From 43d462aebada378c4849c6168575d5861526f43e Mon Sep 17 00:00:00 2001 From: bong-water-water-bong Date: Fri, 23 Oct 2026 13:40:55 -0300 Subject: [PATCH 1/4] llama.cpp pin 851fa6c8d8f2 -> f1670b43c93c: HRX test suite on the runner, #533 case 1, test-hrx-ops kquant routes Fork #126 (publication assertions, closes engine #449), #127 (PAD param dump compile fix + test-hrx-ops on the suite runner), #128 (kquant decode route expectations), #129/#130 (#445 rows at the current pin), #131 (#533 case 1: absent bias/residual slots bind a read-only value - verifier passes, numerically inert). Tests and diagnostics only; the one backend change (#127 PadParams dump branch, #131 placeholder binding) is covered by hrx-backend-test and the model A/B tables in #131. Co-Authored-By: Claude Fable 5.1 --- 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 851fa6c8..f1670b43 160000 --- a/third_party/llama.cpp +++ b/third_party/llama.cpp @@ -1 +1 @@ -Subproject commit 851fa6c8d8f220074ba73c7450ffb2219b27180b +Subproject commit f1670b43c93c3444a5891578c39838907672fe28 From 7b6fbbe12265c0db7c1bd2a649b46cbf88b54825 Mon Sep 17 00:00:00 2001 From: bong-water-water-bong Date: Fri, 23 Oct 2026 14:43:43 -0300 Subject: [PATCH 2/4] Bump llama.cpp: 830e2c161c68 (fork #132, test-hrx-ops route expectations) Fork #132 (tests only): require_matmul_route() accepts the table's kernel or the registered kquant decode family and prints the scheduled sequence when neither is there; stale decode/prefill sibling names and three routed-MoE names replaced from the first device run's log (engine #444). Co-Authored-By: Claude Fable 5.1 --- 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 f1670b43..830e2c16 160000 --- a/third_party/llama.cpp +++ b/third_party/llama.cpp @@ -1 +1 @@ -Subproject commit f1670b43c93c3444a5891578c39838907672fe28 +Subproject commit 830e2c161c68285402e4bbe8e486a58a8e599449 From 03a06ab7ef52d4492a7e6befabf5821db26074c3 Mon Sep 17 00:00:00 2001 From: bong-water-water-bong Date: Fri, 23 Oct 2026 16:09:58 -0300 Subject: [PATCH 3/4] Bump llama.cpp: c3eef46c0fec (fork #134: Flint architecture) Fork #134 ports upstream llama.cpp #30928: Flint (FlintForCausalLM), the 40B-A2.5B MoE coder with hybrid sliding-window/full attention, learned sinks and per-layer RoPE flags. CPU-verified on ryzen (greedy coherence on three prompts, PPL 5.56 on 10 chunks); the HRX build and test-llama-archs on gfx1151 are this gate's job. Co-Authored-By: Claude Fable 5.1 --- 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 830e2c16..c3eef46c 160000 --- a/third_party/llama.cpp +++ b/third_party/llama.cpp @@ -1 +1 @@ -Subproject commit 830e2c161c68285402e4bbe8e486a58a8e599449 +Subproject commit c3eef46c0feca6ff9e3d605cc7951024952f09e2 From c7bfb069fe83970b2e83d2ff766c96b899e031e3 Mon Sep 17 00:00:00 2001 From: bong-water-water-bong Date: Fri, 23 Oct 2026 16:23:30 -0300 Subject: [PATCH 4/4] Bump llama.cpp: f08bd76f672d (fork #135: jinja \xNN escapes) Fork #135 lets the vendored jinja lexer accept \xNN hex escapes, so a chat template that opens with replace("\x00", "") (Flint's does) no longer aborts serve and the onebit backend before the model loads. Co-Authored-By: Claude Fable 5.1 --- 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 c3eef46c..f08bd76f 160000 --- a/third_party/llama.cpp +++ b/third_party/llama.cpp @@ -1 +1 @@ -Subproject commit c3eef46c0feca6ff9e3d605cc7951024952f09e2 +Subproject commit f08bd76f672d06509e053bcdf68bb8fbadc0fb4f