Six tests read family_train_infos() (or a start preflight) without pinning the
host probes, so they only held on a machine with a bf16 accelerator: on a
GPU-less runner the DiT gate empties precision_modes, turns supports_compile
off, and replaces any other preflight message with the no-accelerator note, and
all six failed there. A conftest fixture pins both probes for exactly those
tests, so they assert the family metadata they are about on every host. The
gate's own CPU-only behaviour keeps its dedicated tests.
- The cached-model listing tagged any repo with a model_index.json as
text-to-image, so a community pipeline the image loader's trust rule refuses
still got a row in the Images picker, and a detected-but-untrusted video repo
fell through to that same tag. Gate the image tag on the load path's rule and
hide an untrusted video repo outright.
- A routed diffusion pick only carries a GGUF filename, which is all the chat
picker has, so a curated single-file artifact arrived with no quant and was
loaded as a pipeline: from_pretrained on a repo with no model_index.json. Pass
the page's own catalog spec into the route pick, so a routed pick resolves to
exactly what a direct pick on that page resolves to.
- The dataset mutation endpoints checked is_active() and only then handed their
filesystem work to a thread, so a start reserving in that gap changed captions
or removed images underneath the preflight or the running trainer. The
interlock is now registered for the whole request under the lock reserve()
uses, and a start refuses while a mutation is open rather than waiting on it.
- GIF export held every kept frame as a paletted image before encoding; a clip
may be 2048x2048 for 1024 frames, and at the 12 fps target the step is 1, so
one export click could allocate over 4 GB and take the backend down. Downscale
past 720 px and widen the step to keep at most 300 frames.
- seed accepted any Python int, so an out-of-range one passed every preflight,
evicted the resident models, spawned the trainer and only then died in
torch.manual_seed. Bound it to torch's 64-bit range in the request and config.
- source_revision() scanned the checkpoint root plus text_encoder/tokenizer but
not vae, so swapping or fine-tuning the VAE in place left the conditioning
cache namespace unchanged and a warm run trained against latents from the old
checkpoint. Include the vae directory, like any other component the cached
tensors come from.
- /images/generate answers only when the images are saved, and secure mode's
tunnel caps an origin response near 100 seconds, which a native CPU or a
high-step run passes routinely. The page reported failure while the work kept
running, and a retry would duplicate it. A lost response (fetch rejection or
a gateway status the origin never answered) is now told apart from a refusal:
the page waits out generate-progress and reloads the gallery, so the run it
started still lands.
- The trainer emits the EMA adapter's path with the terminal event, but the
state update dropped it, so neither the run history nor either response
schema carried it and an enabled EMA left nothing discoverable. Keep it, and
show it next to the primary adapter.
- weighting_scheme advertised a choice of timestep sampling; sampling is always
logit-normal and the flag only selects the bell loss weights. Describe what
it does.
Comment-only pass over the Python this PR touches: drop what the code already
says, collapse multi-line explanations that still read on one line, and keep
the reasoning that is not recoverable from the code. No code, docstring
semantics or behaviour changes; verified with an AST comparison against the
previous revision, and the backend suite is unchanged (same 37 environment
failures as before: the API integration tests that need a live keyed server,
the flash-attn install hooks, and the GPU memory fields).
Four fixes from the latest review round:
- The GPU arbiter's chat evictor only cancelled the llama.cpp side. The
orchestrator publishes active_model_name once its worker reports success, so
an in-flight safetensors load was visible only as an entry in loading_models
and finished onto the GPU after ownership had transferred. Cancel every
pending load, and give the safetensors branch the post-load ownership recheck
the GGUF branch already had.
- A manual gpu_layers=0 GGUF load runs on the CPU with the GPUs hidden from the
child, yet it took the arbiter unconditionally: it cancelled a running image
or video generation for a model needing no VRAM, then held CHAT ownership so
the next GPU workload unloaded it for nothing. Gate the acquire on the same
predicate the launch-time CPU-only mask uses, as the image and video loaders
gate on their resolved device.
- The staged-download hook subscribes per repo, not per job, so another job on
the same repo advanced the staged queue (starting a load whose scoped files
were still downloading) or wiped a queue that was still running. Compare the
variant each callback carries, like the chat page's auto-load does.
- The video gallery fetched every record of a page into an object URL that
lives until the page closes: 50 clips at tens to hundreds of MB each, for
cards the user may never scroll to. Fetch a clip as its card nears the strip's
edge, plus the selected one the player needs.
Three from the latest review.
The video download plan always asked for the wide base file list, so an LTX-2.3
pick staged the 2.0 base's VAEs, vocoder and connectors that the checkpoint
supplies itself, while the companion files the 2.3 assembly does read were left
out of the plan and pulled inline at load, outside the panel's progress, cancel
and disk preflight. The plan now recognises a 2.3 pick by name (the load keeps
the authoritative header probe, and under-guessing only falls back to the
load-time pull), narrows the base list, and stages the extras in the same entry
as the checkpoint so one repo stays one scoped job.
A pick routed from the chat picker arrives as ?model= and ?quant= with no picker
metadata, so a bare local .gguf or .safetensors was loaded as a pipeline: an
explicit model_kind wins over the backend's filename sniffing, so it evicted the
resident model and then failed on the missing model_index.json. Both pages now
derive the load kind from the path, the same way their own picker handlers do.
A torchao int8/fp8 build takes adapters only at load time. Switching artifact
inside one family keeps the LoRA selection, since the family did not change,
but the load did not bake it, so the next generation was rejected with 'reload
the model with the adapter selection' while the picker still showed the adapter
as active. The selection is now dropped once per resident build, with a message
saying to pick and load again.
Five fixes from a review pass over the diffusion work.
delete-finetuned rmtree'd a model the Images or Video engine was holding: every
guard on that route is chat-only, and Images loads any local path, so deleting a
local diffusion model under the storage root pulled the weights (and the
companion VAE / text encoders sd.cpp re-reads each generation) out from under a
live pipeline. The cached-model route already refuses this; the trained/exported
one now does too, matching by path rather than repo id, and failing open on a
chat-only install so it cannot block ordinary deletes.
A staged download finishing while its page was hidden loaded the model and
evicted whatever the user was actually using: both diffusion pages stay mounted
behind the router and a load takes the GPU unconditionally. The pick is now held
until its page is on screen again, which is also what chat does.
A scoped download could report success having fetched nothing. With Hugging Face
metadata unavailable no manifest is written, so verification is a no-op, and
snapshot_download returns an existing snapshot folder without downloading when
its own repo_info call fails. A repo already on disk from a full snapshot job
(which ignores *.gguf) therefore completed with no weights and auto-loaded
against them. The requested file list needs no network, so it is checked against
the disk directly.
The XET to HTTP retry reclaimed the job slot without the scoped file list, and
that claim overwrites the stored record, so a later identical scoped start
compared an empty list against the real one and 409'd instead of adopting the
running download.
The DiT accelerator gate probed torch.mps.is_available(), which only exists from
torch 2.5 while the supported floor is 2.4. All three probes shared one
try/except, so on torch 2.4 the AttributeError read as 'no block' and a CPU-only
host still evicted the resident pipeline, downloaded the encoders and died in
the child. Each accelerator is probed on its own now, through
torch.backends.mps.
Six review findings, three of them evict-then-fail orderings:
- The chat load reclaimed the GPU without telling the arbiter it existed. A
chat load holds no llama-server process until its GGUF has downloaded,
which is minutes, so a competing Images/Video acquire in that window
found nothing to cancel, took the GPU, and the chat load then spawned
onto the same device. It now registers an in-flight marker through
acquire_for's register hook (under the arbiter lock, as the image and
video loads do), the evictor cancels a marked load, and the route undoes
itself if ownership moved while it loaded.
- The Hub-download conflict check ran after that handoff, so a GGUF the
download manager already owns destroyed the resident Images/Video
pipeline and then 409'd, having loaded nothing. It moves above the
handoff, together with the marker it handshakes with.
- The image load released the engine router's transition lock before
registering the load, so a second load choosing the other engine could
unload the still-idle engine this one captured; the load then landed on a
deactivated engine, where generate, status, unload and the arbiter's
evictor can no longer reach it. Registration now happens under that lock
and refuses if the engine changed.
- Training a DiT family on a host with no GPU was accepted: nf4 is not a
CPU fallback, its 4-bit load goes through bitsandbytes, which requires
CUDA, XPU or MPS. The start unloaded the working Images pipeline, pulled
the text encoders, and only then died in the child. Rejected before the
teardown now, and /info stops advertising a precision that always 400s.
SDXL keeps its documented fp32-on-CPU path.
- Both diffusion pages kept the routed-pick marker forever, so re-picking
the same checkpoint (after chat evicted it) neither loaded nor cleared
the query string. The marker is released once the query is gone. The
Images key also carried a stray NUL byte, which made the file read as
binary to grep and other tooling.
- diffusers dropped Python 3.9 in 0.38, so the unconditional >=0.39.0 pin
left pip no candidate at all on 3.9 and made every install that composes
the huggingface extras unresolvable there. The floor is conditional now.
Also fixes tests that were already red on the branch: two hand-built
request fakes had gone stale against fields this branch added, and the
handoff-ordering test only failed on a host with fewer than two GPUs.
Four review findings, the first a regression from my own last commit:
- Keying scoped download jobs by a digest of the file set broke the
download manager: it builds that key client-side (it polls and cancels
before any response tells it a key), so it watched and cancelled a key
no worker owned and never fired its ready callback. Keep the derivable
"@scope" key and refuse the second request instead when a live job on
the slot is fetching a different file set -- decided inside the
registry claim, under the lock, so a concurrent claim cannot slip past
it. The manager records the file set on the job as well, so a sibling
quant's transfer is not adopted locally either.
- Both diffusion pages read the route query through a loose useSearch and
both stay mounted once visited, so the hidden one consumed the other's
?model=: it navigated back to its own route and tried to load, say, an
image checkpoint as a video model. Only the visible page consumes it.
- The staged download plan was built without the configured HF token or
the Advanced values the load itself sends. The token matters most: the
backend's Hub metadata lookup is best-effort, so a gated base silently
planned no companion entry and the load pulled those multi-GB files
inline, outside the manager. The memory/quant controls decide whether
the base transformer/ shards are needed at all, and the route dropped
memory_mode, cpu_offload, the prequant path and the LoRA selection
before asking for the plan.
- The video preview kept playing after leaving the page: the keep-alive
layout only hides it, and display:none does not pause a media element,
so a clip the user unmuted kept its audio going over the next page.
Pause on the active transition and do not auto-replay while hidden.
Also completes the hand-built request bodies in the hub download tests:
the scoped-files field this branch added to the route read as an
AttributeError against them, failing five tests.
Five review findings, four of them ways a click did nothing or fetched
the wrong thing:
- A chat pick of a diffusion model routed ggufVariant (a label like
Q4_K_M) in the search param the target page uses verbatim as the GGUF
filename, so the load asked for a file that does not exist. Route
ggufFilename; no filename means a curated non-GGUF pick, loaded as a
pipeline.
- The task-scoped pickers kept only GGUF repos, so the catalog's bf16,
bnb-4bit and single-file fp8 artifacts could not be discovered or
downloaded on the Images and Video pages even though loadSpecFor
knows how to load them. Keep curated artifacts whatever their format,
in Recommended and in Hub search.
- Both pages deduplicated routed selections on the model alone, and
they now stay mounted, so picking the same repo again -- another quant,
or the same one after chat evicted it -- returned early without
loading or clearing the query string. Key on model and quant.
- Every scoped image download shared one @diffusion job key regardless
of the requested files, so switching quant mid-download adopted the
running job: the UI waited on the first file set, then loaded a file
that was never fetched. Include a digest of the file set in the key.
- A scoped plan silently dropped requested files missing from Hub
metadata, and snapshot_download succeeds when an allow pattern matches
nothing, so the job reported completion and triggered a load with
required files absent. Fail the job instead.
Resolves the app-sidebar conflict: main added Hub and Projects rows
inline while this branch renders the nav from navRows in the order and
pin state set under Settings -> Appearance. Kept the data-driven
rendering, having checked both of main's additions are already
represented there - the projects row carries the same icon, label,
active check, handlers and inline New project button.
Main also replaced the sidebar's inline name field with
NewProjectDialog, which owns its own state, so the button no longer
resets a name draft: it sets the move target and opens the dialog, as
main's other call sites do.
* studio: fix Backend CI red on main from an ambiguous ordering anchor
test_load_marker_precedes_hub_guard_and_unload fails on main, so every
open PR against the repo inherits the failure.
Root cause. #7239 (a7761e174) reworked the GGUF GPU-pool validation in
_load_model_impl from "if config.is_gguf and effective_gpu_ids is not
None:" to a bare "if config.is_gguf:", placed earlier in the function
than the GGUF load branch. The test anchors on
source.index("if config.is_gguf:"), a first-match search, so it silently
re-anchored onto the GPU-pool statement. #7251 (95f42bcce) then restored
the assertion "= _resolve_inherited_extra_args(" before
"if config.is_gguf:" against a tree where that anchor already pointed at
the wrong statement, and main went red. Checking out 95f42bcce and
running the suite reproduces the same single failure.
The code is correct. _resolve_inherited_extra_args still runs before the
GGUF load branch and before the hub-download guard that consumes
extra_llama_args for require_mmproj, so the guarantee #7251 protects is
intact; only the assertion is wrong.
Fix. Assert that guarantee behaviourally instead of by source offsets.
The new test drives _load_model_impl over a vision GGUF with a stored
--no-mmproj from a previous same-model load and captures the
require_mmproj the hub guard is called with: inherited --no-mmproj gives
False, nothing to inherit gives True, and an explicit request list wins
over the stored one both ways. Moving the resolution call after the
guard makes the inherited case report True and the test fails, so it
detects the reorder the old assertion was meant to catch, without
depending on how many "if config.is_gguf:" statements the endpoint has.
The surviving marker-before-guard-before-unload assertion had the same
ambiguous anchor for its slice start, silently widening the slice past
the GPU-pool block. It now slices from the "if config.is_gguf:" nearest
above the in-flight marker, which pins the load branch.
The structlog test stub gains a get_logger factory so routes/inference.py
is importable when structlog is absent.
34 pass in tests/test_gguf_load_cache_reuse.py (was 32 pass, 1 fail);
350 pass across it plus test_llama_cpp_mmproj_fallback.py and
test_llama_cpp_mtp_detection.py. A full backend run before and after is
identical apart from this test going from fail to pass.
* studio/tests: repair a pre-existing bare structlog stub before importing routes
* studio/tests: tighten the comments on the new load-ordering coverage
* Tighten comments on the load-ordering coverage for PR #7442
Lets a consumer that reads a deliberate subset of a repo stage it through the
normal download manager. Keyed as "@scope" so it never collides with a quant or
with the repo's full snapshot, and the file list rides the registry so an
XET to HTTP retry respawns the same scoped job.
Reports the repos and exact files a pick needs so the download manager can stage
them with the loader's own file scope. A plain snapshot would add the packaged
root single, transformer shards and fp16 twins the loader never opens.
Both read huggingface_hub's import-time HF_HUB_CACHE, which changing the cache
folder does not update: progress counted the old root while the download wrote to
the new one, and from_pretrained could split one model across both.
* Fix GGUF tool chat server recovery
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover MTP precedence and loosen the replay assertion for PR #7424
Add a regression test for the MTP branch of the tool-loop respawn retry: the
file-wide _make_backend stub forces _maybe_recover_from_mtp_crash to False, so
nothing exercised the case where an MTP crash reload is already claimed and an
ordinary same-config respawn must not run on top of it. Cover both the next
tool-loop request and the final synthesis pass.
Replace the whole-payload equality assertions with a field-wise check. Comparing
the full dict pins max_tokens to the value derived from the dead server's
effective context, so a later fix that rebuilds server-derived defaults after a
respawn would read as a test failure rather than an improvement.
Document that the one-retry budget is per model request, not per chat turn.
* Recover from prefill-time deaths and stop respawn racing the MTP reload
Two gaps in the tool-loop respawn retry, both reproduced before fixing.
A child that exits during prefill has already accepted the socket, so httpx
raises ReadError, WriteError or RemoteProtocolError rather than ConnectError.
Those all arrive before the response opens, which is exactly the window where a
replay is safe, but the helper only caught ConnectError and gave up. Widen the
catch to NetworkError plus RemoteProtocolError. Timeouts stay excluded on
purpose: they mean the server is slow, not dead, and retrying one would spend
the 20 minute first-token budget twice. Windows resets connections where Linux
refuses them, so this also covers the common Windows presentation.
_maybe_recover_from_mtp_crash returns False both when the crash is not an MTP
crash and when an MTP-free reload is already in flight. Callers read that as
permission to respawn, so _respawn_if_dead replayed the crashing MTP kwargs and,
by replacing the process, made the in-flight reload abort on its own newer-load
check. Skip the respawn while that reload owns the corpse. The guard lives in
_respawn_if_dead so the plain chat path gets it too.
Regression tests for both, including a guard against retrying prefill timeouts.
* Release the MTP single-flight claim when the reload never starts
_mtp_runtime_fallback_in_progress is claimed before the reload thread exists, and
only that thread's finally clears it. Two statements ran in between with no unwind
path: re-reading _last_load_kwargs, which an unload can null underneath us, and
Thread.start(), which raises under the thread exhaustion that is exactly the
pressure killing llama-server in the first place. Nothing else ever resets the
flag, so a failure there latched it for the life of the process.
That was survivable before, since respawn ignored the flag. It is not now: the
guard added in db78184be keys off the flag alone, so a latch would silently
disable auto-respawn for every later model, including plain non-MTP ones. Read
the kwargs and process once before claiming, and release the claim if the thread
cannot start.
Restore the whole-payload equality assertions. Comparing field-wise was meant to
leave room for rebuilding server-derived defaults on replay, but the payload is
built once before the retry and re-sent unchanged, so the looser check only
dropped seven real keys and added a vacuous seed comparison.
Also correct the docstring: llama-server flushes its 200 at slot start, so a
death during decode arrives with the response already open. The pre-header window
this covers is an upload still in flight or a request waiting behind busy slots.
* Confirm the child exited before spending the retry
A closing llama-server can beat its own exit status: the socket error arrives while
poll() still reports the process running. _respawn_if_dead then took the alive
branch, handed back the stale _healthy, and the caller read that as a successful
respawn and spent its single retry on the same corpse. When that retry failed,
attempt was no longer 0, so no respawn ever happened and the turn died, with a log
line claiming a respawn that had not occurred. The window matters most for the
pre-header ReadError and RemoteProtocolError shutdowns the retry now covers.
Wait a bounded second for the exit status before calling the child alive. The same
race is already conceded in _maybe_recover_from_mtp_crash, whose recovery thread
polls for 5s because the error can arrive a beat early; 1s here because this runs
on the request path, and a genuinely live server, including one a concurrent caller
has just respawned, still returns promptly.
* Tighten the recovery comments
* Harden the respawn path around concurrent unloads and replacements
Two problems with the reap grace loop, both found by review.
Skip the grace when the server was already replaced. A caller queued on
_respawn_lock behind someone else's respawn woke holding the healthy replacement,
could not tell it from the child its own request had used, and waited out the full
grace. That sleep is under the lock, so the waits serialised: four concurrent
generations cost roughly three grace periods before any retry began. Capture the
process before taking the lock and return early once it has been swapped.
Do not respawn a server that is being torn down on purpose. unload_model() sets
_cancel_event and only clears _last_load_kwargs after the kill, so a request losing
its connection mid-unload could watch that deliberate exit through the grace loop,
read the stale kwargs and load the model straight back; a model switch landing
during the wait was reverted the same way. Re-check the cancel flag and the process
identity under _serial_load_lock before capturing the replay kwargs, matching what
the MTP-crash reload already does.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the respawn comments
* Do not charge the reap grace to a server that is still serving
The grace loop added for the not-yet-reaped race waits on poll(), which for a
live child never returns, so every transient transport error paid the full
_RESPAWN_REAP_GRACE_S. That sleep is held under _respawn_lock, so the cost
serialised: measured 1002 ms for one caller and 8.02 s for eight concurrent ones,
against 0 ms on main. A working install pays this, not a broken one.
A llama-server's listening socket dies with the process, so a loopback connect
separates the two cases in microseconds. Probe it first and return immediately
when the port still accepts; fall through to the grace only when the port is
gone, which is the case the grace exists for. Back to 0.7 ms for one caller and
0.00 s for eight.
Cross-checked on real hardware over Qwen3.5-2B, Llama-3.2-1B, Gemma-3-4B with
mmproj and Qwen3-30B-A3B: decode throughput within noise of main (-0.06%, -3.71%,
+2.57%, +0.29%, against a 54-232% spread between rounds of a single run), output
byte-identical on every round, tool-path recovery restored on the three families
whose model calls the tool, and plain-chat recovery still working on all four.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make the respawn lose to a deliberate unload in every window
Two follow-ups on the respawn path, both reproduced first.
Check _cancel_event before the socket fast path. unload_model sets the flag before
it kills, so the child is still accepting when the probe runs; returning the stale
_healthy there aims the retry at a server that is deliberately going away.
Close the unload TOCTOU. The old cancel check sat under _serial_load_lock, which
unload_model never takes, so an unload could land entirely between that check and
load_model and the captured kwargs would restart a model the user had stopped.
Snapshot the kwargs, the flag and a new _unload_epoch together under _lock, the
lock unload does hold, so a teardown is either wholly before the snapshot or
wholly after it. load_model clears _cancel_event on the way in, so the epoch is
the only evidence that survives; when it moves during the reload the replacement
is unloaded again rather than left running.
_lock stays uncontended across load_model, which would deadlock a plain Lock and
block /status for the length of a load. Error-path latency is unchanged: 0.6 ms
for a live server and 0.00 s for eight concurrent callers.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
- The trainer keyed its persistent conditioning cache on family and
resolution only, while the keys themselves carry just the caption or
image content and crop variant. One cache directory reused for two
checkpoints, or for the same repo at a new revision, let a warm run
skip loading its encoders and train on the other model's embeddings
and latent statistics. Namespace on the base checkpoint and its
resolved revision as well. The revision helper now lives beside the
cache in diffusion_train_extras and the inference wrapper delegates to
it, so the two cannot disagree about what counts as the same source.
- The diffusion learning rate only checked positivity, but 1e309 floats
to inf and satisfies gt, so the route evicted the resident models and
started AdamW with an infinite rate: the first step destroys the
adapter while progress looks normal and the result is saved. Bound it
below 1.0, matching the LLM schema, which rejects inf for the same
reason.
Three review findings:
- The picker filtered every catalog member out of Recommended and Hub
search on the way to canonical group rows, but nothing renders those
rows yet (catalogGroupFitsDevice and groupMatchesQuery are imported and
unused). A task-scoped picker's models list is catalogToModelOptions(),
i.e. group members exclusively, so both lists came back empty and no
curated model could be discovered or downloaded. Keep the artifacts
listed until the grouped UI exists.
- The video delete guard compared only repo_id, so deleting the
companion base of a loaded GGUF video model was allowed even though it
supplies the VAE and text encoders. Compare base_repo too, matching
what the images guard already does for its companions.
- diffusers was declared unversioned while the diffusion stack requires
0.39 (Krea2Pipeline, the cache_context child registries, the Flux2 and
Z-Image pipelines), so an upgrade could keep an older release and
selecting an advertised model failed until the user upgraded by hand.
* studio: shard export checkpoint loads across all visible GPUs
Export checkpoint loading always used unsloth's from_pretrained default of
device_map="sequential", which stacks the whole model on GPU0. On a multi-GPU
host this OOMs GPU0 while the other GPUs sit empty, so a GGUF export that would
comfortably fit across the machine fails with CUDA out of memory (#7053).
Add _multi_gpu_device_map_kwargs(): when the CUDA/ROCm host exposes more than
one visible GPU and get_device_map resolves to "balanced" (the same policy the
inference loader already uses), pass device_map="balanced" to every
from_pretrained in load_checkpoint. In every other case -- single GPU, CPU,
MLX, or any probe failure -- it returns {} so the loader default is untouched.
Fixes#7053
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio/save: reach the UUID/MIG fallback, release sharded models before quantize
Two review fixes on the multi-GPU export sharding:
1. UUID/MIG CUDA_VISIBLE_DEVICES masks resolve to no numeric ids, so the
len(visible) > 1 gate skipped get_device_map entirely and large exports on
those hosts still stacked onto GPU0. An empty id list now routes to
get_device_map(None), whose visible-count fallback exists for exactly this
case; a genuinely GPU-less host still resolves "sequential" and keeps the
loader default.
2. The compressed (FP8/NVFP4) export freed GPU memory before its llm-compressor
subprocess only for single-device models -- a plain .to("cpu") is invalid on
an accelerate-dispatched model, so a multi-GPU-sharded checkpoint stayed
resident on every GPU while the subprocess loaded a second copy. The release
is factored into _offload_model_for_quantize_subprocess /
_restore_model_after_quantize_subprocess: dispatched all-GPU shards get their
accelerate hooks removed, move to CPU, and are re-dispatched over the
recorded hf_device_map afterwards. Maps with cpu/disk targets (already
offloading) and quantized models are left alone, as before.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio/save: budget merged tensors per device, restore hooks if CPU offload fails
Two review fixes on the multi-GPU export path:
1. The LoRA-merge save path budgeted every merged tensor against GPU0
(get_device_properties(0) + unqualified memory_allocated()). A merged tensor
lives on the GPU of its source layer, so for a model sharded across GPUs
(the device_map="balanced" this PR enables) GPU1+ could OOM as their weights
accumulated while only GPU0's headroom was checked. Budget against W's own
device via a per-device cache; single-GPU behavior is unchanged (W on GPU0).
2. _offload_model_for_quantize_subprocess removed the accelerate hooks and then
moved a dispatched model to CPU; if that move raised (host RAM too small for
the sharded checkpoint) the model was left hookless and half-moved, breaking
later exports in the same worker. It now re-dispatches (or, for the
single-device path, moves back) on a failed move before aborting the offload.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio/save: release sharded models before the torchao reload too
The portable torchao FP8/INT8 export freed the in-memory model only when every
parameter sat on one device, then reloaded a second copy with
device_map="auto". A checkpoint loaded through the new multi-GPU export map is
accelerate-dispatched across several GPUs, so that single-device gate never
fired and the original stayed resident on every GPU during the reload -- an OOM
for exactly the models large enough to have needed the sharded load.
It now uses the same _offload_model_for_quantize_subprocess /
_restore_model_after_quantize_subprocess pair as the compressed export, which
removes the accelerate hooks, moves to CPU, and re-dispatches over the recorded
hf_device_map afterwards. Those helpers are extended to XPU as well, since
torchao also runs on Intel GPUs and the path they replace covered both.
* studio/save: release quantized and cpu-spilled shards before quantize reloads
Two cases the release helper skipped outright, both of which leave GPU memory
held while the compressed subprocess or the torchao device_map="auto" reload
allocates a second copy:
- Quantized models. ExportBackend.load_checkpoint loads 4-bit by DEFAULT, so the
common Studio export hit the is_loaded_in_4bit guard and kept a quantized shard
on every visible GPU. They are now attempted like any other model: transformers
refuses .to() for some bitsandbytes builds, but that refusal raises before
anything moves, so the existing recovery path restores the model and returns
None -- best-effort where the stack allows it, old behaviour where it does not.
- Maps that spill to CPU. Any non-GPU target disqualified the whole model even
though the GPU-mapped modules were still resident and are exactly what needs
reclaiming. A cpu spill is safe to move (those weights are already in host RAM)
and is now released; only disk/meta targets are still skipped, because
accelerate keeps those parameters off the model and moving would try to
materialize the whole checkpoint. An all-CPU map is skipped as a no-op.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix multi-GPU offload for PEFT exports and fall back when sharding OOMs (#7215)
The dispatch branch of _offload_model_for_quantize_subprocess never ran for a
PEFT model: the wrapper proxies _hf_hook, so remove_hook_from_submodules raised
AttributeError and the bare except returned None. Studio always loads adapters,
so the new balanced map turned the offload off (0 percent freed against 91.8 on
the sequential path it replaces).
- resolve the real dispatch root before removing or replaying hooks
- snapshot and replay hooks, tensor placements and instance forwards; a plain
re-dispatch rebuilds hooks against the post-PEFT tree (395 to 1379) and drops
the fused kernels accelerate captured into _old_forward before unsloth patched
- drop the accelerator side of tied_params_map so the offload actually frees
- pass skip_keys on the fallback dispatch_model
- log the swallowed exception instead of returning None silently
- guard _unsloth_save_torchao_with_given_config like its two siblings
- retry the export load once on the loader default when the balanced map OOMs,
which happens when a training or chat job already owns the other GPUs
Measured on 4x B200 with Qwen3-0.6B: 89.9 percent freed bf16 and 79.7 percent
4bit under balanced, logits bit-identical, hooks and placements restored
exactly, 184 Params4bit round-tripped unchanged including nested state2.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the original offloaded until the torchao copy is released, and retie shared weights (#7215)
Two follow-ups from review of 8b6b4ca0b.
_unsloth_save_torchao_with_given_config restored the original inside a finally
that ran as soon as from_pretrained returned, so the original and the quantized
copy were both resident while the copy was still being saved. The restore now
sits in an outer finally that covers saving and releasing quantized_model, which
is what the two sibling paths already do.
The dispatch replay did not preserve tied embeddings. A CPU round trip repoints
every tensor and accelerate's tied_params_map is keyed on the old pointer, so
replaying the hooks produced two independent parameters. Reproduced on a tied
Llama: lm_head picked up its own storage, the embedding was duplicated in VRAM,
and an update to one no longer reached the other. The snapshot now records tied
groups (named_parameters(remove_duplicate=False), since the default hides one
half of every pair) and re-ties them after placements are restored.
Verified: tie preserved, no extra storages, live CUDA storage census identical
before and after, updates propagate again, logits bit-identical, and the 4 GPU
invariants unchanged at 89.9 percent freed bf16 and 79.7 percent 4bit.
* Keep meta tensors out of tie groups, restore accelerate move guards, retry CPU spills (#7215)
Four follow-ups from review of a58f1086b.
Meta tensors all report storage pointer 0, and accelerate parks every
CPU-offloaded parameter on meta, so grouping by pointer collapsed them into one
fake tied group. Reproduced with a balanced map that spills two blocks to CPU:
18 meta parameters in a single group with shapes 64x64, 32x64 and 128x64, which
the retie step would have overwritten with the first one. Meta and null-pointer
tensors are now skipped, and the retie also checks shape.
remove_hook_from_submodules deletes the to/cuda/xpu wrappers dispatch_model
installs to stop a caller moving an offloaded model. The snapshot now records
and replays those alongside forward and _old_forward.
The single-device retry only matched OOM, but a balanced map that spills to CPU
is refused by bitsandbytes with a plain ValueError saying modules were dispatched
to the CPU or the disk (transformers quantizers/quantizer_bnb_4bit.py:128), with
no memory wording. That is now retryable too, which matters because Studio loads
4-bit by default and busy secondary GPUs are exactly when balanced spills.
The torchao path dropped the quantized copy at the end of the try, so a failure
in save_pretrained left it resident while the original was restored. The del
moved into the finally, ahead of the restore.
Four regression tests added; suites now 25 and 9.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Retry exports whose multi-GPU load silently offloads to CPU, and clear the failed torchao traceback (#7215)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments for PR #7215
* Keep gradients across the export offload and release the failed torchao copy (#7215)
* Tighten comments for PR #7215
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Daniel Han <unslothai@gmail.com>
Two correctness fixes:
- The cache keyed the checkpoint and its companion base by name only, so
a Hub repo advancing to a new commit, or a local directory updated in
place, kept returning embeddings from the previous text encoder. Pair
both with a revision marker: the locally resolved commit sha for a Hub
repo, config plus text-encoder file stats for a directory. Neither
loads the encoders, so a warm run still keeps them off the GPU.
- flow_shift only checked positivity, but JSON accepts 1e309, which
floats to inf, and inf <= 0 is False while NaN fails every comparison.
The sigma table then evaluates s * u / (1 + (s - 1) * u) as NaN, which
poisons every sampled sigma and saves a corrupted adapter while
progress looks normal. Require a finite value.
Six review findings across the Images page and model scanning:
- The quantized (int8/fp8) load path can only attach LoRA adapters
before quantization, but the frontend load request had no loras field,
so every generation after such a load was rejected and each reload
repeated it. Send the selection with the load.
- build_prequant_checkpoint passed no family to the scheme exclusions
while recording the family in metadata, so a Qwen int8 artifact baked
the short-M text-stream linears and was then rejected wholesale by the
loader's family-keyed check.
- Registering a bare single-file checkpoint directory produced no On
Device row even though the images loader can load it; only its parent
worked. Admit that shape when nothing else matched.
- Unload left the Reapply target set, so the repair path was skipped and
Reapply reloaded the ejected model. Clear it, as the video page does.
- Both FLUX.2 bases were trusted for training but not inference, so
Deploy to Create rejected every FLUX.2 adapter.
- Outpaint allocated the grown canvas before downscaling, exceeding the
browser canvas area cap on a large photo; an over-cap canvas is
unusable, so Extend silently posted a fully transparent image and
mask. Scale the source first.
Four bugs in the batched inference path, all found by review:
- A mixed-prompt batch sent a scalar negative prompt against a prompt
list. Z-Image asserts on the length, and Qwen-Image, Krea 2 and FLUX
true-CFG encode a batch-1 negative against batch-N latents and fail in
the transformer's text/image concat. Broadcast it to match the batch.
- The FBCache step-cache reset sat above the chunk loop. diffusers only
resets that state at the end of a successful call, so a forward that
raised (the OOM the backoff is meant to recover) left its own residual
behind and the halved retry died on a shape mismatch. Reset before
every forward instead.
- The conditioning cache keyed on the checkpoint alone, but a GGUF or
single-file load takes its text encoders from the companion base, so
the same checkpoint against a different base reused the previous
base's embeddings. Key the base too.
- Gallery records stored the base seed and the requested batch size even
when a prompts/seeds list drove the run, so restoring the second image
of seeds=[5, 99] replayed seed 5. List-driven outputs now record as
single-image recipes on their own seed.
Also bound strength above 0: every img2img pipeline derives its step
count from it, so 0 leaves zero denoising steps and either raises or, on
SDXL, crashes on empty latents.
Mount-time recovery handled only phase=completed, so reloading the page
after a multi-minute generation failed left an idle view with no
diagnosis: the backend keeps the terminal failed record only until the
next job, and nothing else survives the reload. Surface it the same way
the poll does, filtering the cancelled sentinel.
Four correctness fixes on the training side:
- The labeling grid read caption sidecars under except OSError, but a
non-UTF-8 sidecar raises UnicodeDecodeError (a ValueError), so one bad
file 500d /diffusion/dataset/{name}/images and the grid could not be
opened to repair it. Read it as no caption, matching the info summary.
- An image past Pillow's own hard limit raises DecompressionBombError,
which derives straight from Exception and so escaped the upload guard's
(OSError, UnidentifiedImageError, ValueError) and returned 500 instead
of the intended 400.
- lora_dropout accepted 1.0, which makes PEFT build nn.Dropout(p=1.0):
lora_A and lora_B receive no gradient and the run saves an untrained
adapter while reporting normal progress. Bound it below 1.0, matching
the LLM request schema.
- The train panel re-seeded the base repo on every dataset refresh
because the family object identity changes on each info fetch, so an
upload or caption save silently replaced the user's chosen base and the
run started on a different model. Track the pick and only re-seed on a
real family change.
* feat(studio): add drag and drop sources to create project
Files dropped on the create-project dialog upload to the new project's
sources as soon as it exists, so a project can start with context instead
of needing a second trip to the Sources tab.
The sidebar and projects page dialogs now reuse NewProjectDialog rather
than each keeping their own copy, and the OCR / caption ingest overrides
move to a shared helper so every upload path sends the same settings.
* fix(studio): harden project source drops
Drops are not filtered by the `accept` attribute the way the picker is, so a
folder or an image would stage and then fail server-side with a confusing
per-file error. Unsupported entries are now refused up front with one message.
Cancel bypassed the dialog's reset, so a discarded name and its staged files
came back on reopen and uploaded into the next project created. Every close
path now goes through one handler.
Long filenames lost their extension in _sanitize_filename and were then
rejected as an unsupported type; the stem is trimmed instead. Adds backend
tests for the project scope, the sanitizer and path stripping.
* fix(studio): address second review pass on source drops
A drop landing on the panel while uploads run was not cancelled, because
pointer-events-none took the panel out of hit testing and nothing else on the
page cancels a file drop. The browser would navigate to the file and kill the
uploads in flight. Drag defaults are now cancelled even while disabled, and the
files are ignored instead.
Name, size and mtime can match for two genuinely different files, so a skipped
duplicate now says so rather than disappearing.
A slow upload could resolve after the dialog unmounted and still navigate,
pulling the user off the page they had moved to. Post-upload work is gated on
the component still being mounted.
* fix(studio): make source drops safe under StrictMode replay
The mount sentinel was only cleared in effect cleanup, so StrictMode's
setup/cleanup/setup replay left it false for good and every create in a dev
build stopped short of closing the dialog or navigating. It is now set on
setup as well.
The pending-sources marker was consumed inside a useState initializer, which
React replays, so the discarded pass ate the flag and the project opened on
Chats. Reading is now a peek and the marker is dropped in an effect.
Identical bytes under two names collapse to one document server-side, which
looked like both files had been added. The upload loop now tracks returned
document ids and says when files were merged.
* fix(studio): guard the route and storage around staged uploads
The sidebar's dialog lives in the root layout and never unmounts on a route
change, so the mount check alone could not stop a slow upload from navigating
the user back to the new project. The route is captured when create is pressed
and compared afterwards, and callers get that answer so the sidebar can still
move a chat while leaving the user where they are.
Reading the vision-pass overrides went straight at localStorage, which throws
outright where storage is blocked. That happened before the upload loop, so a
project was created and every staged source was lost. It now falls back to the
backend defaults, matching loadOptionalBool in the chat runtime store.
read_text raises UnicodeDecodeError, which is not an OSError, so one bad caption
sidecar or video sidecar 500d the info, upload and gallery routes. A flat image
now yields the all-black edge map instead of its own luminance, and the four DiT
loss knobs the trainer implements are declared so model_dump keeps them.
str(None) stored the literal "None" as the caption, so a null row counted as
captioned and would have trained on that text. Also drops an unused import.
- start_training forwards resume_source_run_id to _start_training_impl, which
reads it. Without it every start raised NameError.
- Restore main's anchor in the load-marker order test: the file now has an
earlier `if config.is_gguf:`, so indexing the first one compared the wrong
branch.
- The two diffusion tests that reach diffusers now skip when it is absent,
matching the CPU repo-test env.
- The UI smoke finds nav rows that live in the sidebar's More flyout.
* fix(studio): support hostname-based enterprise proxies
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): strip userinfo from proxy fetch targets
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* ROCm/AMD CI coverage: arch-table parity, native-Linux lib prepend, RDNA4 grouped_mm, discovery-based shell suite
Three merged ROCm fixes shipped without tests, and the CI wiring that
would have run them was gated on files the fixes do not touch.
Tests added (113):
tests/studio/install/test_rocm_arch_table_parity.py (27)
diffs the four duplicated gfx -> AMD pip-index tables across
install.sh, install.ps1, studio/setup.ps1 and install_python_stack.py,
plus the GPU-name -> arch tables and the torch 2.11 pin allowlist.
tests/studio/install/test_rocm_native_linux_lib_dirs.py (26)
covers #7233: system-ROCm lib dirs prepended ahead of bundled
libggml-hip, the /dev/kfd + not-WSL + libhsa gate, the opt-out env
var, root resolution order, and source parity between the two copies.
studio/backend/tests/test_grouped_mm_rdna4_fallback.py (46)
covers #7292: registration on the CUDA dispatch key, grouped and
ungrouped numerics, bias/dtype promotion, and the Linux HIP<7.13 +
RDNA4 name gate, executed from the shipped source rather than a copy.
tests/studio/test_ci_shell_suite_coverage.py (14)
fails if either shell runner goes back to a hardcoded list or skips
a file without a recorded reason.
CI wiring:
studio-backend-ci.yml: add install.sh / install.ps1 to the path filter
(the suites it runs assert against those two files, so install-only
changes -- the shape most AMD/ROCm routing fixes take -- skipped it),
and replace the 13-file hardcoded shell list with directory
discovery. That list had fallen seven files behind, including
test_strixhalo_wsl_reroute.sh, the only shell coverage of the ROCm
WSL reroute, which had never run on a PR.
tests/run_all.sh: same discovery loop so local and CI agree.
* Test review fixes: assert on outcomes, not on the code under test
Self-review of the previous commit found four tests that passed for the
wrong reason.
1. The arch-table parity test pinned expected gfx ids copied out of the
shipped tables, which enshrined three upstream inaccuracies as
correct: RX 9070 (non-XT) is gfx1201 not gfx1200, RX 7800 XT is
gfx1101 not gfx1100, and PRO V710 is gfx1101 not gfx1102 per AMD's
ROCm compatibility matrix. The expectation is now the AMD pip index
leaf -- the thing the tables exist to produce, and what a wrong
answer costs the user. The three known drifts are listed explicitly
with a test asserting they stay cosmetic, i.e. that the wrong and
right ids still map to the same wheel index. That test turns red the
day one of them starts routing users to the wrong wheel.
2. The RDNA4 device-name test extracted the regex from worker.py and
then matched with it, so it could not fail. Widening the pattern --
the dangerous edit, since it forces the slow Python mm fallback onto
RDNA3 users -- would have been silently accepted. It now reads the
live pattern and checks it against fixed cases, plus asserts the
name match stays guarded by `not _lin_arch` and that the name is
lowercased before matching.
3. The CI-coverage test matched a verbatim line of studio-backend-ci.yml,
so reindenting the step would fail the build while a real regression
to a hardcoded list could slip past a reformat. It now parses the
YAML, finds the step by name, and asserts on the glob plus the
absence of individual filenames. The path-filter test likewise reads
the parsed trigger instead of scanning raw text.
4. A set comprehension in the parity helper had a ternary whose branches
were identical.
Mutation-tested: widening the RDNA4 regex, desyncing one copy of the
name table, dropping install.sh from the path filter, and re-skipping
the ROCm WSL shell suite each fail at least two tests. Verified on
Linux (WSL Ubuntu 24.04) with CI's torch pin: 86 + 48 pass.
* Fix three wrong gfx ids in the GPU-name arch tables
The name -> gfx tables disagreed with AMD's ROCm compatibility matrix on
three entries. Corrected against the "Radeon GPU" list at
rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html:
RX 9070, RX 9070 GRE gfx1200 -> gfx1201 (Navi 48, same die as the XT)
RX 7800 XT, RX 7700 XT gfx1100 -> gfx1101 (Navi 32, not Navi 31)
PRO W7700 gfx1100 -> gfx1101
PRO V710 gfx1102 -> gfx1101 (Navi 32, not Navi 33)
No wheel changes for anyone: gfx1200/gfx1201 both resolve to gfx120X-all
and gfx1100/gfx1101/gfx1102 all resolve to gfx110X-all, in all four copies
of the index-family map. That collapse is why the errors survived being
copied into six places -- the leaf-level tests could not see them.
It was not purely cosmetic, though. install.sh's second copy feeds
"Tip: set UNSLOTH_ROCM_GFX_ARCH=<arch>", so a 7800 XT user following the
printed advice exported gfx1100 and made a wrong id authoritative for
every later run. It would also have become a real misroute the moment AMD
split a family across index leaves, as they already do for gfx1151/gfx1150.
Fixed in all six places, which is two more than the table's own "kept in
sync with" comments claim exist:
install.sh _infer_amd_gfx_arch_from_gpu_name
install.sh case "$_gpu_disp_mkt" (banner + env tip; undocumented)
studio/setup.sh
install.ps1
studio/setup.ps1
studio/install_python_stack.py
Ordering is preserved: the gfx1102 arm still precedes gfx1101 in the shell
copies so "RX 7700S" cannot fall onto the "RX 7700" glob, and the
PowerShell copies keep the (?!S) lookahead.
Test changes:
- test_rocm_arch_table_parity.py gains _AMD_DOCUMENTED_ARCH, exact gfx
ids transcribed from AMD rather than from the tables. Agreement between
six copies proves nothing when all six were transcribed from the same
mistake, so the ground truth has to come from outside. Verified it
catches the bug: against the pre-fix tables it fails 6 tests.
- The parity check now covers all six copies. It had four; the two
install.sh copies were being treated as one, and
_WIN_GPU_NAME_ARCH_TABLE was not checked at all.
- test_rocm_support.py's TestGfxArchNameFallback pinned two of the wrong
ids as expected values; updated, and extended with a 9060 XT and a
7900 XTX case so each RDNA3/4 die is represented.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard against unregistered copies of the GPU-name arch table
Counting the copies by hand is what let them drift: the in-code "kept in
sync with" comments claimed four, the arch-id fix found six, and scanning
the tree turns up a seventh.
TestNoUnregisteredArchTable rediscovers the copies from the source tree
instead of trusting a hand-maintained list. A table line is one that names
a card and gives its arch; real tables score 9-17 such lines and the only
other hits in the repo are two single-line prose comments, so the
three-line threshold is not load-bearing. A companion test asserts the
scan still finds the known copies, so the heuristic cannot go blind and
pass by finding nothing.
The seventh copy is tests/_zoo_rocm_spoof.py, the fixture other ROCm tests
build their fake AMD host from. It states the mapping backwards (gfx ->
the name torch should report), which makes it an independent witness: it
had gfx1101 -> RX 7800 XT and gfx1201 -> RX 9070 XT right while all six
installer copies were wrong, and nothing compared the two. Now they are
round-tripped against each other.
RX 6700 XT is pinned as a known divergence rather than normalised. AMD's
compatibility matrix documents no consumer RX 6000 card and no gfx1031 at
all, the installer arm is commented "gfx103X family", and gfx1031 appears
only as an index-family key, never as a value a name table emits. With no
external source to correct against, changing shipped behaviour would be
guesswork. A test fails if the divergence ever disappears, so the
exemption cannot go stale.
Also adds the reverse of the AMD-matrix check: a documented card that
matches no arm anywhere is a silent CPU fallback rather than a wrong id.
This cannot detect hardware nobody transcribed, which would need a live
fetch of AMD's matrix and a non-hermetic suite; the docstring says so
rather than implying coverage that is not there.
Verified on Linux: 478 passed, plus all five new guards mutation-tested
to confirm each fails when its invariant is broken.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Docstring said six copies; the list under it now has seven
* tests: run discovered shell tests with bash, not sh
tests/run_all.sh discovered tests/sh/ instead of listing files, but still
invoked each one with sh. Every file there declares a bash shebang, and on
Debian/Ubuntu /bin/sh is dash: test_apt_distro_prompt.sh,
test_studio_home_node_dir.sh and test_with_llama_cpp_dir_link_behavior.sh
fail on bashisms under dash and pass under bash. The old hand-written list
happened to name only dash-clean files, so switching to discovery is what
surfaced it. Backend CI already used bash, so this was a local-only break.
Guarded by a new test asserting both runners invoke tests/sh/ with bash.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Krackan Point (Radeon 860M/840M) routed to the gfx1150 wheel index
The GPU-name tables map 860M/840M and the Ryzen AI 7 350 / AI 5 340 CPU
strings to gfx1150, but Krackan Point is gfx1152. AMD's own lemonade table
(src/cpp/server/system_info.cpp) maps both Krackan iGPUs to gfx1152.
Unlike the three ids already fixed here, this one is not wheel-neutral:
repo.amd.com publishes gfx1150 and gfx1152 as separate index leaves with
separately built torch wheels, so these laptops were installing wheels
built for a different LLVM target. gfx1152 was absent from the codebase
entirely, so it needed the index-family maps, the torch 2.11 floor lists
(same _grouped_mm bug as gfx1150/1151), the Strix reroute set and the
Windows arch allowlist as well as the seven name tables.
The parity test added in this PR did not catch it because its AMD-matrix
expectations stopped at 890M/880M. Added the APU rows, so the case that
actually changes a wheel is now covered: reverting the tables fails 9
tests naming 860M, 840M and Krackan.
gfx1153 (Ryzen AI 5 430 era) is left alone; AMD publishes no gfx1153
wheel family, so there is nothing to route it to.
Verified: bash -n on both shell installers, PowerShell AST parse on both
.ps1 files, python ast.parse on all touched modules, install suite 1334
passed with no new failures against main, shell suite 20 files.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add gfx1152 to unified-memory classifiers, make parity allowlist set-based
Krackan Point (gfx1152, Radeon 860M/840M) is the third RDNA 3.5 APU and
shares one GPU/system-RAM pool exactly like Strix Point (gfx1150) and
Strix Halo (gfx1151), but only the installers knew about it. The two
runtime classifiers still had two-element arch sets, so a Krackan laptop
got the 0.90 discrete headroom factor on a shared pool and ran llama.cpp
without GGML_CUDA_ENABLE_UNIFIED_MEMORY.
- worker.py _rocm_classify_unified_memory: add gfx1152 to the arch set,
and 860m/840m to the device-name fallback. The NVIDIA GeForce 840M
cannot collide there: the function is only reached under _hw.IS_ROCM.
- llama_cpp.py _amd_apu_wants_unified_memory: add gfx1152 to the arch set.
- Tests for both, including the :sramecc-:xnack- suffix form.
TestGfx211AllowlistParity compared four hardcoded allowlist strings, so
adding gfx1152 to all four installers correctly turned three assertions
red without any installer actually disagreeing with another. Each test
now extracts the set its installer holds and compares it to one EXPECTED
constant. Order and spacing are free, membership is not, and the next
leaf is a one-line edit instead of four.
* [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: Daniel Han <danielhanchen@gmail.com>
_load_model_impl contains more than one `if config.is_gguf:`, so
source.index() returned the earlier one, which belongs to a different check
than the branch the assertion is reasoning about. The inheritance call sits at
line 4543, the earlier branch at 4508 and the branch holding the load marker at
4567, so the comparison read 186995 < 185014 and failed on main.
