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).
The new-format branch of convert_old_quants() writes "{layer_key}.comfy_quant"
markers straight from the checkpoint's _quantization_metadata JSON, ignoring
model_prefix entirely. Checkpoint metadata stores layer keys either with the
full diffusion-model prefix or already stripped of it, and convert_old_quants()
is invoked before and/or after that prefix is stripped from state_dict
(comfy/sd.py calls it up to twice around the strip). A fixed assumption about
which convention is in play silently mismatches the other, so affected layers
never get wrapped in a QuantizedTensor and fall back to plain-dtype storage
(observed as VRAM blowup / manual-cast warnings on NVFP4 checkpoints such as
LTX-Video 2.3).
PR #13328 tried to fix this by reordering comfy/sd.py's calls around the
prefix strip, but that only swaps which convention works and was closed by
the maintainer for risking regressions on checkpoints using the other
convention. This instead makes convert_old_quants() match each layer key
against the state_dict's real key first (zero behavior change when metadata
is already aligned with state_dict), then try adding/stripping model_prefix,
and only fall back to today's blind write when neither matches. No call
sites in comfy/sd.py are touched or reordered. The write is now also
idempotent, guarding against the double-call case rewriting a resolved key.
Adds tests-unit/comfy_quant/test_convert_old_quants_prefix.py covering both
metadata conventions, the legacy scaled_fp8 branch (unaffected), the exact
two-call dance from load_diffusion_model_state_dict, and a documented (xfail)
residual gap: when a checkpoint's real weight keys carry no prefix at all but
its metadata keys do, comfy/sd.py's model_prefix="" call-site argument still
lets model_detection.unet_prefix_from_state_dict get poisoned by the
resulting spurious marker keys. That is pre-existing (reproduces identically
without this patch) and out of scope here since fixing it needs a
comfy/sd.py or model_detection.py change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This avoids name collision (circular imports) for external custom nodes,
for which the comfy path is pushed into sys.path so Python's own logging module
is shadowed otherwise.
fixes: #15229
This priority scheme was broken in the case where you have pin
registration exhaustion while loading a VBAR that gets a big evicition.
The weight would stay in the loaded set but inherit the MRU priority
against other workflow models WRT pin registration which leads to async
offload without pinning.
Fix by universally promiting active pin registration above workflow
pins without concern for the weights/weights-loaded split. This diverges
from the actual budgeting where the split still makes sense.
Changes:
Remove sequential scan hint
Prefer NVML pressure on windows
Add async malloc clamp option (unused by comfy so far)
Workaround AMD windows GPU virtual address space leak
The largest change is the NVML pressure, which works around a cuMemGetInfo
drift from actual VRAM in some circumstances.
Windows has proven this logic works for a long time and there are
corner cases where this materialization actual consumes real RAM
on linux.
Its not as bad as the original windows commit charge surge, but
its still a detectable transient leak. So simplify and unify.
Alan Li
Jukka Seppänen
Barish Ozbay
comfyanonymous
rattus
Alexander Brown
Alex Harper
Alexis Rolland
blepping
endman100
rivadart
Oliver Freyermuth