unsloth/studio/backend/core/inference/diffusion_speed.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

"""Opt-in speed optimisations for the local diffusion backend.

Off by default, so the default render path stays bit-identical to a plain run (the
property the regression harness checks). When the operator opts in, this applies the
near-lossless speedups in the order the diffusers guides recommend
(channels_last + cudnn.benchmark -> compile, with TF32 / fused-QKV under "max"):

  off     - nothing (default; bit-identical reference).
  eager   - everything lossless EXCEPT torch.compile: channels_last VAE +
            cudnn.benchmark + the attention backend + the shared eager monkey-patches
            (fused RMSNorm / AdaLayerNorm + per-arch addcmul fusions, see
            diffusion_eager_patches.py / diffusion_arch_patches.py). The fast first-image
            / casual-use path -- no compile tax to amortise.
  default - LIGHT compile. For a GGUF model: channels_last + cudnn.benchmark +
            torch.compile of ONLY the dequant op chain
            (``torch.compile(dequantize_gguf_tensor, dynamic=True)``) -- the dequant is
            ~70-80% of eager GGUF time, so fusing it gives ~1.24-1.64x for a small
            one-time compile (~7.5-10.4s) and ZERO extra VRAM, resolution-invariant
            (the dequant inputs are fixed-shape weights). For a dense (non-GGUF) model
            there is no dequant, so ``default`` falls back to regional torch.compile of
            the denoiser's repeated block (the only compile lever a dense model has).
  max     - the FULL torch.compile: regional max-autotune compile of the denoiser's
            repeated block (which fuses the GGUF dequant AND the matmul/norm/elementwise
            in one graph -- ~3.2x on the GGUF Z-Image transformer, PSNR ~36 dB vs eager,
            well above the Q4 noise floor) plus TF32 matmul and fused QKV projections.

Tier rationale: ``default`` is the cheap, always-amortising compile (compile just the
hot GGUF dequant; the block stays eager) so the first image is fast and VRAM is
untouched; ``max`` pays the larger regional-compile tax for the bigger warm speedup.
The compiled dequant is deliberately skipped under ``max`` -- the regional block compile
subsumes the dequant fusion (a separately-compiled dequant would be traced into that
graph and break it), so ``max`` runs the stock dequant and lets the block compile it. The
per-family ``supports_torch_compile`` flag and the bf16/CUDA checks gate regional compile.

The backend flags this layer flips (TF32, cudnn.benchmark) are PROCESS-WIDE, so
``snapshot_backend_flags`` / ``restore_backend_flags`` let the caller capture the
prior values at load and restore them at unload, keeping a later ``off`` load
bit-identical instead of inheriting a previous ``max`` run's globals. torch is
imported lazily.
"""

from __future__ import annotations

from typing import Any, Optional

from . import diffusion_gguf_compile as gguf_compile

SPEED_OFF = "off"
SPEED_EAGER = "eager"
SPEED_DEFAULT = "default"
SPEED_MAX = "max"
SPEED_MODES = (SPEED_OFF, SPEED_EAGER, SPEED_DEFAULT, SPEED_MAX)


def snapshot_backend_flags() -> Optional[dict]:
    """Capture the process-wide torch backend flags this layer may mutate, so the
    caller can restore them on unload. None if torch is unavailable."""
    try:
        import torch
        return {
            "matmul_tf32": bool(torch.backends.cuda.matmul.allow_tf32),
            "cudnn_tf32": bool(torch.backends.cudnn.allow_tf32),
            "cudnn_benchmark": bool(torch.backends.cudnn.benchmark),
        }
    except Exception:  # noqa: BLE001 — best-effort; no snapshot -> no restore
        return None


def restore_backend_flags(state: Optional[dict]) -> None:
    """Restore the flags captured by ``snapshot_backend_flags``. No-op on None."""
    if not state:
        return
    try:
        import torch

        torch.backends.cuda.matmul.allow_tf32 = state["matmul_tf32"]
        torch.backends.cudnn.allow_tf32 = state["cudnn_tf32"]
        torch.backends.cudnn.benchmark = state["cudnn_benchmark"]
    except Exception:  # noqa: BLE001 — best-effort restore
        return


def normalize_speed_mode(value: Optional[str]) -> str:
    """Lower/strip a requested speed mode (dashes ok); None / "" -> off."""
    if value is None:
        return SPEED_OFF
    normalized = str(value).strip().lower().replace("-", "_")
    if not normalized:
        return SPEED_OFF
    if normalized not in SPEED_MODES:
        raise ValueError(
            f"Unsupported diffusion speed_mode '{value}'. Use one of: {', '.join(SPEED_MODES)}."
        )
    return normalized


def resolve_speed_mode(value: Optional[str], *, is_gguf: bool) -> str:
    """The effective speed mode when the caller leaves it UNSET (``None``).

    A GGUF model defaults to ``default``: it compiles only the hot dequant op chain
    (~70-80% of eager GGUF time) for ~1.24-1.64x at a small one-time compile and zero
    extra VRAM -- a cheap, always-amortising win whose numeric perturbation sits well
    below the quantisation noise floor (the dequant graph is unchanged, just
    Inductor-fused). A dense (non-GGUF) model stays ``off`` / bit-identical, since there
    compile would be the only source of drift. An explicit value -- including ``"off"``
    -- is always honored verbatim."""
    if value is None:
        return SPEED_DEFAULT if is_gguf else SPEED_OFF
    return normalize_speed_mode(value)


def compile_eligible(target: Any, *, is_gguf: bool, family: Any) -> bool:
    """Whether the denoiser's repeated block should be regionally compiled.

    Only on CUDA (incl. ROCm via supports_default_torch_compile), for a bf16
    transformer, on a compile-friendly family. ``is_gguf`` no longer disqualifies:
    ``compile_repeated_blocks`` runs fine on the GGUF transformer (the per-op
    dequant stays eager, the rest of the block compiles) and is ~2.3x faster, so it
    is kept only for signature/logging compatibility."""
    del is_gguf  # GGUF is compile-eligible now; param kept for call-site compat.
    if not bool(getattr(target, "supports_default_torch_compile", False)):
        return False
    if not bool(getattr(family, "supports_torch_compile", True)):
        return False
    return _is_bfloat16(getattr(target, "dtype", None))


def _is_bfloat16(dtype: Any) -> bool:
    try:
        import torch
        return dtype is torch.bfloat16
    except Exception:
        return str(dtype).endswith("bfloat16")


def apply_speed_optims(
    pipe: Any,
    target: Any,
    *,
    is_gguf: bool,
    family: Any,
    speed_mode: str = SPEED_OFF,
    cache_active: bool = False,
    logger: Any = None,
) -> dict[str, bool]:
    """Apply the opt-in speed optimisations for ``speed_mode`` to a built pipeline,
    BEFORE placement / offload. Returns which optimisations actually engaged. Every
    step is best-effort: a pipeline that doesn't support one is simply skipped."""
    applied = {
        "channels_last": False,
        "cudnn_benchmark": False,
        "tf32": False,
        "fused_qkv": False,
        "compiled": False,
        "compiled_dequant": False,
    }
    mode = normalize_speed_mode(speed_mode)
    if mode == SPEED_OFF:
        return applied

    on_cuda = getattr(target, "device", None) == "cuda"
    family_allows_compile = bool(getattr(family, "supports_torch_compile", True))

    # Lossless: a channels-last VAE speeds up its convolutions with no numeric change.
    applied["channels_last"] = _vae_channels_last(pipe, logger)

    # Near-lossless: let cuDNN autotune the fixed-shape VAE convs (CUDA only). It may
    # pick a different conv algorithm, so it is a "default"-tier (not bit-identical) win.
    if on_cuda:
        applied["cudnn_benchmark"] = _enable_cudnn_benchmark(logger)

    # --- the compile lever, remapped per tier ----------------------------------------
    # default = LIGHT compile: for a GGUF model, compile ONLY the dequant op chain
    #   (~70-80% of eager GGUF time) -- cheap, VRAM-free, resolution-invariant; the
    #   transformer block stays eager. A dense model has no dequant, so default falls
    #   back to the regional block compile (its only compile lever).
    # max = FULL compile: regional max-autotune compile of the repeated denoiser block
    #   (fuses dequant + matmul + norm + elementwise in one graph). It subsumes the
    #   dequant fusion, so we do NOT also install the standalone compiled dequant here.
    # eager = no compile at all.
    if mode == SPEED_DEFAULT:
        if is_gguf and on_cuda and family_allows_compile:
            applied["compiled_dequant"] = gguf_compile.install_compiled_dequant(logger)
        elif compile_eligible(target, is_gguf = is_gguf, family = family):
            applied["compiled"] = _compile_repeated_blocks(
                pipe, logger, max_autotune = False, cache_active = cache_active
            )
    elif mode == SPEED_MAX and compile_eligible(target, is_gguf = is_gguf, family = family):
        applied["compiled"] = _compile_repeated_blocks(
            pipe, logger, max_autotune = True, cache_active = cache_active
        )

    if mode == SPEED_MAX:
        # Near-lossless: TF32 matmul (CUDA only) trades a few mantissa bits for speed.
        if on_cuda:
            applied["tf32"] = _enable_tf32(logger)
        applied["fused_qkv"] = _fuse_qkv(pipe, logger)

    return applied


def _vae_channels_last(pipe: Any, logger: Any) -> bool:
    vae = getattr(pipe, "vae", None)
    if vae is None or not hasattr(vae, "to"):
        return False
    try:
        import torch
        vae.to(memory_format = torch.channels_last)
        return True
    except Exception as exc:  # noqa: BLE001 — optimisation only
        _warn(logger, "channels_last", exc)
        return False


def _compile_repeated_blocks(
    pipe: Any,
    logger: Any,
    *,
    max_autotune: bool = False,
    cache_active: bool = False,
) -> bool:
    transformer = getattr(pipe, "transformer", None)
    fn = getattr(transformer, "compile_repeated_blocks", None)
    if not callable(fn):
        return False
    # default: mode="default" + dynamic=True -- fast cold start, robust to resolution
    # changes (no recompile). max: mode="max-autotune-no-cudagraphs" + dynamic=False --
    # Triton autotuning for a few % more on GEMM/conv-heavy models, at a much longer
    # compile and a recompile per new resolution. The CUDA-graph modes (reduce-overhead
    # / max-autotune) are deliberately NOT used: they crash on the regionally-compiled
    # block because its static output buffer is overwritten across denoise steps.
    #
    # fullgraph drops to False when a step cache is engaged: FBCache's per-step decision is
    # ``@torch.compiler.disable``d, i.e. a graph break, which fullgraph=True rejects ("Skip
    # inlining torch.compiler.disable()d function"). The break is cheap and the rest of the
    # block still compiles.
    kwargs: dict[str, Any] = {"fullgraph": not cache_active, "dynamic": not max_autotune}
    if max_autotune:
        kwargs["mode"] = "max-autotune-no-cudagraphs"
    try:
        fn(**kwargs)
        return True
    except Exception as exc:  # noqa: BLE001 — optimisation only
        _warn(logger, "compile_repeated_blocks", exc)
        return False


def _enable_cudnn_benchmark(logger: Any) -> bool:
    try:
        import torch
        torch.backends.cudnn.benchmark = True
        return True
    except Exception as exc:  # noqa: BLE001 — optimisation only
        _warn(logger, "cudnn_benchmark", exc)
        return False


def _enable_tf32(logger: Any) -> bool:
    try:
        import torch

        torch.backends.cuda.matmul.allow_tf32 = True
        torch.backends.cudnn.allow_tf32 = True
        return True
    except Exception as exc:  # noqa: BLE001 — optimisation only
        _warn(logger, "tf32", exc)
        return False


def _fuse_qkv(pipe: Any, logger: Any) -> bool:
    for owner in (pipe, getattr(pipe, "transformer", None)):
        fn = getattr(owner, "fuse_qkv_projections", None)
        if callable(fn):
            try:
                fn()
                return True
            except Exception as exc:  # noqa: BLE001 — optimisation only
                _warn(logger, "fuse_qkv_projections", exc)
                return False
    return False


def _warn(logger: Any, what: str, exc: Exception) -> None:
    if logger is not None:
        logger.warning("diffusion.speed: %s failed: %s", what, exc)