Qwen3.5 prefix_caching#2256

Open

wenbinc-Bin wants to merge 84 commits into

HabanaAI:aice/v1.22.0from

wenbinc-Bin:qwen3.5_updatePR_prefix_v2

Commits on Feb 11, 2026

Qwen3.5 first patch
wenbinc-Bin
committed

Commits on Feb 12, 2026

Commits on Feb 13, 2026

Fix quant bug.
wenbinc-Bin
committed

Commits on Feb 17, 2026

Commits on Mar 6, 2026

Port Qwen3.5 Configs
Jing
committed

Commits on Mar 16, 2026

Optimize RMSNormGate and GDN

Wei-Lin-Intel
authored and
wenbinc-Bin
committed
Optimize tril in matrix inverse loop

Wei-Lin-Intel
authored and
wenbinc-Bin
committed
Optimize add in for loop

Wei-Lin-Intel
authored and
wenbinc-Bin
committed
Fix model config error
wenbinc-Bin
committed

Commits on Mar 18, 2026

Commits on Mar 20, 2026

reformat v1
Wei-Lin-Intel
committed

Commits on Mar 22, 2026

reformat v2
Wei-Lin-Intel
committed

Commits on Mar 25, 2026

Commits on Mar 31, 2026

support qwen3.5 reasoning and tool parser

huiqiwa
authored and
wenbinc-Bin
committed

Commits on Apr 1, 2026

Fix incorrect updates introduced in fd4ba5f

Jing
authored and
Jing1Ling
committed
Enable chunked prefill for qwen3.5

Jing
authored and
Jing1Ling
committed
fix kwargswq type

Jing
authored and
Jing1Ling
committed
Fix wrong dtype in GDN and add patch for chunked prefill

Wei-Lin-Intel
authored and
Jing1Ling
committed
Fix decode warmup issue

Wei-Lin-Intel
authored and
Jing1Ling
committed
Fix padding issue and t.compile error of torch.index_copy_ on Qwen3.5

Wei-Lin-Intel
authored and
Jing1Ling
committed
Fix qwen3-next GDN dtype

Wei-Lin-Intel
authored and
Jing1Ling
committed
Reduce torch.compile warmup recompilations for Qwen3.5

Wei-Lin-Intel
authored and
Jing1Ling
committed
Revert "fix kwargswq type"
Jing1Ling
committed

Commits on Apr 3, 2026

Commits on Apr 8, 2026

Revert a WA caused by transformer version conflict
wenbinc-Bin
committed

Commits on Apr 10, 2026