Skip to content

armgit / PS5LM

Public Imported from GitHub
0

llama.cpp on a jailbroken PS5: run any GGUF model natively on the console, starting with Qwen 3.8

28.7 KB
Raw Blame

Changelog

Versions follow semver. The app's version is set in ps5/app/version.hpp and shown on the dashboard; each release is tagged v<version> and ships ps5lm-app-v<version>.zip (see scripts/release-app.sh). v0.1.x are the upstream payload releases by cobanov.

[0.11.0] - 2026-11-02

Added

  • Settings > Debug report: one file to attach to a GitHub issue, /data/PS5LM/debug-report.txt and a copy on a USB drive when one is in. It holds the app's version, the system software, where the app keeps its files and the room on each drive, the models and their plans, the downloads, the settings, env.txt and app-args.txt, the last 256 KiB of app.log and llama.log, and which payload ports answer (ELF loader, FTP, kernel log, shell, ps5-exporter, ...). API keys, tokens and passwords are taken out of everything in it. The kernel log goes in only when Include the kernel log is on (off by default), at most its last 1 MiB, read from klogsrv or etaHEN's. Before the first report a dialog says what it holds.

  • Ternary Bonsai 2: a Bonsai 2 preset (PrismML's sampling, 64k context at most) and Bonsai 2 27B in Suggested, upstream's Q2_0 (7.1 GiB). The app's llama.cpp already reads the prism.hadamard.* rotation those files carry: checked on a PC with the 2B and the 27B, on the CPU and on ggml-vulkan with the console's shader paths. On a console, fetched from Suggested (the download carried on across two restarts of the app): a 64k context with a q8_0 cache, 1.36 GiB of the pool left once it serves, about 17 tokens a second, right answers with and without thinking.

  • scripts/bonsai-pq2-to-q2.py: PrismML's PQ2_0 files, which upstream llama.cpp (b11827 too) refuses, to Q2_0, with no loss: each 128-weight block becomes two 64-weight blocks with the same scale and codes. Python only, about 80 s for the 27B; the 2B and the 27B come out byte for byte as the Q2_0 files PrismML publishes. The library says what to do with a PQ2_0 file (and that a PTQ1_0 one does not load) instead of failing at load.

  • scripts/bench-ubatch.py: prompt and generation speed and memory of one model at several prompt batch sizes on the console (2,048-token prompt, each size a restart with the planner's arguments), as a table; and scripts/ps5lm-app.sh memory, which logs the memory pools on request.

  • The app's own storage when the HEN keeps /data closed (#16, #23, #24). The app used to stop there: no log, no settings, every download failing. It now keeps its files in /download0/PS5LM instead, and on the first such start asks, over the screen: use the app's storage, or how to open full storage (OnionHEN's exact_title_ids, etaHEN, other HENs) and start again, which asks the HEN once more. Settings > Sound and storage > Where PS5LM keeps its files shows which one is in use and how full it is. Models saved there stay in the library once full storage opens. Checked on a console with /data made unwritable to the app: the question on the screen, a model searched for, fetched to /download0/PS5LM/models and loaded from there, the choice kept across the restart, and the model still in the library once /data opened again.

  • Two probes for later decisions, out of the app's way (ps5/app/probes.cpp, run from the PC): scripts/ps5lm-app.sh bench-read <model file> measures a drive's sequential and random read speed at the sizes MoE experts come in (64 KiB to 4 MiB, one and four threads, cached and O_DIRECT), and bench-memory logs what the system keeps of the title's memory and the largest block that really allocates. The download and write probes moved there from main.cpp.

  • The TV's request feed shows OpenAI Responses requests (/v1/responses, what codex sends): the question from input, the answer from the response or its output_text deltas. They showed blank before. The API page lists the endpoint, and docs/CLI.md says what it takes.

  • A PC's GPU as a fleet worker. PS5LM_FLEET_GPU=metal (or cuda, vulkan) scripts/build-fleet-host.sh adds that backend, so ggml-rpc-server -d MTL0 serves the PC's GPU next to the consoles. The default host build stays CPU only. Run with a console: a PS5 and a Mac's GPU (Metal), and a PS5 and an RTX 3090 over CUDA and over Vulkan, each serving qwen2.5-0.5b-instruct-q4_k_m.gguf with blocks 0–19 on the PS5. See docs/FLEET.md.

  • scripts/ps5lm-app.sh launch works again, with PS5LM's own launcher payload (ps5/app/launch_app.c, built by build-app.sh) instead of the removed ps5-homebrew-dev-protocol.

Changed

  • The fleet worker is part of the app. There is no separate fleet build or ps5lm-app-fleet zip any more: every PS5LM can be a fleet worker, and stays the normal app unless /data/PS5LM/fleet-worker.txt asks for one, as before. PS5LM_FLEET is gone from scripts/build-app.sh, and CI builds one title again.
  • llama.cpp moves to b11827 (970b51ce7, 2026-10-22), 219 commits on, among them ggml-vulkan's fixes for GPU-visible host memory (#30468) and concurrent first device init (#29766), its strided f16 KV copy on AMD (#27703), and upstream's own tojson floats as Python writes them (#30362, #30483), so patches/0006-jinja-tojson-floats-as-python.patch is gone; patches/0005 is re-based on the new server-http.cpp, and patches/0009 (introspection) on the new server-context.cpp. It builds and links for the console as before.
  • The planner counts the compute buffers' growth with the prompt batch (-ub, about 38 bytes a token per unit of feed-forward width, measured with ggml-vulkan on the console's shader paths), and the batch for language models is one constant, kPromptUbatch, 64. scripts/bench-ubatch.py measured the others on a console (Qwen 3.8 27B UD-IQ2_XXS, 64k, 2,048-token prompt): 124.6 tokens a second at 64, 107.1 at 128, 122.8 at 256, 102.0 at 512, and the larger sizes leave less memory, so it stays 64. See docs/CONSOLE.md.
  • downloadDataSize in sce_sys/param.json is 16 GiB (16384 MiB, was 256 MiB): the size of the app's own storage. The app counts what it holds itself, as a title's statvfs is a stub. The system makes it an 18.8 GB image, /user/download/PPSA99575/download0.dat, at the first start; whether all of that is taken on the drive at once was not measured. Lowering it later frees nothing.
  • A download is refused before it starts, with the reason on the TV, when the drive has no room for it, the downloads already queued to that drive, and 1 GiB to spare. A model and its image projector are queued together or not at all.
  • More memory for models, expected about 0.56 GiB: sce_sys/param.json no longer declares the amm block and the 256 MiB CPU and GPU page tables it carried over from ps5-homebrew-ui's template. With them the system gave the app 11.44 GiB of direct memory (the 12,280,922,112 bytes in #23's log); titles that leave them out, PS5_Vulkan's and PS5_RetroArch's, get 12 GiB, and neither the GL stack nor RADV uses what the blocks configure. The planner reads the memory free, so every model's plan grows with it: with the dashboard up, v0.9.1 had 11.04 GiB free before a model and gave Qwen 3.8 27B UD-Q2_K_XL a 32k context, not the 64k it was suggested with; with the 12 GiB pool it plans 64k again. On a console the pool is 12.0000 GiB and the planner has 11.6 GiB for a model (11.0 before); app.log records the pool and the flexible memory at start (memory at start) and once the model serves (memory when ready).
  • Every line the app writes to app.log starts with the console's local time, as 22:56:01 ps5lm-app: start (#40). Lines from the graphics stack keep their own form.

Fixed

  • A black screen on OnionHEN with PPSA99575 in exact_title_ids (#23, #32, #54). OnionHEN opens /data with a full app jailbreak that also moves the process out of its sandbox, so /app0 and /download0 were gone and the fonts did not load. Once /data is open, the app finds its sandbox in /mnt/sandbox/<title id>_<slot> and reads its own files from there; app.log says when it runs outside its sandbox. Unchanged on etaHEN, which keeps the sandbox. Not yet run on an OnionHEN console.
  • A download that meets a full drive anyway stops at once, "Failed: App storage is full", instead of retrying twenty times.
  • At the end of a download, Models > Downloads said "starting" for the minutes the SHA-256 check takes, so it looked hung (#30). It now says "Checking the SHA-256, 45%". The queue is saved before the check, so a restart during it checks the file again but no longer fetches the last piece first.
  • codex failed on Qwen models with "System message must be at the beginning" (#55): it sends two developer messages after its instructions, and Qwen's template refuses a system message that is not first. patches/0004 folded such messages into the first one only for the Anthropic API (Claude Code); it now does so for Chat Completions, Responses and Anthropic alike, with tests/server_system_messages_test.py checking all three on a host build. docs/CLI.md's codex setup is the one codex 0.166 reads: the profile in its own ~/.codex/ps5.config.toml. Checked with codex 0.166 against a console.
  • Models with sliding-window attention got far less context than fits (#57): the planner counted every attention layer at full length. It now reads <arch>.attention.sliding_window_pattern and counts the window layers at their window, as llama.cpp keeps them. Spark-X2.6 4B (27 of 36 layers keep 512 tokens) plans 256k with a q8_0 cache instead of 128k with q4_0; on a console the server took the 262,144-token context with 2 GiB of the pool left, and found a needle in a 52,821-token prompt. Longest context in Settings goes to 192k and 256k, and the planner tries both.
  • A console serving a model went into rest mode an hour after the last button press: requests from the network are not controller input. Rest mode ends the app, and on a jailbroken console the controllers do not connect again until a restart. While a model is loaded, a download is queued or running, or the app is a fleet worker, it now tells the system every 30 s that it is in use (sceSystemServicePowerTick, as a video player does); with nothing to do, the console's own rest-mode setting applies. app.log says when it starts doing so.
  • scripts/bench-ubatch.py took the old app's last lines, as it stopped, for the restarted one, so a size that loads was reported as not loading, or its prompt went to the stopping server; it never removed app-args.txt (ftpsrv answers a delete with 226, which ftplib refuses), and an error mid-run left the test arguments for every later start. All three are fixed.

[0.10.0] - 2026-11-01

Added

  • Fleet, experimental: one model across several consoles. A PC runs the controller (python3 -m fleet serve) and one llama-server; each PS5 runs the fleet build of the app as an RPC worker that holds a contiguous run of the model's layers on its GPU. Dense llama and qwen2 GGUFs only for now; Qwen 3.8's qwen35 layout is refused by the planner. The fleet build is a separate zip, ps5lm-app-fleet-vX.Y.Z.zip (PS5LM_FLEET=1 scripts/build-app.sh); the normal app has no RPC code. A worker starts only when /data/PS5LM/fleet-worker.txt names the console's own private address, a port below 40000 (the PS5 refuses higher ones) and the acknowledgement, because RPC has no authentication: keep it on an isolated network. docs/FLEET.md is the guide. By @ntindle in #35.
  • An opt-in endpoint for the model's own numbers: POST /v1/introspect answers with the whole logit row at the positions you name, the top candidates with their probabilities, the entropy, and one row of a named graph tensor (l_out, ffn_swiglu, ffn_out, attn_out, attn_norm, ffn_norm, result_norm, result_output) for a named layer. A request can also change one channel in place, with zero, scale or add, at one position or at all of them. messages applies the chat template first. It is off by default: Settings > Developer > "Logits and activations API" (a change restarts the app with the loaded model), or LLAMA_INTROSPECT=1. docs/INTROSPECTION.md is the reference, scripts/ps5lm-introspect.py is a client, and examples/interpret/neuron_ablation.py zeroes the FFN neurons (ffn_swiglu) of one layer one at a time.
  • The patch reads the logits from the context's own output buffer and copies the tensors out of the compute graph with the eval callback that llama.cpp already has, so no work is repeated and no request pays for it unless it asks.

