[Cherry-pick] PRs #2022 #2026 #2032 #1981 #2010 #2043 #2042 #2038 #2050 #2041 #2031 #2061 #1983 #1628 - #2062

Merged

kevalmorabia97 merged 15 commits into

release/0.46.0from

cherry-picks/release-0.46.0

Aug 4, 2026

Commits on Aug 4, 2026

Update documentation guide for ONNX INT4 PTQ on Windows cuda13 host (#2022)

vishalpandya1990
authored and
kevalmorabia97
committed
2485f20
tools/mcp: pin mcp<2 to fix unit CI collection (#2026)

authored and
kevalmorabia97
committed
dda0ca0
fix(export): [NVBug 6525534] preserve nested VLM namespaces (#2032)

meenchen
authored and
kevalmorabia97
committed
51eb1cb
[5726458] [Experimental] Add NVFP4 projection-output-quantizer recipe and HF embedding ONNX export example (#1981)

authored and
kevalmorabia97
committed
7ef683e
Add ModelOpt QAD skill for Slurm workflows (#2010)

mxinO
authored and
kevalmorabia97
committed
b2cf4f6
Skip zero-quota splits in get_dataset_samples (#2043)

kevalmorabia97
and
claude
committed
8c12de8
[skill] evaluation: correct the stale vLLM CUDA-13 image tag rule (#2042)

authored and
kevalmorabia97
committed
bd5f55a
[NVBug: 5987078] Fix unified HF export of compressed NVFP4 weights (--low_memory_mode) (#2038)

authored and
kevalmorabia97
committed
d1a4811
Bump transformers dependency to >=4.57,<5.15 (#2050)
kevalmorabia97
committed
de559f1
[NVBug: 6524370] use sequential device_map for DiffusionGemma (#2041)

authored and
kevalmorabia97
committed
cb2911c
Add Megatron-Bridge prune & quantize launcher pipelines (#2031)

kevalmorabia97
and
claude
committed
90b5b95
Fix saving pruned Nemotron-3-Nano hybrid_override_pattern with MTP or Pipe symbols (#2061)

kevalmorabia97
and
claude
committed
bd856c7
[6425069][ONNX][Autocast] Fix autocast metadata propagation (#1983)

gcunhase
authored and
kevalmorabia97
committed
8cefee9
[6058841] Fix inconsistent tensor types on control-flow (If/Loop/Scan) subgraphs during FP16/BF16 conversion (#1628)

authored and
kevalmorabia97
committed
a943e31
Remove empty 0.47 section from CHANGELOG on the 0.46 release branch
kevalmorabia97
committed
5bbc2d2

Navigation Menu

Uh oh!

[Cherry-pick] PRs #2022 #2026 #2032 #1981 #2010 #2043 #2042 #2038 #2050 #2041 #2031 #2061 #1983 #1628 - #2062

[Cherry-pick] PRs #2022 #2026 #2032 #1981 #2010 #2043 #2042 #2038 #2050 #2041 #2031 #2061 #1983 #1628#2062
kevalmorabia97 merged 15 commits into
release/0.46.0from
cherry-picks/release-0.46.0

Commits on Aug 4, 2026

Update documentation guide for ONNX INT4 PTQ on Windows cuda13 host (#2022)

tools/mcp: pin mcp<2 to fix unit CI collection (#2026)

fix(export): [NVBug 6525534] preserve nested VLM namespaces (#2032)

[5726458] [Experimental] Add NVFP4 projection-output-quantizer recipe and HF embedding ONNX export example (#1981)

Add ModelOpt QAD skill for Slurm workflows (#2010)

Skip zero-quota splits in get_dataset_samples (#2043)

[skill] evaluation: correct the stale vLLM CUDA-13 image tag rule (#2042)

[NVBug: 5987078] Fix unified HF export of compressed NVFP4 weights (--low_memory_mode) (#2038)

Bump transformers dependency to >=4.57,<5.15 (#2050)

[NVBug: 6524370] use sequential device_map for DiffusionGemma (#2041)

Add Megatron-Bridge prune & quantize launcher pipelines (#2031)

Fix saving pruned Nemotron-3-Nano hybrid_override_pattern with MTP or Pipe symbols (#2061)

[6425069][ONNX][Autocast] Fix autocast metadata propagation (#1983)

[6058841] Fix inconsistent tensor types on control-flow (If/Loop/Scan) subgraphs during FP16/BF16 conversion (#1628)

Remove empty 0.47 section from CHANGELOG on the 0.46 release branch