With -lm none the GGUF is read() into our own buffers, but the kernel keeps
the file in page cache afterwards, charged to the pod's cgroup: about 20 GB
after a 35B load. On 2026-10-24 the Talos OOM controller killed the pod right
after a load under exactly that pressure. Those pages are pure waste once the
data is copied.
llama_file::drop_cache() wraps posix_fadvise(POSIX_FADV_DONTNEED) (Linux only,
errors logged at debug). load_all_data() drops each copied tensor's range right
after reading it, so the cache never grows past about one tensor, and drops the
whole file at the final cleanup to catch the metadata and partial pages at
tensor edges. Both only when the loader holds no mapping of the files
(mappings.empty(): not mmap, not mmap+mlock, no lazy tensors), so nothing that
is mapped is ever advised away. Split GGUFs are covered (every file in files).
clip.cpp always copies the mmproj via ifstream, so it drops the whole file
after loading it.
Measured on the pod, gemma-4-E2B UD-Q4_K_XL (2.5 GiB) + mmproj BF16 (0.9 GiB),
CPU only, -lm none, cold cache, same binary with the drops disabled as the
control (4 rounds, rotated order):
- file pages resident after load (mincore): 2431-2444 / 940 MiB -> 0 / 0
- peak model-file residency during load: 2.5 GiB -> 1.2-1.5 GiB (largest tensor)
- pod cgroup file after load: +3.37 GiB -> +0; host MemFree after load
0.6 GB -> 2.5-2.8 GB (MemAvailable counts clean cache, so it cannot show this)
- load time: 5.87 s -> 6.82 s median (+0.95 s); a post-load-only drop is
6.56 s (+0.68 s). The cost is synchronous page freeing that reclaim would
otherwise do later; -lm dio (direct IO) costs the same (+0.9 s) and does not
cover metadata or the mmproj.
- greedy output byte-identical; -lm mmap keeps the model file fully resident
and mapped (2498 MiB, no refaults); -lzm on skips the drop (file mapped).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Comments only, plus one log string:
- the arbiter block comment said the KV cache is never evicted; with
LLAMA_SLEEP_EVICT_KV it is (and the scheduler and scratch with it)
- vram_go_cold said vram_ensure_warm() rebuilds the encoder; only
encode_mtmd_chunk() restores it, and only for an encode
- vram_ensure_warm and the mtmd init block named process_mtmd_chunk; the
restore lives in encode_mtmd_chunk
- handle_sleeping_state said waking happens lazily before decode; it
restores eagerly on the wake transition, before update_slots()
- the vram_cold ordering note now says why it still holds with the inline
want-lock probe
- the sleep log line said "keeping KV cache" even when LLAMA_SLEEP_EVICT_KV
evicts it; it now names what actually happens (the prefix
"server entering sleeping state" that fork-validate.sh counts is kept)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two ways a shutting-down server kept the VRAM arbiter's token, stalling every
other model:
- SIGTERM with a request in flight: terminate() stopped only the task queue.
The HTTP handler kept waiting for results the loop would never send, the main
thread blocked joining it, and the token was released only in the destructor
after that join - i.e. never, until the client went away. Terminate the
result queue too (readers now get "server is shutting down": a 503, or an SSE
error event), and let wait_until_no_sleep give up once stopping.
- SIGTERM while waiting for the token: the loop thread sat in a blocking flock
inside the wake callback, holding mutex_tasks, so terminate() waited for the
current holder to finish and then for a pointless restore of our weights.
Poll the lock non-blocking and abort the wait on shutdown; the aborted wake
stays cold, skips the restore, and the loop exits before update_slots.
Also release the token right after start_loop() returns rather than in the
destructor, and only after the device buffers are actually freed: the old
order unlocked before destroy(), so the next model could start uploading into
VRAM that was still occupied. Freeing at exit uses a new
llama_context_discard_device(), which drops weights, KV, MoE cache and scratch
without capturing host shadows first.
Off production: SIGTERM mid-generation or mid-wait now exits in ~0.2 s with
exit 0 and frees the token within 5-70 ms (previously never / after the
holder's whole generation). Handoff latency and greedy output are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A SIGTERM that lands while the start_loop() thread is inside a sleeping
callback deadlocks the server. The callbacks run under
server_queue::mutex_tasks, and the kernel delivers the signal to that same
thread; the handler calls terminate(), which locks mutex_tasks again, and the
thread waits on itself forever.
With the VRAM arbiter that window is wide - a whole device release, or a
blocking wait for the GPU token - and llama-swap sends SIGTERM to evict
models, so it hit production: an evicted model wedged mid-release, never
released token.lock, and every other model queued behind it. SIGTERM could
not stop it either; only SIGKILL did.
