return None

        from utils.models.model_config import mmproj_matches_model_family

        if not mmproj_matches_model_family(model_path, str(mmproj)):
            logger.warning(
                f"mmproj does not match model family: model={Path(model_path).name} "
                f"mmproj={mmproj.name}"
            )
            return None

        return str(mmproj)

    # ── Lifecycle ─────────────────────────────────────────────────

    def load_model(
        self,
        *,
        # Local mode: pass a path to a .gguf file
        gguf_path: Optional[str] = None,
        # Vision projection (mmproj) for local vision models
        mmproj_path: Optional[str] = None,
        # HF mode: let llama-server download via -hf "repo:quant"
        hf_repo: Optional[str] = None,
        hf_variant: Optional[str] = None,
        hf_token: Optional[str] = None,
        # Common
        model_identifier: str,
        is_vision: bool = False,
        n_ctx: int = 4096,
        chat_template_override: Optional[str] = None,
        cache_type_kv: Optional[str] = None,
        speculative_type: Optional[str] = None,
        spec_draft_n_max: Optional[int] = None,
        n_threads: Optional[int] = None,
        n_gpu_layers: Optional[int] = None,  # Accepted for caller compat, unused
        n_parallel: int = 1,
        extra_args: Optional[List[str]] = None,
    ) -> bool:
        """
        Start llama-server with a GGUF model.

        Two modes:
        - Local: ``gguf_path="/path/to/model.gguf"`` → uses ``-m``
        - HF:    ``hf_repo="unsloth/gemma-3-4b-it-GGUF", hf_variant="Q4_K_M"`` → uses ``-hf``

        In HF mode, llama-server handles downloading, caching, and
        auto-loading mmproj files for vision models.

        Returns True if server started and health check passed.
        """
        # Serialise the whole load so concurrent /load calls never
        # leave two llama-server processes alive (#5401 / #5161). Does
        # not block /unload, /status, /load-progress.
        with self._serial_load_lock:
            # Duplicate /load that raced past the route-level check
            # (the first one hadn't published _healthy=True yet). If the
            # live server already satisfies this request, do nothing.
            if self._already_in_target_state(
                gguf_path = gguf_path,
                model_identifier = model_identifier,
                hf_variant = hf_variant,
                n_ctx = n_ctx,
                cache_type_kv = cache_type_kv,
                speculative_type = speculative_type,
                spec_draft_n_max = spec_draft_n_max,
                chat_template_override = chat_template_override,
                extra_args = extra_args,
                is_vision = is_vision,
            ):
                logger.info(
                    f"load_model: backend already in target state for "
                    f"'{model_identifier}', skipping reload"
                )
                # Retry probe only if a prior attempt didn't complete.
                if not self._audio_probed:
                    try:
                        detected = self._detect_audio_type_strict()
                        self._audio_probed = True
                    except Exception as exc:
                        logger.debug("Fast-path audio probe failed: %s", exc)
                        detected = None
                    if detected in ("snac", "bicodec", "dac"):
                        with self._lock:
                            if not self._healthy:
                                return False
                            try:
                                self.init_audio_codec(detected)
                                self._is_audio = True
                                self._audio_type = detected
                            except Exception as exc:
                                logger.warning(
                                    "Failed to init audio codec '%s': %s",
                                    detected,
                                    exc,
                                )
                                self._audio_probed = False
                                return False
                    elif detected:
                        # csm / whisper / audio_vlm: track type but keep
                        # _is_audio False -- GGUF TTS routing only fires
                        # for snac/bicodec/dac.
                        with self._lock:
                            if not self._healthy:
                                return False
                            self._audio_type = detected
                if not self._healthy:
                    return False
                return True

            self._cancel_event.clear()

            # ── Phase 1: kill old process (under lock, fast) ──────────
            with self._lock:
                self._kill_process()

            binary = self._find_llama_server_binary()
            if not binary:
                raise RuntimeError(
                    "llama-server binary not found. "
                    "Run setup.sh to build it, install llama.cpp, "
                    "or set LLAMA_SERVER_PATH environment variable."
                )
            is_vulkan_backend = self._is_vulkan_backend(binary)

            # ── Phase 2: download (NO lock held, so cancel can proceed) ──
            # Scope HF_HUB_OFFLINE to the download block only when DNS is
            # dead; cleanup runs even on exception so a transient hiccup
            # at the start of one load cannot quarantine future loads.
            if hf_repo:
                with _hf_offline_if_dns_dead():
                    model_path = self._download_gguf(
                        hf_repo = hf_repo,
                        hf_variant = hf_variant,
                        hf_token = hf_token,
                    )
                    # Auto-download mmproj for vision models
                    if is_vision and not mmproj_path:
                        mmproj_path = self._download_mmproj(
                            hf_repo = hf_repo,
                            hf_token = hf_token,
                        )
            elif gguf_path:
                if not Path(gguf_path).is_file():
                    raise FileNotFoundError(f"GGUF file not found: {gguf_path}")
                model_path = gguf_path
            else:
                raise ValueError("Either gguf_path or hf_repo must be provided")

            # Set identifier early so _read_gguf_metadata can use it for DeepSeek detection
            self._model_identifier = model_identifier

            # Read GGUF metadata (context_length, chat_template) -- fast, header only
            self._read_gguf_metadata(model_path)

            # Check cancel after download
            if self._cancel_event.is_set():
                logger.info("Load cancelled after download phase")
                return False

            # Outside ``self._lock`` so /unload, /cancel, /status are
            # not blocked. ``unload_model`` also records the kill, so
            # the frontend /unload+/load Apply path engages the wait
            # here even though no in-process kill happened.
            self._wait_for_vram_settle(since_kill = self._last_kill_monotonic)

            # ── Phase 3: start llama-server (under lock) ──────────────
            with self._lock:
                # Re-check cancel inside lock
                if self._cancel_event.is_set():
                    logger.info("Load cancelled before server start")
                    return False

                self._port = self._find_free_port()

                # Select GPU(s) based on model size + estimated KV cache.
                # Seed safe defaults before GPU probing so the except path
                # still has valid state to publish.
                effective_ctx = n_ctx if n_ctx > 0 else (self._context_length or 0)
                max_available_ctx = self._context_length or effective_ctx
                gpus: list[tuple[int, int]] = []
                try:
                    model_size = self._get_gguf_size_bytes(model_path)
                    gpus = self._get_gpu_free_memory(binary)

                    # Resolve effective context: 0 means let llama-server use the
                    # model's native length.  Only expand to a known native length
                    # if metadata is available; otherwise preserve 0 as a sentinel.
                    if n_ctx > 0:
                        effective_ctx = n_ctx
                    elif self._context_length is not None:
                        effective_ctx = self._context_length
                    else:
                        effective_ctx = 0
                    original_ctx = effective_ctx
                    # Default UI ceiling to the model's native context length.
                    # GPU/VRAM-fit logic below may shrink this if hardware is limited.
                    max_available_ctx = self._context_length or effective_ctx

                    # Will MTP engage on this load? If so, the auto-fit
                    # budget needs to reserve extra VRAM for the draft
                    # model's KV cache + compute graph. Mirrors the
                    # canonical-mode resolver in _build_speculative_flags:
                    # forced mtp / mtp+ngram always engage; auto only
                    # engages on an MTP GGUF >= 3B (sub-3B auto falls
                    # back to ngram-mod which doesn't need headroom);
                    # ngram / ngram-simple / off never engage MTP.
                    _mtp_canonical = _canonicalize_spec_mode(speculative_type)
                    _mtp_effective = _mtp_canonical or "auto"
                    _mtp_size_for_fit = _extract_model_size_b(model_identifier)
                    _mtp_sub_3b_for_fit = (
                        _mtp_size_for_fit is not None and _mtp_size_for_fit < 3.0
                    )
                    _mtp_will_engage = bool(
                        not _extra_args_set_spec_type(extra_args)
                        and (
                            _mtp_effective in ("mtp", "mtp+ngram")
                            or (
                                _mtp_effective == "auto"
                                and (
                                    bool(self._nextn_predict_layers)
                                    or _is_mtp_model_name(model_identifier, model_path)
                                )
                                and not _mtp_sub_3b_for_fit
                            )
                        )
                    )

                    # Auto-cap context to fit in GPU VRAM and select GPUs.
                    #
                    # Two policies depending on whether the user set n_ctx:
                    #
                    # Explicit n_ctx (user chose a context length):
                    #   Honor it. Try the full requested context with _select_gpus
                    #   (which uses as many GPUs as needed). Only cap if it doesn't
                    #   fit on any GPU combination.
                    #
                    # Auto n_ctx=0 (model's native context):
                    #   Prefer fewer GPUs with reduced context over more GPUs,
                    #   since multi-GPU is slower and the user didn't ask for a
                    #   specific context length.
                    gpu_indices, use_fit = None, True
                    explicit_ctx = n_ctx > 0

                    if gpus and self._can_estimate_kv() and effective_ctx > 0:
                        # Compute the largest hardware-aware cap from the model's
                        # native context across all usable GPU subsets (for UI
                        # bounds), independent of the currently requested context.
                        native_ctx_for_cap = self._context_length or effective_ctx
                        if native_ctx_for_cap > 0:
                            ranked_for_cap = sorted(
                                gpus, key = lambda g: g[1], reverse = True
                            )
                            best_cap = 0
                            for n_gpus in range(1, len(ranked_for_cap) + 1):
                                subset = ranked_for_cap[:n_gpus]
                                pool_mib = sum(free for _, free in subset)
                                capped = self._fit_context_to_vram(
                                    native_ctx_for_cap,
                                    pool_mib,
                                    model_size,
                                    cache_type_kv,
                                    n_parallel = n_parallel,
                                    mtp_engaged = _mtp_will_engage,
                                )
                                kv = self._estimate_kv_cache_bytes(
                                    capped, cache_type_kv, n_parallel = n_parallel
                                )
                                total_mib = (model_size + kv) / (1024 * 1024)
                                if total_mib <= pool_mib * 0.90:
                                    best_cap = max(best_cap, capped)
                            if best_cap > 0:
                                max_available_ctx = best_cap
                            else:
                                # Weights exceed 90% of every GPU subset's free
                                # memory, so there is no fitting context. Anchor
                                # the UI's "safe zone" threshold at 4096 (the
                                # spec's default when the model cannot fit) so
                                # the ctx slider shows the "might be slower"
                                # warning as soon as the user drags above the
                                # fallback default instead of never.
                                max_available_ctx = min(4096, native_ctx_for_cap)

                        if explicit_ctx:
                            # Honor the user's requested context verbatim. If it
                            # fits, pin GPUs and skip --fit; if it doesn't, ship
                            # -c <user_ctx> --fit on and let llama-server flex
                            # -ngl (CPU layer offload). The UI is expected to
                            # have surfaced the "might be slower" warning before
                            # the user submitted a ctx above the fit ceiling.
                            requested_total = (
                                model_size
                                + self._estimate_kv_cache_bytes(
                                    effective_ctx, cache_type_kv, n_parallel = n_parallel
                                )
                            )
                            gpu_indices, use_fit = self._select_gpus(
                                requested_total, gpus
                            )
                            # No silent shrink: effective_ctx stays == n_ctx.
                        else:
                            # Auto context: prefer fewer GPUs, cap context
                            # to fit. Same headroom threshold as
                            # _select_gpus (#5106).
                            ranked = sorted(gpus, key = lambda g: g[1], reverse = True)
                            pin_fraction = self._GPU_PIN_VRAM_FRACTION
                            for n_gpus in range(1, len(ranked) + 1):
                                subset = ranked[:n_gpus]
                                pool_mib = sum(free for _, free in subset)
                                capped = self._fit_context_to_vram(
                                    effective_ctx,
                                    pool_mib,
                                    model_size,
                                    cache_type_kv,
                                    n_parallel = n_parallel,
                                    mtp_engaged = _mtp_will_engage,
                                )
                                kv = self._estimate_kv_cache_bytes(
                                    capped, cache_type_kv, n_parallel = n_parallel
                                )
                                total_mib = (model_size + kv) / (1024 * 1024)
                                if total_mib <= pool_mib * pin_fraction:
                                    effective_ctx = capped
                                    gpu_indices = sorted(idx for idx, _ in subset)
                                    use_fit = False
                                    break
                            else:
                                # Native ctx doesn't fit. Drop to 4096 and
                                # re-check before deferring to --fit on:
                                # a model that overflows at 131k may pin
                                # comfortably with a 4096 KV cache (#5106).
                                effective_ctx = min(4096, effective_ctx)
                                if effective_ctx > 0:
                                    for n_gpus in range(1, len(ranked) + 1):
                                        subset = ranked[:n_gpus]
                                        pool_mib = sum(free for _, free in subset)
                                        kv = self._estimate_kv_cache_bytes(
                                            effective_ctx,
                                            cache_type_kv,
                                            n_parallel = n_parallel,
                                        )
                                        total_mib = (model_size + kv) / (1024 * 1024)
                                        if total_mib <= pool_mib * pin_fraction:
                                            gpu_indices = sorted(
                                                idx for idx, _ in subset
                                            )
                                            use_fit = False
                                            break

                    elif gpus:
                        # Can't estimate KV -- fall back to file-size-only check.
                        # Without KV estimation we cannot prove a hardware cap, so
                        # keep the ceiling at the native context (already the default).
                        logger.debug(
                            "Falling back to file-size-only GPU selection",
                            model_size_gb = round(model_size / (1024**3), 2),
                        )
                        gpu_indices, use_fit = self._select_gpus(model_size, gpus)
                        if use_fit and not explicit_ctx:
                            # Weights don't fit on any subset. Default the UI to
                            # 4096 so the slider doesn't land on an unusable native
                            # context. --fit on will flex -ngl at runtime.
                            effective_ctx = (
                                min(4096, effective_ctx) if effective_ctx > 0 else 4096
                            )

                    if effective_ctx < original_ctx:
                        kv_est = self._estimate_kv_cache_bytes(
                            effective_ctx, cache_type_kv, n_parallel = n_parallel
                        )
                        logger.info(
                            f"Context auto-reduced: {original_ctx} -> {effective_ctx} "
                            f"(model: {model_size / (1024**3):.1f} GB, "
                            f"est. KV cache: {kv_est / (1024**3):.1f} GB)"
                        )

                    kv_cache_bytes = self._estimate_kv_cache_bytes(
                        effective_ctx, cache_type_kv, n_parallel = n_parallel
                    )
                    logger.info(
                        f"GGUF size: {model_size / (1024**3):.1f} GB, "
                        f"est. KV cache: {kv_cache_bytes / (1024**3):.1f} GB, "
                        f"context: {effective_ctx}, "
                        f"GPUs free: {gpus}, selected: {gpu_indices}, fit: {use_fit}"
                    )
                except Exception as e:
                    logger.warning(f"GPU selection failed ({e}), using --fit on")
                    gpu_indices, use_fit = None, True
                    effective_ctx = n_ctx  # fall back to original

                launch_mmproj_path = self._resolve_launch_mmproj_path(
                    model_path = model_path,
                    mmproj_path = mmproj_path,
                )
                # Need both a resolved mmproj AND the config vision flag; a stray
                # mmproj passing the family-name heuristic must not flip a non-VLM
                # GGUF into vision mode.
                effective_is_vision = bool(launch_mmproj_path) and bool(is_vision)
                if is_vision and not effective_is_vision:
                    logger.warning(
                        "Vision-capable GGUF loaded without a usable mmproj; "
                        "image input will be disabled for this session"
                    )

                cmd = [
                    binary,
                    "-m",
                    model_path,
                    "--port",
                    str(self._port),
                    "-c",
                    str(effective_ctx) if effective_ctx > 0 else "0",
                    "--parallel",
                    str(n_parallel),
                    "--flash-attn",
                    "on",  # Force flash attention for speed
                    # Error out at n_ctx instead of silently rotating the KV cache; frontend catches it and points the user at "Context Length".
                    "--no-context-shift",
                ]

                if use_fit:
                    cmd.extend(["--fit", "on"])
                elif gpu_indices is not None:
                    # Model fits on selected GPU(s) -- offload all layers
                    cmd.extend(["-ngl", "-1"])

                # -1 = llama.cpp auto-detect (physical cores). Pass explicitly so we
                # do not inherit llama-server's internal default, which has historically
                # varied (hardware concurrency incl. hyperthreads on some builds).
                cmd.extend(
                    ["--threads", str(n_threads if n_threads is not None else -1)]
                )

                # Always enable Jinja chat template rendering for proper template support
                cmd.extend(["--jinja"])

                # KV cache data type
                _valid_cache_types = {
                    "f16",
                    "bf16",
                    "q8_0",
                    "q4_0",
                    "q4_1",
                    "q5_0",
                    "q5_1",
                    "iq4_nl",
                    "f32",
                }
                if cache_type_kv and cache_type_kv in _valid_cache_types:
                    cmd.extend(
                        [
                            "--cache-type-k",
                            cache_type_kv,
                            "--cache-type-v",
                            cache_type_kv,
                        ]
                    )
                    self._cache_type_kv = cache_type_kv
                    logger.info(f"KV cache type: {cache_type_kv}")
                else:
                    self._cache_type_kv = None

                # Speculative decoding (n-gram self-speculation, zero VRAM cost)
                # ngram-mod: ~16 MB shared hash pool, constant memory/complexity,
                # variable draft lengths.  Helps most when the model repeats
                # existing text (code refactoring, summarization, reasoning).
                # For general chat with low repetition, overhead is ~5 ms.
                #
                # Benchmarks from upstream llama.cpp speculative-decoding PRs:
                #   Scenario                        | Without | With    | Speedup
                #   gpt-oss-120b code refactor      | 181 t/s | 446 t/s | 2.5x
                #   Qwen3-235B offloaded            |  12 t/s |  21 t/s | 1.8x
                #   gpt-oss-120b repeat (92% accept)| 181 t/s | 814 t/s | 4.5x
                #
                # Params from llama.cpp server README:
                #   --spec-ngram-mod-n-match 24 (lookup length)
                #   --spec-ngram-mod-n-min 48 --spec-ngram-mod-n-max 64
                #   (MoEs need long drafts; dense models can reduce these)
                # ref: https://github.com/ggml-org/llama.cpp/blob/master/docs/speculative.md
                # ref: https://github.com/ggml-org/llama.cpp/pull/19164
                # ref: https://github.com/ggml-org/llama.cpp/pull/18471
                # draft-mtp: MTP heads on Unsloth's *-MTP GGUFs
                # (llama.cpp #22673). Auto-enabled via nextn_predict_layers,
                # fallback to -MTP in name. GPU: MTP-only. CPU/Mac: chain
                # with ngram-mod. See unsloth.ai/docs/models/qwen3.6#mtp-guide.
                spec_flags = self._build_speculative_flags(
                    speculative_type = speculative_type,
                    spec_draft_n_max = spec_draft_n_max,
                    extra_args = extra_args,
                    model_identifier = model_identifier,
                    model_path = model_path,
                    gpus = bool(gpus),
                    binary = binary,
                )
                cmd.extend(spec_flags)

                # Apply custom chat template override if provided
                self._chat_template_override = chat_template_override
                if chat_template_override:
                    import tempfile

                    flags = detect_reasoning_flags(
                        chat_template_override,
                        self._model_identifier,
                        log_source = "GGUF chat template override",
                    )
                    self._supports_reasoning = flags["supports_reasoning"]
                    self._reasoning_style = flags["reasoning_style"]
                    self._reasoning_always_on = flags["reasoning_always_on"]
                    self._supports_preserve_thinking = flags[
                        "supports_preserve_thinking"
                    ]
                    self._supports_tools = flags["supports_tools"]

                    self._chat_template_file = tempfile.NamedTemporaryFile(
                        mode = "w",
                        suffix = ".jinja",
                        delete = False,
                        prefix = "unsloth_chat_template_",
                    )
                    self._chat_template_file.write(chat_template_override)
                    self._chat_template_file.close()
                    cmd.extend(["--chat-template-file", self._chat_template_file.name])
                    logger.info(
                        f"Using custom chat template file: {self._chat_template_file.name}"
                    )

                # For reasoning models, set default thinking mode.
                # Qwen3.5/3.6 models below 9B (0.8B, 2B, 4B) disable thinking by default.
                # Only 9B and larger enable thinking.
                # Always-on templates ignore the kwarg entirely, so skip.
                if self._supports_reasoning and not self._reasoning_always_on:
                    thinking_default = True
                    mid = (model_identifier or "").lower()
                    if "qwen3.5" in mid or "qwen3.6" in mid:
                        size_val = _extract_model_size_b(mid)
                        if size_val is not None and size_val < 9:
                            thinking_default = False
                    self._reasoning_default = thinking_default
                    reasoning_kw = self._reasoning_kwargs(thinking_default)
                    cmd.extend(
                        [
                            "--chat-template-kwargs",
                            json.dumps(reasoning_kw),
                        ]
                    )
                    logger.info(f"Reasoning model: {reasoning_kw} by default")

                if launch_mmproj_path and effective_is_vision:
                    cmd.extend(["--mmproj", launch_mmproj_path])
                    logger.info(f"Using mmproj for vision: {launch_mmproj_path}")

                # Option C: add --api-key for direct client access when enabled
                import os as _os
                import secrets as _secrets

                if _os.getenv("UNSLOTH_DIRECT_STREAM", "0") == "1":
                    self._api_key = _secrets.token_urlsafe(32)
                    cmd.extend(["--api-key", self._api_key])
                    logger.info(
                        "llama-server started with --api-key for direct streaming"
                    )
                else:
                    self._api_key = None

                # User-supplied pass-through args go last so llama.cpp's
                # last-wins flag parsing lets the user override Studio's
                # auto-set tier-2 flags (e.g. --cache-type-k, --spec-type).
                # The route layer has already validated this list against
                # the managed-flag denylist via validate_extra_args().
                if extra_args:
                    cmd.extend(str(a) for a in extra_args)
                    logger.info(
                        f"Appending user extra args to llama-server: {list(extra_args)}"
                    )

                _log_cmd = list(cmd)
                if "--api-key" in _log_cmd:
                    _ki = _log_cmd.index("--api-key") + 1
                    if _ki < len(_log_cmd):
                        _log_cmd[_ki] = "<redacted>"
                logger.info(f"Starting llama-server: {' '.join(_log_cmd)}")

                # Set library paths so llama-server can find its shared libs and CUDA DLLs
                import os
                import sys

                env = child_env_without_native_path_secret()
                binary_dir = str(Path(binary).parent)

                if sys.platform == "win32":
                    # See _build_windows_path_dirs for ordering. #5106.
                    path_dirs = self._build_windows_path_dirs(
                        binary_dir,
                        sys.prefix,
                        os.environ.get("CUDA_PATH", ""),
                    )
                    existing_path = env.get("PATH", "")
                    env["PATH"] = ";".join(path_dirs) + ";" + existing_path
                else:
                    # Linux: set LD_LIBRARY_PATH for shared libs next to the binary
                    # and CUDA runtime libs (libcudart, libcublas, etc.)
                    import platform

                    lib_dirs = [binary_dir]
                    _arch = platform.machine()  # x86_64, aarch64, etc.

                    # Pip-installed nvidia CUDA runtime libs (e.g. torch's
                    # bundled cuda-bindings).  The prebuilt llama.cpp binary
                    # links against libcudart.so.13 / libcublas.so.13 which
                    # live here, not in /usr/local/cuda.
                    import glob as _glob

                    for _nv_pattern in [
                        os.path.join(
                            sys.prefix,
                            "lib",
                            "python*",
                            "site-packages",
                            "nvidia",
                            "cu*",
                            "lib",
                        ),
                        os.path.join(
                            sys.prefix,
                            "lib",
                            "python*",
                            "site-packages",
                            "nvidia",
                            "cudnn",
                            "lib",
                        ),
                        os.path.join(
                            sys.prefix,
                            "lib",
                            "python*",
                            "site-packages",
                            "nvidia",
                            "nvjitlink",
                            "lib",
                        ),
                    ]:
                        for _nv_dir in _glob.glob(_nv_pattern):
                            if os.path.isdir(_nv_dir):
                                lib_dirs.append(_nv_dir)

                    for cuda_lib in [
                        "/usr/local/cuda/lib64",
                        f"/usr/local/cuda/targets/{_arch}-linux/lib",
                        # Fallback CUDA compat paths (e.g. binary built with
                        # CUDA 12 on a system where default /usr/local/cuda
                        # points to CUDA 13+).
                        "/usr/local/cuda-12/lib64",
                        "/usr/local/cuda-12.8/lib64",
                        f"/usr/local/cuda-12/targets/{_arch}-linux/lib",
                        f"/usr/local/cuda-12.8/targets/{_arch}-linux/lib",
                    ]:
                        if os.path.isdir(cuda_lib):
                            lib_dirs.append(cuda_lib)
                    existing_ld = env.get("LD_LIBRARY_PATH", "")
                    new_ld = ":".join(lib_dirs)
                    env["LD_LIBRARY_PATH"] = (
                        f"{new_ld}:{existing_ld}" if existing_ld else new_ld
                    )

                # Pin to selected GPU(s). On ROCm, llama-server (and any torch
                # in the subprocess) honors HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES;
                # narrowing only CUDA_VISIBLE_DEVICES leaves an AMD child seeing
                # the full HIP/ROCR set the parent inherited.
                if gpu_indices is not None:
                    pinned = ",".join(str(i) for i in gpu_indices)
                    if is_vulkan_backend:
                        # gpu_indices are ggml Vulkan ordinals (see
                        # _get_gpu_free_memory); the Vulkan backend ignores
                        # CUDA_VISIBLE_DEVICES, so pin via its own mask.
                        env["GGML_VK_VISIBLE_DEVICES"] = pinned
                    else:
                        env["CUDA_VISIBLE_DEVICES"] = pinned
                        try:
                            import torch as _torch

                            if getattr(_torch.version, "hip", None) is not None:
                                env["HIP_VISIBLE_DEVICES"] = pinned
                                env["ROCR_VISIBLE_DEVICES"] = pinned
                        except Exception as e:
                            logger.debug(
                                "Failed to set ROCm visibility env vars for child: %s",
                                e,
                            )

                # Defensive kill: if a concurrent load slipped past Phase 1
                # (because its `self._process` was None at the time) and
                # already stored a Popen handle here, drop that orphan
                # before we overwrite the reference. See issue #5161.
                self._kill_process()

                self._stdout_lines = []
                # Tee llama-server output to a dedicated log file so a
                # post-mortem in CI (or after a remote-debug session)
                # has the full subprocess trail even when the parent
                # only stored the last 50 lines. Path lives under the
                # studio home so it ships in the same place all other
                # Studio logs live.
                self._llama_log_fh = None
                try:
                    log_dir = _swa_cache_path().parent / "logs" / "llama-server"
                    log_dir.mkdir(parents = True, exist_ok = True)
                    self._llama_log_path = (
                        log_dir / f"llama-{int(time.time())}-port-{self._port}.log"
                    )
                    self._llama_log_fh = open(
                        self._llama_log_path,
                        "w",
                        encoding = "utf-8",
                        buffering = 1,
                    )
                    logger.info(f"llama-server stdout/stderr -> {self._llama_log_path}")
                except OSError as e:
                    # Best-effort; never block the load on logging.
                    logger.debug(f"Could not open llama-server log file: {e}")
                    self._llama_log_path = None
                self._process = subprocess.Popen(
                    cmd,
                    stdout = subprocess.PIPE,
                    stderr = subprocess.STDOUT,
                    text = True,
                    env = env,
                    **_windows_hidden_subprocess_kwargs(),
                )

                # Start background thread to drain stdout and prevent pipe deadlock
                self._stdout_thread = threading.Thread(
                    target = self._drain_stdout, daemon = True, name = "llama-stdout"
                )
                self._stdout_thread.start()

                # Store the resolved on-disk path, not the caller's kwarg. In
                # HF mode the caller passes gguf_path=None and the real path
                # (``model_path``) is what llama-server is actually mmap'ing.
                # Downstream consumers (load_progress, log lines, etc.) need
                # the path that exists on disk.
                self._gguf_path = model_path
                self._hf_repo = hf_repo
                # For local GGUF files, extract variant from filename if not provided
                if hf_variant:
                    self._hf_variant = hf_variant
                elif gguf_path:
                    try:
                        from utils.models.model_config import _extract_quant_label

                        self._hf_variant = _extract_quant_label(gguf_path)
                    except Exception:
                        self._hf_variant = None
                else:
                    self._hf_variant = None
                self._is_vision = effective_is_vision
                self._model_identifier = model_identifier

                # Store the effective (possibly capped) context separately.
                # Do NOT overwrite _context_length -- it holds the model's native
                # context length from GGUF metadata and is used for display/info.
                self._effective_context_length = (
                    effective_ctx if effective_ctx > 0 else self._context_length
                )
                self._max_context_length = (
                    max_available_ctx
                    if max_available_ctx > 0
                    else self._effective_context_length
                )

                # Wait for llama-server to become healthy
                if not self._wait_for_health(timeout = 600.0):
                    self._kill_process()
                    _gguf = gguf_path or ""
                    _is_ollama = (
                        ".studio_links" in _gguf
                        or os.sep + "ollama_links" + os.sep in _gguf
                        or os.sep + ".cache" + os.sep + "ollama" + os.sep in _gguf
                        or (self._model_identifier or "").startswith("ollama/")
                    )
                    # Only show the Ollama-specific message when the server
                    # output indicates a GGUF compatibility issue, not for
                    # unrelated failures like OOM or missing binaries.
                    if _is_ollama:
                        _output = "\n".join(self._stdout_lines[-50:]).lower()
                        _gguf_compat_hints = (
                            "key not found",
                            "unknown model architecture",
                            "failed to load model",
                        )
                        if any(h in _output for h in _gguf_compat_hints):
                            raise RuntimeError(
                                "Some Ollama models do not work with llama.cpp. "
                                "Try a different model, or use this model directly through Ollama instead."
                            )
                    raise RuntimeError(
                        "llama-server failed to start. "
                        "Check that the GGUF file is valid and you have enough memory."
                    )

                self._healthy = True

                # Commit caller intent only after _healthy=True so a
                # failed startup can't poison the next inheritance check.
                # None keeps prior, [] clears, list sets. Source records
                # the caller's hf_variant (None for local files) so the
                # route's same_source check stays symmetric.
                if extra_args is not None:
                    self._extra_args = list(extra_args)
                    self._extra_args_source = (model_identifier, hf_variant)
                self._requested_n_ctx = int(n_ctx)

                # Catch silent CPU fallback when GPU was intended (#5106).
                self._gpu_offload_active = self._classify_gpu_offload(
                    gpu_indices is not None or use_fit, gpus or []
                )
                if self._gpu_offload_active is False:
                    logger.warning(
                        "llama-server appears to have loaded the model entirely "
                        "on CPU even though Studio detected at least one GPU. "
                        "This usually means the prebuilt binary's GPU backend "
                        "failed to load -- on Windows, cudart64_X.dll / "
                        "cublas64_X.dll could not be resolved. Reinstall the "
                        "Studio llama.cpp prebuilt or install a matching CUDA "
                        "toolkit (issue unslothai/unsloth#5106).",
                    )

                logger.info(
                    f"llama-server ready on port {self._port} "
                    f"for model '{model_identifier}'"
                )

            # Probe outside _lock (interruptible by /unload); init inside.
            self._is_audio = False
            self._audio_type = None
            self._audio_probed = False
            try:
                detected = self._detect_audio_type_strict()
                self._audio_probed = True
            except Exception as exc:
                logger.debug("Audio probe failed: %s", exc)
                detected = None
            if detected in ("snac", "bicodec", "dac"):
                with self._lock:
                    if not self._healthy:
                        return False
                    try:
                        self.init_audio_codec(detected)
                        self._is_audio = True
                        self._audio_type = detected
                    except Exception as exc:
                        # Surface as HTTP 500 -- matches pre-PR contract.
                        logger.warning(
                            "Failed to init audio codec '%s': %s",
                            detected,
                            exc,
                        )
                        self._audio_probed = False
                        return False
            elif detected:
                # csm / whisper / audio_vlm: track type but keep _is_audio
                # False -- GGUF TTS routing only fires for snac/bicodec/dac.
                with self._lock:
                    if not self._healthy:
                        return False
                    self._audio_type = detected

            if not self._healthy:
                return False
            return True

    def _build_speculative_flags(
        self,
        *,
        speculative_type: Optional[str],
        spec_draft_n_max: Optional[int],
        extra_args: Optional[List[str]],
        model_identifier: str,
        model_path: Optional[str],
        gpus: bool,
        binary: Optional[str],
    ) -> List[str]:
        """Return the llama-server flag list for the requested spec mode.

        Side effects: sets ``self._speculative_type`` (resolved internal
        emit), ``self._requested_spec_mode`` (canonical UI mode for the
        status round-trip), and ``self._spec_draft_n_max`` (user override
        only; None when the platform default applies).

        Speculative decoding (n-gram self-speculation, zero VRAM cost):
        ngram-mod uses a ~16 MB shared hash pool, constant memory /
        complexity, variable draft lengths. Helps most when the model
        repeats existing text (code refactor, summarisation, reasoning).
        For general chat with low repetition, overhead is ~5 ms.

        Benchmarks from upstream llama.cpp speculative-decoding PRs:
          Scenario                        | Without | With    | Speedup
          gpt-oss-120b code refactor      | 181 t/s | 446 t/s | 2.5x
          Qwen3-235B offloaded            |  12 t/s |  21 t/s | 1.8x
          gpt-oss-120b repeat (92% accept)| 181 t/s | 814 t/s | 4.5x

        Sub-3B dense MTP regresses vs spec-off because the draft head's
        per-token cost exceeds the acceptance savings at this scale.
        Q4_K_XL clean bench (each prompt once after an unrelated warmup)
        on B200 + x86 CPU:
          0.8B GPU: draft-mtp n=2 = 0.58x vs OFF; ngram-only = 1.10x
          2B   GPU: draft-mtp n=2 = 0.82x vs OFF; OFF or ngram = 1.00x
          0.8B CPU: chained n=2   = 0.86x vs OFF; ngram-only = 1.19x
          2B   CPU: chained n=2   = 0.83x vs OFF; ngram-only = 1.01x
          4B+ GPU/CPU: spec on is a net win (1.08x-1.46x).
        Auto falls back to ngram-mod (zero-VRAM, near-zero idle cost on
        diverse content); forced MTP variants engage anyway and just log
        a warning per the user's choice.
        """
        flags: List[str] = []
        # Reset; emit branches re-set on the resolved emission.
        self._spec_draft_n_max = None
        self._speculative_type = None

        # Canonical UI-facing requested mode: auto / mtp / ngram /
        # mtp+ngram / off / ngram-simple. Legacy values are mapped via
        # _canonicalize_spec_mode (default->auto, draft-mtp->mtp,
        # ngram-mod->ngram, "ngram-mod,draft-mtp"->mtp+ngram).
        canonical_mode = _canonicalize_spec_mode(speculative_type)
        is_mtp_model = bool(self._nextn_predict_layers) or (
            _is_mtp_model_name(model_identifier, model_path)
        )
        user_owns_spec_type = _extra_args_set_spec_type(extra_args)
        _mtp_size_b = _extract_model_size_b(model_identifier)
        _mtp_too_small = _mtp_size_b is not None and _mtp_size_b < 3.0

        if user_owns_spec_type:
            # User --spec-type in extra_args wins outright; suppress
            # auto-emit so we don't emit a duplicate / conflicting
            # spec block. Record requested mode as None.
            self._requested_spec_mode = None
            return flags

        effective_mode = canonical_mode or "auto"
        self._requested_spec_mode = effective_mode

        def _resolved_draft_n_max() -> int:
            # User override wins; else platform default (the B200 / x86
            # clean-sweep sweet spot from PR #5582 is n=2 GPU, n=3 CPU;
            # raising past 3 starts to regress on essay-style
            # low-acceptance prompts).
            if spec_draft_n_max is not None:
                n = int(spec_draft_n_max)
                self._spec_draft_n_max = n
                return n
            return 2 if gpus else 3

        def _emit_mtp(*, chain_ngram: bool) -> bool:
            """Append --spec-type mtp[/draft-mtp][,ngram-mod] + n-max."""
            caps = self.probe_server_capabilities(binary)
            mtp_token = caps.get("mtp_token") if caps else None
            if not mtp_token:
                logger.warning(
                    "Requested MTP speculative decoding but "
                    "llama-server lacks --spec-type mtp/draft-mtp; "
                    "run `unsloth studio update`. Loading without "
                    "speculative decoding."
                )
                return False
            draft_n_max = _resolved_draft_n_max()
            n_max_flag = caps.get("spec_draft_n_max_flag") or "--spec-draft-n-max"
            if chain_ngram:
                ngram_knobs = _build_ngram_mod_flags(caps)
                if ngram_knobs:
                    spec_value = f"ngram-mod,{mtp_token}"
                else:
                    logger.warning(
                        "llama-server lacks ngram-mod tuning "
                        "flags; loading MTP only (no ngram chain)"
                    )
                    spec_value = mtp_token
                flags.extend(
                    [
                        "--spec-type",
                        spec_value,
                        n_max_flag,
                        str(draft_n_max),
                    ]
                )
                flags.extend(ngram_knobs)
            else:
                flags.extend(
                    [
                        "--spec-type",
                        mtp_token,
                        n_max_flag,
                        str(draft_n_max),
                    ]
                )
            self._speculative_type = "draft-mtp"
            chain_label = "chained ngram-mod" if chain_ngram else "MTP-only"
            logger.info(f"Spec decoding: {mtp_token} ({chain_label})")
            return True

        def _emit_ngram_mod() -> bool:
            """Append --spec-type ngram-mod + flag-set knobs."""
            ngram_caps = self.probe_server_capabilities(binary)
            ngram_knobs = _build_ngram_mod_flags(ngram_caps)
            flags.extend(["--spec-type", "ngram-mod"])
            if not ngram_knobs:
                logger.warning(
                    "llama-server lacks ngram-mod tuning "
                    "flags; loading without --spec-ngram-mod-* knobs"
                )
            flags.extend(ngram_knobs)
            self._speculative_type = "ngram-mod"
            logger.info("Spec decoding: ngram-mod")
            return True

        if effective_mode == "off":
            return flags  # nothing to emit
        if effective_mode == "ngram-simple":
            flags.extend(["--spec-type", "ngram-simple"])
            self._speculative_type = "ngram-simple"
            return flags
        if effective_mode == "ngram":
            _emit_ngram_mod()
            return flags
        if effective_mode == "mtp":
            if _mtp_too_small:
                logger.warning(
                    f"Forcing MTP on a {_mtp_size_b:.1f}B model; "
                    "the bench shows draft-mtp regresses below 3B. "
                    "Engaging anyway (user override)."
                )
            elif not is_mtp_model:
                logger.warning(
                    "Forcing MTP on a non-MTP GGUF; llama-server may "
                    "fall back to spec-off if no nextn head is present. "
                    "Engaging anyway (user override)."
                )
            _emit_mtp(chain_ngram = False)
            return flags
        if effective_mode == "mtp+ngram":
            if _mtp_too_small:
                logger.warning(
                    f"Forcing MTP+Ngram on a {_mtp_size_b:.1f}B model; "
                    "the bench shows the chain regresses below 3B. "
                    "Engaging anyway (user override)."
                )
            elif not is_mtp_model:
                logger.warning(
                    "Forcing MTP+Ngram on a non-MTP GGUF; llama-server "
                    "may fall back to ngram-only if no nextn head is "
                    "present. Engaging anyway (user override)."
                )
            _emit_mtp(chain_ngram = True)
            return flags

        # effective_mode == "auto": today's promotion path. llama.cpp
        # #22673: MTP is compatible with mmproj, so there's no vision gate.
        if is_mtp_model and not _mtp_too_small:
            # GPU: MTP-only. CPU/Mac: chain ngram-mod + MTP.
            _emit_mtp(chain_ngram = not gpus)
        elif is_mtp_model and _mtp_too_small:
            # Sub-3B fallback: drop the MTP draft head, keep ngram-mod
            # when the binary supports it.
            _small_caps = self.probe_server_capabilities(binary)
            if _small_caps.get("supports_ngram_mod"):
                logger.info(
                    f"MTP GGUF detected but model size {_mtp_size_b:.1f}B "
                    "is below the 3B speedup threshold; using ngram-mod "
                    "only (zero-VRAM, no draft head). Override via "
                    "--spec-type or the Studio Speculative Decoding "
                    "dropdown."
                )
                _emit_ngram_mod()
            else:
                logger.info(
                    f"MTP GGUF detected but model size {_mtp_size_b:.1f}B "
                    "is below the 3B speedup threshold and the bundled "
                    "llama-server does not advertise ngram-mod; "
                    "auto-disabling speculative decoding."
                )
        else:
            # Non-MTP model: let llama-server choose its default strategy.
            flags.append("--spec-default")
            self._speculative_type = "default"
        return flags

    def _already_in_target_state(
        self,
        *,
        model_identifier: str,
        hf_variant: Optional[str],
        n_ctx: int,
        cache_type_kv: Optional[str],
        speculative_type: Optional[str],
        chat_template_override: Optional[str],
        extra_args: Optional[List[str]],
        is_vision: bool,
        gguf_path: Optional[str] = None,
        spec_draft_n_max: Optional[int] = None,
    ) -> bool:
        """True iff the live server already satisfies these load kwargs.

        Mirrors ``routes/inference.py:_request_matches_loaded_settings``
        but compares raw kwargs so ``load_model`` can short-circuit a
        duplicate /load that raced past the route-level check (#5401).
        """
        if not self.is_loaded:
            return False
        if (self._model_identifier or "").lower() != (model_identifier or "").lower():
            return False
        # Direct-file loads pass hf_variant=None while the backend
        # stores an extracted filename label; compare paths instead
        # to keep the guard symmetric.
        if gguf_path is not None and self._gguf_path:
            try:
                if Path(self._gguf_path).resolve() != Path(gguf_path).resolve():
                    return False
            except OSError:
                return False
        elif (self._hf_variant or "").lower() != (hf_variant or "").lower():
            return False
        if self._requested_n_ctx != int(n_ctx):
            return False

        def _norm(value):
            if value is None:
                return None
            if isinstance(value, str):
                stripped = value.strip().lower()
                return stripped or None
            return value

        if _norm(self._cache_type_kv) != _norm(cache_type_kv):
            return False

        # Compare on the canonical UI-facing mode the user requested.
        # When extra_args carries --spec-type, the route-layer code paths
        # bypass the dropdown anyway and the backend stores
        # _requested_spec_mode = None; the request mirrors that by
        # canonicalising to None.
        if _extra_args_set_spec_type(extra_args):
            req_mode = None
        else:
            req_mode = _canonicalize_spec_mode(speculative_type) or "auto"
        backend_mode = self._requested_spec_mode
        if req_mode != backend_mode:
            return False

        # spec_draft_n_max only matters when an MTP variant is actually
        # engaged. Compare on the resolved spec rather than the requested
        # mode so an Auto request that auto-promoted to draft-mtp under
        # the hood still bounces a reload when the user changes n_max.
        if (
            self._speculative_type == "draft-mtp"
            and spec_draft_n_max is not None
            and int(spec_draft_n_max) != (self._spec_draft_n_max or 0)
        ):
            return False

        if (self._chat_template_override or None) != (chat_template_override or None):
            return False

        # extra_args=None means "no opinion" (inherit semantics handled
        # at the route layer); only an explicit list forces equality.
        if extra_args is not None:
            current = list(self._extra_args) if self._extra_args is not None else []
            if list(extra_args) != current:
                return False
        return True

    def _classify_gpu_offload(
        self,
        expected_gpu: bool,
        detected_gpus: list[tuple[int, int]],
    ) -> Optional[bool]:
        """True if a GPU model buffer was allocated, False if only CPU
        buffers landed despite GPU intent, None when there's no signal
        (no GPU detected, no buffer-size lines, etc.)."""
        if not detected_gpus or not expected_gpu:
            return None
        # llama-server logs one ``... model buffer size = N MiB`` line
        # per backend buffer; CUDA0 / ROCm0 / Metal / Vulkan0 /
        # OpenCL0 / SYCL0 are GPU, CPU / CPU_Mapped are not.
        gpu_markers = ("CUDA", "ROCm", "Metal", "Vulkan", "OpenCL", "SYCL")
        saw_buffer_line = False
        saw_gpu_buffer = False
        for line in self._stdout_lines:
            if "model buffer size" not in line:
                continue
            saw_buffer_line = True
            if any(marker in line for marker in gpu_markers):
                saw_gpu_buffer = True
                break
        if not saw_buffer_line:
            return None
        return saw_gpu_buffer

    def unload_model(self) -> bool:
        """Terminate the llama-server subprocess and cancel any in-flight download."""
        self._cancel_event.set()
        with self._lock:
            self._kill_process()
            logger.info(f"Unloaded GGUF model: {self._model_identifier}")
            self._model_identifier = None
            self._gguf_path = None
            self._hf_repo = None
            self._hf_variant = None
            self._is_vision = False
            self._is_audio = False
            self._audio_type = None
            self._audio_probed = False
            self._port = None
            self._healthy = False
            self._context_length = None
            self._effective_context_length = None
            self._max_context_length = None
            self._chat_template = None
            self._chat_template_override = None
            self._supports_reasoning = False
            self._reasoning_always_on = False
            self._reasoning_style = "enable_thinking"
            self._reasoning_default = True
            self._supports_preserve_thinking = False
            self._supports_tools = False
            self._cache_type_kv = None
            self._speculative_type = None
            self._requested_spec_mode = None
            self._spec_draft_n_max = None
            self._n_layers = None
            self._n_kv_heads = None
            self._n_kv_heads_by_layer = None
            self._n_heads = None
            self._embedding_length = None
            self._kv_key_length = None
            self._kv_value_length = None
            self._sliding_window = None
            self._sliding_window_pattern = None
            self._full_attention_interval = None
            self._kv_lora_rank = None
            self._key_length_mla = None
            self._kv_key_length_swa = None
            self._kv_value_length_swa = None
            self._ssm_inner_size = None
            self._ssm_state_size = None
            self._shared_kv_layers = None
            self._nextn_predict_layers = None
            # Clean up temp chat template file
            if hasattr(self, "_chat_template_file") and self._chat_template_file:
                try:
                    import os

                    os.unlink(self._chat_template_file.name)
                except Exception:
                    pass
                self._chat_template_file = None
            # Free audio codec GPU memory
            if LlamaCppBackend._codec_mgr is not None:
                LlamaCppBackend._codec_mgr.unload()
                LlamaCppBackend._codec_mgr = None
                import torch

                if torch.cuda.is_available():
                    torch.cuda.empty_cache()
            return True

    def _kill_process(self):
        """Terminate the subprocess if running."""
        if self._process is None:
            return
        try:
            self._process.terminate()
            self._process.wait(timeout = 5)
        except subprocess.TimeoutExpired:
            logger.warning("llama-server did not exit on SIGTERM, sending SIGKILL")
            self._process.kill()
            self._process.wait(timeout = 5)
        except Exception as e:
            logger.warning(f"Error killing llama-server process: {e}")
        finally:
            self._process = None
            # Clear healthy so a /load arriving during the replacement
            # server's warm-up window cannot short-circuit against the
            # previous server's health (#5401).
            self._healthy = False
            # Drives _wait_for_vram_settle in the next load_model;
            # set in finally so both in-process and frontend
            # /unload+/load Apply paths record the kill.
            self._last_kill_monotonic = time.monotonic()
            if self._stdout_thread is not None:
                self._stdout_thread.join(timeout = 2)
                self._stdout_thread = None
            fh = getattr(self, "_llama_log_fh", None)
            if fh is not None:
                try:
                    fh.close()
                except Exception:
                    pass
                self._llama_log_fh = None

    @staticmethod
    def _kill_orphaned_servers():
        """Kill orphaned llama-server processes started by studio.

        Only kills processes whose resolved binary lives under a known
        Studio install directory (or matches an exact env-var override)
        to avoid terminating unrelated llama-server instances.

        Mirrors every location that _find_llama_server_binary() can
        return from so that orphans from any supported install path
        are still cleaned up.

        Uses psutil for cross-platform support (Linux, macOS, Windows).
        Falls back to pgrep + /proc/<pid>/exe on Linux when psutil is
        not installed.
        """
        import os
        import signal
        import sys

        try:
            # -- Build the ownership allowlist --------------------------------
            # Two kinds of matches:
            #   exact_binaries  -- env var overrides (exact path match only)
            #   install_roots   -- directory trees that are Studio-owned
            #                      (binary must be *under* one of these)
            install_roots: list[Path] = []

            # Env-mode custom root (mirrors _find_llama_server_binary).
            _is_custom_root = False
            try:
                from utils.paths.storage_roots import studio_root as _sr  # noqa: WPS433

                _resolved_sr = _sr()
                _legacy_studio = Path.home() / ".unsloth" / "studio"
                try:
                    _is_custom_root = _resolved_sr.resolve() != _legacy_studio.resolve()
                except (OSError, ValueError):
                    _is_custom_root = _resolved_sr != _legacy_studio
                if _is_custom_root:
                    install_roots.append(_resolved_sr / "llama.cpp")
            except (ImportError, OSError, ValueError):
                pass

            # Primary install dir (default mode only). Env-mode skips this so
            # a custom-root Studio cannot kill a concurrent default-install
            # Studio's llama-server (same OS user, different install).
            if not _is_custom_root:
                install_roots.append(Path.home() / ".unsloth" / "llama.cpp")

            # Legacy in-tree build dirs (older setup.sh versions)
            project_root = Path(__file__).resolve().parents[4]
            install_roots.append(project_root / "llama.cpp")

            # Legacy: extracted binary
            install_roots.append(project_root / "bin")

            # UNSLOTH_LLAMA_CPP_PATH env var (custom install dir)
            custom_dir = os.environ.get("UNSLOTH_LLAMA_CPP_PATH")
            if custom_dir:
                install_roots.append(Path(custom_dir))

            # LLAMA_SERVER_PATH env var (exact binary path)
            exact_binaries: list[Path] = []
            env_binary = os.environ.get("LLAMA_SERVER_PATH")
            if env_binary:
                try:
                    exact_binaries.append(Path(env_binary).resolve())
                except OSError:
                    pass

            # Resolve all roots so is_relative_to works reliably
            resolved_roots: list[Path] = []
            for root in install_roots:
                try:
                    resolved_roots.append(root.resolve())
                except OSError:
                    pass

            my_pid = os.getpid()

            # -- Enumerate processes -------------------------------------------
            # Prefer psutil (cross-platform).  Fall back to pgrep + /proc on
            # Linux when psutil is not installed.
            try:
                import psutil

                has_psutil = True
            except ImportError:
                has_psutil = False

            if has_psutil:
                for proc in psutil.process_iter(["pid", "name", "exe"]):
                    try:
                        if proc.info["pid"] == my_pid:
                            continue

                        name = proc.info.get("name") or ""
                        if not name.lower().startswith("llama-server"):
                            continue

                        exe = proc.info.get("exe")
                        if not exe:
                            continue

                        exe_path = Path(exe).resolve()

                        # Check ownership: exact binary match OR binary is
                        # under a known install root (proper ancestry, not
                        # substring).
                        is_ours = exe_path in exact_binaries or any(
                            exe_path.is_relative_to(root) for root in resolved_roots
                        )
                        if not is_ours:
                            continue

                        proc.kill()
                        logger.info(
                            f"Killed orphaned llama-server process "
                            f"(pid={proc.info['pid']})"
                        )
                    except (
                        psutil.NoSuchProcess,
                        psutil.AccessDenied,
                        psutil.ZombieProcess,
                    ):
                        pass
            else:
                # -- Fallback: pgrep + /proc/<pid>/exe (Linux only) -----------
                if sys.platform != "linux":
                    return
                result = subprocess.run(
                    ["pgrep", "-a", "-f", "llama-server"],
                    capture_output = True,
                    text = True,
                    timeout = 5,
                    env = child_env_without_native_path_secret(),
                )
                if result.returncode != 0:
                    return

                for line in result.stdout.strip().splitlines():
                    parts = line.strip().split(None, 1)
                    if len(parts) < 2:
                        continue
                    pid = int(parts[0])
                    if pid == my_pid:
                        continue

                    # Resolve the actual executable.  /proc/<pid>/exe is a
                    # symlink to the real binary and avoids all cmdline-
                    # parsing ambiguities (spaces in paths, argv rewriting).
                    # Fall back to the first cmdline token when /proc is
                    # unavailable.
                    proc_exe = Path(f"/proc/{pid}/exe")
                    try:
                        binary = proc_exe.resolve(strict = True)
                    except (OSError, ValueError):
                        cmdline = parts[1]
                        token = cmdline.split()[0] if cmdline.strip() else ""
                        if not token:
                            continue
                        binary = Path(token).resolve(strict = False)

                    owned = binary in exact_binaries or any(
                        binary.is_relative_to(root) for root in resolved_roots
                    )
                    if not owned:
                        continue

                    try:
                        os.kill(pid, signal.SIGKILL)
                        logger.info(f"Killed orphaned llama-server process (pid={pid})")
                    except ProcessLookupError:
                        pass
                    except PermissionError:
                        pass
        except Exception:
            logger.warning("Error during orphan server cleanup", exc_info = True)

    def _cleanup(self):
        """atexit handler to ensure llama-server is terminated."""
        self._kill_process()

    def _wait_for_health(self, timeout: float = 120.0, interval: float = 0.5) -> bool:
        """
        Poll llama-server's /health endpoint until it responds 200.

        Also monitors subprocess for early exit/crash.
        """
        deadline = time.monotonic() + timeout
        url = f"http://127.0.0.1:{self._port}/health"

        while time.monotonic() < deadline:
            # Check if process crashed
            if self._process.poll() is not None:
                # Give the drain thread a moment to collect final output
                if self._stdout_thread is not None:
                    self._stdout_thread.join(timeout = 2)
                output = "\n".join(self._stdout_lines[-50:])
                logger.error(
                    f"llama-server exited with code {self._process.returncode}. "
                    f"Output: {output[:2000]}"
                )
                return False

            try:
                resp = httpx.get(url, timeout = 2.0)
                if resp.status_code == 200:
                    return True
            except (
                httpx.ConnectError,
                httpx.TimeoutException,
                # ReadError covers TCP RST mid-read while llama-server is
                # still binding the port (Windows: WinError 10054). The
                # crash-detection branch above catches a real exit; this
                # one keeps a transient socket close from masking it.
                httpx.ReadError,
                httpx.RemoteProtocolError,
                httpx.WriteError,
            ):
                pass

            time.sleep(interval)

        logger.error(f"llama-server health check timed out after {timeout}s")
        return False

    # ── Message building (OpenAI format) ──────────────────────────

    @staticmethod
    def _parse_tool_calls_from_text(content: str) -> list[dict]:
        """Thin wrapper around the shared parser in tool_call_parser
        so safetensors and llama_cpp pick up the same fixes."""
        return _shared_parse_tool_calls_from_text(content)

    @staticmethod
    def _build_openai_messages(
        messages: list[dict],
        image_b64: Optional[str] = None,
    ) -> list[dict]:
        """
        Build OpenAI-format messages, optionally injecting an image_url
        content part into the last user message for vision models.

        If no image is provided, returns messages as-is.
        """
        if not image_b64:
            return messages

        # Find the last user message and convert to multimodal content parts
        result = [msg.copy() for msg in messages]
        last_user_idx = None
        for i, msg in enumerate(result):
            if msg["role"] == "user":
                last_user_idx = i

        if last_user_idx is not None:
            text_content = result[last_user_idx].get("content", "")
            result[last_user_idx]["content"] = [
                {"type": "text", "text": text_content},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/png;base64,{image_b64}",
                    },
                },
            ]

        return result

    # ── Generation (proxy to llama-server) ────────────────────────

    @staticmethod
    def _iter_text_cancellable(
        response: "httpx.Response",
        cancel_event: Optional[threading.Event] = None,
    ) -> Generator[str, None, None]:
        """Iterate over an httpx streaming response with cancel support.

        Checks cancel_event between chunks and on ReadTimeout.  The
        cancel watcher in _stream_with_retry also calls response.close()
        on cancel, which unblocks iter_text() once the response exists.
        During normal streaming llama-server sends tokens frequently,
        so the cancel check between chunks is the primary mechanism.
        """
        text_iter = response.iter_text()
        while True:
            if cancel_event is not None and cancel_event.is_set():
                response.close()
                return
            try:
                chunk = next(text_iter)
                yield chunk
            except StopIteration:
                return
            except httpx.ReadTimeout:
                # No data within the timeout window -- just loop back
                # and re-check cancel_event.
                continue

    @staticmethod
    @contextlib.contextmanager
    def _stream_with_retry(
        client: "httpx.Client",
        url: str,
        payload: dict,
        cancel_event: Optional[threading.Event] = None,
        headers: Optional[dict] = None,
    ):
        """Open an httpx streaming POST with cancel support.

        Sends the request once with a long read timeout (120 s) so
        prompt processing (prefill) can finish without triggering a
        retry storm.  The previous 0.5 s timeout caused duplicate POST
        requests every half second, forcing llama-server to restart
        processing each time.

        A background watcher thread provides cancel by closing the
        response when cancel_event is set.  Limitation: httpx does not
        allow interrupting a blocked read from another thread before
        the response object exists, so cancel during the initial
        header wait (prefill phase) only takes effect once headers
        arrive.  After that, response.close() unblocks reads promptly.
        In practice llama-server prefill is 1-5 s for typical prompts,
        during which cancel is deferred -- still much better than the
        old retry storm which made prefill slower.
        """
        if cancel_event is not None and cancel_event.is_set():
            raise GeneratorExit

        # Background watcher: close the response if cancel is requested.
        # Only effective after response headers arrive (httpx limitation).
        _cancel_closed = threading.Event()
        _response_ref: list = [None]

        def _cancel_watcher():
            while not _cancel_closed.is_set():
                if cancel_event.wait(timeout = 0.3):
                    # Cancel requested. Keep polling until the response object
                    # exists so we can close it, or until the main thread
                    # finishes on its own (_cancel_closed is set in finally).
                    while not _cancel_closed.is_set():
                        r = _response_ref[0]
                        if r is not None:
                            try:
                                r.close()
                                return
                            except Exception as e:
                                logger.debug(
                                    f"Error closing response in cancel watcher: {e}"
                                )
                        # Response not created yet -- wait briefly and retry
                        _cancel_closed.wait(timeout = 0.1)
                    return

        watcher = None
        if cancel_event is not None:
            watcher = threading.Thread(
                target = _cancel_watcher, daemon = True, name = "prefill-cancel"
            )
            watcher.start()

        try:
            # Long read timeout so prefill (prompt processing) can finish
            # without triggering a retry storm.  Cancel during both
            # prefill and streaming is handled by the watcher thread
            # which closes the response, unblocking any httpx read.
            prefill_timeout = httpx.Timeout(
                connect = 30,
                read = 120.0,
                write = 10,
                pool = 10,
            )
            with client.stream(
                "POST",
                url,
                json = payload,
                timeout = prefill_timeout,
                headers = headers,
            ) as response:
                _response_ref[0] = response
                if cancel_event is not None and cancel_event.is_set():
                    raise GeneratorExit
                yield response
                return
        except (httpx.ReadError, httpx.RemoteProtocolError, httpx.CloseError):
            # Response was closed by the cancel watcher
            if cancel_event is not None and cancel_event.is_set():
                raise GeneratorExit
            raise
        finally:
            _cancel_closed.set()

    def generate_chat_completion(
        self,
        messages: list[dict],
        image_b64: Optional[str] = None,
        temperature: float = 0.6,
        top_p: float = 0.95,
        top_k: int = 20,
        min_p: float = 0.01,
        max_tokens: Optional[int] = None,
        repetition_penalty: float = 1.0,
        presence_penalty: float = 0.0,
        stop: Optional[list[str]] = None,
        cancel_event: Optional[threading.Event] = None,
        enable_thinking: Optional[bool] = None,
        reasoning_effort: Optional[str] = None,
        preserve_thinking: Optional[bool] = None,
    ) -> Generator[str | dict, None, None]:
        """
        Send a chat completion request to llama-server and stream tokens back.

        Uses /v1/chat/completions — llama-server handles chat template
        application and vision (multimodal image_url parts) natively.

        Yields cumulative text (matching InferenceBackend's convention).
        """
        if not self.is_loaded:
            raise RuntimeError("llama-server is not loaded")

        openai_messages = self._build_openai_messages(messages, image_b64)

        payload = {
            "messages": openai_messages,
            "stream": True,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k if top_k >= 0 else 0,
            "min_p": min_p,
            "repeat_penalty": repetition_penalty,
            "presence_penalty": presence_penalty,
        }
        # Pass enable_thinking / reasoning_effort / preserve_thinking per-request
        _reasoning_kw = self._request_reasoning_kwargs(
            enable_thinking, reasoning_effort, preserve_thinking
        )
        if _reasoning_kw is not None:
            payload["chat_template_kwargs"] = _reasoning_kw
        # Default cap to the model's effective context length when known,
        # otherwise the conservative floor. The wall-clock backstop below
        # keeps a stuck model from running indefinitely either way.
        payload["max_tokens"] = (
            max_tokens
            if max_tokens is not None
            else (self._effective_context_length or _DEFAULT_MAX_TOKENS_FLOOR)
        )
        payload["t_max_predict_ms"] = _DEFAULT_T_MAX_PREDICT_MS
        if stop:
            payload["stop"] = stop
        payload["stream_options"] = {"include_usage": True}

        url = f"{self.base_url}/v1/chat/completions"
        cumulative = ""
        in_thinking = False
        _stream_done = False
        _metadata_usage = None
        _metadata_timings = None

        try:
            # _stream_with_retry uses a 120 s read timeout so prefill
            # can finish.  Cancel during streaming is handled by the
            # watcher thread (closes the response on cancel_event).
            stream_timeout = httpx.Timeout(connect = 10, read = 0.5, write = 10, pool = 10)
            _auth_headers = (
                {"Authorization": f"Bearer {self._api_key}"} if self._api_key else None
            )
            with httpx.Client(
                timeout = stream_timeout, limits = httpx.Limits(max_keepalive_connections = 0)
            ) as client:
                with self._stream_with_retry(
                    client,
                    url,
                    payload,
                    cancel_event,
                    headers = _auth_headers,
                ) as response:
                    if response.status_code != 200:
                        error_body = response.read().decode()
                        raise RuntimeError(
                            f"llama-server returned {response.status_code}: {error_body}"
                        )

                    buffer = ""
                    has_content_tokens = False
                    reasoning_text = ""
                    for raw_chunk in self._iter_text_cancellable(
                        response, cancel_event
                    ):
                        buffer += raw_chunk
                        while "\n" in buffer:
                            line, buffer = buffer.split("\n", 1)
                            line = line.strip()

                            if not line:
                                continue
                            if line == "data: [DONE]":
                                if in_thinking:
                                    if has_content_tokens:
                                        # Real thinking + content: close the tag
                                        cumulative += "</think>"
                                        yield cumulative
                                    else:
                                        # Only reasoning_content, no content tokens:
                                        # the model put its entire reply in reasoning
                                        # (e.g. Qwen3 always-think mode). Show it
                                        # as the main response, not as a thinking block.
                                        cumulative = reasoning_text
                                        yield cumulative
                                _stream_done = True
                                break  # exit inner while
                            if not line.startswith("data: "):
                                continue

                            try:
                                data = json.loads(line[6:])
                                # Capture server timings/usage from final chunks
                                _chunk_timings = data.get("timings")
                                if _chunk_timings:
                                    _metadata_timings = _chunk_timings
                                _chunk_usage = data.get("usage")
                                if _chunk_usage:
                                    _metadata_usage = _chunk_usage
                                choices = data.get("choices", [])
                                if choices:
                                    delta = choices[0].get("delta", {})

                                    # Handle reasoning/thinking tokens
                                    # llama-server sends these as "reasoning_content"
                                    # Wrap in <think> tags for the frontend parser
                                    reasoning = delta.get("reasoning_content", "")
                                    if reasoning:
                                        reasoning_text += reasoning
                                        if not in_thinking:
                                            cumulative += "<think>"
                                            in_thinking = True
                                        cumulative += reasoning
                                        yield cumulative

                                    token = delta.get("content", "")
                                    if token:
                                        has_content_tokens = True
                                        if in_thinking:
                                            cumulative += "</think>"
                                            in_thinking = False
                                        cumulative += token
                                        yield cumulative
                            except json.JSONDecodeError:
                                logger.debug(
                                    f"Skipping malformed SSE line: {line[:100]}"
                                )
                        if _stream_done:
                            break  # exit outer for
                    if _metadata_usage or _metadata_timings:
                        _metadata_usage = _backfill_usage_from_timings(
                            _metadata_usage, _metadata_timings
                        )
                        yield {
                            "type": "metadata",
                            "usage": _metadata_usage,
                            "timings": _metadata_timings,
                        }

        except httpx.ConnectError:
            raise RuntimeError("Lost connection to llama-server")
        except Exception as e:
            if cancel_event is not None and cancel_event.is_set():
                return
            raise

    # ── Tool-calling agentic loop ──────────────────────────────

    def generate_chat_completion_with_tools(
        self,
        messages: list[dict],
        tools: list[dict],
        temperature: float = 0.6,
        top_p: float = 0.95,
        top_k: int = 20,
        min_p: float = 0.01,
        max_tokens: Optional[int] = None,
        repetition_penalty: float = 1.0,
        presence_penalty: float = 0.0,
        stop: Optional[list[str]] = None,
        cancel_event: Optional[threading.Event] = None,
        enable_thinking: Optional[bool] = None,
        reasoning_effort: Optional[str] = None,
        preserve_thinking: Optional[bool] = None,
        max_tool_iterations: int = 25,
        auto_heal_tool_calls: bool = True,
        tool_call_timeout: int = 300,
        session_id: Optional[str] = None,
    ) -> Generator[dict, None, None]:
        """
        Agentic loop: let the model call tools, execute them, and continue.

        Yields dicts with:
          {"type": "status", "text": "Searching: ..."/"Reading: ..."}   -- tool status updates
          {"type": "content", "text": "token"}            -- streamed content tokens (cumulative)
          {"type": "reasoning", "text": "token"}          -- streamed reasoning tokens (cumulative)
        """
        from core.inference.tools import execute_tool

        if not self.is_loaded:
            raise RuntimeError("llama-server is not loaded")

        conversation = list(messages)
        url = f"{self.base_url}/v1/chat/completions"
        _accumulated_completion_tokens = 0
        _accumulated_predicted_ms = 0.0
        _accumulated_predicted_n = 0

        def _strip_tool_markup(text: str, *, final: bool = False) -> str:
            if not auto_heal_tool_calls:
                return text
            return strip_tool_call_markup(text, final = final)

        # XML prefixes that signal a tool call in content.
        # Empty when auto_heal is disabled so the buffer never
        # speculatively holds content for XML detection.
        _TOOL_XML_SIGNALS = (
            ("<tool_call>", "<function=") if auto_heal_tool_calls else ()
        )
        _MAX_BUFFER_CHARS = 32

        # ── Duplicate tool-call detection ────────────────────────
        # Track recent (tool_name, arguments) hashes to detect loops
        # where the model repeats the exact same call.  Retries after
        # a transient failure are allowed (only block when the previous
        # identical call succeeded).
        _tool_call_history: list[tuple[str, bool]] = []  # (key, failed)

        # ── Re-prompt on plan-without-action ─────────────────
        # When the model describes what it intends to do (forward-looking
        # language) without actually calling a tool, re-prompt once.
        # Only triggers on responses that signal intent/planning -- a
        # direct answer like "4" or "Hello!" will not match.
        # Pattern is compiled once at module level (_INTENT_SIGNAL).
        _reprompt_count = 0

        # Reserve extra iterations for re-prompts so they don't
        # consume the caller's tool-call budget.  Only add the
        # extra slot when tool iterations are actually allowed.
        _extra = _MAX_REPROMPTS if max_tool_iterations > 0 else 0
        for iteration in range(max_tool_iterations + _extra):
            if cancel_event is not None and cancel_event.is_set():
                return

            # Build payload -- stream: True so we detect tool signals
            # in the first 1-2 chunks without a non-streaming penalty.
            payload = {
                "messages": conversation,
                "stream": True,
                "stream_options": {"include_usage": True},
                "temperature": temperature,
                "top_p": top_p,
                "top_k": top_k if top_k >= 0 else 0,
                "min_p": min_p,
                "repeat_penalty": repetition_penalty,
                "presence_penalty": presence_penalty,
                "tools": tools,
                "tool_choice": "auto",
            }
            _reasoning_kw = self._request_reasoning_kwargs(
                enable_thinking, reasoning_effort, preserve_thinking
            )
            if _reasoning_kw is not None:
                payload["chat_template_kwargs"] = _reasoning_kw
            payload["max_tokens"] = (
                max_tokens
                if max_tokens is not None
                else (self._effective_context_length or _DEFAULT_MAX_TOKENS_FLOOR)
            )
            payload["t_max_predict_ms"] = _DEFAULT_T_MAX_PREDICT_MS
            if stop:
                payload["stop"] = stop

            try:
                _auth_headers = (
                    {"Authorization": f"Bearer {self._api_key}"}
                    if self._api_key
                    else None
                )

                # ── Speculative buffer state machine ──────────────────
                # BUFFERING: accumulating content, checking for tool signals
                # STREAMING: no tool detected, yielding tokens to caller
                # DRAINING:  tool signal found, silently consuming rest
                _S_BUFFERING = 0
                _S_STREAMING = 1
                _S_DRAINING = 2

                detect_state = _S_BUFFERING
                content_buffer = ""  # Raw content held during BUFFERING
                content_accum = ""  # All content tokens (for tool parsing)
                reasoning_accum = ""
                cumulative_display = ""  # Cumulative text yielded (with <think>)
                in_thinking = False
                has_content_tokens = False
                tool_calls_acc = {}  # Structured delta.tool_calls fragments
                has_structured_tc = False
                _iter_usage = None
                _iter_timings = None
                _stream_done = False
                _last_emitted = ""

                stream_timeout = httpx.Timeout(
                    connect = 10,
                    read = 0.5,
                    write = 10,
                    pool = 10,
                )
                with httpx.Client(
                    timeout = stream_timeout,
                    limits = httpx.Limits(max_keepalive_connections = 0),
                ) as client:
                    with self._stream_with_retry(
                        client,
                        url,
                        payload,
                        cancel_event,
                        headers = _auth_headers,
                    ) as response:
                        if response.status_code != 200:
                            error_body = response.read().decode()
                            raise RuntimeError(
                                f"llama-server returned {response.status_code}: "
                                f"{error_body}"
                            )

                        raw_buf = ""
                        for raw_chunk in self._iter_text_cancellable(
                            response,
                            cancel_event,
                        ):
                            raw_buf += raw_chunk
                            while "\n" in raw_buf:
                                line, raw_buf = raw_buf.split("\n", 1)
                                line = line.strip()

                                if not line:
                                    continue
                                if line == "data: [DONE]":
                                    # Flush thinking state for STREAMING
                                    if detect_state == _S_STREAMING and in_thinking:
                                        if has_content_tokens:
                                            cumulative_display += "</think>"
                                            yield {
                                                "type": "content",
                                                "text": _strip_tool_markup(
                                                    cumulative_display,
                                                    final = True,
                                                ),
                                            }
                                        else:
                                            cumulative_display = reasoning_accum
                                            yield {
                                                "type": "content",
                                                "text": cumulative_display,
                                            }
                                    _stream_done = True
                                    break  # exit inner while
                                if not line.startswith("data: "):
                                    continue

                                try:
                                    chunk_data = json.loads(line[6:])
                                    _ct = chunk_data.get("timings")
                                    if _ct:
                                        _iter_timings = _ct
                                    _cu = chunk_data.get("usage")
                                    if _cu:
                                        _iter_usage = _cu

                                    choices = chunk_data.get("choices", [])
                                    if not choices:
                                        continue

                                    delta = choices[0].get("delta", {})

                                    # ── Structured tool_calls ──
                                    tc_deltas = delta.get("tool_calls")
                                    if tc_deltas:
                                        # Once visible content has been
                                        # emitted, do not reclassify this
                                        # turn as a tool call.
                                        if _last_emitted:
                                            continue
                                        has_structured_tc = True
                                        detect_state = _S_DRAINING
                                        for tc_d in tc_deltas:
                                            idx = tc_d.get("index", 0)
                                            if idx not in tool_calls_acc:
                                                tool_calls_acc[idx] = {
                                                    "id": tc_d.get("id", f"call_{idx}"),
                                                    "type": "function",
                                                    "function": {
                                                        "name": "",
                                                        "arguments": "",
                                                    },
                                                }
                                            elif tc_d.get("id"):
                                                # Update ID if real one
                                                # arrives on a later delta
                                                tool_calls_acc[idx]["id"] = tc_d["id"]
                                            func = tc_d.get("function", {})
                                            if func.get("name"):
                                                tool_calls_acc[idx]["function"][
                                                    "name"
                                                ] += func["name"]
                                            if func.get("arguments"):
                                                tool_calls_acc[idx]["function"][
                                                    "arguments"
                                                ] += func["arguments"]
                                        continue

                                    # ── Reasoning tokens ──
                                    # Only yield in STREAMING state. In BUFFERING
                                    # and DRAINING, accumulate silently so we don't
                                    # corrupt the consumer's prev_text tracker
                                    # (routes/inference.py never resets prev_text
                                    # between tool iterations).
                                    reasoning = delta.get("reasoning_content", "")
                                    if reasoning:
                                        reasoning_accum += reasoning
                                        if detect_state == _S_STREAMING:
                                            if not in_thinking:
                                                cumulative_display += "<think>"
                                                in_thinking = True
                                            cumulative_display += reasoning
                                            yield {
                                                "type": "content",
                                                "text": cumulative_display,
                                            }

                                    # ── Content tokens ──
                                    token = delta.get("content", "")
                                    if token:
                                        has_content_tokens = True
                                        content_accum += token

                                        if detect_state == _S_DRAINING:
                                            pass  # accumulate silently

                                        elif detect_state == _S_STREAMING:
                                            if in_thinking:
                                                cumulative_display += "</think>"
                                                in_thinking = False
                                            cumulative_display += token
                                            cleaned = _strip_tool_markup(
                                                cumulative_display,
                                            )
                                            if len(cleaned) > len(_last_emitted):
                                                _last_emitted = cleaned
                                                yield {
                                                    "type": "content",
                                                    "text": cleaned,
                                                }

                                        elif detect_state == _S_BUFFERING:
                                            content_buffer += token
                                            stripped_buf = content_buffer.lstrip()
                                            if not stripped_buf:
                                                continue

                                            # Check tool signal prefixes
                                            is_prefix = False
                                            is_match = False
                                            for sig in _TOOL_XML_SIGNALS:
                                                if stripped_buf.startswith(sig):
                                                    is_match = True
                                                    break
                                                if sig.startswith(stripped_buf):
                                                    is_prefix = True
                                                    break

                                            if is_match:
                                                detect_state = _S_DRAINING
                                            elif (
                                                is_prefix
                                                and len(stripped_buf)
                                                < _MAX_BUFFER_CHARS
                                            ):
                                                pass  # keep buffering
                                            else:
                                                # Not a tool -- flush buffer
                                                detect_state = _S_STREAMING
                                                # Flush any reasoning accumulated
                                                # during BUFFERING phase
                                                if reasoning_accum:
                                                    cumulative_display += "<think>"
                                                    cumulative_display += (
                                                        reasoning_accum
                                                    )
                                                    cumulative_display += "</think>"
                                                cumulative_display += content_buffer
                                                cleaned = _strip_tool_markup(
                                                    cumulative_display,
                                                )
                                                if len(cleaned) > len(_last_emitted):
                                                    _last_emitted = cleaned
                                                    yield {
                                                        "type": "content",
                                                        "text": cleaned,
                                                    }

                                except json.JSONDecodeError:
                                    logger.debug(
                                        f"Skipping malformed SSE line: {line[:100]}"
                                    )
                            if _stream_done:
                                break  # exit outer for

                # ── Resolve BUFFERING at stream end ──
                if detect_state == _S_BUFFERING:
                    stripped_buf = content_buffer.lstrip()
                    if (
                        stripped_buf
                        and auto_heal_tool_calls
                        and any(s in stripped_buf for s in _TOOL_XML_SIGNALS)
                    ):
                        detect_state = _S_DRAINING
                    elif content_accum or reasoning_accum:
                        detect_state = _S_STREAMING
                        if content_buffer:
                            # Flush any reasoning accumulated first
                            if reasoning_accum:
                                cumulative_display += "<think>"
                                cumulative_display += reasoning_accum
                                cumulative_display += "</think>"
                            cumulative_display += content_buffer
                            yield {
                                "type": "content",
                                "text": _strip_tool_markup(
                                    cumulative_display,
                                    final = True,
                                ),
                            }
                        elif reasoning_accum and not has_content_tokens:
                            # Reasoning-only response (no content tokens):
                            # show reasoning as plain text, matching
                            # the final streaming pass behavior for
                            # models that put everything in reasoning.
                            cumulative_display = reasoning_accum
                            yield {
                                "type": "content",
                                "text": cumulative_display,
                            }
                    else:
                        return

                # ── STREAMING path: no tool call ──
                if detect_state == _S_STREAMING:
                    # Safety net: check for XML tool signals in content.
                    # The route layer resets prev_text on tool_start, so
                    # post-tool synthesis streams correctly even if
                    # content was already emitted before the tool XML.
                    _safety_tc = None
                    if auto_heal_tool_calls and any(
                        s in content_accum for s in _TOOL_XML_SIGNALS
                    ):
                        _safety_tc = self._parse_tool_calls_from_text(
                            content_accum,
                        )
                    if not _safety_tc:
                        # ── Re-prompt on plan-without-action ──
                        # If the model described what it intends to do
                        # (forward-looking language) without calling any
                        # tool, nudge it to act.  Only fires once per
                        # request and only on short responses that
                        # contain intent signals -- a direct answer
                        # like "4" or "Hello!" won't trigger this.
                        # Use content if available, otherwise fall back
                        # to reasoning text (reasoning-only stalls).
                        _stripped = content_accum.strip()
                        if not _stripped:
                            _stripped = reasoning_accum.strip()
                        if (
                            tools
                            and _reprompt_count < _MAX_REPROMPTS
                            and 0 < len(_stripped) < _REPROMPT_MAX_CHARS
                            and _INTENT_SIGNAL.search(_stripped)
                        ):
                            _reprompt_count += 1
                            logger.info(
                                f"Re-prompt {_reprompt_count}/{_MAX_REPROMPTS}: "
                                f"model responded without calling tools "
                                f"({len(_stripped)} chars)"
                            )
                            conversation.append(
                                {
                                    "role": "assistant",
                                    "content": _stripped,
                                }
                            )
                            conversation.append(
                                {
                                    "role": "user",
                                    "content": (
                                        "STOP. Do NOT write code or explain. "
                                        "You MUST call a tool NOW. "
                                        "Call web_search or python immediately."
                                    ),
                                }
                            )
                            # Accumulate tokens and timing from this iteration
                            _fu_r = (
                                _backfill_usage_from_timings(_iter_usage, _iter_timings)
                                or {}
                            )
                            _accumulated_completion_tokens += _fu_r.get(
                                "completion_tokens", 0
                            )
                            _it_r = _iter_timings or {}
                            _accumulated_predicted_ms += _it_r.get("predicted_ms", 0)
                            _accumulated_predicted_n += _it_r.get("predicted_n", 0)
                            yield {"type": "status", "text": ""}
                            continue

                        # Content was already streamed.  Yield metadata.
                        yield {"type": "status", "text": ""}
                        _fu = (
                            _backfill_usage_from_timings(_iter_usage, _iter_timings)
                            or {}
                        )
                        _fc = _fu.get("completion_tokens", 0)
                        _fp = _fu.get("prompt_tokens", 0)
                        _tc = _fc + _accumulated_completion_tokens
                        if (
                            _iter_usage
                            or _iter_timings
                            or _accumulated_completion_tokens
                        ):
                            _mt = dict(_iter_timings) if _iter_timings else {}
                            if _accumulated_predicted_ms or _accumulated_predicted_n:
                                _mt["predicted_ms"] = (
                                    _mt.get("predicted_ms", 0)
                                    + _accumulated_predicted_ms
                                )
                                _tn = (
                                    _mt.get("predicted_n", 0) + _accumulated_predicted_n
                                )
                                _mt["predicted_n"] = _tn
                                _tms = _mt["predicted_ms"]
                                if _tms > 0:
                                    _mt["predicted_per_second"] = _tn / (_tms / 1000.0)
                            yield {
                                "type": "metadata",
                                "usage": {
                                    "prompt_tokens": _fp,
                                    "completion_tokens": _tc,
                                    "total_tokens": _fp + _tc,
                                },
                                "timings": _mt,
                            }
                        return

                    # Safety net caught tool XML -- treat as tool call
                    tool_calls = _safety_tc
                    content_text = _strip_tool_markup(
                        content_accum,
                        final = True,
                    )
                    logger.info(
                        f"Safety net: parsed {len(tool_calls)} tool call(s) "
                        f"from streamed content"
                    )
                else:
                    # ── DRAINING path: assemble tool_calls ──
                    tool_calls = None
                    content_text = content_accum
                    if has_structured_tc:
                        # Filter out incomplete fragments (e.g. from
                        # truncation by max_tokens or disconnect).
                        tool_calls = [
                            tool_calls_acc[i]
                            for i in sorted(tool_calls_acc)
                            if (
                                tool_calls_acc[i]
                                .get("function", {})
                                .get("name", "")
                                .strip()
                            )
                        ] or None
                    if (
                        not tool_calls
                        and auto_heal_tool_calls
                        and any(s in content_accum for s in _TOOL_XML_SIGNALS)
                    ):
                        tool_calls = self._parse_tool_calls_from_text(
                            content_accum,
                        )
                    if tool_calls and not has_structured_tc:
                        content_text = _strip_tool_markup(
                            content_text,
                            final = True,
                        )
                    if tool_calls:
                        logger.info(
                            f"Parsed {len(tool_calls)} tool call(s) from "
                            f"{'structured delta' if has_structured_tc else 'content text'}"
                        )
                    if not tool_calls:
                        # DRAINING but no tool calls (false positive).
                        # Merge accumulated metrics from prior tool
                        # iterations so they are not silently dropped.
                        yield {"type": "status", "text": ""}
                        if content_accum:
                            # Strip leaked tool-call XML before yielding
                            content_accum = _strip_tool_markup(
                                content_accum, final = True
                            )
                        if content_accum:
                            yield {"type": "content", "text": content_accum}
                        _fu = (
                            _backfill_usage_from_timings(_iter_usage, _iter_timings)
                            or {}
                        )
                        _fc = _fu.get("completion_tokens", 0)
                        _fp = _fu.get("prompt_tokens", 0)
                        _tc = _fc + _accumulated_completion_tokens
                        if (
                            _iter_usage
                            or _iter_timings
                            or _accumulated_completion_tokens
                        ):
                            _mt = dict(_iter_timings) if _iter_timings else {}
                            if _accumulated_predicted_ms or _accumulated_predicted_n:
                                _mt["predicted_ms"] = (
                                    _mt.get("predicted_ms", 0)
                                    + _accumulated_predicted_ms
                                )
                                _tn = (
                                    _mt.get("predicted_n", 0) + _accumulated_predicted_n
                                )
                                _mt["predicted_n"] = _tn
                                _tms = _mt["predicted_ms"]
                                if _tms > 0:
                                    _mt["predicted_per_second"] = _tn / (_tms / 1000.0)
                            yield {
                                "type": "metadata",
                                "usage": {
                                    "prompt_tokens": _fp,
                                    "completion_tokens": _tc,
                                    "total_tokens": _fp + _tc,
                                },
                                "timings": _mt,
                            }
                        return

                # ── Execute tool calls ──
                _accumulated_completion_tokens += (
                    _backfill_usage_from_timings(_iter_usage, _iter_timings) or {}
                ).get("completion_tokens", 0)
                _it = _iter_timings or {}
                _accumulated_predicted_ms += _it.get("predicted_ms", 0)
                _accumulated_predicted_n += _it.get("predicted_n", 0)

                assistant_msg = {"role": "assistant", "content": content_text}
                if tool_calls:
                    assistant_msg["tool_calls"] = tool_calls
                conversation.append(assistant_msg)

                for tc in tool_calls or []:
                    func = tc.get("function", {})
                    tool_name = func.get("name", "")
                    raw_args = func.get("arguments", {})

                    if isinstance(raw_args, str):
                        try:
                            arguments = json.loads(raw_args)
                        except (json.JSONDecodeError, ValueError):
                            if auto_heal_tool_calls:
                                arguments = {"query": raw_args}
                            else:
                                arguments = {"raw": raw_args}
                    else:
                        arguments = raw_args

                    if tool_name == "web_search":
                        _ws_url = (arguments.get("url") or "").strip()
                        if _ws_url:
                            _parsed = urlparse(_ws_url)
                            if _parsed.scheme in ("http", "https") and _parsed.hostname:
                                _ws_host = _parsed.hostname
                                if _ws_host.startswith("www."):
                                    _ws_host = _ws_host[4:]
                                status_text = f"Reading: {_ws_host}"
                            else:
                                status_text = "Reading page..."
                        else:
                            status_text = f"Searching: {arguments.get('query', '')}"
                    elif tool_name == "python":
                        preview = (
                            (arguments.get("code") or "").strip().split("\n")[0][:60]
                        )
                        status_text = (
                            f"Running Python: {preview}"
                            if preview
                            else "Running Python..."
                        )
                    elif tool_name == "terminal":
                        cmd_preview = (arguments.get("command") or "")[:60]
                        status_text = (
                            f"Running: {cmd_preview}"
                            if cmd_preview
                            else "Running command..."
                        )
                    else:
                        status_text = f"Calling: {tool_name}"
                    yield {"type": "status", "text": status_text}

                    yield {
                        "type": "tool_start",
                        "tool_name": tool_name,
                        "tool_call_id": tc.get("id", ""),
                        "arguments": arguments,
                    }

                    # ── Duplicate call detection ──────────────
                    # str(dict) is stable here: arguments always comes from
                    # json.loads on the same model output within one request,
                    # so insertion order is deterministic (Python 3.7+).
                    _tc_key = tool_name + str(arguments)
                    _prev = _tool_call_history[-1] if _tool_call_history else None
                    if _prev and _prev[0] == _tc_key and not _prev[1]:
                        result = (
                            "You already made this exact call. "
                            "Do not repeat the same tool call. "
                            "Try a different approach: fetch a URL "
                            "from previous results, use Python to "
                            "process data you already have, or "
                            "provide your final answer now."
                        )
                    else:
                        _effective_timeout = (
                            None if tool_call_timeout >= 9999 else tool_call_timeout
                        )
                        # Guard against the model emitting a tool not in the
                        # per-request advertised set: filtered MCP names, a
                        # built-in the caller opted out of, or a stale name
                        # from a prior turn. Mirrors the safetensors loop's
                        # allowed_tool_names check.
                        _allowed = {
                            (t.get("function") or {}).get("name")
                            for t in (tools or [])
                            if (t.get("function") or {}).get("name")
                        }
                        if _allowed and tool_name not in _allowed:
                            result = (
                                f"Error: tool '{tool_name}' is not enabled "
                                "for this request. Use one of the enabled "
                                "tools or provide a final answer."
                            )
                        else:
                            result = execute_tool(
                                tool_name,
                                arguments,
                                cancel_event = cancel_event,
                                timeout = _effective_timeout,
                                session_id = session_id,
                            )

                    yield {
                        "type": "tool_end",
                        "tool_name": tool_name,
                        "tool_call_id": tc.get("id", ""),
                        "result": result,
                    }

