Add a ModelAttentionBackend node to manually select the attention for models in the workflows. Currently supports pytorch attention or comfy kitchen attention.
Add --use-ck-attention to enable comfy kitchen attention as the default attention backend for all models (might break some).
VAEDecode unwraps a NestedTensor latent (video/audio pair) to its
video component before calling vae.decode(). VAEDecodeTiled skipped
this unwrap and passed the NestedTensor straight into
vae.decode_tiled(), which fails deep in the MiniMax H3 video VAE when
a real tensor's .to() is called with the NestedTensor as an argument.
Fixes#15468.
* Implement tags_all/tags_any/tags_none on the assets list API (BE-6600)
Adds the three canonically-named tag filter params to GET /api/assets and
GET /api/assets/tags/refine:
- tags_all: asset carries every tag (replaces include_tags)
- tags_any: asset carries at least one tag (new)
- tags_none: asset carries no tag (replaces exclude_tags)
Clauses intersect; tags_none always wins. include_tags/exclude_tags remain
as permanent deprecated aliases and behave exactly as before when used on
their own.
Invalid combinations return 400 INVALID_TAG_FILTER, but only when the
request uses at least one new-name parameter (non-empty after
normalisation):
- mixed spellings of one slot (include_tags with tags_all, exclude_tags
with tags_none)
- the same tag in the effective all-list and none-list (query can never
match)
Old-names-only requests gain no new error paths: include_tags=a&exclude_tags=a
still returns an empty 200. tags_any/tags_none overlap stays valid (dead
term, not a dead query).
* Address review findings: positional-compat, deprecation metadata, test matrix
- Move any_tags to the end of the four touched signatures: inserting it
mid-signature silently misbound pre-existing positional callers (e.g.
a caller passing name_contains positionally would have it consumed as
any_tags).
- Mark include_tags/exclude_tags Field(deprecated=True) on both list
schemas so generated schema metadata matches the contract, not just a
comment (schemas_out.py already uses this form for Asset.name).
- Add tests: legal cross-slot old/new combinations, repeated query-key
concatenation (pins Core behavior; outside the cross-platform
contract), tags_any two-page cursor consistency (total/has_more/
no-overlap), refine-route mixed-spelling rejection + legacy-conflict
preservation, and schema deprecation metadata.
* Pin tag-value opacity: case-sensitive matching, byte-exact conflict check
The prod tag survey (~/comfy/prod-model-tag-shape.md) found live
case-distinct tag pairs (SEEDVR2/seedvr2) that resolve differently, so
the contract now states tag values are opaque byte-strings. Pin that:
case-distinct tags filter separately, and a case-distinct all/none pair
is not an INVALID_TAG_FILTER conflict.
* Document tags_all/tags_any/tags_none in openapi.yaml, deprecate aliases
Add the three tag-filter parameters to both listAssets and
getAssetTagHistogram parameter blocks and mark include_tags/exclude_tags
deprecated: true, keeping the spec in step with the runtime schemas so
generated clients can discover the new filters while the aliases stay
present for existing consumers.
* Move schemas_in import to module scope in test_list_filter
Review feedback: no import cycle requires the local import.
* Silence per-request DeprecationWarning in the tag-filter remap shim
Reading the deprecated include_tags/exclude_tags fields by attribute
fires pydantic's DeprecationWarning on every list/refine request even
for callers using only the new names. The warning is aimed at API
clients, not the server's own remap; read via model_dump instead.
* Cap tag-filter lists at 100 entries, all spellings
Review finding: unbounded tag lists fan out into one correlated EXISTS
per tag on both page and count statements. Cap each list at 100
normalized entries with 400 INVALID_TAG_FILTER naming the parameter.
Applies to the legacy spellings as well — a deliberate, decided
exception to the old-names-behave-identically rule, since a cap only on
new names would leave the same fan-out reachable through the aliases.
* Strip process narration from comments
Comments carried decision dates, contract cross-references, and review
context. Keep only the constraints the code cannot show, one line each.
Mirrors the decode fix: encode_tiled is identical to encode for
handles_tiling VAEs, so the OOM-retry path needs to free other models'
memory the same way decode's retry path already does.
For VAEs that set handles_tiling (MiniMax H3 video, SeedVR2), the tiled
decode fallback ends up calling the exact same decode path as the
regular decode that just OOM'd, so retrying without changing anything
about available memory just fails identically. Free memory held by
other loaded models (including dynamically streamed weights) before
the retry, keeping only the VAE itself loaded, so the retry actually
has a chance to succeed.
chelsealong
comfyanonymous
Alexander Brown
Alex Harper
Christian Byrne
Robin Huang
Daxiong (Lin)
Jukka Seppänen
comfyanonymous
Alexander Piskun
Alexis Rolland
Simon Pinfold
blepping
chaObserv
claude[bot]
Terry Jia
Comfy Org PR Bot