* studio: stream export worker output into the export dialog
The Export Model dialog only showed a spinner on the "Exporting..."
button while the worker subprocess was doing the actual heavy lifting.
For Merged to 16bit and GGUF / Llama.cpp exports this meant several
minutes (or more, for large models) of opaque silence, with no way to
tell whether save_pretrained_merged, convert_hf_to_gguf.py, or
llama-quantize was making progress.
This adds a live terminal-style output panel inside the export dialog,
rendered just above the Cancel / Start Export buttons and scrollable
with auto-follow-tail. It shows stdout and stderr from both the worker
process itself and any child process it spawns (GGUF converter,
llama-quantize), coloured by stream.
Backend
- core/export/worker.py: new _setup_log_capture(resp_queue) installed
before LogConfig.setup_logging. It saves the original stdout/stderr
fds, creates pipes, os.dup2's the write ends onto fds 1 and 2 (so
every child process inherits the redirected fds), and spins up two
daemon reader threads. Each thread reads bytes from a pipe, echoes
them back to the original fd (so the server console keeps working),
splits on \n and \r, and forwards each line to the resp queue as
{"type":"log","stream":"stdout|stderr","line":...,"ts":...}.
PYTHONUNBUFFERED=1 is set so nested Python converters flush
immediately.
- core/export/orchestrator.py:
- Thread-safe ring buffer (collections.deque, maxlen 4000) with a
monotonically increasing seq counter. clear_logs(),
get_logs_since(cursor), get_current_log_seq(), is_export_active().
- _wait_response handles rtype == "log" by appending to the buffer
and continuing the wait loop. Status messages are also surfaced as
a "status" stream so users see high level progress alongside raw
subprocess output.
- load_checkpoint, _run_export, and cleanup_memory now wrap their
bodies with the existing self._lock (previously unused), clear the
log buffer at the start of each op, and flip _export_active in a
try/finally so the SSE endpoint can detect idle.
- routes/export.py:
- Wrapped every sync orchestrator call (load_checkpoint,
cleanup_memory, export_merged_model, export_base_model,
export_gguf, export_lora_adapter) in asyncio.to_thread so the
FastAPI event loop stays free during long exports. Without this
the new SSE endpoint could not be served concurrently with the
blocking export POST.
- New GET /api/export/logs/stream SSE endpoint. Honors
Last-Event-ID and a since query param for reconnect, emits log /
heartbeat / complete / error events, uses the id field to carry
the log seq so clients can resume cleanly. On first connect
without an explicit cursor it starts from the current seq so old
lines from a previous run are not replayed.
Frontend
- features/export/api/export-api.ts: streamExportLogs() helper that
authFetches the SSE endpoint and parses id / event / data fields
manually (same pattern as streamTrainingProgress in train-api.ts).
- features/export/components/export-dialog.tsx:
- Local useExportLogs(exporting) hook that opens the SSE stream on
exporting transitions to true, accumulates up to 4000 lines in
component state, and aborts on cleanup.
- New scrollable output panel rendered above DialogFooter, only
shown for Merged to 16bit and GGUF / Llama.cpp (LoRA adapter is
a fast disk write with nothing to show). Dark terminal styling
(bg-black/85, emerald text, rose for stderr, sky for status),
max-height 14rem, auto-scrolls to the bottom on new output but
stops following if the user scrolls up. A small streaming / idle
indicator is shown next to the panel title.
- DialogContent widens from sm:max-w-lg to sm:max-w-2xl when the
output panel is visible so the logs have room to breathe.
Verified
- Python smoke test (tests/smoke_export_log_capture.py): spawns a
real mp.get_context("spawn") process, installs _setup_log_capture,
confirms that parent stdout prints, parent stderr prints, AND a
child subprocess invoked via subprocess.run (both its stdout and
stderr) are all captured in the resp queue. Passes.
- Orchestrator log helpers tested in isolation: _append_log,
get_logs_since (with and without a cursor), clear_logs not
resetting seq so reconnecting clients still progress. Passes.
- routes.export imports cleanly in the studio venv and /logs/stream
shows up in router.routes.
- bun run build: tsc -b plus vite build, no TypeScript errors.
No existing export behavior is changed. If the subprocess, the SSE
endpoint, or the frontend hook fails, the export itself still runs to
completion the same way it did before, with or without logs visible.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* export dialog: trim bootstrap noise, scope logs per screen, show realpath
Several follow-ups to the live export log work:
1. Worker bootstrap noise (transformers venv activation, Unsloth banner,
"Top GGUF/hub models" lists, vision detection, 2k-step weight load
bar) is dropped from the export-dialog stream. A threading.Event
gate in worker.py defaults closed and only opens once _handle_export
actually starts; until then the reader thread still echoes lines to
the saved console fd for debugging but does not push them onto the
resp_queue. The orchestrator already spawns a fresh subprocess for
every checkpoint load, so the gate is naturally reset between runs.
2. tqdm in non-tty mode defaults to a 10s mininterval, which makes
multi-step bars look frozen in the panel. Set TQDM_MININTERVAL=0.5
in the worker env so any tqdm-driven progress emits more often.
3. The dialog's useExportLogs hook now also clears its line buffer
when exportMethod or open changes, so re-opening the dialog into a
different action's screen no longer shows the previous action's
saved output. A useElapsedSeconds tick + "Working Xs" badge in the
log header gives users a visible sign that long single-step phases
(cache copies, GGUF conversion) are still running when no new lines
are arriving.
4. ExportBackend.export_{merged,base,gguf,lora} now return
(success, message, output_path); the worker forwards output_path on
each export_*_done response, the orchestrator's _run_export passes
it to routes/export.py, which surfaces it via
ExportOperationResponse.details.output_path. The dialog's Export
Complete screen renders the resolved on-disk realpath under "Saved
to" so users can find their exported model directly.
* fix(cli): unpack 3-tuple return from export backend
ExportOrchestrator.export_{merged,base,gguf,lora} now return
(success, message, output_path) so the studio dialog can show
the on-disk realpath. The CLI still unpacked 2 values, so every
`unsloth export --format ...` crashed with ValueError before
reporting completion. Update the four call sites and surface
output_path via a "Saved to:" echo.
* fix(studio): anchor export log SSE cursor at run start
The export dialog SSE defaulted its cursor to get_current_log_seq()
at connect time, so any line emitted between the POST that kicks
off the export and the client opening the stream was buffered with
seqs 1..k and then skipped (seq <= cursor). Long-running exports
looked silent during their first seconds.
Snapshot _log_seq into _run_start_seq inside clear_logs() and
expose it via get_run_start_seq(). The SSE default cursor now uses
that snapshot, so every line emitted since the current run began
is reachable regardless of when the client connects. Old runs
still can't leak in because their seqs are <= the snapshot.
* fix(studio): reconnect export log SSE on stream drop
useExportLogs launched streamExportLogs once per exporting
transition and recorded any drop in .catch(). Long GGUF exports
behind a proxy with an idle kill-timeout would silently lose the
stream for the rest of the run even though the backend already
supports Last-Event-ID resume. The "retry: 3000" directive emitted
by the backend is only meaningful to native EventSource; this
hook uses a manual fetch + ReadableStream parse so it had no
effect.
Wrap streamExportLogs in a retry loop that tracks lastSeq from
ExportLogEvent.id and passes it as since on reconnect. Backoff is
exponential with jitter, capped at 5s, reset on successful open.
The loop stops on explicit backend `complete` event or on effect
cleanup.
* fix(studio): register a second command so Typer keeps `export` as a subcommand
The CLI export unpacking tests wrap `unsloth_cli.commands.export.export`
in a fresh Typer app with a single registered command. Typer flattens a
single-command app into that command, so the test's
`runner.invoke(cli_app, ["export", ckpt, out, ...])` treats the leading
`"export"` token as an unexpected extra positional argument -- every
parametrized case failed with:
Got unexpected extra argument (.../out)
Register a harmless `noop` second command so Typer preserves subcommand
routing and the tests actually exercise the 3-tuple unpack path they
were written to guard.
Before: 4 failed
After: 4 passed
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: studio-install <studio@local.install>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Roland Tannous <rolandtannous@gravityq.ai>
* Studio: anchor ctx-slider warning threshold at 4096 when weights exceed VRAM
The chat settings sheet's ctx slider reads `max_context_length` from
`/api/inference/status` and renders
Exceeds estimated VRAM capacity (N tokens). The model may use
system RAM.
when the user drags the slider above that value. For models whose
weights fit on some GPU subset, `_max_context_length` was already set
to the binary-search cap and the warning fired correctly.
For models whose weights exceed 90% of every GPU subset's free memory
(e.g. MiniMax-M2.7-GGUF at 131 GB on a 97 GB GPU), the ceiling-probe
loop never matched a subset, so `max_available_ctx` stayed at the
native context (e.g. 196608). The slider ran all the way to native
with no indication that any value above the 4096 spec default would
trigger `--fit on` and degrade performance.
Anchor `max_available_ctx` at `min(4096, native_context_length)` when
no subset fits, so the warning fires at the right threshold and the
user sees the correct safe-zone / warning-zone split:
Before (MiniMax-M2.7 on 97 GB GPU):
slider 0 .. 196608, warning threshold = 196608 (never fires)
After:
slider 0 .. 196608, warning threshold = 4096 (fires correctly)
No frontend changes required: `chat-settings-sheet.tsx` already
consumes `ggufMaxContextLength` (= status.max_context_length) as the
warning threshold and `ggufNativeContextLength` as the slider max.
Adds tests/test_llama_cpp_max_context_threshold.py covering
weights-exceed-VRAM (single / multi-GPU), a native-ctx below the 4096
fallback case (don't lie about supported ctx), fittable-model
regressions (small / multi-GPU / tiny on huge GPU), and the
`max_context_length` property's fallback semantics.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: make GGUF disk-space preflight cache-aware
The pre-download disk check in LlamaCppBackend.load_model compared the
repo's total GGUF size against free disk without crediting bytes
already present in the Hugging Face cache. Re-loading a large cached
model (e.g. MiniMax-M2.7-GGUF at 131 GB) then failed cold with
"Not enough disk space to download any variant" whenever free disk
was below the full weight footprint, even though nothing actually
needed to be downloaded.
Subtract bytes already on disk via try_to_load_from_cache before
comparing against free space. A partial blob (interrupted download) is
not credited, so a second attempt still allocates room to finish the
download. The log line now also surfaces how much is already cached.
Adds tests/test_llama_cpp_cache_aware_disk_check.py covering the
fully-cached, partial-cache-insufficient-disk, partial-cache-enough-disk,
cold-cache, incomplete-blob, and zero-size-path-info cases. Sparse
tempfiles keep the GB-scale scenarios cheap to simulate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: honor explicit GGUF ctx and default to 4096 when weights exceed VRAM
The load-time auto-fit in LlamaCppBackend.load_model had two issues for
models whose weights do not fit on any GPU subset (the common case for
large MoE GGUFs such as MiniMax-M2.7, Qwen3.5-397B-A17B, etc.):
1. Auto mode (max_seq_length=0) left effective_ctx at the model's native
context when no subset passed the 90% fit check. The UI slider then
landed on e.g. 196608 for MiniMax-M2.7, far above anything usable.
Default the auto-pick to 4096 so the UI starts at a sane value; the
slider ceiling stays at the native context so the user can still
opt in to longer contexts and receive the "might be slower" warning.
2. Explicit ctx was silently shrunk when weights fit but the requested
KV overflowed the 90% budget. The shrink loop emitted -c <capped>
-ngl -1 without informing the caller, so a user who had opted into
a longer context via the UI never actually got it. Drop the shrink
loop on the explicit path and emit -c <user_ctx> --fit on instead,
letting llama-server flex -ngl (CPU layer offload).
Adds tests/test_llama_cpp_context_fit.py covering both paths, the
file-size-only fallback when KV metadata is missing, non-regression on
fittable auto-pick, and platform-agnostic input shape.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* [Studio] Install flash attn at setup time for linux
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* cleanup changes
Signed-off-by: Datta Nimmaturi <venkatadattasainimmaturi@gmail.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Test cases
* wheel_utils: narrow url_exists exceptions and log at debug level
---------
Signed-off-by: Datta Nimmaturi <venkatadattasainimmaturi@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: Roland Tannous <rolandtannous@gravityq.ai>
* Show non exported models in chat UI
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Distinguish b/w LoRa and full fine tune saves. Cleanup
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
* Studio: add API key authentication for programmatic access
External users want to hit the Studio API (chat completions with tool
calling, training, export, etc.) without going through the browser
login flow. This adds sk-unsloth- prefixed API keys that work as a
drop-in replacement for JWTs in the Authorization: Bearer header.
Backend:
- New api_keys table in SQLite (storage.py)
- create/list/revoke/validate functions with SHA-256 hashed storage
- API key detection in _get_current_subject before the JWT path
- POST/GET/DELETE /api/auth/api-keys endpoints on the auth router
Frontend:
- /api-keys page with create form, one-time key reveal, keys table
- API Keys link in desktop and mobile navbar
- Route registered with requireAuth guard
Zero changes to any existing route handler -- every endpoint that uses
Depends(get_current_subject) automatically works with API keys.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use actual origin in API key usage examples
The examples on /api-keys were hardcoded to localhost:8888 which is
wrong for remote users. Use window.location.origin so the examples
show the correct URL regardless of where the user is connecting from.
* Add `unsloth studio run` CLI command for one-liner model serving
Adds a `run` subcommand that starts Studio, loads a model, creates an
API key, and prints a ready-to-use curl command -- similar to
`ollama run` or `vllm serve`.
Usage: unsloth studio run -m unsloth/Qwen3-1.7B-GGUF --gguf-variant UD-Q4_K_XL
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add end-to-end tests for `unsloth studio run` and API key usage
Tests the 4 usage examples from the API Keys page:
1. curl basic (non-streaming) chat completions
2. curl streaming (SSE) chat completions
3. OpenAI Python SDK streaming completions
4. curl with tools (web_search + python)
Also tests --help output, invalid key rejection, and no-key rejection.
All 7 tests pass against Qwen3-1.7B-GGUF.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add /v1/completions, /v1/embeddings, /v1/responses endpoints and --parallel support
- llama_cpp.py: accept n_parallel param, pass to llama-server --parallel
- run.py: plumb llama_parallel_slots through to app.state
- inference.py: add /completions and /embeddings as transparent proxies to
llama-server, add /responses as application-level endpoint that converts
to ChatCompletionRequest; thread n_parallel through load_model
- studio.py: set llama_parallel_slots=4 for `unsloth studio run` path
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make /v1/responses endpoint match OpenAI Responses API format
The existing /v1/responses shim returned Chat Completions format, which
broke OpenAI SDK clients using openai.responses.create(). This commit
replaces the endpoint with a proper implementation that:
- Returns `output` array with `output_text` content parts instead of
`choices` with `message`
- Uses `input_tokens`/`output_tokens` instead of `prompt_tokens`/
`completion_tokens` in usage
- Sets `object: "response"` and `id: "resp_..."`
- Emits named SSE events for streaming (response.created,
response.output_text.delta, response.completed, etc.)
- Accepts all OpenAI Responses API fields (tools, store, metadata,
previous_response_id) without erroring -- silently ignored
- Maps `developer` role to `system` and `input_text`/`input_image`
content parts to the internal Chat format
Adds Pydantic schemas for request/response models and 23 unit tests
covering schema validation, input normalisation, and response format.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add Anthropic-compatible /v1/messages endpoint (#4981)
* Add Anthropic-compatible /v1/messages endpoint with tool support
Translate Anthropic Messages API format to/from internal OpenAI format
and reuse the existing server-side agentic tool loop. Supports streaming
SSE (message_start, content_block_delta, etc.) and non-streaming JSON.
Includes offline unit tests and e2e tests in test_studio_run.py.
* Add enable_tools, enabled_tools, session_id to /v1/messages endpoint
Support the same shorthand as /v1/chat/completions: enable_tools=true
with an optional enabled_tools list uses built-in server tools without
requiring full Anthropic tool definitions. session_id is passed through
for sandbox isolation. max_tokens is now optional.
* Strip leaked tool-call XML from Anthropic endpoint content
Apply _TOOL_XML_RE to content events in both streaming and
non-streaming tool paths, matching the OpenAI endpoint behavior.
* Emit custom tool_result SSE event in Anthropic stream
Adds a non-standard tool_result event between the tool_use block close
and the next text block, so clients can see server-side tool execution
results. Anthropic SDKs ignore unknown event types.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Split /v1/messages into server-side and client-side tool paths
enable_tools=true runs the existing server-side agentic loop with
built-in tools (web_search/python/terminal). A bare tools=[...] field
now triggers a client-side pass-through: client-provided tools are
forwarded to llama-server and any tool_use output is returned to the
caller with stop_reason=tool_use for client execution.
This fixes Claude Code (and any Anthropic SDK client) which sends
tools=[...] expecting client-side execution but was previously routed
through execute_tool() and failing with 'Unknown tool'.
Adds AnthropicPassthroughEmitter to convert llama-server OpenAI SSE
chunks into Anthropic SSE events, plus unit tests covering text
blocks, tool_use blocks, mixed, stop reasons, and usage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix httpcore GeneratorExit in /v1/messages passthrough stream
Explicitly aclose aiter_lines() before the surrounding async with
blocks unwind, mirroring the prior fix in external_provider.py
(a41160d3) and cc757b78's RuntimeError suppression.
* Wire stop_sequences through /v1/messages; warn on tool_choice
Plumb payload.stop_sequences to all three code paths (server-side
tool loop, no-tool plain, client-side passthrough) so Anthropic SDK
clients setting stop_sequences get the behavior they expect. The
llama_cpp backend already accepted `stop` on both generate_chat_
completion and generate_chat_completion_with_tools; the Anthropic
handler simply wasn't passing it.
tool_choice remains declared on the request model for Anthropic SDK
compatibility (the SDK often sets it by default) but is not yet
honored. Log a structured warning on each request carrying a non-
null tool_choice so the silent drop is visible to operators.
* Wire min_p / repetition_penalty / presence_penalty through /v1/messages
Align the Anthropic endpoint's sampling surface with /v1/chat/completions.
Adds the three fields as x-unsloth extensions on AnthropicMessagesRequest
and threads them through all three code paths: server-side tool loop,
no-tool plain, and client-side passthrough.
The passthrough builder emits "repeat_penalty" (not "repetition_penalty")
because that is llama-server's field name; the backend methods already
apply the same rename internally.
* Fix block ordering and prev_text reset in non-streaming tool path
_anthropic_tool_non_streaming was building the response by appending
all tool_use blocks first, then a single concatenated text block at
the end — losing generation order and merging pre-tool and post-tool
text into one block. It also never reset prev_text between synthesis
turns, so the first N characters of each post-tool turn were dropped
(where N = length of the prior turn's final cumulative text).
Rewrite to build content_blocks incrementally in generation order,
matching the streaming emitter's behavior: deltas within a turn are
merged into the trailing text block, tool_use blocks interrupt the
text sequence, and prev_text is reset on tool_end so turn N+1 diffs
against an empty baseline.
Caught by gemini-code-assist[bot] review on #4981.
* Make test_studio_run.py e2e tests pytest-compatible
Add a hybrid session-scoped studio_server fixture in conftest.py that
feeds base_url / api_key into the existing e2e test functions. Three
invocation modes are now supported:
1. Script mode (unchanged) — python tests/test_studio_run.py
2. Pytest + external server — point at a running instance via
UNSLOTH_E2E_BASE_URL / UNSLOTH_E2E_API_KEY env vars, no per-run
GGUF load cost
3. Pytest + fixture-managed server — pytest drives _start_server /
_kill_server itself via --unsloth-model / --unsloth-gguf-variant,
CI-friendly
The existing _start_server / _kill_server helpers and main() stay
untouched so the script entry point keeps working exactly as before.
Test function signatures are unchanged — the (base_url, api_key)
parameters now resolve via the new fixtures when running under
pytest.
* Rename test_studio_run.py -> test_studio_api.py
The file is entirely about HTTP API endpoint testing (OpenAI-compatible
/v1/chat/completions, Anthropic-compatible /v1/messages, API key auth,
plus a CLI --help sanity check on the command that runs the API). None
of its tests cover training, export, chat-UI, or internal-Python-API
concerns.
The old name misleadingly suggested "tests for the unsloth studio run
CLI subcommand" — the new name reflects the actual scope.
Updates:
- git mv the file (rename tracked, history preserved)
- Rewrite opening docstring to state the API surface focus and call
out what is explicitly out of scope
- Update all 4 Usage-block path references to the new filename
- LOG_FILE renamed to test_studio_api.log
- conftest.py fixture import rewritten from test_studio_run to
test_studio_api, plus 7 docstring/comment references updated
No functional changes to test logic, signatures, or main().
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix httpcore asyncgen cleanup in /v1/messages and /v1/completions
The earlier fix in 985e92a9 was incomplete: it closed aiter_lines()
explicitly but still used `async with httpx.AsyncClient()` /
`async with client.stream()` inside the generator. When the generator
is orphaned (e.g. client disconnects mid-stream and Starlette drops
the StreamingResponse iterator without explicitly calling aclose()),
Python's asyncgen finalizer runs the cleanup in a DIFFERENT task than
the one that originally entered the httpx context managers. The
`async with` exits then trigger httpcore's HTTP11ConnectionByteStream
.aclose(), which enters anyio.CancelScope.__exit__ with a mismatched
task and raises RuntimeError("Attempted to exit cancel scope in a
different task"). That error escapes any user-owned try/except
because it happens during GC finalization.
Replace `async with` with manual client/response lifecycle in both
/v1/messages passthrough and /v1/completions proxy. Close the
response and client in a finally block wrapped in
`try: ... except Exception: pass`. This suppresses RuntimeError (and
other Exception subclasses) from the anyio cleanup noise while
letting GeneratorExit (a BaseException, not Exception) propagate
cleanly so the generator terminates as Python expects.
Traceback observed in user report:
File ".../httpcore/_async/connection_pool.py", line 404, in __aiter__
yield part
RuntimeError: async generator ignored GeneratorExit
...
File ".../anyio/_backends/_asyncio.py", line 455, in __exit__
raise RuntimeError(
RuntimeError: Attempted to exit cancel scope in a different task
* Expand unsloth studio run banner with SDK base URL and more curl examples
Add an explicit "OpenAI / Anthropic SDK base URL" line inside the info
box so SDK users don't accidentally copy the bare server URL (without
/v1) into their OpenAI/Anthropic SDK constructors and hit 404s.
Replace the single /v1/chat/completions curl example with three
labeled blocks: chat/completions, Anthropic /messages, and OpenAI
Responses. The Anthropic example includes max_tokens (Anthropic SDKs
require it even though Studio accepts None).
All examples derived from a computed sdk_base_url so the /v1 prefix
stays in sync if the public path ever changes.
* Hash API keys with HMAC-SHA256 + persistent server secret
Stores the HMAC secret in a new app_secrets singleton table. Fixes
CodeQL py/weak-sensitive-data-hashing alert on storage.py:74-76,
394-395. Refresh tokens stay on plain SHA-256 (unchanged _hash_token)
so existing user sessions survive upgrade — API keys are new on this
branch so there is no migration.
* Use PBKDF2 for API key hashing per CodeQL recommendation
HMAC-SHA256 was still flagged by py/weak-sensitive-data-hashing.
Switch to hashlib.pbkdf2_hmac, which is in CodeQL's recommended
allowlist (Argon2/scrypt/bcrypt/PBKDF2). Persistent server-side
salt stays in app_secrets for defense-in-depth. 100k iterations to
match auth/hashing.py's password hasher.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: Roland Tannous <rolandtannous@gravityq.ai>
* Pin bitsandbytes to continuous-release_main on ROCm for 4-bit decode fix
bitsandbytes 0.49.2 on PyPI ships with a broken 4-bit GEMV kernel on
every ROCm target:
- CDNA (gfx90a / gfx942 / gfx950 = MI210 / MI300X / MI350) via a
broken blocksize=32/64 warp64 GEMV kernel whose tests were
explicitly skipped with ROCM_WARP_SIZE_64 guards because the
code was known broken.
- RDNA3 / RDNA3.5 (gfx1100-1103 / gfx1150-1152) via a compile-time
BNB_WARP_SIZE macro in the host-side dispatch that resolves to
64 when the multi-arch wheel is compiled with CDNA as the
primary target, so num_blocks is wrong on RDNA and half the GEMV
output is never written.
At decode shape (1, 1, hidden) both bugs produce NaN. Training is
unaffected because training shapes are (batch, seq_len > 1, hidden)
and never touch the GEMV path. The crash during autoregressive
inference surfaces as _assert_async_cuda_kernel in torch.multinomial
which on HIP becomes a hard HSA_STATUS_ERROR_EXCEPTION instead of
a clean Python error.
Both bugs are fixed by bitsandbytes commit 713a3b8 ("[ROCm] Enable
blocksize 32 4-bit quantization and GEMV kernels on AMD CDNA",
PR #1887, merged 2026-03-09) which replaces BNB_WARP_SIZE with a
runtime hipDeviceGetAttribute query and ships a working CDNA warp64
kernel. That commit has not shipped to PyPI yet, but
continuous-release_main wheels are published on every push to bnb
main via GitHub Releases.
Point the ROCm install path at the continuous-release_main x86_64 and
aarch64 wheels and fall back to PyPI >=0.49.1 when the pre-release is
unreachable (offline installs, firewalled hosts, or architectures not
covered by the pre-release wheels). Drop the pin once bnb cuts a
0.50+ tag on PyPI.
Verified on MI300X (gfx942, ROCm 7.2, torch 2.10.0+rocm7.1): direct
bnb GEMV shape test now returns 0.0078 max abs error at seq_len=1
(no NaN) vs NaN on 0.49.2, and full Unsloth + for_inference + 4-bit
sampling generation works end-to-end.
NVIDIA / CPU / Mac / Windows paths are unaffected -- the helper is
gated on the ROCm torch index and platform.machine() respectively.
* Drop Studio ROCm 16-bit fallback now that bnb 0.50+ fixes 4-bit decode
The 16-bit fallback in studio/backend/core/inference/inference.py was
added as a workaround for a bug that this PR already fixes at the
install layer: bitsandbytes <= 0.49.2 has a broken 4-bit GEMV kernel
on every ROCm target, which NaNs at decode shape (seq_len=1) and
crashes autoregressive inference. bnb PR #1887 (commit 713a3b8, in
0.50.0.dev0+, pinned by install.sh / install_python_stack.py in this
PR) restores correct 4-bit decode on MI300X and verified working
end-to-end with full Unsloth + for_inference + sampling.
Revert the dual code path so ROCm and NVIDIA both go through the
normal FastLanguageModel.from_pretrained + for_inference flow:
- Remove the conditional `from unsloth import` that skipped the
import on ROCm. The monkey-patches it was trying to avoid were
never the cause of the crash; bnb 4-bit GEMV was.
- Remove the `if _hw_module.IS_ROCM:` branch in load_model that
loaded with plain transformers + PEFT + bfloat16, and the
`_resolve_fp16_base` helper it relied on.
- Remove the `get_chat_template is not None` fallback in
_load_chat_template_info -- get_chat_template is now always
imported.
- Refactor the audio/vision ROCm guard to check _hw_module.IS_ROCM
directly instead of the removed _IS_ROCM_ENV global. Audio and
vision on ROCm still need separate validation (FastVisionModel
and the CSM audio codecs were never tested on HIP) so the guard
stays for now.
Add _bnb_rocm_4bit_ok() as a runtime safety net for users who
install from this PR before the install.sh bnb pin kicks in, or
whose installer fell back to the PyPI pin because the continuous-
release wheel was unreachable. When the installed bnb is < 0.50 on
ROCm, force load_in_4bit=False and strip any -unsloth-bnb-4bit /
-bnb-4bit suffix from the model path so a pre-quantized repo
resolves to its FP16 sibling instead of pulling bnb back in via
the repo's quantization_config. LoRA adapters whose base is a
pre-quantized repo on old bnb will still fail inside Unsloth's
loader -- the only real fix there is `unsloth studio update`.
Verified on MI300X (gfx942, ROCm 7.2, torch 2.10.0+rocm7.1):
- HAPPY path (bnb 0.50.0.dev0, load_in_4bit=True, pre-quantized
repo): loads in 4-bit via the fixed GEMV, generation returns
"Paris." for greedy and sampling.
- SAFETY-NET path (simulated old bnb, suffix-stripped to the
FP16 sibling, load_in_4bit=False): loads in bf16, generation
returns "Paris." for greedy and sampling.
Net diff is ~45 lines smaller than the pre-revert state because
the entire plain-transformers 16-bit branch is gone.
* Cache _bnb_rocm_4bit_ok() with functools.cache
load_model() can be called many times in a single session but the bnb
version and hardware state cannot change at runtime, so memoise the
check. First call is ~1.9 ms (dominated by the lazy `import bitsandbytes`
inside the try block), subsequent calls drop to sub-microsecond dict
lookups. Zero behavioral change.
* Shorten verbose bnb/ROCm comments
Comment-only cleanup across install.sh, studio/install_python_stack.py,
and studio/backend/core/inference/inference.py. No behavioral change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove _bnb_rocm_4bit_ok safety net from inference.py
Studio's ROCm support is brand new (PR #4720, merged today) and every
fresh install pulls the bnb continuous-release_main wheel via
install.sh / install_python_stack.py in this same PR. There are no
existing ROCm Studio installs carrying bnb < 0.50, so the defensive
version-check fallback is guarding against a scenario that cannot
actually occur. Delete the helper, the functools import, and the
safety-net block -- inference.py now calls FastLanguageModel.from_pretrained
directly with no ROCm branching.
* Drop audio/vision ROCm guard in inference.py — verified unblocked by bnb fix
Vision inference was blocked by the same bnb 4-bit GEMV bug that affected
text inference (vision models use bnb 4-bit for the LM backbone). With
bnb 0.50+ pinned in install.sh / install_python_stack.py, vision works
end-to-end on MI300X: Llama-3.2-11B-Vision-Instruct-unsloth-bnb-4bit
loaded in 4-bit via FastVisionModel + for_inference returns a correct
answer to a multimodal prompt.
Audio (CSM) was never actually blocked by HIP — on this hardware CSM
loads and runs its backbone forward pass fine with bnb 0.50, then fails
during generate() with a transformers-level kwarg validation mismatch
in generation_csm.py (`backbone_last_hidden_state` rejected). That's a
pre-existing transformers/CSM integration bug that reproduces identically
on NVIDIA, so the ROCm-gated guard was never actually protecting users
from anything HIP-specific.
Remove the combined audio/vision guard and the now-unused _hw_module
import. Also restore the one-word "Can be" in an inline comment that
drifted during the earlier comment-shortening pass, so the inference.py
delta vs pre-#4720 is exactly the max_seq_length<=0 crash fix and
nothing else.
* Shorten max_seq_length=0 guard comment to one line
---------
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Add ROCm detection to install.sh and expand shell tests
Add AMD ROCm GPU detection to get_torch_index_url() in install.sh.
When nvidia-smi is not found, probe for ROCm via amd-smi, /opt/rocm
version file, hipconfig, dpkg-query, and rpm.
Includes validation guard for malformed _rocm_tag, Debian epoch prefix
stripping, ROCm 7.2+ cap to rocm7.1 index, bitsandbytes AMD install,
and status messaging. Shell tests expanded to 23 cases.
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Add ROCm torch reinstall support to install_python_stack.py
Add _detect_rocm_version() and _ensure_rocm_torch() to detect when a
Linux host has ROCm but the venv received CPU-only torch, and reinstall
with the correct ROCm wheels. Covers ROCm 6.0 through 7.1 with a
30-second timeout on the torch GPU probe subprocess.
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Add ROCm support to llama.cpp prebuilt installer
Add has_rocm field to HostInfo, extend detect_host() to probe for ROCm
via hipcc/amd-smi/rocm-smi/ROCM_PATH, and route ROCm hosts to upstream
prebuilts (Linux ROCm 7.2 prebuilt with source fallback, Windows HIP
prebuilt with CPU fallback). Add linux-rocm and windows-hip install
kinds to runtime_patterns_for_choice().
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Add IS_ROCM hardware flag and fix AMD error message
Add IS_ROCM flag to hardware.py detect_hardware() (set when
torch.version.hip is present, DeviceType stays CUDA). Export IS_ROCM
from __init__.py. Add "rocm" key to get_package_versions().
Replace "We do not support AMD" error in tokenizer_utils.py with a
helpful message pointing to ROCm installation docs.
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Add comprehensive ROCm support test suite (68 tests)
Add tests/studio/install/test_rocm_support.py covering all ROCm code
paths across install_llama_prebuilt.py, install_python_stack.py,
hardware.py, tokenizer_utils.py, and install.sh. All tests use mocks
and run without AMD hardware.
Covers: asset selection (11), runtime patterns (5), HostInfo (4),
ROCm version detection (9), torch reinstall (9), index mapping (8),
hardware flag (8), tokenizer message (2), install.sh structure (10),
and live regression (1).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden ROCm support: probe error handling, version cap, validation
Address review findings from 8 independent reviewers:
- Wrap _ensure_rocm_torch() torch probe in try/except for
TimeoutExpired and OSError so a hung or broken torch import does not
crash the installer (8/8 reviewers flagged this)
- Add torch>=2.4,<2.11.0 version cap to the ROCm reinstall path to
prevent installing unsupported torch 2.11.0 from the rocm7.1 index
- Use with-statement for file reads in _detect_rocm_version() to avoid
resource leaks
- Handle ROCM_PATH="" correctly (use `or "/opt/rocm"` instead of
default parameter to avoid relative path resolution)
- Strengthen shell validation guard from rocm[0-9] to rocm[1-9] to
reject rocm0.x tags that would produce nonexistent PyTorch index URLs
- Switch shell version cap from blocklist to allowlist (rocm6.*|rocm7.0*
|rocm7.1* pass through, everything else caps to rocm7.1) so future
ROCm 10+ does not fall through to a nonexistent index
- Add sorted() to _ROCM_TORCH_INDEX lookup for defensive ordering
- Fix test_probe_timeout_handled: replace zero-assertion test with
proper assertions verifying reinstall proceeds after timeout
* Clean up rocm_paths list construction in detect_host()
Filter None from the ROCM_PATH env var lookup at list construction time
instead of relying on the inline `if p` guard in the any() call.
* Require actual AMD GPU presence before selecting ROCm paths
All 8 reviewers across 2 cycles independently flagged that ROCm
detection used toolkit/filesystem hints (hipcc, /opt/rocm, rocm-core)
as a proxy for GPU presence, which would misroute CPU-only or NVIDIA
hosts that happen to have ROCm tools installed.
Now all 3 detection points (install.sh, install_python_stack.py,
install_llama_prebuilt.py) probe for an actual AMD GPU before
entering the ROCm path:
- install.sh: check rocminfo for gfx* GPU names, or amd-smi list
for device rows, before version detection
- install_python_stack.py: new _has_rocm_gpu() function probes
rocminfo and amd-smi list before _ensure_rocm_torch() proceeds
- install_llama_prebuilt.py: detect_host() probes rocminfo/amd-smi
list instead of just checking tool existence or directory paths
Also:
- Shell test mock amd-smi now handles "list" subcommand
- Python tests updated to mock _has_rocm_gpu where needed
- Added test_no_gpu_with_rocm_tools_skips to verify the new guard
- Test index lookups now use sorted() to match production code
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden hipconfig version parsing and torch probe compatibility
- Add parts[1].isdigit() check in hipconfig version parsing to handle
versions like "6.3-HIP" where the minor component has non-numeric
suffix (strip "-" prefix before int() conversion)
- Use getattr() in torch probe subprocess to safely handle old or
custom torch builds that may lack torch.version.hip/cuda attributes
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strengthen AMD GPU detection and add NVIDIA precedence guard
- Change amd-smi list detection from any-non-empty-output to requiring
"gpu" marker in output, matching the shell-side NR>1 check. Prevents
false positives from header-only amd-smi list output.
- Add nvidia-smi check at the top of _ensure_rocm_torch() so mixed
AMD+NVIDIA hosts preserve NVIDIA precedence (matching install.sh and
install_llama_prebuilt.py behavior).
- Apply the same amd-smi marker fix to install_llama_prebuilt.py
detect_host() for consistency.
* Add Windows-specific ROCm/HIP detection in detect_host()
The previous detect_host() ROCm check used rocminfo and amd-smi list
which are Linux-only tools. On Windows, has_rocm would always be False,
making the Windows HIP prebuilt path at line 1794 unreachable.
Now detect_host() uses platform-specific detection:
- Linux: rocminfo (check for gfx GPU names) or amd-smi list
- Windows: hipinfo.exe, amd-smi, or amdhip64.dll on PATH
This allows Windows AMD users to get the HIP prebuilt binary instead
of silently falling through to the CPU prebuilt.
* Add AMD ROCm gaps: Mamba/SSM source builds, GPU monitoring, Windows messaging, RDNA expansion
- worker.py: Add HIP detection to causal-conv1d/mamba-ssm probe, check
for hipcc before ROCm source builds, improve status messages and error
reporting, add timeout and uv support for the source build fallback
- amd.py: New AMD GPU monitoring module via amd-smi metric --json,
mirroring nvidia.py structure (utilization, temperature, power, VRAM)
- hardware.py: Branch to amd.py when IS_ROCM is True for GPU utilization,
visible GPU queries, and physical GPU count
- install_python_stack.py: Detect AMD GPUs on Windows and warn that
ROCm-enabled PyTorch must be installed manually
- kernels/utils.py: Expand is_rdna() to cover RDNA2 (gfx1030-1032),
RDNA3 (gfx1102-1103), RDNA3.5 (gfx1150-1152) alongside existing entries
- tests: Add 32 new tests covering all changes (95/95 pass)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden ROCm detection, fix VRAM heuristic, and expand RDNA2 coverage
- Windows ROCm detection: validate actual GPU presence via hipinfo/amd-smi
output markers instead of just checking tool existence on PATH
- _ensure_rocm_torch: validate nvidia-smi actually reports a GPU before
giving NVIDIA precedence (fixes AMD-only hosts with stale NVIDIA tools)
- amd.py _parse_numeric: handle dict-shaped metric objects from newer
amd-smi versions ({"value": 10, "unit": "W"}) and strip MiB/GiB units
- amd.py VRAM heuristic: raise threshold from 100k to 10M to correctly
handle MI300X (192 GB = 196608 MB) and other high-VRAM GPUs
- amd.py visible GPU: use AMD-reported GPU IDs instead of enumerate index
so non-dense sets like CUDA_VISIBLE_DEVICES=1,3 report correctly
- install.sh: add ROCm <6.0 minimum version guard (no PyTorch wheels
exist for older versions); fix rocm7.1* glob to not match rocm7.10+
- is_rdna: add gfx1033-1036 for RDNA2 mobile GPUs (RX 6600M etc.)
