Skip to content

docs(serving): correct gather NaN root cause + record the hardening - #2148

Merged
joelteply merged 1 commit into
canaryfrom
docs/gather-root-cause-and-hardening
Aug 4, 2026
Merged

docs(serving): correct gather NaN root cause + record the hardening#2148
joelteply merged 1 commit into
canaryfrom
docs/gather-root-cause-and-hardening

Conversation

@joelteply

Copy link
Copy Markdown
Contributor

supports_op/dispatch mismatch was the real cause; alignment + fence work lands as hardening with measured defaults.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo

…ch mismatch, not addressing

The doc credited the alignment invariant; the actual 5090 root cause was a gate
that accepted shapes CUDA dispatches to MMQ (no gather), which reads the
contiguous staging the consume-arm intentionally leaves unpopulated for aliased
experts. Records the generalizable rule (supports_op must mirror the dispatch
predicate exactly) and what the hunt bought as hardening: the identity-table
bisect instrument, fence precision, and opt-in event retirement with the
measured reason it is not the default.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo
@joelteply
joelteply merged commit ea125d0 into canary Aug 4, 2026
2 checks passed
@joelteply
joelteply deleted the docs/gather-root-cause-and-hardening branch August 4, 2026 03:29
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant