Script 'mail_helper' called by obssrc
Hello community,
here is the log from the commit of package python-compressed-tensors for
openSUSE:Factory checked in at 2026-10-28 17:47:29
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Comparing /work/SRC/openSUSE:Factory/python-compressed-tensors (Old)
and /work/SRC/openSUSE:Factory/.python-compressed-tensors.new.2252680
(New)
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "python-compressed-tensors"
Wed Oct 28 17:47:29 2026 rev:5 rq:1393178 version:0.20.0
Changes:
--------
---
/work/SRC/openSUSE:Factory/python-compressed-tensors/python-compressed-tensors.changes
2026-09-28 10:47:28.926540425 +0200
+++
/work/SRC/openSUSE:Factory/.python-compressed-tensors.new.2252680/python-compressed-tensors.changes
2026-10-28 17:47:45.016083139 +0100
@@ -1,0 +2,27 @@
+Tue Oct 27 06:31:10 UTC 2026 - Martin Pluskal <[email protected]>
+
+- Update to 0.20.0:
+ * Add MXFP6 (E2M3) quantization with mxfp6-pack-quantized format,
+ and a MXFP4xFP8 quantization scheme
+ * Add layerwise compression and decompression, and an expert
+ sparsity config for MoE models
+ * Add FP8 block dequantizer from_pretrained and fuse the quantize
+ and FP4-pack Triton kernels
+ * Offload: support device_map="disk", reuse safetensors handles
+ via disk_load_context, group subgraph staging reads, and batch
+ metadata exchange in distributed dispatch; fix onloading of
+ offloaded children in set_onload_device
+ * convert_checkpoint gains max_workers="auto" by default and
+ quieter file-copying logs; support MTP checkpoint export
+ * Quantization fixes: NVFP4 on A100, 4-bit unpack by splitting
+ bytes (also on big-endian hosts), contiguous packed INT
+ tensors, qparam init on disk, weight transforms and
+ dense-saving of quantized Linear/Embedding subclasses,
+ router bias pruning fix
+ * Parse CT_ENFORCE_EAGER as a boolean, so =0/=false disables it;
+ skip default-valued fields in QuantizationArgs serialization
+ * Drop Python 3.10: minimum is now 3.11
+ * Many more fixes and improvements; see upstream's release notes
+ for the full list
+
+-------------------------------------------------------------------
Old:
----
compressed_tensors-0.19.0.tar.gz
New:
----
compressed_tensors-0.20.0.tar.gz
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Other differences:
------------------
++++++ python-compressed-tensors.spec ++++++
--- /var/tmp/diff_new_pack.uxCXXO/_old 2026-10-28 17:47:46.001124347 +0100
+++ /var/tmp/diff_new_pack.uxCXXO/_new 2026-10-28 17:47:46.005124514 +0100
@@ -18,7 +18,7 @@
%{?sle15_python_module_pythons}
Name: python-compressed-tensors
-Version: 0.19.0
+Version: 0.20.0
Release: 0
Summary: Library for utilization of compressed safetensors of neural
network models
License: Apache-2.0
@@ -62,7 +62,7 @@
# make this noarch package fail to build on every other architecture.
# Most of the test suite additionally requires a GPU.
-%files %{python_files}
+%files %python_files
%license LICENSE
%doc README.md
%{python_sitelib}/compressed_tensors
++++++ compressed_tensors-0.19.0.tar.gz -> compressed_tensors-0.20.0.tar.gz
++++++
++++ 8335 lines of diff (skipped)
