* Move and adaptation of GDN opt from Qwen3next to Qwen3.5 (sgl-project#215) * rm 3*copy after casual conv, do q/k/v split in causal_conv1d_update_split_qkv kernel during decode time * Add fused Sigmoid and mul triton kernel * add fused gdn gating and sigmoid func for prefill phase * Add fused causal conv1d fwd split qkv kernel * Add fused cumsum_kkt, solve_tril, and merge_recompute kernels for prefill optimization * Increase parallel params for chunk_gated_delta_rule_fwd_kernel * add chunk o kernel opt * Add sigmoid and mul for shared expert module --------- Co-authored-by: yixionghuo <66975556+yixionghuo@users.noreply.github.com> * [Fix] Load weights error in compressed-tensors * Fuse 4 GDN input projections (qkv/z/b/a) into single GEMM (sgl-project#219) * Revert "Fuse 4 GDN input projections (qkv/z/b/a) into single GEMM (sgl-project#219)" (sgl-project#223) This reverts commit 17b0749. * Fuse 4 GDN input projections (qkv/z/b/a) into single GEMM (sgl-project#225) * [Fix] Core dump issue in stress test --------- Co-authored-by: zachYao <zayao@amd.com> Co-authored-by: root <root@hjbog-srdc-29.amd.com>

64% faster awq #229

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions