Radeon AI PRO R9700 (gfx1201, RDNA4) support #1781

Activity

1. mehdi-saeedi commented on Oct 8, 2026

   

   mehdi-saeedi

   on Oct 8, 2026

   CollaboratorAuthor

   More actions

   Review round addressed. @chaojhou reviewed every Hyperloom PR here; all of the must-fix items are now pushed, and the PR descriptions were updated where the behaviour changed.

   • fix(forge): ignore rocminfo's generic ISA names when detecting the GPU target #1773: merged main to clear the campaign_config.py conflict with Read kernelforge workspace state only in the shape the current code writes #1768. Approved, unchanged otherwise.
   • fix(specialists): spawn the Claude CLI the installer recorded #1775: one Claude CLI resolver for every agent path, in roles/claude.py, reading HYPERLOOM_CLAUDE_CLI_PATH, then GEAK_CLAUDE_BIN, then PATH; both variables documented.
   • fix(preflight): expand ${VAR} references in the custom-header settings #1776: header expansion now runs against the same synthesized Anthropic view claude_sdk_env_options uses, through one shared helper, so a reference to the credential that is not set no longer resolves to empty.
   • fix(profile): keep GPU events in the traces of spawned vLLM engines #1777: folded into a single profile start-up shim with one chain loader and one injection; both broad excepts removed.
   • fix(prelude): keep PRELUDE open until the initial roofline books #1779: the hold now covers the ENABLEMENT exit as well as PRELUDE, through one shared predicate helper, with both fields in enablement_baseline. Also merged main for the roofline executor refactor ([Bad smell] refactor(roofline): lift the profile retry loop out of _execute #1702).
   • feat(gpu): support the Radeon AI PRO R9700 (gfx1201) with vLLM #1782: the two MI behaviour changes are listed under Observable effect, the rebaseline test and its claim are dropped, and r9700 is documented in SKILL.md and references/frameworks.md.

   The MAGPIE_REF bump stays a separate follow-up, opened once this round settles: it moves Magpie for every board, while #1782 is additive for all of them, so the two carry different risk.
2. chaojhou commented on Oct 22, 2026

   

   chaojhou

   on Oct 22, 2026

   Collaborator

   More actions

   Every item in this tracking issue has landed on main:

   • R9700 identity, vLLM gfx12 routing and the GEAK product/ISA handoff: feat(gpu): support the Radeon AI PRO R9700 (gfx1201) with vLLM #1782.
   • Bring-up fixes: fix(gemm-tune): keep the TunableOp recording envs in the vLLM shape capture #1772, fix(forge): ignore rocminfo's generic ISA names when detecting the GPU target #1773, fix(install): find the image Python in /opt/python/bin #1774, fix(specialists): spawn the Claude CLI the installer recorded #1775, fix(preflight): expand ${VAR} references in the custom-header settings #1776, fix(profile): keep GPU events in the traces of spawned vLLM engines #1777, fix(prelude): keep PRELUDE open until the initial roofline books #1779, feat(kernel): surface missing kernel-tuning tools at preflight and from GEAK #1780, fix(phases): treat a GEAK timeout as a terminal KERNEL result #1783 and docs(profile): scope the start-up shim's note to the stack it was observed on #1799.
   • GEAK fixes: fix(e2e_workflow): adopt a head's entity_kind from its profiled row GEAK#498, README: own-GPU-machine docker example uses the harbor proxy image instead of the pullable primussafe ref #500, feat(kernel-opt): parallel GEAK/OOB ladders + per-attempt cache isolation; flagship CI hardening #502 and ci: restore optimize-submit manual dispatch #504. These reach installs because GEAK_REF defaults to main.
   • MAGPIE_REF moved to 6a1ee26f in Use Magpie's native AgentX contract for optimization #1922 (install.sh:183). That commit carries the gfx12 runner vllm_gfx12.sh, from Magpie#100, add image-version-check #110 and Long-Term Memory #112.
   • TRACELENS_REF moved to 614c597b in Drive the TraceLens Agent from a YAML run config #1952 (install_kernel_tools.sh:96). That commit contains TraceLens#1096, Capture Platform Config and warn on CPU settings that aren't optimal  #1097 and fix(kernel): harden Forge runtime targeting and E2E integration #1098.

   Closing as resolved. The older open RDNA4 PRs (#1296, #1757, #1032) were not part of this checklist and are triaged separately.
3. chaojhou

   closed this as completedon Oct 22, 2026
4. github-actions

   added

   awaiting-author-responseWaiting for issue author to reply (auto-managed)Waiting for issue author to reply (auto-managed)

   on Oct 22, 2026