The branch is now located from the load marker itself, which is the landmark
the rest of the test already relies on, so the assertion compares the
inheritance call against the branch that actually guards it. The slice used by
the following assertions is anchored the same way, which also tightens them:
they previously searched from the earlier branch to end of file.
The invariant is unchanged and still has teeth: moving the inheritance call
after the branch makes the assertion fail.
Co-authored-by: danielhanchen <unslothai@gmail.com>
* fix(studio/colab): survive ipykernel OutStream close() during startup
Unsloth Studio crashed at server startup on Colab with:
Unsloth Studio failed to start: 'OutStream' object has no attribute
'watch_fd_thread'
Root cause:
- Colab's ipykernel OutStream is created with watchfd=False, so it never
gains a watch_fd_thread. The OutStream.close() in the affected ipykernel
versions joins that thread unconditionally and raises AttributeError
(ipython/ipykernel#867).
- _setup_server_disk_logging() replaces sys.stdout/sys.stderr with a tee.
That changes the console object identity, so Colab's absl logging handler
(which captured the original OutStream and whose close() deliberately skips
sys.stdout/sys.stderr) no longer treats it as the live console.
- run_server builds uvicorn.Config(...), whose configure_logging runs
logging.config.dictConfig -> logging.shutdown, closing every existing
handler. The absl handler then calls close() on the orphaned OutStream and
the AttributeError propagates out of uvicorn.Config and aborts startup.
Fix:
- Before installing the tee, harden the displaced console streams' close() so
only the ipykernel#867 AttributeError is swallowed; a healthy close() runs
unchanged and any other error still propagates. The buggy close() raises
before it nulls pub_thread, so the stream stays fully usable.
- Give _TeeStream its own close() that flushes the log copy and forwards
close() to the wrapped console stream best-effort, so a handler that
captured the tee cannot crash startup either.
Add regression tests reproducing the exact path (an absl-style handler closing
a watchfd=False OutStream stand-in during logging.shutdown) and asserting the
tee/console path survives and keeps logging.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Show the Colab login password in the shareable link card
* Tighten Colab card comments for PR #7404
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make the Colab tunnel URL clickable and emphasise the password
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Narrow the console close() hardening to the watch_fd_thread AttributeError
* Put the Colab password on its own line so selection excludes the label
* Keep the Colab password as plain selectable text
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: reset quantized KV cache to f16 when flash-attn-off fallback fires
Studio force-enables --flash-attn on for GGUF launches. On a hard startup
or first-decode crash it retries via _with_flash_attn_off, which flipped FA
off but left --cache-type-k/-v untouched. A quantized KV cache (q8_0, q4_0,
q4_1, q5_0, q5_1, iq4_nl) requires flash attention in llama.cpp, so the retry
itself aborted at init with 'V cache quantization requires flash_attn' instead
of recovering.
Reset any quantized --cache-type-k/-v to f16 in the FA-off fallback path so
the retry can actually launch. Non-quantized types (f16, bf16, f32) run fine
without flash attention and are left unchanged. Handles long and short flag
forms and both space and equals syntax, rewriting in place to preserve list
length. Adds pytest coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: FA-off fallback resets only the quantized V cache and drops env-only V cache
Only the V cache requires flash attention in llama.cpp; a quantized K cache
runs fine without it. Restrict the FA-off crash-recovery reset to the V axis
(main and draft) so a memory-constrained config keeps its quantized K cache
instead of risking an OOM on the recovery. Also drop an inherited quantized V
cache set purely through the environment (LLAMA_ARG_CACHE_TYPE_V /
LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_V) at the FA-off retry sites, which the argv
rewrite cannot reach, so the child falls back to the f16 default rather than
aborting.
* Studio: normalize underscore V-cache aliases in the FA-off fallback
llama.cpp rewrites '_' to '-' for any '--' long option before matching,
so a pass-through --cache_type_v q8_0 enables a quantized V cache just
like --cache-type-v. The FA-off crash-recovery reset only matched the
hyphenated spelling, so the underscore alias slipped through and the
retry still aborted with "V cache quantization requires flash_attn".
Canonicalize the flag name the same way before matching (short flags and
the type value are untouched).
* [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>
Adds a "Sidebar navigation" section to Settings -> Appearance, above the
existing profile-menu customizer, with the same drag-to-reorder + switch UI.
- New sidebarNav preference: one { id, pinned } entry per navigable row
(projects, hub, images, train, video, recipes, export), array order = render
order. Defaults match the shipped layout, so an untouched install is unchanged.
- Unpinning moves a row into the More flyout rather than hiding it, so no page
becomes unreachable. New chat and Search stay fixed as actions.
- app-sidebar now renders from one navRows descriptor map, so a pinned row and
its flyout counterpart cannot drift; the More row appears only when something
is unpinned and highlights off whatever it actually holds.
- Mirrored in the backend PersonalizationCustomization: without it the model's
extra="ignore" would drop the field, and because sync replaces local state
with the server's copy once customization is saved, the user's pin order would
reset on the next sync. The validator dedupes and back-fills like sidebarMenu
but preserves the client's order, since here order is meaningful.
Frontend typecheck, i18n parity and catalog checks pass; 32 personalization
tests pass, including a round-trip asserting a reordered list survives a save.
Resolve the drift between PR #6763 and current main:
- deletion: main moved cached-model deletion into hub/services/models/deletion.py,
so the PR's Images/Video in-use guards move there too as _diffusion_blocks_delete
and _video_blocks_delete, keeping main's fail-closed 503 contract.
- llama_keepwarm: take main's rewrite, re-apply the PR's image/video inference
suffixes so a generation in flight blocks an idle unload.
- routes/training: keep main's sidecar-swap 409 and resume_source_run_id, run
start_training in the worker thread the PR's unload hook needs.
- model picker: main renamed components/assistant-ui/model-selector ->
features/model-picker/... and rewrote pickers.tsx, so the PR's picker work is
ported onto main's version (task/catalog props, task gating of hub + cached +
local rows, single-device expanderGpuGb, fine-tuned section hidden when scoped)
rather than reverting main's pinned-models and per-model-config work.
- images/video pages: imports repointed at the new model-selector path.
- tests: delete-guard tests retargeted at the deletion service.
Typecheck, i18n parity and model-catalog checks pass.
The gguf order fix that landed on main dropped the only assertion
covering the prerequisite that llama_extra_args inheritance runs before
the GGUF branch: the inherited value (a carried --no-mmproj) shapes the
hub guard's require_mmproj, so a future reorder could reject a load
over an mmproj download the inherited arguments would disable. The
comment also misattributed the inheritance site to
_guard_chat_load_against_training.
The assertion is restored anchored on the call form
"= _resolve_inherited_extra_args(", which pins the endpoint's call site
(the bare name would match the function definition, which always
precedes the endpoint, making the check vacuous), and the comment now
names the real inheritance site. 32 tests pass.
* Fix reasoning-only Qwen3.6 completions in Studio
* Address reasoning-only review findings
* [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>
* feat(studio): add DoRA support to studio
* fix: added use_dora fast encoder LoraConfig and gated use_dora on AdapterMethod
* fix(studio) serverside normalization for use_dora=true - add note documenting use_dora is silently dropped on diffusion
* fix: dora button disabled on mac, add preflight guard on GGUF lora export, mismatch now correctly falls through to existing error instead of silently no-opping
* Studio: add dora to the WizardState LoRA variant union for consistency
* Reject --use_dora on the MLX (Apple Silicon) CLI path
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
