unsloth/studio/backend/utils/models/model_config.py
Daniel Han 0533efe3f8
Harden model fetching (#6391)
* Harden model fetching: consent gate for trust_remote_code

Add a load-path consent gate that scans a model's auto_map repository code
before it executes and blocks CRITICAL/HIGH findings unless the user pins
approval of that exact code version. Capability detection stays code-free,
reading raw config.json instead of AutoConfig.

- Scan config.json and tokenizer_config.json auto_map, nested local helpers,
  and external owner/name--module repos; fail closed on partial downloads.
- Gate inference, training, and export workers, including the MLX path and a
  LoRA's base model, and report requires_trust_remote_code from the raw config
  so chat and auto-load surface the dialog.
- Verify trusted-org auto-enable against the Hub with the request token and key
  the verdict cache by token; reject local-path and spoofed names.
- Add a consent dialog showing the flagged file, line, and surrounding code.
- Thread hf_token through the scan and load paths for gated repos.

* Address review: token handling, tokenizer/LoRA scan coverage, rollback

- Send the HF token for remote-code scans in the POST body, not the URL, so it
  never lands in a log or browser history.
- Collect tokenizer_config.json auto_map files directly instead of relying only
  on the repo file listing.
- Resolve a LoRA's base model for the validate flag and the scan endpoint so the
  dialog scans the code the workers actually gate.
- Pass the request token to the training YAML trusted-org auto-enable.
- Resend a previously approved fingerprint when rolling back to a custom-code
  model after a failed switch.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Consent UX: drop legacy chat toggle, fix decline copy, purge declined downloads

The per-model consent dialog is now the single approval path for custom
(auto_map) code in chat, so three leftovers from before it existed are removed:

- Remove the "Enable custom code" switch from Chat Settings and stop persisting
  trust_remote_code, so a previously saved blanket-on cannot linger and load a
  model without going through per-version review. The flag stays as an internal
  YAML/preset default (e.g. first-party auto-enable); the load path still gates
  every custom-code load on a fingerprint only the dialog produces.
- Reword the decline message and the auto-load toast to describe approving the
  model's code from the dialog, not a missing settings toggle.
- On decline, purge the repo the scan downloaded so untrusted code is not left
  on disk. A new /api/models/discard-remote-code endpoint deletes only a
  metadata-only cache entry the scan created; it refuses local paths, loaded
  models, and any repo with weight files cached, so a model the user already had
  or pre-downloaded is always left untouched. The frontend only calls it when
  the scan reported created_by_scan.

Adds discard-endpoint tests (delete metadata-only, refuse on weights/gguf,
refuse local, no-op when not cached) and a created_by_scan payload assertion.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Export: remove the user-facing trust remote code toggle

The Export page kept a "Trust remote code" switch (default on) next to the HF
token field. Like chat, custom (auto_map) code should be approved per model
through the load-time review dialog, not a persistent blanket switch, so the
toggle is removed. The export load path already routes through the same consent
dialog: an HF source now starts with trust_remote_code off and only enables it
when the user approves the scanned code in the dialog (a local checkpoint the
user exported stays trusted by default). With the dialog unreachable and no
approval, an HF source loads with trust_remote_code off, which fails closed
rather than running unreviewed code.

* Block loads of repos with unsafe files using Hugging Face's security scan

The trust_remote_code consent gate covers one load-time RCE vector (a repo's
auto_map Python). It does not cover the other: a malicious pickle inside a weight
file (pytorch_model.bin, *.pkl, *.dat) deserializes during from_pretrained even
with trust_remote_code False, so a repo with a normal config plus a poisoned
pickle slips past the existing gate.

Add a metadata-only malware gate that uses Hugging Face's own scan (picklescan +
ClamAV), read via model_info(securityStatus=True).security_repo_status. It never
downloads, opens, or unpickles the flagged files; it only reads the Hub's verdict
and surfaces the flagged file names. New evaluate_file_security runs
unconditionally (independent of trust_remote_code) in every load path (inference,
training SFT/MLX, export), blocking the load when a file is flagged
unsafe/suspicious/malicious. The /remote-code-scan preflight and the validate
endpoint also report the result so the consent dialog opens as a hard block (no
override) listing the flagged files, even for a repo with no custom code.

Policy: hard block with no user override; fail open when the scan is unavailable
(offline/unscanned) so legitimate loads are not broken; no first-party exemption
(a poisoned pickle in a compromised trusted repo still blocks); local paths and
GGUF are skipped (no Hub scan, non-pickle format). Blocking does not gate on
scansDone, since that is often false for clean repos and a file already flagged
unsafe is unsafe regardless.

Adds test_file_security.py covering the block/allow/fail-open/skip matrix.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: scan list-form tokenizer auto_map, gate unsafe files on all load paths

Fixes from a 10-reviewer pass on the model-fetching hardening:

- The remote-code scanner skipped tokenizer auto_map encoded as a [slow, fast]
  list (transformers' standard tokenizer shape, e.g.
  {"AutoTokenizer": ["owner/repo--tokenization_x.Slow", null]}). External
  tokenizer code in that form was never fetched, scanned, or fingerprinted, so an
  AutoTokenizer(trust_remote_code=True) load could run it. _auto_map_refs now
  flattens string, list, and nested values. Adds a regression test.

- Compare-mode chat loads and background auto-load only gated on
  requires_trust_remote_code, so a repo flagged unsafe by the Hub scan but with no
  custom code skipped the hard-block dialog. Both now also gate on
  requires_security_review, matching the main chat path.

- The /remote-code-scan and /validate routes collapsed a LoRA adapter to its base
  before the malware scan, so unsafe files in the adapter repo itself were missed
  in the pre-load review (the workers already scan both). Both routes now run the
  file-security scan over the adapter and the base.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Require approval for all HIGH remote code, fail closed when unscannable

Tighten the load-time security gates based on review:

Consent gate
- HIGH-severity auto_map code now requires explicit, per-version approval for
  every repo, including first-party unsloth/nvidia. The org is no longer a
  blanket bypass: a compromised first-party repo with HIGH code still warrants
  review. CRITICAL stays a hard block; clean code still loads after the consent
  prompt.
- Fail closed when auto_map code is present but cannot be fully fetched or
  listed to scan (gated, offline, transient, or a repo-listing failure that
  could hide an imported helper). We cannot fingerprint code we cannot see, so
  this is a non-approvable block, retryable once the repo is reachable.
- Scan auto_map from every config that can carry one (model, tokenizer, image
  and feature processor, processor, video processor), not just config.json and
  tokenizer_config.json, so a custom-processor model is not missed. The file
  list is the single source of truth in remote_code_scan and is pinned to the
  transformers filename constants by a guard test.
- Distinguish a genuine 404 (config truly absent) from a transient error: only
  the latter forces a scan, so a repo with no config is correctly a no-op.

Malware gate
- Scan a remote repo even when its name ends in .gguf; only local paths skip the
  Hub scan, so a repo cannot dodge the scan by naming itself "*.gguf".
- Correct the docstring: a file already flagged unsafe blocks regardless of
  scansDone; the only fail-open path is an unavailable scan.

Coverage
- Resolve a remote LoRA adapter's base model (not just local directories) so the
  base, where the code and weights actually execute, is scanned in validate,
  the scan route, and the training and export workers.
- Gate the embedding training path (FastSentenceTransformer) with the malware
  and consent checks, matching the other load paths.

Tests updated and added for each change.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Scope malware gate to the load-path vector; stop false-blocking first-party models

Follow-up hardening from a second review pass + a broad live model matrix
(unsloth/* , nvidia/* , third-party, and the eicar malware repo).

Malware / unsafe-file gate
- Scope the block to the actual RCE vector: a root-level file in a code-executing
  format. from_pretrained deserializes weight files at the repo ROOT, so a flag is
  only a load-path pickle vector there. Two exclusions, because neither is loaded:
  inert formats (safetensors is tensor-only, gguf is non-pickle, configs/text/
  images) and files in subdirectories. This keeps eicar blocked (its *.pkl/*.dat/
  eicar_test_file sit at the repo root) while no longer false-blocking legitimate
  first-party repos: nvidia/Nemotron-H-8B-Base-8K ships root safetensors plus NeMo
  pickle checkpoints under nemo/ that the loader never touches, and the Hub flags
  both; the gate previously hard-blocked it.
- Unknown / future non-"safe" levels now fail closed (block) instead of being
  silently allowed, so Hub schema drift cannot introduce a bypass; in-progress
  ("pending"/"scanning"/"error") levels stay non-blocking to avoid false blocks.

Consent gate
- Ignore a STALE own-repo auto_map target that is absent from the repo listing (an
  older config pointing at a file the repo no longer ships) instead of failing the
  whole repo closed as unscannable. The present .py are still fully scanned, which
  is the stronger coverage, and a file that is not there cannot execute. This
  unblocks first-party models like unsloth/PaddleOCR-VL (its tokenizer_config.json
  names processing_ppocrvl.py while the repo ships processing_paddleocr_vl.py). A
  referenced .py that IS present but cannot be fetched, and a repo-listing failure,
  still fail closed.

Remote LoRA base resolution
- Distinguish a genuine 404 (not a LoRA / repo absent -> None) from a transient
  error: the transient case is retried once, then logged as a WARNING (a missed
  base is scanned by neither gate) rather than silently skipped.

Discard endpoint
- Treat .onnx and .ckpt as weights so a repo whose only heavy artifact is one of
  those is never eligible for the declined-download purge.

Tests added for each: load-path scoping (safetensors/subdir/Nemotron-H shapes,
unknown-level fail-closed, pending non-block), stale own-repo auto_map ref, remote
LoRA transient retry, and the empty-config-list (all-404 -> []) semantics.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Make LoRA-base transient-warning test robust to logging backend

Assert on the logger object directly instead of capsys, so the test does not
depend on whether the real structlog logger or the module-stub logger is active
(which varies with test collection order).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Allow a repo with auto_map but no executable code (e.g. GGUF) instead of blocking

A config can declare an auto_map yet the repo ship NO executable .py -- most
commonly a GGUF repo whose config.json carries an auto_map copied from the original
model (e.g. unsloth/Llama-3_1-Nemotron-Ultra-253B-v1-GGUF references
modeling_decilm.py, which the GGUF-only repo does not contain). A GGUF model loads
through llama.cpp, which never executes auto_map, and transformers cannot run a file
that is not present, so there is nothing to scan and trust_remote_code is a no-op.

The fail-closed change treated this empty result the same as "code is present but we
could not fetch it" and hard-blocked the load. Distinguish the two: repo_remote_code_files
now RAISES RemoteCodeUnscannable when code is present but cannot be fully fetched or
listed (offline / gated / transient / a present .py that 404s / a listing failure),
and returns an empty dict only when the listing succeeded and the repo genuinely ships
no executable .py. The consent gate blocks on the exception (fail closed) and allows the
empty case as a no-op. Real unscannable code still hard-blocks; eicar and CRITICAL/HIGH
custom code are unaffected.

Verified against all 37 unsloth/*Nemotron* models (two GGUF repos were false-blocked,
now load) and the existing matrix (eicar still blocks; DeepSeek-OCR / NVLM-D-72B still
prompt approvable consent). Tests updated to expect the raise for unscannable cases and
added for the no-executable-code no-op.

* Ignore vestigial auto_map in GGUF repos (llama.cpp never runs it)

A GGUF repo's config.json is often copied verbatim from the original
transformers model, auto_map and all, but a GGUF load goes through
llama.cpp which never executes auto_map, so the config is inert. Treat
a direct .gguf reference, and a repo that ships .gguf weights with no
.safetensors, as having no remote code so the consent flow is never
triggered. A mixed repo with both .gguf and .safetensors is still gated,
since the safetensors variant would load through transformers where
auto_map does run. The check sits behind the existing auto_map-present
gate so normal models pay no extra repo listing.

* Add scanner-result copy to the remote-code consent dialog

Make the consent dialog state the scan outcome in plain language for
every model. When the static scan finds nothing, reassure the user with
'Our automatic scanner did not flag any worrying files, but please
double check.' (shown only for the clean, approvable case). When the
scan flags custom code or unsafe files, label the list with 'Our
automatic scanner flagged issues including:'. The Hugging Face
attribution for unsafe files stays in the dialog description.

* Close GGUF-suffix consent bypass for repo ids ending in .gguf

The .gguf short-circuit in _config_has_auto_map skipped the scan for any
model name ending in .gguf, including a bare two-segment repo id like
'evil/model.gguf'. Such a repo can still ship safetensors plus auto_map
Python that transformers would execute, so skipping the scan was an
asymmetric bypass (file_security already scans those repos). Restrict the
short-circuit to genuine direct GGUF file references via
_is_direct_gguf_file_ref: a local .gguf path, or a remote repo_id plus
filename (three or more segments). A two-segment repo id named *.gguf now
falls through to the config scan and _is_gguf_repo file inspection, so it
only skips consent when it actually ships .gguf weights and no safetensors.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Align consent dialog body with the title and fix narrow-width overflow

The scan results (the 'Our automatic scanner...' label, finding/unsafe
cards, and the clean-scan reassurance) sat at the dialog's left padding
while the title and description were indented past the status icon, so
the body did not line up under the description. Move the title,
description and results into one column to the right of the icon so they
share a left edge, and let that column fill its width so the description
no longer wraps early.

Also stop a wide code snippet from pushing the dialog off-screen on
narrow viewports: AlertDialogHeader is a grid with place-items-center,
which sized the content row to its content; give the row w-full so it
fills the track, and add min-w-0 down the results chain so the snippet
scrolls inside its card instead of widening the dialog. Verified aligned
and contained from mobile portrait through ultrawide.

* Treat a repo as GGUF-only only when it ships no transformers weights

_is_gguf_repo excluded only .safetensors, so a repo with a .gguf and a
pytorch_model.bin (or .pt/.pth/.h5/.msgpack/.onnx/.ckpt) and no
safetensors was treated as GGUF-only and skipped the consent scan, even
though transformers can load that weight set and execute the repo's
auto_map code. Require the absence of ANY transformers-loadable weight
before treating the repo as a llama.cpp-only GGUF load. A genuine
GGUF-only repo (only .gguf) is still inert; a mixed repo with any pickle
or safetensors weight is gated. Adds a regression test across all the
non-safetensors weight formats.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Block flagged subdir weight shards referenced by a root index

The malware gate treated every subdirectory file as non-loadable, but
from_pretrained deserializes a subdir shard a root index references
(pytorch_model.bin.index.json -> shards/...-00001-of-00002.bin). Read the
root weight indexes and block a flagged subdir pickle the weight_map
points at; a flagged subdir pickle no index lists (NeMo nemo/*.distcp)
stays non-blocking, and an inconclusive index lookup fails closed.

* Pass hf_token to the export checkpoint load

ExportBackend.load_checkpoint scanned with hf_token in the worker but
loaded the weights unauthenticated, so a gated/private checkpoint passed
preflight then 401'd at from_pretrained. Add hf_token to load_checkpoint
and forward token to every from_pretrained branch; the worker passes the
command's hf_token.

* Scope created_by_scan to every HF cache the discard searches

created_by_scan used get_cache_path (active HF_HUB_CACHE only) while
/discard-remote-code deletes across active, legacy, and default caches. A
repo the user already had in a legacy/default cache was marked
scan-created and deleted on decline. Check all three caches for the repo
dir before declaring the scan created it.

* Scan the full .py closure of external auto_map repos

An auto_map cross-repo ref (owner/name--module.Class) only had its entry
file downloaded, but transformers also fetches that file's relative
imports from the same repo, so a dangerous helper.py was left outside the
scanned fingerprint. List each external repo's .py and scan the whole set
(plus the referenced entry files); fail closed if the repo cannot be
listed or fetched.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fail closed when a weight index cannot be fully read

_indexed_shard_paths treated a partial result as definitive: if one weight
index read cleanly but another failed transiently, it returned the shard
paths it did see. A flagged subdirectory pickle listed only by the index we
could not read would then be classed as "not a load input" and skipped,
re-opening the very fail-open this guard was added to close.

Return None whenever any index read is inconclusive, even if another read
cleanly, so the caller blocks the already-flagged subdir pickle. A repo that
ships no index files raises EntryNotFoundError for each (never inconclusive)
and still returns an empty set.

* Match cached repos case-insensitively in the created_by_scan guard

_repo_in_any_hf_cache resolved casing only against the active cache and then
probed every cache with an exact directory name. A case-variant already
present in a legacy or default cache (models--Unsloth--Foo for a scan of
unsloth/foo) was missed, so the repo was marked created_by_scan and deleted
on decline -- but discard_remote_code_download deletes case-insensitively,
so that delete would hit the user's pre-existing cache entry. Detect
case-insensitively too, mirroring the deletion path.

* Skip remote-code and security review for selected GGUF variants

validate_model ran the trust_remote_code and Hugging Face security-scan
preflight against the repo even when the selected artifact is a .gguf. A
GGUF loads through llama.cpp, which never executes the repo's auto_map
Python and never deserializes root pickle weights, so repo-level Transformers
artifacts (a config.json with auto_map, or an unsafe pytorch_model.bin next
to the .gguf in a mixed repo) are inert for that load. Gating the GGUF on
them is a false positive. Run both preflights only for non-GGUF loads.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Scope the malware gate to actual load roots and serialized files

Two fixes to evaluate_file_security so it neither misses a load-path pickle nor
false-blocks an inert file:

- Honor subdirectory load roots. Spark-TTS / BiCodec call from_pretrained on the
  snapshot's LLM subdirectory, so a flagged pickle directly under it is a
  root-level load artifact there. A new load_subdirs parameter (set from the
  model's audio type via security_load_subdirs) reclassifies those files relative
  to the load root and looks for weight indexes under it, so a flagged shard in
  that subdir is no longer skipped as "not root-level".
- Exempt source files. A root .py is never deserialized by from_pretrained;
  executable repo code runs only through auto_map, which the remote-code consent
  gate scans. Flagging a Python helper here would false-block a repo that merely
  ships a build or train script.

* Scan a LoRA adapter and base as one consent unit, and gate MEDIUM code

A LoRA load runs both the adapter's and the base's repo code. The consent gate
scanned them separately and pinned one fingerprint per repo, so an adapter that
shipped its own auto_map code was either never shown in the dialog (which only
saw the base) or impossible to approve with the base's fingerprint.

evaluate_remote_code_consent_for_targets now scans all of a load's repos as a
single combined unit and pins ONE fingerprint over the union of their code, so
approving the load approves every repo's code together. evaluate_remote_code_consent
becomes a thin single-target wrapper, and an unscannable target fails the whole
load closed.

Also gate MEDIUM findings: like HIGH they now block pending pinned approval, so a
direct API caller cannot run flagged code by setting trust_remote_code=True
without consenting. Only a clean scan loads without a fingerprint.

* Preflight a LoRA load's adapter and base as one combined consent scan

scan_model_remote_code rewrote a LoRA adapter to its base and scanned only the
base for remote code, so the dialog never surfaced an adapter's own auto_map
code. Scan the adapter and base together through
preflight_remote_code_consent_for_targets, which pins one combined fingerprint
the worker gate accepts. The malware preflight is also scoped to each target's
load subdirectories.

* Apply combined consent and subdir-aware malware scan in load workers

Each load worker (inference, export, training) evaluated remote-code consent
once per target with a single shared fingerprint, so a LoRA adapter that ships
its own auto_map code could not be approved by the base's fingerprint. They now
scan the adapter and base together via evaluate_remote_code_consent_for_targets,
which pins one combined fingerprint over the union of their code. The malware
scan in each worker is also scoped to the model's load subdirectories so a
flagged pickle under a from_pretrained load subdir is not missed.

* Report a consistent trust_remote_code requirement after a model loads

validate_model reports requires_trust_remote_code from the YAML default OR the
raw auto_map, but the load, already-loaded, and status responses reported only
the YAML default. A custom-code model approved and loaded via auto_map was then
reported as not requiring trust_remote_code, so the frontend stored false and a
later retry or rollback sent trust_remote_code=false and failed.

A shared resolver reports the same requirement for a loaded model (a value
stored at load time, else the trust_remote_code the load used, else the YAML
default, else the raw auto_map check), and the load response persists it so the
status and already-loaded paths stay consistent. The selected-GGUF security
review is also scoped to the model's load subdirectories.

* Run the consent gate on training resume and for YAML-only trust_remote_code

Three frontend gaps left a model loading without the trust_remote_code it needs:

- The shared consent helper returned early when the scan found no auto_map and no
  unsafe files, dropping a requirement that comes from a model's Studio YAML
  default (e.g. GLM-4.7-Flash). It now grants the caller's requirement with an
  empty pin instead of sending trust_remote_code=false.
- Resume-from-history called startTraining directly with no consent gate, so a
  resumed run whose model needs custom code (or an old run with no approved
  fingerprint) hit the worker block with no dialog. It now runs the same gate as
  a fresh start.
- HF export passed requiresTrustRemoteCode=false for every HF source, so a
  YAML-only model could not flip the flag before export. It now signals the
  requirement for HF sources.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Cover both LoRA repos in validate, report GGUF as inert, purge all declined repos

Three follow-on gaps from the combined adapter+base consent work:

- validate_model resolved requires_trust_remote_code from the base alone, so a
  LoRA adapter that ships its OWN auto_map code (with a plain base) was reported
  as not needing trust_remote_code and the consent dialog never opened. It now
  checks the [adapter, base] target set, matching the scan route and the workers
  (which already gate both) and the security review already running over both.

- The already-loaded, loaded, and status responses for a selected GGUF reported
  requires_trust_remote_code from the model's YAML default. A GGUF loads through
  llama.cpp, which never executes the repo's auto_map Python, so the requirement
  is inert for that load. They now report False, matching validate_model (which
  already skips both gates for GGUF) so a status refresh cannot flip the flag
  back on.

- The remote-code scan downloads both the adapter's and the base's config, but
  created_by_scan tracked only the primary, so a base the scan was first to pull
  into the cache was left on disk when the user declined. The scan now reports
  scan_created_repos (every repo it newly cached) and the decline cleanup purges
  each; created_by_scan stays for older clients. The frontend falls back to the
  primary flag when the list is absent.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Scan the repo the load fetches, purge external code on decline, harden consent pins

Six follow-on hardening fixes from a fresh review pass over the gate:

- The malware gate scanned the literal "Spark-TTS-0.5B/LLM" alias, but the trainer
  downloads it as unsloth/Spark-TTS-0.5B and loads LLM/, so the alias 404'd and
  failed open, missing a flagged LLM/ pickle. evaluate_file_security now resolves
  the alias to the repo the loader fetches and scans LLM/ as a load root.

- security_load_subdirs relied only on tokenizer detection, which fails on an
  unresolved alias or offline; it now also honors the Studio YAML audio_type
  default, so a BiCodec LLM/ load root is not missed.

- The remote-code scan downloads external auto_map repos (owner/name--module.Class),
  but the decline cleanup tracked only the model/adapter/base, leaving the external
  untrusted code cached. The scan now enumerates external auto_map repos and reports
  the ones it created in scan_created_repos, so a decline purges them too.

- External auto_map refs failed the whole load closed on a stale or mis-derived
  dotted ref (sub.mod.py vs the real sub/mod.py) even though the actual file was
  present and scanned. They now drop such refs when the repo listing is real, exactly
  like the own-repo path; an empty/incomplete listing still fetches and fails closed.

- The combined consent fingerprint keyed code by the raw target string, so the scan
  endpoint's canonicalized casing and a worker's raw user input produced different
  pins for identical code, rejecting a valid approval. Hub repo ids are now folded to
  lowercase in the key (local paths stay case-sensitive), so the pin tracks the code.

- Export threaded hf_token into the weight load but not into detect_audio_type /
  is_vision_model, so a gated multimodal base 404'd in detection and fell through to
  the text loader. Both probes now use the same token.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Thread the token through check-vision and guard the gate's parallel sites

The /check-vision endpoint classified a model without the hf_token, so a gated or
private vision model 404'd in the probe and was reported as a plain text model --
the same dropped-token shape as the export probes, at a sibling site. It now passes
the token like the neighboring /check-embedding endpoint.

Add deterministic consistency guards (tests/test_security_gate_consistency.py) that
enumerate the gate's parallel sites mechanically instead of relying on a review to
spot a missed sibling: every is_vision_model / is_embedding_model / detect_audio_type
caller under routes/ and core/ must thread the token, every GGUF response must report
trust_remote_code via the resolver or False (never the raw YAML default), and every
load worker that runs the malware or consent gate must resolve the LoRA base. A new
site that drops the token or mis-reports the requirement now fails CI directly.

* Narrow the LLM alias rewrite and make audio detection token-aware

Three fixes from the confirmatory review, one a regression from the previous round:

- _load_scan_target rewrote EVERY remote repo ending in "/LLM" to unsloth/<parent>,
  so a real third-party repo named "<owner>/LLM" was scanned as unsloth/<owner>
  while the loader still fetched the real repo -- a fail-open hole introduced when
  the Spark-TTS alias handling was added. It now rewrites only a registry-known
  bicodec alias; every other "/LLM" repo is scanned as itself.

- detect_audio_type cached results under the bare model name, so an unauthenticated
  probe of a gated/private repo cached None and poisoned a later authenticated call
  with the token. The cache is now keyed by (normalized_name, token_fingerprint),
  matching the vision cache.

- The training fallback /check-vision call dropped the hf_token, misclassifying a
  gated/private VLM when the config endpoint failed. It now passes the token, like
  the getModelConfig call it falls back from; checkEmbeddingModel takes the token too.

Extend the consistency guards: every capability cache must be keyed by a tuple
including the token, so a cache re-declared as Dict[str, ...] fails CI.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Document the broad .py scan as deliberate and enforce it with a test

The remote-code scanner scans every .py in a repo once an auto_map exists, not
just the auto_map entry's static import closure. This is intentional: the entry
module can reach a sibling via an absolute import, importlib, or exec, none of
which a static relative-import closure follows, so closure-only scanning would be
a real bypass of a load-time RCE gate. The broad scan never under-scans; the cost
is that an unrelated benign script can over-block, which is the safe failure
direction (HIGH stays approvable; only CRITICAL hard-blocks).

Spell this out at both the local and remote scan sites so the choice reads as
deliberate, and add a test asserting an unrelated, never-imported .py is still
scanned -- so a future narrowing to the static closure fails CI.

* Purge a declined remote LoRA adapter the scan downloaded

scan_model_remote_code probed the created-by-scan state AFTER resolving the base,
but get_base_model_from_lora_identifier downloads a remote adapter's own
adapter_config.json, so the adapter looked already-cached and was dropped from
scan_created_repos. On decline the adapter -- including the auto_map .py the
preflight fetched -- was left on disk, defeating the "untrusted code is not left
on disk" guarantee for the adapter itself.

Snapshot the primary's cache state BEFORE base resolution and use it when marking
the adapter scan-created; on any probe error treat it as pre-existing so a decline
never deletes it. The base and external repos are unaffected (their configs are not
downloaded before their own probe). Add a test that models the mid-scan download
side effect, which the prior static-stub tests did not.

* Clear remote-code approval when the training model changes

Switching the training model from an approved custom-code model to a clean one
kept the previous model's trust_remote_code=true and approved fingerprint in the
store: setSelectedModel reset visionImageSize on a true switch but not the
remote-code approval. The clean model then trained with trust_remote_code=true,
which bypasses the compiler and disables fused cross-entropy.

Reset trustRemoteCode and approvedRemoteCodeFingerprint on a true model switch.
The new model's own YAML default is re-applied by loadAndApplyModelDefaults, and a
custom-code model still re-opens the consent dialog before training starts, so the
only change is that a clean model no longer inherits a stale approval.

* Trim verbose comments across the model-fetching hardening changes

Condense the explanatory comments and docstrings introduced across the
trust_remote_code consent gate, the malware/unsafe-file gate, the remote-code
scanner, the load workers, the model routes, and the security frontend into
fewer, tighter lines while preserving every security rationale (fail-open vs
fail-closed direction, the deliberate broad-scan anti-bypass note, the
empty-vs-unscannable distinction, stale-ref handling, and the alias-rewrite
spoof guard).

Comments and docstrings only. No code, logic, identifiers, or test behaviour
changed; verified comment-only via the AST/TypeScript checker (40/40), with the
backend test suite and frontend tsc green.

* Do not cache transient audio-detection failures

detect_audio_type cached _detect_audio_from_tokenizer's result
unconditionally, so a transient read failure (network error or 5xx,
returned as None) poisoned the cache and the later successful probe never
ran. Mirror the vision cache: _detect_audio_from_tokenizer now returns
(audio_type, definitive) and the caller caches only definitive results.

A read that succeeds with no audio tokens, or clean 404s for every
tokenizer path, stays a cacheable None; only a genuine transient failure
(connection error, timeout, 5xx, malformed body) skips the cache so the
next call retries.

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
2026-06-18 05:39:52 -07:00

2816 lines
105 KiB
Python

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""Model and LoRA configuration handling."""
from dataclasses import dataclass
from typing import Optional, Dict, Any
from utils.paths import (
normalize_path,
is_local_path,
is_model_cached,
get_cache_path,
resolve_cached_repo_id_case,
outputs_root,
exports_root,
resolve_output_dir,
resolve_export_dir,
)
from utils.utils import without_hf_auth
from utils.models.gguf_metadata import (
is_mmproj_by_metadata,
pairing_score,
read_gguf_general_metadata,
)
import structlog
from loggers import get_logger
import os
import re
import subprocess
import sys
from pathlib import Path
from typing import List, Tuple
import hashlib
import json
import threading
import yaml
from utils.native_path_leases import child_env_without_native_path_secret
from utils.subprocess_compat import (
windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
)
logger = get_logger(__name__)
def _env_offline() -> bool:
"""True if HF_HUB_OFFLINE or TRANSFORMERS_OFFLINE is set to a truthy value."""
return os.environ.get("HF_HUB_OFFLINE", "").lower() in (
"1",
"true",
"yes",
) or os.environ.get("TRANSFORMERS_OFFLINE", "").lower() in ("1", "true", "yes")
# ── Model size extraction ────────────────────────────────────
import re as _re
_MODEL_SIZE_RE = _re.compile(r"(?:^|[-_/])(\d+\.?\d*)\s*([bm])(?:$|[-_/])", _re.IGNORECASE)
# MoE active-parameter pattern: "A3B", "A3.5B", etc.
_ACTIVE_SIZE_RE = _re.compile(r"(?:^|[-_/])a(\d+\.?\d*)\s*([bm])(?:$|[-_/])", _re.IGNORECASE)
# Gemma 3n/4 effective-parameter pattern: "E2B", "E4B" -- the runtime
# footprint (MatFormer + per-layer embeddings), which is the size that
# matters for size-gated policies like sub-3B speculative-decoding fallback.
_EFFECTIVE_SIZE_RE = _re.compile(r"(?:^|[-_/])e(\d+\.?\d*)\s*([bm])(?:$|[-_/])", _re.IGNORECASE)
def extract_model_size_b(model_id: str) -> float | None:
"""Extract model size in billions from a model identifier.
Prefers MoE active-parameter notation (e.g. ``A3B`` in
``Qwen3.5-35B-A3B``), then Gemma effective-parameter notation
(e.g. ``E2B``), over total params. Handles ``B`` (billions) and
``M`` (millions) suffixes.
"""
mid = (model_id or "").lower()
# First match wins, in priority order: active > effective > total.
for pattern in (_ACTIVE_SIZE_RE, _EFFECTIVE_SIZE_RE, _MODEL_SIZE_RE):
m = pattern.search(mid)
if m:
val = float(m.group(1))
return val / 1000.0 if m.group(2).lower() == "m" else val
return None
# Maps equivalent model names to their canonical YAML config file.
# Format: "canonical_model_name.yaml": [equivalent model names].
# Canonical filename derives from the first model name in each list.
MODEL_NAME_MAPPING = {
# ── Embedding models ──
"unsloth_all-MiniLM-L6-v2.yaml": [
"unsloth/all-MiniLM-L6-v2",
"sentence-transformers/all-MiniLM-L6-v2",
],
"unsloth_bge-m3.yaml": [
"unsloth/bge-m3",
"BAAI/bge-m3",
],
"unsloth_embeddinggemma-300m.yaml": [
"unsloth/embeddinggemma-300m",
"google/embeddinggemma-300m",
],
"unsloth_gte-modernbert-base.yaml": [
"unsloth/gte-modernbert-base",
"Alibaba-NLP/gte-modernbert-base",
],
"unsloth_Qwen3-Embedding-0.6B.yaml": [
"unsloth/Qwen3-Embedding-0.6B",
"Qwen/Qwen3-Embedding-0.6B",
"unsloth/Qwen3-Embedding-4B",
"Qwen/Qwen3-Embedding-4B",
],
# ── Other models ──
"unsloth_answerdotai_ModernBERT-large.yaml": [
"answerdotai/ModernBERT-large",
],
"unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml": [
"unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit",
"unsloth/Qwen2.5-Coder-7B-Instruct",
"Qwen/Qwen2.5-Coder-7B-Instruct",
],
"unsloth_codegemma-7b-bnb-4bit.yaml": [
"unsloth/codegemma-7b-bnb-4bit",
"unsloth/codegemma-7b",
"google/codegemma-7b",
],
"unsloth_ERNIE-4.5-21B-A3B-PT.yaml": [
"unsloth/ERNIE-4.5-21B-A3B-PT",
],
"unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml": [
"unsloth/ERNIE-4.5-VL-28B-A3B-PT",
],
"tiiuae_Falcon-H1-0.5B-Instruct.yaml": [
"tiiuae/Falcon-H1-0.5B-Instruct",
"unsloth/Falcon-H1-0.5B-Instruct",
],
"unsloth_functiongemma-270m-it.yaml": [
"unsloth/functiongemma-270m-it-unsloth-bnb-4bit",
"google/functiongemma-270m-it",
"unsloth/functiongemma-270m-it-unsloth-bnb-4bit",
],
"unsloth_gemma-2-2b.yaml": [
"unsloth/gemma-2-2b-bnb-4bit",
"google/gemma-2-2b",
],
"unsloth_gemma-2-27b-bnb-4bit.yaml": [
"unsloth/gemma-2-9b-bnb-4bit",
"unsloth/gemma-2-9b",
"google/gemma-2-9b",
"unsloth/gemma-2-27b",
"google/gemma-2-27b",
],
"unsloth_gemma-3-4b-pt.yaml": [
"unsloth/gemma-3-4b-pt-unsloth-bnb-4bit",
"google/gemma-3-4b-pt",
"unsloth/gemma-3-4b-pt-bnb-4bit",
],
"unsloth_gemma-3-4b-it.yaml": [
"unsloth/gemma-3-4b-it-unsloth-bnb-4bit",
"google/gemma-3-4b-it",
"unsloth/gemma-3-4b-it-bnb-4bit",
],
"unsloth_gemma-3-27b-it.yaml": [
"unsloth/gemma-3-27b-it-unsloth-bnb-4bit",
"google/gemma-3-27b-it",
"unsloth/gemma-3-27b-it-bnb-4bit",
],
"unsloth_gemma-3-270m-it.yaml": [
"unsloth/gemma-3-270m-it-unsloth-bnb-4bit",
"google/gemma-3-270m-it",
"unsloth/gemma-3-270m-it-bnb-4bit",
],
"unsloth_gemma-3n-E4B-it.yaml": [
"unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit",
"google/gemma-3n-E4B-it",
"unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit",
],
"unsloth_gemma-3n-E4B.yaml": [
"unsloth/gemma-3n-E4B-unsloth-bnb-4bit",
"google/gemma-3n-E4B",
],
"unsloth_gemma-4-31B-it.yaml": [
"unsloth/gemma-4-31B-it",
"google/gemma-4-31B-it",
],
"unsloth_gemma-4-26B-A4B-it.yaml": [
"unsloth/gemma-4-26B-A4B-it",
"google/gemma-4-26B-A4B-it",
],
"unsloth_gemma-4-E2B-it.yaml": [
"unsloth/gemma-4-E2B-it",
"google/gemma-4-E2B-it",
],
"unsloth_gemma-4-E4B-it.yaml": [
"unsloth/gemma-4-E4B-it",
"google/gemma-4-E4B-it",
],
"unsloth_gemma-4-31B.yaml": [
"unsloth/gemma-4-31B",
"google/gemma-4-31B",
],
"unsloth_gemma-4-26B-A4B.yaml": [
"unsloth/gemma-4-26B-A4B",
"google/gemma-4-26B-A4B",
],
"unsloth_gemma-4-E2B.yaml": [
"unsloth/gemma-4-E2B",
"google/gemma-4-E2B",
],
"unsloth_gemma-4-E4B.yaml": [
"unsloth/gemma-4-E4B",
"google/gemma-4-E4B",
],
"unsloth_gpt-oss-20b.yaml": [
"openai/gpt-oss-20b",
"unsloth/gpt-oss-20b-unsloth-bnb-4bit",
"unsloth/gpt-oss-20b-BF16",
],
"unsloth_gpt-oss-120b.yaml": [
"openai/gpt-oss-120b",
"unsloth/gpt-oss-120b-unsloth-bnb-4bit",
],
"unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml": [
"unsloth/granite-4.0-350m",
"ibm-granite/granite-4.0-350m",
"unsloth/granite-4.0-350m-bnb-4bit",
],
"unsloth_granite-4.0-h-micro.yaml": [
"ibm-granite/granite-4.0-h-micro",
"unsloth/granite-4.0-h-micro-bnb-4bit",
"unsloth/granite-4.0-h-micro-unsloth-bnb-4bit",
],
"unsloth_LFM2-1.2B.yaml": [
"unsloth/LFM2-1.2B",
],
"unsloth_llama-3-8b-bnb-4bit.yaml": [
"unsloth/llama-3-8b",
"meta-llama/Meta-Llama-3-8B",
],
"unsloth_llama-3-8b-Instruct-bnb-4bit.yaml": [
"unsloth/llama-3-8b-Instruct",
"meta-llama/Meta-Llama-3-8B-Instruct",
],
"unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml": [
"unsloth/Meta-Llama-3.1-8B-bnb-4bit",
"unsloth/Meta-Llama-3.1-8B-unsloth-bnb-4bit",
"meta-llama/Meta-Llama-3.1-8B",
"unsloth/Meta-Llama-3.1-70B-bnb-4bit",
"unsloth/Meta-Llama-3.1-8B",
"unsloth/Meta-Llama-3.1-70B",
"meta-llama/Meta-Llama-3.1-70B",
"unsloth/Meta-Llama-3.1-405B-bnb-4bit",
"meta-llama/Meta-Llama-3.1-405B",
],
"unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml": [
"unsloth/Meta-Llama-3.1-8B-Instruct-unsloth-bnb-4bit",
"unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit",
"meta-llama/Meta-Llama-3.1-8B-Instruct",
"unsloth/Meta-Llama-3.1-8B-Instruct",
"RedHatAI/Llama-3.1-8B-Instruct-FP8",
"unsloth/Llama-3.1-8B-Instruct-FP8-Block",
"unsloth/Llama-3.1-8B-Instruct-FP8-Dynamic",
],
"unsloth_Llama-3.2-3B-Instruct.yaml": [
"unsloth/Llama-3.2-3B-Instruct-unsloth-bnb-4bit",
"meta-llama/Llama-3.2-3B-Instruct",
"unsloth/Llama-3.2-3B-Instruct-bnb-4bit",
"RedHatAI/Llama-3.2-3B-Instruct-FP8",
"unsloth/Llama-3.2-3B-Instruct-FP8-Block",
"unsloth/Llama-3.2-3B-Instruct-FP8-Dynamic",
],
"unsloth_Llama-3.2-1B-Instruct.yaml": [
"unsloth/Llama-3.2-1B-Instruct-unsloth-bnb-4bit",
"meta-llama/Llama-3.2-1B-Instruct",
"unsloth/Llama-3.2-1B-Instruct-bnb-4bit",
"RedHatAI/Llama-3.2-1B-Instruct-FP8",
"unsloth/Llama-3.2-1B-Instruct-FP8-Block",
"unsloth/Llama-3.2-1B-Instruct-FP8-Dynamic",
],
"unsloth_Llama-3.2-11B-Vision-Instruct.yaml": [
"unsloth/Llama-3.2-11B-Vision-Instruct-unsloth-bnb-4bit",
"meta-llama/Llama-3.2-11B-Vision-Instruct",
"unsloth/Llama-3.2-11B-Vision-Instruct-bnb-4bit",
],
"unsloth_Llama-3.3-70B-Instruct.yaml": [
"unsloth/Llama-3.3-70B-Instruct-unsloth-bnb-4bit",
"meta-llama/Llama-3.3-70B-Instruct",
"unsloth/Llama-3.3-70B-Instruct-bnb-4bit",
"RedHatAI/Llama-3.3-70B-Instruct-FP8",
"unsloth/Llama-3.3-70B-Instruct-FP8-Block",
"unsloth/Llama-3.3-70B-Instruct-FP8-Dynamic",
],
"unsloth_Llasa-3B.yaml": [
"HKUSTAudio/Llasa-1B",
"unsloth/Llasa-3B",
],
"unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml": [
"unsloth/Magistral-Small-2509",
"mistralai/Magistral-Small-2509",
"unsloth/Magistral-Small-2509-bnb-4bit",
],
"unsloth_Ministral-3-3B-Instruct-2512.yaml": [
"unsloth/Ministral-3-3B-Instruct-2512",
],
"unsloth_mistral-7b-v0.3-bnb-4bit.yaml": [
"unsloth/mistral-7b-v0.3-bnb-4bit",
"unsloth/mistral-7b-v0.3",
"mistralai/Mistral-7B-v0.3",
],
"unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml": [
"unsloth/Mistral-Nemo-Base-2407-bnb-4bit",
"unsloth/Mistral-Nemo-Base-2407",
"mistralai/Mistral-Nemo-Base-2407",
"unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit",
"unsloth/Mistral-Nemo-Instruct-2407",
"mistralai/Mistral-Nemo-Instruct-2407",
],
"unsloth_Mistral-Small-Instruct-2409.yaml": [
"unsloth/Mistral-Small-Instruct-2409-bnb-4bit",
"mistralai/Mistral-Small-Instruct-2409",
],
"unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml": [
"unsloth/mistral-7b-instruct-v0.3-bnb-4bit",
"unsloth/mistral-7b-instruct-v0.3",
"mistralai/Mistral-7B-Instruct-v0.3",
],
"unsloth_Qwen2.5-1.5B-Instruct.yaml": [
"unsloth/Qwen2.5-1.5B-Instruct-unsloth-bnb-4bit",
"Qwen/Qwen2.5-1.5B-Instruct",
"unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit",
],
"unsloth_Nemotron-3-Nano-30B-A3B.yaml": [
"unsloth/Nemotron-3-Nano-30B-A3B",
],
"unsloth_orpheus-3b-0.1-ft.yaml": [
"unsloth/orpheus-3b-0.1-ft",
"unsloth/orpheus-3b-0.1-ft-unsloth-bnb-4bit",
"canopylabs/orpheus-3b-0.1-ft",
"unsloth/orpheus-3b-0.1-ft-bnb-4bit",
],
"OuteAI_Llama-OuteTTS-1.0-1B.yaml": [
"OuteAI/Llama-OuteTTS-1.0-1B",
"unsloth/Llama-OuteTTS-1.0-1B",
"unsloth/llama-outetts-1.0-1b",
"OuteAI/OuteTTS-1.0-0.6B",
"unsloth/OuteTTS-1.0-0.6B",
"unsloth/outetts-1.0-0.6b",
],
"unsloth_PaddleOCR-VL.yaml": [
"unsloth/PaddleOCR-VL",
],
"unsloth_Phi-3-medium-4k-instruct.yaml": [
"unsloth/Phi-3-medium-4k-instruct-bnb-4bit",
"microsoft/Phi-3-medium-4k-instruct",
],
"unsloth_Phi-3.5-mini-instruct.yaml": [
"unsloth/Phi-3.5-mini-instruct-bnb-4bit",
"microsoft/Phi-3.5-mini-instruct",
],
"unsloth_Phi-4.yaml": [
"unsloth/phi-4-unsloth-bnb-4bit",
"microsoft/phi-4",
"unsloth/phi-4-bnb-4bit",
],
"unsloth_Pixtral-12B-2409.yaml": [
"unsloth/Pixtral-12B-2409-unsloth-bnb-4bit",
"mistralai/Pixtral-12B-2409",
"unsloth/Pixtral-12B-2409-bnb-4bit",
],
"unsloth_Qwen2-7B.yaml": [
"unsloth/Qwen2-7B-bnb-4bit",
"Qwen/Qwen2-7B",
],
"unsloth_Qwen2-VL-7B-Instruct.yaml": [
"unsloth/Qwen2-VL-7B-Instruct-unsloth-bnb-4bit",
"Qwen/Qwen2-VL-7B-Instruct",
"unsloth/Qwen2-VL-7B-Instruct-bnb-4bit",
],
"unsloth_Qwen2.5-7B.yaml": [
"unsloth/Qwen2.5-7B-unsloth-bnb-4bit",
"Qwen/Qwen2.5-7B",
"unsloth/Qwen2.5-7B-bnb-4bit",
],
"unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml": [
"unsloth/Qwen2.5-Coder-1.5B-Instruct-bnb-4bit",
"Qwen/Qwen2.5-Coder-1.5B-Instruct",
],
"unsloth_Qwen2.5-Coder-14B-Instruct.yaml": [
"unsloth/Qwen2.5-Coder-14B-Instruct-bnb-4bit",
"Qwen/Qwen2.5-Coder-14B-Instruct",
],
"unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml": [
"unsloth/Qwen2.5-VL-7B-Instruct",
"Qwen/Qwen2.5-VL-7B-Instruct",
"unsloth/Qwen2.5-VL-7B-Instruct-unsloth-bnb-4bit",
],
"unsloth_Qwen3-0.6B.yaml": [
"unsloth/Qwen3-0.6B-unsloth-bnb-4bit",
"Qwen/Qwen3-0.6B",
"unsloth/Qwen3-0.6B-bnb-4bit",
"Qwen/Qwen3-0.6B-FP8",
"unsloth/Qwen3-0.6B-FP8",
],
"unsloth_Qwen3-4B-Instruct-2507.yaml": [
"unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit",
"Qwen/Qwen3-4B-Instruct-2507",
"unsloth/Qwen3-4B-Instruct-2507-bnb-4bit",
"Qwen/Qwen3-4B-Instruct-2507-FP8",
"unsloth/Qwen3-4B-Instruct-2507-FP8",
],
"unsloth_Qwen3-4B-Thinking-2507.yaml": [
"unsloth/Qwen3-4B-Thinking-2507-unsloth-bnb-4bit",
"Qwen/Qwen3-4B-Thinking-2507",
"unsloth/Qwen3-4B-Thinking-2507-bnb-4bit",
"Qwen/Qwen3-4B-Thinking-2507-FP8",
"unsloth/Qwen3-4B-Thinking-2507-FP8",
],
"unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml": [
"unsloth/Qwen3-14B-Base",
"Qwen/Qwen3-14B-Base",
"unsloth/Qwen3-14B-Base-bnb-4bit",
],
"unsloth_Qwen3-14B.yaml": [
"unsloth/Qwen3-14B-unsloth-bnb-4bit",
"Qwen/Qwen3-14B",
"unsloth/Qwen3-14B-bnb-4bit",
"Qwen/Qwen3-14B-FP8",
"unsloth/Qwen3-14B-FP8",
],
"unsloth_Qwen3-32B.yaml": [
"unsloth/Qwen3-32B-unsloth-bnb-4bit",
"Qwen/Qwen3-32B",
"unsloth/Qwen3-32B-bnb-4bit",
"Qwen/Qwen3-32B-FP8",
"unsloth/Qwen3-32B-FP8",
],
"unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml": [
"Qwen/Qwen3-VL-8B-Instruct-FP8",
"unsloth/Qwen3-VL-8B-Instruct-FP8",
"unsloth/Qwen3-VL-8B-Instruct",
"Qwen/Qwen3-VL-8B-Instruct",
"unsloth/Qwen3-VL-8B-Instruct-bnb-4bit",
],
"sesame_csm-1b.yaml": [
"sesame/csm-1b",
"unsloth/csm-1b",
],
"Spark-TTS-0.5B_LLM.yaml": [
"Spark-TTS-0.5B/LLM",
"unsloth/Spark-TTS-0.5B",
],
"unsloth_tinyllama-bnb-4bit.yaml": [
"unsloth/tinyllama",
"TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
],
"unsloth_whisper-large-v3.yaml": [
"unsloth/whisper-large-v3",
"openai/whisper-large-v3",
],
}
# Reverse lookup: model_name -> canonical_filename
_REVERSE_MODEL_MAPPING = {}
for canonical_file, model_names in MODEL_NAME_MAPPING.items():
for model_name in model_names:
_REVERSE_MODEL_MAPPING[model_name.lower()] = canonical_file
def load_model_config(
model_name: str,
use_auth: bool = False,
token: Optional[str] = None,
trust_remote_code: bool = False,
):
"""Load model config with optional authentication control.
``trust_remote_code`` defaults to ``False``: capability detection and
metadata lookups must never execute a model repo's ``auto_map`` Python.
Deliberate remote-code loads pass the flag explicitly through
``FastLanguageModel.from_pretrained`` with the user's own consent.
"""
from transformers import AutoConfig
if token:
return AutoConfig.from_pretrained(
model_name, trust_remote_code = trust_remote_code, token = token
)
if not use_auth:
# No auth, for public model checks
with without_hf_auth():
return AutoConfig.from_pretrained(
model_name,
trust_remote_code = trust_remote_code,
token = None,
)
# Default auth (cached tokens)
return AutoConfig.from_pretrained(
model_name,
trust_remote_code = trust_remote_code,
)
# Detection sets come from the installed transformers registry, unioned with a
# small curated set of auto_map VLMs (DeepSeek-OCR, Kimi, phi3_v) whose arch is
# repo-defined and absent from the registry. ForConditionalGeneration is NOT a
# vision signal (overloaded across text/audio/vision); ForVisionText2Text is.
_VLM_ARCH_SUFFIXES = ("ForVisionText2Text",)
_CURATED_REMOTE_VLM_TYPES = frozenset(
{
"phi3_v",
"llava",
"llava_next",
"llava_onevision",
"internvl_chat",
"cogvlm2",
"minicpmv",
"gemma4",
"deepseek_vl_v2",
"kimi_k25",
}
)
# Fallbacks used only if the transformers registry import fails.
_FALLBACK_AUDIO_MODEL_TYPES = frozenset({"csm", "whisper"})
def _build_detection_sets():
"""Return (vlm_model_types, vlm_class_names, audio_model_types) from the
installed transformers registry, unioned with the curated repo-code VLM
set. Reads only static name dicts -- no model is loaded, no code runs.
Falls back to curated/hardcoded values if transformers is unavailable.
"""
try:
from transformers.models.auto import modeling_auto as _ma
def _names(attr):
d = getattr(_ma, attr, None)
return dict(d) if d else {}
itt = _names("MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES")
v2s = _names("MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES")
vlm_types = set(itt) | set(v2s) | set(_CURATED_REMOTE_VLM_TYPES)
vlm_classes = set(itt.values()) | set(v2s.values())
audio_types: set = set()
for attr in (
"MODEL_FOR_CTC_MAPPING_NAMES",
"MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES",
"MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES",
"MODEL_FOR_TEXT_TO_WAVEFORM_MAPPING_NAMES",
"MODEL_FOR_TEXT_TO_SPECTROGRAM_MAPPING_NAMES",
"MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES",
):
audio_types |= set(_names(attr))
audio_types |= set(_FALLBACK_AUDIO_MODEL_TYPES)
return frozenset(vlm_types), frozenset(vlm_classes), frozenset(audio_types)
except Exception as exc: # pragma: no cover - defensive
logger.warning("Could not build detection sets from transformers: %s", exc)
return (
frozenset(_CURATED_REMOTE_VLM_TYPES),
frozenset(),
frozenset(_FALLBACK_AUDIO_MODEL_TYPES),
)
_VLM_MODEL_TYPES, _VLM_CLASS_NAMES, _AUDIO_ONLY_MODEL_TYPES = _build_detection_sets()
# Pre-computed .venv_t5 paths and backend dir for subprocess version switching.
# Vision check uses the Gemma 4 5.5 sidecar for existing Gemma 4 architectures.
from utils.paths.storage_roots import studio_root as _studio_root # noqa: E402
_VENV_T5_DIR = str(_studio_root() / ".venv_t5_550")
_BACKEND_DIR = str(Path(__file__).resolve().parent.parent.parent)
def _is_vlm(config) -> bool:
architectures = getattr(config, "architectures", None) or []
model_type = getattr(config, "model_type", None)
explicit_vision = (
hasattr(config, "vision_config")
or hasattr(config, "img_processor")
or hasattr(config, "image_token_index")
or hasattr(config, "projector_config")
)
# Audio-only models are vision only if they carry an explicit vision sub-config.
if model_type in _AUDIO_ONLY_MODEL_TYPES and not explicit_vision:
return False
return (
explicit_vision
or any(x in _VLM_CLASS_NAMES for x in architectures)
or any(isinstance(x, str) and x.endswith(_VLM_ARCH_SUFFIXES) for x in architectures)
or model_type in _VLM_MODEL_TYPES
)
def _raw_config_has_vision_config(
model_name: str, hf_token: Optional[str] = None
) -> Optional[bool]:
try:
if is_local_path(model_name):
config_path = Path(normalize_path(model_name)).expanduser() / "config.json"
else:
from huggingface_hub import hf_hub_download
config_path = Path(
hf_hub_download(
repo_id = model_name,
filename = "config.json",
token = hf_token,
)
)
config = json.loads(config_path.read_text())
architectures = config.get("architectures") or []
model_type = config.get("model_type")
explicit_vision = (
"vision_config" in config
or "img_processor" in config
or "image_token_index" in config
or "projector_config" in config
)
# Audio-only models are vision only if they carry an explicit vision sub-config.
if model_type in _AUDIO_ONLY_MODEL_TYPES and not explicit_vision:
return False
return (
explicit_vision
or any(isinstance(x, str) and x in _VLM_CLASS_NAMES for x in architectures)
or any(isinstance(x, str) and x.endswith(_VLM_ARCH_SUFFIXES) for x in architectures)
or model_type in _VLM_MODEL_TYPES
)
except Exception as exc:
logger.warning("Could not read config.json for '%s': %s", model_name, exc)
return None
# why: inline _is_vlm and constants are prepended so the subprocess stays
# self-contained and does not import the parent backend module graph.
_VISION_CHECK_INLINE_HELPERS = (
"_VLM_ARCH_SUFFIXES = " + repr(tuple(_VLM_ARCH_SUFFIXES)) + "\n"
"_VLM_MODEL_TYPES = " + repr(set(_VLM_MODEL_TYPES)) + "\n"
"_VLM_CLASS_NAMES = " + repr(set(_VLM_CLASS_NAMES)) + "\n"
"_AUDIO_ONLY_MODEL_TYPES = " + repr(set(_AUDIO_ONLY_MODEL_TYPES)) + "\n"
"def _is_vlm(config):\n"
" architectures = getattr(config, 'architectures', None) or []\n"
" model_type = getattr(config, 'model_type', None)\n"
" explicit_vision = (\n"
" hasattr(config, 'vision_config')\n"
" or hasattr(config, 'img_processor')\n"
" or hasattr(config, 'image_token_index')\n"
" or hasattr(config, 'projector_config')\n"
" )\n"
" if model_type in _AUDIO_ONLY_MODEL_TYPES and not explicit_vision:\n"
" return False\n"
" return (\n"
" explicit_vision\n"
" or any(x in _VLM_CLASS_NAMES for x in architectures)\n"
" or any(isinstance(x, str) and x.endswith(_VLM_ARCH_SUFFIXES) for x in architectures)\n"
" or model_type in _VLM_MODEL_TYPES\n"
" )\n"
)
# Subprocess script run with transformers 5.x active. Takes model_name and
# token via argv, prints JSON result to stdout.
_VISION_CHECK_SCRIPT = (
r"""
import sys, os, json
os.environ["TOKENIZERS_PARALLELISM"] = "false"
# Activate transformers 5.x
venv_t5 = sys.argv[1]
backend_dir = sys.argv[2]
model_name = sys.argv[3]
token = sys.argv[4] if len(sys.argv) > 4 and sys.argv[4] != "" else None
sys.path.insert(0, venv_t5)
if backend_dir not in sys.path:
sys.path.insert(0, backend_dir)
"""
+ _VISION_CHECK_INLINE_HELPERS
+ r"""
try:
from transformers import AutoConfig
# Capability detection never executes model repo code.
kwargs = {"trust_remote_code": False}
if token:
kwargs["token"] = token
config = AutoConfig.from_pretrained(model_name, **kwargs)
is_vlm = _is_vlm(config)
model_type = getattr(config, "model_type", None)
archs = getattr(config, "architectures", [])
print(json.dumps({"is_vision": is_vlm, "model_type": model_type,
"architectures": archs}))
except Exception as exc:
print(json.dumps({"error": str(exc)}))
sys.exit(1)
"""
)
def _is_vision_model_subprocess(model_name: str, hf_token: Optional[str] = None) -> Optional[bool]:
"""Run is_vision_model in a subprocess with transformers 5.x.
Spawns a clean subprocess with .venv_t5/ on sys.path so AutoConfig
recognizes newer architectures. Returns True/False for definitive results,
or None for transient failures (timeouts, subprocess errors), which are not
cached so they can be retried.
"""
token_arg = hf_token or ""
try:
result = subprocess.run(
[
sys.executable,
"-c",
_VISION_CHECK_SCRIPT,
_VENV_T5_DIR,
_BACKEND_DIR,
model_name,
token_arg,
],
capture_output = True,
text = True,
timeout = 60,
env = child_env_without_native_path_secret(),
**_windows_hidden_subprocess_kwargs(),
)
if result.returncode != 0:
stderr = result.stderr.strip()
logger.warning(
"Vision check subprocess failed for '%s': %s",
model_name,
stderr or result.stdout.strip(),
)
return None
data = json.loads(result.stdout.strip())
if "error" in data:
logger.warning(
"Vision check subprocess error for '%s': %s",
model_name,
data["error"],
)
return None
is_vlm = data["is_vision"]
logger.info(
"Vision check (subprocess, transformers 5.x) for '%s': "
"model_type=%s, architectures=%s, is_vision=%s",
model_name,
data.get("model_type"),
data.get("architectures"),
is_vlm,
)
return is_vlm
except subprocess.TimeoutExpired:
logger.warning("Vision check subprocess timed out for '%s'", model_name)
return None
except Exception as exc:
logger.warning("Vision check subprocess failed for '%s': %s", model_name, exc)
return None
def _token_fingerprint(token: Optional[str]) -> Optional[str]:
"""SHA256 digest of the token for use as a cache key (avoids storing the
raw bearer token in process memory)."""
if token is None:
return None
return hashlib.sha256(token.encode("utf-8")).hexdigest()
# Cache vision detection per session to avoid repeated subprocess spawns.
# Keyed by (normalized_model_name, token_fingerprint) to handle gated models.
# Only definitive results are cached; transient failures (network, timeouts)
# are NOT cached so they can be retried.
_vision_detection_cache: Dict[Tuple[str, Optional[str]], bool] = {}
_vision_cache_lock = threading.Lock()
def is_vision_model(model_name: str, hf_token: Optional[str] = None) -> bool:
"""
Detect vision-language models (VLMs) via architecture in config. Works for
fine-tuned models since they inherit the base architecture.
Models needing transformers 5.x are checked in a .venv_t5/ subprocess.
Results are cached per (model_name, token_fingerprint) for the process
lifetime; transient failures are not cached so they can be retried.
Args:
model_name: Model identifier (HF repo or local path)
hf_token: Optional HF token for gated/private models
"""
# Local GGUF models are served by llama-server. Their multimodal
# capability comes from a companion mmproj, not a Transformers config.
# Do not cache this lookup: a projector may be added beside an existing
# weight file after it was first inspected.
if is_local_path(model_name):
local_path = normalize_path(model_name)
gguf_file = detect_gguf_model(local_path)
if gguf_file:
companion_root = _local_gguf_companion_search_root(local_path, gguf_file)
mmproj_file = detect_mmproj_file(gguf_file, search_root = companion_root)
is_vision = mmproj_file is not None
logger.debug(
"Local GGUF vision check for '%s': mmproj=%s, is_vision=%s",
gguf_file,
mmproj_file,
is_vision,
)
return is_vision
# Normalize model name so different casings of the same repo share a key
try:
if is_local_path(model_name):
resolved_name = normalize_path(model_name)
else:
resolved_name = resolve_cached_repo_id_case(model_name)
except Exception as exc:
logger.debug(
"Could not normalize model name '%s' for cache key: %s",
model_name,
exc,
)
resolved_name = model_name
cache_key = (resolved_name, _token_fingerprint(hf_token))
# Lock-free fast path for cache hits. Sentinel distinguishes "key not found"
# from "value is False" in a single atomic dict.get() call.
_MISS = object()
cached = _vision_detection_cache.get(cache_key, _MISS)
if cached is not _MISS:
return cached
# Compute outside the lock so long-running detection isn't serialized across
# models. Two concurrent calls may both run, but produce the same result.
result = _is_vision_model_uncached(resolved_name, hf_token)
# Only cache definitive results; None is a transient failure, retry later.
if result is not None:
with _vision_cache_lock:
_vision_detection_cache[cache_key] = result
return result
return False
def _is_vision_model_uncached(model_name: str, hf_token: Optional[str] = None) -> Optional[bool]:
"""Uncached vision detection; use is_vision_model() instead.
Returns True/False for definitive results, or None on transient errors
(network, timeout, subprocess failure) so the caller knows not to cache.
"""
# Try the raw-config reader FIRST (code-free, version-independent): it classifies
# repo-code VLMs like DeepSeek-OCR via declarative vision_config with no remote-code
# execution or transformers-5.x subprocess.
raw = _raw_config_has_vision_config(model_name, hf_token = hf_token)
if raw is not None:
return raw
# Raw read failed transiently: fall back to AutoConfig with remote code DISABLED
# (in a transformers-5.x subprocess when the main process can't parse the arch).
from utils.transformers_version import needs_transformers_5
if needs_transformers_5(model_name):
logger.info(
"Model '%s' needs transformers 5.x -- checking vision via subprocess",
model_name,
)
return _is_vision_model_subprocess(model_name, hf_token = hf_token)
try:
config = load_model_config(model_name, use_auth = True, token = hf_token)
if _is_vlm(config):
model_type = getattr(config, "model_type", None)
archs = getattr(config, "architectures", None) or []
logger.info(
"Model %s detected as VLM (model_type=%s, architectures=%s)",
model_name,
model_type,
archs,
)
return True
return False
except Exception as e:
logger.warning(f"Could not determine if {model_name} is vision model: {e}")
# Permanent failures (not found, gated, bad config) cache as False;
# transient ones (network, timeout) should not.
try:
from huggingface_hub.errors import RepositoryNotFoundError, GatedRepoError
except ImportError:
try:
from huggingface_hub.utils import (
RepositoryNotFoundError,
GatedRepoError,
)
except ImportError:
RepositoryNotFoundError = GatedRepoError = None
if RepositoryNotFoundError is not None and isinstance(
e, (RepositoryNotFoundError, GatedRepoError)
):
return False
if isinstance(e, (ValueError, json.JSONDecodeError)):
return False
return None
VALID_AUDIO_TYPES = ("snac", "csm", "bicodec", "dac", "whisper", "audio_vlm")
# Keyed by (normalized_name, token_fingerprint) like the vision cache, so an
# unauthenticated miss (None) cannot poison a later authenticated lookup.
_audio_detection_cache: Dict[Tuple[str, Optional[str]], Optional[str]] = {}
# Tokenizer token patterns → audio_type (all 6 types from tokenizer_config.json)
_AUDIO_TOKEN_PATTERNS = {
"csm": lambda tokens: "<|AUDIO|>" in tokens and "<|audio_eos|>" in tokens,
"whisper": lambda tokens: "<|startoftranscript|>" in tokens,
# Gemma 3n: <audio_soft_token>; Gemma 4: <|audio|> (not csm's <|AUDIO|>).
"audio_vlm": lambda tokens: "<audio_soft_token>" in tokens or "<|audio|>" in tokens,
"bicodec": lambda tokens: any(t.startswith("<|bicodec_") for t in tokens),
"dac": lambda tokens: (
"<|audio_start|>" in tokens
and "<|audio_end|>" in tokens
and "<|text_start|>" in tokens
and "<|text_end|>" in tokens
),
"snac": lambda tokens: (sum(1 for t in tokens if t.startswith("<custom_token_")) > 10000),
}
def detect_audio_type(model_name: str, hf_token: Optional[str] = None) -> Optional[str]:
"""Detect if a model is an audio model and return its type.
Works for any model via tokenizer_config.json special tokens.
Returns an audio_type string ('snac', 'csm', 'bicodec', 'dac', 'whisper',
'audio_vlm') or None.
"""
# Normalize casing + include the token fingerprint (mirrors is_vision_model).
try:
if is_local_path(model_name):
resolved_name = normalize_path(model_name)
else:
resolved_name = resolve_cached_repo_id_case(model_name)
except Exception:
resolved_name = model_name
cache_key = (resolved_name, _token_fingerprint(hf_token))
if cache_key in _audio_detection_cache:
return _audio_detection_cache[cache_key]
result, definitive = _detect_audio_from_tokenizer(model_name, hf_token)
# Cache only definitive results; a transient read failure stays None and retries.
if definitive:
_audio_detection_cache[cache_key] = result
if result:
logger.info(f"Model {model_name} detected as audio model: audio_type={result}")
return result
def _detect_audio_from_tokenizer(
model_name: str, hf_token: Optional[str] = None
) -> Tuple[Optional[str], bool]:
"""Detect audio type from tokenizer special tokens.
Checks local HF cache first, then fetches tokenizer_config.json from HF;
examines added_tokens_decoder for distinctive patterns.
Returns (audio_type_or_None, definitive). definitive is False only on a
transient read failure (network/timeout/5xx) so the caller skips caching and
retries; a successful read with no audio tokens is a definitive None.
"""
def _check_token_patterns(tok_config: dict) -> Optional[str]:
added = tok_config.get("added_tokens_decoder", {})
if not added:
return None
token_contents = [v.get("content", "") for v in added.values()]
for audio_type, check_fn in _AUDIO_TOKEN_PATTERNS.items():
if check_fn(token_contents):
return audio_type
return None
read_any = False # parsed at least one tokenizer_config -> a None is definitive
# 1) Local HF cache first (works for gated/offline models)
try:
repo_dir = get_cache_path(model_name)
if repo_dir is not None and repo_dir.exists():
snapshots_dir = repo_dir / "snapshots"
if snapshots_dir.exists():
for snapshot in snapshots_dir.iterdir():
for tok_path in [
"tokenizer_config.json",
"LLM/tokenizer_config.json",
]:
tok_file = snapshot / tok_path
if tok_file.exists():
tok_config = json.loads(tok_file.read_text())
read_any = True
result = _check_token_patterns(tok_config)
if result:
return result, True
except Exception as e:
logger.debug(f"Could not check local cache for {model_name}: {e}")
# 2) Fall back to HuggingFace API
try:
import requests
import os
except Exception:
return None, read_any
paths_to_try = ["tokenizer_config.json", "LLM/tokenizer_config.json"]
token = hf_token or os.environ.get("HF_TOKEN")
headers = {"Authorization": f"Bearer {token}"} if token else {}
transient = False # a fetch failed for a non-404 reason (network/5xx)
for tok_path in paths_to_try:
url = f"https://huggingface.co/{model_name}/resolve/main/{tok_path}"
try:
resp = requests.get(url, headers = headers, timeout = 15)
except Exception as e:
logger.debug(f"Could not fetch {tok_path} for {model_name}: {e}")
transient = True
continue
if resp.status_code == 404:
continue # genuinely absent on this path
if not resp.ok:
transient = True # 5xx/403/etc -- can't tell, don't cache
continue
try:
tok_config = resp.json()
except Exception as e:
logger.debug(f"Bad tokenizer_config for {model_name}/{tok_path}: {e}")
transient = True
continue
read_any = True
result = _check_token_patterns(tok_config)
if result:
return result, True
# No audio tokens: definitive unless every attempt failed transiently.
return None, (read_any or not transient)
def is_audio_input_type(audio_type: Optional[str]) -> bool:
"""True if an audio_type accepts audio input: whisper (ASR), audio_vlm (Gemma3n)."""
return audio_type in ("whisper", "audio_vlm")
def _is_mmproj(filename: str) -> bool:
"""Check if a GGUF filename is a vision projection (mmproj) file."""
return "mmproj" in filename.lower()
def _is_mtp_drafter(path: str) -> bool:
"""True for a separate-file MTP drafter (speculative head), a companion
to the main model rather than a selectable quant: the repo-root
``mtp-*.gguf`` or the ``MTP/`` subdir copies (Gemma 4).
Mirrors hub.utils.gguf.is_mtp_drafter_path (utils cannot import hub).
Must be excluded everywhere mmproj is, or the drafter leaks into variant
menus (a phantom quant) and quant-matched file lookups -- e.g. a ``Q8_0``
request must not resolve to ``MTP/...-Q8_0-MTP.gguf``, which sorts ahead
of the real weight.
"""
p = path.lower()
if not p.endswith(".gguf"):
return False
name = p.rsplit("/", 1)[-1]
return name.startswith("mtp-") or "/mtp/" in f"/{p}"
# Family tokens for #5347's filename fallback. Lowercase; order irrelevant.
_MODEL_FAMILY_TOKENS: tuple[str, ...] = (
"qwen",
"gemma",
"llama",
"mistral",
"ministral",
"magistral",
"devstral",
"phi",
"deepseek",
"internvl",
"minicpm",
"llava",
"glm",
"yi",
"command-r",
"molmo",
"pixtral",
"smolvlm",
"moondream",
"granite",
"ovis",
"nemotron",
"kimi",
"nanonets",
"cosmos",
"mimo",
"apriel",
"lfm",
)
# Word-bounded match: a letter on either side disqualifies (stops ``phi``
# matching ``sapphire``, ``yi`` matching ``tiny``).
_FAMILY_TOKEN_RE_CACHE: Dict[str, "_re.Pattern[str]"] = {}
def _family_token_re(token: str) -> "_re.Pattern[str]":
pat = _FAMILY_TOKEN_RE_CACHE.get(token)
if pat is None:
pat = _re.compile(rf"(?:^|[^a-z])({_re.escape(token)})(?:[^a-z]|$)")
_FAMILY_TOKEN_RE_CACHE[token] = pat
return pat
def _detect_family_token(filename: str) -> Optional[str]:
"""Leftmost-position match; ties prefer the longer token."""
name = filename.lower()
best: Optional[tuple[int, int, str]] = None # (start, -len, token)
for token in _MODEL_FAMILY_TOKENS:
m = _family_token_re(token).search(name)
if m is None:
continue
key = (m.start(1), -len(token), token)
if best is None or key < best:
best = key
return None if best is None else best[2]
def mmproj_matches_model_family(model_path: str, mmproj_path: str) -> bool:
"""Launcher guard: True unless both filenames carry recognised family
tokens that disagree."""
model_fam = _detect_family_token(Path(model_path).name)
mmproj_fam = _detect_family_token(Path(mmproj_path).name)
if model_fam is None or mmproj_fam is None:
return True
return model_fam == mmproj_fam
def _shared_prefix_len(a: str, b: str) -> int:
n = min(len(a), len(b))
for i in range(n):
if a[i] != b[i]:
return i
return n
def _is_gguf_filename(filename: str) -> bool:
return filename.lower().endswith(".gguf")
def _iter_gguf_files(directory: Path, recursive: bool = False):
if not directory.is_dir():
return
iterator = directory.rglob("*") if recursive else directory.iterdir()
for f in iterator:
if f.is_file() and _is_gguf_filename(f.name):
yield f
def detect_mmproj_file(path: str, search_root: Optional[str] = None) -> Optional[str]:
"""Find the mmproj GGUF for a model.
``path``: directory or a .gguf file. ``search_root``: optional ancestor
to also walk (snapshot layouts where the weight is in ``snapshot/BF16/``
but the projector sits at ``snapshot/``). Returns the projector path or
``None``."""
p = Path(path)
start_dir = p.parent if p.is_file() else p
if not start_dir.is_dir():
return None
# Walk incrementally so a sibling subdir's mmproj cannot leak in.
seen: set[Path] = set()
scan_order: list[Path] = []
def _add(d: Path) -> None:
try:
resolved = d.resolve()
except OSError:
return
if resolved in seen or not resolved.is_dir():
return
seen.add(resolved)
scan_order.append(resolved)
_add(start_dir)
# Ollama's .studio_links/foo.gguf -> blobs/sha256-...: also scan target dir.
try:
if p.is_symlink() and p.is_file():
target_parent = p.resolve().parent
if target_parent.is_dir():
_add(target_parent)
except OSError:
pass
if search_root is not None:
try:
root_resolved = Path(search_root).resolve()
start_resolved = start_dir.resolve()
if root_resolved == start_resolved or (
start_resolved.is_relative_to(root_resolved)
if hasattr(start_resolved, "is_relative_to")
else str(start_resolved).startswith(str(root_resolved) + "/")
):
cur = start_resolved
while cur != root_resolved and cur.parent != cur:
cur = cur.parent
_add(cur)
if cur == root_resolved:
break
except OSError:
pass
candidates: list[Path] = []
seen_resolved: set[Path] = set()
for d in scan_order:
for f in _iter_gguf_files(d):
try:
resolved = f.resolve()
except OSError:
continue
if resolved in seen_resolved:
continue
# Prefer ``general.type=='mmproj'``, else filename.
meta = read_gguf_general_metadata(str(resolved))
by_meta = is_mmproj_by_metadata(meta)
if by_meta is True or (by_meta is None and _is_mmproj(f.name)):
seen_resolved.add(resolved)
candidates.append(resolved)
if not candidates:
return None
# Directory path: no model name to compare against; legacy behaviour.
if not p.is_file():
return str(candidates[0])
# Stage 1: GGUF metadata. Stage 2: filename family token (#5347).
model_stem = p.stem.lower()
model_family = _detect_family_token(p.name)
weight_meta = read_gguf_general_metadata(str(p))
scored: list[tuple[int, Path]] = []
for c in candidates:
cand_meta = read_gguf_general_metadata(str(c))
meta_score = pairing_score(weight_meta, cand_meta)
if meta_score == -1:
logger.info(f"detect_mmproj_file: dropped {c.name} (metadata mismatch)")
continue
if meta_score == 0 and model_family is not None:
# Unrecognised candidate family is a wildcard (``mmproj-F16.gguf``).
cand_family = _detect_family_token(c.name)
if cand_family is not None and cand_family != model_family:
logger.info(
f"detect_mmproj_file: dropped {c.name} "
f"(filename family {cand_family!r} vs model {model_family!r})"
)
continue
scored.append((meta_score, c))
if not scored:
return None
# Score first, then longest shared prefix, then shorter stem.
best = max(
scored,
key = lambda sc: (
sc[0],
_shared_prefix_len(model_stem, sc[1].stem.lower()),
-len(sc[1].stem),
),
)
return str(best[1])
def detect_mtp_file(path: str, search_root: Optional[str] = None) -> Optional[str]:
"""Find the separate MTP drafter (``mtp-*.gguf``) for a local GGUF model.
The drafter that pairs with the main weights sits at the repo/snapshot
root (Gemma 4); the weight itself may be at the root or in a quant subdir,
so scan the weight's directory and ``search_root``. Matches by the
``mtp-`` filename prefix unsloth uses for ``-hf`` auto-discovery -- the
same signal as the HF download path. Repos that bake the head into the
main GGUF (Qwen) have no such sibling, so this returns None.
Pairs by name so a multi-model folder can't attach a foreign drafter:
unsloth names the drafter ``mtp-<model>.gguf`` where ``<model>`` prefixes
the weight filename across all Gemma 4 repos (e.g.
``mtp-gemma-4-12B-it.gguf`` next to ``gemma-4-12B-it-qat-Q4_0.gguf``).
An unmatched drafter is skipped (fail-safe: no MTP).
"""
p = Path(path)
weight_name = p.name.lower() if p.suffix.lower() == ".gguf" else None
start_dir = p.parent if p.is_file() else p
dirs = [start_dir]
if search_root is not None:
dirs.append(Path(search_root))
for d in dirs:
try:
entries = sorted(d.iterdir())
except OSError:
continue
for f in entries:
name = f.name.lower()
if not (name.startswith("mtp-") and name.endswith(".gguf")):
continue
stem = name[len("mtp-") : -len(".gguf")]
if not stem or (weight_name is not None and not weight_name.startswith(stem)):
continue
try:
if f.is_file():
return str(f.resolve())
except OSError:
continue
return None
def detect_gguf_model(path: str) -> Optional[str]:
"""Check if a local path is or contains a GGUF model file.
Handles a direct .gguf path or a directory of .gguf files. Skips mmproj
files (pass those via ``--mmproj``; see :func:`detect_mmproj_file`). Returns
the .gguf path or None. For HF repos, use detect_gguf_model_remote().
"""
p = Path(path)
# Case 1: direct .gguf file
if p.suffix.lower() == ".gguf":
# Companions are not models: rejecting a drafter here also keeps
# detect_mtp_file from pairing the same file with itself
# (-m drafter --model-draft drafter). Include the immediate parent
# dir so the MTP/ subdir copies are caught -- the basename alone
# (...-MTP.gguf) doesn't match the predicate's mtp- prefix.
rel = f"{p.parent.name}/{p.name}"
quant = _extract_quant_label(rel)
if _is_mmproj(p.name) or _is_mtp_drafter(rel) or _is_big_endian_gguf_path(rel, quant):
return None
# Extension is authoritative: don't gate on is_file()/exists(), which
# can fail in the Windows lock window after llama-server is killed.
try:
is_dir = p.is_dir()
except OSError:
is_dir = False # stat() unavailable in the lock window
if not is_dir:
return str(p.absolute()) # absolute() keeps symlink names readable
# Directory named "*.gguf": fall through to the dir scan below.
# Case 2: directory containing .gguf files (skip mmproj / MTP drafter)
if p.is_dir():
gguf_files = []
for f in _iter_gguf_files(p):
context_rel = f"{f.parent.name}/{f.name}"
quant = _extract_quant_label(context_rel)
if (
_is_mmproj(f.name)
or _is_mtp_drafter(context_rel)
or _is_big_endian_gguf_path(context_rel, quant)
):
continue
gguf_files.append(f)
gguf_files.sort(key = lambda f: f.stat().st_size, reverse = True)
if gguf_files:
return str(gguf_files[0].resolve())
return None
# Preferred GGUF quant levels, descending priority. UD (Unsloth Dynamic)
# variants beat standard quants on quality per bit; repos without UD fall back
# to standard quants. Ordered by size/quality tradeoff, not raw quality.
_GGUF_QUANT_PREFERENCE = [
# UD variants (best quality per bit) -- Q4 is the sweet spot
"UD-Q4_K_XL",
"UD-Q4_K_L",
"UD-Q5_K_XL",
"UD-Q3_K_XL",
"UD-Q6_K_XL",
"UD-Q6_K_S",
"UD-Q8_K_XL",
"UD-Q2_K_XL",
"UD-IQ4_NL",
"UD-IQ4_XS",
"UD-IQ3_S",
"UD-IQ3_XXS",
"UD-IQ2_M",
"UD-IQ2_XXS",
"UD-IQ1_M",
"UD-IQ1_S",
# Standard quants (fallback for non-Unsloth repos)
"Q4_K_M",
"Q4_K_S",
"Q5_K_M",
"Q5_K_S",
"Q6_K",
"Q8_0",
"Q3_K_M",
"Q3_K_L",
"Q3_K_S",
"Q2_K",
"Q2_K_L",
"IQ4_NL",
"IQ4_XS",
"IQ3_M",
"IQ3_XXS",
"IQ2_M",
"IQ1_M",
"F16",
"BF16",
"F32",
]
def _pick_best_gguf(filenames: list[str]) -> Optional[str]:
"""Pick the best GGUF file: quant levels in _GGUF_QUANT_PREFERENCE order, else first .gguf."""
gguf_files = [f for f in filenames if f.lower().endswith(".gguf")]
if not gguf_files:
return None
for quant in _GGUF_QUANT_PREFERENCE:
for f in gguf_files:
if quant in f:
return f
return gguf_files[0]
@dataclass
class GgufVariantInfo:
"""A single GGUF quantization variant from a HuggingFace repo."""
filename: str # e.g., "gemma-3-4b-it-Q4_K_M.gguf"
quant: str # e.g., "Q4_K_M" (extracted from filename)
size_bytes: int # file size
def _extract_quant_label(filename: str) -> str:
"""
Extract quant label like Q4_K_M, IQ4_XS, BF16 from a GGUF filename.
Examples:
"gemma-3-4b-it-Q4_K_M.gguf""Q4_K_M"
"model-IQ4_NL.gguf""IQ4_NL"
"model-BF16.gguf""BF16"
"model-UD-IQ1_S.gguf""UD-IQ1_S"
"model-UD-TQ1_0.gguf""UD-TQ1_0"
"MXFP4_MOE/model-MXFP4_MOE-0001.gguf""MXFP4_MOE"
"Qwen3.6-IQ4_XS-3.53bpw.gguf""IQ4_XS-3.53bpw"
"""
import re
basename = filename.rsplit("/", 1)[-1]
# Strip .gguf and any shard suffix (-00001-of-00010)
stem = re.sub(r"-\d{3,}-of-\d{3,}", "", basename.rsplit(".", 1)[0])
quant_re = (
r"(UD-)?" # Optional UD- prefix (Ultra Discrete)
r"(MXFP[0-9]+(?:_[A-Z0-9]+)*" # MXFP variants: MXFP4, MXFP4_MOE
r"|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?" # IQ variants: IQ4_XS, IQ4_NL, IQ1_S
r"|TQ[0-9]+_[0-9]+" # Ternary quant: TQ1_0, TQ2_0
r"|Q[0-9]+_K_[A-Z]+" # K-quant: Q4_K_M, Q3_K_S
r"|Q[0-9]+_[0-9]+" # Standard: Q8_0, Q5_1
r"|Q[0-9]+_K" # Short K-quant: Q6_K
r"|BF16|F16|F32)" # Full precision
# Optional bits-per-weight modifier so repos that ship multiple
# files at the same base quant (e.g. byteshape's IQ4_XS at 3.53,
# 3.97, 4.19 bpw) don't collapse into a single merged variant.
r"(-[0-9]+(?:\.[0-9]+)?bpw)?"
)
match = re.search(quant_re, stem, re.IGNORECASE)
# Subdir layouts like ``BF16/foo.gguf`` keep the quant in the directory,
# not the basename. Check parent dirs too so the label matches the
# snapshot-relative path produced elsewhere.
if not match and "/" in filename:
parents = filename.rsplit("/", 1)[0]
for segment in reversed(parents.split("/")):
m = re.search(quant_re, segment, re.IGNORECASE)
if m:
match = m
break
if match:
prefix = match.group(1) or ""
bpw = match.group(3) or ""
return f"{prefix}{match.group(2)}{bpw}"
# Fallback: last hyphen-separated segment
return stem.split("-")[-1]
_BIG_ENDIAN_GGUF_FILENAME_RE = re.compile(r"(^|[-_])be(?:[._-]|$)", re.IGNORECASE)
_GGUF_KNOWN_QUANT_RE = re.compile(
r"(UD-)?"
r"(MXFP[0-9]+(?:_[A-Z0-9]+)*"
r"|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?"
r"|TQ[0-9]+_[0-9]+"
r"|Q[0-9]+_K_[A-Z]+"
r"|Q[0-9]+_[0-9]+"
r"|Q[0-9]+_K"
r"|BF16|F16|F32)",
re.IGNORECASE,
)
def _is_big_endian_gguf_path(path: str, quant: str = "") -> bool:
normalized = path.replace("\\", "/")
name = normalized.rsplit("/", 1)[-1]
stem = name.rsplit(".", 1)[0].lower()
quant_key = quant.strip().lower()
quant_index = stem.find(quant_key) if quant_key else -1
parent = normalized.rsplit("/", 1)[0].lower() if "/" in normalized else ""
quant_in_parent_only = (
bool(parent)
and quant_index < 0
and (
(quant_key and quant_key in parent)
or (not quant_key and _GGUF_KNOWN_QUANT_RE.search(parent) is not None)
)
)
for match in _BIG_ENDIAN_GGUF_FILENAME_RE.finditer(stem):
if quant_index >= 0 and quant_index < match.start():
return True
tail = stem[match.end() :].lstrip("._-")
if not tail or _GGUF_KNOWN_QUANT_RE.search(tail) is None:
return not quant_in_parent_only
return False
def _local_gguf_companion_search_root(selected_path: str, gguf_file: str) -> str:
"""Directory to scan upward from for local GGUF companion files."""
import re
selected = Path(selected_path)
gguf_path = Path(gguf_file)
if selected.suffix.lower() != ".gguf":
return selected_path
gguf_dir = gguf_path.parent
if not gguf_dir.name:
return str(gguf_dir)
quant_dir_re = (
r"(UD-)?("
r"MXFP[0-9]+(?:_[A-Z0-9]+)*"
r"|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?"
r"|TQ[0-9]+_[0-9]+"
r"|Q[0-9]+_K_[A-Z]+"
r"|Q[0-9]+_[0-9]+"
r"|Q[0-9]+_K"
r"|BF16|F16|F32"
r")"
)
if re.fullmatch(quant_dir_re, gguf_dir.name, re.IGNORECASE):
return str(gguf_dir.parent)
return str(gguf_dir)
def _iter_hf_cache_snapshots(repo_id: str):
"""Yield HF cache snapshot dirs for *repo_id*, newest first.
Empty if HF_HUB_CACHE is missing, the repo isn't cached, or has no
snapshots. Repo name match is case-insensitive to handle casing drift
between download time and lookup.
"""
try:
from huggingface_hub import constants as hf_constants
except Exception:
return
cache_dir = Path(hf_constants.HF_HUB_CACHE)
target = f"models--{repo_id.replace('/', '--')}".lower()
repo_dir: Optional[Path] = None
try:
if not cache_dir.is_dir():
return
for entry in cache_dir.iterdir():
if entry.is_dir() and entry.name.lower() == target:
repo_dir = entry
break
except OSError:
return
if repo_dir is None:
return
snapshots = repo_dir / "snapshots"
try:
if not snapshots.is_dir():
return
snap_dirs = [s for s in snapshots.iterdir() if s.is_dir()]
except OSError:
return
snap_dirs.sort(key = lambda s: s.stat().st_mtime, reverse = True)
yield from snap_dirs
def _list_gguf_variants_from_hf_cache(repo_id: str) -> Optional[tuple[list[GgufVariantInfo], bool]]:
"""Variants from the local HF cache snapshot, or None if not cached."""
for snap in _iter_hf_cache_snapshots(repo_id):
variants, has_vision = list_local_gguf_variants(str(snap))
if variants or has_vision:
return variants, has_vision
return None
def list_gguf_variants(
repo_id: str, hf_token: Optional[str] = None
) -> tuple[list[GgufVariantInfo], bool]:
"""List all GGUF quant variants in a HF repo.
Separates main model files from mmproj (vision projection) files; mmproj
presence flags a vision-capable model.
Returns:
(variants, has_vision): non-mmproj GGUF variants + vision flag.
"""
from huggingface_hub import model_info as hf_model_info
# Offline: skip the API and serve from cache
if _env_offline():
cached = _list_gguf_variants_from_hf_cache(repo_id)
if cached is not None:
return cached
try:
info = hf_model_info(repo_id, token = hf_token, files_metadata = True)
except Exception as e:
# Permanent errors (deleted/gated/bad revision) must surface to the
# caller; serving stale cache would mask the real cause. Matches the
# early-return in ``detect_gguf_model_remote``.
if type(e).__name__ in (
"RepositoryNotFoundError",
"GatedRepoError",
"RevisionNotFoundError",
"EntryNotFoundError",
):
raise
# API failed transiently; fall back to local snapshot if fully downloaded.
cached = _list_gguf_variants_from_hf_cache(repo_id)
if cached is not None:
logger.warning(
"HF API unreachable for %s (%s); using local cache snapshot.",
repo_id,
e.__class__.__name__,
)
return cached
raise
variants: list[GgufVariantInfo] = []
has_vision = False
quant_totals: dict[str, int] = {} # quant -> total bytes
quant_first_file: dict[str, str] = {} # quant -> first filename (display)
for sibling in info.siblings:
fname = sibling.rfilename
if not fname.lower().endswith(".gguf"):
continue
size = sibling.size or 0
# mmproj files are vision projections, not main model files
if "mmproj" in fname.lower():
has_vision = True
continue
# MTP drafters are speculative-decoding companions, not quants.
if _is_mtp_drafter(fname):
continue
quant = _extract_quant_label(fname)
if _is_big_endian_gguf_path(fname, quant):
continue
quant_totals[quant] = quant_totals.get(quant, 0) + size
if quant not in quant_first_file:
quant_first_file[quant] = fname
for quant, total_size in quant_totals.items():
variants.append(
GgufVariantInfo(
filename = quant_first_file[quant],
quant = quant,
size_bytes = total_size,
)
)
# Sort by size descending (largest = best quality first); pinning and OOM
# demotion happen client-side where GPU VRAM info exists.
variants.sort(key = lambda v: -v.size_bytes)
return variants, has_vision
def _resolve_gguf_dir(p: Path) -> Optional[Path]:
"""Resolve a path to the directory containing GGUF variants.
Directory *p* returns directly. A ``.gguf`` file whose parent dir has
model metadata (``config.json`` or ``adapter_config.json``) returns the
parent -- all GGUFs there belong to the same model. Returns ``None`` for
loose standalone GGUFs (no config) to avoid cross-wiring unrelated models.
"""
if p.is_dir():
return p
if p.is_file() and p.suffix.lower() == ".gguf":
parent = p.parent
if (
(parent / "config.json").exists()
or (parent / "adapter_config.json").exists()
or (parent / "export_metadata.json").exists()
):
return parent
return None
def list_local_gguf_variants(directory: str) -> tuple[list[GgufVariantInfo], bool]:
"""List GGUF quant variants in a local directory.
Like :func:`list_gguf_variants` but reads the filesystem. Aggregates shard
sizes by quant label so split GGUFs appear as one variant.
Returns:
(variants, has_vision): non-mmproj GGUF variants + vision flag.
"""
p = _resolve_gguf_dir(Path(directory))
if p is None:
return [], False
quant_totals: dict[str, int] = {}
quant_first_file: dict[str, str] = {}
has_vision = False
# Recurse so variant-specific subdirs (e.g. ``BF16/...gguf`` used by
# some HF GGUF repos for the largest quants) are picked up. Result
# filenames keep the relative subpath so ``_find_local_gguf_by_variant``
# can locate the file again.
for f in sorted(_iter_gguf_files(p, recursive = True)):
if _is_mmproj(f.name):
has_vision = True
continue
try:
size = f.stat().st_size
except OSError:
size = 0
# Use the relative path so ``BF16/foo.gguf`` and ``Q4_K_M/foo.gguf``
# get distinct quant labels instead of collapsing on basename.
rel = f.relative_to(p).as_posix()
if _is_mtp_drafter(rel):
continue
quant = _extract_quant_label(rel)
if _is_big_endian_gguf_path(rel, quant):
continue
quant_totals[quant] = quant_totals.get(quant, 0) + size
if quant not in quant_first_file:
quant_first_file[quant] = rel
variants = [
GgufVariantInfo(
filename = quant_first_file[q],
quant = q,
size_bytes = s,
)
for q, s in quant_totals.items()
]
variants.sort(key = lambda v: -v.size_bytes)
return variants, has_vision
def _find_local_gguf_by_variant(directory: str, variant: str) -> Optional[str]:
"""Find the GGUF file in *directory* matching a quantization *variant*.
For sharded GGUFs (multiple files sharing a quant label), returns the
first shard (sorted by name), which is what ``llama-server -m`` expects.
Returns the resolved absolute path, or ``None`` if no match.
"""
p = _resolve_gguf_dir(Path(directory))
if p is None:
return None
# Recurse so variants under a quant-named subdir (e.g.
# ``BF16/foo-BF16-00001-of-00002.gguf``) are found. Match the relative
# path so the quant label can come from the dir name when the basename
# omits it.
matches = []
for f in _iter_gguf_files(p, recursive = True):
rel = f.relative_to(p).as_posix()
if _is_mmproj(f.name) or _is_mtp_drafter(rel):
continue
quant = _extract_quant_label(rel)
if quant != variant or _is_big_endian_gguf_path(rel, quant):
continue
matches.append(f)
matches.sort()
if matches:
return str(matches[0].resolve())
return None
def _detect_gguf_from_hf_cache(repo_id: str) -> Optional[str]:
"""Best GGUF filename for *repo_id* from the local HF cache, or None.
Excludes mmproj (vision projector) files so a partial cache holding only
the projector cannot route it as the main model.
"""
for snap in _iter_hf_cache_snapshots(repo_id):
rel_files = []
for f in _iter_gguf_files(snap, recursive = True):
rel = f.relative_to(snap).as_posix()
quant = _extract_quant_label(rel)
if _is_mmproj(f.name) or _is_mtp_drafter(rel) or _is_big_endian_gguf_path(rel, quant):
continue
rel_files.append(rel)
if rel_files:
return _pick_best_gguf(rel_files)
return None
def detect_gguf_model_remote(repo_id: str, hf_token: Optional[str] = None) -> Optional[str]:
"""Return the best GGUF filename in a HF repo, or None.
Retries (3 attempts, 1s/2s/4s backoff) on transient HF Hub failures: a
silent None would make the caller treat a GGUF-only repo as non-GGUF and
fall through to MLX on Apple Silicon. Offline falls back to the local cache.
"""
import time
from huggingface_hub import model_info as hf_model_info
if _env_offline():
cached = _detect_gguf_from_hf_cache(repo_id)
if cached is not None:
return cached
last_err: Optional[Exception] = None
for attempt in range(3):
try:
info = hf_model_info(repo_id, token = hf_token)
repo_files = []
for sibling in info.siblings:
fname = sibling.rfilename
if not fname.lower().endswith(".gguf"):
continue
quant = _extract_quant_label(fname)
if (
_is_mmproj(fname)
or _is_mtp_drafter(fname)
or _is_big_endian_gguf_path(fname, quant)
):
continue
repo_files.append(fname)
return _pick_best_gguf(repo_files)
except Exception as e:
last_err = e
# 404 / RepoNotFound is permanent -- don't retry
err_name = type(e).__name__
if err_name in (
"RepositoryNotFoundError",
"GatedRepoError",
"RevisionNotFoundError",
"EntryNotFoundError",
):
logger.debug(f"Could not check GGUF files for '{repo_id}': {e}")
return None
if attempt < 2:
time.sleep(2**attempt)
# All attempts failed; fall back to local cache for offline users.
cached = _detect_gguf_from_hf_cache(repo_id)
if cached is not None:
logger.warning(
"HF API unreachable for '%s' (%s); using local cache to detect GGUF.",
repo_id,
type(last_err).__name__ if last_err else "unknown",
)
return cached
logger.warning(f"Could not check GGUF files for '{repo_id}' after 3 attempts: {last_err}")
return None
def download_gguf_file(
repo_id: str,
filename: str,
hf_token: Optional[str] = None,
) -> str:
"""Download a specific GGUF file from a HF repo; returns the local path."""
from huggingface_hub import hf_hub_download
local_path = hf_hub_download(
repo_id = repo_id,
filename = filename,
token = hf_token,
)
return local_path
# Cache embedding detection per session to avoid repeated HF API calls
_embedding_detection_cache: Dict[tuple, bool] = {}
def is_embedding_model(model_name: str, hf_token: Optional[str] = None) -> bool:
"""Detect embedding/sentence-transformer models via HF metadata.
Combines three signals: "sentence-transformers" or "feature-extraction" in
tags, or pipeline_tag in {"sentence-similarity", "feature-extraction"}.
Catches models like gte-modernbert whose library_name is "transformers".
Args:
model_name: Model identifier (HF repo or local path)
hf_token: Optional HF token for gated/private models
Returns:
True if embedding model, else False (default for local paths or errors).
"""
cache_key = (model_name, hf_token)
if cache_key in _embedding_detection_cache:
return _embedding_detection_cache[cache_key]
# Local paths: check for sentence-transformer marker (modules.json)
if is_local_path(model_name):
local_dir = normalize_path(model_name)
is_emb = os.path.isfile(os.path.join(local_dir, "modules.json"))
_embedding_detection_cache[cache_key] = is_emb
return is_emb
try:
from huggingface_hub import model_info as hf_model_info
info = hf_model_info(model_name, token = hf_token)
tags = set(info.tags or [])
pipeline_tag = info.pipeline_tag or ""
is_emb = (
"sentence-transformers" in tags
or "feature-extraction" in tags
or pipeline_tag in ("sentence-similarity", "feature-extraction")
)
_embedding_detection_cache[cache_key] = is_emb
if is_emb:
logger.info(
f"Model {model_name} detected as embedding model: "
f"pipeline_tag={pipeline_tag}, "
f"sentence-transformers in tags={('sentence-transformers' in tags)}, "
f"feature-extraction in tags={('feature-extraction' in tags)}"
)
return is_emb
except Exception as e:
logger.warning(f"Could not determine if {model_name} is embedding model: {e}")
_embedding_detection_cache[cache_key] = False
return False
def _has_model_weight_files(model_dir: Path) -> bool:
"""Return True when a directory contains loadable model weights."""
for item in model_dir.iterdir():
if not item.is_file():
continue
suffix = item.suffix.lower()
if suffix == ".safetensors":
return True
if suffix == ".gguf":
return "mmproj" not in item.name.lower()
if suffix == ".bin":
name = item.name.lower()
if (
name.startswith("pytorch_model")
or name.startswith("model")
or name.startswith("adapter_model")
or name.startswith("consolidated")
):
return True
return False
def _detect_training_output_type(model_dir: Path) -> Optional[str]:
"""Classify a Studio training output as LoRA or full finetune."""
adapter_config = model_dir / "adapter_config.json"
adapter_model = model_dir / "adapter_model.safetensors"
if adapter_config.exists() or adapter_model.exists():
return "lora"
config_file = model_dir / "config.json"
if config_file.exists() and _has_model_weight_files(model_dir):
return "merged"
return None
def _looks_like_lora_adapter(model_dir: Path) -> bool:
return model_dir.is_dir() and (
(model_dir / "adapter_config.json").exists()
or any(model_dir.glob("adapter_model*.safetensors"))
or any(model_dir.glob("adapter_model*.bin"))
)
def scan_trained_models(outputs_dir: str = str(outputs_root())) -> List[Tuple[str, str, str]]:
"""Scan outputs folder for trained Studio models.
Returns:
List of (display_name, model_path, model_type), where model_type is
"lora" for adapter runs or "merged" for full finetunes.
"""
trained_models = []
outputs_path = resolve_output_dir(outputs_dir)
if not outputs_path.exists():
logger.warning(f"Outputs directory not found: {outputs_dir}")
return trained_models
try:
for item in outputs_path.iterdir():
if item.is_dir():
model_type = _detect_training_output_type(item)
if model_type is None:
continue
display_name = item.name
model_path = str(item)
trained_models.append((display_name, model_path, model_type))
logger.debug("Found trained model: %s (%s)", display_name, model_type)
# Sort by mtime, newest first
trained_models.sort(key = lambda x: Path(x[1]).stat().st_mtime, reverse = True)
logger.info(
"Found %s trained models in %s",
len(trained_models),
outputs_dir,
)
return trained_models
except Exception as e:
logger.error(f"Error scanning outputs folder: {e}")
return []
def scan_exported_models(
exports_dir: str = str(exports_root()),
) -> List[Tuple[str, str, str, Optional[str]]]:
"""Scan exports folder for exported models (merged, LoRA, GGUF).
Supports two layouts: two-level {run}/{checkpoint}/ (merged & LoRA) and
flat {name}-finetune-gguf/ (GGUF).
Returns:
List of (display_name, model_path, export_type, base_model), where
export_type is "lora" | "merged" | "gguf".
"""
results = []
exports_path = resolve_export_dir(exports_dir)
if not exports_path.exists():
return results
try:
for run_dir in exports_path.iterdir():
if not run_dir.is_dir():
continue
# Flat GGUF export (e.g. exports/gemma-3-4b-it-finetune-gguf/).
# Skip mmproj (vision projection) files — not loadable as main models.
gguf_files = [f for f in _iter_gguf_files(run_dir) if not _is_mmproj(f.name)]
if gguf_files:
base_model = None
export_meta = run_dir / "export_metadata.json"
try:
if export_meta.exists():
meta = json.loads(export_meta.read_text())
base_model = meta.get("base_model")
except Exception:
pass
display_name = run_dir.name
model_path = str(gguf_files[0])
results.append((display_name, model_path, "gguf", base_model))
logger.debug(f"Found GGUF export: {display_name}")
continue
# Two-level: {run}/{checkpoint}/
for checkpoint_dir in run_dir.iterdir():
if not checkpoint_dir.is_dir():
continue
adapter_config = checkpoint_dir / "adapter_config.json"
config_file = checkpoint_dir / "config.json"
has_weights = any(checkpoint_dir.glob("*.safetensors")) or any(
checkpoint_dir.glob("*.bin")
)
has_gguf = any(_iter_gguf_files(checkpoint_dir))
base_model = None
export_type = None
if adapter_config.exists():
export_type = "lora"
try:
cfg = json.loads(adapter_config.read_text())
base_model = cfg.get("base_model_name_or_path")
except Exception:
pass
elif config_file.exists() and has_weights:
export_type = "merged"
export_meta = checkpoint_dir / "export_metadata.json"
try:
if export_meta.exists():
meta = json.loads(export_meta.read_text())
base_model = meta.get("base_model")
except Exception:
pass
elif has_gguf:
export_type = "gguf"
gguf_list = list(_iter_gguf_files(checkpoint_dir))
# checkpoint_dir first, then run_dir (export.py writes
# metadata to the top-level export dir)
for meta_dir in (checkpoint_dir, run_dir):
export_meta = meta_dir / "export_metadata.json"
try:
if export_meta.exists():
meta = json.loads(export_meta.read_text())
base_model = meta.get("base_model")
if base_model:
break
except Exception:
pass
display_name = f"{run_dir.name} / {checkpoint_dir.name}"
model_path = str(gguf_list[0]) if gguf_list else str(checkpoint_dir)
results.append((display_name, model_path, export_type, base_model))
logger.debug(f"Found GGUF export: {display_name}")
continue
else:
continue
# Fallback: base model from ./outputs/{run_name}/adapter_config.json
if not base_model:
outputs_adapter_cfg = resolve_output_dir(run_dir.name) / "adapter_config.json"
try:
if outputs_adapter_cfg.exists():
cfg = json.loads(outputs_adapter_cfg.read_text())
base_model = cfg.get("base_model_name_or_path")
except Exception:
pass
display_name = f"{run_dir.name} / {checkpoint_dir.name}"
model_path = str(checkpoint_dir)
results.append((display_name, model_path, export_type, base_model))
logger.debug(f"Found exported model: {display_name} ({export_type})")
results.sort(key = lambda x: Path(x[1]).stat().st_mtime, reverse = True)
logger.info(f"Found {len(results)} exported models in {exports_dir}")
return results
except Exception as e:
logger.error(f"Error scanning exports folder: {e}")
return []
def get_base_model_from_checkpoint(checkpoint_path: str) -> Optional[str]:
"""Read the base model name from a local training or checkpoint directory."""
try:
checkpoint_path_obj = Path(checkpoint_path)
adapter_config_path = checkpoint_path_obj / "adapter_config.json"
if adapter_config_path.exists():
with open(adapter_config_path, "r") as f:
config = json.load(f)
base_model = config.get("base_model_name_or_path")
if base_model:
logger.info("Detected base model from adapter_config.json: %s", base_model)
return base_model
config_path = checkpoint_path_obj / "config.json"
if config_path.exists():
with open(config_path, "r") as f:
config = json.load(f)
for key in ("model_name", "_name_or_path"):
base_model = config.get(key)
if base_model and str(base_model) != str(checkpoint_path_obj):
logger.info(
"Detected base model from config.json (%s): %s",
key,
base_model,
)
return base_model
# TODO: torch.load default weights_only=True (torch >= 2.6) rejects pickled TrainingArguments; re-enable via safe_globals or weights_only=False once threat model allows.
# training_args_path = checkpoint_path_obj / "training_args.bin"
# if training_args_path.exists():
# try:
# import torch
#
# training_args = torch.load(training_args_path)
# if hasattr(training_args, "model_name_or_path"):
# base_model = training_args.model_name_or_path
# logger.info(
# "Detected base model from training_args.bin: %s", base_model
# )
# return base_model
# except Exception as e:
# logger.warning(f"Could not load training_args.bin: {e}")
dir_name = checkpoint_path_obj.name
if dir_name.startswith("unsloth_"):
parts = dir_name.split("_")
if len(parts) >= 2:
model_parts = parts[1:-1]
base_model = "unsloth/" + "_".join(model_parts)
logger.info("Detected base model from directory name: %s", base_model)
return base_model
logger.warning(f"Could not detect base model for checkpoint: {checkpoint_path}")
return None
except Exception as e:
logger.error(f"Error reading base model from checkpoint config: {e}")
return None
def get_base_model_from_lora(lora_path: str) -> Optional[str]:
"""Read the base model name from a LoRA adapter's config, or None."""
try:
lora_path_obj = Path(lora_path)
if not _looks_like_lora_adapter(lora_path_obj):
return None
# adapter_config.json first
adapter_config_path = lora_path_obj / "adapter_config.json"
if adapter_config_path.exists():
with open(adapter_config_path, "r") as f:
config = json.load(f)
base_model = config.get("base_model_name_or_path")
if base_model:
logger.info(f"Detected base model from adapter_config.json: {base_model}")
return base_model
# Fallback: try training_args.bin (requires torch)
# TODO: torch.load default weights_only=True (torch >= 2.6) rejects pickled TrainingArguments; also an remote code execution sink for third-party LoRAs via this route, re-enable behind a trust check if needed.
# training_args_path = lora_path_obj / "training_args.bin"
# if training_args_path.exists():
# try:
# import torch
#
# training_args = torch.load(training_args_path)
# if hasattr(training_args, "model_name_or_path"):
# base_model = training_args.model_name_or_path
# logger.info(
# f"Detected base model from training_args.bin: {base_model}"
# )
# return base_model
# except Exception as e:
# logger.warning(f"Could not load training_args.bin: {e}")
# Last resort: parse from dir name (unsloth_<model>_<timestamp>)
dir_name = lora_path_obj.name
if dir_name.startswith("unsloth_"):
parts = dir_name.split("_")
if len(parts) >= 2:
model_parts = parts[1:-1] # Skip "unsloth" and timestamp
base_model = "unsloth/" + "_".join(model_parts)
logger.info(f"Detected base model from directory name: {base_model}")
return base_model
logger.warning(f"Could not detect base model for LoRA: {lora_path}")
return None
except Exception as e:
logger.error(f"Error reading base model from LoRA config: {e}")
return None
def get_base_model_from_lora_identifier(
identifier: str, hf_token: Optional[str] = None
) -> Optional[str]:
"""Resolve a LoRA adapter's base model for a LOCAL dir OR a REMOTE HF repo.
``get_base_model_from_lora`` only reads a local adapter directory (it requires
``is_dir()``). The SECURITY gates must also follow a *remote* adapter's base,
because the base model's code / weights are what execute on load: an attacker's
adapter repo can point ``base_model_name_or_path`` at a base carrying a poisoned
pickle or HIGH auto_map code. For a remote repo id we fetch ONLY the small
``adapter_config.json`` (metadata; never a weight file) and read the base. Use
this in the gate paths so a remote LoRA base is scanned, not just the adapter.
Returns the base model id, or ``None`` when the identifier is not a LoRA adapter
or the base cannot be determined (the caller still scans the identifier itself).
A genuine 404 (no ``adapter_config.json`` / repo absent) is distinguished from a
transient error: the latter is retried once, then logged as a WARNING (a missed
base would be scanned by neither gate), so a network blip does not silently and
invisibly skip the base.
"""
# Local path: reuse the existing directory reader (identical behavior).
try:
if is_local_path(identifier):
return get_base_model_from_lora(identifier)
except Exception:
return get_base_model_from_lora(identifier)
# Remote repo id: read base_model_name_or_path from adapter_config.json only.
from huggingface_hub import hf_hub_download
from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError
last_exc = None
for _attempt in range(2): # one retry: a transient blip must not skip the base
try:
cfg_path = hf_hub_download(
identifier, "adapter_config.json", token = hf_token if hf_token else None
)
except (EntryNotFoundError, RepositoryNotFoundError):
# No adapter_config.json -> not a resolvable LoRA; caller scans the identifier.
return None
except Exception as exc: # transient / auth / network -> retry once
last_exc = exc
continue
try:
with open(cfg_path, "r") as f:
base_model = json.load(f).get("base_model_name_or_path")
except Exception as exc:
logger.warning("Could not parse adapter_config.json for '%s': %s", identifier, exc)
return None
if base_model:
logger.info(
"Detected base model from remote adapter_config.json (%s): %s",
identifier,
base_model,
)
return base_model # may be None if the key is absent (still a valid answer)
# Both attempts failed transiently: log loudly -- a missed base is gated by neither gate.
logger.warning(
"Could not resolve remote LoRA base for '%s' after retry (%s); its base, if "
"any, will not be added to the security scan targets.",
identifier,
type(last_exc).__name__ if last_exc else "unknown",
)
return None
# Status indicators that appear in UI dropdowns
UI_STATUS_INDICATORS = [" (Ready)", " (Loading...)", " (Active)", ""]
def load_model_defaults(model_name: str) -> Dict[str, Any]:
"""Load default training parameters for a model from a YAML file.
Looks in configs/model_defaults/ (incl. subfolders) by model name or its
MODEL_NAME_MAPPING aliases, else falls back to default.yaml. Returns the
parameter dict, or {} if none found.
"""
try:
script_dir = Path(__file__).parent.parent.parent
defaults_dir = script_dir / "assets" / "configs" / "model_defaults"
# Check the mapping first
if model_name.lower() in _REVERSE_MODEL_MAPPING:
canonical_file = _REVERSE_MODEL_MAPPING[model_name.lower()]
for config_path in defaults_dir.rglob(canonical_file):
if config_path.is_file():
with open(config_path, "r", encoding = "utf-8") as f:
config = yaml.safe_load(f) or {}
logger.info(f"Loaded model defaults from {config_path} (via mapping)")
return config
# For local paths (e.g. /home/.../Spark-TTS-0.5B/LLM from
# adapter_config.json, or C:\Users\...\model on Windows), match the
# last 1-2 path components against the registry (e.g. "Spark-TTS-0.5B/LLM").
_is_local_path = is_local_path(model_name)
# Normalize Windows backslash paths so Path().parts splits correctly
# on POSIX/WSL hosts (pathlib treats backslashes as literals on Linux).
_normalized = normalize_path(model_name) if _is_local_path else model_name
if model_name.lower() not in _REVERSE_MODEL_MAPPING and _is_local_path:
parts = Path(_normalized).parts
for depth in [2, 1]:
if len(parts) >= depth:
suffix = "/".join(parts[-depth:])
if suffix.lower() in _REVERSE_MODEL_MAPPING:
canonical_file = _REVERSE_MODEL_MAPPING[suffix.lower()]
for config_path in defaults_dir.rglob(canonical_file):
if config_path.is_file():
with open(config_path, "r", encoding = "utf-8") as f:
config = yaml.safe_load(f) or {}
logger.info(
f"Loaded model defaults from {config_path} (via path suffix '{suffix}')"
)
return config
# Exact model name match (backward compatibility). For local paths,
# use only the dir basename to avoid passing absolute paths (e.g.
# C:\...) into rglob, which raises "Non-relative patterns are
# unsupported" on Windows.
_lookup_name = Path(_normalized).name if _is_local_path else model_name
model_filename = _lookup_name.replace("/", "_") + ".yaml"
# Search subfolders and root
for config_path in defaults_dir.rglob(model_filename):
if config_path.is_file():
with open(config_path, "r", encoding = "utf-8") as f:
config = yaml.safe_load(f) or {}
logger.info(f"Loaded model defaults from {config_path}")
return config
# Fall back to default.yaml
default_config_path = defaults_dir / "default.yaml"
if default_config_path.exists():
with open(default_config_path, "r", encoding = "utf-8") as f:
config = yaml.safe_load(f) or {}
logger.info(f"Loaded default model defaults from {default_config_path}")
return config
logger.warning(f"No default config found for model {model_name}")
return {}
except Exception as e:
logger.error(f"Error loading model defaults for {model_name}: {e}")
return {}
@dataclass
class ModelConfig:
"""Configuration for a model to load."""
identifier: str # Clean model identifier (org/name or path)
display_name: str # Original UI display name
path: str # Normalized filesystem path
is_local: bool # Local file vs HF model?
is_cached: bool # Already in HF cache?
is_vision: bool # Vision model?
is_lora: bool # LoRA adapter?
is_gguf: bool = False # GGUF model?
is_audio: bool = False # TTS audio model?
audio_type: Optional[str] = None # Audio codec type: 'snac', 'csm', 'bicodec', 'dac'
has_audio_input: bool = False # Accepts audio input (ASR/speech understanding)
gguf_file: Optional[str] = None # Full path to the .gguf file (local mode)
gguf_mmproj_file: Optional[str] = None # Full path to the mmproj .gguf file (vision projection)
gguf_mtp_file: Optional[str] = None # Full path to the separate MTP drafter (local mode)
gguf_hf_repo: Optional[str] = (
None # HF repo ID for -hf mode (e.g. "unsloth/gemma-3-4b-it-GGUF")
)
gguf_variant: Optional[str] = None # Quantization variant (e.g. "Q4_K_M")
base_model: Optional[str] = None # Base model (for LoRAs)
@classmethod
def from_lora_path(
cls,
lora_path: str,
hf_token: Optional[str] = None,
) -> Optional["ModelConfig"]:
"""Create ModelConfig from a local LoRA adapter path, auto-detecting the
base model from adapter config.
Args:
lora_path: Path to the LoRA adapter directory
hf_token: HF token for vision detection
"""
try:
lora_path_obj = Path(lora_path)
if not lora_path_obj.exists():
logger.error(f"LoRA path does not exist: {lora_path}")
return None
base_model = get_base_model_from_lora(lora_path)
if not base_model:
logger.error(f"Could not determine base model for LoRA: {lora_path}")
return None
is_vision = is_vision_model(base_model, hf_token = hf_token)
audio_type = detect_audio_type(base_model, hf_token = hf_token)
display_name = lora_path_obj.name
identifier = lora_path # path is the identifier for local LoRAs
return cls(
identifier = identifier,
display_name = display_name,
path = lora_path,
is_local = True,
is_cached = True, # local LoRAs are always cached
is_vision = is_vision,
is_lora = True,
is_audio = audio_type is not None and audio_type != "audio_vlm",
audio_type = audio_type,
has_audio_input = is_audio_input_type(audio_type),
base_model = base_model,
)
except Exception as e:
logger.error(f"Error creating ModelConfig from LoRA path: {e}")
return None
@classmethod
def from_identifier(
cls,
model_id: str,
hf_token: Optional[str] = None,
is_lora: bool = False,
gguf_variant: Optional[str] = None,
) -> Optional["ModelConfig"]:
"""Create ModelConfig from a clean model identifier (HF repo or local
path), for FastAPI routes that send sanitized paths.
Args:
model_id: Clean model identifier (HF repo name or local path)
hf_token: Optional HF token for vision detection on gated models
is_lora: Whether this is a LoRA adapter
gguf_variant: Optional GGUF quant variant (e.g. "Q4_K_M") to load
via -hf for remote repos; None auto-selects via _pick_best_gguf().
Returns:
ModelConfig or None if it cannot be created.
"""
if not model_id or not model_id.strip():
return None
identifier = model_id.strip()
is_local = is_local_path(identifier)
path = normalize_path(identifier) if is_local else identifier
# Add unsloth/ prefix for shorthand HF models
if not is_local and "/" not in identifier:
identifier = f"unsloth/{identifier}"
path = identifier
# Reuse a cached case-variant's exact repo_id spelling to avoid
# one-time re-downloads after #2592.
if not is_local:
resolved_identifier = resolve_cached_repo_id_case(identifier)
if resolved_identifier != identifier:
logger.info(
"Using cached repo_id casing '%s' for requested '%s'",
resolved_identifier,
identifier,
)
identifier = resolved_identifier
path = resolved_identifier
# Auto-detect GGUF models (check before LoRA/vision detection)
if is_local:
if gguf_variant:
gguf_file = _find_local_gguf_by_variant(path, gguf_variant)
else:
gguf_file = detect_gguf_model(path)
if gguf_file:
display_name = Path(gguf_file).stem
logger.info(f"Detected local GGUF model: {gguf_file}")
# Vision: check base model, then look for mmproj
mmproj_file = None
gguf_is_vision = False
gguf_dir = Path(gguf_file).parent
# Is this a vision model, per export metadata?
base_is_vision = False
meta_path = gguf_dir / "export_metadata.json"
if meta_path.exists():
try:
meta = json.loads(meta_path.read_text())
base = meta.get("base_model")
if base and is_vision_model(base, hf_token = hf_token):
base_is_vision = True
logger.info(f"GGUF base model '{base}' is a vision model")
except Exception as e:
logger.debug(f"Could not read export metadata: {e}")
# Direct file selections may point into a quant subdir while
# mmproj-*.gguf lives at the snapshot root.
companion_root = _local_gguf_companion_search_root(path, gguf_file)
mmproj_file = detect_mmproj_file(gguf_file, search_root = companion_root)
if mmproj_file:
gguf_is_vision = True
logger.info(f"Detected mmproj for vision: {mmproj_file}")
elif base_is_vision:
logger.warning(f"Base model is vision but no mmproj file found in {gguf_dir}")
# Separate MTP drafter sibling (Gemma 4), mirroring mmproj.
mtp_file = detect_mtp_file(gguf_file, search_root = companion_root)
if mtp_file:
logger.info(f"Detected MTP drafter: {mtp_file}")
return cls(
identifier = identifier,
display_name = display_name,
path = path,
is_local = True,
is_cached = True,
is_vision = gguf_is_vision,
is_lora = False,
is_gguf = True,
gguf_file = gguf_file,
gguf_mmproj_file = mmproj_file,
gguf_mtp_file = mtp_file,
)
else:
# Does the HF repo contain GGUF files?
gguf_filename = detect_gguf_model_remote(identifier, hf_token = hf_token)
if gguf_filename:
# Preflight: verify llama-server binary exists before a multi-GB
# download. include_denied: a transiently locked binary still
# exists (the lock clears long before the download finishes; the
# load itself reports a still-locked binary distinctly).
from core.inference.llama_cpp import LlamaCppBackend
if not LlamaCppBackend._find_llama_server_binary(include_denied = True):
raise RuntimeError(
"llama-server binary not found — cannot load GGUF models. "
"Run setup.sh to build it, or set LLAMA_SERVER_PATH."
)
# list_gguf_variants() detects vision & resolves the variant
variants, has_vision = list_gguf_variants(identifier, hf_token = hf_token)
variant = gguf_variant
if not variant: # auto-select best quant
variant_filenames = [v.filename for v in variants]
best = _pick_best_gguf(variant_filenames)
if best:
variant = _extract_quant_label(best)
else:
variant = "Q4_K_M" # Fallback — llama-server's own default
display_name = f"{identifier.split('/')[-1]} ({variant})"
logger.info(
f"Detected remote GGUF repo '{identifier}', "
f"variant={variant}, vision={has_vision}"
)
return cls(
identifier = identifier,
display_name = display_name,
path = identifier,
is_local = False,
is_cached = False,
is_vision = has_vision,
is_lora = False,
is_gguf = True,
gguf_file = None,
gguf_hf_repo = identifier,
gguf_variant = variant,
)
# Auto-detect LoRA for local paths (adapter_config.json on disk)
if not is_lora and is_local:
detected_base = (
get_base_model_from_lora(path) if _looks_like_lora_adapter(Path(path)) else None
)
if detected_base:
is_lora = True
logger.info(f"Auto-detected local LoRA adapter at '{path}' (base: {detected_base})")
# Auto-detect LoRA for remote HF models. When offline, huggingface_hub
# raises OfflineModeIsEnabled in ~0ms; we fall through to the cache.
if not is_lora and not is_local:
try:
from huggingface_hub import model_info as hf_model_info
info = hf_model_info(identifier, token = hf_token)
repo_files = [s.rfilename for s in info.siblings]
if "adapter_config.json" in repo_files:
is_lora = True
logger.info(f"Auto-detected remote LoRA adapter: '{identifier}'")
except Exception as e:
logger.debug(f"Could not check remote LoRA status for '{identifier}': {e}")
# API may have failed; adapter_config.json could still be cached.
if not is_lora:
for snap in _iter_hf_cache_snapshots(identifier):
if (snap / "adapter_config.json").is_file():
is_lora = True
logger.info(f"Auto-detected cached LoRA adapter: '{identifier}'")
break
# Handle LoRA adapters
base_model = None
if is_lora:
if is_local:
# Local LoRA: read adapter_config.json from disk
base_model = get_base_model_from_lora(path)
else:
# Remote LoRA: fetch adapter_config.json from HF
try:
from huggingface_hub import hf_hub_download
config_path = hf_hub_download(identifier, "adapter_config.json", token = hf_token)
with open(config_path, "r") as f:
adapter_config = json.load(f)
base_model = adapter_config.get("base_model_name_or_path")
if base_model:
logger.info(f"Resolved remote LoRA base model: '{base_model}'")
except Exception as e:
logger.warning(
f"Could not download adapter_config.json for '{identifier}': {e}"
)
if not base_model:
logger.warning(f"Could not determine base model for LoRA '{path}'")
return None
check_model = base_model
else:
check_model = identifier
vision = is_vision_model(check_model, hf_token = hf_token)
audio_type_val = detect_audio_type(check_model, hf_token = hf_token)
has_audio_in = is_audio_input_type(audio_type_val)
display_name = Path(path).name if is_local else identifier.split("/")[-1]
return cls(
identifier = identifier,
display_name = display_name,
path = path,
is_local = is_local,
is_cached = is_model_cached(identifier) if not is_local else True,
is_vision = vision,
is_lora = is_lora,
is_audio = audio_type_val is not None and audio_type_val != "audio_vlm",
audio_type = audio_type_val,
has_audio_input = has_audio_in,
base_model = base_model,
)
@classmethod
def from_ui_selection(
cls,
dropdown_value: Optional[str],
search_value: Optional[str],
local_models: list = None,
hf_token: Optional[str] = None,
is_lora: bool = False,
) -> Optional["ModelConfig"]:
"""Create a ModelConfig from UI dropdown/search selections (base models and LoRAs)."""
selected = None
if search_value and search_value.strip():
selected = search_value.strip()
elif dropdown_value:
selected = dropdown_value
if not selected:
return None
display_name = selected
# Resolve display names via the 'local_models' parameter
if " (Active)" in selected or " (Ready)" in selected:
clean_display_name = selected.replace(" (Active)", "").replace(" (Ready)", "")
if local_models:
for local_display, local_path in local_models:
if local_display == clean_display_name:
selected = local_path
break
# Strip all UI status indicators to get the final identifier
identifier = selected
for status in UI_STATUS_INDICATORS:
identifier = identifier.replace(status, "")
identifier = identifier.strip()
is_local = is_local_path(identifier)
path = normalize_path(identifier) if is_local else identifier
# Add unsloth/ prefix for shorthand HF models
if not is_local and "/" not in identifier:
identifier = f"unsloth/{identifier}"
path = identifier
if not is_local:
resolved_identifier = resolve_cached_repo_id_case(identifier)
if resolved_identifier != identifier:
identifier = resolved_identifier
path = resolved_identifier
# Keep existing local GGUF selections on the llama-server path. This
# constructor is still used by older inference helpers and must not
# describe a .gguf weight file as loadable by FastVisionModel.
if is_local and not is_lora and detect_gguf_model(path):
gguf_config = cls.from_identifier(path, hf_token = hf_token)
if gguf_config is not None:
gguf_config.display_name = display_name
return gguf_config
# --- Base Model and Vision Detection ---
base_model = None
is_vision = False
if is_lora:
# A LoRA MUST have a base model.
base_model = get_base_model_from_lora(path)
if not base_model:
logger.warning(
f"Could not determine base model for LoRA '{path}'. Cannot create config."
)
return None # cannot proceed without a base model
# A LoRA's vision capability comes from its base model.
is_vision = is_vision_model(base_model, hf_token = hf_token)
else:
# Base model: check its own vision status.
is_vision = is_vision_model(identifier, hf_token = hf_token)
from utils.paths import is_model_cached
is_cached = is_model_cached(identifier) if not is_local else True
return cls(
identifier = identifier,
display_name = display_name,
path = path,
is_local = is_local,
is_cached = is_cached,
is_vision = is_vision,
is_lora = is_lora,
base_model = base_model, # None for base models, set for LoRAs
)