Skip to content

Created by mistake - #47026

Closed
JackChuang wants to merge 1 commit into
sgl-project:mainfrom
bytedance-iaas:horenc/fix-glm53-norope-slot0
Closed

JackChuang wants to merge 1 commit into
sgl-project:mainfrom
bytedance-iaas:horenc/fix-glm53-norope-slot0

Conversation

@JackChuang

@JackChuang JackChuang commented Oct 28, 2026 •

Copy link
Copy Markdown
Contributor

Created by mistake

… writer

set_mla_kv_buffer_kernel_norope wrote every loc, including the padding
slot 0 that the RoPE writers skip since sgl-project#36003. Padding rows can carry
NaN K since sgl-project#42689, and the FP8 TileLang sparse decode reads row 0 for
unused top-k entries, so models with qk_rope_head_dim = 0 (GLM-5.3-Flash)
decoded NaN logits with an FP8 KV cache. Skip the slot in the kernel and
in the torch-native fallback, and run the reserved-slot tests with
rope_dim = 0.
@JackChuang JackChuang closed this Oct 28, 2026
@JackChuang JackChuang changed the title [Fix] kvcache: Skip the reserved padding slot in the zero-RoPE MLA KV writer SGL#47013 Created by mistake Oct 28, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant