Repository navigation
bump-xdna: correct the XRT path in the PR's test recipe - #82
Merged
Merged
Conversation
build-xdna.sh installs there; ONEBIT_XRT_ROOT=<prefix> fails to configure (found verifying #76). XILINX_XRT points XRT at the new driver plugin for the lane test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PR Reviewer Guide 🔍(Review updated until commit 45c2963)Here are some key observations to aid the review process:
|
|
Persistent review updated to latest commit a806941 |
|
Persistent review updated to latest commit b0cd956 |
|
Persistent review updated to latest commit 45c2963 |
This was referenced Oct 5, 2026
bong-water-water-bong
added a commit
that referenced
this pull request
Oct 5, 2026
…to our fork (Loom NaN backport) (#320) * Bump HRX: llama.cpp 522dab4789ce (long-prompt fix), hrx-system to our fork 98d05d94a9f9 third_party/llama.cpp e57beb9721af -> 522dab4789ce (fork #89 runtime overhead, #90: graph inputs never written back to host — every multi-ubatch prompt was wrong since #82 — and the non-replay upload race). third_party/hrx-system moves from ROCm/hrx-system 51b1739ae5fd to our fork 1bit-MONSTER/hrx-system 1bit/main 98d05d94a9f9 = the same AMD commit plus the GFX11 wave64 lane-mask drain backport (#313 NaN) and two libhrx device knobs. bump-hrx.yml now moves only the llama.cpp pin; hrx-system is rebased by hand. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * registry: regenerated at llama.cpp 522dab4789ce Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> --------- Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
This was referenced Oct 7, 2026
bong-water-water-bong
pushed a commit
that referenced
this pull request
Oct 14, 2026
… root-caused a NaN in the same family BUILT 05:52, compile caveat discharged: the patch applied to the real tree and the build exited 0 (graph-executor.cpp.o recompiled, libggml-hrx.so relinked), and the artefact was verified rather than inferred - bin/libggml-hrx.so.0.18.0 contains BOTH GGML_HRX_PRODUCER_PROBE and GGML_HRX_PRODUCER_DEVICE, object and library both timestamped 05:52, source backed up at /tmp/graph-executor.cpp.orig (sha256 cba0419f55e748ed...). Run 315dev launched behind the gate-aware wrapper that holds no lock and no GTT. Still NOT claimed: that it arms anything at run time - the armed/anomaly lines are that evidence and do not exist yet. THE SIBLING FIX: a fork pin-bump (fb24c1f, Oct 5) turned out to describe a root-caused NaN fix in exactly this family - llama.cpp e57beb9721af -> 522dab4789ce with fork #90: 'graph inputs never written back to host - every multi-ubatch prompt was wrong since #82 - and the non-replay upload race'; and hrx-system moved to our fork 98d05d94a9f9 carrying 'the GFX11 wave64 lane-mask drain backport (#313 NaN)'. The fork merge reads: '522dab478 Merge pull request #90 from 1bit-MONSTER/1bit/kv-nan-rootcause / ggml-hrx: graph inputs are never written back to host; non-replay programs wait for their input upload'. Two consequences: it is independent evidence that the family is real and reachable (a KV-NaN root cause was found in the host write-back path); and the engine pin's tree ALREADY contains a PendingHostWriteback path whose final step is memcpy(writeback.host_destination, writeback.mapped_source, size) - a DEVICE MAPPING copied into a HOST DESTINATION (runtime/command-program-executor.cpp) - i.e. the exact line where 'the device wrote these bytes' and 'the mapping still holds something else' are decided, in the pin this work runs on. NOT established, and it decides whether #90 is a fix or already shipped: 522dab4789ce versus the engine pin e44c9d01a. The local clone is SHALLOW - 522dab4789ce has no parents in it and e44c9d01a is absent - so merge-base cannot run. If 522dab4789ce is an ancestor of the pin, #90 is already shipped and #315's remaining faults are a different path; if it is a descendant, #90 is a forward-only fix candidate. Settling it needs a full fetch or the engine repo's submodule, and it is worth doing before any other run.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Found while verifying #76:
build-xdna.sh <prefix>installs XRT at<prefix>/root/opt/xilinx/xrt, so the recipe's-DONEBIT_XRT_ROOT=<prefix>fails to configure. The recipe now uses the install path, and setsXILINX_XRTso the lane test loads the new driver plugin.🤖 Generated with Claude Code