From e515265d9dec60297f9f9cba034365596875f9fd Mon Sep 17 00:00:00 2001 From: bong-water-water-bong Date: Sun, 8 Nov 2026 08:04:14 -0400 Subject: [PATCH 1/2] registry: decline RMSN as a reviewed non-alias, so the census reports it instead of alerting The 2026-11-08 census run found RMSN as a new uncovered class (DuoNeural/RMSN-193M-Soliton-Reasoning-SFT). Read from the model's own config and card rather than the class name: model_type rmsn, dim 1024, depth 16, slots 64, seq_len 512, vocab 49152 on the cosmo2 tokenizer, 243.7 M parameters, and no attention/FFN keys -- the card names it a Resonant Metriplectic Sheaf Network (port-Hamiltonian/metriplectic/Lyapunov tags, its own runtime, val_lyap in the config). A continuous-time system, not a Transformer with extra keys, and nothing at the pin registers the class or an rmsn gguf architecture, so a mapping would claim a graph the engine cannot run. Recorded with watch: report and a dated note in docs/arch-gaps.md, which turns the alert into a REVIEWED line in the census. The two REMOTE-CODE classes from the same run are deliberately left out (single-uploader policy: the alert is the review trigger) and the 14 UNVERIFIABLE gated repos get no entries (retried by the watcher; an "unknown" row could never resolve). Both decisions are written down in the doc rather than left implicit. tools/registry_build.py --check-gaps: 34 reviewed classes, none mapped without a reason, rc 0 (the registry_gaps ctest and census.yml run the same check). Fixes #1451 --- docs/arch-gaps.md | 21 +++++++++++++++++++++ registry/significant.json | 5 +++++ 2 files changed, 26 insertions(+) diff --git a/docs/arch-gaps.md b/docs/arch-gaps.md index 43a36c814..e37f51213 100644 --- a/docs/arch-gaps.md +++ b/docs/arch-gaps.md @@ -993,3 +993,24 @@ in `registry/significant.json`. project's own pipeline, and nothing at the pin registers the class. The V4.6 text tower is supported, so this is a converter-and-proof job rather than an architecture gap, but it is not one registration: the head and the vision side would both have to be verified. + +## Census run 2026-11-08 (`engine #1451`) + +* `RMSN` (`DuoNeural/RMSN-193M-Soliton-Reasoning-SFT`) — **declined, added to `registry/significant.json` with + `"watch": "report"`** so the census prints `. REVIEWED` instead of `!! UNCOVERED`. Read from the model's own + `config.json` and card, not inferred from the class name: `architectures: ["RMSN"]`, `model_type: rmsn`, + `dim 1024`, `depth 16`, `slots 64`, `seq_len 512`, `vocab_size 49152` on `HuggingFaceTB/cosmo2-tokenizer`, + 243.7 M parameters — and **no attention or FFN keys at all**. The card names it *Resonant Metriplectic Sheaf + Network*, tagged port-Hamiltonian / metriplectic / Lyapunov / sheaf, with its own GitHub runtime, and the config + carries a Lyapunov diagnostic (`val_lyap`) next to perplexity. That is a continuous-time dynamical system, not a + Transformer with a few extra keys, and nothing at the pin registers the class or an `rmsn` gguf architecture + (`conversion/` has no module for it, `src/llama-arch.cpp` has no arch), so a registry entry would claim a graph + the engine cannot run. +* The two `. REMOTE-CODE` classes in the same run (`Lfm3V9ForCausalLM` from `LTM-Lab/LFM3-1.7B-v9`, + `PhobosScratchLM` from `marsnext/Phobos-1`) are **deliberately not** added, per the policy recorded above: each + was seen from a single uploader, and the `. REMOTE-CODE` line plus the `REMOTE_CODE_SPREAD` alert is the review + trigger — silencing it now would remove the prompt. Both would need their config read the same way before anyone + maps them. +* The 14 `UNVERIFIABLE` rows are **not** a registry problem and get no entries: they are gated repos with no + readable `config.json`, so the watcher retries them. Adding "unknown" dispositions would create rows that can + never resolve and would hide the day one of them becomes readable. diff --git a/registry/significant.json b/registry/significant.json index 8bd3014cb..0568e8b44 100644 --- a/registry/significant.json +++ b/registry/significant.json @@ -152,6 +152,11 @@ "model_type": "grt", "family": "GRT rotational/coupling decoder (MatterWaveAI grt-30m-code, 8 models in registry/census.json). Not an alias of anything mapped: the config reads model_type grt with a rotational coupling mixer (n_qkv_rotations 11, n_mixer_rotations 1, mor_router_mode rotational, mor_delta_scale 0.5, coupling_init_mode diagonal, angle_init_scale 0.1), learnable_rope with rope_placement stream, residual_gate_mode bounded_depth, no residual norm, and it declares transformers_version 5.23.0.dev0 - a research decoder, not a llama/qwen shape with extra keys. Nothing at the pin registers the class or a 'grt' gguf architecture (conversion/ has no grt module, src/llama-arch.cpp has no grt arch), so mapping it would claim a graph the engine cannot run. Declined with this reason on 2026-11-07 under engine #1328; the census reports it as . REVIEWED instead of !! UNCOVERED. Remove this entry and watch: report if a backend ever ports it.", "watch": "report" + }, + "RMSN": { + "model_type": "rmsn", + "family": "RMSN, Resonant Metriplectic Sheaf Network (DuoNeural physical-informed research decoder: model_type rmsn with dim/depth/slots/seq_len and a Lyapunov diagnostic in the config, card tags port-hamiltonian/metriplectic/lyapunov/sheaf, its own GitHub runtime - a continuous-time dynamical system, not a Transformer with attention/FFN tensors). Not an alias of anything mapped: config.json declares architectures [RMSN] with vocab_size 49152 on the HuggingFaceTB/cosmo2-tokenizer, dim 1024, depth 16, slots 64, seq_len 512, 243.7 M parameters and no attention/FFN keys, and nothing at the pin registers the class or an 'rmsn' gguf architecture (conversion/ has no module for it, src/llama-arch.cpp has no arch), so mapping it would claim a graph the engine cannot run. Declined with this reason on 2026-11-08 under engine #1451; the census reports it as . REVIEWED instead of !! UNCOVERED. Remove this entry and watch: report if a backend ever ports it.", + "watch": "report" } } } From afb5e7392fce3d3bd45d676ddc8430d9ff4d16f7 Mon Sep 17 00:00:00 2001 From: bong-water-water-bong Date: Sun, 8 Nov 2026 08:44:22 -0400 Subject: [PATCH 2/2] hrx: stop compiling the builder's HSA runtime path into 1bit, and red the gate on this machine's tree c33's roll failed its own package gate with one string read out of the packed `1bit`: /home//wt/c33-payload/hrx-toolchain/.../libhsa-runtime64.so.1. cmake/hrx.cmake globs the runtime out of the build's toolchain, and CMakeLists.txt compiled that absolute path in as ONEBIT_HRX_LIBHSA -- so the path is a build-machine path by construction (engine#1459). Three consequences: the tarball bytes depend on where the build ran (c32 baked /opt/rocm-therock and carried 0 /home/ strings, c33 baked its own workdir and carried 1, same source); it writes the builder's tree into a public artifact; and serve can only use the value when the path exists on the target, which it does not -- the appliance installs the runtime flat at /opt/rocm-therock/lib, so serve falls through to its scan anyway. - the macro is compiled in only when -DONEBIT_HRX_LIBHSA_BAKE= is passed, the explicit opt-in for a builder whose layout is the target's; - ONEBIT_HRX_LIBHSA stays a cache value: os/mini/mkimage.sh:1020 reads it to decide which runtime to ship; - the gate row in scripts/weekly.sh red on `/home/` only, which is why c32 shipped a builder path under /opt and passed. It now reds on the build's own roots ($SRC, $W) and deliberately not on /opt/rocm-therock/..., which is also the appliance's deployed path and therefore legitimate. Gate checked synthetically: a string under $SRC/ -> fail (n=2); the deployed /opt/rocm-therock/.../_rocm_sdk_core/lib path -> pass (n=0). bash -n clean on weekly.sh, build.sh and vm-check.sh; the two OS comments describing the old behaviour are updated in the same commit. Fixes #1459 --- CMakeLists.txt | 11 ++++++++++- cmake/hrx.cmake | 9 ++++++++- os/v3/puppy/build.sh | 9 +++++---- os/v3/puppy/vm-check.sh | 5 +++-- scripts/weekly.sh | 13 +++++++++++-- 5 files changed, 37 insertions(+), 10 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index d85427bf3..3c8c8e073 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -423,8 +423,17 @@ if(ONEBIT_HRX) add_dependencies(onebit llama_hrx) endif() onebit_build_relative(ONEBIT_HRX_SERVER _onebit_hrx_server) - target_compile_definitions(onebit PRIVATE ONEBIT_HRX_SERVER="${_onebit_hrx_server}" ONEBIT_HRX_LIBHSA="${ONEBIT_HRX_LIBHSA}" + target_compile_definitions(onebit PRIVATE ONEBIT_HRX_SERVER="${_onebit_hrx_server}" ONEBIT_HRX_ALLOW_NO_SVM=${ONEBIT_HRX_ALLOW_NO_SVM}) + # The HSA runtime path is NOT compiled in (engine#1459): the glob below finds it under the builder's toolchain, so + # baking it puts that build directory into a public artifact (c33's roll failed its own gate with one + # `/home//wt/.../libhsa-runtime64.so.1` string), it makes the tarball bytes depend on where the build ran, and + # serve can only use it when the path exists on the target -- which it does not, so it falls through to its own scan + # either way. ONEBIT_HRX_LIBHSA stays a cache value because os/mini/mkimage.sh reads it to decide what to ship; + # ONEBIT_HRX_LIBHSA_BAKE is the explicit opt-in for a builder whose layout *is* the target's. + if(ONEBIT_HRX_LIBHSA_BAKE) + target_compile_definitions(onebit PRIVATE ONEBIT_HRX_LIBHSA="${ONEBIT_HRX_LIBHSA_BAKE}") + endif() if(ONEBIT_SERVE_TEST_GGUF) foreach(_dev hrx cpu) add_test(NAME serve_e2e_${_dev} diff --git a/cmake/hrx.cmake b/cmake/hrx.cmake index 08ce35c49..1e4a9445d 100644 --- a/cmake/hrx.cmake +++ b/cmake/hrx.cmake @@ -103,9 +103,16 @@ if(ONEBIT_BINARY_DROP_HRX_BIN AND EXISTS "${ONEBIT_BINARY_DROP_HRX_BIN}/llama-se "TheRock HSA runtime HRX loads (IREE_HAL_AMDGPU_LIBHSA_PATH)" FORCE) message(STATUS "ONEBIT_HRX: no libhsa-runtime64.so.1 under ${ONEBIT_HRX_TOOLCHAIN}; the build " "proceeds, but HRX needs TheRock's on the machine that runs it (pass " - "-DONEBIT_HRX_LIBHSA= to bake it in)") + "-DONEBIT_HRX_LIBHSA= to tell mkimage what to ship, or " + "-DONEBIT_HRX_LIBHSA_BAKE= to compile it into '1bit' as well)") endif() endif() + set(ONEBIT_HRX_LIBHSA_BAKE "" CACHE FILEPATH + "Compile this HSA runtime path into '1bit'. Empty by default: ONEBIT_HRX_LIBHSA is the toolchain path the build " + "FOUND and stays build metadata because os/mini/mkimage.sh reads it to decide what to ship, but a path compiled " + "into the binary is the builder's own tree in a public artifact and is dead on the appliance, where serve falls " + "through to its /opt/rocm-therock scan anyway (engine#1459). Set this only when the build directory's layout is " + "the target's.") set(ONEBIT_HRX_SERVER "${ONEBIT_BINARY_DROP_HRX_BIN}/llama-server" CACHE FILEPATH "HRX llama-server 1bit runs" FORCE) set(ONEBIT_HRX_ALLOW_NO_SVM 1) diff --git a/os/v3/puppy/build.sh b/os/v3/puppy/build.sh index 623226dbe..d0419f0c1 100755 --- a/os/v3/puppy/build.sh +++ b/os/v3/puppy/build.sh @@ -29,10 +29,11 @@ # libhsa-runtime64.so.1 plus every sibling it needs at load time (the TheRock `librocm_sysdeps_*` set and # librocprofiler-register.so.0). Cycle 13's board boot measured why the set and not one file (engine#715 + # the 2 Nov hardware run): -# * `ONEBIT_HRX_LIBHSA` is a CMake compile definition, not a variable - nothing in the engine reads it at -# run time (CMakeLists.txt:279, cmake/hrx.cmake). app/serve.cpp:201 resolves in the order --hrx-libhsa, -# the BAKED build-machine path if it exists, then a recursive scan of /opt/rocm-therock. On the appliance -# the baked path is a release-build directory that does not exist, so the scan is the only route left; +# * `ONEBIT_HRX_LIBHSA` is NOT a compile definition any more (engine#1459): the path the build finds in its own +# toolchain stays in the CMake cache for os/mini/mkimage.sh to ship the runtime from, and it is compiled in only +# when a builder passes -DONEBIT_HRX_LIBHSA_BAKE=. app/serve.cpp:201 resolves in the order --hrx-libhsa, a +# baked path if one was asked for and exists, then a recursive scan of /opt/rocm-therock, so on the appliance the +# scan is the route; # * the pinned libhsa carries `RUNPATH=$ORIGIN` and NEEDED librocm_sysdeps_{elf,drm,drm_amdgpu,numa} + # librocprofiler-register, so a lone file cannot load even when handed to the loader: measured `ldd` = 4 # `not found`. With the release kit's own directory on the path: 0 not found. diff --git a/os/v3/puppy/vm-check.sh b/os/v3/puppy/vm-check.sh index 9d70aaaa0..4d5582547 100755 --- a/os/v3/puppy/vm-check.sh +++ b/os/v3/puppy/vm-check.sh @@ -343,8 +343,9 @@ rows = [ "[ \"$n\" = 3 ] && grep -q 'trap .one_rescue' /initrd/init && echo init-guard-3-paths", "init-guard-3-paths"), # Cycle 14, item 1 - the GPU blocker the board found: HRX registered no device because nothing ever opened - # libhsa-runtime64.so.1. `ONEBIT_HRX_LIBHSA` is a CMake compile definition, not a variable (measured in - # CMakeLists.txt:279 and app/serve.cpp:201), so the only runtime route on an appliance is the engine"'"'s + # libhsa-runtime64.so.1. Nothing bakes an HSA path into `1bit` by default (engine#1459: the build's own + # toolchain path is build metadata for mkimage, and -DONEBIT_HRX_LIBHSA_BAKE is the opt-in), so the runtime + # route on an appliance is the engine"'"'s # scan of /opt/rocm-therock - and even handed the file, the pinned libhsa cannot load without its # $ORIGIN siblings. This row asserts all three halves on the artifact: the set at the path the scan walks, # no unresolved dependency, and a dlopen by BARE soname inside the guest (what libggml-hrx does when the diff --git a/scripts/weekly.sh b/scripts/weekly.sh index d9c7e000e..9441cc771 100755 --- a/scripts/weekly.sh +++ b/scripts/weekly.sh @@ -450,8 +450,17 @@ package_checks() { # package_checks done | grep -E '\.(c|cc|cpp|cxx|h|hpp|py)$' | xargs -r -n1 basename | LC_ALL=C sort -u) n=0; [ -z "$privsrc" ] || n=$(strings -a "$f" | grep -c -F -f <(printf '%s\n' "$privsrc") || true) [ "$n" = 0 ] || { echo "FAIL the packed 1bit names a private source file ($n strings)"; b1=1; } - n=$(strings -a "$f" | grep -c '/home/' || true) - [ "$n" = 0 ] || { echo "FAIL the packed 1bit names a build-machine home ($n strings)"; b1=1; } + # Red on THIS machine's tree, not on one prefix that happens to be common (engine#1459): the row used to test + # for `/home/`, so c32 shipped a builder path under `/opt` and passed. Test the build's own roots instead -- and + # deliberately not `/opt/rocm-therock/...`, which is also the appliance's deployed runtime path, so baking that + # one is legitimate. + n=0 + for _root in "$SRC" "$W"; do + [ -n "$_root" ] || continue + _c=$(strings -a "$f" | grep -c -F "$_root" || true) + n=$((n + ${_c:-0})) + done + [ "$n" = 0 ] || { echo "FAIL the packed 1bit names a build-machine tree ($n strings under \$SRC/\$W: $SRC, $W)"; b1=1; } if [ $b1 = 0 ]; then echo "ok the packed 1bit: no symbol table, no private source names, no build-machine paths"; else bad=1; fi fi for f in LICENSE NOTICE LICENSE-BINARY NOTICE-BINARY; do