Script 'mail_helper' called by obssrc
Hello community,
here is the log from the commit of package python-comfy-kitchen for
openSUSE:Factory checked in at 2026-09-29 17:53:26
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Comparing /work/SRC/openSUSE:Factory/python-comfy-kitchen (Old)
and /work/SRC/openSUSE:Factory/.python-comfy-kitchen.new.383539 (New)
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "python-comfy-kitchen"
Tue Sep 29 17:53:26 2026 rev:6 rq:1381374 version:0.2.36
Changes:
--------
---
/work/SRC/openSUSE:Factory/python-comfy-kitchen/python-comfy-kitchen.changes
2026-09-23 14:36:01.536013244 +0200
+++
/work/SRC/openSUSE:Factory/.python-comfy-kitchen.new.383539/python-comfy-kitchen.changes
2026-09-29 17:55:39.953918041 +0200
@@ -1,0 +2,28 @@
+Tue Sep 29 08:27:17 UTC 2026 - Martin Pluskal <[email protected]>
+
+- Update to 0.2.36:
+ * New Huawei Ascend NPU backend. It registers only when torch-npu
+ and an Ascend device are present, so list_backends() reports it
+ unavailable on this build and the import is a no-op without
+ torch-npu
+ * Eager backend: fp16_conv3d and group_norm_silu_pad3d accept an
+ out= destination, and the new fp16_conv3d_out and
+ group_norm_silu_pad3d_out ops write straight into an
+ NDHWC-ordered output or view, so a tiled conv needs no per-tile
+ copies
+ * group_norm_silu_pad3d gains zero_pad= for models whose
+ convolutions zero-pad the spatial border instead of
+ reflecting it
+ * W6A8 grouped weight quantization next to W4A8, in the same
+ packed row layout with the bit depth implied by the row width
+ * The is_available() probes (sol_attn, flash decode, INT8
+ attention, fused DeltaNet) now take a torch.device rather than
+ an int index and report False for a non-CUDA device; a caller
+ passing an integer index breaks
+ * Fused DeltaNet decode now declines below SM75, which the shared
+ memory check alone used to wave through on Volta
+ * Remainder of the release is CUDA and HIP kernel work (int8
+ attention, W4A8 GEMM, 256-wide flash decode), not compiled in
+ this CPU-only build
+
+-------------------------------------------------------------------
Old:
----
comfy-kitchen-0.2.35.tar.gz
New:
----
comfy-kitchen-0.2.36.tar.gz
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Other differences:
------------------
++++++ python-comfy-kitchen.spec ++++++
--- /var/tmp/diff_new_pack.EBfZrs/_old 2026-09-29 17:55:40.877956577 +0200
+++ /var/tmp/diff_new_pack.EBfZrs/_new 2026-09-29 17:55:40.879956660 +0200
@@ -18,7 +18,7 @@
%{?sle15_python_module_pythons}
Name: python-comfy-kitchen
-Version: 0.2.35
+Version: 0.2.36
Release: 0
Summary: Fast kernel library for ComfyUI (CPU/eager backend)
# Legal-Review-Notice: NOTICE names the NVIDIA CUDA runtime EULA for a
++++++ comfy-kitchen-0.2.35.tar.gz -> comfy-kitchen-0.2.36.tar.gz ++++++
++++ 11950 lines of diff (skipped)