- worker.py: increase ROCm source build timeout from 600s to 1800s;
fix success log message for ROCm source builds
- Tests: update mocks for _has_usable_nvidia_gpu, add RDNA2 target asserts
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add HIP_VISIBLE_DEVICES support, unit-aware VRAM parsing, Windows GPU validation
- hardware.py: check HIP_VISIBLE_DEVICES and ROCR_VISIBLE_DEVICES on ROCm
before falling back to CUDA_VISIBLE_DEVICES, so multi-GPU AMD setups with
HIP-specific env vars report the correct visible device set
- amd.py: add _parse_memory_mb() that reads "unit" from dict-shaped amd-smi
JSON (e.g. {"value": 192, "unit": "GiB"}) and converts to MB correctly;
fixes MI300X VRAM misreported as 0.19 GB instead of 192 GB
- install_python_stack.py: Windows AMD warning now validates actual GPU
presence via hipinfo/amd-smi output markers before printing
- install_llama_prebuilt.py: restore amdhip64.dll fallback for Windows HIP
detection after tool-based checks, so Windows HIP installs without CLI
tools on PATH are still detected
- hardware.py: fix IS_ROCM comment to accurately describe its role
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix HIP_VISIBLE_DEVICES empty-string handling in GPU visibility spec
Use explicit None checks instead of Python `or` operator when reading
HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES, so that an empty string
("") is correctly honored as "no visible GPUs" rather than silently
falling through to CUDA_VISIBLE_DEVICES on mixed ROCm+CUDA systems.
* Fix IS_ROCM test assertion for multi-line formatting
* Cap torchvision/torchaudio versions, remove amdhip64.dll fallback, fix visible GPU count
- Cap torchvision<0.26.0 and torchaudio<2.11.0 alongside torch<2.11.0 in
both install.sh and install_python_stack.py to prevent resolver from
selecting incompatible companion packages from ROCm wheel index
- Remove amdhip64.dll fallback in Windows ROCm detection (DLL presence
without hipinfo/amd-smi is not proof of GPU existence)
- Fix get_visible_gpu_count() to use _get_parent_visible_gpu_spec() which
respects HIP_VISIBLE_DEVICES/ROCR_VISIBLE_DEVICES on ROCm hosts
* Attribute is_rdna() RDNA2/3/3.5/4 expansion to PR #4428
The is_rdna() expansion to cover RDNA2 (gfx1030-1036), RDNA3
(gfx1100-1103), RDNA3.5 (gfx1150-1152), and RDNA4 (gfx1200-1201)
architectures is based on the original work from PR #4428.
Co-authored-by: GoldenGrapeGentleman <yueyuan@amd.com>
Co-authored-by: billishyahao <bill.he@amd.com>
* Support AMD Radeon for studio (#4770)
Co-authored-by: Iswarya Alex <iswarya.alex@amd.com>
* Remove ROCm test files from main PR
Move test_rocm_support.py and shell test additions to a separate PR
to keep the main ROCm support PR focused on implementation changes.
* Fix installer and hardware detection issues for PR #4720
- Fix empty _tri_arg passed to uv pip install in Radeon path (causes
"Empty field is not allowed for PEP508" error)
- Fix Radeon fallback: use ROCm index instead of CPU-only when
repo.radeon.com is unreachable (TORCH_INDEX_URL already has ROCm)
- Use $TORCH_CONSTRAINT in fallback paths instead of hardcoded strings
- Fix _pick_radeon_wheel: relax suffix to match manylinux_2_28_x86_64
wheels (AMD Radeon repo does not use bare linux_x86_64 platform tag)
- Fix IS_ROCM export: use __getattr__ so callers always see the live
value after detect_hardware() runs
- Fix apply_gpu_ids: set HIP_VISIBLE_DEVICES and ROCR_VISIBLE_DEVICES
on ROCm so _get_parent_visible_gpu_spec picks up narrowed GPU set
- Fix _parse_memory_mb: distinguish GB (1000 MB) from GiB (1024 MiB)
- Add amd-smi version as a fallback in _detect_rocm_version
- Fix trailing whitespace and missing newline at EOF in install.sh
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix GPU detection false positives and add missing health groups
- Fix _has_rocm_gpu() false positive: require "GPU: <number>" data rows
from amd-smi list, not just header containing "gpu"
- Apply same fix in detect_host() in install_llama_prebuilt.py
- Add runtime_payload_health_groups for linux-rocm and windows-hip so
partial/corrupt ROCm/HIP prebuilt installs are properly detected
- Add bitsandbytes install to Radeon fallback paths (was only in the
success path, skipped when repo.radeon.com was unreachable)
- Keep DEVICE/CHAT_ONLY as direct imports in __init__.py (matching main)
and only use __getattr__ for IS_ROCM
* Fix _ensure_rocm_torch and Windows AMD warning false positives
- _ensure_rocm_torch: only skip when HIP is already present, not for
CUDA builds (which are unusable on AMD-only hosts). Fixes the case
where a venv has a stale CUDA wheel and the repair step is skipped.
- Windows AMD warning: use GPU data row check (same as Linux fix) to
avoid false positives from amd-smi list header-only output.
* Fix amd-smi GPU detection for GPU[N] output format
Older amd-smi versions output "GPU[0] : Card series: ..." instead of
"GPU: 0". The regex now matches both "GPU: <digit>" and "GPU[<digit>"
formats to detect actual GPU data rows.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden AMD GPU detection against false positives
- install.sh: replace weak amd-smi list check (awk 'NR>1 && NF') with
strict pattern matching GPU data rows (/^GPU[[:space:]]*[:\[]/)
- All files: reject rocminfo gfx000 (CPU HSA agent) by requiring
gfx[1-9] instead of gfx[0-9] in the rocminfo GPU probe
- Fixes false positives on hosts with ROCm tools but no AMD GPU
* Remove duplicate comment from pre-commit merge
* Refactor: deduplicate AMD detection, consolidate bitsandbytes, clean up imports
- Extract _has_amd_rocm_gpu() shell function to avoid duplicating the
rocminfo/amd-smi GPU detection logic in get_torch_index_url and
the Radeon auto-detect block
- Consolidate bitsandbytes install into a single case block after torch
install (was duplicated 4 times across Radeon success/fallback paths)
- Move math and re imports to top of amd.py (were inline in functions)
- Add _smi_query() helper in hardware.py to centralize IS_ROCM backend
selection for get_gpu_utilization and get_visible_gpu_utilization
Addresses Gemini code review suggestions.
* Fix VRAM parsing for string values and GB/GiB consistency
- Extract unit from string-valued VRAM fields (e.g. "192 GiB") so
_parse_memory_mb correctly applies the unit multiplier instead of
treating the value as bare MB
- Treat GB and GiB identically (both as binary x1024) since GPU tools
including amd-smi use binary units even when labeling them "GB"
- Fixes incorrect VRAM reporting on MI300-class cards (was showing
~0.19 GB instead of 192 GB for string-valued outputs)
* Add --no-cache to uv for ROCm HIP source builds
Avoid stale cache artifacts from partial HIP source builds when
uv is used for causal-conv1d/mamba-ssm compilation on ROCm.
The pip path already uses --no-cache-dir; this adds the uv equivalent
(--no-cache) only when is_hip is True.
* Fix critical: initialize _amd_gpu_radeon before case block
_amd_gpu_radeon was only set inside the */rocm*) case arm, so on
NVIDIA/CPU/macOS paths where TORCH_INDEX_URL does not contain "rocm",
the variable was unbound. With set -u (nounset) enabled, this crashes
the installer for every non-AMD user.
Move initialization to before the case block so it is always defined.
* Fix Windows AMD: route has_rocm hosts to HIP prebuilt path
resolve_release_asset_choice was selecting windows-cpu for all Windows
x86_64 hosts including those with has_rocm=True. Windows AMD users
should fall through to resolve_upstream_asset_choice which tries the
HIP prebuilt first. Add "not host.has_rocm" guard to the published
windows-cpu selection.
* Harden ROCm detection, Radeon wheel fallback, and HIP visibility
Addresses review findings from parallel reviewers on PR #4720:
- install.sh: add _has_usable_nvidia_gpu() helper requiring nvidia-smi -L
to actually list a GPU before treating the host as NVIDIA. Fixes the
stale-nvidia-smi-on-PATH regression where AMD-only hosts fell into the
CUDA branch.
- install.sh: fix hipconfig awk blocks to propagate a non-zero exit code
when the output is not a recognisable version string, so the ||-chain
continues to dpkg-query / rpm instead of terminating early.
- install.sh: fail-closed on Radeon wheel fallback. When torch,
torchvision or torchaudio is missing from the Radeon repo for the
active Python tag, fall back to the standard ROCm index instead of
silently mixing Radeon wheels with PyPI defaults. Quote all wheel
arguments individually so wheel filenames cannot be word-split or
glob-expanded.
- install_llama_prebuilt.py: detect_host() now requires nvidia-smi -L to
list a GPU before setting has_physical_nvidia. Routes AMD ROCm hosts
with a broken leftover nvidia-smi to the ROCm path instead of
misclassifying them as NVIDIA.
- install_llama_prebuilt.py: scan upstream assets for any rocm-<version>
prebuilt instead of hard-coding rocm-7.2, so ROCm 6.x / 7.0 / 7.1 / 7.3+
users pick up a matching upstream prebuilt when one exists.
- install_llama_prebuilt.py: validate_server() adds --n-gpu-layers 1 for
linux-rocm and windows-hip hosts, so new HIP prebuilts are preflighted
on the GPU path instead of passing validation on CPU only.
- install_llama_prebuilt.py: restore the published windows-cpu fallback
for AMD Windows hosts without a HIP prebuilt so hash-approved bundles
are still preferred over the raw upstream CPU asset.
- install_python_stack.py: drop the /opt/rocm / hipcc gate in
_ensure_rocm_torch() and rely on _has_rocm_gpu(). Runtime-only ROCm
installs (package-managed minimal installs, Radeon software) that ship
amd-smi / rocminfo without hipcc can now repair a CPU-only venv via
"unsloth studio update". Adds an explicit IS_WINDOWS / IS_MACOS guard.
- studio/backend/utils/hardware/amd.py: honour HIP_VISIBLE_DEVICES /
ROCR_VISIBLE_DEVICES / CUDA_VISIBLE_DEVICES in
get_primary_gpu_utilization(). A process restricted to GPU 2 now
reports metrics for GPU 2 instead of physical GPU 0. Tighten the plain
bytes unit detection to an explicit allowlist.
- studio/backend/utils/hardware/hardware.py: route
get_backend_visible_gpu_info()'s backend_cuda_visible_devices field
through a helper that reads HIP_VISIBLE_DEVICES on ROCm. Drop the
unconditional "(rocm=False)" suffix in apply_gpu_ids() logs.
* Fix round 2 regressions: ROCm validate_server and Windows HIP routing
Follow-up to 810b833b addressing review findings on the first round of
hardening commits:
- install_llama_prebuilt.py validate_server: gate --n-gpu-layers on the
resolved install_kind instead of host.has_rocm. AMD Windows hosts
without a HIP prebuilt fall back to windows-cpu and must not be
validated with GPU layers; thread install_kind through from the
caller.
- install_llama_prebuilt.py resolve_release_asset_choice: reinstate the
"not has_rocm" guard on the published windows-cpu bundle so AMD
Windows hosts reach resolve_upstream_asset_choice() where the new
HIP prebuilt path lives. Prefer a published windows-hip bundle first
when one exists, fall through to upstream HIP + upstream CPU
otherwise.
- install_llama_prebuilt.py detect_host: also set has_physical_nvidia
when the secondary --query-gpu block confirms a working NVIDIA GPU,
so older nvidia-smi versions without -L support do not silently skip
the Linux diagnostics that key off has_physical_nvidia.
- install_llama_prebuilt.py: drop redundant "import re as _re" /
"import re as _re_rocm" local aliases in favour of the existing
top-level "import re".
- install_python_stack.py _ensure_rocm_torch: run the AMD
bitsandbytes install unconditionally after the HIP-torch probe so
"unsloth studio update" on venvs that already have ROCm torch still
gains the AMD bitsandbytes build.
- install.sh: add a non-x86_64 early-exit to get_torch_index_url() so
aarch64 / arm64 Linux hosts do not hit the ROCm wheel index
(PyTorch only publishes ROCm wheels for linux_x86_64).
- install.sh: add bitsandbytes install to the migrated-environment
branch so upgrades pick it up for ROCm hosts instead of only the
fresh-install path.
- install.sh: in the Radeon wheel path, pass version constraints +
--no-index --find-links to uv instead of explicit wheel URLs so a
version-compatible torch / torchvision / torchaudio triple is
resolved, rather than picking the highest-version wheel for each
package independently.
- studio/backend/utils/hardware/amd.py _first_visible_amd_gpu_id: fall
through to lower-priority visibility env vars when the first entry
is malformed (leading comma, all-whitespace first token) instead of
silently returning GPU 0.
* Fix round 3 findings: x86_64 guard, ROCm version clip, Radeon deps
Address issues surfaced by the round 3 reviewers on top of 8636fa63:
- install_python_stack.py _ensure_rocm_torch: add the same `x86_64`
guard that install.sh already has. Linux aarch64 / arm64 ROCm hosts
must skip the repair path entirely; PyTorch only publishes ROCm
wheels for linux_x86_64, and without this guard
`unsloth studio update` aborts with a missing-wheel error on non
x86_64 hosts.
- install_llama_prebuilt.py resolve_upstream_asset_choice: add a
best-effort _detect_host_rocm_version() helper (reading
/opt/rocm/.info/version, amd-smi version, hipconfig --version) and
filter rocm_candidates to entries whose major.minor is <= host
version. Falls back to the newest candidate only when no compatible
one exists, so a ROCm 6.4 host downloads rocm-6.4 instead of being
handed the numerically newest rocm-7.2 bundle (which fails preflight
and forces a source build).
- install.sh: remove the round 2 --no-index switch from the Radeon
wheel branch. --no-index forced uv to ignore PyPI entirely, which
broke transitive dependency resolution (filelock, sympy, networkx,
jinja2, fsspec, setuptools, typing-extensions, ...) on a fresh venv.
Restore the round 1 explicit wheel URL invocation but add a
torch / torchvision / torchaudio version-pair sanity check so a
mismatched trio (e.g. torch 2.9.1 + torchvision 0.23.0 + torchaudio
2.9.0) falls back to the standard ROCm index instead of installing a
broken combination.
- install_python_stack.py _ensure_rocm_torch: restructure the
"tag is None" path so it no longer short-circuits the bitsandbytes
install. On a ROCm runtime older than anything in
_ROCM_TORCH_INDEX, print the "no wheel" warning but still run the
AMD bitsandbytes install.
- studio/backend/core/training/worker.py: restore the pre-PR
"no timeout" behaviour for non-HIP causal-conv1d / mamba-ssm source
builds. The round 2 "timeout = 1800 if is_hip else 300" cap aborts
slow non-HIP builds (Linux aarch64, unsupported torch/CUDA combos)
after 5 minutes; omit timeout for the non-HIP branch so the cap
only applies to ROCm source builds.
* Fix round 4 findings: apply_gpu_ids env inheritance, Radeon X.Y, bitsandbytes gate
Address remaining issues surfaced by the round 4 reviewers:
- studio/backend/utils/hardware/hardware.py apply_gpu_ids: mirror the
selection into HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES whenever
the caller already had a ROCm visibility env var set, not only when
IS_ROCM has already been set by detect_hardware(). Training and
inference workers call apply_gpu_ids() before detect_hardware()
runs, so the old guard would leave a forked ROCm worker with a
stale HIP_VISIBLE_DEVICES mask that no longer matched the
narrowed CUDA_VISIBLE_DEVICES selection.
- install.sh get_radeon_wheel_url: accept X.Y ROCm versions in
addition to X.Y.Z. The `/opt/rocm/.info/version` file and some
hipconfig versions report only two components, and the Radeon
repository publishes both rocm-rel-X.Y.Z/ and rocm-rel-X.Y/
directories, so treating X.Y as invalid caused Radeon hosts to fall
back to the generic ROCm index even when a matching AMD wheel set
existed.
- install_python_stack.py _ensure_rocm_torch: only install the AMD
bitsandbytes build when the venv actually has a ROCm-compatible
torch (either already present or just installed by this function).
Previously the bitsandbytes install ran unconditionally, which
could leave an AMD bitsandbytes layered on top of a CPU/CUDA torch
on hosts where the ROCm runtime is older than any entry in
_ROCM_TORCH_INDEX. Also add --force-reinstall so an existing
CPU/CUDA bitsandbytes is replaced by the AMD build during upgrades.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix gemini findings: amd-smi metric envelope validation and dict-wrapped GPU id
Two medium-severity defensive fixes from the gemini-code-assist review on
the AMD monitoring backend:
1. _extract_gpu_metrics may return a dict where every value is None when
amd-smi succeeds (zero exit) but the JSON envelope contains no usable
fields (error response, unsupported card). The new _has_real_metrics
helper lets get_primary_gpu_utilization surface available:False and
lets get_visible_gpu_utilization skip ghost device rows so the UI
does not render placeholder cards with empty numbers.
2. Newer amd-smi versions wrap scalar fields as {"value": 0, "unit":
"none"}, including the per-GPU id. The previous int(raw_id) call
silently fell back to the enumeration index in that case, losing the
real GPU id. Routing raw_id through the existing _parse_numeric
helper handles bare ints, floats, strings, and the dict shape
uniformly, with a debug log on parse failure.
* Fix gemini round 2 findings: explicit length guard on ROCm version file parser
Both _detect_rocm_version (install_python_stack.py) and
_detect_host_rocm_version (install_llama_prebuilt.py) read /opt/rocm/.info/version
or $ROCM_PATH/lib/rocm_version, split on "." and unconditionally accessed
parts[1]. The surrounding broad `except Exception: pass` already swallowed
the resulting IndexError, so a one-component file like "6\n" did fall
through to the next detection source -- but the control flow relied on
exception handling instead of an explicit check.
Add `if len(parts) >= 2:` guards in both helpers so the loop falls through
on its own without raising. Behaviour is unchanged for the common multi-
component case; the previously-silent IndexError path becomes an explicit
no-op.
* Fix gemini round 3: include has_rocm in validate_server fallback path
When validate_server is called without an explicit install_kind (older
call sites that have not been updated), the fallback was only enabling
--n-gpu-layers for NVIDIA and macOS arm64 hosts. AMD ROCm Linux hosts
fell through to the CPU validation path even though the prebuilt being
exercised was a HIP binary.
Add host.has_rocm to the fallback expression so the GPU offload flag is
applied consistently with the install_kind=='linux-rocm' / 'windows-hip'
branches above.
* Fix gemini round 4: remove risky bytes-vs-MB heuristic in _parse_memory_mb
The previous heuristic divided any bare number above 10_000_000 by
1024*1024 on the assumption that large unit-less values were bytes.
This misclassified small VRAM allocations: 5 MB of used VRAM reported
as 5_242_880 bytes without a unit would be taken at face value and
render as 5_242_880 MB (~5 TB) in the monitoring UI.
Modern amd-smi always provides explicit units (MiB/GiB dict form),
and legacy amd-smi returns bare numbers in MB -- the heuristic never
had a real workload to handle. Drop it and default to MB for bare
numeric input, keeping the existing unit-aware branches for dict /
string inputs unchanged.
The unrelated gemini suggestion to "default minor to 0" in the
amd-smi version awk parser was intentionally NOT applied: rocm7.0
and rocm7.1 ship different wheel sets, so silently substituting 0
for a missing minor could install the wrong wheels. The existing
reject-and-fall-through behaviour is safer.
* Fix gemini round 5: POSIX compliance and leading-comma visibility parsing
Three medium findings from gemini-code-assist addressed in this commit:
1. _pick_radeon_wheel used grep -o and sort -V, both GNU extensions
that are not in POSIX and break on BSD/BusyBox coreutils. install.sh
has a #!/bin/sh shebang so the whole pipeline was rewritten as a
single awk script that extracts all href="..." hits on each line,
filters to wheels matching the package prefix and python tag, and
picks the newest version via zero-padded lexical comparison. No
external sort or grep is needed.
2. _first_visible_amd_gpu_id in the AMD monitoring backend treated a
leading comma (e.g. HIP_VISIBLE_DEVICES=",1") as "fall through to
the next env var", which is surprising given the clear intent to
narrow to device 1. Filter empty tokens after the split and return
the first real one. An all-commas value ("," / ",,,") still falls
through because no real tokens exist; the empty-string and "-1"
explicit-zero cases are unchanged.
The unrelated amd-smi version awk parser suggestion was not applied
(see round 4 commit message for rationale: defaulting a missing minor
to 0 could silently install the wrong ROCm wheel set).
* Fix 20-reviewer.py findings: base drift, Radeon %2B, dpkg/rpm fallback, bnb, backend label
Consolidated fix batch from a 20-parallel reviewer.py run on the current
head. Each fix is drawn from a high-consensus finding and addresses a
real bug or feature gap, not a stylistic preference.
1. install.sh: bump `unsloth>=2026.4.2` -> `unsloth>=2026.4.4` at five
call sites so this branch no longer regresses main's version floor
(main bumped to 2026.4.4 in #4876). Without this, merging 4720 would
silently downgrade the minimum version pin for fresh installs.
2. install.sh: URL-decode Radeon wheel names before extracting the
torch / torchvision / torchaudio version strings. Real wheel URLs
from repo.radeon.com are percent-encoded ("torch-2.10.0%2Brocm7.2.0...")
so the previous `[+-]` terminator in the sed regex never matched,
`_torch_ver` stayed empty, `_radeon_versions_match` stayed false,
and every Radeon consumer install silently fell back to the generic
ROCm index. Now decode %2B -> + first, then extract, then validate.
3. install.sh: the two AMD bitsandbytes install lines were running
`uv pip install "bitsandbytes>=0.49.1"` without `--force-reinstall`,
so upgrades where the venv already has a CPU/CUDA bitsandbytes
satisfying the constraint would keep the stale non-AMD wheel. Add
`--force-reinstall --no-cache-dir` to both call sites, matching the
pattern already used in install_python_stack.py::_ensure_rocm_torch.
4. install_python_stack.py and install_llama_prebuilt.py: add
`dpkg-query -W rocm-core` and `rpm -q rocm-core` fallbacks to the
Python-side ROCm version detectors so they match the chain in
install.sh::get_torch_index_url. Package-managed ROCm installs
(Debian/Ubuntu/RHEL/Fedora distro packages) can expose GPUs via
rocminfo/amd-smi but still lack /opt/rocm/.info/version, hipconfig,
or amd-smi `version` output -- without these fallbacks, `unsloth
studio update` on such hosts returned None and skipped the ROCm
torch repair. Also strip the dpkg epoch prefix ("1:6.3.0-1") before
parsing so epoch-annotated packages parse correctly.
5. hardware.py: add a `_backend_label(device)` helper that returns
"rocm" when IS_ROCM is set and the device is DeviceType.CUDA, and
use it for every `"backend": ...` emission in JSON responses served
to the Studio frontend. Internally we still represent ROCm hosts as
DeviceType.CUDA (ROCm torch reuses the whole torch.cuda.* API
surface), but the user-facing API now correctly reports "rocm" on
AMD boxes instead of labeling them as "cuda".
All 250 simulation scenarios pass (was 233 before this batch: added 17
new regression tests covering the version pin, %2B decoding, bnb
force-reinstall flags, dpkg/rpm fallback presence, and the
_backend_label helper's four-way truth table).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix gemini round 6 + URL audit: amd.py defensive checks, rocm6.5+ clip to 6.4
Two rounds of fixes in one commit, plus a full URL audit of every PyPI /
download.pytorch.org / repo.radeon.com reference the PR introduces.
amd.py (4 medium gemini findings on commit b3627bc2):
1. _extract_gpu_metrics used `and vram_total_mb` as part of the vram_util
gate. The follow-up `vram_total_mb > 0` already handles the division
guard, but the truthiness check was redundant and slightly surprising
for a 0.0 valid value. Replace with explicit `is not None and > 0`
for both vram_util and power_util.
2. get_physical_gpu_count called `data.get("gpu", ...)` without guarding
for non-dict envelopes. A scalar / string JSON response from amd-smi
would raise AttributeError. Add an isinstance(data, dict) check and
return None for unexpected shapes.
3. get_visible_gpu_utilization had the same .get() exposure on the outer
envelope. Rewrite the gpu_list extraction as an explicit
list/dict/else cascade so a malformed scalar envelope produces
gpu_list=[data] and continues without raising.
4. The same function's per-entry loop also called gpu_data.get() on
whatever was inside gpu_list. If a scalar ever leaks into the list
(directly or via the previous fix's fallback), _extract_gpu_metrics
would raise on the first .get() inside the helper. Skip non-dict
entries in the loop before extracting metrics.
install.sh (URL audit finding, previously flagged by 20-reviewer as #13):
5. get_torch_index_url used `rocm6.*` in the rocm tag case statement,
which matched rocm6.5 and rocm6.6 and emitted
download.pytorch.org/whl/rocm6.5 -- which returns HTTP 403 because
PyTorch only publishes rocm 5.7, 6.0-6.4, 7.0-7.2. Enumerate the
supported 6.x minors explicitly and add a rocm6.* fallback branch
that clips to rocm6.4 (the last supported 6.x wheel set).
URL audit results (all URLs PR 4720 references):
- 14/14 download.pytorch.org/whl/{cpu,cu118,cu124,cu126,cu128,cu130,
rocm6.0..6.4,rocm7.0..7.2} return HTTP 200.
