Skip to content

Feature/split moe#1458

Merged
xhuohai merged 94 commits intodev/3.0from
feature/split_MoE
Nov 10, 2025
Merged

Feature/split moe#1458
xhuohai merged 94 commits intodev/3.0from
feature/split_MoE

Conversation

@curioyang
Copy link
Copy Markdown
Contributor

@curioyang curioyang commented Oct 20, 2025

This pull request introduces support for the new SparseExperts custom operation to the NTT module, including its definition, code generation, and evaluation logic. The changes also add utility methods and update service registration to ensure the new operation is properly integrated.

SparseExperts operation integration

  • Added the new SparseExperts custom operation class in CustomOp/CPU/SparseExperts.cs, defining its parameters, properties, and display logic.
  • Implemented the SparseExperts functional API in CustomOp/CPU/Functional.cs, allowing users to create and use the operation in expressions.

Code generation and evaluation

  • Updated KernelCSourceConvertVisitor.cs to handle code generation for SparseExperts and TopK operations, including buffer handling and argument flattening.
  • Added a utility method FlattenTuple to assist with argument processing in code generation.
  • Registered the SparseExpertsEvaluator in the NTT module service configuration to enable runtime evaluation of the new operation.

@curioyang curioyang marked this pull request as ready for review October 28, 2025 02:30
@github-actions
Copy link
Copy Markdown

github-actions bot commented Oct 28, 2025

Test Results

3 321 tests  ±0   3 321 ✅ ±0   2h 18m 59s ⏱️ + 4m 35s
    6 suites ±0       0 💤 ±0 
    6 files   ±0       0 ❌ ±0 

Results for commit 38e3d90. ± Comparison against base commit 23b6ba4.

♻️ This comment has been updated with latest results.

@xhuohai xhuohai merged commit bb35bab into dev/3.0 Nov 10, 2025
25 checks passed
@xhuohai xhuohai deleted the feature/split_MoE branch November 10, 2025 08:42
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.

4 participants