Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

18.2 KB
Raw Blame
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.

"""Engine-specific launch policy and selected-device memory budgeting."""

from __future__ import annotations

from dataclasses import dataclass
import subprocess

from utils.subprocess_compat import windows_hidden_subprocess_kwargs
import json
from pathlib import Path


def tool_parser_for_template(template, engine):
    """Select a native output parser from the model's own tool syntax."""
    if not isinstance(template, str) or "tools" not in template:
        return None
    formats = (
        (("<tool_call>", "<function="), "qwen3_coder", "qwen3_coder"),
        (("<tool_call>", "<arg_key>"), "glm45", "glm"),
        (("<tool_call>", "arguments"), "hermes", "qwen"),
        (("[TOOL_CALLS]",), "mistral", "mistral"),
        (("<|python_tag|>",), "llama3_json", "llama3"),
        (("<|channel|>", "<|message|>"), "openai", "gpt-oss"),
        (("<|tool_calls_section_begin|>",), "kimi_k2", "kimi_k2"),
        (("<minimax:tool_call>",), "minimax_m2", "minimax-m2"),
        (("<|tool▁calls▁begin|>",), "deepseek_v3", "deepseekv3"),
    )
    for markers, vllm, sglang in formats:
        if all(marker in template for marker in markers):
            return vllm if engine == "vllm" else sglang
    return None


_SGLANG_REASONING_PARSERS = {
    "openai_gptoss": "gpt-oss",
    "minimax_m2": "minimax",
    "deepseek_v3": "deepseek-v3",
    "deepseek_r1": "deepseek-r1",
}


def reasoning_parser_for_template(template, engine):
    """Reasoning parser for the model's own thinking markup; None leaves the launch unchanged."""
    parser = _vllm_reasoning_parser(template)
    return parser if engine == "vllm" else _SGLANG_REASONING_PARSERS.get(parser, parser)


def _vllm_reasoning_parser(template):
    if not isinstance(template, str):
        return None
    if "<|channel|>" in template and "<|message|>" in template:
        return "openai_gptoss"
    if "<think>" not in template or "</think>" not in template:
        return None
    for marker, parser in (
        ("<arg_key>", "glm45"),
        ("<minimax:tool_call>", "minimax_m2"),
        ("<|tool_calls_section_begin|>", "kimi_k2"),
    ):
        if marker in template:
            return parser
    # Per vLLM's reasoning_outputs.md: V3.1 / V3.2 gate on a `thinking` variable and answer untagged when
    # it is off (deepseek_v3); R1, QwQ and Qwen3 Thinking-2507 open <think> in the prompt (deepseek_r1).
    if "<|Assistant|>" in template:
        if "reasoning_effort" in template:
            return None
        return "deepseek_v3" if "thinking" in template.replace("think>", "") else "deepseek_r1"
    if "<|im_start|>" in template:
        return "qwen3" if "enable_thinking" in template else "deepseek_r1"
    return None


def _release_at_least(version: str | None, floor: str) -> bool:
    from packaging.version import InvalidVersion, Version
    try:
        return version is not None and Version(version) >= Version(floor)
    except InvalidVersion:
        return False


@dataclass(frozen = True)
class EngineAdapter:
    name: str
    module: str
    model_option: str
    context_option: str
    memory_option: str
    extra_args: tuple[str, ...] = ()
    exact_token_count: bool = False

    def key_environment(self, key):
        """The API key travels in the environment, never argv, which any local user can read."""
        return {"VLLM_API_KEY" if self.name == "vllm" else "UNSLOTH_ENGINE_API_KEY": key}

    def environment(self, gpu_count):
        if self.name == "vllm":
            # vLLM 0.30+ DeepSeek-OCR encoder kernel reads a plain-float LOG2E global, a NameError on
            # Triton 3.7 (unslothai/unsloth#3911). The knob only lifts that error and is not in the cache key.
            env = {"TRITON_ALLOW_NON_CONSTEXPR_GLOBALS": "1"}
            if gpu_count > 1:
                env["VLLM_HOST_IP"] = "127.0.0.1"
            return env
        if self.name == "sglang" and gpu_count > 1:
            # Else SGLang rejects GPUs whose free memory differs >10%; KV pools size from the minimum anyway.
            return {
                "SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK": "0",
                # CuTe RMSNorm compiles for rank zero's arch only; CUDA path handles mixed Ampere/Ada.
                "FLASHINFER_USE_CUDA_NORM": "1",
            }
        return {}

    def command(
        self,
        python,
        model,
        port,
        key,
        context,
        memory_fraction,
        gpu_count = 1,
        options = None,
        trust_remote_code = False,
        served_model_name = None,
    ):
        options = options or {}
        if options.get("gpu_memory_utilization"):
            # A user ceiling only lowers Studio's measured budget: above it the engine would OOM.
            memory_fraction = min(memory_fraction, options["gpu_memory_utilization"])
        tool_args = []
        if options.get("tool_parser"):
            tool_args = ["--tool-call-parser", options["tool_parser"]]
            if self.name == "vllm":
                tool_args.append("--enable-auto-tool-choice")
        if options.get("reasoning_parser"):
            tool_args += ["--reasoning-parser", options["reasoning_parser"]]
        mode = options.get("parallelism", "tensor")
        parallel_args = ["--tensor-parallel-size", str(gpu_count if mode == "tensor" else 1)]
        if mode != "tensor":
            flag = "--pipeline-parallel-size" if mode == "pipeline" else "--data-parallel-size"
            parallel_args += [flag, str(gpu_count)]
        if gpu_count > 1:
            parallel_args += (
                [
                    "--data-parallel-backend"
                    if mode == "data"
                    else "--distributed-executor-backend",
                    "mp",
                ]
                if self.name == "vllm"
                else ["--enable-p2p-check"]
            )
        precision = options.get("precision", "auto")
        precision_args = []
        legacy_torchao = not (
            self.name == "sglang" and _release_at_least(options.get("engine_version"), "0.5.18")
        )
        if not legacy_torchao and precision in ("int8", "int4"):
            # SGLang 0.5.18 removed --torchao-config, the only load-time INT8 / INT4 path it had.
            raise ValueError(
                f"SGLang {options.get('engine_version')} cannot convert weights to "
                f"{precision.upper()} when loading. Choose FP8 or Model default, or use vLLM."
            )
        if precision in ("bf16", "fp16"):
            precision_args = ["--dtype", "bfloat16" if precision == "bf16" else "float16"]
        elif self.name == "vllm" and (
            precision in ("int8", "fp8") or (precision == "int4" and mode == "pipeline")
        ):
            # Native online INT8 converts only MoE experts; online FP8 / BNB INT4 return invalid output.
            config = {
                "_type": {
                    "int4": "Int4WeightOnlyConfig",
                    "int8": "Int8WeightOnlyConfig",
                    "fp8": "Float8WeightOnlyConfig",
                }[precision],
                "_version": 1 if precision == "int8" else 2,
                "_data": {"set_inductor_config": True},
            }
            if precision == "int4":
                config["_data"].update(
                    {
                        "group_size": 32,
                        "int4_packing_format": {
                            "_type": "Int4PackingFormat",
                            "_data": "TILE_PACKED_TO_4D",
                        },
                        "int4_choose_qparams_algorithm": {
                            "_type": "Int4ChooseQParamsAlgorithm",
                            "_data": "HQQ",
                        },
                    }
                )
            precision_args = [
                "--quantization",
                "torchao",
                "--hf-overrides",
                json.dumps({"quantization_config_dict_json": json.dumps(config)}),
            ]
        elif precision == "int4":
            precision_args = (
                ["--quantization", "bitsandbytes"]
                if self.name == "vllm"
                else ["--torchao-config", "int4wo-32"]
            )
        elif precision == "int8":
            precision_args = ["--torchao-config", "int8wo"]
        elif precision == "fp8":
            precision_args = (
                ["--torchao-config", "fp8wo"]
                if options.get("disable_cuda_graph") and legacy_torchao
                else ["--quantization", "fp8"]
            )
        if self.name == "sglang" and "--torchao-config" in precision_args:
            # SGLang's Blackwell default (cutedsl TGV GEMM) has no TorchAO tensor-subclass dispatch.
            precision_args.extend(["--bf16-gemm-backend", "torch"])
        if options.get("disable_cuda_graph"):
            precision_args.extend(
                ["--enforce-eager"]
                if self.name == "vllm"
                else ["--disable-cuda-graph", "--disable-piecewise-cuda-graph"]
                if legacy_torchao
                # 0.5.18 folded both into per-phase backends and removed the piecewise flag.
                else [
                    "--cuda-graph-backend-decode",
                    "disabled",
                    "--cuda-graph-backend-prefill",
                    "disabled",
                ]
            )
        if self.name == "sglang":
            entrypoint = [
                str(Path(__file__).with_name("sglang_server.py")),
                self.model_option,
                model,
            ]
        else:
            launcher = str(Path(__file__).with_name("vllm_server.py"))
            entrypoint = (
                [launcher, "vllm.entrypoints.cli.main", "serve", model]
                if mode == "data"
                else [launcher, self.module, self.model_option, model]
            )
        return [
            python,
            "-I",
            *entrypoint,
            self.context_option,
            str(context),
            self.memory_option,
            str(memory_fraction),
            *parallel_args,
            *precision_args,
            *tool_args,
            *(
                ["--max-num-seqs", str(options["max_num_seqs"])]
                if self.name == "vllm" and options.get("max_num_seqs")
                else []
            ),
            *(
                [
                    "--kv-cache-dtype",
                    # SGLang names the FP8 format; vLLM's "fp8" is the same E4M3 cache.
                    "fp8_e4m3" if self.name == "sglang" else options["kv_cache_dtype"],
                ]
                if options.get("kv_cache_dtype")
                else []
            ),
            *(
                ["--random-seed" if self.name == "sglang" else "--seed", str(options["seed"])]
                if options.get("seed") is not None
                else []
            ),
            *(["--trust-remote-code"] if trust_remote_code else []),
            # Prefix caching is on by default in both; this only reports its hits in usage.
            "--enable-prompt-tokens-details" if self.name == "vllm" else "--enable-cache-report",
            *(
                vllm_speculative_args(model)
                if self.name == "vllm" and not options.get("no_speculative_config")
                else []
            ),
            *self.extra_args,
            "--host",
            "127.0.0.1",
            "--port",
            str(port),
            "--served-model-name",
            served_model_name or model,
            "--load-format",
            options.get("load_format", "auto"),
        ]

    def progress(self, line):
        lowered = line.lower()
        if "loading model weights" in lowered or "loading safetensors" in lowered:
            return "loading_weights"
        if any(word in lowered for word in ("cuda graph", "compil", "warmup")):
            return "warming_up"
        return None


ADAPTERS = {
    "vllm": EngineAdapter(
        "vllm",
        "vllm.entrypoints.openai.api_server",
        "--model",
        "--max-model-len",
        "--gpu-memory-utilization",
        ("--generation-config", "vllm"),
        True,
    ),
    "sglang": EngineAdapter(
        "sglang",
        "sglang.launch_server",
        "--model-path",
        "--context-length",
        "--mem-fraction-static",
        ("--attention-backend", "triton", "--sampling-backend", "pytorch"),
        True,
    ),
}


def _wsl_guest_to_host(path: str, nt: "bool | None" = None) -> str:
    """`/mnt/c/x` (the path the Windows WSL engine is given) back to `C:\\x` for host-side reads."""
    import os
    import re

    if (os.name == "nt" if nt is None else nt) and re.match(r"^/mnt/[a-zA-Z](/|$)", path):
        return path[5].upper() + ":\\" + path[7:].replace("/", "\\")
    return path


def vllm_speculative_args(model) -> list[str]:
    """`--speculative-config` for a merged Studio export that carries a drafter trained with the
    fine-tune (drafter.json): its own MTP head, or the EAGLE-3 drafter in `<export>/eagle3`."""
    import json
    import os

    if not isinstance(model, str):
        return []
    host = _wsl_guest_to_host(model)
    if not os.path.isdir(host):
        return []
    try:
        with open(os.path.join(host, "drafter.json"), encoding = "utf-8") as f:
            info = json.load(f)
    except (OSError, ValueError):
        return []
    if not isinstance(info, dict):
        return []
    # A drafter vLLM cannot run (llama.cpp-only EAGLE-3 targets, some MTP families) must not stop
    # the model from starting.
    if isinstance(info.get("engines"), dict) and not info["engines"].get("vllm"):
        return []
    steps = info.get("draft_steps") or 3
    if info.get("strategy") == "mtp":
        spec = {"method": "mtp", "num_speculative_tokens": int(steps)}
    elif info.get("strategy") == "eagle3" and os.path.isfile(
        os.path.join(host, "eagle3", "config.json")
    ):
        spec = {
            "method": "eagle3",
            # The path vLLM sees: the WSL guest's when the model path was translated for it.
            "model": (
                model.rstrip("/") + "/eagle3"
                if host != model
                else os.path.join(os.path.abspath(model), "eagle3")
            ),
            "num_speculative_tokens": int(steps),
        }
    else:
        return []
    return ["--speculative-config", json.dumps(spec, separators = (",", ":"))]


def launch_arguments(
    engine,
    python,
    model,
    port,
    key,
    context,
    memory_fraction = 0.8,
    tensor_parallel_size = 1,
):
    try:
        adapter = ADAPTERS[engine]
    except KeyError:
        raise ValueError("Unknown inference engine") from None
    return adapter.command(python, model, port, key, context, memory_fraction, tensor_parallel_size)


def memory_reserve_mib(engine: str, options: dict | None) -> int:
    """Reserve for allocations the engine does not budget; the warmup grows with vocabulary and card
    size."""
    options = options or {}
    precision = options.get("precision", "auto")
    torchao = engine == "vllm" and (
        precision in ("int8", "fp8")
        or (precision == "int4" and options.get("parallelism", "tensor") == "pipeline")
    )
    return (3072 if engine == "vllm" else 4096) + (6144 if torchao else 0)


RESERVE_SHARE = 0.06


MANAGED_ENGINE_BUDGET_FLOOR = 0.80


def gpu_memory_fraction(
    gpu_ids: list[int],
    reserve_mib: int = 512,
    reserve_share: float = 0.0,
    measure = None,
) -> float:
    """Budget every selected GPU after the previous resident stops; reserve the larger of
    reserve_mib and reserve_share per card for unbudgeted allocations; measure gives (total,
    free) MiB (default nvidia-smi); an unreadable device is an actionable failure, never a
    fallback to a larger engine default."""
    from utils.vram_budget_settings import get_vram_budget_fraction
    try:
        rows = (measure or _nvidia_memory_rows)(gpu_ids)
        if len(rows) != len(gpu_ids):
            raise ValueError("Could not measure every selected GPU")
        # vLLM / SGLang cannot offload to the CPU, so a GGUF-only low budget must not shrink them;
        # the per-load engine_gpu_memory_utilization still caps lower.
        fraction = max(get_vram_budget_fraction(), MANAGED_ENGINE_BUDGET_FLOOR)
        for total, free in rows:
            reserve = max(reserve_mib, reserve_share * total)
            if total <= 0 or free <= reserve or free > total:
                raise ValueError("Insufficient available GPU memory")
            # One fraction for all ranks: the most constrained GPU bounds it.
            fraction = min(fraction, (free - reserve) / total)
        if fraction < 0.05:
            raise ValueError("Insufficient available GPU memory")
        return int(fraction * 1000) / 1000
    except (OSError, ValueError, subprocess.SubprocessError) as exc:
        raise RuntimeError(
            "Could not reserve memory on every selected GPU. Free GPU memory or select fewer GPUs and retry."
        ) from exc


def _nvidia_memory_rows(gpu_ids: list[int]) -> list[tuple[float, float]]:
    # R570+ MIG ids are CUDA devices: budget each slice; nvidia-smi --id takes physical ids.
    try:
        from utils.hardware import mig
        slices = mig.rows_for(gpu_ids)
    except Exception:
        slices = None
    if slices is not None:
        if not all(float(r.get("memory_total_mib") or 0) > 0 for r in slices):
            raise RuntimeError(
                "The selected MIG GPU instance has no memory reading right now because the GPU "
                "probe did not answer. Retry in a few seconds."
            )
        return [
            (float(r["memory_total_mib"]), float(r.get("memory_free_mib") or 0)) for r in slices
        ]
    from utils.hardware.nvidia import _nvidia_smi_executable

    result = subprocess.run(
        [
            _nvidia_smi_executable(),
            "--id",
            ",".join(str(gpu_id) for gpu_id in gpu_ids),
            "--query-gpu=memory.total,memory.free",
            "--format=csv,noheader,nounits",
        ],
        capture_output = True,
        **windows_hidden_subprocess_kwargs(),
        text = True,
        encoding = "utf-8",
        errors = "replace",
        timeout = 60,
        check = True,
    )
    return [
        tuple(float(value.strip()) for value in row.split(","))
        for row in result.stdout.strip().splitlines()
    ]