* studio: reserve MTP draft VRAM in GGUF auto-fit
Auto-fit advertised a context (for example ~110k for the Qwen3.6-27B MTP
GGUF) that fit on paper but OOMed mid-generation or during tool calls once
MTP speculative decoding was active. The MTP draft path's VRAM was reserved
as a flat 5% of total VRAM, which tracks neither of the two real costs: the
MTP head keeps its own attention KV cache that grows with context, and the
speculative verification buffer grows with --spec-draft-n-max. On the
hybrid Mamba/attention Qwen3.6 models the main KV is small, so auto-fit
happily kept a near-native context while the draft path pushed the load
over budget at runtime.
Replace the flat fraction with a byte-accurate, context- and n_max-aware
reserve sized from GGUF dims: draft KV from nextn_predict_layers and the
attention dims at f16 (llama.cpp's MTP draft context uses f16 KV regardless
of the main cache type), plus a verify buffer per embedding-unit per draft
token. The reserve is evaluated per candidate context inside the fit binary
search and added to every pin/fit check, including the tensor-parallel
planner and its even-split decision. Coefficients were calibrated against
llama-server VRAM measurements on the Qwen3.6-27B MTP GGUF (RMS 14 MiB).
The flat fraction remains as a fallback when GGUF dims are unavailable, so
non-MTP loads are unchanged. The budget now also engages when the user wires
MTP through extra args (--spec-type draft-mtp, including chains), reads the
effective draft depth from --spec-draft-n-max or the legacy --draft-max with
extras taking precedence over the first-class field, reserves a separate
drafter's weights when supplied via --model-draft/--spec-draft-model/-md,
and mirrors _build_speculative_flags so it never reserves for MTP the launch
resolver will not emit (needs a head/drafter and a binary that supports
--spec-type mtp).
Adds tests/test_mtp_vram_budget.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: total-based VRAM budget + deterministic compute-graph buffer
Build on the byte-accurate MTP reserve with three changes that make the
GGUF auto-fit budget deterministic across architectures and recover usable
context, especially for MTP models on a single tight card.
1. Total-based budget. Cap GPU occupancy at a fraction of TOTAL VRAM rather
than a fraction of FREE VRAM, and raise the fraction from 0.90 to 0.95:
budget = free - (1 - 0.95) * total (per GPU, summed for a pool)
The reserve is now absolute (a fixed slice of the card) instead of
shrinking as the GPU fills, so a partly-used GPU keeps a constant cushion
for compute/CUDA/verify buffers instead of over-promising context and
spilling to CPU at runtime. _get_gpu_memory() reads memory.total alongside
memory.free; _fit_context_to_vram, _select_gpus and the load_model pool
loops thread the totals through. Multi-GPU layer-split pools
sum(free_i - 0.05*total_i); tensor mode reserves per device.
2. Deterministic compute-graph buffer. Replace the flat 5 GB/device tensor
reserve (a magic constant that over-reserved about 8x on a 27B model) with
_estimate_compute_buffer_bytes, sized from GGUF dims and the launch flags:
out = n_vocab * n_ubatch * 4 # vocab-width output buffer
act = 4 * n_embd * n_ubatch * 4 # activation scratch
pipeline_per_device = act + out * (n_parallel - 1)
tensor_per_device = 2*act + out * n_parallel
The buffer is context-independent and scales with --parallel (serving
slots), not with how the model is split across GPUs. It is now reserved in
BOTH multi-GPU paths (layer split folds one buffer into the pooled
footprint; tensor mode reserves it per device). The flat 5 GB stays only as
a fallback when vocab/embedding dims are unavailable. Calibrated against
llama-server measurements (parallel 1/2/4/8 give 36/492/1388/3220 MiB on a
single GPU; about 600 MiB/device tensor); the estimate is a small upper
bound.
3. GGUF parsing. Read vocab size (tokenizer tokens array length) and
feed_forward_length for the compute-buffer estimate.
Effect on the Qwen3.6-27B MTP Q6_K case (MTP on): a single 32 GB card at
about 31 GB free advertises f16 23k to 64k, q8_0 44k to 115k, q4_0 82k to
200k; 2x 24 GB tensor mode recovers the full 262k window for f16 (was about
134k). Validated on hardware: 1x 32 GB f16 at 64768 loads at 29.3 GB / 120
t/s; 2x 23 GB tensor f16 at 262144 loads at 22.2 GB/device / 98 t/s; both
within 0.4% of the estimate. Adds test_compute_buffer.py and updates the
KV/context-fit/MTP-budget tests for the 0.95 constant and the new budget.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten comments in the VRAM auto-fit changes
Condense the docstrings and inline comments added by this PR (internal backend
helpers): drop restated-signature docstrings, fold multi-line block comments to
one or two lines, and remove notes that just repeat the code. No behavior change
(AST-verified comment/docstring-only via comment_tools.py); the backend test
suite is unchanged and green.
* studio: address review findings in the VRAM auto-fit budget
Five fixes from a parallel-reviewer pass on this PR; all confirmed against the
real functions and covered by new tests.
- Tensor mode now honors the total-based VRAM cap. _plan_tensor_parallel took
total_by_idx and budgets each GPU at free - (1-frac)*total, mirroring the
layer-split paths; previously it fit against raw free and could spend the 5%
safety cushion on a partly-used multi-GPU box (reproduced ~3.3 GB over).
- Draft K and V cache types are parsed and accounted independently. A one-sided
override (e.g. --cache-type-k-draft q4_0, V left f16) no longer applies the
small quant to both axes and under-reserves the f16 axis. The embedded-head
formula sizes per axis; the separate-drafter path uses the heavier type so it
never under-reserves.
- The compute-graph buffer honors a user --ubatch / --ubatch-size / -ub override
(parsed and threaded into every _estimate_compute_buffer_bytes call and the
tensor planner); it previously always assumed the 512 default, under-reserving
up to ~8x at --ubatch 4096.
- GPU ranking uses the usable budget (free - (1-frac)*total) instead of raw free
in _select_gpus and both auto-context subset loops, so a more-used large card
no longer outranks a less-used small card that has more usable room.
Adds regression tests for each (tensor total cap, ubatch reserve scaling, split
K/V no-under-reserve, --ubatch parser, usable-ranking GPU selection). Full
targeted backend suite green (321 passed).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: gate tensor-parallel admission on usable VRAM budget
The tensor-parallel GPU admission filters still used raw free VRAM after
the total-based budget landed, an asymmetric fix: a partly-used large card
can clear the per-device compute-buffer reserve on raw free while its usable
budget (free - (1-frac)*total) does not, so the planner admitted it and the
even split could emit a near-zero weight slice for a GPU that should have
been excluded.
- _plan_tensor_parallel: admit GPUs by usable budget, not raw free (move the
_usable helper above the filter).
- load_model: admit the tensor set by _gpu_usable, and downgrade to layer
split when the pooled usable budget cannot hold weights plus per-device
compute buffers (the planner can only floor the context, not stop an
overcommitted launch).
Adds regression tests: planner drops a GPU whose usable budget is below the
reserve, and a source-level check that load_model admits on the usable
budget and carries the pooled-weight downgrade.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: size the MTP reserve for the user's overriding drafter
A user --model-draft passed in extra_args is appended last and wins at the
llama-server launch, but the VRAM budget preferred Studio's auto-detected
drafter (mtp_draft_path or extras), so a larger custom drafter was
under-reserved. Flip the precedence to extras-first, matching the draft-depth
(n_max) resolution two lines above. Adds a source-level regression test.
* studio: account for MTP reserve in tensor gate, restore 2-col GPU probe
Two issues found by re-review of the prior fix:
- The tensor-parallel capacity gate only checked the model weights against the
pooled budget, not the MTP reserve. A separate-drafter MTP load whose weights
fit but weights + drafter do not could still launch overcommitted in tensor
mode. Add the non-shrinkable MTP reserve (drafter weights + floor draft KV, or
the flat 2 GiB fallback when dims are unavailable) to the gate.
- The nvidia-smi probe was switched to a three-column query (index,free,total)
for the total-based budget but required exactly three columns, so a driver or
mock returning the legacy two-column "index,free" was dropped and the probe
fell through to the real GPUs. Accept two columns (total 0) and treat an
unknown total as the legacy free*fraction in _select_gpus.
Tests: tensor gate asserts the MTP term is included; _get_gpu_memory parses both
two- and three-column output; the existing two-column GPU-detection mocks pass
again.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: keep the VRAM cushion in tensor planning when GPU totals are unknown
_plan_tensor_parallel fell back to raw free VRAM when a GPU's total was
unavailable (a two-column nvidia-smi probe reporting total 0), while
_select_gpus and the load_model ranking both fall back to free*fraction. That
let tensor planning spend the 5% cushion the rest of the fit preserves and
over-advertise context in exactly that path. Align the fallback to
free*_CTX_FIT_VRAM_FRACTION. Updates the no-totals planner test expectations
(now free*frac) and adds a regression test that total 0 keeps the cushion.
* studio: honor LLAMA_ARG_* env overrides and HF draft flags in the VRAM budget
The budget parsed llama-server flags only from the request's extra_args, but the
child process inherits Studio's full environment (child_env_without_native_path_secret
copies os.environ), and llama-server honors LLAMA_ARG_* env vars for the same
options. So a service-level override the child acts on was invisible to the fit,
which could then advertise a context/GPU set that OOMs at load.
- _extra_args_n_ubatch: fall back to LLAMA_ARG_UBATCH (drives the compute buffer;
an unseen 4096 vs the 512 default under-reserves ~8x).
- _extra_args_mtp_draft_path: also recognize the HF draft-repo flags
(--spec-draft-hf/-hfd/-hfrd/--hf-repo-draft) and fall back to
LLAMA_ARG_SPEC_DRAFT_MODEL / LLAMA_ARG_SPEC_DRAFT_HF_REPO. An HF repo isn't a
local file so it can't be sized, but recognizing it routes to the flat reserve
instead of mis-sizing Studio's auto/embedded drafter.
- _extra_args_draft_cache_types: fall back to
LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_K/_V per axis.
CLI extra_args win over env (they are appended last at launch). Each parser takes
an injectable env for deterministic tests. Adds env-fallback and HF-flag tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: review polish - drop non-flag --ubatch, harden GPU probe, document buffer
Non-blocking items from a second review pass; no behavior change in the common path:
- _extra_args_n_ubatch: drop --ubatch; the binary only accepts --ubatch-size/-ub,
so parsing --ubatch implied support it does not have (it would over-reserve for a
launch that fails on the unknown flag).
- _get_gpu_memory: skip a malformed nvidia-smi line instead of letting one bad line
raise and drop the whole NVIDIA probe to the torch fallback.
- _estimate_compute_buffer_bytes: document that the per-slot output-buffer model
assumes a small n_outputs_max (chat decode); it would under-count for
embeddings / --logits-all / reranking, which Studio does not run on this path.
* studio: honor LLAMA_ARG_SPEC_TYPE when deciding the MTP reserve
_extra_args_requests_mtp only checked extra_args, but the child inherits Studio's
env and llama-server honors LLAMA_ARG_SPEC_TYPE. So a service-level
LLAMA_ARG_SPEC_TYPE=draft-mtp would run MTP while the fit skipped the draft
reserve and could advertise a context/GPU set that OOMs at load. Recognize the
env value (CLI still wins). Completes the env-override coverage alongside ubatch,
draft model, and draft cache types. Adds an env regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: reserve VRAM for non-MTP model-based draft modes too
The draft reserve only engaged for MTP. A user passing a non-MTP model-based
draft mode (--spec-type draft-simple / draft-eagle3) with a --model-draft loads
a separate draft model whose weights + KV consume GPU memory, but the fit
reserved nothing and could OOM at load. Engage the existing drafter reserve for
those modes when extras (or LLAMA_ARG_SPEC_TYPE) name a drafter; ngram-* load no
model and are unaffected. Purely additive (reserves where there was none).
Adds parser + gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: floor quantized embedded MTP draft KV at f16; fix two test issues
Address PR review feedback (three findings):
1. Quantized embedded MTP draft KV was underpriced. The embedded head is a
single draft layer, so llama.cpp cannot amortize quantized-KV overhead over
many layers the way the main model does: a quantized draft KV (e.g.
--spec-draft-type-k q4_0) actually fits LESS context than f16, not more
(ggml-org/llama.cpp#24102, where a collaborator recommends f16 for the draft
KV). Pricing q4_0 at 0.5625 of an element (~28% of f16) under-reserved, so a
quantized override could advertise a context that shrinks or OOMs at load.
Floor the embedded draft KV bytes-per-element at f16 (quantized types priced
as f16, f32 still its full 4 bytes). The separate multi-layer drafter, where
quantization does amortize, keeps the user's real type.
2. test_load_model_reserves_for_non_mtp_draft_modes asserted an exact one-line
source substring that pre-commit black wrapped across lines, breaking CI.
Strip whitespace before matching so the check survives any line-wrapping.
3. test_compute_buffer.py installed a partial httpx stub via setdefault that, if
collected before test_kv_cache_estimation.py, leaked into sys.modules without
HTTPError/Response and could break the transformers introspection tier by
collection order. Adopt the sister file's pattern: only stub when real httpx
is absent, and include the full symbol set.
Updates the affected draft-KV tests to assert the f16 floor.
* studio: guard httpx stub in test_mtp_vram_budget too
test_mtp_vram_budget.py installed a partial httpx stub via setdefault that, like
test_compute_buffer.py before it, lacked HTTPError/Response and could leak into
sys.modules ahead of tests that need huggingface_hub/transformers, breaking the
introspection tier by collection order. Apply the same guard used by
test_kv_cache_estimation.py: only stub when real httpx is absent, with the full
symbol set.
* studio: per-device layer-split reserve, effective spec-type, drafter weights, KV restore
Address PR review feedback (four findings in the auto-fit budget):
A. Reserve the per-device layer-split overhead. A layer (pipeline) split allocates
a fixed per-device overhead (CUDA context + per-device compute scratch) on every
participating GPU, beyond the slot-scaling compute buffer that is conserved across
the split. Measured ~0.9 GB/device on the Qwen3.6-27B GGUF (b9625), independent of
--parallel: layer-split TOTAL VRAM grew +894 MiB (parallel=8) / +946 MiB
(parallel=1) per extra GPU, ~linear to +2.6 GB at 4 GPUs. The fit folded a single
compute buffer for all subset sizes, so a k-GPU layer split was short by
~(k-1)*0.9 GB and could pin a context that fits the pool on paper but OOMs a device.
Reserve (k-1) * _PIPELINE_PER_DEVICE_OVERHEAD_MIB per subset in the layer-split fit;
k=1 adds nothing, so single-GPU sizing (and the validated benchmark rows) is unchanged.
B. Track the effective --spec-type. _extra_args_requests_mtp returned true on the
first MTP-ish --spec-type and consulted LLAMA_ARG_SPEC_TYPE even when a CLI
--spec-type was present, contrary to llama.cpp (last CLI value wins; a CLI flag
overrides the env). So `--spec-type draft-mtp --spec-type ngram-mod` or a non-MTP
CLI value with a stale MTP env over-reserved a drafter the launch won't load
(shrinking context / selecting extra GPUs). Route both detectors through a new
_effective_spec_type helper.
C. Keep known drafter weights in the fallback reserve. When a separate drafter's KV
metadata can't be sized, _estimate_mtp_overhead_bytes returned None and discarded
the drafter's known weight bytes, falling back to the flat 5% reserve; a drafter
larger than that cushion could launch over budget and OOM. Reserve the known
weights even when KV sizing fails (None only when nothing is known).
D. Restore quantized KV on tensor->layer-split downgrade. The tensor attempt drops a
quantized KV cache (tensor mode aborts on it). When the GPU-count or capacity gate
then downgrades to layer split -- which supports quantized KV -- the dropped type
was lost and the launch used f16, using more VRAM and shrinking context. Remember
the dropped type and restore it on downgrade (the launch re-emits it from the var).
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: per-device overhead in GPU pin, skip CPU draft, gate env spec-type
Address PR review feedback (three follow-up findings):
F1. Reserve the per-device layer-split overhead in the pin path too. The earlier
per-device reserve was added to the auto-context fit loops but not to
_select_gpus, which the explicit-ctx and file-size-only paths use to PIN GPUs
with -ngl -1 (no --fit fallback). A 2+ GPU pin within ~1 GiB/extra-GPU of the
budget could OOM a device at load. Add a per_device_overhead_bytes arg to
_select_gpus so a k-GPU pin must hold model + (k-1)*overhead; pass the pipeline
overhead at both pin call sites. Single-GPU pins are unchanged.
F2. Don't charge a CPU-offloaded drafter against the GPU budget. A user passing
--spec-draft-ngl 0 or --spec-draft-device none/cpu keeps the separate draft
model's weights + KV on CPU, but the budget still charged the full drafter GGUF
size, auto-reducing context or downgrading GPU selection. Detect the CPU-offload
flags and drop the separate drafter (and its flat fallback) from the budget; an
embedded head follows the main -ngl and is unaffected.
F3. Consult LLAMA_ARG_SPEC_TYPE only when it can reach the child. llama-server's CLI
args override env, and _build_speculative_flags emits a --spec-type/--spec-default
for every UI mode except "off". So a stale MTP env on a non-MTP model (auto mode)
made the fit reserve MTP that the emitted --spec-default disables, shrinking
context / picking extra GPUs. Gate the env consult on "no user --spec-type and UI
mode off"; the MTP-model auto path still engages via Studio's own detection.
Adds regression tests for each.
* studio: drafter budget precedence and --spec-default in effective spec-type
Two spec-precedence fixes surfaced by an independent multi-reviewer pass:
R3. Size the drafter the launch actually loads. _mtp_draft_for_budget consulted
LLAMA_ARG_SPEC_DRAFT_MODEL (via _extra_args_mtp_draft_path's env fallback)
before Studio's resolved mtp_draft_path, but _build_speculative_flags emits
--model-draft mtp_draft_path, which overrides the env at launch. With a stale
(smaller) env drafter, the budget under-reserved and could OOM. Order the
budget by what actually launches: CLI extras --model-draft (appended last,
wins), then Studio's emitted mtp_draft_path (when MTP engages and the user
doesn't own --spec-type), then the env drafter.
R4. Treat --spec-default as a CLI spec override in _effective_spec_type. It only
recognized --spec-type, so extras=["--spec-default"] with LLAMA_ARG_SPEC_TYPE=
draft-mtp fell through to the env and over-reserved MTP, even though the CLI
--spec-default overrides the env to a non-MTP default. Recognize it as a CLI
spec flag (resolves to "default", non-MTP) that suppresses the env fallback.
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: refine MTP draft reserve (parallel slots, last-wins, KV cushion, ranking)
Address PR review feedback (five follow-up findings, all edges of this session's
earlier MTP/auto-fit changes):
G1. Price the separate drafter's KV per --parallel slot. _mtp_draft_kv_bytes called
the drafter's _estimate_kv_cache_bytes with the default n_parallel=1, but the
drafter is served under the main model's slot count; a sliding-window drafter
(Gemma) grows KV per slot and was under-reserved. Thread n_parallel through the
draft KV / overhead estimate and the fit closure.
G2. Honor last-wins for the draft-offload flags. _extra_args_draft_offloaded_to_cpu
returned True on the first CPU value, so --spec-draft-ngl 0 --spec-draft-ngl -1
(final = GPU) wrongly dropped the drafter reserve while the server kept it on
GPU -> OOM. Decide on the final value of each flag only.
G3. Keep the flat cushion when only the drafter weights could be sized. The weights
fallback installs mtp_overhead_fn, which made callers drop the flat MTP reserve,
leaving the still-unsized draft KV with no cushion. Keep the flat fraction on in
that weights-only case, on top of the byte-accurate weights.
G4. Rank auto/cap GPU subsets by the active budget fraction. The ranking used a
hard-coded 0.95 while the fit tests _pin_fraction (lowered by the flat MTP
reserve); on mixed-total GPUs that could order subsets differently and pick a
worse plan. Rank with the same fraction the fit uses.
G5. Keep the embedded-head flat reserve under a draft CPU-offload flag. F2's
not-_draft_on_cpu guard also dropped the reserve for an embedded MTP head, which
is part of the main model and stays on GPU regardless of --spec-draft-ngl. Only
suppress the flat reserve for a CPU-offloaded separate drafter (no embedded head).
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: keep GPU on non-integer total, keep tensor flat reserve for weights-only
Two review findings:
- _get_gpu_memory dropped a whole GPU when nvidia-smi reported a non-integer
memory.total ("N/A" on some drivers / MIG / vGPU): index, free and total were
parsed in one try/except that skipped the line on any ValueError, so the GPU
vanished from the probe and the load could silently spill to CPU. Parse index
and free (required) first, then total separately, defaulting to 0 (the fit then
uses the free*frac path for that GPU). Adds N/A and bad-free test cases.
- Tensor planning skipped the flat MTP reserve for a weights-only drafter (file
size known, KV unsizable): the capacity gate used the byte floor whenever
mtp_overhead_fn was set, so it reserved only the drafter weights and no draft
KV. Tensor mode has no --fit valve, so that could overcommit and OOM. Keep the
flat reserve (never below the byte floor) in the weights-only case too, mirroring
the layer-split _mtp_kv_unsized handling. Adds a regression test.
(A third suggestion -- fold --batch-size into the compute-buffer reserve -- was
checked on hardware and declined: -b 8192 -ub 512 used identical VRAM to the
default at -c 64000, so the logical batch does not size the graph buffer; the
estimate correctly uses the physical micro-batch.)
* studio: budget the main KV from LLAMA_ARG_CACHE_TYPE env when Studio emits none
The child inherits LLAMA_ARG_CACHE_TYPE_K / LLAMA_ARG_CACHE_TYPE_V, but Studio
emits --cache-type-k/-v only when the param or extras set the type. When neither
does, a heavier env type (f32) reaches the child while the auto-fit budget
assumed the f16 default, under-reserving the main KV and risking OOM at the
advertised context. This is the one main-KV axis that lacked the env-aware
handling the other axes already have (spec-type, draft model, draft cache type,
ubatch).
load_model now adopts the heavier of the two env types when it exceeds f16 (only
f32 does), and the launch re-emits it so child and budget stay byte-consistent.
Quantized env types are <= f16 and remain safely over-reserved by the default,
so they are left untouched (no change). A single value is used because the
budget's KV estimate has one cache_type_kv knob, matching parse_cache_override's
existing key/value collapse.
Adds _env_main_cache_type_for_budget plus regression tests covering f32 adoption,
the K/V heavier-of collapse, quantized/unknown no-ops, and the load_model source
precedence.
* studio: budget tensor parallel when LLAMA_ARG_SPLIT_MODE env selects it
Studio emits --split-mode tensor only on its tensor branch; the default
layer-split path emits nothing and resolve_tensor_parallel consults only extras.
The child inherits LLAMA_ARG_SPLIT_MODE, so a tensor env on a layer-split plan
silently runs the child tensor-parallel (heavier per-device compute buffer)
while the budget reserved only the layer-split per-device overhead, under-
reserving on multi-GPU.
load_model now flips the plan to tensor when extras do not set a split mode and
the env selects tensor, so Studio plans, reserves, and emits tensor consistently.
The flip is one-directional (guarded on not tensor_parallel and no extras
split-mode) so an existing tensor plan is never downgraded and extras keep
precedence. Other env modes (layer/row/none) are not a runtime-heavier surprise
and are left untouched.
Adds _env_split_mode_is_tensor plus unit and load_model source-level tests.
* studio: reconcile inherited llama.cpp env with the budgeted launch decision
Addresses a review pass over the VRAM auto-fit work. The budget now sizes the
right amount, but the child process inherits LLAMA_ARG_* env (see
child_env_without_native_path_secret), and a few axes could still run the child
in a mode Studio neither chose nor budgeted.
Mixed known/unknown GPU totals over-advertised the pooled layer-split budget.
_pool_budget_mib pooled free and total separately, so an unknown-total GPU
(MIG/vGPU/N/A) contributed its full free with no cushion when mixed with
known-total GPUs (~(1-frac)*free over-advertise, about 500 MiB in a two-GPU
case). It now sums each GPU's own usable budget, and the layer-split fit calls
take that as an absolute budget (budget_frac=1.0, total_mib=None) so the fit and
the footprint check agree. All-known-total pools are unchanged.
LLAMA_ARG_SPLIT_MODE=tensor survived a tensor-to-layer downgrade. The downgrade
only stripped CLI extras, so the inherited env still ran the child tensor while
Studio budgeted layer split. When the final decision is layer split, a non-layer
inherited split mode (and any paired LLAMA_ARG_TENSOR_SPLIT) is now cleared from
the child env.
Inherited quantized LLAMA_ARG_CACHE_TYPE_K/_V crashed tensor mode. Tensor mode
aborts on a quantized KV cache; Studio drops a quantized cache_type_kv for the
tensor attempt but the inherited env reached the child anyway. When the final
decision is tensor split, a quantized cache-type env is now cleared so the child
uses the tensor-safe default that was budgeted.
Env-derived cache budget no longer mutates the emitted launch flags. An env-only
main KV type now informs the budget only; it is not re-emitted, so an asymmetric
K=f32,V=f16 env reaches the child as set instead of being rewritten to symmetric
--cache-type-k/-v f32.
Adds source-level regression tests for all four and confirms the documented
single-GPU/tensor/pipeline numbers are byte-identical before and after.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten comments in the VRAM auto-fit code
Compress the verbose docstrings and inline comments added by this work to
succinct 2-4 line versions, drop restated/obvious ones, and cut duplicated
rationale across the two tensor-downgrade branches. Keeps the non-obvious intent
(env-inheritance precedence, the #24102 embedded-draft floor, the per-device
overhead and pool-budget rationale) while removing roughly 120 lines of comment
text from llama_cpp.py. Also trims the few longest test-comment blocks; concise
per-test scenario notes are left intact.
No logic change: verified with comment_tools.py check --strip-docstrings (code-
only signature unchanged vs the prior commit) and the full backend suite still
passes (824).
* studio: lock in env-drafter engagement for the separate-draft reserve
A review suggested an env-provided LLAMA_ARG_SPEC_DRAFT_MODEL would skip the
draft reserve and OOM. It does not: the gate's _extra_args_mtp_draft_path(extra_args)
call defaults env=None, which consults os.environ, so an env-only drafter still
sets _user_draft_via_extras and is sized via _env_draft_for_budget. Add a source
guard that the gate keeps the env-inclusive form (not extras-only env={}) and a
behavioral test mirroring the reviewed scenario, so a future cleanup can't
regress it. No production change.
* studio: carry the unsized MTP reserve and env split/offload into tensor planning
Addresses a review pass over the multi-GPU and env-inheritance paths.
Tensor planner dropped the unsized draft-KV cushion. When a separate drafter has
known weights but unreadable KV metadata, _plan_tensor_parallel receives a
non-None weights-only mtp_overhead_fn and applied the flat 2 GiB reserve only for
the no-fn case, so its binary search spent the unsized-KV cushion on context and
over-advertised. Add mtp_flat_reserve_bytes (subtracted from the pooled budget and
the even-split check), and pass it from load_model whenever _mtp_kv_unsized. The
layer path and the tensor pre-gate already kept this cushion.
Stale LLAMA_ARG_TENSOR_SPLIT survived in tensor mode. When the planner picks an
even split it emits no --tensor-split, so an inherited tensor-split env reached the
child and overrode the budgeted split. The layer downgrade branch cleared it; the
tensor branch now does too.
Env-only draft CPU offload was ignored. _extra_args_draft_offloaded_to_cpu checked
extras but not LLAMA_ARG_N_GPU_LAYERS_DRAFT, so an env-offloaded drafter was still
charged GPU budget and under-advertised context. It now consults that env (the
device flag has no env), called with env=os.environ.
Layer-split compute buffer had no fallback when GGUF dims are missing. The estimate
returns 0 then, so the layer path folded no buffer while the tensor path falls back
to the flat reserve. Use the flat reserve for the layer path too (a safe upper
bound, since the tensor buffer >= the layer one).
All four are gated on conditions the documented benchmarks don't hit; the
single-GPU/tensor/pipeline reconfirm numbers are byte-identical, and the full
backend suite passes (830) with regression tests for each fix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: share the env-aware tensor decision across load and dedup matchers
A review pass found the inherited-LLAMA_ARG_SPLIT_MODE=tensor flip lived only
in load_model, so the two duplicate-load matchers disagreed with it.
Consolidate the decision into _effective_tensor_parallel (extras + toggle, then
flip on when extras set no split mode and the child inherits a tensor split
env). load_model, the backend matcher (_already_in_target_state) and the route
matcher (_request_matches_loaded_settings) now all call it. Before, an env-driven
tensor server compared against resolve_tensor_parallel (env-blind) in both
matchers, so a follow-up load that should dedup was seen as a mismatch and the
healthy server was needlessly killed and reloaded.
Also finish the tensor cache-type handling: when the tensor attempt drops a
quantized KV it now re-adopts a heavier inherited env cache type (f32) for the
budget, mirroring the initial adoption; and the two layer-split downgrades clear
_cache_type_from_env so the restored quantized type is actually re-emitted rather
than left to a stale inherited env.
All gated on inherited env the documented benchmarks don't set; the single-GPU,
tensor and pipeline reconfirm numbers are byte-identical, and the full backend
suite passes (832) with unit + source regression tests for the shared helper and
the route matcher.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: complete the env-aware tensor/spec handling across all paths
A second review pass found the env-aware tensor/MTP handling was applied
asymmetrically: some paths inherited LLAMA_ARG_* env, others didn't. Three real
follow-ups, plus a small consolidation so the env semantics live in one place.
1. Tensor fallback ignored the inherited tensor env. load_with_tensor_fallback
computed its retry gate with the env-blind resolve_tensor_parallel, so an
env-only tensor load (toggle off, no --split-mode extra) that crashed on a
tensor-incompatible GGUF re-raised instead of retrying layer split. It now
uses the env-aware decision; and since the inherited env would otherwise
re-engage tensor on the retry (CLI args persist, the env does too), the retry
forces --split-mode layer (CLI wins over env) so it can't re-crash.
2. Duplicate-load matchers looped reloads after a tensor->layer downgrade. Both
matchers compared the env-expanded tensor decision against the loaded server,
but load_model may downgrade tensor to layer (capacity/buffer) and scrub the
child env. The still-set parent env then made every identical request look
like a mismatch, killing and reloading a healthy layer server. Add
_tensor_parallel_matches_loaded, which only lets an inherited tensor env raise
a match against a server that actually launched tensor; a downgraded server
matches the same request (an identical load would downgrade the same way).
3. MTP binary-capability fallback leaked an inherited LLAMA_ARG_SPEC_TYPE. When
the binary lacks MTP, _emit_mtp degraded but emitted no spec flag, so an
inherited LLAMA_ARG_SPEC_TYPE=draft-mtp still reached the child and attempted
MTP the gate had budgeted off. It now emits --spec-default (CLI wins over env)
like the sibling no-head / non-MTP fallbacks.
Consolidation: moved _env_split_mode_is_tensor / _effective_tensor_parallel into
llama_server_args.py (with the new _tensor_parallel_matches_loaded) so the
lightweight tensor_fallback module can share them without importing llama_cpp;
llama_cpp re-exports them for back-compat.
All gated on inherited env the documented benchmarks don't set; the single-GPU,
tensor and pipeline reconfirm numbers are byte-identical, and the full backend
suite passes (883) with regression tests for each fix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: budget the heavier axis of asymmetric --cache-type-k/-v extras
A review pass found the explicit-extras counterpart of the env cache-type fix.
load_model adopts the heavier inherited LLAMA_ARG_CACHE_TYPE_K/_V env for the
reserve, but the explicit-extras path used resolve_cache_type_kv, which collapses
both axes to one last-wins value. So extras such as
--cache-type-k f32 --cache-type-v f16 (lighter axis last) budgeted f16 for both
axes while the child allocates f32 on K, over-advertising context and
re-opening the OOM path this PR closes.
Add parse_cache_override_per_axis (keeps the K/V last-wins values apart) and
_extra_args_main_cache_type_for_budget (the heavier of the two by bytes/elem),
and budget from it. The user's extras are appended last and win per axis at the
child, so this only raises the reserve; the emitted command and the asymmetric
child cache are unchanged, and the common single-axis / symmetric cases resolve
to the same type as before.
Reconfirm numbers (single-GPU table, tensor, pipeline) are byte-identical, and
the full backend suite passes (892) with per-axis parser and heavier-axis budget
regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fix tensor-safety masking and strip inherited HF drafter selectors
A review pass found two more env/extras edge cases on the speculative and tensor
cache paths.
Tensor-safety could miss a quantized axis. The previous change budgets the
heavier-by-bytes cache type, but that masks a quantized axis paired with a
heavier one: --cache-type-k f16 --cache-type-v q4_0 resolves to f16, so the
tensor-safety block did not fire and the q4_0 axis survived into tensor mode,
which aborts on quantized KV. Test each explicit --cache-type-k/-v axis (not just
the budget type) so any quantized axis drops the cache for the tensor attempt.
Inherited HF drafter selectors were not stripped. _extra_args_mtp_draft_path
treats --spec-draft-hf / -hfd / -hfrd / --hf-repo-draft as drafter selectors, but
_SPEC_FLAGS only stripped the local --model-draft selectors, so on an inherited-
extras Apply a stale HF drafter survived and last-wins-overrode Studio's
re-derived spec choice. Add the HF aliases to _SPEC_FLAGS. The per-drafter tuning
knobs (--spec-draft-type-*, -ngld, --spec-draft-device) are intentionally left in
place: the VRAM budget reads them via the same parsers the child honors, so they
stay consistent on inherit, and stripping them would silently move a CPU-offloaded
drafter back onto the GPU.
A third flagged item -- that the HF draft env var should be LLAMA_ARG_HFD_REPO --
was a false positive from a stale manpage; the bundled binary's common/arg.cpp
sets LLAMA_ARG_SPEC_DRAFT_HF_REPO for --spec-draft-hf, which the code already
uses, so it is left unchanged.
Reconfirm numbers (single-GPU table, tensor, pipeline) are byte-identical, and
the full backend suite passes (899) with regression tests for both fixes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: preserve asymmetric cache on tensor downgrade and skip CPU-drafter reserve
A review pass found two more tensor-path edges, one a regression from the
per-axis cache change.
Tensor-to-layer downgrade collapsed asymmetric cache extras. The per-axis
tensor-safety check strips an asymmetric --cache-type-k/-v (tensor rejects
quantized KV), but the downgrade restored only the scalar heavier type, so a
layer fallback silently rewrote --cache-type-k q4_0 --cache-type-v f16 to
symmetric f16/f16 even though layer split supports the original. Save the
original extras before the tensor strip and restore them verbatim (minus the
user --split-mode) on both downgrade points; the budget still uses the heavier
scalar, the child gets the real asymmetric cache. Before the per-axis change this
case happened to survive (last-wins was f16, untouched), so this restores that.
Tensor mode reserved GPU VRAM for a CPU-offloaded drafter. The layer path drops
the flat MTP reserve when the only drafter is a separate CPU one with no embedded
head, but the tensor capacity gate and planner still charged it, under-advertising
context. Gate the tensor reserve on the same condition via _mtp_reserves_gpu.
Reconfirm numbers (single-GPU table, tensor, pipeline) are byte-identical (both
fixes are gated on conditions the benchmarks don't hit), and the full backend
suite passes (901) with regression tests for each.
* studio: drop now-unused llama_server_args imports from llama_cpp
The refactor re-pointed load_model and the matchers off resolve_tensor_parallel /
resolve_cache_type_kv and moved the env split-mode helper into llama_server_args,
leaving those three names imported but unused in llama_cpp. The repo's import-hoist
safety-net lint blocks that, so drop them; the env split-mode test now imports
_env_split_mode_is_tensor from its real home (llama_server_args).
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: Add Tensor-Parallel llama.cpp support
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Tensor-Parallel fallback and GPU selection
* Studio: reconcile split-mode extras and harden tensor-split planning
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: reconcile split-mode extras in backend duplicate-load guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: preserve inherited non-tensor split modes on reload
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor cancellation in tensor fallback, preserve tensor mode on rollback, and don't raise an explicit small context
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: reconcile split-mode in reload check and strip it on tensor downgrade
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strip --tensor-split alongside --split-mode so inherited ratios don't override the tensor planner
An inherited or stale --tensor-split in llama_extra_args was appended after
Studio's computed --tensor-split and won last in llama.cpp, re-introducing the
asymmetric-GPU OOM tensor mode is meant to prevent. Group -ts/--tensor-split
into the split-mode shadow set so it is stripped on inherit and on the layer
fallback; parse_split_mode_override still keys on the mode value only.
* Drop quantized KV for the tensor attempt and report native max context
Tensor mode aborts on a quantized KV cache, so a user with q8_0/q4_1 etc. who
enabled Tensor Parallelism silently fell back to layer split. Clear the cache
type (and strip inherited/explicit --cache-type) for the tensor attempt only;
the layer fallback re-runs with tensor off and keeps the user's choice.
Also report max_available_ctx from the native context, not an explicit small
-c, so the context slider no longer warns too early in tensor mode.
* Reconcile inherited split-mode extras in the already-loaded check
When a same-model load omitted llama_extra_args, the tensor comparison resolved
the raw (None) request and treated an inherited --split-mode tensor server as a
mismatch, forcing a needless reload. Compare using the stored extras stripped
the same way the reload strips them.
* Pass tensor_parallel through compare-mode loads
The generalized compare path loaded each GGUF without tensor_parallel, so
compare ran layer split even with the toggle on and left the settings sheet
stale. Send the toggle and hydrate the loaded state from the response, matching
the main chat and recipe load paths.
* Add --tensor-parallel flag to unsloth studio run
The headless one-liner could only reach tensor mode by passing --split-mode
tensor as a raw llama.cpp extra. Add a first-class --tensor-parallel/
--no-tensor-parallel option that sets the tensor_parallel field on the
/api/inference/load payload, forwarded through the studio-venv re-exec like the
other polarity flags. Matches the web UI toggle and the API field.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* fix(studio): inherit llama_extra_args and honor --no-mmproj
Reloading the same GGUF from the UI without gguf_variant no longer drops
CLI pass-through args like --no-mmproj. Skip mmproj download and launch
when --no-mmproj is present in llama_extra_args.
Co-authored-by: Cursor <cursoragent@cursor.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): tighten GGUF llama_extra_args variant inheritance guard
Reject inherited CLI args when the request changes gguf_variant or when
omitted variant resolves differently from the stored extra_args source.
Co-authored-by: Cursor <cursoragent@cursor.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Treat --no-mmproj-auto and --mmproj-auto with last-wins parsing for PR #5902
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: support separate-file MTP GGUF drafters (Gemma 4)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix review findings for separate-file MTP drafters
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: pair local MTP drafters by name and include them in reload dedup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: manage --model-draft in extras and reject MTP/ copies as models
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Trim and tighten code comments and docstrings across the repository. Comment-only: every changed file verified code-identical to main via AST/token comparison.
Trim and tighten code comments and docstrings across studio/ Python. Comment-only: every changed file verified code-identical to main via AST/token comparison.
Raise ruff line-length to 100 and extend the local pre-commit format pipeline (def-signature magic-comma normalization, short multi-line assert collapse, kwarg '=' spacing, blank-line-after-short-import removal, adjacent string-literal / f-string+plain merge, redundant-pass pruning). Every transform re-checks the file AST and is dropped if it would differ; the whole-repo reformat is verified AST-identical per file and idempotent.
* fix: honor --ctx-size and other forwarded args from `unsloth studio run` in Studio's context-fit logic
* refactor: extract resolve_requested_ctx as single source of truth
The test helper was reimplementing the two-line
'ctx_override = parse_ctx_override(...); requested_ctx = ctx_override
if ctx_override is not None else n_ctx' pattern locally, so the test
asserted against its own reimplementation rather than production logic.
Extract the conditional into resolve_requested_ctx and have both the
production caller and the test use it.
* fix(studio): honor pass-through cache type flags in KV VRAM estimate
Studio's KV cache VRAM estimate computed from the first-class
cache_type_kv even when the user passed -ctk/--cache-type-k/-ctv/
--cache-type-v via extras. Those flags reached llama-server fine
(last-wins on the CLI) but the pre-launch estimate kept using the
default f16 bytes-per-element, so GPU placement decisions could be
off when the user lowered cache precision via pass-through.
Adds parse_cache_override + resolve_cache_type_kv in llama_server_args.py
(mirroring parse_ctx_override / resolve_requested_ctx), wires both into
load_model alongside the existing ctx resolution, and adds focused
unit tests for the parser + resolver.
Follow-up to @rolandtannous review on #5815.
---------
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Studio: expose --parallel / -np on `unsloth studio run`
The CLI was hardcoding `llama_parallel_slots=4` in `run_kwargs` at
`unsloth_cli/commands/studio.py`, leaving users unable to tune the
concurrent decode slot count even though the engine, KV-cache math,
and `studio.backend.run.run_server(llama_parallel_slots=...)`
plumbing all already accepted any N. This change adds a `--parallel`
/ `--n-parallel` / `-np` typer option (default 4 -- matches the
previous hardcoded value), forwards it into `run_kwargs`, and pins
the new surface with 4 unit tests.
Per-request state in `routes/inference.py` is already isolated
(`cancel_event` and `prev_text` are per-request locals in every
streaming handler; the `_lock` / `_serial_load_lock` only wrap
load/unload, not chat completions), so no concurrency refactor is
needed alongside this -- the engine layer already handles N
concurrent requests on one loaded model when llama-server is told
to.
Range guards: 1 <= N <= 64. With higher N each slot gets ctx/N KV
cache; users tuning this should be aware that per-call context
shrinks proportionally.
`unsloth studio` (the bare default command, no subcommand) still
defaults to llama_parallel_slots=1 via `run_server`'s own default;
this PR does not change that path -- it only exposes the knob on the
one-liner `studio run` command that already silently used 4.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Forward --parallel through venv re-exec and drop colliding short aliases
`unsloth studio run` re-execs into the Studio venv when invoked from
outside it (the common path). The arg-builder forwards every typer
option but the new --parallel, so the child re-execs at the default 4
and any user value is silently dropped. Worse: pre-PR users who
already pass `-np N` as a pass-through extra (where llama.cpp's
last-wins parsing made it stick) silently lose N after this PR lands.
Forward --parallel explicitly in the re-exec arg list.
While auditing the re-exec path, also drop the colliding 1-char
short aliases -m (--model) and -f (--frontend) plus the redundant
-hfr. Click's short-option clustering had been silently mis-parsing
~11 llama-server short flags via the pass-through path: -fa as
`-f a`, -mg 0 as `-m g` + stray 0, -fitt 1024 as `-f itt` + stray
1024, -hff path as `-f f` + stray `-h path`, -cmoe / -cram / -sm /
-ncmoe etc. The docstring promise ("any flag this command does not
recognize is forwarded verbatim") was silently violated.
-hf (2-char) is kept because Click treats multi-char shorts atomically
(no clustering of -hff / -hfv / -hffv / -hft) and -hf is documented
in basics/api/README.md. --model / --hf-repo / --frontend long forms
all unchanged. studio_default keeps -f because it has no pass-through.
Tests:
- test_studio_run_parallel_flag.py: 8 new re-exec coverage cases
(all 3 aliases, 3 platforms via sys.platform mock, pre-PR `-np`
regression, mixed with pass-through extras).
- test_studio_run_short_alias_clashes.py (new): surface checks that
the removed shorts cannot reappear, plus 11 parametrized cases
proving each previously-broken llama-server short flag now passes
through verbatim, plus a happy-path test that documented -hf still
works for `org/repo:variant` syntax.
All 27 tests pass. Negative test (revert either fix) shows the new
tests catch the regression.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix stale studio run docstring describing rejected llama-server flags
The pre-PR docstring listed --port, -c / --ctx-size, --api-key, -ngl,
--jinja, --flash-attn, --no-context-shift as "rejected with HTTP 400",
but only --port and --api-key (plus other networking / auth / model
identity / single-model UI flags) are actually in
studio/backend/core/inference/llama_server_args.py's denylist. -c /
-ngl / --jinja / --flash-attn / --no-context-shift are pass-through
and last-wins-override Studio's auto-set value.
Rewrite the docstring to match the real denylist groups and point at
the canonical source. Also add --parallel to one of the examples now
that it is a first-class flag.
* ci: broaden Linux + narrow Windows llama.cpp runtime patterns + trim #5741 comments (#5746)
* ci: broaden Linux llama.cpp runtime pattern to lib*.so*
#5741 patched the explicit Linux pattern list to add
``libllama-*-impl.so*`` after ggml-org/llama.cpp#23462 (between
b9279 and b9283) split each binary's entry code into a paired
``lib<binary>-impl.so`` shared library. Same class of upstream
repackaging will hit us again whenever a new shared lib is added.
Mirror what macOS already does and replace the per-lib list with a
single ``lib*.so*`` glob. ``copy_globs`` (line 3614) unions
patterns, so the per-variant ``libggml-cuda.so*`` / ``libggml-hip.so*``
entries were never filtering anything; the spec lives in
``runtime_payload_health_groups`` (line 5209) which keeps the
explicit minimum-required list per variant.
Dry-run against b9296-bin-ubuntu-x64.tar.gz: 40 files copied (all
ggml, llama, mtmd, impl variants + the two binaries we ship), 22
skipped (other CLIs, rpc-server, LICENSE). Functionally equal to
the post-#5741 set.
* cleanup: trim #5741 comments on the pydantic split
Comments added in #5741 explained the original bug in full each
time. They are mostly redundant with the commit message and the PR.
Trim them to one short paragraph per site.
No behavior change.
* ci: narrow Windows runtime pattern to llama-server.exe + llama-quantize.exe
Studio only invokes llama-server and llama-quantize. Mac and Linux
already filter to those two binaries; Windows was the odd one out
with ``*.exe`` copying every CLI upstream ships (llama-cli,
llama-bench, llama-mtmd-cli, ...).
Dry-run on b9296 (win cpu-x64, cpu-arm64, cuda-13.1, hip-radeon):
20 unused EXEs skipped per variant, all DLLs (incl. the new
llama-*-impl.dll family) still copied via ``*.dll``.
``existing_install_matches_choice`` already checks llama-server.exe
exists explicitly (line 5297), so the health gate is unchanged.
* Lower default weight_decay in RL config from 0.01 to 0.001 (#5747)
In full FT, AdamW weight decay shrinks the parameter directly so the
implicit prior is W -> 0. In LoRA the trained parameters are A and B
while the effective weight is W = W_init + (alpha/r) * B @ A; decaying
A and B separately drives BA -> 0, hence W -> W_init rather than 0.
The previous default of 0.01 inherited from full-FT recipes adds a
measurable pull on the merged adapter back toward the base model over
a few thousand steps. 0.001 keeps a small Frobenius-norm prior on
||A||^2 + ||B||^2 for numerical stability without meaningfully biasing
the merged weight toward init, and aligns with the value used across
the unsloth notebook templates.
* Studio: strip orphan tool_call XML leaking into visible content (#5735)
* Studio: strip orphan tool_call XML from streamed visible content
The speculative-buffer state machine in
`studio/backend/core/inference/llama_cpp.py` can slice a tool_call XML
block between the silent DRAINING path and the user-visible
content_accum, depending on when in the model's emission the BUFFERING
-> STREAMING -> DRAINING transitions fire. Three leak shapes were
observed in a 2026-05-22 sweep of 900 Qwen3.5 / Qwen3.6 GGUF runs:
Pre-fix XML leak rate: 20/900 (2.22%), concentrated 6.7% on the
larger Q8 / MTP configs:
Qwen3.6-35B-A3B Q8_0 4/60 (6.7%)
Qwen3.6-35B-A3B-MTP Q4 4/60 (6.7%)
Qwen3.5-35B-A3B Q8_0 3/60 (5.0%)
Qwen3.6-27B Q8_0 3/60 (5.0%)
The existing `_TOOL_XML_RE` only matched well-formed
`<tool_call>...</tool_call>` and `<function=...></function>` pairs, so
unterminated openings (close was DRAINED) and orphan closes (opening
was DRAINED) survived the strip and reached the user.
Fix relaxes the regex to also strip:
1. Orphan opening up to end-of-string: `(?:</tool_call>|\Z)`
2. Orphan closing tag: bare `</tool_call>` / `</function>`
Verified on the full sweep: 20/900 -> 0/900 (100% of detected leaks
eliminated). 16 unit tests in `test_tool_xml_strip.py` pin all three
leak shapes plus the well-formed cases, plus parametrised checks on
the 5 actual real-world leak samples from the sweep data.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strip tail-only </parameter> orphan + tighten regex
The 2026-05-22 gdpval sweep surfaced a 4th XML-leak shape not caught
by the earlier regex: a bare `</parameter>\n\n` at end-of-buffer (7
of 192 trials, all Qwen3.5-27B + a few Qwen3.6-27B). The model emits
the full `<tool_call><function=...><parameter=...>...content...
</parameter></function></tool_call>` envelope, the speculative buffer
DRAINS the opening tags as intended, but EOS (max_tokens cutoff)
truncates the outer `</function></tool_call>` close, leaving just
`</parameter>` as the visible tail.
We strip this ONLY when end-anchored (`\s*\Z`) so legitimate
mid-text uses (user code samples, documentation discussing the
Qwen tool-call XML shape) survive. Verified on the 192-trial
gdpval corpus: before=7, after=0.
While at it, fold the five top-level alternations into three by
sharing tag-name and prefix subgroups:
<tool_call>... + <function=\w+>... + --> <(?:tool_call|function=\w+)>...
</tool_call> | </function> --> </(?:tool_call|function)>
Semantically identical (verified by replay over the 192-trial
corpus + adversarial inputs, 0 diffs) and 1.34x faster on real
workloads. Backtracking-safety pinned by two new perf guards
(256KB '<' spam, 1000x orphan opens).
Tests: 16 -> 28 (6 new functional + 4 well-formed-vs-orphan +
2 perf guards).
* Tighten comments in XML-strip regex and tests
Code says what it does; comments were repeating it. Strip the verbose
explanations down to the WHY-only bits (engine quirk, tail-anchor
rationale, real-world source of each test sample). No code changes.
inference.py: 21 -> 12 lines around _TOOL_XML_RE
test_tool_xml_strip.py: 343 -> 259 lines (-84)
Tests: 28/28 still pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Address review: deny pass-through --parallel, preserve legacy short aliases, fix test harness
Round 1 review fixes for #5737:
1. Deny --parallel / --n-parallel / -np in the pass-through validator.
Without this, `unsloth studio run --model X --parallel 8 -- --parallel
999` would last-win-override the running llama-server slot count while
Studio's app.state.llama_parallel_slots and KV-cache fitting stay at
the typer value (8), so the resource plan and the running process
disagree. Also bypasses the typer 1..64 range guard. Reject so the
only path is the first-class typer flag.
2. Backwards-compat shim for -m / -hfr / -f. Dropping the short aliases
from typer broke any script using `unsloth studio run -m X` or
`-hfr Y` or `-f dist`. Add _consume_legacy_short_aliases which pops
EXACT whole-token matches (or `-x=value` inline form) from ctx.args
into the corresponding typer parameter. Clustered tokens (`-fa`,
`-mg`, `-fitt`, ...) are left in the pass-through tail unchanged.
--model becomes Optional with an explicit missing-required check
after the preprocessor so legacy `-m X` still satisfies the
"must specify a model" requirement.
3. Drop mix_stderr from CliRunner. Typer 0.25.1 / Click 8.4.1 removed
the kwarg; the test harness raised TypeError before exercising the
PR behaviour. Tests run cleanly on current and older Typer/Click.
4. Correct the -np regression test docstring. Pre-PR `-np 8` was
clustered by Click as `-p 8` (port=8) + stray `-n`, silently
breaking the port binding -- not "passed through as 8 slots". The
post-PR assertion (child gets --parallel 8) is unchanged.
5. Update studio run docstring listing rejected flags so it now
correctly includes --parallel / -np / --n-parallel.
New tests:
- test_llama_server_args.py: parametrized denylist coverage for
--parallel / --n-parallel / -np including equals-form, including
out-of-range bypass attempts (999, 0). is_managed_flag flips True.
- test_studio_run_short_alias_clashes.py: legacy -m / -hfr / -f
promote to typer params; --model X + -m Y conflict errors; clustered
-mg / -fa / -fitt still pass through (the original bug fix holds).
132 tests pass (98 backend + 34 cli).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Extend legacy-alias shim tests for repo:variant, inline value form, and missing model
Three additional edge cases for the -m / -hfr / -f preprocessor:
- `-m unsloth/foo:UD-Q4_K_XL` round-trips through both the preprocessor
and _split_repo_variant so the child sees --model + --gguf-variant.
- `-m=foo` inline value form is promoted just like `-m foo`.
- Missing --model after the preprocessor raises typer.Exit(2) cleanly
(replacing typer's pre-PR required-flag enforcement now that --model
is Optional to allow the legacy promotion path).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scrub .github/workflows for staging push (matches staging base)
* Fix studio CLI argv handling and pass-through docstring drift
- studio/backend/core/inference/llama_server_args.py: drop the stale
``-np``/``--parallel`` entry from the docstring's pass-through tunable
list. These flags moved into _DENYLIST_GROUPS so the docstring now
contradicts the validator and would mislead future maintainers
debugging the ValueError from validate_extra_args(["--parallel","8"]).
The deleted wording was introduced by dbea77e34 ("Studio: forward
llama-server args from `unsloth studio run`, activate `unsloth run`,
and allow passing model:quant to load models") when --parallel was
still a documented pass-through; the same commit's "quant" reference
is about the model:quant syntax, unrelated to the parallel slot
wording being deleted here.
- unsloth_cli/commands/studio.py: add _expand_attached_np_short next to
_consume_legacy_short_aliases. Both work around Click's short-option
clustering for this command -- the legacy preprocessor for `-m` / `-f`
/ `-hfr` and this one for the attached `-np<N>` form. Click clusters
`-np8` as `-n -p 8` because `-p` is the typer short for `--port`,
silently setting port=8 and dropping the parallel value; rewriting the
attached form into separated `-np <N>` in sys.argv before Click
parses preserves the user's value. Space/equals forms (`-np 8`,
`-np=8`) already work and are left alone.
- unsloth_cli/__init__.py: import _expand_attached_np_short from the
studio command and run it only when argv[0] looks like the unsloth
console-script or workspace cli.py, so importing this module from a
notebook or pytest run does not mutate the caller's argv.
* Tighten the -np canonicaliser comments
Drop the helper's co-location sentence (location is self-evident from
grep) and shorten the entry-gate rationale to one short sentence
covering the why.
* Sync .github/workflows with upstream author branch
* Sync .github/workflows with upstream author branch
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Bump install.sh / install.ps1 pin to unsloth>=2026.5.7 (#5753)
PyPI release unsloth 2026.5.7 is now live. Bumps the pinned floor in
install.sh and install.ps1 from unsloth>=2026.5.6 to unsloth>=2026.5.7
so fresh installs resolve to the new wheel.
Tagged on main as v0.1.416-beta.
* Catch attached `-np<N>` form in backend pass-through validator
The CLI-side `_expand_attached_np_short` rewrites `-np8` to `-np 8`
before Click parses, but HTTP /load `llama_extra_args=["-np8"]` goes
straight to `validate_extra_args` which only matched the exact token.
Reproducer: `validate_extra_args(["-np8"])` previously returned
`["-np8"]` instead of raising; once forwarded to llama-server it
last-win-overrode Studio's slot count while
`app.state.llama_parallel_slots` stayed at the typer value.
Normalise `-np<digits>` to `-np` in `_flag_name` so the denylist
catches the attached form alongside `-np`, `-np=8`, `--parallel`,
`--parallel=8`, and `--n-parallel`. Tests parametrize the new form
including out-of-range values.
* Restore _consume_legacy_short_aliases unit tests + _expand_attached_np_short tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restore .github/workflows from origin/main
Earlier merge from claude_review's staging-scrub commits accidentally
deleted production CI workflows. Restore them to main's state.
* Scrub .github/workflows for staging push (matches staging base)
* Sync .github/workflows with upstream author branch
* Round 5+6: broaden -np gate to exact basenames + runtime parallel test
Reviewer-flagged improvements squashed into one commit so the auto-push
review bot doesn't keep stomping the branch:
- unsloth_cli/__init__.py: exact-basename match instead of
endswith('cli.py'). Covers unsloth, unsloth.exe, unsloth-cli,
unsloth-cli.exe, cli.py, unsloth-cli.py. A third-party mycli.py that
happens to import unsloth_cli no longer has its argv mutated.
- unsloth_cli/tests/test_studio_run_parallel_flag.py: parametrised
runtime test (N in {1, 4, 8, 64}) that fakes the in-venv path and
asserts run_server is invoked with llama_parallel_slots=N.
Complements the existing source-text check so refactors that preserve
runtime semantics don't trip a false failure.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Round 7: respect '--' end-of-options and reject flag-as-value
Round 7 reviewer flagged three legitimate edge cases:
- _expand_attached_np_short rewrote post-'--' tokens. Convention: '--'
ends option processing; payload after it is raw. Stop the loop there.
- _consume_legacy_short_aliases promoted post-'--' legacy aliases for
the same reason. Treat post-'--' tail as raw.
- Legacy '-m -fa' silently consumed '-fa' as the model name, hiding
the real CLI shape error. Reject any next-token that starts with '-'
(except the lone '-' stdin/path sentinel) with a clear BadParameter.
Also expanded the missing-model error string to mention the still-
supported legacy '-m' / '-hfr' aliases so users hitting that diagnostic
on legacy scripts get the right migration hint.
Added four regression tests covering each new behaviour.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Round 8: soften flag-as-value to long-form only + normalise is_managed_flag
Round 8 reviewer flagged two cleanups:
- _consume_legacy_short_aliases rejected any next token starting with
'-' as a flag, which would break legitimate values like '-foo'
(path or model name with leading dash). Narrow the rejection to
'--long' tokens only; '-x' short forms still pass through.
- is_managed_flag did raw _DENYLIST membership while validate_extra_args
goes through _flag_name first, so '-np8' / '--parallel=8' /
'--port=9000' classified as not-managed by the helper but rejected
by the validator. Route is_managed_flag through _flag_name so the
two helpers agree on every form callers might use.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Round 9: also catch -np-1 / -np+1 signed attached forms in denylist
Round 9 reviewer noticed _flag_name normalised -np<digits> but missed
signed variants -np-1 and -np+1, so validate_extra_args waved them
through while rejecting --parallel -1. llama.cpp would error out on
negative slot counts anyway, but the validator should classify every
form of the managed flag identically so the boundary is consistent.
* Round 10: signed -np in CLI canonicaliser + reject empty inline aliases
Round 10 reviewer flagged two real issues:
- _expand_attached_np_short rewrote only -np<digits>; signed forms
-np-1 / -np+1 fell through. Backend _flag_name already classifies
them as managed, so the CLI rewriter must too -- otherwise Click
clusters -np-1 into -n -p -1 (port=-1) and never reaches the
backend validator at all.
- -m= / -hfr= / -f= empty inline forms were accepted and produced
--model '' / --frontend '' (then Path('') silently became '.') on
re-exec. Reject empty inline values at the preprocessor with a
clear BadParameter so the malformed input fails fast.
Both behaviours pinned with parametrised regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Expose --parallel on plain `unsloth studio` for API-path parity
The PR added --parallel to `unsloth studio run` but the plain
`unsloth studio` callback (used for API-only / bare-server launches)
still hardcoded llama_parallel_slots to its run_server default. With
--parallel now denied as a llama_extra_args pass-through, that flow
had no first-class way to raise concurrency.
- unsloth_cli/commands/studio.py: add --parallel / --n-parallel typer
Option (default 4, range 1..64) to studio_default, forward through
the venv re-exec, and pass llama_parallel_slots= to run_server in
the in-venv path.
- studio/backend/run.py: argparse --parallel / --n-parallel with the
same range guard so the spawned child accepts the forwarded flag.
- unsloth_cli/tests/test_studio_run_parallel_flag.py: test pins the
new option presence, aliases, default and range guards.
* Round 12: narrow entry-point gate, preserve pre-PR plain-studio default, drop brittle source-text test
Three Opus subagent reviewers (security / backcompat / code-quality)
flagged the same handful of real issues. Consensus fixes:
- unsloth_cli/__init__.py: narrow the -np canonicaliser gate to just
{unsloth, unsloth.exe} (the only pyproject-declared console_script).
The previous cli.py / unsloth-cli.py entries would silently rewrite
sys.argv for any third-party myproj/cli.py that happens to import
unsloth_cli. Dev users running python cli.py ... -np N still work
via the space form, which parses without the rewrite.
- unsloth_cli/commands/studio.py + studio/backend/run.py: restore the
pre-PR llama_parallel_slots default of 1 on plain unsloth studio and
python studio/backend/run.py. unsloth studio run keeps its
hardcoded-pre-PR default of 4. Without this, my earlier API-path
parity commit silently dropped per-call context to ctx/4 for the
plain-studio flow.
- unsloth_cli/tests/test_studio_run_parallel_flag.py: drop the brittle
source-text grep test (test_run_kwargs_use_parallel_value). The
parametrised runtime test test_in_venv_path_passes_parallel_to_run_server
already pins the same intent against actual behaviour.
- unsloth_cli/tests/test_studio_run_short_alias_clashes.py: pin the
narrow entry-point gate with a parametrised negative test covering
seven third-party argv[0] basenames (cli.py, /path/myproj/cli.py,
pytest, unsloth-cli, etc.). Re-broadening the gate now trips a
test instead of silently mutating an unrelated CLI's argv.
* Round 13: shared parallel constants, denylist invariant test, defence-in-depth
Three Opus subagent reviewers (adversarial-user / maintenance /
cross-file consistency) flagged a consistent set of cleanups; folded
into one commit to avoid the pre-commit.ci force-push race.
unsloth_cli/commands/studio.py:
- Extract _PARALLEL_MIN / _PARALLEL_MAX / _PARALLEL_DEFAULT_RUN /
_PARALLEL_DEFAULT_PLAIN module-level constants and use them in both
typer Options (plain studio_default = 1, studio run = 4).
- _expand_attached_np_short now rewrites -np<junk> when the suffix
starts with a digit (or signed digit) so '-np8x' surfaces as a
clean '-np takes an int' typer error instead of a baffling
'--port invalid' complaint after Click clusters '-n -p 8x'.
- Re-exec forwarding emits --load-in-4bit / --no-load-in-4bit
explicitly in both directions; previously the True default relied
on both layers sharing the same default forever.
- run() docstring now explicitly says --parallel / -np pass-through
via llama_extra_args is denied (use the typer flag above).
studio/backend/run.py:
- Mirror the parallel constants and route the argparse default,
range check, and error message through them. Help text mentions
the asymmetry with 'unsloth studio run' so direct-launch dev users
aren't confused by Default 1 in isolation.
studio/backend/core/inference/llama_server_args.py:
- _flag_name strips surrounding whitespace before denylist lookup so
a caller can't slip a managed flag past the boundary with a
trailing space (the trimmed form is what downstream parsers see).
Tests:
- New typer-aliases-subset-of-denylist invariant: every alias the
typer Option claims as --parallel on run() MUST be in the backend
parallel denylist group. Catches the failure mode where someone
adds a new alias and forgets the boundary.
- Extended denylist parametrize to cover ~14 previously untested
aliases (-mu, -dr, -hfv/-hfrv/-hffv family, -mmu, full --ui group,
--models-preset / --models-autoload / --no-models-autoload).
- Whitespace-padded denylist rejection (' --parallel', '-np ', etc).
- --load-in-4bit re-exec test pinning both polarities + default.
- -np<junk> argv rewriter regression tests.
- Cross-reference headers between the two test files.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: repair mlx studio base export save_method (#5727)
* Round 14: align backend -np recogniser with CLI rewriter + reject parent --parallel
Round 14 (reviewer.py --parallel 20 with gpt-5.3-codex-spark) flagged
two real P1s and a stale-rebase warning. All three addressed.
- studio/backend/core/inference/llama_server_args.py: widen
_flag_name so -np<digit-prefix> with trailing junk (-np8x,
-np-1foo, -np+1bar, -np9zzz) classifies as managed flag -np,
matching the CLI _expand_attached_np_short rewriter. Without this,
POST /api/inference/load with llama_extra_args=['-np8x'] slipped
past the boundary while the CLI canonicalised the same form. The
two sides now agree on every digit-prefix form.
- unsloth_cli/commands/studio.py: reject --parallel on the
studio group when a subcommand is invoked. Pre-PR the studio
callback had no --parallel; my Round 12 addition made
'unsloth studio --parallel 8 run ...' silently drop the 8
because typer doesn't propagate parent options into subcommand
kwargs. Now errors with exit 2 and a message pointing the
operator at the correct invocation
('unsloth studio run --parallel 8 ...').
- Picked up origin/main via merge (parent commit 0caf0526): the
pre-flight stale-rebase detector found 2 lines on main in
studio/backend/core/export/export.py missing from PR HEAD.
Merged cleanly with no conflicts.
Tests:
- Parametrised denylist coverage for -np<digit-prefix>+junk forms.
- New runtime test confirms exit 2 + helpful error when the group
--parallel is supplied alongside an invoked subcommand.
- Test that the default group --parallel value still lets a
subcommand resolve (no false-positive rejection).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten code comments across --parallel PR
Comment-only pass over the seven PR-touched files; trim verbose
docstrings, collapse multi-line section dividers, and drop
redundant prose that the code already conveys. No behaviour change.
* Studio: trim remaining verbose docstrings missed in last pass
Shorten the test_studio_run_parallel_flag.py module docstring and
the `Re-exec arg-builder coverage` block. No behaviour change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: second comment-tightening pass across PR-touched code
Trim docstrings and inline comments in studio.py, run.py,
llama_server_args.py, and unsloth_cli/__init__.py. No behaviour change;
all 215 tests still pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: deny --embedding / --rerank / --tools pass-through
`--embedding` and `--rerank` flip llama-server into single-endpoint
mode, which breaks Studio's /v1/chat/completions hop. llama-server's
own `--tools` flag silently stacks on top of Studio's tool policy
resolved by `--enable-tools` / `--disable-tools`.
Add all three (plus the `--embeddings` / `--reranking` plural aliases)
to the boundary denylist so HTTP /load and pass-through extras both
reject them cleanly instead of silently desyncing the server surface.
Test added to the existing `test_denylist_rejects_all_aliases`
parametrize. 220 tests pass.
* Studio: make PR-touched tests robust to minimal envs + Windows
Two cross-OS CI findings:
1. `test_typer_parallel_aliases_are_subset_of_backend_denylist` was
doing `from core.inference.llama_server_args import _DENYLIST_GROUPS`
which triggers `core/inference/__init__.py` and pulls in the full
backend chain (fastapi / structlog / loggers / utils.hardware).
The invariant only needs the constants tuple, so load the module
directly via `importlib.util.spec_from_file_location` -- the test
now runs with just typer + pytest installed.
2. `test_legacy_frontend_alias_still_promotes_to_frontend` asserted
the literal string `"/tmp/dist"` after the value round-trips through
`Path()`. On Windows `str(Path("/tmp/dist"))` is `"\tmp\dist"`, so
the assertion tripped on the same logical path. Compare via
`Path(x) == Path("/tmp/dist")` so the test passes on every OS.
Both surfaced by the staging-4 cross-OS CI; no production-code change.
220 tests still pass locally.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: load llama_server_args.py directly in its unit tests
Same fix as the previous CLI-test commit: import the module via
`importlib.util.spec_from_file_location` instead of
`from core.inference.llama_server_args import ...`, so the test no
longer needs the full backend chain (fastapi / structlog / loggers /
utils.hardware) installed via `core/inference/__init__.py`.
The boundary validator is intentionally dependency-free; its unit
tests should reflect that.
* Fix test_main_composer_has_dir_auto anchor after PR #5784
PR #5784 ("Improve image generation UI") rewrote the message-input
textarea's static `aria-label="Message input"` into a JSX conditional
`aria-label={overlay ? "Image edit instructions" : "Message input"}`
but did not update the RTL bidi-attribute regression test, leaving
the literal-string `find('aria-label="Message input"')` anchor with
no match. The `Repo tests (CPU)` job has been red on main since.
Anchor on the inner `"Message input"` string literal instead -- it
survives both spellings and still pins the same textarea element so
the `dir="auto"` assertion has the right block to inspect.
Verified by re-running the exact CI command:
954 passed, 3 skipped, 23 deselected (was 948 passed, 1 failed).
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Long Yixing <longyixing331@gmail.com>
* studio: emit one comma-chained --spec-type for CPU/Mac MTP path
llama-server takes a single --spec-type whose value may be
comma-separated to chain implementations (e.g. ngram-mod,draft-mtp).
The CPU/Mac MTP branch in LlamaCppBackend.load_model was passing
--spec-type twice in the same invocation, which is not the documented
chaining mechanism and silently drops one of the two specs depending
on llama.cpp's argv handling.
Collapse the pair to --spec-type ngram-mod,{mtp_token} and update the
stale _extra_args_set_spec_type docstring that claimed llama-server
accumulates repeated --spec-type. Update the matching pass-through
fixture in test_llama_server_args.py.
* studio: align MTP ngram-mod knobs with llama.cpp upstream defaults
Two correctness fixes against the llama.cpp server README:
1. The CPU/Mac comma-chained branch was emitting
--spec-ngram-mod-n-max 6 with --spec-ngram-mod-n-min 48, which is
nonsensical (min > max). Per the upstream default the value is 64.
2. The standalone ngram-mod branch was emitting --spec-ngram-size-n,
--draft-min, --draft-max. llama.cpp removed those arg aliases for
ngram-mod (they live only on the ngram-simple / map families now);
the correct knobs are --spec-ngram-mod-n-match / n-min / n-max.
Also refresh the inline comment block to point at the server README
rather than the older docs/speculative.md draft- aliases.
* Studio: auto-enable MTP speculative decoding for MTP GGUFs
Detect Unsloth's MTP (multi-token-prediction) GGUFs and auto-emit the
right --spec-type draft-mtp flags for llama-server (llama.cpp PR
#22673), so users get the speedup without configuration.
Detection prefers the GGUF metadata field <arch>.nextn_predict_layers
(verified on Qwen3.6-27B-MTP-GGUF / qwen35 and Qwen3.6-35B-A3B-MTP-GGUF
/ qwen35moe). Falls back to a -MTP marker in the identifier / filename
so HF-mode loads can detect MTP from the repo name before the GGUF is
downloaded.
Flag presets follow the Unsloth MTP guide:
GPU: --spec-type draft-mtp --spec-draft-n-max 6
CPU/Mac: --spec-type draft-mtp --spec-draft-n-max 3 \
--spec-type ngram-mod --spec-ngram-mod-n-match 24 \
--spec-ngram-mod-n-min 48 --spec-ngram-mod-n-max 6
User overrides win: if the caller passes --spec-type / --spec-default
via unsloth run / unsloth studio run pass-through (or HTTP
llama_extra_args), the auto-emit steps aside so llama-server only sees
the user's flag. Scalar tuning knobs like --spec-draft-n-max compose
with the auto preset via llama-server's last-wins parsing.
_already_in_target_state mirrors the same promotion so a repeat /load
with unchanged settings against an MTP backend running draft-mtp
short-circuits cleanly instead of forcing a reload.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: serialise GGUF reload and inherit unsloth-run extra args
Closes#5401.
Three related GGUF reload bugs reproduced against `unsloth studio run -m unsloth/Qwen3-0.6B-GGUF --gguf-variant Q4_K_M --top-k 20 --seed 42`:
1. The `POST /api/inference/load` already-loaded short-circuit only compared `model_identifier` and `hf_variant`. A same-(model, variant) Apply that flipped `cache_type_kv` / `speculative_type` / `chat_template_override` / `max_seq_length` / `llama_extra_args` returned `status="already_loaded"` and the new setting silently never reached llama-server.
2. The frontend chat-settings Apply path POSTs `/unload` then `/load` without round-tripping `llama_extra_args`. Every reload after `unsloth run --some-flag X` quietly dropped `--some-flag X` from the spawned `llama-server` command line.
3. `LlamaCppBackend.load_model` released `_lock` between Phase 1 (kill) and Phase 3 (spawn) so two concurrent loads each passed Phase 1 with `self._process is None`. Both ran Phase 2 (download), both reached Phase 3, and the Phase 3 defensive `_kill_process()` from #5171 collapsed them to one survivor only after both `subprocess.Popen` calls had landed. For the 86 GB MoE in #5161 / the model in #5401 the overlap window was tens of seconds, long enough to OOM the host. With a 0.6B model the pgrep timeline showed two simultaneous PIDs for 3.3 s on `main`.
Fix:
`studio/backend/core/inference/llama_cpp.py`
* Add `self._serial_load_lock = threading.Lock()`. The whole body of `load_model` runs under this lock so two concurrent `/api/inference/load` requests are strictly sequential. The fine-grained `_lock` and the Phase 3 defensive `_kill_process()` from #5171 are kept as a second layer. `/unload`, `/status`, and `/load-progress` are unaffected because they only touch the fine-grained lock or read properties.
* Add `self._extra_args` plus an `extra_args` property, written inside `load_model` whenever the caller supplies a non-`None` value. `unload_model()` deliberately does not reset it so the route layer can inherit the args across the frontend's `/unload` + `/load` gap.
`studio/backend/routes/inference.py`
* Add `_request_matches_loaded_settings(request, llama_backend)` that compares `max_seq_length`, `cache_type_kv`, `speculative_type`, `chat_template_override`, and `llama_extra_args` between the incoming request and the live backend. Same-(model, variant) requests whose runtime settings differ now fall through to a real reload instead of returning `already_loaded`. A missing `llama_extra_args` field on the request is treated as "inherit current", so the short-circuit still fires when the only difference is the frontend not echoing the CLI flags back.
* GGUF load branch inherits `llama_extra_args` from `llama_backend.extra_args` when the request omits the field, re-validates through `validate_extra_args`, and forwards the result to `load_model(...)`. An explicit `[]` from the caller is still honoured as "clear".
Verified end to end against a live `unsloth studio run` instance:
| Scenario | Before | After |
| --------------------------------------------------------------- | --------- | ------------------------------------------------------------------------ |
| `/load` same (model, variant, settings) | 1 PID, `already_loaded` | unchanged |
| `/load` same model, variant, new `cache_type_kv=q8_0` ctx=8192 | `already_loaded`, settings dropped | `loaded`, `/status` reports the new settings, new server has `-c 8192 --cache-type-k q8_0 --top-k 20 --seed 42` |
| Frontend Apply `/unload` + `/load`, new settings, no `llama_extra_args` field | Drops `--top-k 20 --seed 42` | Preserves `--top-k 20 --seed 42` |
| `/unload` + two parallel `/load` | Two PIDs for 3.3 s | Max simultaneous count = 1 across the full pgrep timeline |
| `/load` with `llama_extra_args=[]` (explicit clear) | n/a | `loaded`, new server has no `--top-k` / `--seed` |
| `/load` with `llama_extra_args=["--top-k","30","--seed","7"]` (override) | n/a | `loaded`, new server has the supplied flags |
`pytest studio/backend/tests` is green except for one pre-existing terminal-width-sensitive assertion (`test_studio_api.py::test_help_output`) and the pre-existing `test_studio_api.py` fixture errors that fail on unmodified main too. No new regressions.
* Studio: track requested n_ctx so Auto-slider flips trigger a reload
Review feedback on PR #5427 from gemini-code-assist.
The original short-circuit compared ``request.max_seq_length`` against
``llama_backend.context_length`` (the effective context). VRAM-fit
logic can cap the running server below what the caller asked for, so
this comparison incorrectly returns ``already_loaded`` when the user
flips the slider from an explicit length (e.g. 8192) back to "Auto"
(0): the explicit request was capped to, say, 4096, and the new "Auto"
request reads ``backend.context_length == 4096`` and decides nothing
changed.
Track the originally requested ``n_ctx`` on the backend instead and
compare against that. ``requested_n_ctx == 0`` means the last load
asked for the model's native length; ``request.max_seq_length == 0``
matches it.
Verified in the sandbox suite (now 90 tests):
- ``test_explicit_to_auto_triggers_reload`` -- loaded with explicit
8192, then Apply with ``max_seq_length=0`` falls through to a real
reload and the new server runs at the native 40960.
- ``test_auto_to_explicit_triggers_reload`` -- inverse direction.
- ``test_explicit_to_same_explicit_short_circuits`` -- re-Apply with
the same explicit value still short-circuits (no needless reload).
- Existing scenarios (kv change, spec change, template change, extra
args inherit, parallel-load stress, frontend Apply flow) unchanged.
``pytest studio/backend/tests`` still green on the same set of tests;
the pre-existing ``test_help_output`` failure and ``test_studio_api``
fixture errors are unaffected.
* Studio: tighten comments in the 5401 fix
Trim the verbose explanatory comments and docstrings introduced in
f9cbec3b and dd0b1d58 down to one-line summaries. The "why" still
points at issue #5401; the multi-paragraph rationale belonged in the
PR body, not the source. No behaviour change.
* ci: retrigger after zoo drift + IPython fixes landed in main
* ci: retrigger Mac Studio UI CI after transient fetch flake
* Studio: address six P2 followups on the 5401 reload PR
Tightens the inheritance and serial-load paths to close the six P2
findings raised by codex-connector on PR #5427 against `f9cbec3b` /
`dd0b1d58`.
1. Re-check loaded state before killing queued loads. Two duplicate
`/api/inference/load` requests both pass the route-level
`is_loaded` gate before the first publishes `_healthy = True`. The
second waits on `_serial_load_lock`, enters Phase 1, and tears down
the just-spawned llama-server for a redundant full reload. Added
`LlamaCppBackend._already_in_target_state(...)` and a short-circuit
at the top of the serial-lock block: if the live server already
satisfies the kwargs, return True without killing.
2. Don't inherit CLI overrides that shadow new first-class settings.
`unsloth run -c 4096` is a permitted pass-through; the validator
docs explicitly call out `-c`/`--ctx-size`. Stored in `_extra_args`
and appended after Studio's own flags, the inherited `-c 4096`
silently won the last-wins parse against a new
`max_seq_length=8192`. Added `strip_shadowing_flags` in
`llama_server_args.py` (covers `-c`, `--cache-type-k/v`, `--spec-*`,
`--chat-template*`, `--jinja`/`--no-jinja`) and the route runs the
inherited list through it before validate + forward.
3. Restrict inherited llama args to the same GGUF model. `_extra_args`
is deliberately preserved across `unload_model()` for the chat-
settings Apply flow (`/unload` + `/load` with no `llama_extra_args`
field). Now also track `_extra_args_source = (model_identifier,
hf_variant)` so the route can refuse cross-model inheritance.
`LlamaCppBackend.extra_args_source` exposes the tuple.
4. Persist extras only after a successful load. `_extra_args` was
written at the top of `load_model` before Popen + health check, so
a failed startup left bad args in place to poison the next UI
retry. The write (along with `_requested_n_ctx`) is now deferred
until after `_healthy = True`.
5. Ignore speculative diffs for vision loads. `load_model` silently
gates speculative decoding on `not is_vision`, so the backend's
`_speculative_type` stays `None` for vision models. The route's
comparator now normalises the request's value to `"off"` when
`llama_backend.is_vision` to avoid a no-op reload of a vision
server every time the dropdown defaults to `default`. The
`_already_in_target_state` helper applies the same rule.
6. Wait for the replacement server before short-circuiting. `_kill_process`
did not clear `_healthy`; the new first-class settings
(`_cache_type_kv`, `_speculative_type`, `_chat_template_override`)
are written under `_lock` BEFORE Popen + `_wait_for_health`. A
duplicate `/load` arriving during the new server's warm-up window
could short-circuit against the not-yet-healthy replacement and the
caller would start inference against a server that was still
loading. `_kill_process` now sets `_healthy = False` in its
`finally` block so `is_loaded` returns False from the moment the
old server is killed until the new one finishes warm-up.
Tests:
- Sandbox suite under `./temp/sim_5401/` extended to 136 tests (was
90): new unit coverage for `strip_shadowing_flags` (12 cases),
`_kill_process` clears `_healthy`, `extra_args_source` lifecycle and
cross-model behaviour, failed-load preserving prior extras, and the
duplicate-load short-circuit at `load_model` level. New live
integration cases verify shadow-strip via `pgrep` on the live
llama-server cmdline, cross-model refusal, and PID stability across
a duplicate-load race. All 136 pass.
- `pytest studio/backend/tests --deselect test_studio_api.py`:
1079 passed, 46 skipped, identical to the pre-change count. The
pre-existing `test_studio_api.py` fixture errors and the
terminal-width-sensitive `test_help_output` are unaffected.
- Ruff: clean on the three modified files.
* Studio: tighten GGUF reload inheritance and duplicate-load guard
Re-narrow llama_extra_args to None after validate_extra_args when the
incoming request omitted the field, so the backend can distinguish
"caller omitted, inherit prior load" from "caller explicitly cleared
to []". Without this a queued duplicate /load reaches the backend as
[] and fails _already_in_target_state's exact-equality check, killing
the just-started llama-server. The pass-through validate call from
the original "forward llama-server args from unsloth studio run /
unsloth run" change is preserved as-is; only the post-pass narrowing
is new. Cross-source loads now explicitly clear extras so a model
switch can't accidentally inherit via the backend's "no opinion"
semantics.
Store the caller's hf_variant kwarg (None for local GGUF files) in
_extra_args_source instead of the derived self._hf_variant
(an extracted filename quant label like "Q4_K_M"). Same-source check
in the route is now symmetric for HF and direct-file loads.
Add gguf_path to _already_in_target_state and prefer on-disk path
identity when both backend and caller have a path. This stops the
duplicate-load guard from killing a healthy server on repeat local
loads (where hf_variant is None on the caller side but extracted on
the backend side).
Split shadow-flag stripping into per-group toggles (context / cache /
spec / template). The route now opts into stripping only the groups
whose first-class field was actually set on the incoming request, so
an inherited --chat-template-file survives an Apply that omits
chat_template_override. _request_matches_loaded_settings detects
shadowing extras on the inherit path and falls through to a real
reload so the strip can run.
Mark --spec-default, --jinja, --no-jinja as boolean inside the
shadow stripper so the value-consuming heuristic no longer eats the
following positional token.
* Studio: trim comments around GGUF reload inheritance
* Studio: cover GGUF reload inheritance and shadow-flag stripping
* Studio: drop redundant issue refs from inheritance comments
* Studio: drop redundant issue refs from inheritance comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: key inheritance source off resolved gguf_variant
codex-connector P2 on PR #5427cd14cae1: the inheritance gate at
``routes/inference.py:696`` compared the stored ``source[1]`` against
``request.gguf_variant``, but the HF branch loaded with
``hf_variant = config.gguf_variant`` (the *resolved* variant after
ModelConfig auto-pick). When the caller omitted ``gguf_variant`` on a
follow-up Apply, ``source[1] == "Q4_K_M"`` but
``(request.gguf_variant or "") == ""``, ``same_source`` returned False,
and the chat-settings Apply silently dropped CLI pass-through flags
for every auto-pick / local-file load.
Fix both sides of the comparison to key off ``config.gguf_variant``:
* The route compares ``source[1]`` to ``config.gguf_variant`` (the
resolved label) rather than the request field.
* The local-mode load_model call now passes
``hf_variant = config.gguf_variant`` so ``_extra_args_source``
stores the same string the route reads back. The HF branch already
did this.
Sandbox: added test_source_records_caller_variant_not_extracted_label
to lock the storage key contract.
``pytest studio/backend/tests --deselect test_studio_api.py``:
1100 passed, identical to pre-change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: deny upstream --ui family on llama-server pass-through
The validator's web-UI block named only ``--webui`` / ``--no-webui``,
which is llama.cpp's pre-rename spelling. Current upstream
(``tools/server/README.md``) uses ``--ui`` / ``--no-ui`` plus
``--ui-config``, ``--ui-config-file``, and ``--ui-mcp-proxy`` /
``--no-ui-mcp-proxy``. Without these in the denylist a user could
``unsloth run --ui`` and enable llama-server's built-in web UI on
the port Studio's reverse proxy targets, breaking the UI surface.
Keep the legacy ``--webui`` group so the validator still rejects
old binaries that haven't been re-spelled.
Cross-referenced against the README's full flag list; this was the
only gap for the post-#5401 inheritance / shadow-strip work. Pass-
through flags from every other README category (sampling, jinja,
ctx, cache, threads, GPU, reasoning, grammar, chat-template-kwargs)
already validate cleanly; sandbox suite exercises ~60 of them in
the new ``test_08_llama_server_pass_through.py``.
``pytest studio/backend/tests --deselect test_studio_api.py``:
1100 passed, identical to pre-change.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: forward unknown CLI args directly to llama-server
`unsloth studio run --model X --top-k 20 --chat-template-file foo.jinja`
now passes the unknown flags through to the llama-server subprocess.
Adds a denylist for flags Studio manages (port, -m, -c, --api-key, -ngl,
--flash-attn, --no-context-shift, --jinja, GPU-fit, model-identity, ...)
that returns HTTP 400 on collision. HTTP callers can supply the same
list via LoadRequest.llama_extra_args.
* Studio: accept `--model org/repo:variant` shorthand in `unsloth studio run`
Mirrors llama.cpp's `-hf <repo>:<quant>` and ollama's pull syntax so
`unsloth studio run --model unsloth/gpt-oss-20b-GGUF:UD-Q4_K_XL` is
equivalent to `--model unsloth/... --gguf-variant UD-Q4_K_XL`. Local
paths and Windows drive letters are preserved verbatim. If both an
embedded variant and an explicit `--gguf-variant` are given and they
disagree, the command fails with a clear error.
* Studio: register `unsloth run` as alias for `unsloth studio run`
Top-level `unsloth run --model ...` is now equivalent to
`unsloth studio run --model ...`. Same context_settings, so unknown
flags continue to pass through to llama-server.
* Studio: let users override soft-managed llama-server flags from CLI
Trims the denylist to flags Studio fundamentally cannot share with
the user (model identity, --host/--port/--path/--api-prefix,
--api-key, --ssl-*, --webui, --models-*). Soft-managed flags --
-c/--ctx-size, --parallel, --flash-attn, --no-context-shift,
--jinja, -ngl, -t/--threads, --fit* -- now pass through and override
Studio's auto-set version via llama.cpp's last-wins CLI parsing.
Lets users tune their run on the spot:
unsloth run --model X -c 131072 --parallel 1 --threads 32
* Studio: accept `-hf` / `-hfr` / `--hf-repo` as aliases for `--model`
Matches llama-server's `-hf <repo>:<quant>` spelling so users coming
from llama.cpp can use the same flag. Typer claims the aliases before
the pass-through validator runs, so the HTTP-API denylist on those
flags is unaffected.
unsloth run -hf unsloth/gpt-oss-20b-GGUF:UD-Q4_K_XL
Daniel Han