From 230dafb3692d21f5d76530f44acc1be4ebd53c09 Mon Sep 17 00:00:00 2001 From: pi-hrx-sched Date: Mon, 9 Nov 2026 10:40:09 -0400 Subject: [PATCH 1/2] bump llama.cpp to 8f124f906: the live pool check is on by default (engine #1417) Moves third_party/llama.cpp b970d85ba5 -> 8f124f906c, the tip of 1bit/hrx-vulkan-patched. Everything in the range is a merged, reviewed fork PR: 16e0b5e ggml-hrx: the live pool check is ON by default, so a neighbour growing GTT is refused at admission (engine #1417) <- fork PR #254, this issue's fix 8f124f9 ggml-hrx: a typo must not disarm the pool guardrail (own review of #254) 090fda5 hrx: say which pool state produced the graph split when strict placement refuses (#240) 932485c kv-share: report the tensor table a prefill provider writes into (engine #753) (#253) 70b8fe4 ggml-hrx: declare device_free_total before its first use, so the pin compiles (#252) c89d750 test-backend-ops: space the index-list scatter's scores (#242) 4db2eef ggml-hrx: spend only what a refusal proves (engine #1502) (#251) 6663e37 ggml-hrx: an opt-in prefill path over the repacked q6_k tiles (engine #856) (#250) d27648e ggml-hrx: a refused allocation now shrinks the pool budget (engine #1502) (#249) The last two commits of that list are why the pin, not a cherry-pick, is the delivery: #254 was merged and then corrected in 8f124f9, whose own message records the defect - pool_live_check_enabled() parsed the knob with std::atoi, so 'true', 'yes' or a misspelling read as 0 and turned the guardrail OFF. Only an explicit 0/false/no switches it off now. Pinning #254's merge commit would have shipped the guardrail with the hole in it. Verified: git merge-base --is-ancestor b970d85ba5 8f124f906 -> yes, and scripts/check_pins.py origin/main passes. The fork-side test (11 arms built and run in #254, plus the fail-safe arm from 8f124f9) is the test for the fix itself; a pin bump changes no engine code, so there is nothing here to compile - hw-gate runs on this bump-hrx/ branch by convention. registry/architectures.json's sources sha still names b970d85ba; the file is generated by the daily census workflow, which rewrites it on its next run, so this PR does not hand-edit it. --- third_party/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/third_party/llama.cpp b/third_party/llama.cpp index b970d85ba..8f124f906 160000 --- a/third_party/llama.cpp +++ b/third_party/llama.cpp @@ -1 +1 @@ -Subproject commit b970d85ba5a9be396a3d271f9de8d8dd090aa148 +Subproject commit 8f124f906c67bd74f0f0a712836ef997938e92e0 From 3a69f211289bd94c1ff143b12d5322936da7e53c Mon Sep 17 00:00:00 2001 From: pi-hrx-sched Date: Mon, 9 Nov 2026 11:01:11 -0400 Subject: [PATCH 2/2] registry: regenerate at the new pin (only the sources sha moves) The build job failed on my first push at 'registry_pins': registry/architectures.json: llama.cpp (hrx) recorded at b970d85ba, pinned at 8f124f906 a pin moved without regenerating the registry: check out the submodules at their pins and run tools/registry_build.py I had written in the PR body that the file is generated and would be rewritten by the daily census run, so I left it alone. That was wrong: tools/registry_build.py --check-pins asserts the recorded sha EQUALS the pin, and ctest runs it as registry_pins (CMakeLists.txt:718), so a pin bump that skips the regeneration cannot merge. Regenerated with the submodule checked out at 8f124f906: wrote registry/architectures.json: hrx 298, npu 9, architectures 298, mapped 298 and the only delta is the sources sha - the counts and every architecture entry are unchanged, so none of the nine commits in this bump touched the architecture surface. --check-pins and --check-gaps both pass here now. --- registry/architectures.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/registry/architectures.json b/registry/architectures.json index 7a2667f17..6cca862db 100644 --- a/registry/architectures.json +++ b/registry/architectures.json @@ -1,7 +1,7 @@ { "about": "HF architecture -> GGUF architecture and the backends whose code accepts it. Generated by tools/registry_build.py from the pinned sources; do not edit.", "sources": { - "llama.cpp (hrx)": "b970d85ba5a9be396a3d271f9de8d8dd090aa148" + "llama.cpp (hrx)": "8f124f906c67bd74f0f0a712836ef997938e92e0" }, "counts": { "hrx": 298,