Make the handler only write the signal number to a self-pipe, which is
async-signal-safe, and run shutdown_handler on a dedicated thread. terminate()
there simply waits until the loop thread drops mutex_tasks, after which the
loop exits normally. The thread is joined before the objects the handler uses
are destroyed. If pipe() fails the old direct call is kept.
Reproduced 6/6 off production (SIGTERM during release and during the token
wait); with the fix 8/8 clean exits, and handoffs, image requests and greedy
output are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The doorbell was edge-triggered: a waiter wrote <arena>/doorbell/<pid> and the
holder's warden turned the inotify event into a queue flag. A ring carried no
notion of "still wanted", so it was lost whenever the holder was not already warm
and listening - during its model load (the watch does not exist yet, and the
kernel does not queue events for a watch that is not there) or inside
restore_device - and any ring that survived past the point its sender had been
satisfied caused a spurious release. Lost rings wedged the waiter permanently,
because flock(LOCK_EX) never times out and, with idle-sleep disabled, the holder
had no other reason to release. The wait blocks under mutex_tasks, so the whole
server stopped answering while /health still returned 200.
Express the request as kernel state instead. A waiter holds <arena>/want.lock
shared while it waits and drops it once it owns the token; the sleep decision
probes that lock non-blocking and releases the GPU while anyone is waiting. The
probe needs its own fd - flock treats two open file descriptions of one file
independently, so probing on the waiter's fd would convert our own lock rather
than conflict with it. Nothing can be missed, nothing goes stale, and a waiter
that dies is cleaned up by the kernel.
Probe from should_sleep() on the loop thread rather than from a warden thread.
Routing it through a flag is what made the first attempts fail: start_loop()
holds mutex_tasks from should_sleep() through the callbacks to the wait, so a
warden's request_yield() blocks on that mutex and is admitted only after the flag
has been consumed, latching a release for the next wake. Reading live state where
the decision is made has no edge to latch, and drops the warden's poll latency.
Two sleep-path bugs this exposed: wait_until_no_sleep() waited on !sleeping but
the loop clears req_stop_sleeping on the way in, so a loop that slept again
before the waiter ran stranded it forever - re-ask on every wake. And a task
queued after the waiter saw us awake could not wake us by itself, so sleep now
also breaks on a non-empty queue. Hold the GPU for 100 ms after a wake: the
request that woke us is not queued yet, and yielding at once only sends it round
again.
Measured on the RX 580 pod, two servers contending, 60 alternating handoffs:
0 stranded, median 0.309 s against the doorbell's 0.314 s.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
The vision encoder sits idle in VRAM on every text request. With
LLAMA_MMPROJ_ONDEMAND=1 the server keeps the encoder in RAM (released
after load) and brings it into VRAM just-in-time before mtmd_batch_encode,
releasing it again afterwards - so the encoder's VRAM is free for KV /
expert cache on the common text-only path.
When the encoder does not fit (that VRAM has been claimed), evict the LLM
backbone weights first: they are not needed while the encoder runs (the
decode that uses them happens afterwards), their host shadow is read-only
(cheap free, no D2H), and the KV / prompt cache is left untouched. Order
on the way back matters: free the encoder before re-uploading the weights
so the peak stays within VRAM.
Also wire the encoder release/restore into the VRAM arbiter: vram_go_cold
releases it; the just-in-time encode-path restore replaces the previous
restore in vram_ensure_warm.
Assisted-by: Claude
Add clip_release_device/clip_restore_device (and mtmd_release_device/
mtmd_restore_device wrappers over the vision + audio contexts) that free
the multimodal encoder's device weight buffer to a read-only host shadow
and rebuild it on demand, using the same shadow/free/reallocate pattern
as llama_model weights. No-op for a CPU-backed encoder. This lets the
server drop the ~hundreds-of-MiB vision encoder from VRAM when it is not
encoding an image.
Assisted-by: Claude
With LLAMA_SLEEP_EVICT_KV the KV cache device buffers are freed when a model
goes cold. update_slots() touches the KV before the decode-time vram_ensure_warm
(e.g. SWA models create a checkpoint that reads the KV via ggml_backend_tensor_get),
so the KV must already be resident by then. Move the restore to the sleep-wake
handler (handle_sleeping_state(false)), which runs before update_slots, fixing a
GGML_ASSERT(buffer) crash on iSWA models (gemma) when KV eviction is enabled.
