Skip to content

docs/releases.md says the 7 Loom build-machine paths are fixed; the nested build never receives a prefix map, so they are not #1496

Closed as not planned

Description

@bong-water-water-bong

What the merged docs say (docs/releases.md:147-150, from #1462, merged):

The three mechanisms are fixed (engine#1461), and the payload is now checked rather than described: … llama.cpp passes the same flags into its nested GGML_HRX_CMAKE_ARGS (llama.cpp#243), which is where the 7 Loom strings came from …

What the mechanism actually does. The nested ggml-hrx-deps ExternalProject_Add - the build that compiles Loom - receives an argument list that forwards the build type and the compilers, and not the flags:

GGML_HRX_CMAKE_ARGS =
    -DCMAKE_BUILD_TYPE=…  -DCMAKE_C_COMPILER=…  -DCMAKE_CXX_COMPILER=…
    -DIREE_BUILD_TESTS=OFF  -DIREE_BUILD_BENCHMARKS=OFF  -DIREE_HAL_DRIVER_*=…
    [-DIREE_ROCM_PATH=…]  [-DFETCHCONTENT_BASE_DIR=…]

There is no -DCMAKE_C_FLAGS / -DCMAKE_CXX_FLAGS in it (checked in the fork's ggml/src/ggml-hrx/CMakeLists.txt on 2026-11-08). A prefix map set on the parent build therefore does not reach the nested one, so Loom's __FILE__ values stay absolute.

And the strings are still there, measured out of the c33 image's shipped SFS (267726ac…):

payload member /home/ occurrences
bin/lemond 126
bin/lemonade 13
1bit/hrx/llama/bin/libloomc.so.0 7

(Note for anyone re-measuring: 1bit/hrx/llama/lib/libloomc.so.0.0.1 reads 0 - the file carrying the 7 is 1bit/hrx/llama/bin/libloomc.so.0, a regular file. Two paths, one name.)

What is genuinely fixed: the Lemonade half. scripts/build-lemonade.sh now passes four maps (-ffile-prefix-map/-fdebug-prefix-map for the source tree and the build directory, #1464), and the author measured lemond rebuilt from the merged script at 0 /home/ against the 126 above. Worth keeping in the docs for a second reason: cmake does not echo -DCMAKE_CXX_FLAGS, so the arrival of the maps is answerable only from CMakeCache.txt.

Why this matters rather than being pedantic: #1461 is closed as fixed, and the release docs now tell a reader that the payload no longer names the build machine. It still does, in one library, and the payload is therefore not byte-reproducible - which is the property the qualification paragraph exists to protect. A closed issue plus a merged doc asserting a fix is the shape this release lane spends its time disproving: the statement was checked against another statement, not against the artifact.

Suggested correction for docs/releases.md:147-150:

The Lemonade half is fixed (engine#1464: four prefix maps, 126 → 0 /home/ strings in lemond, measured). The Loom half is not: the nested ggml-hrx-deps build receives no CMAKE_C_FLAGS/CMAKE_CXX_FLAGS through GGML_HRX_CMAKE_ARGS, so the 7 strings in libloomc.so.0 survive a rebuild. The payload is therefore not yet byte-reproducible, and scripts/weekly.sh's build_path_strings check is what stops a roll claiming otherwise.

What would actually close it (not this lane's call, and not needed for c33): choose one of - forward the flags through the nested argument list; or set the map inside the nested project; or compile Loom in the parent build. Any of the three is a fork change; this lane's contribution is the measurement that says the current doc claims a fix the mechanism cannot perform.

Labelled as a docs defect first: the release body for c33 already says "the Lemonade half is fixed (126 → 0), 7 remain in libloomc, so the payload is not yet byte-reproducible and the qualification stands", so the artifact of record and the docs disagree with each other right now.

Activity

  1. bong-water-water-bong commented on Nov 8, 2026

    @bong-water-water-bong
    CollaboratorAuthor

    Confirmed on the pinned submodule, not a branch, and the number is zero:

    third_party/llama.cpp  0ce01dafcf35
    $ sed -n '/set(GGML_HRX_CMAKE_ARGS/,/)/p' ggml/src/ggml-hrx/CMakeLists.txt | grep -cE 'CMAKE_C_FLAGS|CMAKE_CXX_FLAGS'
    0
    

    The list forwards CMAKE_BUILD_TYPE, CMAKE_C_COMPILER, CMAKE_CXX_COMPILER and the IREE driver switches - so the engine's -ffile-prefix-map (cmake/hrx.cmake:73-76, which does map CMAKE_BINARY_DIR where Loom is fetched) stops at the llama.cpp boundary and the nested ggml-hrx-deps-build compiles Loom with absolute paths. That is the whole of the 7.

    And one consequence worth writing into the issue, because it is already live on main: scripts/weekly.sh's build_path_stringsnow walks **every** packed ELF, so a payload rolled from this tree will **fail** that row (the 7 are inopt/1bit-package/1bit/hrx/llama/bin/libloomc.so.0). The gate is right and the leak is real - #1461 was closed on a sentence rather than on that measurement, and the ordering trap in #1461's own text was 'maps first, row last'. So either the fork forwards the two flags (plus the engine's CMAKE_BINARY_DIR` map, already there) before the next roll, or the next roll reds on this row and the red is the correct answer.

    Opened #1497 to correct docs/releases.md, which currently says all three mechanisms are fixed; it carries these numbers and says plainly that the roll is expected to fail until this lands.

  2. bong-water-water-bong commented on Nov 8, 2026

    @bong-water-water-bong
    CollaboratorAuthor

    Correction to my comment above, which was measured at a pin that has since moved. I wrote that GGML_HRX_CMAKE_ARGS carries zero occurrences of CMAKE_C_FLAGS/CMAKE_CXX_FLAGS, measured at third_party/llama.cpp 0ce01dafcf35. That was right about that commit and is wrong about this issue now: main pins the fork at 6ea38d0457d3 - "ggml-hrx: pass the parent's C/CXX flags into the deps ExternalProject (engine#1461) (#243)" - and there the flags are in the base set, not a conditional:

    20  set(GGML_HRX_CMAKE_ARGS
    31      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
    32      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
    

    and main's cmake/hrx.cmake carries -ffile-prefix-map=${CMAKE_BINARY_DIR}=build beside the source-tree map.

    So the claim this issue was filed against - that the docs say the Loom half is fixed - is true at main's pin, and the thing that was wrong is my measurement's vintage. The libloomc.so.0 strings in the c33 payload are real and ship (7 of them: this artifact predates both halves), which is the part of your report that stands and belongs in the release body; what does not stand is that it cannot be fixed here.

    One nuance worth keeping from pi-release's review, because I got it half right: at 0ce01dafc the flags do appear in that CMakeLists.txt - in a GCC-15-conditional list(APPEND) at 68-69 - so a whole-file grep -c returns 2 while the base set() contains none. A number can be exactly right about its scope and still mislead when someone reads the scope as 'the file'.

  3. bong-water-water-bong commented on Nov 8, 2026

    @bong-water-water-bong
    CollaboratorAuthor

    Closing this as overtaken by the tree — it was right when filed and is false as of main's pin, and the useful output is the nuance, not the issue.

    What I filed: docs/releases.md says the 7 Loom build-machine paths are fixed, and the nested build never receives a prefix map.

    Measured now on main, third_party/llama.cpp 6ea38d0457d3 (2026-11-08 09:45, "ggml-hrx: pass the parent's C/CXX flags into …"):

    ggml/src/ggml-hrx/CMakeLists.txt:20  set(GGML_HRX_CMAKE_ARGS
                                     31  -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}      <- inside the BASE set
                                     32  -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
    cmake/hrx.cmake:82   string(APPEND ONEBIT_HRX_PATH_MAP_FLAGS " -ffile-prefix-map=${CMAKE_BINARY_DIR}=build")
    

    So both halves are in the tree, the docs sentence is true, and #1497 (the correction) was closed by its author for exactly that reason.

    The part worth keeping, because it is what made this hard to see: at the pin my report cited — 0ce01dafcf35 — the flags do appear in that same file, but only in a GCC-15-conditional list(APPEND) (lines 68-69), not in the base set. So a whole-file grep -c 'CMAKE_C_FLAGS' returns 2 and reads like a contradiction of "no forwarding", while the base set genuinely forwards none. A number can be exactly right about its scope and still mislead when a reader takes the scope for the file.

    And the artifact side is unchanged and worth keeping in the record: c33's payload still ships all 146 (bin/lemond 126, bin/lemonade 13, 1bit/hrx/llama/bin/libloomc.so.0 7) because it was cut before both halves landed — so "not byte-reproducible" is still true of that image, and the release body for c33 says so, names 6ea38d0457d3 as the fix, and calls the next roll's 0 a prediction rather than a measurement. If the next roll's build_path_strings row reds anyway, that is the finding to open a fresh issue for.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions