Skip to content

fix: retain select last token InfiniOps adapter - #1538

Merged
voltjia merged 1 commit into
mainfrom
fix/select-last-token-infiniops-selection
Aug 26, 2026
Merged

fix: retain select last token InfiniOps adapter#1538
voltjia merged 1 commit into
mainfrom
fix/select-last-token-infiniops-selection

Conversation

@voltjia

@voltjia voltjia commented Aug 26, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • declare add and index_select as the InfiniOps dependencies of the composed select_last_token_hidden adapter
  • retain that adapter when both aligned operators are selected through ops.json

Impact

Without an explicit dependency mapping, adapter filtering treats select_last_token_hidden as a direct InfiniOps operator. Since the aligned implementation is composed from add and index_select, the filter removes it and InfiniCore falls back to the legacy InfiniOp implementation even when both required InfiniOps operators are available.

This change only corrects source selection. It does not modify the adapter implementation, operator interfaces, or submodules.

Validation

  • git diff --check
  • validated from InfiniRT 6b256e0ab085, InfiniOps c8e15ebe9da4, InfiniCore 847c4c04fe18 plus this change, and InfiniLM 80bb09ecebc9
  • built and installed InfiniRT and InfiniOps independently, then configured InfiniCore as an installed-package consumer on NVIDIA and MetaX
  • ran Qwen3-0.6B paged-attention inference with greedy decoding on both platforms
  • NVIDIA temporary route instrumentation reported only SELECT_ROUTE=infiniops for select_last_token_hidden
  • MetaX GDB route counts reported select_infiniops=1, select_infiniop=0, and random_infiniop=0; aligned Argmax and linked flash-attention providers were exercised

@voltjia
voltjia requested a review from a team August 26, 2026 08:16
@voltjia
voltjia merged commit bb6aa58 into main Aug 26, 2026
10 checks passed
@voltjia
voltjia deleted the fix/select-last-token-infiniops-selection branch August 26, 2026 08:28
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants