LLAMA_MOE_IDS_SKIP=0 forces the ids readback in sched_copy_experts (the
upstream path) and =log reports each skip on stderr, so validation can show
that the skip engages at prefill and that output is byte-identical with and
without it in one build. Read once (function-local statics); unset changes
nothing. Not set in production.
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>
The MoE cache (--moe-cache-mib, #29887) keeps the most recently used host
experts in device banks. The on-demand VRAM path (release_device /
restore_device, used by the server VRAM arbiter and the on-demand mmproj path)
did not know about it, so its banks were never freed on handoff. Restoring them
by reallocation alone would also be wrong: the banks come back zeroed while the
slot maps still mark experts as cached, so the model would silently read zero
weights.
Free the banks on release and empty the cache instead of shadowing it: re-init
every LRU, re-point its slot-map pointers at the host slot maps and clear those
to -1. The banks are reallocated (which also re-inits their views) and cleared
on restore, before the scheduler is rebuilt, and the cache refills from the
host experts as before. Release runs in both modes, weights-only and evict_kv.
device_resident() now covers the cache too, for the early returns in
release/restore, the auto-restore in decode() and
llama_context_weights_resident().
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
release_device(evict_kv) drops the scheduler, so a context evicted by the VRAM
arbiter and then destroyed reaches the destructor with a null sched. The upstream
"compute buffer size matches expectation" loop calls
ggml_backend_sched_get_buffer_size() unconditionally, whose GGML_ASSERT(sched)
then aborts: the server exits 134 instead of 0 on every shutdown taken while
cold, which under llama-swap makes an ordinary stop look like a crashed child and
leaves a ggml backtrace in the log each time. It also skipped the rest of the
destructor, so the cold teardown path had never actually run to completion.
Guard at the call site rather than relaxing the assert - the assert is right, and
every other caller reserves the scheduler first. Same shape as the null guards in
synchronize(), memory_breakdown() and the released-buffer iteration.
The loop is a diagnostic size comparison, already reported by sched_reserve() at
load, so skipping it when cold loses nothing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
On device release with KV eviction, also call ggml_backend_free_scratch()
on each backend so a cold model drops its Vulkan compute preallocations
(reallocated lazily on the next compute via restore).
Also guard llama_context::memory_breakdown() against a freed scheduler:
evict_kv release resets sched to null, so a cold model that is then torn
down (e.g. terminated by the process manager) hit GGML_ASSERT(sched) in
ggml_backend_sched_get_buffer_type. Skip the compute-buffer accounting
when sched is null.
Assisted-by: Claude
The recurrent (SSM/conv) state of hybrid models (e.g. Qwen3.5) was left
resident when a model's device buffers were released for on-demand VRAM
sharing - llama_memory_recurrent::release_device_buffers() was a no-op
default. Implement it (and restore_device_buffers) with the same
capture-host-shadow / free / reallocate pattern as llama_kv_cache, so
llama_memory_hybrid now evicts both its attention KV and its recurrent
state. The state is read-write, so its shadow is recaptured on every
release.
Assisted-by: Claude
Under LLAMA_SLEEP_EVICT_KV a cold model still held the scheduler's worst-case
compute buffer (hundreds of MiB, e.g. ~700 MiB for gemma-26B) plus the resident
experts. release_device(evict_kv) now also frees the sched (sched.reset() +
sched_need_reserve), and restore_device() rebuilds it via sched_reserve(), so a
fully-evicted model holds essentially no VRAM (gemma cold: 7147 -> 51 MiB).
Trades a heavier re-warm (weights+KV H2D + sched reserve) for the freed VRAM;
weights-only mode is unchanged (fast switch, keeps KV+compute).
Assisted-by: Claude
release_device_weights()/release_device_buffers() leave a null buffer in
ctxs_bufs while the device memory is released for on-demand VRAM sharing.
The memory-breakdown / total_size / clear paths iterated these and called
ggml_backend_buffer_get_size()/get_type()/clear() on the null buffer, tripping
GGML_ASSERT(buffer) and aborting on shutdown of a cold model. Skip null buffers
in llama_model::memory_breakdown and llama_kv_cache::{memory_breakdown,
total_size,clear}.
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
Offloaded expert weights are uploaded per split, and to decide which experts to
upload the copy callback reads the routing ids back to the host. The ids are
produced on the same device we are about to upload to, so the readback forces a
full pipeline flush - 47 of them per eval on a 48-layer MoE.
Once the batch draws enough experts the readback stops telling us anything. At
2048 tokens x 10 experts over 256 experts every expert comes back used, so the
used set is all true and the copy is a single whole-tensor range anyway.
Skip the readback when ggml_nelements(ids) >= 4 * n_expert and mark every expert
used. Uploading an expert that no id selects cannot change the result, since
mul_mat_id only reads the rows the ids point at, so this stays exact. With the
MoE cache on, the all-used set still goes through moe_cache->copy_experts, which
copies the cached experts device-to-device from the host slot maps and uploads
the rest - still exact, and no readback is needed for it. Decode is unaffected:
n_ids there is the number of experts per token, far below the threshold. The
small-batch MoE cache slot path has its own readback and is not touched.
The st.ids/st.used caching is unchanged: a repeat call with the same ids tensor
reuses the all-true set, which is what the threshold would decide again.
Port of dbc449afd ("sched: skip the MoE routing-ids readback when the batch
uses every expert"), whose target block in ggml_backend_sched_compute_splits was
removed upstream (#29943) in favour of this copy callback.
Measured (original, before the port) on an RX 580 (Vulkan, Polaris) running
Laguna-S-2.1 118B IQ2_M with all experts host-resident: pp2048 +4.7% at depth
32k, +0.9% at 16k, tg32 +3%.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
The generic scale pass in llama-model.cpp loads the per-expert
NVFP4 scale2 tensors (ffn_{up,gate,down}_exps_s) and the per-tensor
MXFP8 scales for the dense FFN and the shared expert, but
minimax-m3.cpp never forwarded any of them to the graph.
build_moe_ffn() defaults these to nullptr, so build_lora_mm_id()
skipped its w_s branch entirely and the routed-expert weights came
out global_scale times too large, swamping the residual (reported
G ~= 15123 for M3.1 Flash) and producing token soup.
Forward all scale tensors at the three call sites, matching the
pattern used by flint and the other NVFP4-capable MoE models.
Co-authored-by: origin-labs-ai <origin-labs-ai@users.noreply.github.com>
The amax POOL_1D reduction spans the whole matmul row, so a device cannot
compute it from an axis-0 split. The Meta split-state case reported
MIRRORED, which gives each device a partial max of its own slice. Drop
that case and deny -sm tensor for the arches which build the amax graph,
see [TAG_MOE_AMAX_SCALE].
Assisted-by: dsh:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (MacBook, M5 Max, 128GB)
Add Soprano-80M and Soprano-1.1-80M (https://huggingface.co/ekwek/Soprano-1.1-80M)
as an audio generation model: a Qwen3 backbone plus a Vocos decoder that runs
as a gen-only mmproj.
- convert: detect Soprano checkpoints and export the backbone as qwen3 and the
decoder.pth Vocos weights as the mmproj
- vocab: add the soprano pre-tokenizer (lowercase, whitespace collapse,
per-digit split, [UNK] fallback)
- mtmd: add the Vocos graph and ISTFT, allow loading a gen-only mmproj for
Soprano and include the gen-audio context in mtmd_get_memory_usage
- mtmd-helper: add the Soprano pipeline; the decoder runs over the whole
utterance, so stream mode sends the audio in one chunk after flush()
- tts: document Soprano usage
Assisted-by: Codex
Assisted-by: Claude Opus 5.5
* qwen4exp : build the QSA selection without the kq_mask
* clean
* lang
* qwen4exp : reshape the fused QSA selection to the kq_mask streams
ggml_top_k_mask with a single keep row returns [n_kv, n_tokens], while
build_attn_qsa expects the kq_mask layout [n_kv, n_tokens/n_stream, 1, n_stream].
With several streams the shape assert failed at graph reserve.
* test the QSA optimization
* reword
* update build_fused_top_k_mask instead of handwiring
* simplify
* keep il at the end for consistency
qwen4exp published the hc-wide residual [n_embd, hc, n_tokens] as
t_layer_inp, but the drafter reads layer inputs as [n_embd, n_tokens]
rows: llama_context sizes each embd_layer_inp buffer as n_embd * n_batch
and the DFlash encoder input is n_extract * n_embd wide. Evaluating a
DFlash draft over a qwen4exp target therefore aborts on the first request
in llama_context::extract_layer_inputs.
