Skip to content

fix(ds4): retain wide heterogeneous feature capture - #577

Merged
davide221 merged 1 commit into
mainfrom
codex/generic-heterogeneous-stage-planner
Aug 13, 2026
Merged

fix(ds4): retain wide heterogeneous feature capture#577
davide221 merged 1 commit into
mainfrom
codex/generic-heterogeneous-stage-planner

Conversation

@davide221

@davide221 davide221 commented Aug 3, 2026

Copy link
Copy Markdown
Contributor

Summary

  • keep batched DSpark feature capture enabled during sparse heterogeneous prefill
  • preserve required capture boundaries for exact and unsupported hybrid paths
  • cover the supported and fallback cases with focused unit tests

This PR is stacked on #575 and contains one focused change: 41 additions and 7 deletions across three files.

Validation

  • test_deepseek4_unit: passed
  • sparse heterogeneous context qualification: passed
  • deterministic response hashes: matched
  • local strict syntax and diff checks: passed

@cubic-dev-ai cubic-dev-ai Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

All reported issues were addressed across 35 files

Reply with feedback, questions, or to request a fix.

Re-trigger cubic

Comment thread server/src/deepseek4/deepseek4_graph.cpp Outdated
Comment thread server/deps/llama.cpp/ggml/src/ggml-cuda/fattn.cu
Comment thread server/src/common/heterogeneous_stage_planner.cpp Outdated
Comment thread server/deps/llama.cpp/ggml/include/ggml.h Outdated
Comment thread server/scripts/qualify_ds4_q5_r9700_strix.sh Outdated
Comment thread server/scripts/qualify_ds4_q5_r9700_strix.sh Outdated
Comment thread server/scripts/analyze_rocprof_overlap.py Outdated
Comment thread server/docs/DS4_R9700_STRIX_PROFILE.md Outdated
Comment thread server/src/common/heterogeneous_stage_planner.cpp Outdated
@davide221
davide221 force-pushed the codex/ds4-overlapped-moe-balance branch from 17597e3 to e5fb838 Compare August 12, 2026 18:20
@davide221
davide221 force-pushed the codex/generic-heterogeneous-stage-planner branch from 1e54c5d to c029f90 Compare August 12, 2026 18:20
@davide221 davide221 changed the title feat(moe): add a generic heterogeneous stage planner fix(ds4): retain wide heterogeneous feature capture Aug 12, 2026
@davide221
davide221 force-pushed the codex/ds4-overlapped-moe-balance branch from e5fb838 to 3240627 Compare August 12, 2026 18:46
@davide221
davide221 force-pushed the codex/generic-heterogeneous-stage-planner branch from c029f90 to a209c34 Compare August 12, 2026 18:46
@davide221
davide221 force-pushed the codex/ds4-overlapped-moe-balance branch from 3240627 to 805d5a7 Compare August 12, 2026 19:36
@davide221
davide221 force-pushed the codex/generic-heterogeneous-stage-planner branch from a209c34 to 3ff79a6 Compare August 12, 2026 19:36
@davide221
davide221 force-pushed the codex/ds4-overlapped-moe-balance branch from 805d5a7 to a9c8e7a Compare August 12, 2026 19:39
@davide221
davide221 force-pushed the codex/generic-heterogeneous-stage-planner branch from 3ff79a6 to a46f877 Compare August 12, 2026 19:39
@davide221
davide221 force-pushed the codex/ds4-overlapped-moe-balance branch from a9c8e7a to bae0822 Compare August 12, 2026 20:11
@davide221
davide221 force-pushed the codex/generic-heterogeneous-stage-planner branch from a46f877 to 5dd1cfc Compare August 12, 2026 20:11
@davide221
davide221 force-pushed the codex/ds4-overlapped-moe-balance branch from bae0822 to c358b9d Compare August 13, 2026 13:02
@davide221
davide221 force-pushed the codex/generic-heterogeneous-stage-planner branch from 5dd1cfc to b69f52e Compare August 13, 2026 13:02
@davide221

Copy link
Copy Markdown
Contributor Author

@cubic-dev-ai review the current head after the branch rebase. Please focus on correctness, maintainability, security, and compatibility with the updated base.

@cubic-dev-ai

cubic-dev-ai Bot commented Aug 13, 2026

Copy link
Copy Markdown
Contributor

@cubic-dev-ai review the current head after the branch rebase. Please focus on correctness, maintainability, security, and compatibility with the updated base.

@davide221 I have started the AI code review. It will take a few minutes to complete.

@cubic-dev-ai cubic-dev-ai Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No issues found across 3 files

Re-trigger cubic

@davide221
davide221 force-pushed the codex/generic-heterogeneous-stage-planner branch from b69f52e to 0f83e60 Compare August 13, 2026 13:55
@davide221

Copy link
Copy Markdown
Contributor Author

@cubic-dev-ai review the current rebased head. Please check correctness, maintainability, security, and interaction with the updated heterogeneous prefill path.

@cubic-dev-ai

cubic-dev-ai Bot commented Aug 13, 2026

Copy link
Copy Markdown
Contributor

@cubic-dev-ai review the current rebased head. Please check correctness, maintainability, security, and interaction with the updated heterogeneous prefill path.

@davide221 I have started the AI code review. It will take a few minutes to complete.

@cubic-dev-ai cubic-dev-ai Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No issues found across 3 files

Re-trigger cubic

@davide221
davide221 force-pushed the codex/ds4-overlapped-moe-balance branch from 50f9f2b to ceaf788 Compare August 13, 2026 14:04
@davide221
davide221 force-pushed the codex/generic-heterogeneous-stage-planner branch from 0f83e60 to 2102a4c Compare August 13, 2026 14:18
@davide221
davide221 changed the base branch from codex/ds4-overlapped-moe-balance to main August 13, 2026 14:18
@davide221
davide221 merged commit 49daa1c into main Aug 13, 2026
1 check passed
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.

1 participant