Enable PCP for GDN models - #3386

Open

wenxindongwork wants to merge 7 commits into

mainfrom

Commits on Sep 6, 2026

Qwen3.5 hybrid GDN end-to-end under PCP
wenxindongwork
committed
17431fb
Fix batched-einsum QuantLinearConfig; address review on _is_attention and quantized matmul rank
wenxindongwork
committed
e672977
Rebase fixup: pcp_vs_tp_benchmark uses renamed PCP axis
wenxindongwork
committed
85d62e4
Remove BATCH axis, use DENSE_DATA (same value) for KV-cache batch dim
wenxindongwork
committed
05de637
Trim and simplify PCP sharding comments per review
wenxindongwork
committed
2250dd9
Apply pre-commit (yapf) line-wrapping after BATCH->DENSE_DATA rename
wenxindongwork
committed
67a0897
VllmQuantLinearConfig: warn + default to replicated for unknown layer types
wenxindongwork
committed
7a3c35d