Assisted-by: Claude
Before uploading a model's weights, acquire the shared VRAM token (ring the
doorbell so any resident model releases first). Previously load uploaded weights
to VRAM before the arbiter was active, so loading a large model while another
(e.g. the warm 4B task model) held VRAM could exceed the budget and OOM.
- add vram_arena_open() (idempotent flock/doorbell setup) and
vram_acquire_for_load(), called from load_model() before
common_init_from_params().
- a coordinated load now stays warm holding the token and serves its first
request without a re-warm (drop the init-time go_cold cold-start).
Assisted-by: Claude
The request handler only calls wait_until_no_sleep() (which wakes a server out
of its sleeping state) when sleep_idle_seconds >= 0. But the VRAM arbiter's
cross-process doorbell can put a server to sleep even when idle-sleep is
disabled, so without this a doorbell-slept server would never wake and requests
to it would hang until timeout.
Do not bypass the wake path when LLAMA_SLEEP_VRAM_ONLY is set, so the arbiter no
longer depends on --sleep-idle-seconds being configured.
Assisted-by: Claude
Extend on-demand device residency to the KV cache so that when a model's KV
plus another model would not fit in VRAM, the KV can also be evicted to a host
shadow (D2H on release, H2D on restore) instead of only the weights.
- llama_memory_i: add release_device_buffers()/restore_device_buffers()
(default no-op). Implemented in llama_kv_cache (D2H shadow of the live
ctxs_bufs, freed and reallocated like the weights); llama_memory_hybrid and
llama_kv_cache_iswa delegate to their child caches.
- llama_context::release_device(evict_kv): also evict the memory's device
buffers when requested; restore_device() rebuilds them. Public API
llama_context_release_device gains an evict_kv flag.
- server: LLAMA_SLEEP_EVICT_KV=1 enables it. Off by default (weights-only),
since the KV shadow adds a D2H/H2D copy of the live cache each cycle.
Validated on RX 580 (Vulkan), 4B @ 32k ctx: weights-only cold VRAM 1750 MB
(KV stays); weights+KV cold VRAM 726 MB (KV freed, ~1 GB reclaimed). KV
survives the round-trip: prompt cache reused after the cycle (prompt_n 4 vs
42), correct output.
Assisted-by: Claude
Add release/restore of a model's GPU weight buffers (keeping a host shadow
and the KV cache) so several always-loaded llama-server processes can
time-share a single GPU without reloading or losing the prompt cache.
- llama-model: release_device_weights()/restore_device_weights() capture a
compact host shadow (stable iteration order, view-skipping) and free then
realloc the device weight buffers; weights_resident() query.
- llama-context: release_device()/restore_device() wrappers; decode() auto-
restores; public C API llama_context_release_device/restore_device.
- server: LLAMA_SLEEP_VRAM_ONLY makes idle-sleep release only the VRAM weights
(not a full unload/reload). A cross-process flock token in LLAMA_VRAM_ARENA
enforces "resident iff holds token"; an inotify doorbell forces the holder
to release on contention. The warden thread only touches the task queue, so
releases run on the loop thread and never race a decode.
Validated on RX 580 (Vulkan): two models share 8GB, never both resident,
correct output under contention, KV cache preserved (no re-prefill).
Assisted-by: Claude
* mtmd: support hibiki realtime duplex
* fix and clean up
* fix long session quality
* fix ci
* tests : add the moshi fusion baseline for Metal
The Metal fusion check found 5 RMS_NORM+MUL fusions in the new moshi
arch and no baseline for them.
---------
Co-authored-by: Pascal <admin@serveurperso.com>
* convert: qwen3-asr: put "language ..." output into thinking block
* convert: qwen3-asr: read the chat template from models/templates
The converter embedded its own copy of Qwen3-ASR.jinja. Read the
bundled file instead, as the Kimi-K3 and Kokoro converters do, so
the GGUF and test-chat share a single template.
---------
Co-authored-by: Pascal <admin@serveurperso.com>
* mtmd: drama3 voice cloning
The Fish S2 codec encoder (fish_s2_spkenc) now also serves drama3:
- conversion: drama3 writes the codec encoder and the audio encoder KV, shared with fish_s2
- fish_s2_spkenc: outputs the codes of the clip (gen-side reference) and, in a drama3_gen mmproj,
builds the drama3 backbone rows (E_lm[semantic] + E_cb[1][code 1]) / sqrt(2)
- drama3_gen tower: teacher forcing when the caller gives the codes of the next frame
- drama3 pipeline: reference-bracket-v1 prompt with one reference (vLLM-Omni fish_omni), the
reference frames go through the backbone as codec rows and through the tower with their codes
- mtmd_gen_state_get_ref_codes(): read-only access to the reference codes of a gen state
Claude-Session: https://claude.ai/code/session_01PpX3DmaxNSmJbxGaxnUBrX
* fix correctness
* clean up
* rm redundant code
* fish_s2: rm unused import
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
llama-tts built the mtmd context from mtmd_context_params_default(), which hardcodes 4 threads, and never copied n_threads over, so -t never reach the mtmd side.
Copy params.cpuparams.n_threads into mtmd_params.n_threads, as mtmd-cli and the server do; with no -t this now follows the resolved default.
Verified: -t 1 now runs single-threaded, -t 4 is unchanged, and the produced WAV is identical.
Assisted-by: Claude
- Allow cls.output.weight to be absent from the mmproj. It is an optional
Qwen3 reranking head that the YuE2 NAR decoder never uses.
- Without this fix, mtmd_init_from_file() fails with
"tensor cls.output.weight is missing from the mmproj", preventing
audio generation from starting.
* cuda: add scatter for batched tensor writes
* update test case to mirror get_n test case
* save restore timing logs
* refactor code into batch-copy.cu file
* llama-bench : size the batch to the prompt for models without memory
Assisted-by: Claude Opus 5.5
* server : read up to 16 lfm3-d1 questions in one pass
Assisted-by: Claude Opus 5.5
* server: fix lora and recurrent
* tests: run the hybrid SWA checkpoint test with one token per ubatch
With and without the cache, the prompt was not split into the same
ubatches, and the generated tokens are decoded one at a time while the
uncached run processes them in a batch. On CUDA this moves the top
logprobs of the tiny model by up to 0.03 and swaps near ties, so the
test failed while the cache behaved as expected.
With one token per ubatch both runs go through the same kernels and
the output is compared exactly.
---------
Co-authored-by: Pascal <admin@serveurperso.com>
* jinja: impl scope-chain key lookup
* add test for call_statement lambda
* chat: add tts_dict for kokoro G2P template
* more testing
* value_float_t range check
* no json-in-meta, add llama_model_meta_arr_str/_n
* mtmd: add kokoro tts support
* detailed guide about G2P and text norm port
* refine G2P
* clean up
* migrate batch api
* nits
* Added support for Qwen3-TTS ICL/reference mode
Assisted-By: Claude Opus 5.5
* fix correctness
* attach transcription to bitmap
* nits
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* sampling : fix crash when both ignore_eos and grammar are on
* server : rework, throw 400 when grammar and ignore_eos are on
* sampling : move the check into the sampler (review suggestion)
* tests : remove redundant tests
* ui: replace deprecated packages and fix npm audit findings
Swap the mdast stub for @types/mdast, move the katex override to 0.19.0,
and override glob, source-map and html-encoding-sniffer under
workbox-build and http-server so installs no longer pull deprecated
packages. Apply the non-breaking npm audit fixes to the lockfile.
Assisted-by: Claude (Opus 5.5 Medium)
* ui: migrate Llama Chat to SvelteKit 3 and its built-in service worker
Move to SvelteKit 3, Vite 8, TypeScript 6 and Storybook 10.6. The
SvelteKit options move from svelte.config.js into sveltekit() in
vite.config.ts, $lib imports become #lib subpath imports from
package.json, and $app/environment becomes $app/env. Modules named
server.* or living under a server/ folder are server-only in SvelteKit 3,
so the llama-server components, store and enums move to llama-server
names.
@vite-pwa/sveltekit and Workbox give way to SvelteKit's own service
worker in src/service-worker, registered by usePwa(), which precaches the
app files and leaves API requests to the network. A small plugin emits
manifest.webmanifest. static/sw.js removes the worker of older builds, and
llama-server serves a self-removing service-worker.js as well as sw.js
when the built-in UI is off or replaced.
Assisted-by: Claude (Opus 5.5 Medium)
* ui: use SvelteKit 3's update detection and Svelte 5's current patterns
A page the service worker doesn't control now learns of a new build
from SvelteKit itself: `updated` from $app/state checks
_app/version.json on navigation, on focus and every hour. It replaces
the version Llama Chat kept in localStorage and compared by hand.
Reload on the update prompt pulls in the new worker and activates it
before reloading, so one click lands on the new build, whether or not
a worker controlled the page.
Following Svelte's current guidance:
- usePwa keeps its state in $state rather than a store, and no effect
checks `browser`
- nearViewport, autoResizeTextarea and the custom CSS injection are
attachments ({@attach}) rather than actions (use:)
- class lists use arrays rather than class: directives
- the chat contexts come from createContext, typed, in place of
setContext/getContext with string keys
- ActionIconCopyToClipboard takes $props() rather than `export let`
Assisted-by: Claude (Opus 5.5 Medium)
* ui: move to ESLint 10
ESLint 9 is out of support. Every plugin the app uses supports 10:
typescript-eslint, eslint-plugin-svelte, perfectionist,
simple-import-sort, eslint-plugin-storybook and eslint-config-prettier.
@eslint/js and @eslint/compat follow.
ESLint 10's recommended rules ask for two things the code now does:
errors rethrown from a catch keep the caught error as their `cause`,
and no variable starts with a value that every path overwrites.
Assisted-by: Claude (Opus 5.5 Medium)
* ui: override minimatch to clear npm audit's last finding
ESLint 10.14, @eslint/config-array 0.23.5 and typescript-eslint 8.72
depend on minimatch 10, which GHSA-g3hv-75wg-m249 (denial of service
via globparts amplification) affects below 11.0.1. Their fixed releases
move to minimatch 11 but are all younger than the repo's 7-day
min-release-age, so an override pins minimatch 11.0.5, the newest the
rule allows. The three only call new Minimatch(), match() and
minimatch(), which minimatch 11 keeps. npm audit now finds nothing.
Assisted-by: Claude (Opus 5.5 Medium)
* mtmd: support YuE2-3B
* improve speed
* add CFG
* fix streaming
* add cfg control from api
* check ctx size
* fix n_predict
* add flow_steps config
* nits
* fix batch api
* nits
* adapt
* mtmd: yue2 samples from its own saved logits
The context logits belong to the last decode of any slot, so with
several server slots the score and the unguided codes were sampled
from another sequence. Keep the logits of each own decode in
save_logits(), which also does the guidance mix.
* mtmd: let the yue2 snake chain fuse
The beta reshape landed between sqr and mul, so the mul, sin, sqr,
mul, add chain was not consecutive and the backends never fused it.
Expand both reshapes first, like qwen3tts does.
---------
Co-authored-by: Pascal <admin@serveurperso.com>
* ui: move tools/ui to ui/apps/chat
The chat app moves to its own root folder, ui/apps/chat, as the first step of splitting the UI into one library
(ui/lib) and the apps built on it (ui/apps). The code is unchanged; every path that pointed at tools/ui now points at
ui/apps/chat: CMake adds it from tools/, and the Dockerfiles, the Nix package, the workflows, the labeler, CODEOWNERS,
.editorconfig, the git hooks, the dev and demo scripts and the docs follow it.
Assisted-by: Claude (Opus 5.5 Medium)
* ui: name Llama Chat's package @ggml-org/llama-chat
It was still called llama-ui, the name of the library that now sits beside it
in ui/lib (@ggml-org/llama-ui). The package stays private; the app's display name
and the llama-ui.zip build artifact are unchanged.
Assisted-by: Claude (Opus 5.5 Medium)
* ui: call the built-in web UI Llama Chat in docs, help text and comments
The app in tools/ui is Llama Chat; Llama UI is the component library it is
built from. Prose that said "Web UI", "web UI" or "WebUI" for the app now
names it: the READMEs, SECURITY.md, the --ui help text, the CMake option
description, one server log line, code comments and the Storybook intro.
Names that others depend on are unchanged: the deprecated --webui flag
aliases, the LLAMA_*_WEBUI CMake and environment names, the webui_settings
field, the storage migration keys and the Nix attributes.
Assisted-by: Claude (Opus 5.5 Medium)
* ui: name tools/ui as a former path of Llama Chat in the code-review skill
Changes under the old tools/ui path, as well as tools/server/webui, still
get the Llama Chat checklist.
Assisted-by: Claude (Opus 5.5 Medium)
* ui: name Llama Chat's path in the security policy's exclusions
Llama Chat no longer lives under tools/server/, so the exclusion listed
under it now names ui/apps/chat. What is covered stays the same.
Assisted-by: Claude (Opus 5.5 Medium)
The reasoning effort probe already renders every candidate level, but
it dropped the ones the template raises on, so a request with such a
level (e.g. max on Qwen3.8) reached the template and came back as a
500 on /v1/chat/completions, /v1/responses and /v1/messages.
Keep the raising levels in reasoning_effort_rejected and reject them
in oaicompat_chat_params_parse with an invalid_argument, which the
server answers with a 400 listing the supported levels. Aliases and
levels a template ignores are not rejected, so high on Qwen3.8
still renders as xhigh.
Also trim the output_config.effort test to a single level.