[FIX_FOR_VLLM_CUSTOM=5b39b268f506150dbab38f6f6c04b7c843e37c07] Fix upstream regressions: MoE refactor, DeepSeek V4 router, KV offload HMA#1403

Merged

iboiko-habana merged 10 commits into

vllm-project:mainfrom

pawel-olejniczak:fix/vllm-hourly-29-4

May 6, 2026

Commits on May 5, 2026

[FIX_FOR_VLLM_CUSTOM=2c06cf3486] Fix MoE runner import rename and LoRA classes tuple (upstream #40560, #35077)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=4d51588e23] Fix FusedMoE test helper for DeepSeek V4 router API (upstream #40860)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=8a8c9b564e] Sync offloading test utils with upstream per-job store completion (upstream #39186, #39403)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=4d51588e23] Add hash_indices_table to HPU create_fused_moe_router (upstream #40860)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=809d83c2dc] Fix MoE test calls: _forward_dispatch → forward (upstream #40560)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=8a8c9b564e] Adapt to OffloadingConnectorMetadata per-job API (upstream #39186)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=a250f1bd5f] Implement add_shrink and add_expand in HPU punica wrapper (upstream #35077)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=809d83c2dc] Fix INC FP8 DeepSeek V2 crash after MoE runner refactor (upstream #40560)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=626daa2076] Accept synthetic_mode kwarg in HPU rejection_sample (upstream #40662)
pawel-olejniczak
committed
[FIX_FOR_VLLM_CUSTOM=4d51588e23] Pass input_ids through patched MoE forward calls (upstream #40860)
pawel-olejniczak
committed