Collapse the streams with a plain mean before exposing them, as deepseek4
already does for its hyper-connection residual, and build the mean only
for the layers a drafter actually requested.
Assisted-by: pi:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (MacBook, M5 Max, 128GB)
* nemotron : support non-standard GQA, BPE tokenizer and domyn-small pre-type
* gguf-py : write Metaspace prepend_scheme as GGUF metadata
* vocab : parse metaspace prepend_scheme into enum, rename key
* vocab : use metaspace prepend_scheme in detokenize, add TODOs for other tokenizers
* vocab : use constant for metaspace escaped space
---------
Co-authored-by: Martin Cimmino <martin.cimmino@domyn.com>
* Add Flint model (FlintForCausalLM). Assisted by Codex
* cleanup
* one more
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* model : add SenseNova U1 understanding support
Add support for the dense SenseNova U1 understanding branch, including
the U1.5-8B-MoT model.
- Add HF-to-GGUF conversion for NEOChatModel, excluding MoT generation
tensors and concatenating temporal and spatial Q/K normalization
weights.
- Register the sensenova_u1 GGUF architecture and custom spatial RoPE
metadata.
- Implement temporal/spatial Q/K normalization with 4-axis RoPE.
- Keep U1-specific text position handling in the model implementation.
- Add synthetic architecture coverage for sensenova_u1.
Only the dense understanding branch is supported; generation
branches are intentionally rejected during conversion.
* conversion : treat one-expert SenseNova U1 configs as dense
* mtmd : add SenseNova U1 vision support
Add conversion support for the SenseNova U1 vision encoder and implement
its CPU mtmd projector graph.
The graph applies patch embedding, GELU, adjacent-pair 2D RoPE, spatial
downsampling, and dense projection for dynamic-resolution image inputs.
* mtmd : fix SenseNova U1 vision token layout
* model : clean up SenseNova U1 RoPE metadata
Replace the hard-coded spatial RoPE key with shared GGUF and LLM_KV
definitions used by conversion, model loading, and architecture tests.
Clarify why SenseNova U1 retains its custom text-position input: U1 uses
[t, 0, 0, 0], while the generic M-RoPE input uses [t, t, t, 0].
* mtmd : simplify SenseNova U1 projector integration
Map the SenseNova U1 dense projection to the standard mm.0 weight and
bias tensors and reuse the shared projector fields in the loader and graph.
Read vision parameters from the checkpoint configuration and remove the
U1-specific tensor names and model fields while preserving its strided
projection, bias, and row-major token layout.
* mtmd : fix SenseNova U1 image positions and attention
* fix spacing
* mtmd : restore generic test input chunks
* mtmd : restore SenseNova U1 integration after upstream merge
* Update conversion/sensenova_u1.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Update src/models/sensenova-u1.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* model : remove SenseNova U1 BF16 activation casts
Use standard backend precision handling for QKV, FFN, and output
projections instead of casting inputs based on the output tensor type.
Assisted-by: Codex
* Update src/models/sensenova-u1.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* llama: create new fused op for QSA scoring
* enroll glm5.3-next too
* tests : initialize the remaining tensors in test_kpool_mask
The harness sentinels were left uninitialized, so leftover memory could read as NaN and fail the NaN check.
* ggml: generalize the k-pool mask op into ggml_top_k_mask
The k-pool mask op only covered the pooled QSA selection of qwen4exp
and glm5-next, while the DSA models and the DeepSeek V4 CSA build the
same top-k mask by hand with fill, set_rows and add.
ggml_top_k_mask(mask, top_k, score, blocks, extra) keeps the mask
cells of the live picks and the extra cells, and sets the rest to
-inf. Every row tensor is 4D in the layout of top_k, and the mask is
broadcast to those rows, so a selection can also be made per head.
score, blocks and extra are optional: without score every pick is
live, without blocks a pick is a cell, and extra holds the cells a
row keeps without a pick.
llm_graph_context::build_fused_top_k_mask views the row tensors in
the layout of kq_mask and registers the fused node. The k-pool
models, the DSA attention of build_attn and the DeepSeek V4 CSA mask
use it when the probe enables the op, and keep their current path on
backends without it.
---------
Co-authored-by: Pascal <admin@serveurperso.com>
* llama : scale moe down projection per expert token
Replace the global amax rescaling for the MoE down projection with a per-expert-token factor computed by a POOL_1D max reduction along the intermediate dimension. Undo the scale on the output with the same exact power-of-two factor. Handle POOL_1D in the Meta split-state logic so a split along axis 0 becomes a replicated per-device partial result, enabling tensor parallelism.
