Retirement work item for the governed standards exception in collector/evidence_exceptions.yaml (expires 2026-10-31; see PR #1486 review S2).
Default rc23 serving on SM100/103 selects mHC.fused_hc (previous site post_mapping + next site pre_mapping in one autotuned pipeline), while the shipped mhc_module_perf tables measure the standalone pre/post decomposition (kernel_source: trtllm_mhc_pre_* / trtllm_mhc_post_mapping). The approximation is conservative (over-charges the boundary) with a ceiling of mHC's share of DSV4 layer cost (9-14% on b200); the actual Blackwell fused/unfused delta is unmeasured.
Scope of this issue:
- collect
fused_hc on SM100/103 at the pinned rc23 runtime (new table or variant column — producer+consumer contract change, owner approval required)
- switch the SDK DSV4 model accounting to bill the fused boundary on Blackwell without double-counting normalization or pre/post
- regenerate provenance and affected support-matrix rows
- delete the standards exception entry
🤖 Generated with Claude Code
Retirement work item for the governed standards exception in
collector/evidence_exceptions.yaml(expires 2026-10-31; see PR #1486 review S2).Default rc23 serving on SM100/103 selects
mHC.fused_hc(previous site post_mapping + next site pre_mapping in one autotuned pipeline), while the shippedmhc_module_perftables measure the standalone pre/post decomposition (kernel_source: trtllm_mhc_pre_* / trtllm_mhc_post_mapping). The approximation is conservative (over-charges the boundary) with a ceiling of mHC's share of DSV4 layer cost (9-14% on b200); the actual Blackwell fused/unfused delta is unmeasured.Scope of this issue:
fused_hcon SM100/103 at the pinned rc23 runtime (new table or variant column — producer+consumer contract change, owner approval required)🤖 Generated with Claude Code