Subcache nodes (from node expansion) now participate in external
provider store/lookup. Previously skipped to avoid duplicates, but
the cost of missing partial-expansion cache hits outweighs redundant
stores — especially with looping behavior on the horizon.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
CacheContext is imported from _caching and re-exported for use by
caching.py. Add noqa comment to satisfy the linter.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Address review feedback from guill:
- Rename _contains_nan to _contains_self_unequal, use not (x == x)
instead of math.isnan to catch any self-unequal value
- Remove Unhashable and repr() fallbacks from _canonicalize; raise
ValueError for unknown types so _serialize_cache_key returns None
and external caching is skipped (fail-closed)
- Update tests for renamed function and new fail-closed behavior
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Address review feedback:
- Move CacheProvider/CacheContext/CacheValue definitions to
comfy_api/latest/_caching.py (source of truth for public API)
- comfy_execution/cache_provider.py re-exports types from there
- Build _providers_snapshot eagerly on register/unregister instead
of lazy memoization in _get_cache_providers
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add docstring with usage example to Caching class matching the
convention used by sibling APIs (Execution.set_progress, ComfyExtension)
- Remove non-deterministic pickle fallback from _serialize_cache_key;
return None on JSON failure instead of producing unretrievable hashes
- Move cache_provider imports to top of execution.py (no circular dep)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Strip verbose docstrings and section banners to match existing minimal
documentation style used throughout the codebase.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Make all docstrings and comments generic for the OSS codebase.
Remove references to Kubernetes, Redis, GCS, pods, and other
infrastructure-specific terminology.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ExecutionList in graph.py calls output_cache.get() and .set() from
sync methods (is_cached, cache_link, get_cache). These cannot await
the now-async get/set. Add get_local/set_local that bypass external
providers and only access the local dict — which is all graph
traversal needs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The lifecycle notification method was importing the old non-prefixed
names (has_cache_providers, get_cache_providers, logger) which no
longer exist after the API cleanup.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add try/except to _build_context, return None when hash fails
- Return None from _serialize_cache_key on total failure (no id()-based fallback)
- Replace hex-like test literal with non-secret placeholder
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove unused CacheContext and _serialize_cache_key imports from
caching.py (now handled by _build_context helper)
- Update test_cache_provider.py to use _-prefixed internal names
- Update tests for new CacheContext.cache_key_hash field (str)
- Make MockCacheProvider methods async to match ABC
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Make on_lookup/on_store async on CacheProvider ABC
- Simplify CacheContext: replace cache_key + cache_key_bytes with
cache_key_hash (str hex digest)
- Make registry/utility functions internal (_prefix)
- Trim comfy_api.latest.Caching exports to core API only
- Make cache get/set async throughout caching.py hierarchy
- Use asyncio.create_task for fire-and-forget on_store
- Add NaN gating before provider calls in Core
- Add await to 5 cache call sites in execution.py
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* model_management: Remove non-comfy dynamic _v caster
* Force pre-load non-comfy weights to GPU in ModelPatcherDynamic
Non-comfy weights may expect to be pre-cast to the target
device without in-model casting. Previously they were allocated in
the vbar with _v which required the _v fault path in cast_to.
Instead, back up the original CPU weight and move it directly to GPU
at load time.
* draft zeta (z-image pixel space)
* revert gitignore
* model loaded and able to run however vector direction still wrong tho
* flip the vector direction to original again this time
* Move wrongly positioned Z image pixel space class
* inherit Radiance LatentFormat class
* Fix parameters in classes for Zeta x0 dino
* remove arbitrary nn.init instances
* Remove unused import of lru_cache
---------
Co-authored-by: silveroxides <ishimarukaito@gmail.com>
Comfy Aimdo 0.2.4 fixes a VRAM buffer alignment issue that happens in
someworkflows where action is able to bypass the pytorch allocator
and go straight to the cuda hook.
This was previously considering the pool of dynamic models as one giant
entity for the sake of smart memory, but that isnt really the useful
or what a user would reasonably expect. Make Dynamic VRAM properly purge
its models just like the old --disable-smart-memory but conditioning
the dynamic-for-dynamic bypass on smart memory.
Re-enable dynamic smart memory.
Multi-step samplers (eg. dpmpp_2s_ancestral) call the model at intermediate sigma values not present in the schedule. This caused set_step to crash with "No sample_sigmas matched current timestep" when context windows were enabled.
The fix is to keep self._step from the last exact match when a substep sigma is encountered, since substeps are still logically part of their parent step and should use the same context windows.
Co-authored-by: ozbayb <17261091+ozbayb@users.noreply.github.com>
* sd: add support for clip model reconstruction
* nodes: SetClipHooks: Demote the dynamic model patcher
* mp: Make dynamic_disable more robust
The backup need to not be cloned. In addition add a delegate object
to ModelPatcherDynamic so that non-cloning code can do
ModelPatcherDynamic demotion
* sampler_helpers: Demote to non-dynamic model patcher when hooking
* code rabbit review comments
Allow non QuantizedTensor layer to set want_requant to get the post lora
calculation stochastic cast down to the original input dtype.
This is then used by the legacy fp8 Linear implementation to set the
compute_dtype to the preferred lora dtype but then want_requant it back
down to fp8.
This fixes the issue with --fast fp8_matrix_mult is combined with
--fast dynamic_vram which doing a lora on an fp8_ non QT model.
Users should either use the cu126 one or the regular one (cu130 at the moment)
The cu128 portable is still included in the latest github release but I will stop including it as soon as it becomes slightly annoying to deal with. This might happen as soon as next week.
