llama.cpp on a jailbroken PS5: run any GGUF model natively on the console, starting with Qwen 3.8
Fleet validation — 2026-10-21–22
A record of the original validation. Since then the RPC patches moved to
patches/fleet/, the worker became part of every app build (from v0.11.0; v0.10.0 had a separatePS5LM_FLEET=1build), andscripts/test-fleet-native.shandtest-fleet-title.shwere folded intoscripts/test-native-build.sh(FLEET.md).
Result
A working host-side increment: one model split over two real CPU RPC processes,
including the actual python3 -m fleet serve path. Experimental PS5 worker source
is integrated. The complete RPC-enabled native PS5 title now cross-builds and
passes static ELF/FSELF validation. No console inference has been tested.
At the time of the initial local validation below, no public listener or console
deployment was performed. Hosted CI results are reported separately in the PR
checks. This is not a hardware-validated PS5 release.
Revisions
| Component | Commit |
|---|---|
| PS5LM base | c4595784ce75c1d58777b11ef812d3e087e74296 |
| llama.cpp | 04f7226a2c12c7f382c545a6db5a034266aa53cb |
| ps5-homebrew-ui | bf44d80384d5b8f8f4f213cf98d748887211328d |
All existing llama.cpp patches 0001–0009 and new patches 0010–0012 were applied.
fleet-build.json records the exact patch hashes and local host binary checksum.
The wire protocol remains upstream RPC 8.0.0; a handshake is not revision
attestation. Both ends must be built from the same sources.
Passed
Final aggregate command (Linux sandbox, no system-wide package installation):
CMAKE="$PWD/.deps/host-tools/cmake/data/bin/cmake" \
NINJA="$PWD/.deps/host-tools/bin/ninja" \
scripts/test-fleet.sh --rpc
Public CMake 4.4.4 and Ninja 1.13.2 wheels were installed under .deps/host-tools
because those host tools were absent. On a normal configured system the command
is simply scripts/test-fleet.sh --rpc. NumPy and PyYAML supply fixture generation only.
| Check | Outcome / scope |
|---|---|
| Python unit/security suite | 52 passed: GGUF parsing, unsafe/unknown layouts, download policy/framing/cleanup (mocked HTTPS), budgets, KV, split mapping, preflight failures, binary verification, SIGTERM child cleanup |
| Worker configuration | 43 checks passed, no listener |
| Worker runtime/file tests | 19 checks passed, fake ggml backend, no GPU/socket |
| Worker sanitizers | ASan + UBSan passed; LeakSanitizer disabled because sandbox ptrace prevents it |
| Worker compile | Linux C++17 compile against real pinned ggml headers with -Wall -Wextra -Werror |
| Existing model planner | Original model_plan_test synthetic-header test passed |
| CPU/RPC build | llama-server, llama-cli, ggml-rpc-server, test-rpc-multi-server built |
| Upstream RPC regression | test-rpc-multi-server passed |
| Server-side NO_COMM regression | Rank 0 and rank 1 peer commands refused; ordinary device-count RPC remained usable; test ports loopback/reserved |
| Real two-worker inference | Passed, not mocked; actual tensor allocation and graph execution on both processes |
| Actual fleet CLI | Passed manifest → preflight → plan → llama-server → inference with new workers |
| Cleanliness | git diff --check, shell/Python syntax checks passed; no leftover llama-server/RPC test processes |
Inference evidence
Fixture: 231,136-byte, four-block dense llama GGUF with deterministic random
untrained weights, 32-dimensional embeddings, 4 query heads, 2 KV heads,
256-token model context. It is generated locally by tests/make_fleet_fixture.py;
no trained model was downloaded. Tests used context 128, one sequence, FP16 KV,
two CPU RPC workers, and tensor-split 3,7.
Verbose direct-RPC logs assign:
| Logical layer | Device |
|---|---|
| Blocks 0, 1 | RPC0 (worker A) |
| Blocks 2, 3 | RPC1 (worker B) |
| Output slot 4 | RPC1 |
Each direct worker made 3 buffer allocations and 9 graph compute/recompute calls. The local reference, direct two-worker execution, and actual fleet CLI all returned token IDs:
282, 131, 193, 131, 193, 131, 193, 131
All 40 top-five log-probability values per run matched the local baseline, with
maximum absolute error 0.0 here (test tolerance 1e-5). This proves the tested
execution path, not useful text quality, larger-model correctness or performance.
Evidence is in build/fleet-final-validation.log, build/fleet-smoke/result.json,
its worker/controller logs and JSON responses, plus
build/fleet-host/fleet-build.json. Re-running the scripts regenerates this
evidence locally; generated models, binaries, logs and SDK archives are not
committed to the repository.
CI preparation — 2026-10-22
Re-ran scripts/test-fleet.sh --rpc successfully after adding the bounded
GitHub Actions workflow: the same 52 Python, 43 configuration and 19 runtime
checks passed, as did both RPC regressions and the actual fleet CLI inference
comparison (maximum log-probability error 0.0). The workflow passed actionlint
v1.7.7 and YAML permission/event checks. Actions use verified commit pins;
the fixture dependencies are pinned in tests/requirements-fleet.txt. This is local pre-publication evidence;
refer to the PR checks for the hosted run.
Hosted dependency correction — 2026-10-22
The first fork run built all C++ targets and passed the safety/RPC regression
checks, but fixture generation exposed a missing transitive gguf import:
PyYAML. The workflow now installs a pinned fixture requirements file, and the
test script checks imports before compiling. The complete suite was re-run
successfully in a fresh Python virtual environment containing only those
requirements; both inference comparisons again had maximum error 0.0. See the
PR/fork Actions run for the hosted result on the current commit.
Initial full PS5 title build attempt — 2026-10-21
The upstream tree had no checked-in CI workflow. Its documented route is a Linux cross-build using PS5_Vulkan, its pinned forks, RADV and homebrew SDK tooling. The workflow now separates host inference, native-object compilation and the full-title cross-build. The first two jobs alone do not prove a title build. The evidence below records the initial local full-title build attempt; the SDK/runtime gaps were subsequently resolved for the native-object gate below.
- Ran
bash scripts/build-app.sh: stopped because UI submodule was uninitialized. - Initialized its pinned commit and read its
AGENTS.md; reran the official build. - The upstream OpenGL SDK endpoint returned HTTP 404:
https://github.com/blackbearreloaded/ps5-opengl/releases/download/v1.0.0/ps5-opengl-sdk-1.0.0.tar.gz - Followed the alternative already documented in
docs/BUILDING.md:https://github.com/MyLuxy/overdrive-deps/releases/download/ps5-opengl-v1.0.0/ps5-opengl-sdk-1.0.0.tar.gz - Downloaded exactly 226,263,775 bytes, bounded by 250 MB / 120 seconds. Computed
archive SHA-256 matched the submodule's pinned value:
f93643c04c843d56143b00951df1f8042ea7706ae9f19e4e9abf158e1ead77c5. - Ran official
fetch-opengl-sdk.sh: extraction and pinned manifest verification passed (manifest SHA-256f4b91f672be037fbac3f82494f1225deaf4c227a03f37ac3ffa56abb213b943f). - Reran official
scripts/build-app.sh: exit 2, exact next blocker:.deps/src/PS5_Vulkan/.deps/native/radv-release/lib/libvulkan_radeon.ps5.amissing.
At that initial attempt, also absent under .deps/src/PS5_Vulkan:
build/host/ps5-native-tool,
runtime/libc.prx, .deps/native/ps5-payload-sdk/bin/prospero-lld and its C++
headers. The compiler did not run in that initial full-title attempt. Downloading
the standalone legacy payload SDK
would not replace the native Vulkan title dependencies. No unverified PS5 binary
is included.
Reproduction: follow docs/BUILDING.md to prepare the pinned PS5_Vulkan/Mesa/SDK
sources and host packages, apply patches/mesa, run PS5_Vulkan's
tools/setup-native-dependencies.sh, tools/build-radv.sh release and make app,
then run this repository's scripts/build-app.sh. The verified OpenGL mirror
resolves the removed archive, not the remaining RADV setup.
Build evidence: build/fleet-ps5-build-attempt*.log,
build/fleet-opengl-mirror-download.log, build/fleet-opengl-mirror-verify.log.
Native-target follow-up — 2026-10-22
Using the public documented bootstrap (not a proprietary SDK or firmware dump):
| Component | Pinned input |
|---|---|
| PS5_Vulkan | f410e67a0241b3e8edcdbcd2a9b285dfc4a412ed |
| PS5_PayloadSDK fork | b83202be73e930050e00de0f1b4d9ec46c0391bf |
| Base homebrew SDK v0.42 archive SHA-256 | 8cfbc7cd5811e719eb4f0c47eea668d3dc7b40bc8ab11c4a5031d40c23ec02da |
| zlib 1.3.2 archive SHA-256 | bb329a0a2cd0274d05519d61c667c062e06990d72e125ee2dfa8de64f0119d16 |
fleet_worker.cpp, patched ggml-rpc.cpp, and patched transport.cpp compile
successfully through PS5LM's actual ps5cxx wrapper. Target macros include
__PROSPERO__, __SCE__ and __FreeBSD__=9; outputs are x86-64 relocatable ELF
objects, not executable PS5 titles. The worker's compile uses -Werror.
The clean-room runtime reproducer builds two copies and verifies upstream's
recorded hashes, including signed libc.prx:
e6ff45d16adf687855cc3b33b0c8a4132b6504360b221e0a34c7e99fb3ba0036.
No Sony implementation or console dump is downloaded. Local Clang 19 validation
passed; the repeatable scripts/test-fleet-native.sh also passed from a fresh
dependency cache using Clang 18.1.8, which is the object-only CI compiler family.
The separate native job's status is in the fork Actions checks.
The object-only check must not be described as a full build, console boot, Vulkan validation or PS5 emulation. The full-title follow-up below is separate. QEMU's documented system/user modes do not provide the PS5-specific AGC/loader/direct-memory environment, so a generic Linux/FreeBSD VM is not an equivalent hardware gate.
Full native title cross-build — 2026-10-22
scripts/test-fleet-title.sh passed locally using public Debian 13/Clang 19
packages and the documented upstream build scripts:
- Mesa base
1ba9874f478657ccf6da72ce55f0260164cc0bd0plus all threepatches/mesacommits yields deterministic source revision9e0a7c63bc6d4b6bcb60831163bb3fe9f46e91ea. - All 389 host shader-tool tasks and 794 PS5 RADV tasks completed; the archive includes the actual PS5 winsys, not a stub or host Vulkan replacement.
- The real title CMake build exposed
strnlenhidden by_XOPEN_SOURCE=600. Patch 0011 replaces it with the equivalent bounded C++ scan. The faster native check now uses those actual CMake feature-test macros, so it catches regressions. - The patched Vulkan/RPC/server libraries and native console dashboard compiled and linked into a complete title folder. Optional downloaded browser assets are omitted, not the console UI. No trained model or proprietary firmware was downloaded.
eboot.elfandeboot.binboth pass upstreamtools/inspect.shwith zero static errors: FreeBSD OSABI 9, ABI 2, type0xfe10, aligned LOADs and required process metadata. The native tool's FSELF check also reports valid integrity.
The first local passing eboot.bin was 100,329,398 bytes, SHA-256
58a2bd3aec133b8d1bffe7a63f534fb3ff8271857ca3bf74b595055b263a0317.
That identifies this local artifact, not a promise of bit-identical builds across
paths/compiler versions. Only the upstream clean-room runtime has a fixed
reproduction checksum. Package binaries, SDKs and logs are not committed.
Evidence: build/fleet-native-full/title-validation.log and
build/fleet-title/result.json. Reproduce using the package list in the
workflow's digest-pinned Debian ps5-title job and scripts/test-fleet-title.sh.
The fresh hosted run, rather than this local cache-backed aggregate, establishes
whether the entire CI bootstrap works from scratch; see fork Actions results.
This is a real cross-build/package check, not a console boot or hardware test.
Review fixes and remaining boundaries
Independent review plus real host execution found and corrected:
- Missing binary checksum in the host manifest.
- Wrapper SIGTERM orphaning the server.
- 16-node plans exceeding the pinned CLI's accepted split count (now max 15).
- Unmodeled shared RoPE weight duplication (those layouts now rejected).
- Confusing RPC buffer names with CLI device names (actual names RPC0, RPC1).
- Server peer-communication commands opening an additional wildcard listener (server-side NO_COMM rejection plus forced worker/controller setting).
- PS5/BSD socket
sin_leninitialization and SIGPIPE portability concerns.
Not verified: actual console boot/runtime import resolution, driver/device memory allocations, Vulkan kernels, actual worker-screen rendering, real network drops on PS5, thermal behaviour, sustained inference, large GGUF models or multi-console performance. Fresh HTTPS GGUF acquisition is covered by offline downloader tests, not a live download/inference integration run. The OpenGL SDK download is separate and does not validate the model acquisition path.
Qwen 3.8 27B is not supported by this fleet estimator: PS5LM lists its GGUF
architecture as qwen35, whose hybrid/recurrent memory model is intentionally not
guessed. See FLEET.md for the narrow supported layouts and extension path.
Next hardware gate
With a title built as above: use an isolated wired network, two consoles, one controller and a small admitted dense model. Verify both RPC handshakes, actual per-device allocation logs, identical greedy output against a local reference, memory high-water marks, disconnect/close recovery and title restarts between models. Only then test a model too large for one console; use measured memory reserves and benchmark prefill/decode separately. Nothing in the current host results substitutes for that gate.