Trim the comments across the diffusion backend
Comment-only pass over the Python this PR touches: drop what the code already says, collapse multi-line explanations that still read on one line, and keep the reasoning that is not recoverable from the code. No code, docstring semantics or behaviour changes; verified with an AST comparison against the previous revision, and the backend suite is unchanged (same 37 environment failures as before: the API integration tests that need a live keyed server, the flash-attn install hooks, and the GPU memory fields).
This commit is contained in:
Daniel Han
parent
ea3c3822a1
commit
36df317293
113 changed files with 3389 additions and 4087 deletions
|
|
@ -144,8 +144,8 @@ def test_explicit_threshold_overrides_quant(monkeypatch):
|
|||
|
||||
|
||||
def test_non_cachemixin_runs_uncached(monkeypatch):
|
||||
# A transformer without enable_cache (e.g. Z-Image) must NOT install the standalone hook
|
||||
# -- its pipeline opens no cache_context, so it runs uncached instead of crashing at gen.
|
||||
# A transformer without enable_cache (e.g. Z-Image) must NOT install the standalone hook: its
|
||||
# pipeline opens no cache_context, so it runs uncached instead of crashing at generation.
|
||||
rec: dict = {}
|
||||
_stub_diffusers(monkeypatch, hook_recorder = rec)
|
||||
t = _NonCacheMixinTransformer()
|
||||
|
|
@ -154,10 +154,9 @@ def test_non_cachemixin_runs_uncached(monkeypatch):
|
|||
|
||||
|
||||
def test_pipeline_without_cache_context_runs_uncached(monkeypatch):
|
||||
# A CacheMixin transformer whose PIPELINE never opens a cache_context (Flux Kontext /
|
||||
# img2img / inpaint / controlnet reuse the CacheMixin FluxTransformer2DModel) must run
|
||||
# uncached -- otherwise the First-Block-Cache hook raises "No context is set" on the
|
||||
# first forward, crashing every default generation.
|
||||
# A CacheMixin transformer whose PIPELINE never opens a cache_context (Flux Kontext / img2img /
|
||||
# inpaint / controlnet reuse FluxTransformer2DModel) must run uncached, else the First-Block-Cache
|
||||
# hook raises "No context is set" on the first forward.
|
||||
_stub_diffusers(monkeypatch)
|
||||
t = _MixinTransformer()
|
||||
assert apply_step_cache(_NoCtxPipe(t), mode = "fbcache") is None
|
||||
|
|
@ -172,8 +171,8 @@ def test_incompatible_model_runs_uncached(monkeypatch):
|
|||
|
||||
|
||||
def test_enable_cache_failure_rolls_back_partial_hooks(monkeypatch):
|
||||
# enable_cache can raise after hooking some blocks; the reported-uncached model
|
||||
# must not actually run half-cached, so the failure path calls disable_cache.
|
||||
# enable_cache can raise after hooking some blocks; the reported-uncached model must not actually
|
||||
# run half-cached, so the failure path calls disable_cache.
|
||||
_stub_diffusers(monkeypatch)
|
||||
t = _MixinTransformer(fail = True)
|
||||
t.disabled = False
|
||||
|
|
@ -201,9 +200,9 @@ def test_missing_transformer_is_none(monkeypatch):
|
|||
|
||||
|
||||
def test_diffusers_unavailable_runs_uncached(monkeypatch):
|
||||
# no diffusers import -> best-effort returns None, load proceeds uncached. Block the
|
||||
# hooks module too: the config import falls back to diffusers.hooks, which a REAL
|
||||
# earlier import in the test session may have left cached in sys.modules.
|
||||
# No diffusers import: best-effort returns None and the load proceeds uncached. Block the hooks
|
||||
# module too, since the config import falls back to it and a real earlier import may have left it
|
||||
# cached in sys.modules.
|
||||
monkeypatch.setitem(sys.modules, "diffusers", None)
|
||||
monkeypatch.setitem(sys.modules, "diffusers.hooks", None)
|
||||
t = _MixinTransformer()
|
||||
|
|
@ -228,9 +227,8 @@ def test_effective_steps_txt2img_is_full_count():
|
|||
|
||||
|
||||
def test_effective_steps_low_strength_shrinks_below_the_bar():
|
||||
# A 28-step upscale at strength 0.35 denoises int(9.8) = 9 steps (diffusers get_timesteps
|
||||
# floors the product), which is below FBCACHE_MIN_STEPS -> the auto policy must NOT engage
|
||||
# FBCache there.
|
||||
# A 28-step upscale at strength 0.35 denoises int(9.8) = 9 steps (diffusers floors the product),
|
||||
# below FBCACHE_MIN_STEPS, so the auto policy must NOT engage FBCache.
|
||||
eff = effective_denoise_steps(28, 0.35)
|
||||
assert eff == 9
|
||||
assert eff < FBCACHE_MIN_STEPS
|
||||
|
|
@ -239,33 +237,31 @@ def test_effective_steps_low_strength_shrinks_below_the_bar():
|
|||
def test_effective_request_strength_uses_pipe_default_when_omitted():
|
||||
import inspect
|
||||
|
||||
# txt2img (no init image) or a pipe without the strength kwarg -> full trajectory (None).
|
||||
# txt2img or a pipe without the strength kwarg gives the full trajectory (None).
|
||||
assert effective_request_strength(None, False, True, 0.6) is None
|
||||
assert effective_request_strength(0.5, True, False, None) is None
|
||||
# img2img with an explicit strength -> that value.
|
||||
assert effective_request_strength(0.2, True, True, 0.6) == 0.2
|
||||
# img2img with an OMITTED strength -> the pipe's own signature default (< 1), so the auto
|
||||
# policy keys on the real (short) trajectory, not the full step count. This is the fix:
|
||||
# int(28 * 0.6) = 16 real steps, not 28.
|
||||
# img2img with an OMITTED strength uses the pipe's own signature default, so the auto policy keys
|
||||
# on the real (short) trajectory: int(28 * 0.6) = 16 real steps, not 28.
|
||||
s = effective_request_strength(None, True, True, 0.6)
|
||||
assert s == 0.6
|
||||
assert effective_denoise_steps(28, s) == 16
|
||||
# A non-numeric signature default (inspect.Parameter.empty) falls back to the full count.
|
||||
# A non-numeric signature default falls back to the full count.
|
||||
assert effective_request_strength(None, True, True, inspect.Parameter.empty) is None
|
||||
assert effective_request_strength(None, True, True, None) is None
|
||||
|
||||
|
||||
def test_effective_steps_matches_diffusers_get_timesteps():
|
||||
# Mirror diffusers exactly: it denoises init_timestep = min(int(num_inference_steps *
|
||||
# strength), num_inference_steps) steps (the product is floored, not rounded).
|
||||
# Mirror diffusers exactly: it denoises min(int(steps * strength), steps) steps (floored).
|
||||
for steps, strength in [(28, 0.35), (28, 0.8), (50, 0.5), (20, 0.99), (30, 0.1)]:
|
||||
expected = max(1, min(int(steps * strength), steps))
|
||||
assert effective_denoise_steps(steps, strength) == expected
|
||||
|
||||
|
||||
def test_toggle_stays_off_for_low_strength_workflow(monkeypatch):
|
||||
# End to end: a 28-step request would engage FBCache, but at strength 0.35 the
|
||||
# effective ~10 steps keep it uncached.
|
||||
# End to end: a 28-step request would engage FBCache, but at strength 0.35 the effective ~10 steps
|
||||
# keep it uncached.
|
||||
_stub_diffusers(monkeypatch)
|
||||
t = _ToggleTransformer()
|
||||
mode = maybe_toggle_step_cache(_pipe(t), steps = effective_denoise_steps(28, 0.35))
|
||||
|
|
@ -294,8 +290,8 @@ def test_normalize_auto_is_a_distinct_state():
|
|||
|
||||
|
||||
def test_apply_treats_stray_auto_as_off(monkeypatch):
|
||||
# AUTO must be resolved by the loader; if it ever reaches the engage call the
|
||||
# load runs uncached instead of crashing.
|
||||
# AUTO must be resolved by the loader; if it ever reaches the engage call the load runs uncached
|
||||
# instead of crashing.
|
||||
_stub_diffusers(monkeypatch)
|
||||
t = _MixinTransformer()
|
||||
assert apply_step_cache(_pipe(t), mode = "auto") is None
|
||||
|
|
@ -432,8 +428,8 @@ def test_arming_is_idempotent(monkeypatch):
|
|||
|
||||
|
||||
def test_arming_skips_uncompiled_blocks(monkeypatch):
