Skip to content

Document decoder residual paths and dead RoPE values - #12

Draft
rui-ren with Copilot wants to merge 1 commit into
mainfrom
copilot/ruirentensorrt-mobius
Draft

Document decoder residual paths and dead RoPE values#12
rui-ren with Copilot wants to merge 1 commit into
mainfrom
copilot/ruirentensorrt-mobius

Conversation

Copilot AI commented Sep 6, 2026

Copy link
Copy Markdown

Documents the decoder graph topology, including producer-consumer edges, residual branches, merge points, and fused-GQA dead-value handling.

  • Residual paths

    • Traces both pre-norm attention and MLP residual branches.
    • Identifies the two Add merge points validated by the existing structural test.
  • Dead-value analysis

    • Explains why discarded RoPE Gather outputs are dead.
    • Shows how direct cos_cache and sin_cache consumption avoids unnecessary graph nodes.
hidden → RMSNorm → Attention ─┐
   └──────── residual ────────┴→ Add
                               ↓
          RMSNorm → MLP ──────┐
   └──────── residual ────────┴→ Add → output

No Mobius source files are modified; the deliverable is the issue graph trace.

Copilot AI changed the title [WIP] Add topological graph tracing for ONNX Document decoder residual paths and dead RoPE values Sep 6, 2026
Copilot AI requested a review from rui-ren September 6, 2026 17:01
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.

ONNX for TensorRT — Day 5: Topological Graph Tracing, Residual Paths, and Dead Values

2 participants