Fixed

  • Downloads keep their speed past 2 GB. The console slows a title's own writes to about 1.2 MB/s after the first ~2.2 GB, while a payload's keep going, so the app now writes each piece through a loopback ftpsrv: in file order, with REST/STOR at its place, done once ftpsrv answers 226. It uses the user's ftpsrv on 2121 when a test write through it works, else its own copy, ps5lm-ftpsrv on 127.0.0.1:2123, which it sends to elfldr when needed and which leaves other ftpsrv copies on 2121 and 2122 running. With an ftpsrv that has no EPSV it uses PASV. When no ftpsrv answers, the app writes the file itself as before. Ported from PS5SD (#41).
  • Models saved to an M.2 or USB drive show up in the library wherever that drive is mounted. The library looked only in /mnt/ext0, /mnt/usb0 and /mnt/usb1, while downloads can go to any drive the exporter reports; it now also reads the download folder and every mounted drive's PS5LM/models (#43).
  • The SHA-256 check at the end of a download shows how far it has read, once a second, instead of a line that did not move for minutes on a large file. The queue is saved when the check starts, so a restart during the check repeats only the check, not the download (#30, #51).

[0.9.1] - 2026-10-20

Fixed

  • A full log queue no longer stops the server. llama.cpp's log holds 512 lines, and the app's writer prints each one to app.log and llama.log; when it fell behind, the server thread waited for room, so /slots and /metrics stopped answering while /health still did, the hang in #20. On the console a full queue now drops the line and the writer says how many it dropped (patches/0008-log-never-blocks.patch, by @bilawalriaz in #28).
  • Storage on consoles where the HEN does not open /data to the app by itself (#16, #23, #24). The app asks the HEN at every start, the restart for a model included, by writing {"PID":n} to /download0/etahen_jailbreak, which OnionHEN honours for a title on its allowlist: add PPSA99575 to app_jailbreak.exact_title_ids in /data/OnionHEN/config.ini. If /data stays closed, a notification on the TV says so and names the OnionHEN setting. From @abhi245y's PS5LM Plus.
  • /data/PS5LM is created when it is missing.

Changed

  • ps5-homebrew-ui and ps5-opengl are gone from GitHub: the submodule now comes from a mirror at the same pinned commit, and docs/BUILDING.md says where the ps5-opengl SDK archive is still found.

[0.9.0] - 2026-10-16

Added

  • A new screen, designed from scratch: Home (the model, its speed, and every request and its answer streaming as it happens), Models (on this PS5, suggested models measured on the console, Hugging Face search and downloads), API (the endpoints and an example) and Settings (with the logs under Diagnostics). Manrope type, interface sounds from PocketVibe, and a full-screen splash while the app restarts for a model.
  • Decision models: llama.cpp's /v1/systemone (TypeSafe's API, the one Ollaya serves) with d1, d1 Omni, Clef, OpenJev and others. The planner gives them a 2048-token batch and a 4k context.
  • MiniCPM-V 4.7's image projector, and the other models of llama.cpp 04f7226.
  • /data/PS5LM/env.txt: driver and ggml switches set before the GPU opens, for measuring.
  • Console page: temperature, power, GPU and CPU load charted over the last two minutes or the last hour, and a table of every reading.
  • Home's request feed reads as a chat, and the console can ask itself: Square sends a demo request, Triangle opens a keyboard (with Turkish letters) for a message of your own.
  • Settings with several choices open a list under the row; Cross picks, Circle cancels.

Changed

  • The app is PS5LM under its own title ID, PPSA99575, with a new icon and home screen background. PS5LM Plus (PPSA99581) stays @abhi245y's app.
  • llama.cpp moves to 04f7226; RADV carries three of PS5CEMU-HAR's driver patches (patches/mesa), among them freed buffers waiting for the GPU.
  • Models with an image projector get a 2048-token ubatch, so images are no longer cut to 64 tokens.

Fixed

  • The app opens on Home, not on the model list, and draws its splash before the GPU opens, so a restart for a model no longer shows a black screen.
  • Home shows the load as it happens, how much of the model is on the GPU, instead of a still splash, and offers Demo and Message only once the server answers: requests sent while the model loaded were lost.
  • One sound for each kind of action, at one pitch: no two actions share a sound by accident.
  • A float in any request ("temperature": 0.7) stopped the server: the console's localeconv() reports an empty decimal point. The app now gives the C locale's.
  • An image projector llama.cpp cannot read no longer stops the server: the app restarts once and loads the model text only.
  • scripts/build-app.sh fetches the OpenGL SDK through ps5-homebrew-ui's script, creates build/, and no longer needs ImageMagick.

Known issues

  • A decision model can stop answering after about 55 requests in a row on the console (the server waits on the GPU); restart the app or load the model again. Language models are not affected.

[0.8.3] - 2026-10-13

Changed

  • The app is now called PS5LM Plus, to tell it apart from Mert Cobanov's original PS5LM payload: on the home screen, in the dashboard header and in the console's notifications. The title ID (PPSA99581), the data folder (/data/PS5LM) and the repository stay as they were, so updating keeps settings, models and downloads.
  • The dashboard header shows the chat's address without "Chat at http://", so it clears the page tabs with any IP address.

[0.8.2] - 2026-10-13

Changed

  • contentVersion in sce_sys/param.json now follows the app's version (major+1, minor, patch: 0.8.2 is 01.008.002). It had stayed at 01.000.000 since the first release, so stores such as ProsperoStore could not tell an installed copy was out of date. scripts/release-app.sh refuses to package a release whose contentVersion doesn't match.
  • README: new screenshots from v0.8.1 (Qwen 3.8 27B generating, the library, the loaded-model panel) and a shorter text.

[0.8.1] - 2026-10-13

Fixed

  • The chat server stopped answering after the network dropped: when the console's link went down, accept() failed with SCE_NET_EINACTIVEDISABLED (errno 163, the same 0x804101a3 a dropped download reported), cpp-httplib took that as fatal and closed the listening socket, and the model stayed loaded with nothing to reach it. llama-server now binds again once a second until it works (patches/0005-listener-survives-network-drops.patch). Checked on the console: the listener came back within a second after each of dozens of drops. A reply streaming at the moment of a drop is still cut.

Added

  • README: the web UI with a sample chat.

Found

  • The repeated "Ethernet disconnected" notifications are the console's link going down and up (eth0: link state changed to DOWN in the kernel log, up to eight times in a few seconds), with or without PS5LM running: a cable, port or switch problem, not the app.

[0.8.0] - 2026-10-13

Added

  • Downloads page (the fifth on L1/R1): a queue kept in /data/PS5LM/downloads.json, each entry with its speed, time left and state. Cross pauses or resumes (or tries a failed one again), Square twice cancels and deletes the partial file, or clears a finished one.
  • Downloads survive a dropped connection, a pause and the app restarting: the file is fetched in 32 MiB pieces on two connections, the part done from its start is saved with the queue, and a lost connection is retried on its own (4, 8, 16, 32, then 60 s apart, up to 20 times in a row). The SHA-256 is checked once the file is whole.
  • Get models: Square checks whether the file under the cursor fits, and the row says CHECKING while its header is read.
  • A busy screen (a spinner and what is happening) while a model is switched or unloaded, before the restart for "unload and download", and while the models are planned again after a context or cache change.
  • README rewritten around the app: features, the payloads it needs, controls. The payload's README moved to docs/PAYLOAD.md.

Fixed

  • A download that lost its connection failed and threw away what it had (2.0 of 10.4 GB, sceHttp2ReadData 0x804101a3); now it carries on.
  • A connection that went quiet could hang a download: requests now time out after 30 s without data and are asked again.
  • Settings froze the app for seconds on every change, each one reading every model's header again. Only the context and cache settings re-plan now, in the background behind the busy screen; the others apply at once.

Measured (download speed on the console)

  • The PC on the same network: 19.6 MB/s from Hugging Face on one connection.
  • libSceHttp2 (HTTP/2), one connection: 2.8 to 5.9 MB/s, varying more from run to run than with any setting. Reads of 256 KiB gave 0.65 MB/s (each read waits about half a second), 4 MiB and larger 3 to 6 MB/s. Four requests on one template share one connection (3.5 to 6 MB/s); on four templates, made one at a time, 2.8 MB/s; on two, 4.2 MB/s. Made in four threads at once, the templates crashed the app.
  • libSceHttp2 takes at most 4 requests (512 KiB pool); 8 or 16 are refused (0x817b121c).
  • libSceHttp (HTTP/1.1): 0.44 MB/s per connection whatever the read or receive block size, adding up with more connections (4: 1.6, 8: 2.5 MB/s, one of the eight refused).
  • In the app, two pieces at once on their own templates: 7.0 to 9.8 MB/s for a 1.9 GiB file (v0.7: 3.8 MB/s), still well under the PC's speed. Going further needs sockets and TLS of our own, for a later release.

Checked on the console

  • A 1.9 GiB download: paused (the partial file stopped growing), resumed from the saved 128 MiB, closed and relaunched mid-download (carried on from 905 MiB), finished with its SHA-256 matching.
  • Get models during a download: search, listing and fit checks work.
  • The busy screen on a context change; the library re-planned behind it.
  • Payloads stayed up throughout.

[0.7.0] - 2026-10-10

Added

  • Downloads check free space first: the file and 1 GiB to spare must fit on the drive. When it does not, a dialog offers the drive with the most room for this one download; the download itself refuses too, for the after-restart path.
  • M.2 and every mounted USB drive: Download to in Settings steps through the drives that are mounted (internal, /mnt/ext0, /mnt/usb0..3), the STORAGE tile and the usage details list each one, and the library groups models by drive. Models in PS5LM/models on the M.2 drive are found too.
  • Before a download while a model holds the memory (under 2 GiB free), a dialog offers to unload it first: the app restarts without the model and the download starts on its own.

Fixed

  • Storage figures were made up: a title's statvfs is a stub that reports 64 GiB with 16 GiB free for every path, and statfs is not exported to titles. Free and total space now come from ps5-exporter (bundled since v0.4.0), which reads the kernel's; internal storage is /user. Without the exporter the space shows as unknown and does not block a download.
  • Get models opened after a restart now lists the popular repositories.
  • A download without a checksum survives the unload-and-download restart.

Checked on the console

  • Real space: internal 166.7 of 627.6 GiB free, USB 461 of 477 GiB, as ps5-exporter reports them; the drive's 33 MB second partition is left out.
  • A 500 GiB job refused on internal storage; a download to the USB drive; Download to stepping between internal and USB, saved; unload and download (SmolLM2 Q2_K); the scratch size warning and Clear.
  • No M.2 drive is fitted, so that path is checked in the preview only.

[0.6.0] - 2026-10-10

Added

  • Get models page (the fourth on L1/R1): Hugging Face's GGUF repositories, most downloaded or by family (L2/R2), each file's size and whether it fits, planned from its first 16 MiB, and downloads to internal storage or USB (Settings), SHA-256 checked, added to the library when done. HTTPS through the system's own libSceHttp2 and libSceSsl. Checked on the console: SmolLM2-135M F16 (271 MB) at 3.4 MB/s.
  • Chat tools: the browser chat's model can read, write, edit and search files (llama-server's built-in tools), confined to /data/PS5LM/scratch by patches/0003-tools-root.patch. Settings has a switch, a size warning (default 2 GiB) and a Clear action.
  • Images: a vision model's mmproj file beside it is loaded with it (--mmproj), and the planner counts its memory.
  • The planner can plan from the first bytes of a GGUF (a range request), the base of the model market's fit column.

Measured

  • Writes from the app: stdio 18.5 MB/s, write() 232 MB/s on /data (USB: 293 and 621 MB/s). Downloads use write().
  • Chat tools on the console: a write to /data/homebrew lands in the scratch folder; relative paths start there. Gemma 4 E4B with its projector described a test image correctly.

[0.5.0] - 2026-10-10

Added

  • docs/CLI.md: opencode, codex and Claude Code against the model on the console, each checked on a small edit with Qwen3.8-27B UD-IQ2_XXS.

Fixed

  • Claude Code got HTTP 500 on every request: it sends system messages in the middle of a conversation and Qwen's template refuses them. patches/0004 folds them into the first system message.

[0.4.0] - 2026-10-10

Added

  • POWER card: the SoC's power (CPU, GPU and memory together), its history, energy used this session and joules per generated token; expanded, min, average and max, fan duty and every temperature sensor. The reading is sceKernelGetSocPowerConsumption's low 32 bits in milliwatts (22 W idle, 86 W while the 27B generates), through ps5-exporter.
  • USAGE card: GPU and CPU on one chart; expanded, GPU, CPU, memory, storage and the models on disk.

Changed

  • USAGE replaces the GPU and CPU tiles; THERMALS shows the hottest sensor.

Fixed

  • Speed and GPU use during a long reply: llama-server's counters move only when a request ends, so the live figures come from the slot's token count.

[0.3.0] - 2026-10-10

Added

  • Pages on L1/R1: Dashboard, Settings and Logs, with tabs in the header.
  • Settings page (rail and rows after the kit's Control Room design): load a model at launch and which one, the planner's longest context and KV cache type, interface sounds; saved to /data/PS5LM/settings.json.
  • Logs page: app.log and llama.log, scrollable, with an errors-only filter.
  • Model details on Cross over the Model tile: every llama-server argument, the plan, cache and preset; Triangle opens the library.
  • THERMALS tile (SoC and CPU temperature, fan duty), STORAGE tile (free space on /data and USB), and the CPU clock. A title may not read the sensors (0x80020002), so the app bundles Marice's ps5-exporter (GPL-3.0, pinned v0.2.0) and sends it to elfldr when it is not running.
  • scripts/ps5lm-app.sh press <button>: drive the screens from a PC.

Changed

  • One MEMORY tile replaces MEMORY and GPU MEMORY (the GPU and CPU share one pool); CONTEXT takes in the session counts.

[0.2.0] - 2026-10-10

First release of the native app (title PPSA99581), from this fork.

Added

  • Native PS5 app running llama-server on the GPU through ggml-vulkan and RADV, with an 11.44 GiB unified device heap.
  • Dashboard on the TV (ps5-homebrew-ui): generation speed, GPU busy, CPU use, memory pool, GPU memory split, session, context and model tiles.
  • Model library with internal storage and USB drive sections; models load straight from USB.
  • Planner that picks context length and KV cache type for the memory free.
  • Sampling presets for Qwen 3.5/3.6/3.8, Gemma 4, gpt-oss, Granite, Nemotron, Mistral and Llama.
  • Model switching by restarting the app (in-place unload leaks GPU memory).
  • scripts/ps5lm-app.sh to drive the app from a PC, and the version on the dashboard and in the log.

Fixed

  • Unloading during a reply no longer hangs the app: it waits at most 10 s for llama-server before restarting.

Changed

  • A plain launch loads no model and opens the library.