unsloth/studio/backend/core/inference/diffusion_cache.py
Daniel Han de2f22df2b perf(image): compile numeric parity, cache-hook compile arming, FBCache toggle crash fix, TE fp8 zero-row guard
Applies the video round-2 accuracy findings to the image diffusion stack and fixes
two real image-path bugs found while measuring. All numbers B200, production
settings (family default steps/guidance, 1024px, seed 42, 4 fixed prompts), LPIPS
(AlexNet) via the new scripts/image_speedmem_bench.py, which drives the production
lever functions in the loader's own order.

- inductor precision parity: emulate_precision_casts=True on the regional-compile
  path (fused pointwise kernels keep fp32 intermediates where eager rounds to bf16
  between ops). Pairwise LPIPS of the compiled tier vs the same-stack eager tier:
  Qwen-Image 0.019 to 0.006 at identical speed (72.4 vs 72.5 ms/step), FLUX.1-dev
  0.046 to 0.029 at +2% step time (69.8 vs 68.3, reproduced), FLUX.2-klein-4B
  0.018 to 0.017 at identical speed. Snapshot/restored with the other process-wide
  backend flags so an off load never inherits it.
- cache x compile composition: re-point each cache hook's fn_ref.original_forward
  at a torch.compile'd wrapper of the same bound method (armed only where the
  speed layer compiled the block; restored before every disable_cache and before
  the partial-hook cleanup). Qwen-Image FBCache computed steps 91.8 to 71.2 ms
  (back at the uncached compiled rate), 1.21x end to end (7.36 to 6.06 s per 4
  images); FLUX.1-dev already traced through its FBCache hook and is measured
  neutral (same-process armed vs unarmed latents bit-identical). Skip counts
  within noise (13 vs 11 of 76; pairwise LPIPS 0.005).
- FBCache mid-session toggle crash: diffusers 0.39 caches the HookRegistry child
  list on first cache_context use, so an uncached generation followed by a
  20+-step generation (the auto toggle path) enabled hooks the context never
  reached and crashed with "No context is set" (reproduced live on FLUX.1-dev).
  Invalidate the stale child cache after every enable_cache.
- TE fp8_dynamic zero-row guard: torchao per-row fp8 derives a per-output-channel
  scale from the row amax, so an all-zero weight row is 0/0 = NaN. SDXL's
  text_encoder_2 (OpenCLIP bigG) ships exactly such a row, and every explicit
  fp8_dynamic SDXL render came out black; keep zero-row Linears dense (LPIPS
  0.976 black to 0.096 working). Other families' encoders have no such rows and
  are byte-identical.
- No AUTO TE quant exists on the image branch (text_encoder_quant defaults dense,
  explicit-only), so the video round's auto-dense retune has no image analogue;
  the explicit lever's cost is now measured (TE fp8_dynamic alone, LPIPS vs
  bit-exact: Qwen-Image 0.038, FLUX.1-dev 0.084, SDXL 0.096; no speed win, VRAM
  -6.5 GB on Qwen-Image) for the docs.

Tests: 96 passing across the cache/speed/precision suites (11 new arming, 2
child-registry, 2 zero-row, 4 inductor-flag); ruff clean.
2026-07-10 16:07:46 +00:00

381 lines
18 KiB
Python

# 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 step caching for the diffusion transformer (First-Block-Cache).
Across denoising steps a DiT's output changes little once the trajectory settles, so most of
the transformer can be reused. First-Block-Cache (FBCache) computes the first block, and if
its residual barely changed from the previous step (within ``threshold``) it skips the
remaining blocks and reuses their cached output. diffusers ships it natively
(``transformer.enable_cache(FirstBlockCacheConfig(...))`` for CacheMixin models, or the
standalone ``apply_first_block_cache`` hook).
Measured on Flux.1-dev (28 steps, 1024px, B200): ~1.4x on top of torch.compile (2.83 ->
2.03 s) at LPIPS ~0.08 vs the no-cache output -- deep inside the speed-for-quality bar.
OFF by default and a deliberate per-load opt-in, because the win scales with step count: a
few-step distilled model (e.g. Z-Image-Turbo at ~8 steps) has almost no headroom and a
single skipped step is a large fraction of the trajectory, so caching is for many-step
models (Flux / Qwen-Image). It composes with torch.compile only with ``fullgraph=False``
(the cache's compiler-disabled decision is a graph break), which the speed layer switches to
automatically when a cache is engaged. Best-effort: an incompatible model (e.g. a transformer
whose block signature the hook does not recognise) is caught and the load proceeds uncached.
torch / diffusers imported lazily.
"""
from __future__ import annotations
from typing import Any, Optional
TC_OFF = "off"
TC_AUTO = "auto"
TC_FBCACHE = "fbcache"
TC_MODES = (TC_FBCACHE,)
# FBCache residual thresholds: higher skips more steps (faster, lower quality). The dense
# bf16 default; a quantised transformer shifts the residual distribution, so it needs a
# higher threshold for the cache to trigger at all (per ParaAttention's fp8 guidance).
DEFAULT_FBCACHE_THRESHOLD = 0.08
QUANT_FBCACHE_THRESHOLD = 0.12
# The auto policy's step-count bar: FBCache's win scales with step count (each skipped
# step is a larger quality hit on a short trajectory), so auto engages it only at 20+
# steps -- full "dev"-style schedules (28+) qualify, distilled turbo models (4-9) never do.
FBCACHE_MIN_STEPS = 20
def normalize_transformer_cache(value: Optional[str]) -> Optional[str]:
"""Lower/strip a requested cache mode; None / "" / "none" / "off" -> None (disabled),
"auto" -> TC_AUTO (the loader decides from the step count).
Raises ValueError for an unsupported value so a bad request is rejected cheaply."""
if value is None:
return None
normalized = str(value).strip().lower().replace("-", "_")
if not normalized or normalized in ("none", "off"):
return None
if normalized == TC_AUTO:
return TC_AUTO
if normalized not in TC_MODES:
raise ValueError(
f"Unsupported transformer_cache '{value}'. Use one of: off, auto, "
f"{', '.join(TC_MODES)}."
)
return normalized
def _invalidate_child_registry_cache(transformer: Any) -> None:
"""Drop the HookRegistry's cached child-registry list after (un)installing hooks.
``cache_context`` propagates the state context through ``_get_child_registries``,
which diffusers 0.39 caches on first use. An UNCACHED generation already calls
``cache_context`` (the pipeline wraps every denoise call), creating the
transformer-level registry with an EMPTY cached child list -- so a later
``enable_cache`` (the auto step-count toggle engaging FBCache mid-session) installs
block hooks that ``_set_context`` never reaches, and the first cached forward dies
with "No context is set". Invalidate the stale cache so the next ``cache_context``
rebuilds it over the freshly hooked blocks. Best-effort and cheap (one attribute)."""
registry = getattr(transformer, "_diffusers_hook", None)
if registry is not None and getattr(registry, "_child_registries_cache", None) is not None:
try:
registry._child_registries_cache = None
except Exception: # noqa: BLE001 -- diffusers internals moved; leave as-is
pass
# diffusers' cache hook registry names whose compute branch we re-point at a compiled
# inner forward (leader = the measuring first block, block = the remaining ones); both
# hook families share the fn_ref layout.
_CACHE_HOOK_NAMES = (
"mag_cache_leader_block_hook",
"mag_cache_block_hook",
"fbc_leader_block_hook",
"fbc_block_hook",
)
def _compile_hooked_block_inners(transformer: Any, logger: Any = None) -> int:
"""Restore the regional compile on cache-hooked blocks' COMPUTED steps.
``enable_cache`` replaces each block's ``forward`` with the hook's ``new_forward``
(stashing the pre-hook bound method in ``fn_ref.original_forward``), whose skip
decision is data-dependent Python: MagCache ``@torch.compiler.disable``s the whole
``new_forward`` (recursive -- the compute branch runs EAGER), and even FBCache's
traceable ``new_forward`` graph-breaks around its disabled threshold decision,
which on some archs (measured: Qwen-Image) drops the compute branch's call into
``original_forward`` out of the compiled region -- the block's regional compile
artifact (``_compiled_call_impl``) is never reached and the cache forfeits the
compile win on every non-skipped step. An explicitly ``torch.compile``d callable
re-enables
dynamo for its own extent even inside a disabled frame, so re-pointing
``fn_ref.original_forward`` at a compiled wrapper of the same bound method restores
compiled compute steps while the skip decision stays eager exactly as designed.
Measured (B200, scripts/image_speedmem_bench.py): Qwen-Image FBCache computed steps
91.8 -> 71.2 ms (= the uncached compiled rate), 1.21x end to end; FLUX.1-dev is
neutral (its FBCache ``new_forward`` happens to trace, so computed steps were
already compiled -- same-process armed vs unarmed latents bit-identical); on the
video DiT balanced MagCache went 39.4 -> 26.9 s at 50 steps.
Only blocks the speed layer actually compiled are armed (``_compiled_call_impl``
guard -- eager tiers stay untouched), and only when ``original_forward`` is a plain
bound method (a stacked hook chain, e.g. offload, captures a partial and is
skipped). Idempotent via the ``_unsloth_orig_inner`` marker; best-effort. Returns
the number of hooks armed."""
try:
import torch
except Exception: # noqa: BLE001 -- no torch, nothing to arm
return 0
armed = 0
try:
for module in transformer.modules():
registry = getattr(module, "_diffusers_hook", None)
if registry is None or getattr(module, "_compiled_call_impl", None) is None:
continue
hooks = getattr(registry, "hooks", None) or {}
for name in _CACHE_HOOK_NAMES:
hook = hooks.get(name)
fn_ref = getattr(hook, "fn_ref", None) if hook is not None else None
orig = getattr(fn_ref, "original_forward", None)
if orig is None or getattr(hook, "_unsloth_orig_inner", None) is not None:
continue
if getattr(orig, "__self__", None) is None:
continue # not the plain bound method; arming would miss the block
# fullgraph=False / dynamic=True: a cache is active by definition (its
# decision points graph-break) and this matches the default tier the
# regional compile used. Dynamo caches per code object, so re-arming
# after a toggle is effectively free (~0.03 s).
fn_ref.original_forward = torch.compile(orig, fullgraph = False, dynamic = True)
hook._unsloth_orig_inner = orig
armed += 1
except Exception as exc: # noqa: BLE001 -- best-effort: the cache still works eager
_warn(logger, "cache-hook inner compile", exc)
return armed
if armed and logger is not None:
logger.info(
"diffusion.cache: %d cache-hooked block(s) armed with compiled inner forwards",
armed,
)
return armed
def _restore_hooked_block_inners(transformer: Any) -> None:
"""Undo ``_compile_hooked_block_inners``: put the plain bound methods back and clear
the markers. MUST run before ``disable_cache`` -- ``remove_hook`` splices
``fn_ref.original_forward`` back into ``module.forward``, and leaving the compiled
wrapper there would pin a stale compiled callable onto the uncached path."""
try:
modules = list(transformer.modules())
except Exception: # noqa: BLE001 -- not a torch module (tests/fakes): nothing armed
return
for module in modules:
registry = getattr(module, "_diffusers_hook", None)
if registry is None:
continue
hooks = getattr(registry, "hooks", None) or {}
for name in _CACHE_HOOK_NAMES:
hook = hooks.get(name)
orig = getattr(hook, "_unsloth_orig_inner", None) if hook is not None else None
if orig is None:
continue
try:
hook.fn_ref.original_forward = orig
hook._unsloth_orig_inner = None
except Exception: # noqa: BLE001 -- per-hook best-effort
pass
def _pipeline_opens_cache_context(pipe: Any) -> bool:
"""Whether the pipeline enters ``transformer.cache_context(...)`` in its denoise loop.