Assisted-by: pi:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)
* cont : simplify comments
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : refactor amax src1 scaling result
Return the rescaled activation and the optional scale tensor through a small struct instead of an output pointer.
Assisted-by: pi:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)
* cont : add TODO + clean-up
* llama : move LFM2 shortconv and FFN blocks into a shared LFM base
* Update src/models/lfm-base.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* model : add H2O-Lightning-4B decision model
Adds the h2o-lightning decision type for h2oai/h2o-lightning-4b
(Qwen3.5-4B with a merged decision LoRA).
None of the existing types match this model. It answers with " A"
rather than "A", uses up to 255 labels, and was trained on Python
json.dumps text, so floats like 12.0 must render as 12.0.
The converter writes the decision metadata and the systemone template
from the model's serve_config.json and refuses a config that doesn't
match. Tests skip unless the tiny test model is available locally.
Assisted-by: Claude Opus 5.5
* rm redundant code
* put labels to gguf
---------
Co-authored-by: Jonathan C. McKinney <jon.mckinney@h2o.ai>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* model : drop the redundant precision overrides in the Kolibri-1 router
Follow-up to the review in #30673: the router matmul works with the
default precision, so both ggml_prec_set_* calls can go away.
Assisted-by: pi:llama.cpp/ggml-org/Qwen3.8-Flash-Next-GGUF (MacBook, M5 Max, 128GB)
* cont : fix comment
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* model : add Kolibri-1 (Aleph-Alpha) support
Kolibri-1 is a bilingual German/English MoE reasoning model with 78B
parameters and 3.46B active per token: 50 layers, 384 routed experts plus
one shared expert.
Attention interleaves four sliding layers at a 513 window with one full
layer. The sliding layers rotate with NeoX RoPE, the full layers are NoPE.
GQA runs 48 query heads against 4 KV heads at head_dim 128 with a
QK-RMSNorm, and every block carries four RMSNorms, so the attention and the
MLP output are normalized before they rejoin the residual stream. Routing
picks the experts from logits + expert_bias and weights them with the
sigmoid of the unbiased logits, without renorm.
The converter reuses AfmoeModel for the per-expert checkpoint layout and
the expert selection bias, and writes the shared expert and the per layer
sliding window pattern. The released FP8 checkpoint needs no special code,
it goes through the generic weight_scale_inv path. Three tensor names are
added to the mapping (post_attn_norm, post_ffn_norm,
moe.router.expert_bias) and the 128k vocabulary registers the qwen2
pre-tokenizer by hash. The arch is registered in test-llama-archs, with GQA
and the sliding window pattern.
Assisted-by: DeepSeek Harness (DeepSeek V4.1 Flash)
* Update src/models/kolibri1.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Add fusion records for kolibri1
* Update src/models/kolibri1.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
The scheduler uploads each view of a host input on its own, so the per-layer reshape of new_pool_idxs (and of sel_mask in glm5-next) was copied once per layer.
* model: add Muse Spark Open (text, vision and audio)
Muse Spark Open is a latent space MoE: the residual stream stays at 8192
while the 256 experts live in 4096, with a projection down and up around
them, top-8 sigmoid routing with a correction bias. Attention interleaves
three sliding layers at a 2048 window with rope for one global layer
without rope, with a sigmoid output gate and a weightless QK-norm.
Residuals are norm preserving, their gate is folded at conversion time.
Vision reuses the Muse Glimmer tower. Audio is a causal convolutional
pyramid with no attention and no spectrogram, four stages folding 16, 6,
5 and 4 samples into the channel dim, one embedding per 80 ms of 24 kHz.
The quantizer keeps the latent projections and the attention gate at
Q8_0, a little over 1% of the weights, since the whole residual stream
passes through them and the low-bit types turned the output into a
repeating loop. The model saver writes the post norm epsilon, and the
arch is registered in test-llama-archs.
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Muse Spark: drop the redundant forward expand
* Muse Spark: drop the redundant forward expand and tighten the comments
* Update src/llama-quant.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Update src/llama-graph.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Update conversion/muse_spark.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Update conversion/muse_spark.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Update conversion/muse_spark.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
* Update conversion/muse_spark.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>