Commit Graph

119 Commits

Author SHA1 Message Date
alex 303eaa945e fix(kyojin): load calibrated HC sidecars from the model directory
CI / check (push) Has been cancelled
Bead: nix-overlay-kzh
2026-11-19 14:58:32 +03:00
alex 8389c88d04 perf(kyojin): finish completed streams before the next prefill
CI / check (push) Has been cancelled
Bead: nix-overlay-8gs
2026-11-19 11:22:29 +03:00
alex 848a6ec9c4 perf(kyojin): give warm replies a bounded prefill grace
Bead: nix-overlay-3k7
2026-11-19 11:02:32 +03:00
alex ba836e1e75 perf(kyojin): split large prefix writes under disk pressure
Bead: nix-overlay-3c1
2026-11-19 10:33:29 +03:00
alex 42c2a416fe perf(kyojin): support direct writes for large prefix segments
Bead: nix-overlay-2q7
2026-11-19 01:34:07 +03:00
alex 635f9f7dfc fix(kyojin): reuse queued prefix pages across agent turns
Bead: nix-overlay-7p6
2026-11-19 01:28:42 +03:00
alex 7843d79e76 fix(kyojin): release prefix buffers before the writer waits
Bead: nix-overlay-9hu
2026-11-19 00:41:57 +03:00
alex 41a50393f9 perf(kyojin): allow blocking HIP waits in the Qwen server
CI / check (push) Has been cancelled
Bead: nix-overlay-j2g
2026-11-18 21:16:42 +03:00
alex 2b8eddf7d2 fix(kyojin): Stop direct expert reads at unaligned EOF
Bead: nix-overlay-qy9
2026-11-18 18:34:42 +03:00
alex 6156a0d904 perf(kyojin): Skip prefix segments that cannot fit on disk
Bead: nix-overlay-tb3
2026-11-18 18:31:54 +03:00
alex afd0c4b84d perf(kyojin): Let short replies finish before a long prefill
CI / check (push) Has been cancelled
Give two-session Qwen serving a bounded, opt-in decode grace for large cold prompts. Returning to normal interleaving at the deadline bounds the wait when the running reply takes longer than expected.

Bead: nix-overlay-qfi
2026-11-18 15:19:23 +03:00
alex 7bb05ea685 feat(strata): build the MMQ prompt path
CI / check (push) Has been cancelled
Strata 0.2.14 leaves llama.cpp's MMQ prompt kernels out by default, so
STRATA_MMQ_KQ does nothing and UD-Q5_K_XL's Q6_K layer runs on Strata's
int8 kernel. On ms-s1-max, 0.2.15 with the gfx1151 prompt set and
STRATA_MMQ_KQ=1 gives first-token logits bit-identical to 0.2.13 over
15 cold prompts at the same speed (~1180 tok/s at 16K); without MMQ an
agent prompt reached KL 0.199 against 0.2.13. The engine grows from 39
to 54 MB.

Bead: nix-overlay-481
2026-11-16 21:42:33 +03:00
alex b3287c926f feat(strata): bump 0.2.13 -> 0.2.15
Strata 0.2.14 runs the prompt on its own kernels and vendors llama.cpp
(third_party/strata-llama), so the separate llama.cpp fetch for the
vision encoder goes. The build keeps upstream's halo flags: MMQ stays
out (STRATA_MMQ defaults to OFF), 47 of 48 UD-Q5_K_XL layers run on the
fused kernels, and the ROCm 10.1 hipBLASLt table still ships.

UD-Q5_K_XL on ms-s1-max, 16K cold prompts: 715 -> ~810 tok/s on the
defaults, ~1180 tok/s with the gfx1151 prompt set (unchanged).

Bead: nix-overlay-ydg
2026-11-16 20:51:14 +03:00
alex d14f6d5e37 fix(strata): ship a hipBLASLt table calibrated for ROCm 10.1
Upstream's gfx1151-hipblaslt-100401.txt was calibrated on TheRock 7.14.1.
therock-rocm 10.1.0rc3 reports the same hipBLASLt version, so the engine
accepted the table, but its solution ids do not exist in that build and
every prompt GEMM fell back to the heuristic. The table now comes from
tools/hip/tune_hipblaslt run on gfx1151 against therock-rocm 10.1 and is
installed in bin/tuning/, where the engine finds it by itself.

UD-Q5_K_XL, 17.5K cold prompt on ms-s1-max: 418 -> 700-725 tok/s.

Bead: nix-overlay-2yv
2026-11-16 20:22:23 +03:00
alex 598985f7c9 chore(awgen): bump 0.3.2 -> 0.3.3
CI / check (push) Has been cancelled
Upstream tagged v0.3.3; refreshed src and cargoHash. Build verified with
nix build .#awgen and awgen --version reports 0.3.3.
2026-11-13 14:13:50 +03:00
alex 7672ff56d3 feat(purpleflow): add PurpleFlow desktop app
CI / check (push) Has been cancelled
2026-11-13 10:40:13 +03:00
alex f2abdc96f9 fix(rocm): Address package review findings
CI / check (push) Has been cancelled
2026-11-12 23:55:09 +03:00
alex 16166f115b feat(rocm): Vendor SDK recipe and update to 10.1
Maintain the gfx1151 SDK locally using the official AMD archive. Use the TheRock patchelf pin and adapt LLVM auxiliary tools for pure Nix builds.

Verify SDK installation, HIP and rocBLAS GPU operations, and rebuild its consumers. Document the internal rc3 stamp and upstream migration findings.
2026-11-12 23:40:48 +03:00
alex 00dc8c2c8c feat(strixite): Add ROCm 10 build for gfx1151 2026-11-12 21:57:26 +03:00
alex 3d04ce81ad feat(crucible): Add Strix Halo inference package 2026-11-12 21:41:56 +03:00
alex a30c7d6c3a feat(strata): Add ROCm 10 build for gfx1151 2026-11-12 21:28:24 +03:00
alex 85bd7421de feat(llm): package upstream inference engines
Centralize engine pins, compiler wrappers and NPU packages.
Use upstream Kyojin v1.8.10 and drop the Thunderbolt dependency.

Offline evaluation passed; workstation builds remain pending.

Bead: nix-overlay-v09
2026-11-12 20:36:52 +03:00
Marmel Coder 4d08fafddc feat(codex-oauth-proxy): package private-state user service 2026-11-12 16:40:15 +03:00
alex 0feb16a0c0 fix(codex-oauth-proxy): support gpt-6.1-terra
1.1.1 omits the model from normalizeModel's allowlist, so requests for gpt-6.1-terra are rewritten to gpt-5.5. ChatGPT accounts are not entitled to gpt-5.5, so the request fails with model_not_found.

Add the model to the allowlist, allowed reasoning efforts, and default effort. Covered by the transform tests.
2026-11-12 15:47:53 +03:00
Marmel Coder 55f6a0b7ed feat(codex-oauth-proxy): add 1.1.1 Codex OAuth proxy package
Packages dvcrn/codex-oauth-proxy, which exposes a ChatGPT Codex account through OpenAI-compatible APIs, as a buildGoModule derivation. Builds only cmd/codex-oauth-proxy because the worker package is js/wasm code, and passes -buildvcs=false because the Nix source has no .git. Upstream Go tests run offline, so doCheck is enabled.
2026-11-12 02:04:06 +03:00
alex 19774c73f3 fix(codex-api-proxy): propagate sys.path to daemon child
CI / check (push) Has been cancelled
2026-11-09 22:22:25 +03:00
alex 9ffe8b006f feat(codex-api-proxy): add 0.1.6 local Codex API proxy 2026-11-09 22:08:09 +03:00
alex 589936cd8d feat(marmel): add arbitrary HTTP headers patch
CI / check (push) Has been cancelled
Apply upstream 56d0b1f so top-level, specialist, validator and MCP request headers can be configured. nix build .#marmel passes with the patch.
2026-11-09 10:03:47 +03:00
alex 44e6d90a77 feat(magg): add magg 1.3.0 MCP aggregator package
CI / check (push) Has been cancelled
Add magg, an MCP aggregator / meta-MCP server, to the overlay. It is a
Python package on the hatchling backend, built with
python3Packages.buildPythonApplication from the v1.3.0 source tarball,
so no upstream release artifacts are fetched. Ships both console
scripts declared upstream: magg (the aggregator CLI/server) and mbro
(the MCP browser). License is AGPL-3.0-or-later.

The install check runs the upstream pytest suite with
-m "not integration", because the integration-marked tests need a live
server. A forced nix build .#magg --rebuild passes with 329 passed,
3 skipped, 3 deselected, 0 failed, and both binaries smoke-test clean.

Also adds the README and AGENTS.md package rows, the upstream and build
research notes, and the verification report in docs/magg_build_report.md.
2026-11-08 12:48:30 +03:00
alex e569243ae4 chore(marmel): bump to afcc1dc (main), give checks a usable TMPDIR
CI / check (push) Has been cancelled
Track upstream main at afcc1dc971ae4acd59b157f031d4bdbc9d52a7d0
(2026-11-07), 22 commits past the de3e99c8 pin. Upstream still ships no
tags or releases and Cargo.toml still reads version 1.0.0, so the dated
0-unstable-* string stays. cargoHash moves on its own: the new
crates/zim-search workspace member grew Cargo.lock from 399 to 621
entries, and cargo vendor hashes the whole workspace.

The real packaging work is the TMPDIR override appended to preCheck.
Upstream's new harness::fs::path_tests scratch-root helper puts its
scratch dirs under <cwd>/target/ and first asserts that root lies
outside std::env::temp_dir(). The build sandbox exports TMPDIR=/build,
which is an ancestor of the unpacked source at /build/source, so that
premise check tripped and all 8 path-security tests failed with
"scoped root /build/source/target/marmel-path-tests-N sits inside the
temp allow-list entry /build". Pointing TMPDIR at
/tmp/marmel-check-tmp, outside $NIX_BUILD_TOP, restores upstream's
premise, so all 8 tests keep running for real instead of being filtered
out of cargoTestFlags. Build-time only; the installed binary is
untouched.

Two derivation comments carried stale facts and are corrected:

- meta.license claimed the repository shipped no LICENSE file and that
  Cargo.toml had no license field. Upstream c98179d023 (Na1w/marmel#40)
  added a root LICENSE file and declared license = "MIT" in Cargo.toml,
  so the comment now records that MIT is asserted in the README,
  Cargo.toml and LICENSE alike; lib.licenses.mit was already the right
  value.
- the cargoTestFlags comment quoted "the lib suite passes at 1055
  tests". Roughly 500 test fns landed since (TUI, fs-path, PTY and
  dispatch-integrity suites), so the figure is dropped rather than
  refreshed: nothing checks it and it rots on every bump.

Also adds the two research notes behind the bump under docs/: the
de3e99c8..afcc1dc9 upstream delta audit and the update-tooling findings.

Verified: clean nix build .#marmel including the serialised test suite
(--test-threads=1) — 2146 passed, 0 failed.
2026-11-07 16:54:02 +03:00
alex 1edac7a69d feat(agnix): ship agnix-lsp and agnix-mcp with the CLI
CI / check (push) Has been cancelled
Editors and MCP clients need the language server and the MCP server, and
upstream's release job builds all three crates in one cargo invocation.
Build -p agnix-cli -p agnix-lsp -p agnix-mcp in the same derivation so the
workspace compiles once. Closes nix-overlay-ocf.
2026-11-03 15:33:28 +03:00
alex 9c39cff0b0 feat(agnix): add agnix 0.60.3 built from source
CI / check (push) Has been cancelled
Add the agnix agent-config linter to the overlay. Built with
rustPlatform.buildRustPackage from the v0.60.3 source tarball
(cargoBuildFlags -p agnix-cli), so no upstream pre-built release
binaries are fetched. Closes nix-overlay-wfv.
2026-11-03 15:14:49 +03:00
alex 43199e5b8b feat(pdfcraft): add PdfCraft 0.18.0 built from source
CI / check (push) Has been cancelled
Packages the storytold/pdfcraft Cargo workspace at tag v0.18.0: the pdfcraft
desktop app and pdfcraft-cli, with the desktop entry, AppStream metadata,
icons, licences and the two CC-BY-SA OCR models the app reads from
share/pdfcraft/models. No upstream binaries are downloaded.

Two deviations from upstream's build instructions:

- The three [patch.crates-io] entries upstream's own Cargo.lock records as
  [[patch.unused]] (egui_demo_lib, egui_inspection, egui_system_font_provider)
  have no package in nixpkgs' partial vendor tree to resolve against, so cargo
  aborts. postPatch drops them; they are unused, and cargo drops the matching
  lock entries itself, so the build runs without --locked.

- nixos-unstable's rustc 1.98.1 and 1.99.0 reject every AVX-VNNI intrinsic
  ("intrinsic signature mismatch for llvm.x86.avx512.vpdpbusd/vpwssd"), and
  rten-gemm, pic-scale and zlib-rs all use one. The package therefore builds
  against a new nixpkgs-rust195 input (nixos-26.05, rustc 1.95.0), the same
  pattern as nixpkgs-torch211. Follow-up: nix-overlay-g4w.

Also adds a Home Manager module (programs.pdfcraft) to
homeManagerModules.default, and documents both in the README.
2026-11-03 13:48:55 +03:00
alex d05ca58ee6 chore(kyojin): bump 1.7.1 -> 1.8.5
CI / check (push) Has been cancelled
Upstream 1.8.4/1.8.5 are iGPU decode and returning-session speedups with no
packaging change: setup.py still reads EXL3_ROCM_DEV_INCLUDE/EXL3_HIP_DEFINES,
the pyproject runtime deps are unchanged, and tools/{glm,mimo,qwen}/serve.py
are still the serve entry points. Only the pin and source hash move.

Verified on gfx1151: nix build .#kyojin, importing exllamav3_ext against the
pinned ROCm torch 2.11, kyojin-serve-qwen --help, and nix build
.#kyojin.tests.jit.
2026-11-02 16:16:00 +03:00
alex 7056685370 feat(memdoor): add memdoor 1.769.0
CI / check (push) Has been cancelled
Add guregodevo/memdoor, a Go coding/automation agent, built from source:

- the binary embeds the React UI through web/embed.go (`//go:embed
  all:dist`), and upstream gitignores web/dist, so the bundle is built
  with buildNpmPackage and copied into web/dist before the Go build
- CGO stays off: the SQLite driver is modernc.org/sqlite, so FTS5 needs
  no C flags and the binary is static
- upstream's whole Go suite runs (buildGoModule would trim checks to
  subPackages, where cmd/cli has no tests); it needs git on PATH and
  separate scratch dirs for HOME and the Go temp root, because one test
  writes ~/.memdoor and ProjectRoot walks up looking for that marker

Track the package in beads as nix-overlay-yxj.
2026-11-02 01:53:57 +03:00
alex 89b2f8a725 chore(awgen): bump 0.3.1 -> 0.3.2
CI / check (push) Has been cancelled
Upstream tagged v0.3.2 (5f97208) in alex/awgen. The fix prices allowed-IP
progress per progress instead of per whole route list; dependencies are
unchanged. cargoHash still moves because the crate version line in
Cargo.lock feeds the vendor staging hash.

Verified: nix build .#awgen is green, and result/bin/awgen --version
prints 0.3.2 (upstream now sets the crate version from the tag, so the
stale "prints 0.1.0" caveat from the 0.3 bump no longer applies).
2026-10-26 09:00:01 +03:00
alex 559994a66f chore(awgen): bump 0.3 -> 0.3.1
CI / check (push) Has been cancelled
Upstream tagged v0.3.1; refresh src and cargo hashes (dependencies
changed between the two tags). Verified with nix build .#awgen and
result/bin/awgen --version.
2026-10-25 17:36:56 +03:00
alex 387edce283 chore(awgen): bump 0.2 -> 0.3
CI / check (push) Has been cancelled
Upstream tagged v0.3 (3d10c5f) in alex/awgen. The dependency set is
unchanged, so cargoHash stays the same; only the source hash needed a
recompute for the new tag.

Verified: nix build .#awgen is green, cargo checkPhase reports 29 passed
/ 0 failed, and the resulting binary runs. Upstream Cargo.toml still
declares 0.1.0, so `awgen --version` prints 0.1.0 while the derivation
version (taken from the tag) is 0.3.
2026-10-25 17:06:26 +03:00
alex e2f24b0876 chore(packages): bump pins to current upstream releases
- awg-tool 0.4.0 -> 0.5.0
- awgen 0.1 -> 0.2
- codeaf 0.7.1 -> 0.8.2
- ds4 -> ee7e840 (2026-10-21)
- freebuff 0.2.19 -> 0.2.34
- graphify 0.9.77 -> 0.9.95 (add pyyaml dependency)
- hipengine 0.6.1 -> 0.8.3
- kubernetes-mcp-server 0.0.67 -> 0.0.69 (build with go_1_27)
- kyojin 1.3 -> 1.7.1 (drop obsolete setup.py/pyproject patches, ship
  gr_mix_hip from the package, add ninja runtime dependency)
- marmel -> de3e99c (2026-10-23)
- omniroute 3.8.50 -> 3.8.52
- open-code-review 1.12.12 -> 1.13.3 (build with go_1_27)
- pilotdeck 2026.10.17 -> 2026.10.20
- radar 1.16.2 -> 1.18.0
- traycer 1.4.2 -> 1.5.1

Verified: nix build passes for every bumped package; kyojin.tests.jit
compiles all five JIT kernels.

🤖 Generated with Codebuff
Co-Authored-By: Codebuff <noreply@codebuff.com>
2026-10-24 14:09:34 +03:00
alex fc00731a45 chore(llama-cpp): remove package
Drop the llama-cpp derivation and its README entry.

🤖 Generated with Codebuff
Co-Authored-By: Codebuff <noreply@codebuff.com>
2026-10-24 14:09:11 +03:00
alex eefef7cce2 feat(awgen): add AmneziaWG configuration generator package
CI / check (push) Has been cancelled
Package alex/awgen (offline, stateless Rust CLI for AmneziaWG configs) from
the upstream Gitea repository. Source is fetched with fetchFromGitea and
pinned to the v0.1 tag; the cargo dependency hash is pinned as well.

Also list the package in the README package table and project tree.
2026-10-20 21:55:24 +03:00
alex 5310b47924 feat(pilotdeck): build Electron desktop app from source
CI / check (push) Has been cancelled
Replace the prebuilt .deb web-only package with the real Electron
application built from upstream v2026.10.17:

- compile apps/desktop main process (tsc) and launch it with
  nixpkgs electron_42 in app-dir mode; upstream env overrides
  (PILOTDECK_DESKTOP_RUNTIME_ROOT, PILOTDECK_DESKTOP_NODE) select the
  staged runtime and bundled Node.js
- replicate the upstream runtime staging (build-runtime.mjs): prod-only
  hoisted node_modules from the dedicated runtime lockfile via a second
  fetchPnpmDeps, dist/skills/ui tree, source-import rewrite, pinned Cua
  driver, upstream runtime verification probes
- share electron-updater into the app by expanding pnpm virtual store
  scopes of its dependency closure
- wrap bin/pilotdeck (desktop, --no-sandbox, ozone hint auto) and keep
  bin/pilotdeck-server (headless web server) on the same runtime
- add pilotdeck.desktop entry, hicolor icon, startupWMClass

Verified: nix build passes, app launches under Xvfb reaching
"PilotDeck Web UI ready" with the window mapped as WM_CLASS
pilotdeck-desktop; headless server answers on :3001.
2026-10-20 00:13:31 +03:00
alex a46a2bb967 fix(pilotdeck): build from source
CI / check (push) Has been cancelled
Replace the prebuilt Debian package with the documented source build. Build the web UI and gateway with pnpm, stage the pinned Cua driver, and start the web UI from the wrapper.
2026-10-19 21:52:37 +03:00
alex fd74633e8d feat: add PilotDeck package
CI / check (push) Has been cancelled
Package the upstream Linux Electron release for x86_64 and aarch64 as
pilotdeck. The Debian payload includes the bundled Node.js runtime and web
UI; the wrapper adds the Electron runtime libraries to LD_LIBRARY_PATH and
installs a desktop entry.

README lists the new package.
2026-10-19 21:07:13 +03:00
alex adbed5f3c0 chore(marmel): bump to b08923ea (main), drop obsolete patches
CI / check (push) Has been cancelled
Upstream fixed both reasons for the local patches, so neither is needed:

- 62cb75634c6e (#16) reports a missing TLS trust store instead of
  panicking through reqwest::Client::default(); the env-cleared test
  child now stays reachable without SSL_CERT_FILE, which is what
  session-boot-trust-store.patch existed to work around.
- b08923eaef60 (#17) adds /nix/store and /gnu/store to
  BUILTIN_READONLY_PATHS; the checkPhase rebuild confirms the store rule
  works on a Nix host: test_internal_sandbox_cross_directory_rename and
  test_internal_sandbox_extra_readonly_paths_never_grant_write run for
  real, no SKIP line is printed, and only the /etc/resolv.conf DNS test
  stays filtered. Lib suite: 1055 passed, 1 filtered; all integration
  suites green.

passthru.updateScript now passes the bare attribute name: nix-update
rejects ".#marmel" with AttributePathError, so the documented bump path
was broken. Verify with `nix-update --flake marmel --use-update-script`.
2026-10-19 09:52:45 +03:00
alex cb3f338e57 chore(marmel): bump to 981b35b (main)
CI / check (push) Has been cancelled
Track upstream main at 981b35b0d356e56de429ab4198e427ccfbe7d85f
(2026-10-17). Cargo.lock is unchanged between the two revisions, so
cargoHash stays; src hash regenerated by nix-update. Version string
follows nix-update's 0-unstable-* dated scheme now.

Both downstream patches still apply and upstream issues stay open
(Na1w/marmel#14, #13, fix PR #16 unmerged), so nix-overlay-pnb and
nix-overlay-nn2 remain blocked. Lib suite now passes at 1030 tests
with the same single DNS sandbox filter; comment updated.
2026-10-18 14:51:58 +03:00
alex 51dd95c19c fix(marmel): allow /nix/store in the Landlock sandbox
The upstream ruleset allow-lists FHS paths only (/usr, /bin, /lib, /opt,
/etc, /var), and every executable here resolves below the store: /bin/sh
is a symlink into /nix/store/<hash>-bash-5.3p9/bin/sh. Landlock matches
the resolved target rather than the symlinked view, so the harness
re-entry (`marmel --internal-sandbox-exec`) cannot exec the shell and
every shell tool call comes back as "Failed to exec shell in sandbox:
Permission denied (os error 13)" with exit 1. The sandbox is on by
default, so the only workaround on a Nix host is MARMEL_DISABLE_SANDBOX=1,
which drops the protection altogether.

The patch adds "/nix/store" to the read-only+execute list, guarded by
exists() like its neighbours. Store paths are world-readable by design,
so the added reach is nothing a local user did not already have. Checked
against the built binary: sandboxed exec works, git runs, workspace
writes work, ~/.ssh and /root stay blocked.

Tested: nix build .#marmel passes, lib suite 966 passed / 0 failed with
one filter. test_internal_sandbox_cross_directory_rename no longer needs
a skip and passes; test_internal_sandbox_exec_dev_null_and_dns stays
skipped for its own reason, the no-network build sandbox has no
/etc/resolv.conf. Temporary downstream fix, drop it once upstream takes
the store path (nix-overlay-5ml).
2026-10-16 16:19:18 +03:00
alex 892f30e82a fix(marmel): bump to 3f238ff, give e2e child a trust store
CI / check (push) Has been cancelled
Track main at 3f238ff (per-role reasoning_effort, #12), 21 commits past
d6627d7 ("1.0.0 (#6)"). cargoHash is unchanged: Cargo.lock is identical
across that range, so the vendored tree is byte-for-byte the same.

The new tests/test_session_boot.rs spawns the real binary with
env_clear(), which drops SSL_CERT_FILE. rustls-native-certs is then left
with a fixed list of /etc paths to probe, none of which exist in the Nix
sandbox, so the child dies in reqwest::Client::new() and four exit-code
assertions observe the panic's 101. The patch forwards the runner's
SSL_CERT_FILE to the child and leaves every assertion intact.

The underlying defect is upstream's rather than Nix's:
src/llm/client.rs builds its client with unwrap_or_default() on a
Result<Client, reqwest::Error>, and reqwest's Client::default() is
Client::new(), which repeats the same build and .expect()s it. The
fallback can therefore never recover; it turns a reportable "no CA
certificates" error into a panic. Reproduced on a stock busybox image
with no /etc/ssl at all, so it hits any minimal container, not just
sandboxed builds.

Tested: 1251 tests pass across 31 targets, 0 failures.
2026-10-16 14:32:15 +03:00
alex 0813f55762 fix(kyojin): Isolate JIT and runtime toolchains
CI / check (push) Has been cancelled
Allow a pinned JIT SDK without rebuilding PyTorch or the extension.
Keep runtime device bitcode unchanged: exporting TheRock bitcode into
the serving process crashes CLR 7.2 on a basic tensor operation.

Disable unsupported expandable segments and check all five HIP kernel
modules, with an optional GPU matmul and module-loading regression test.
2026-10-08 17:17:56 +03:00
alex 4c24610aae fix(kyojin): serve wrappers get a C compiler, hipcc and the JIT sources
CI / check (push) Has been cancelled
env.sh assumes a distro box: a system cc for Triton's runtime C builds, a hipcc
for the JIT HIP kernels and the repo root on PYTHONPATH. The wrappers now export
all three (, ,  for the amdgcn bitcode,
PYTHONPATH=${src}/gr for gr_mix_hip).

The wheel ships only the files pyproject lists, so the .hip kernel sources, the
qsa_proof marker gating the QSA prefill kernel, the dense-GEMM tuning seed and
gr/ never reached site-packages: kernels fell back to Triton and every start
re-tuned dense GEMM for 7 minutes.

Verified on the Strix Halo box: server READY, start-up 713 s -> 103 s, warm-up
418 s -> 4 s, 584-token prompt served over /v1/chat/completions with no
'unavailable' fallbacks.
2026-10-08 15:57:52 +03:00