[FIX_FOR_VLLM_CUSTOM=3975eb6de6ea914b9d7b27fd517e0c971ddeb6fc] Fix upstream breakages: NIXL connector, TpKVTopology rename, MoE refactor, transformers v5#1377

Merged

iboiko-habana merged 11 commits into

vllm-project:mainfrom

pawel-olejniczak:fix/vllm-hourly-tpkvtopo-rename

Apr 29, 2026

Commits on Apr 28, 2026

[FIX_FOR_VLLM_CUSTOM=b47840019e61a3983c8144066a99c843d177947d] Fix TpKVTopology rename to TransferTopology and update offloading tests
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=b47840019e] Fix FusedMoE reduce_results removal and forward_dispatch rename (upstream PR #35949)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=b47840019e] Fix DefaultMoERunner method renames (forward_entry, maybe_reduce_output)
pawel-olejniczak
committed
Import transformers to avoid partial-init races
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=4b7869d6bc] Fix MoERunner import rename and remove double TP reduction in MoE models
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=b47840019e] Fix MoE runner import: use MoERunnerBase from moe_runner_base
pawel-olejniczak
committed
Update transformers to v5
pawel-olejniczak
committed
Revert "Import transformers to avoid partial-init races"
pawel-olejniczak
committed
Fix lm-eval GSM8K accuracy 0.0 with transformers v5
pawel-olejniczak
committed
Fix patch nixl_utils for HPU platform to use nixl._api instead of rixl._api
pawel-olejniczak
committed
Exclude transformers 5.6.* to fix DeepSeek-V2-Lite tokenizer regression
pawel-olejniczak
committed