Skip to content

fix(rocm): distinguish CUDA and ROCm in the MTP policy cache key - #1933

Open
cestercian wants to merge 1 commit into
lablup:mainfrom
cestercian:cursor/fix-mtp-policy-cuda-rocm-key-fcd0
Open

cestercian wants to merge 1 commit into
lablup:mainfrom
cestercian:cursor/fix-mtp-policy-cuda-rocm-key-fcd0

Conversation

@cestercian

Copy link
Copy Markdown
Contributor

Summary

Non-Apple hosts used to share the hardware label `Unknown-0c`, so a CUDA MTP policy verdict was reused on ROCm. `hardware_label()` now prefixes backend identity (`cuda` / `rocm`, plus architecture when present) on non-Apple hosts. Apple labels stay unchanged. Old `Unknown-0c` hint files are not migrated (they record no vendor).

Closes #1887

Test plan

  • `cargo test --profile test-fast --no-default-features -p mlxcel --lib server::batch::mtp_policy` — 51 passed

Non-Apple hosts used to share the hardware label Unknown-0c, so a
CUDA verdict was reused on ROCm. Prefix the backend (and architecture
when present) on non-Apple hosts. Leave the Apple spelling unchanged
and do not migrate old hint files: they record no vendor, so a mapping
would guess.

Closes lablup#1887
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.

fix(rocm): CUDA and ROCm hosts share one MTP policy cache key

1 participant