                    # Nudge model to try a different approach on errors
                    _error_prefixes = (
                        "Error",
                        "Search failed",
                        "Execution error",
                        "Blocked:",
                        "Exit code",
                        "Failed to fetch",
                        "Failed to resolve",
                        "No query provided",
                    )
                    _is_error = isinstance(result, str) and result.lstrip().startswith(
                        _error_prefixes
                    )
                    _tool_call_history.append((_tc_key, _is_error))
                    # Strip image sentinel before feeding result to the LLM
                    # (the full result with sentinel is still yielded via
                    # tool_end so the frontend can extract image paths).
                    _result_content = result
                    if "\n__IMAGES__:" in _result_content:
                        _result_content = _result_content.rsplit("\n__IMAGES__:", 1)[0]
                    if _is_error:
                        _result_content = (
                            _result_content + "\n\nThe tool call encountered an issue. "
                            "Please try a different approach or rephrase your request."
                        )

                    tool_msg = {
                        "role": "tool",
                        "name": tool_name,
                        "content": _result_content,
                    }
                    tool_call_id = tc.get("id")
                    if tool_call_id:
                        tool_msg["tool_call_id"] = tool_call_id
                    conversation.append(tool_msg)

                # Clear tool status badge before next generation iteration
                yield {"type": "status", "text": ""}
                # Continue the loop to let model respond with context
                continue

            except httpx.ConnectError:
                raise RuntimeError("Lost connection to llama-server")
            except Exception as e:
                if cancel_event is not None and cancel_event.is_set():
                    return
                raise

        # ── Tool iteration cap reached -- synthesize final answer ──
        # The model used all iterations without producing a final text
        # response. Inject a nudge so the final streaming pass produces
        # a useful answer instead of continuing to request tools.
        if max_tool_iterations > 0:
            conversation.append(
                {
                    "role": "user",
                    "content": (
                        "You have used all available tool calls. Based on "
                        "everything you have found so far, provide your final "
                        "answer now. Do not call any more tools."
                    ),
                }
            )

        # Clear status
        yield {"type": "status", "text": ""}

        # Final streaming pass with the full conversation context
        stream_payload = {
            "messages": conversation,
            "stream": True,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k if top_k >= 0 else 0,
            "min_p": min_p,
            "repeat_penalty": repetition_penalty,
            "presence_penalty": presence_penalty,
        }
        _reasoning_kw = self._request_reasoning_kwargs(
            enable_thinking, reasoning_effort, preserve_thinking
        )
        if _reasoning_kw is not None:
            stream_payload["chat_template_kwargs"] = _reasoning_kw
        stream_payload["max_tokens"] = (
            max_tokens
            if max_tokens is not None
            else (self._effective_context_length or _DEFAULT_MAX_TOKENS_FLOOR)
        )
        stream_payload["t_max_predict_ms"] = _DEFAULT_T_MAX_PREDICT_MS
        if stop:
            stream_payload["stop"] = stop
        stream_payload["stream_options"] = {"include_usage": True}

        cumulative = ""
        _last_emitted = ""
        in_thinking = False
        has_content_tokens = False
        reasoning_text = ""
        _metadata_usage = None
        _metadata_timings = None
        _stream_done = False

        try:
            stream_timeout = httpx.Timeout(connect = 10, read = 0.5, write = 10, pool = 10)
            _auth_headers = (
                {"Authorization": f"Bearer {self._api_key}"} if self._api_key else None
            )
            with httpx.Client(
                timeout = stream_timeout, limits = httpx.Limits(max_keepalive_connections = 0)
            ) as client:
                with self._stream_with_retry(
                    client,
                    url,
                    stream_payload,
                    cancel_event,
                    headers = _auth_headers,
                ) as response:
                    if response.status_code != 200:
                        error_body = response.read().decode()
                        raise RuntimeError(
                            f"llama-server returned {response.status_code}: {error_body}"
                        )

                    buffer = ""
                    for raw_chunk in self._iter_text_cancellable(
                        response, cancel_event
                    ):
                        buffer += raw_chunk
                        while "\n" in buffer:
                            line, buffer = buffer.split("\n", 1)
                            line = line.strip()

                            if not line:
                                continue
                            if line == "data: [DONE]":
                                if in_thinking:
                                    if has_content_tokens:
                                        cumulative += "</think>"
                                        yield {
                                            "type": "content",
                                            "text": _strip_tool_markup(
                                                cumulative, final = True
                                            ),
                                        }
                                    else:
                                        cumulative = reasoning_text
                                        yield {"type": "content", "text": cumulative}
                                _stream_done = True
                                break  # exit inner while
                            if not line.startswith("data: "):
                                continue

                            try:
                                chunk_data = json.loads(line[6:])
                                # Capture server timings/usage from final chunks
                                _chunk_timings = chunk_data.get("timings")
                                if _chunk_timings:
                                    _metadata_timings = _chunk_timings
                                _chunk_usage = chunk_data.get("usage")
                                if _chunk_usage:
                                    _metadata_usage = _chunk_usage
                                choices = chunk_data.get("choices", [])
                                if choices:
                                    delta = choices[0].get("delta", {})

                                    reasoning = delta.get("reasoning_content", "")
                                    if reasoning:
                                        reasoning_text += reasoning
                                        if not in_thinking:
                                            cumulative += "<think>"
                                            in_thinking = True
                                        cumulative += reasoning
                                        yield {"type": "content", "text": cumulative}

                                    token = delta.get("content", "")
                                    if token:
                                        has_content_tokens = True
                                        if in_thinking:
                                            cumulative += "</think>"
                                            in_thinking = False
                                        cumulative += token
                                        cleaned = _strip_tool_markup(cumulative)
                                        # Only emit when cleaned text grows (monotonic).
                                        if len(cleaned) > len(_last_emitted):
                                            _last_emitted = cleaned
                                            yield {"type": "content", "text": cleaned}
                            except json.JSONDecodeError:
                                logger.debug(
                                    f"Skipping malformed SSE line: {line[:100]}"
                                )
                        if _stream_done:
                            break  # exit outer for
                    _final_usage = _metadata_usage or {}
                    _final_completion = _final_usage.get("completion_tokens", 0)
                    _final_prompt = _final_usage.get("prompt_tokens", 0)
                    _total_completion = (
                        _final_completion + _accumulated_completion_tokens
                    )
                    if _metadata_usage or _metadata_timings:
                        _merged_timings = (
                            dict(_metadata_timings) if _metadata_timings else {}
                        )
                        if _accumulated_predicted_ms or _accumulated_predicted_n:
                            _merged_timings["predicted_ms"] = (
                                _merged_timings.get("predicted_ms", 0)
                                + _accumulated_predicted_ms
                            )
                            _total_predicted_n = (
                                _merged_timings.get("predicted_n", 0)
                                + _accumulated_predicted_n
                            )
                            _merged_timings["predicted_n"] = _total_predicted_n
                            _total_predicted_ms = _merged_timings["predicted_ms"]
                            if _total_predicted_ms > 0:
                                _merged_timings["predicted_per_second"] = (
                                    _total_predicted_n / (_total_predicted_ms / 1000.0)
                                )
                        yield {
                            "type": "metadata",
                            "usage": {
                                "prompt_tokens": _final_prompt,
                                "completion_tokens": _total_completion,
                                "total_tokens": _final_prompt + _total_completion,
                            },
                            "timings": _merged_timings,
                        }

        except httpx.ConnectError:
            raise RuntimeError("Lost connection to llama-server")
        except Exception as e:
            if cancel_event is not None and cancel_event.is_set():
                return
            raise

    # ── TTS support ────────────────────────────────────────────

    def detect_audio_type(self) -> Optional[str]:
        """Detect audio/TTS codec; swallows errors (use _strict variant to distinguish)."""
        try:
            return self._detect_audio_type_strict()
        except Exception as e:
            logger.debug(f"Audio type detection failed: {e}")
            return None

    def _detect_audio_type_strict(self) -> Optional[str]:
        """Codec name on match, None on definitive non-audio, raises on transport/JSON errors."""
        if not self.is_loaded:
            return None
        _auth_headers = (
            {"Authorization": f"Bearer {self._api_key}"} if self._api_key else None
        )
        with httpx.Client(timeout = 10, headers = _auth_headers) as client:

            def _detok(tid: int) -> str:
                # Non-200 means "marker not in vocab" -- keep probing.
                # Transport / JSON errors still raise.
                r = client.post(f"{self.base_url}/detokenize", json = {"tokens": [tid]})
                if r.status_code != 200:
                    return ""
                return r.json().get("content", "")

            def _tok(text: str) -> list[int]:
                r = client.post(
                    f"{self.base_url}/tokenize",
                    json = {"content": text, "add_special": False},
                )
                if r.status_code != 200:
                    return []
                return r.json().get("tokens", [])

            # Check codec-specific tokens (not generic ones that may exist in non-audio models)
            if "<custom_token_" in _detok(128258) and "<custom_token_" in _detok(
                128259
            ):
                return "snac"
            if len(_tok("<|AUDIO|>")) == 1 and len(_tok("<|audio_eos|>")) == 1:
                return "csm"
            if len(_tok("<|startoftranscript|>")) == 1:
                return "whisper"
            if len(_tok("<audio_soft_token>")) == 1:
                return "audio_vlm"
            if (
                len(_tok("<|bicodec_semantic_0|>")) == 1
                and len(_tok("<|bicodec_global_0|>")) == 1
            ):
                return "bicodec"
            if len(_tok("<|c1_0|>")) == 1 and len(_tok("<|c2_0|>")) == 1:
                return "dac"
        return None

    # Prompt format per codec: (template, stop_tokens, needs_token_ids)
    # Matches prompts in InferenceBackend._generate_snac/bicodec/dac
    _TTS_PROMPTS = {
        "snac": (
            "<custom_token_3>{text}<|eot_id|><custom_token_4>",
            ["<custom_token_2>"],
            True,
        ),
        "bicodec": (
            "<|task_tts|><|start_content|>{text}<|end_content|><|start_global_token|>",
            ["<|im_end|>", "</s>"],
            False,
        ),
        "dac": (
            "<|im_start|>\n<|text_start|>{text}<|text_end|>\n<|audio_start|><|global_features_start|>\n",
            ["<|im_end|>", "<|audio_end|>"],
            False,
        ),
    }

    _codec_mgr = None  # Shared AudioCodecManager instance

    def init_audio_codec(self, audio_type: str) -> None:
        """Load the audio codec at model load time (mirrors non-GGUF path)."""
        import torch
        from core.inference.audio_codecs import AudioCodecManager

        if LlamaCppBackend._codec_mgr is None:
            LlamaCppBackend._codec_mgr = AudioCodecManager()

        device = "cuda" if torch.cuda.is_available() else "cpu"
        model_repo_path = None

        # BiCodec needs a repo with BiCodec/ weights — download canonical SparkTTS
        if audio_type == "bicodec":
            from huggingface_hub import snapshot_download
            import os

            repo_path = snapshot_download(
                "unsloth/Spark-TTS-0.5B", local_dir = "Spark-TTS-0.5B"
            )
            model_repo_path = os.path.abspath(repo_path)

        LlamaCppBackend._codec_mgr.load_codec(
            audio_type, device, model_repo_path = model_repo_path
        )
        logger.info(f"Loaded audio codec for GGUF TTS: {audio_type}")

    def generate_audio_response(
        self,
        text: str,
        audio_type: str,
        temperature: float = 0.6,
        top_p: float = 0.95,
        top_k: int = 50,
        min_p: float = 0.0,
        max_new_tokens: int = 2048,
        repetition_penalty: float = 1.1,
    ) -> tuple:
        """
        Generate TTS audio via llama-server /completion + codec decoding.
        Returns (wav_bytes, sample_rate).
        """
        if audio_type not in self._TTS_PROMPTS:
            raise RuntimeError(f"GGUF TTS does not support '{audio_type}' codec.")

        tpl, stop, need_ids = self._TTS_PROMPTS[audio_type]

        payload: dict = {
            "prompt": tpl.format(text = text),
            "stream": False,
            "n_predict": max_new_tokens,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k if top_k >= 0 else 0,
            "min_p": min_p,
            "repeat_penalty": repetition_penalty,
        }
        if stop:
            payload["stop"] = stop
        if need_ids:
            payload["n_probs"] = 1

        _auth_headers = (
            {"Authorization": f"Bearer {self._api_key}"} if self._api_key else None
        )
        with httpx.Client(
            timeout = httpx.Timeout(300, connect = 10), headers = _auth_headers
        ) as client:
            resp = client.post(f"{self.base_url}/completion", json = payload)
            if resp.status_code != 200:
                raise RuntimeError(
                    f"llama-server returned {resp.status_code}: {resp.text}"
                )

        data = resp.json()
        token_ids = (
            [p["id"] for p in data.get("completion_probabilities", []) if "id" in p]
            if need_ids
            else None
        )

        import torch

        device = "cuda" if torch.cuda.is_available() else "cpu"
        return LlamaCppBackend._codec_mgr.decode(
            audio_type, device, token_ids = token_ids, text = data.get("content", "")
        )