|
||||
# An eager-tier load has no _compiled_call_impl: the hook must stay untouched
|
||||
# (compiling the inner would ADD compile where the user chose eager).
|
||||
# An eager-tier load has no _compiled_call_impl: the hook must stay untouched (compiling the inner
|
||||
# would ADD compile where the user chose eager).
|
||||
_stub_torch_compile(monkeypatch)
|
||||
block, hook, orig = _hooked_block(compiled = False)
|
||||
assert _compile_hooked_block_inners(_fake_dit([block])) == 0
|
||||
|
|
@ -441,8 +437,8 @@ def test_arming_skips_uncompiled_blocks(monkeypatch):
|
|||
|
||||
|
||||
def test_arming_skips_partial_captured_inner(monkeypatch):
|
||||
# A stacked hook chain (e.g. group offload) captures a functools.partial, not the
|
||||
# plain bound method; arming would compile the wrong layer of the chain.
|
||||
# A stacked hook chain (e.g. group offload) captures a functools.partial, not the plain bound
|
||||
# method; arming would compile the wrong layer of the chain.
|
||||
_stub_torch_compile(monkeypatch)
|
||||
block, hook, orig = _hooked_block(bound = False)
|
||||
assert _compile_hooked_block_inners(_fake_dit([block])) == 0
|
||||
|
|
@ -450,8 +446,8 @@ def test_arming_skips_partial_captured_inner(monkeypatch):
|
|||
|
||||
|
||||
def test_arming_covers_every_cache_hook_family(monkeypatch):
|
||||
# FBCache is the image cache today, but the hook-name table already covers the
|
||||
# MagCache layout too (same fn_ref shape), so a future mode arms for free.
|
||||
# FBCache is the image cache today, but the hook-name table already covers the MagCache layout
|
||||
# (same fn_ref shape), so a future mode arms for free.
|
||||
_stub_torch_compile(monkeypatch)
|
||||
names = (
|
||||
"mag_cache_leader_block_hook",
|
||||
|
|
@ -478,8 +474,8 @@ def test_restore_tolerates_fakes_without_modules():
|
|||
|
||||
|
||||
def test_apply_step_cache_arms_compiled_blocks_on_toggle(monkeypatch):
|
||||
# The generation-time toggle engages the cache AFTER the load already compiled the
|
||||
# blocks; apply_step_cache must arm the fresh hooks itself.
|
||||
# The generation-time toggle engages the cache AFTER the load already compiled the blocks, so
|
||||
# apply_step_cache must arm the fresh hooks itself.
|
||||
_stub_diffusers(monkeypatch)
|
||||
_stub_torch_compile(monkeypatch)
|
||||
block, hook, orig = _hooked_block()
|
||||
|
|
@ -496,8 +492,8 @@ def test_apply_step_cache_arms_compiled_blocks_on_toggle(monkeypatch):
|
|||
|
||||
|
||||
def test_toggle_disable_restores_inners_before_disable(monkeypatch):
|
||||
# remove_hook splices fn_ref.original_forward back into module.forward, so the
|
||||
# compiled wrapper must be swapped out BEFORE disable_cache runs.
|
||||
# remove_hook splices fn_ref.original_forward back into module.forward, so the compiled wrapper
|
||||
# must be swapped out BEFORE disable_cache runs.
|
||||
_stub_diffusers(monkeypatch)
|
||||
order = []
|
||||
|
||||
|
|
@ -518,8 +514,8 @@ def test_toggle_disable_restores_inners_before_disable(monkeypatch):
|
|||
|
||||
|
||||
def test_enable_failure_restores_inners_before_partial_disable(monkeypatch):
|
||||
# enable_cache can fail after hooking (and arming) some blocks; the partial-hook
|
||||
# cleanup must un-arm them before disable_cache splices original_forward back.
|
||||
# enable_cache can fail after hooking (and arming) some blocks; the partial-hook cleanup must
|
||||
# un-arm them before disable_cache splices original_forward back.
|
||||
_stub_diffusers(monkeypatch)
|
||||
order = []
|
||||
|
||||
|
|
@ -544,9 +540,9 @@ def test_enable_failure_restores_inners_before_partial_disable(monkeypatch):
|
|||
|
||||
|
||||
def test_enable_invalidates_stale_child_registry_cache(monkeypatch):
|
||||
# diffusers 0.39 caches the child-registry list on first cache_context use; an
|
||||
# UNCACHED generation already populates it (empty), so a later toggle-time
|
||||
# enable_cache would install hooks the context never reaches ("No context is set").
|
||||
# diffusers 0.39 caches the child-registry list on first cache_context use, and an UNCACHED
|
||||
# generation already populates it (empty), so a later toggle-time enable_cache would install hooks
|
||||
# the context never reaches ("No context is set").
|
||||
_stub_diffusers(monkeypatch)
|
||||
t = _MixinTransformer()
|
||||
t._diffusers_hook = types.SimpleNamespace(_child_registries_cache = ["stale"])
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue