[Model] Support uneven TP3 Qwen3.5/Qwen3Next layouts#47759

Draft

s0l wants to merge 13 commits into

vllm-project:mainfrom

s0l:exp/tp3-dcp-gdn-128k

Commits on Jul 10, 2026

Fix CUDA graph profiling KV cache sizing

s0l
and
Codex
committed
Support uneven TP3 Qwen3.5 attention partitions

s0l
and
Codex
committed
Handle packed weights in padded TP shards

s0l
and
Codex
committed
Align Mamba prefill chunks to DCP block geometry

s0l
and
Codex
committed
Honor language-model-only for Qwen3.5

s0l
and
Codex
committed
Support uneven TP dense Qwen MLP partitions
s0l
committed
Fix context parallel graph profiling import
s0l
committed
Support non-power-of-two context parallel correction
s0l
committed