The First-Block-Cache hook requires it at run time, and a CacheMixin transformer alone
does NOT guarantee it: Flux Kontext / img2img / inpaint / controlnet reuse the CacheMixin
FluxTransformer2DModel but never open a cache_context. Read from the pipeline ``__call__``
source, resolved off the instance so a per-expert proxy view (``_SecondDiTView``)
delegates to the real pipe; if it cannot be read, report False so the cache stays off."""
import inspect
call = getattr(pipe, "__call__", None)
if call is None:
return False
try:
src = inspect.getsource(call)
except (OSError, TypeError):
return False
# Match the actual call `cache_context(` -- a bare mention in a comment/docstring lacks
# the paren, so this does not false-positive on prose.
return "cache_context(" in src
def apply_step_cache(
pipe: Any,
*,
mode: Optional[str],
threshold: Optional[float] = None,
quant_active: bool = False,
logger: Any = None,
) -> Optional[str]:
"""Engage step caching on ``pipe.transformer``. Returns the mode actually engaged, or
None when disabled / unsupported (the load then runs uncached). ``threshold`` overrides
the default; ``quant_active`` raises the default so the cache still triggers on a
quantised transformer. Best-effort: never raises for an incompatible model."""
mode = normalize_transformer_cache(mode)
if mode is None or mode == TC_AUTO:
# AUTO must be resolved by the loader (step-count policy) before reaching the
# engage call; treat a stray auto as off rather than crashing the load.
return None
transformer = getattr(pipe, "transformer", None)
if transformer is None:
return None
thr = (
threshold
if threshold is not None
else (QUANT_FBCACHE_THRESHOLD if quant_active else DEFAULT_FBCACHE_THRESHOLD)
)
# Engage only via the transformer's native enable_cache (the diffusers CacheMixin path):
# the lower-level apply_first_block_cache hook would install on a non-CacheMixin
# transformer too (e.g. Z-Image), whose pipeline opens no cache_context and would crash
# the first generation -- so a model without enable_cache runs uncached per the
# best-effort contract instead of being reported as cached and then failing.
enable_cache = getattr(transformer, "enable_cache", None)
if not callable(enable_cache):
_warn(logger, mode, RuntimeError("transformer has no cache_context (not a CacheMixin)"))
return None
# A CacheMixin transformer is necessary but NOT sufficient: the First-Block-Cache hook
# raises "No context is set" on the first forward unless the PIPELINE wraps its denoise
# loop in transformer.cache_context(...). Flux Kontext / img2img / inpaint / controlnet
# reuse the CacheMixin FluxTransformer2DModel yet their __call__ opens no cache_context,
# so engaging FBCache there would crash every default generation -- run uncached instead.
if not _pipeline_opens_cache_context(pipe):
_warn(
logger, mode, RuntimeError("pipeline __call__ opens no cache_context; running uncached")
)
return None
try:
try:
from diffusers import FirstBlockCacheConfig
except ImportError: # older diffusers exports it only from diffusers.hooks
from diffusers.hooks import FirstBlockCacheConfig
config = FirstBlockCacheConfig(threshold = thr)
enable_cache(config)
# enable_cache AFTER the pipe has already run leaves a stale cached child-registry
# list on the transformer's HookRegistry; the block hooks just installed would then
# never receive the cache context. Must follow every enable_cache.
_invalidate_child_registry_cache(transformer)
# If the blocks are already regionally compiled (the generation-time toggle
# path: compile ran at load), re-point the fresh hooks' compute branch at
# compiled inners; the load path (cache before compile) is armed by
# _compile_repeated_blocks instead. No-op when nothing is compiled.
_compile_hooked_block_inners(transformer, logger)
try:
transformer._unsloth_step_cache = f"{mode}@{thr}"
except Exception: # noqa: BLE001 — marker is best-effort
pass
if logger is not None:
logger.info("diffusion.cache: %s engaged (threshold=%s)", mode, thr)
return mode
except Exception as exc: # noqa: BLE001 — incompatible model -> run uncached
# enable_cache can fail after hooking some blocks; drop any partial hooks so
# the reported-uncached model doesn't actually run half-cached. Any armed
# compiled inners must be restored FIRST (remove_hook splices original_forward
# back into module.forward).
_restore_hooked_block_inners(transformer)
try:
transformer.disable_cache()
except Exception: # noqa: BLE001
pass
_warn(logger, mode, exc)
return None
def effective_denoise_steps(steps: int, strength: Optional[float]) -> int:
"""The number of steps diffusers ACTUALLY denoises for a request.
An image-conditioned workflow with ``strength`` < 1 (img2img / upscale / inpaint) runs
only a fraction of ``num_inference_steps``: diffusers' ``get_timesteps`` computes
``init_timestep = min(int(num_inference_steps * strength), num_inference_steps)`` and
denoises exactly ``init_timestep`` steps -- the product is FLOORED, not rounded. The auto
step-cache policy must key on THIS count -- e.g. a 28-step upscale at strength 0.35 runs
``int(9.8) = 9`` real steps, exactly the short trajectory FBCache should stay off (each
skipped step is a large quality hit). ``strength`` None (txt2img / reference) or >= 1 -> the
full count.
"""
s = int(steps)
if strength is None or float(strength) >= 1.0:
return s
return max(1, min(int(s * float(strength)), s))
def effective_request_strength(
request_strength: Optional[float],
has_init_image: bool,
pipe_accepts_strength: bool,
pipe_default_strength: Any,
) -> Optional[float]:
"""The strength the pipe will ACTUALLY apply, for keying the auto step-cache policy.
Only image-conditioned pipelines that take ``strength`` apply it (txt2img / a pipe without
the kwarg run the full trajectory -> None). When the request omits ``strength`` the loader
does NOT pass the kwarg, so the pipe runs its OWN signature default (< 1 for every img2img /
inpaint pipeline here, e.g. 0.6); the policy must key on that default, not the full step
count, or FBCache engages on a fraction of the advertised steps. A non-numeric default
(``inspect.Parameter.empty``) falls back to the full count (None).
"""
if not (has_init_image and pipe_accepts_strength):
return None
if request_strength is not None:
return request_strength
return pipe_default_strength if isinstance(pipe_default_strength, (int, float)) else None
def maybe_toggle_step_cache(
pipe: Any,
*,
steps: int,
quant_active: bool = False,
threshold: Optional[float] = None,
logger: Any = None,
) -> Optional[str]:
"""Generation-time enable/disable for an AUTO cache decision, keyed on the actual
step count: engage FBCache at ``FBCACHE_MIN_STEPS`` or more, run uncached below it.
Idempotent (the ``_unsloth_step_cache`` marker tracks the engaged state), so calling
it on every generation is cheap. Only the loader's auto path calls this; an explicit
user choice is never toggled. Returns the mode now active (or None when uncached)."""
transformer = getattr(pipe, "transformer", None)
if transformer is None:
return None
engaged = getattr(transformer, "_unsloth_step_cache", None)
want = int(steps) >= FBCACHE_MIN_STEPS
if want and not engaged:
return apply_step_cache(
pipe,
mode = TC_FBCACHE,
threshold = threshold,
quant_active = quant_active,
logger = logger,
)
if not want and engaged:
disable_cache = getattr(transformer, "disable_cache", None)
if callable(disable_cache):
try:
# Before remove_hook splices fn_ref.original_forward back into
# module.forward: the compiled inner wrappers must not leak onto the
# uncached path.
_restore_hooked_block_inners(transformer)
disable_cache()
transformer._unsloth_step_cache = None
if logger is not None:
logger.info(
"diffusion.cache: fbcache disengaged (auto: %s steps < %s)",
steps,
FBCACHE_MIN_STEPS,
)
return None
except Exception as exc: # noqa: BLE001 — keep the cache rather than crash
_warn(logger, "fbcache disable", exc)
return TC_FBCACHE
return TC_FBCACHE if engaged else None
def _warn(logger: Any, what: str, exc: Exception) -> None:
if logger is not None:
logger.warning("diffusion.cache: %s unavailable (%s); running uncached", what, exc)