- 9/9 repo.radeon.com/rocm/manylinux/rocm-rel-{5.7,6.0,6.1,6.2,6.3,
6.4,7.0,7.1,7.2}/ return HTTP 200.
- X.Y.Z patch directories exist for 7.0.2, 7.1.1, 7.2.1 but NOT for
6.3.0, 6.4.0, 6.2.1 -- install.sh already handles this via the X.Y.Z
-> X.Y fallback sed in the Radeon wheel install block.
- Docs links (rocm.docs.amd.com, docs.unsloth.ai AMD guide) and the
llama.cpp GitHub releases API endpoint all return 200.
Test suite: 255 -> 258. New regression coverage:
- U17: get_physical_gpu_count tolerates scalar amd-smi envelope
- U18: get_visible_gpu_utilization tolerates scalar envelope
- U19a-c: vram_util / power_util return None on zero total, but
vram_total_gb still echoes 0.0 (not None)
- A_rocm{6.5,6.6,6.9}_clips_to_rocm64: install.sh clips unsupported
6.x minors to rocm6.4 instead of producing a 403 index URL
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix reviewer.py round 2: tokenizer AMD multi-GPU, --no-torch bnb, main.py backend label
Three high-confidence findings from a second 20-parallel reviewer.py run
on commit 7effb3ae. Triaged 15 total findings and applied the three that
were confirmed as real bugs; the rest were either false positives (e.g.
"migrated AMD venv not repaired" -- _ensure_rocm_torch runs downstream
via setup.sh regardless), design decisions (e.g. visibility mask env
vars not consulted in installer detection), or edge cases the existing
fallback logic already handles.
1. unsloth/tokenizer_utils.py [6/20]: the multi-GPU guard's shell probe
runs `nvidia-smi --query-gpu=memory.used`, catches the failure, then
only raises if `torch.cuda.is_available()` is False. On ROCm torch,
torch.cuda.is_available() returns True (ROCm reuses the torch.cuda.*
API), so the guard becomes dead code on AMD hosts and multi-GPU AMD
setups slip through even though unsloth does not support them yet.
Add a torch.cuda.device_count() > 1 fallback inside the except so
AMD multi-visible-device setups are flagged consistently with the
original CUDA memory check.
2. install.sh [1/20]: the fresh-install bitsandbytes block for AMD ROCm
ran unconditionally when TORCH_INDEX_URL matched `*/rocm*`, even when
SKIP_TORCH=true (from --no-torch or Intel Mac auto-detect). A user
running `install.sh --no-torch` on an AMD host would still pull in
bitsandbytes despite explicitly asking for GGUF-only mode. Wrap the
case block in an outer `[ "$SKIP_TORCH" = false ]` guard.
3. studio/backend/main.py [3/20]: the /api/system endpoint returned
`"device_backend": get_device().value`, which is "cuda" on ROCm
hosts (because ROCm torch piggybacks on torch.cuda). Other endpoints
(hardware.py) already use the _backend_label helper which swaps
"cuda" -> "rocm" when IS_ROCM. Route /api/system through the same
helper so the Studio UI reports the backend consistently across all
endpoints.
4. studio/backend/tests/test_utils.py: update test_backend_matches_device
to call _backend_label(get_device()) instead of raw get_device().value
so the test matches the new contract and still passes on CUDA hosts.
Tests: 258 -> 261. New regression coverage:
- X08 main.py /api/system uses _backend_label
- X09 tokenizer multi-GPU guard has device_count() fallback
- X10 fresh-install bnb case block gated on SKIP_TORCH=false
* fix: prevent bitsandbytes from overwriting ROCm torch with CUDA wheels
During install, bitsandbytes was installed without --no-deps, causing
uv to resolve torch from PyPI (CUDA build) and silently overwrite the
ROCm wheels that were just installed in the previous step.
This happened in three places:
- install.sh: bitsandbytes install in both migrated and fresh paths
- install_python_stack.py: bitsandbytes install inside _ensure_rocm_torch()
Additionally, multiple install steps in install_python_stack.py (extras,
overrides, studio deps) can pull in CUDA torch via transitive
dependencies. A final _ensure_rocm_torch() call at the end of the
install sequence ensures ROCm torch is always in place at runtime.
All changes are gated behind ROCm-specific conditions and do not affect
NVIDIA, CPU-only, macOS, or Windows install paths.
Tested on AMD Instinct MI300X VF with ROCm 7.2.0 -- confirms
torch==2.10.0+rocm7.1 with HIP 7.1.25424 after install.
* fix: ROCm inference fallback -- skip Unsloth patching and bnb 4-bit on HIP
On AMD ROCm (HIP), two issues prevent the normal Unsloth inference path:
1. Unsloth's global monkey-patching of transformers model classes
(LlamaRotaryEmbedding, attention modules) triggers
_assert_async_cuda_kernel crashes on HIP during generation.
Training uses different code paths and works fine.
2. bitsandbytes 4-bit matmul kernels also trigger HIP assertion
failures on MI300X (CDNA3 / gfx942), even without Unsloth patching.
This commit adds a ROCm-specific inference fallback that:
- Skips importing Unsloth at module level (prevents global patching)
- Loads models in 16-bit with plain transformers + PEFT instead
- Resolves pre-quantized model names (e.g. "xxx-bnb-4bit" -> "xxx")
since pre-quantized HF repos still trigger bnb codepaths
- Guards get_chat_template calls (unavailable without Unsloth import)
- Fixes max_seq_length=0 being passed to from_pretrained (GGUF
semantics don't apply to transformers path)
The NVIDIA path is completely unchanged -- Unsloth import and
for_inference() optimization remain active. GGUF inference (via
llama-server/HIP) is unaffected since it never imports Python model
classes. AMD GPUs typically have large VRAM (e.g. 192GB on MI300X)
so 16-bit loading is practical for inference.
Tested on AMD Instinct MI300X VF (ROCm 7.2, HIP 7.1.25424):
- Simple generation: PASS
- Compare mode (base vs finetuned): PASS
- GGUF inference + tool calling: PASS (unaffected by this change)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: guard audio/vision inference on ROCm, remove unused import
- Add clear RuntimeError for audio/vision model inference on ROCm
(these paths use Unsloth's FastModel/FastVisionModel which would
crash on HIP; GGUF inference is the supported path on AMD)
- Remove unused `import os as _os` from the ROCm changes
* fix: amd-smi parsing for newer output format (gpu_data wrapper, mem_usage, temperature)
amd-smi on recent ROCm versions (7.x) wraps metric output in a
{"gpu_data": [...]} envelope instead of returning a raw list. This
caused get_primary_gpu_utilization() and get_visible_gpu_utilization()
to fail silently (returning available=False) because the GPU data
dict was never unwrapped.
Additionally:
- VRAM data moved from "vram" to "mem_usage" with "total_vram" /
"used_vram" keys. Added fallback key lookup.
- Temperature "edge" sensor returns "N/A" on MI300X VF; the previous
dict.get() chain returned the "N/A" string instead of falling
through to "hotspot". Changed to a loop that checks each key until
a parseable value is found.
Tested on AMD Instinct MI300X VF (ROCm 7.2, amd-smi 24.x):
- GPU utilization: 0% (idle), up to 100% during training
- Temperature: 40-44C (from hotspot sensor)
- VRAM: 0.28/191.69 GB (idle)
- Power: 158-211W draw
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Bug fix detecting radeon (#4940)
* Bug fix detecting radeon
* Expanding GPU target for gfx1100*
* Generalize gfx family-prefix filter to cover gfx10/gfx12 as well
rocminfo on ROCm 6.1+ emits LLVM generic-family ISA lines alongside the
specific GPU (e.g. gfx11-generic next to gfx1100). The outer grep captures
the bare family prefix from the generic line, and passing that to
-DGPU_TARGETS breaks the HIP build because clang only accepts specific
gfxNNN ids.
The previous filter only special-cased gfx11. Generalize it so any bare
2-digit family prefix (gfx10, gfx11, gfx12, ...) is dropped whenever a
specific sibling target is present in the same list. No real AMD GPU has
a 2-digit gfx id, so the filter can only ever drop family prefixes and
never a real target.
Covers the existing gfx11 cases unchanged, and extends the same fix to
gfx10-1-generic / gfx10-3-generic (RDNA1/2) and gfx12-generic (RDNA4),
which would otherwise hit the same build failure on newer rocminfo.
---------
Co-authored-by: Iswarya Alex <iswarya.alex@amd.com>
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
---------
Co-authored-by: Eda Z <eda.zhou@amd.com>
Co-authored-by: GoldenGrapeGentleman <yueyuan@amd.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: billishyahao <bill.he@amd.com>
Co-authored-by: Iswarya Alex <47045679+iswaryaalex@users.noreply.github.com>
Co-authored-by: Iswarya Alex <iswarya.alex@amd.com>
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
* split venv_t5 into venv_t5_530 and venv_t5_550 for tiered transformers 5.x support
* fix bfloat16 crash on T4 for FORCE_FLOAT32 models and disable trust_remote_code auto-enable for native t5 models
* revert FORCE_FLOAT32 dtype change
* restrict trust_remote_code auto-enable to Nemotron models only
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* use config.json model_type for tier detection, add unsloth/nvidia namespace guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Revert "[pre-commit.ci] auto fixes from pre-commit.com hooks"
This reverts commit fb43d468e2.
* Revert "use config.json model_type for tier detection, add unsloth/nvidia namespace guard"
This reverts commit fc49ae2453.
* add unsloth/nvidia namespace guard to Nemotron trust_remote_code auto-enable
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* reorder tier checks: all substring matches before config.json fetches
* extract shared activate_transformers_for_subprocess into transformers_version.py
* narrow Nemotron trust_remote_code to nemotron_h/nemotron-3-nano, add to export worker
* clean venv_t5 dirs before re-install in setup.sh, clarify version alias comment
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* run venv_t5 migration outside deps fast-path gate in both setup scripts
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Add vision detection cache to is_vision_model() to avoid redundant subprocess spawns
is_vision_model() is called 4-5 times per training run for the same model
with zero caching. For transformers 5.x models, each call spawns a full
subprocess (~6s each). This adds a module-level _vision_detection_cache dict
following the same pattern as the existing _audio_detection_cache used by
detect_audio_type(). The function is refactored into a thin cache wrapper
around _is_vision_model_uncached(), saving ~12s per training run.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Include hf_token in vision cache key for gated model correctness
Cache key is now (model_name, hf_token) instead of just model_name.
This prevents stale False results when an unauthenticated probe for a
gated model is followed by an authenticated call.
* Remove test file from main PR - will be submitted separately
* Fix vision cache: normalize model names and skip caching transient failures
- Normalize model names in cache key using resolve_cached_repo_id_case()
to avoid duplicate entries for different casings of the same HF repo
(aligns with case normalization from #4822)
- Return None instead of False on transient failures (network errors,
subprocess timeouts, HF API issues) so the cache layer can distinguish
"definitely not a vision model" from "failed to check"
- Only cache definitive True/False results; transient failures are retried
on the next call instead of being permanently locked in as False
* Refine failure handling: cache deterministic failures, guard normalization
- Subprocess non-zero exit, JSON errors, and general exceptions return
False (deterministic, cached) instead of None (retryable). Only
subprocess.TimeoutExpired returns None since timeouts are transient.
- Wrap cache key normalization in try/except so resolve_cached_repo_id_case
or normalize_path failures fall back to raw model_name instead of
crashing callers.
* Harden vision detection cache: fix transient failure handling, thread safety, token security
- All subprocess failure paths now return None (transient) instead of False,
preventing permanent misclassification of VLMs after temporary HF/auth/network errors
- Use SHA256 fingerprint for hf_token in cache key instead of raw bearer token
- Add threading.Lock with double-checked locking to prevent thundering herd
of concurrent subprocess spawns for the same uncached model
- Distinguish permanent failures (RepositoryNotFoundError, GatedRepoError,
ValueError) from transient ones in _is_vision_model_uncached
- Pass resolved/normalized model name to detection (not just cache key)
- Log normalization fallback at debug level instead of silent swallow
- Thread hf_token through callers in routes/models.py and trainer.py
that previously omitted it
* Refine lock strategy and token fingerprint
- Move detection computation outside the lock to avoid serializing
long-running subprocess spawns (60s timeout) and HF API calls across
all concurrent model checks. Lock is now only held for cache writes.
- Use full SHA256 digest for token fingerprint instead of truncated
16-char prefix to eliminate collision risk.
* Fix huggingface_hub import fallback and use atomic cache read
- Add fallback import path for RepositoryNotFoundError/GatedRepoError
from huggingface_hub.utils (older hub versions) when .errors is
not available
- Use sentinel-based dict.get() for single atomic cache read instead
of two-step in/[] pattern (future-proof for no-GIL runtimes)
* [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: add speculative decoding support (ngram-mod, on by default)
Enable n-gram speculative decoding for GGUF models in Unsloth Studio.
Uses llama.cpp's ngram-mod mode which gives 10-40% faster generation
with zero VRAM cost via a 4MB fixed hash table that auto-resets on
low acceptance rates.
Backend:
- Add speculative_type field to LoadRequest, LoadResponse, and
InferenceStatusResponse pydantic models
- Add speculative_type parameter to LlamaCppBackend.load_model()
with allowlist validation (ngram-simple, ngram-mod)
- Pass --spec-type, --spec-ngram-size-n 16, --draft-max 24 flags
to llama-server when ngram-mod is active
- Default to ngram-mod for non-vision GGUF models server-side
- Silently skip speculative decoding for vision models (unsupported
in llama.cpp server-context.cpp)
Frontend:
- Add speculative_type to TS API types
- Add speculativeType/loadedSpeculativeType to chat runtime store
with default value of "ngram-mod"
- Add On/Off toggle in Model settings section (GGUF only, hidden
for vision models), included in dirty check for Apply/Reset
- Wire speculative_type through model load request and response
- Restore speculative type state on page refresh/reconnect
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: remove server-side speculative decoding override
The backend was overriding speculative_type=None to "ngram-mod" for
non-vision GGUF models, which prevented users from disabling spec
decoding via the UI toggle. The frontend store already defaults to
"ngram-mod", so the backend fallback was redundant and blocked the
explicit "Off" setting.
* fix: use recommended ngram-mod params from llama.cpp docs
Update speculative decoding params to match the recommended values
from llama.cpp docs (docs/speculative.md):
--spec-ngram-size-n 24 (was 16, docs say small n not recommended)
--draft-min 48 (was 0)
--draft-max 64 (was 24, docs note MoEs need long drafts)
Also fix comment: ngram-mod uses ~16 MB (4M entries * 4 bytes),
not 4 MB.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* add benchmark table and references to speculative decoding comment
Include speedup numbers from llama.cpp PRs #18471 and #19164 as an
inline comment so future readers understand the expected gains.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(studio): harden sandbox security for terminal and python tools
The existing command blocklist used naive str.split() which is trivially
bypassable via quoting, full paths, nested shells, variable expansion,
and cross-tool pivoting through Python os.system/subprocess. Fixes#4818.
Changes:
- Replace str.split() blocklist with shlex.split() + os.path.basename()
tokenization and regex scanning at shell command boundaries
- Add sanitized subprocess environment (_build_safe_env) that strips
credentials (HF_TOKEN, WANDB_API_KEY, GH_TOKEN, AWS_*, etc.) and
restricts PATH to /usr/local/bin:/usr/bin:/bin
- Add PR_SET_NO_NEW_PRIVS via prctl on Linux so sudo/su/pkexec fail
at the kernel level regardless of how they are invoked
- Add RLIMIT_NPROC (256) and RLIMIT_FSIZE (100MB) to prevent fork
bombs and disk filling attacks
- Extend AST safety checker to detect os.system(), os.popen(),
subprocess.run/Popen/call/check_output, os.exec*, os.spawn* calls
containing blocked commands or dynamic (non-literal) arguments
- Add cross-platform support: cmd.exe on Windows, bash on Unix;
CREATE_NO_WINDOW flag on Windows, preexec_fn on Unix
- Expand blocklist from 7 to 14 commands: add su, chown, passwd,
mount, umount, fdisk, kill, killall, pkill
- Apply all layers to both _bash_exec and _python_exec
Zero measurable performance overhead -- shlex parsing and a single
prctl syscall per subprocess fork.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix review findings: exception_catching dead code, false positives, process substitution
- Include exception_catching reasons in _check_code_safety so bare
except-in-loop timeout evasion is actually blocked (was computed in
_check_signal_escape_patterns but never read by the caller)
- Remove base.split() inner loop that caused false positives on quoted
text arguments containing blocked words (e.g. echo "kill this process")
- Add targeted nested shell detection for bash/sh/zsh -c arguments
instead, which catches bash -c 'sudo whoami' without false positives
- Add <() process substitution to the regex character class so
diff <(rm -rf /path) is also caught
- Fix error message to say "unsafe patterns" instead of specifically
mentioning signal manipulation when other categories trigger
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review feedback: regex paths, keyword args, list element scanning
- Regex now matches blocked commands after optional path prefix at shell
boundaries (catches ls; /usr/bin/sudo and similar)
- Nested shell detection uses os.path.basename so bash -c "/bin/rm" is
caught
- AST checker now inspects keyword arguments (not just positional) so
subprocess.run(args="sudo ...", shell=True) is detected
- List elements in subprocess calls are now checked via
_find_blocked_commands for consistency (catches subprocess.run(["bash",
"-c", "rm -rf /"]))
- Dynamic argument check uses _is_safe_literal that validates list
contents are all string literals
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix nested shell scan to only check the script body, not positional args
bash -c 'script' arg0 arg1 -- only tokens[i+1] is the script body;
subsequent tokens are $0, $1 positional parameters passed to the script
and are not executed as shell commands. Scanning all remaining tokens
caused false positives.
* Add subshell parentheses to regex command boundary detection
(sudo whoami) was not caught because ( was not in the regex character
class for shell command boundaries. Add ( to the set alongside ;, &,
|, backtick, newline.
* Address high-priority review findings from 7 parallel reviewers
- Track from-imports of dangerous functions (from os import system,
from subprocess import run as r, etc.) via shell_exec_aliases dict
so bare-name calls are detected by the AST checker
- Include the active Python interpreter and virtualenv directories
in the sanitized PATH so pip, uv, and Studio packages remain
accessible in the sandbox
- Add Windows-specific blocked commands (rmdir, takeown, icacls,
runas, powershell, pwsh) only on win32 platform
- Add os.posix_spawn and os.posix_spawnp to _SHELL_EXEC_FUNCS
- Handle tuple literals same as list literals in AST argument
inspection (both _extract_strings_from_list and _is_safe_literal)
* Fix false positive on check=True kwargs and recursive nested shell scanning
- Only inspect command-carrying keyword arguments (args, command,
executable, path, file) in the AST checker, not control flags like
check=True, text=True, capture_output=True which are booleans and
were incorrectly flagged as non-literal dynamic arguments
- Replace split() in nested shell detection with recursive call to
_find_blocked_commands so that quoted commands (bash -c '"sudo"
whoami') and semicolons (bash -c "sudo;ls") within nested shells
are properly detected through the full shlex + regex pipeline
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Move preexec_fn imports to module level and use find_library for libc
Addresses two Gemini review findings:
1. preexec_fn thread safety: _sandbox_preexec previously imported ctypes
and resource inside the function body, which runs between fork() and
exec() in the child process. In a multi-threaded server, this could
deadlock if the import machinery locks were held by another thread at
fork time. Now all imports and the libc handle are resolved once at
module load time, so _sandbox_preexec only calls C-level functions
(prctl, setrlimit) with no Python import activity.
2. Hardcoded libc.so.6 path: replaced with ctypes.util.find_library("c")
which works on glibc (libc.so.6), musl (libc.musl-*.so.1), and other
Linux distributions where libc has a different soname.
* Apply Gemini style suggestions: combined regex, dict.fromkeys, constant hoisting
- Combine per-word regex loop into a single re.findall with alternation
pattern, avoiding repeated regex compilation and searching
- Replace manual dedup loop with dict.fromkeys for PATH entries
- Hoist _CMD_KWARGS frozenset out of visit_Call to avoid recreating it
on every AST node visit
* Add cmd /c nested shell detection for Windows parity
The nested shell scan only checked for Unix shells (bash -c, sh -c, etc).
Add cmd /c and cmd.exe /c detection so that Windows nested shell
invocations are also recursively scanned for blocked commands. The token
scan already catches blocked commands at any position, so this is
defense-in-depth for consistency across platforms.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle combined shell flags (-lc, -xc) and interleaved flags (--login -c)
The nested shell scan only matched token == "-c" with the immediately
preceding token being a shell name. This missed:
- Combined flags: bash -lc 'rm ...' (-lc ends with c, is a valid
combined flag meaning -l -c)
- Interleaved flags: bash --login -c 'sudo ...' (--login sits between
bash and -c)
Now matches any short flag ending in 'c' (e.g. -lc, -xc, -ic) and
walks backwards past intermediate flags to find the shell binary.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix /bin/bash bypass, remove RLIMIT_NPROC, reduce AST false positives
Addresses three high-consensus findings from 20-reviewer pass:
1. /bin/bash -c 'sudo whoami' bypassed nested shell scan because the
backwards flag-skip logic treated paths starting with / as flags.
Now only skips tokens starting with - as Unix flags; on Windows
only skips short /X flags (not /bin/bash style paths). [9/20]
2. RLIMIT_NPROC=256 caused subprocess.run to fail with EAGAIN because
Linux enforces NPROC per real UID, not per process tree. Removed
RLIMIT_NPROC entirely; RLIMIT_FSIZE and PR_SET_NO_NEW_PRIVS remain
as the primary resource and privilege controls. [5/20]
3. AST checker rejected safe dynamic subprocess usage like
cmd=["git","status"]; subprocess.run(cmd) as shell_escape_dynamic.
Now only flags dynamic args for shell-string functions (os.system,
os.popen, subprocess.getoutput, etc.) or when shell=True is
explicitly set. List-based subprocess calls with shell=False (the
default) do not pass through a shell and are not flagged. [12/20]
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle Windows drive letter paths and .exe extensions in command detection
Gemini review found that Windows absolute paths (C:\Windows\System32\
shutdown.exe) and executable extensions (.exe, .com, .bat, .cmd) were
not handled:
- Token scan now strips .exe/.com/.bat/.cmd extensions before checking
the blocklist, so sudo.exe matches sudo, shutdown.bat matches shutdown
- Regex pattern now includes optional Windows drive letter prefix
([a-zA-Z]:[/\\]) and optional executable extension suffix, so commands
after shell metacharacters with full Windows paths are also caught
* Handle **kwargs dict expansion, non-literal shell=, and except Exception false positive
Addresses three findings from second 20-reviewer pass:
1. **kwargs dict expansion (9/20): subprocess.run(**{"args": "rm ...",
"shell": True}) bypassed the AST checker because **kwargs were
treated as opaque. Now expands literal dict **kwargs to inspect
their keys, and flags opaque **kwargs (variable dicts) as unsafe.
2. Non-literal shell= values (7/20): shell=variable was treated as
shell=False (safe). Now any shell= value that is not literally
False is treated as potentially True (conservative default).
3. except Exception false positive (1/20): except Exception in a loop
was flagged as timeout evasion, but Exception does not catch
SystemExit or KeyboardInterrupt which are used for timeout
enforcement. Narrowed to only flag except BaseException and
except TimeoutError in loops.
* [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>
Fixes#4809
On a new Studio chat, the first tool call could start before the frontend
initializes the thread ID. That meant the first request could go out without
a session_id, so the backend started the tool in the shared sandbox root
instead of the chat's session sandbox.
Frontend:
- Eagerly initialize the thread when switching to a new chat
- Resolve the thread ID once at request time and keep it stable through
async model-load waits
- Disable ActiveThreadSync during new-chat initialization to prevent
stale thread IDs from being written back
- Add error handling for thread initialization failures
- Clear activeThreadId on all compare-mode entry paths to prevent
cross-session leakage
- Fix exitCompare to restore context usage from the saved view
- Coerce falsy thread IDs to undefined for consistent backend/frontend
fallback behavior
- Use _default as the image sessionId fallback to match the backend
Backend:
- Use ~/studio_sandbox/_default when a request arrives without a session_id
* fix(studio): reuse HF cached repo casing to prevent duplicate downloads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Move cache case resolution tests to separate PR
Tests for resolve_cached_repo_id_case and get_model_config case resolution
belong in their own PR to keep this change focused on the runtime fix.
* fix(studio): debug-log HF_HUB_CACHE fallback in path_utils
* Fix stale memoization in resolve_cached_repo_id_case
- Check exact-case path before memo to ensure a newly-appeared exact
match always wins over a previously memoized variant
- Validate memoized entries still exist on disk before returning them
to prevent stale results when cache dirs are deleted/recreated
* Minor cleanups for cache case resolution
- Use .is_dir() instead of .exists() for exact-case cache check
(cache entries are always directories)
- Remove redundant fallback in _detect_audio_from_tokenizer since
get_cache_path already handles case resolution and returns None
when the model is not cached
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* feat: allow non-LLM recipes to run without provider block
* feat: reorder execution tabs and add generation-aware data tab empty state
* fix: add accessibility attrs to data tab spinner and use literal ellipsis
* fix(studio): use shared spinner, stub provider, and hide unused LLM metrics
Backend: inject stub model provider for sampler-only recipes so
DataDesigner init does not reject empty provider lists.
Frontend: use shared Spinner component, hide LLM columns metric
and model usage card when recipe has no LLM columns.
* Fix tab reset and terminal auto-scroll regressions for PR #4805
Reset detailTab to "data" when switching between executions so
the Data tab default is applied consistently, not only on first
mount. Also add detailTab to the terminal scroll effect deps so
auto-scroll-to-bottom fires when the user opens the Overview tab
after landing on Data.
* Guard terminal scroll reset to only fire on Overview tab
The previous scroll effect ran on every tab switch, which could
reset the user's manual scroll position if they scrolled up in
the terminal and briefly switched tabs. Now the scroll-to-bottom
and sticky-bottom reset only fires when navigating to the
Overview tab.
* Use None for stub provider api_key instead of literal string
The stub ModelProvider that satisfies the DataDesigner registry
for non-LLM recipes should not carry a fake credential string.
Using None avoids sending an Authorization header if the provider
is ever inadvertently invoked.
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Differentiate web_search query searches from URL fetches in the Studio chat UI.
Backend (llama_cpp.py):
- Emit "Reading: hostname" for URL fetches and "Searching: query" for query searches in SSE status events
- Only show hostname for valid http/https URLs; schemeless/non-http URLs get "Reading page..." generic fallback
- Strip www. prefix for consistency with the frontend
Frontend (tool-ui-web-search.tsx):
- Tool card shows "Read hostname" / "Reading hostname..." for URL fetches
- Shows "Searched query" / "Searching for query..." for query searches
- Uses new URL() with protocol check; falls back to "Read page" / "Reading page..." for non-http URLs
* fix(studio): lazy-import AutoConfig in model_config.py to fix transformers 5.x version switch
Move `from transformers import AutoConfig` from module level to inside
load_model_config() where it is actually used.
model_config.py is transitively imported at module load time via:
core/inference/__init__ → llama_cpp → utils.models → model_config
In inference subprocesses (mp.spawn), this chain runs before
_activate_transformers_version() can prepend .venv_t5/ to sys.path.
The eager import caches transformers 4.57.6 in sys.modules, and the
subsequent sys.path change has no effect — Python always checks
sys.modules before sys.path.
Making the import lazy ensures transformers is not loaded until after
version activation, so the subprocess picks up the correct version.
* fix(studio): also lazy-import extract_model_size_b in llama_cpp.py
Belt-and-suspenders: make the import that originally triggered the
chain lazy as well, so future module-level AutoConfig additions in
utils.models cannot reintroduce the problem.
* [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>
Gemma 4 is a native transformers 5.5 model and does not need
trust_remote_code=True. The auto-enable logic (added for NemotronH)
was catching all transformers 5.x models, including Gemma 4.
When trust_remote_code=True, unsloth_compile_transformers() returns
early without running the compiler. This disables the fused cross
entropy patch, causing logged training loss to be inflated by the
gradient_accumulation_steps factor.
Exclude models matching "gemma-4" or "gemma4" from the auto-enable
so the compiler runs and applies fused cross entropy correctly.
The model list merge order was `top_gguf + top_hub + static_models`,
which meant the HF download-ranked models always came first. New models
like Gemma 4 have low download counts and were not in the HF top-40,
so they got buried after 80 other models despite being at the top of
the curated static defaults in defaults.py.
Flip the merge to `static_models + top_gguf + top_hub` so editorial
picks (new model launches, promoted models) always appear first in the
Recommended section, with HF popularity backfilling after.
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Small GGUF models (<9B) frequently generate full code or lengthy
explanations instead of calling tools, bypassing the existing
plan-without-action re-prompt mechanism. Three issues:
1. _REPROMPT_MAX_CHARS=500 was too low -- models that output full
HTML/code responses (often 1000+ chars) never triggered the
re-prompt at all, since it only fires on short responses.
2. _MAX_REPROMPTS=1 gave the model only one chance to comply.
Small models often need 2-3 nudges before switching from
text generation to tool calling.
3. The re-prompt text ("Please use the available tools...") was
too polite for small models to follow reliably.
4. Tool-calling detection missed chat templates using Jinja
whitespace-trimming syntax ({%- if tools -%}) since only
({%- if tools %}) and ({% if tools %}) were checked.
Changes:
- Raise _REPROMPT_MAX_CHARS from 500 to 2000 so longer responses
(code blocks, multi-paragraph plans) still trigger re-prompts
- Raise _MAX_REPROMPTS from 1 to 3 for more retry budget
- Use direct, imperative re-prompt language that small models
follow more reliably ("STOP. You MUST call a tool NOW.")
- Strengthen the system prompt tool nudge to explicitly forbid
outputting code blocks (redirect to the python tool instead)
- Add Jinja whitespace-trimmed variants to the tool_markers
list so all template styles are detected correctly
* UI Changes
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove unrelated test file
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* feat(studio): display images from Python tool execution in chat UI
When the model calls the Python tool to create a matplotlib plot or
other image file, the image now displays inline in the chat output
instead of being invisible to the user.
Backend:
- Detect new image files (png/jpg/gif/webp/bmp) after Python subprocess
completes by diffing os.listdir before/after execution
- Append __IMAGES__ sentinel to tool result for frontend consumption
- Strip sentinel before injecting result into LLM context (role: tool)
so the model never sees file paths
- Add GET /sandbox/{session_id}/{filename} endpoint with JWT auth
(header or query param), path traversal protection, extension
allowlist, realpath containment check, and nosniff header
Frontend:
- Parse __IMAGES__ sentinel in tool_end SSE events, create structured
result with text/images/sessionId
- Render <img> tags in Python tool UI pointing at the sandbox endpoint
Also fixes a bug where SyntaxError in user code was misreported as
"unsafe code detected" instead of showing the actual Python traceback.
The _check_code_safety function now lets SyntaxError pass through to
the subprocess for a proper error message.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): improve SVG detection and strip XML preamble
Handle <?xml ...?> declarations before <svg> tags in code fences,
strip XML declaration from SVGs before data URI rendering, and
update the sloth suggestion prompt to request showing code.
* fix(studio): persist parentId so retries survive reload
The append() handler was destructuring only { message } from
ExportedMessageRepositoryItem and discarding parentId. When loading
a saved thread, load() used ExportedMessageRepository.fromArray()
which chains all messages sequentially, flattening retry branches
into a linear list.
Now append() writes parentId to the MessageRecord, and load()
reconstructs the tree when parentIds are present. Old threads
without parentId fall back to the existing fromArray() behavior.
* fix(studio): address review findings for image display and retry persistence
Image detection:
- Use mtime comparison instead of filename-only diff so overwritten
files (e.g. plt.savefig("chart.png") called twice) are detected
Sentinel parsing:
- Use rsplit/lastIndexOf instead of split/indexOf so user code that
prints __IMAGES__: does not collide with the backend sentinel
Mixed legacy/new threads:
- For old messages without a stored parentId, infer sequential parent
from the previous message instead of null, preventing multiple roots
Sandbox endpoint:
- Change Cache-Control from "public, max-age=3600" to "private,
no-store" since these are authenticated responses
---------
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(studio): prevent small models from stalling on tool-calling tasks
Small GGUF models (< 9B params) in "Think, Search, Code" mode would
often describe what they planned to do ("Let me create this dashboard")
and then stop generating without ever calling a tool.
Three changes:
1. Simplify web_tips for small models: remove the "fetch its full content
by calling web_search with the url parameter" guidance for models < 9B.
This multi-step instruction causes small models to plan elaborate
search-then-fetch-then-code sequences they cannot reliably execute.
2. Add "always call tools directly" imperative to the system prompt nudge
so models act immediately instead of narrating their intentions.
3. Add plan-without-action re-prompt in the agentic loop: when the model
emits planning text (matching patterns like "let me", "I'll", etc.)
without calling any tool, inject a nudge asking it to call the tool
and continue the loop. Capped at 2 re-prompts per request.
Benchmarked with Qwen3.5-4B-GGUF (N=5 trials per variant):
- Baseline: 40% of requests had any tool call
- Combined fix: 100% of requests had at least one tool call
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(studio): allow context length slider to reach model's native limit
The context length slider was hard-capped to the VRAM-estimated maximum,
preventing users from requesting higher context even though the backend
already handles it safely (multi-GPU selection, --fit fallback). Expose
the model's native context length from GGUF metadata as a separate API
field and use it as the slider ceiling instead. Add an amber warning
when the selected context exceeds the estimated VRAM capacity.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Raise VRAM budget to 90% and add native_context_length tests
Increase the GPU memory utilization threshold from 70% to 90% across
_select_gpus and _fit_context_to_vram, allowing longer context lengths
before VRAM capping kicks in.
Add 33 tests for the native_context_length feature covering the backend
property, context value separation invariants, Pydantic models, route
completeness, edge cases, and cross-platform binary I/O.
* [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 SSL handshake failures (SSLV3_ALERT_HANDSHAKE_FAILURE, CERTIFICATE_VERIFY_FAILED) when fetching HTTPS pages by introducing _PinnedHTTPSConnection that separates TCP connect (to pinned IP) from TLS handshake (with real hostname for SNI/cert verification)
- Fix SSRF DNS-rebinding vulnerability: previous impl swapped conn.host before connect(), causing fresh DNS resolution; new subclass keeps TCP pinned to validated IP
- Fix SPA/JS-rendered doc sites returning empty content by rotating real browser User-Agents (Chrome/Firefox/Safari)
- Strip nav/footer from HTML-to-Markdown output so article content is not buried under navigation chrome
- Increase raw fetch cap from 64KB to 512KB so SSR article content is reached on GitBook/Docusaurus/Next.js pages
- Fix IPv6 address bracketing in URL netloc construction
- Hoist SSL context, handler classes, and stdlib imports to module level (created once, not per-call)
- Use consistent UA across redirect hops to avoid breaking session-aware bot detection
* feat(studio): architecture-aware KV cache VRAM estimation
Replace the single legacy formula (2 * n_kv_heads * head_dim * n_layers
* n_ctx * bpe) with 5-path estimation that reads 8 additional GGUF
metadata fields:
1. MLA (DeepSeek-V2/V3, GLM-4.7, GLM-5, Kimi-K2.5) -- K-only cache
using compressed KV latent + RoPE; no separate V allocation
2. Hybrid Mamba (Qwen3.5-27B, Qwen3.5-35B-A3B) -- only attention
layers (1 in N) carry KV; Mamba layers have none
3. Sliding Window (Gemma-3, gpt-oss) -- SWA layers cache
min(ctx, window) tokens instead of the full context
4. Standard GQA -- uses explicit key_length/value_length from GGUF
instead of embed // n_heads (which is wrong for many models)
5. Legacy fallback -- identical to old formula for old GGUFs
New GGUF fields parsed: attention.key_length, attention.value_length,
attention.sliding_window, full_attention_interval,
attention.kv_lora_rank, attention.key_length_mla, ssm.inner_size,
ssm.state_size.
Validated against 9 real GGUF files (72/72 field checks pass).
The legacy formula was off by +682% for Gemma-3 and -81% for
DeepSeek-V3.1.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix MLA fallback and SWA global/local ratio heuristic
Two fixes based on review findings:
1. MLA fallback now uses key_length_mla from GGUF metadata instead of
hardcoded rope_dim=64. Falls back to 64 only when key_length_mla is
absent. This ensures correct estimates for MLA variants that use
rope dimensions other than 64.
2. SWA global/local layer ratio changed from 50/50 to 1/4 (25% global,
75% SWA). Most sliding window architectures have predominantly local
layers (Gemma-3 uses ~17% global, gpt-oss uses ~50%). The 1/4
heuristic is closer to the common case and still a large improvement
over the legacy formula which ignores SWA entirely.
* Tighten _can_estimate_kv gate and treat sliding_window=0 as disabled
Two additional fixes from review round 1 (5/8 and 4/8 reviewer consensus):
1. _can_estimate_kv now requires BOTH key_length AND value_length for
the explicit-dims path. Previously key_length alone was enough,
which could cause silent fallthrough to the legacy formula with
fabricated defaults (n_kv=1, head_dim=128) when value_length was
absent from the GGUF.
2. SWA path now requires sliding_window > 0. Some GGUFs use 0 as a
disabled sentinel. Without this guard, min(ctx, 0) would zero out
all SWA layer contributions, severely underestimating KV cache.
* Fix MLA n_kv safety and use ceiling division for hybrid path
Addresses Gemini Code Assist review findings:
1. MLA path now uses n_kv_mla = n_kv_heads or 1 (not n_heads). This
prevents a 128x overestimate for DeepSeek-V3 if head_count_kv is
absent from the GGUF (n_heads=128 would have been used instead).
2. Hybrid path now uses ceiling division for attention layer count.
This prevents undercounting by 1 when n_layers is not perfectly
divisible by full_attention_interval.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix: clear tool status badge immediately after tool execution
The tool status timer badge (Searching 1s, 2s...) persisted after
tool calls finished because the status clear event was only sent
at the start of the next generation iteration, not after tool
execution completed.
Backend: yield status clear after all tools finish in the agentic
loop iteration, before continue starts the next generation pass.
Frontend: debounce badge visibility by 300ms so sub-second tool
calls dont flash the badge.
* Fix debounce regression for consecutive tool calls
Only apply the 300ms show-delay when transitioning from idle to
tool-active. When switching between consecutive tools in the same
turn (e.g. web_search -> python), keep the badge visible immediately
so it does not flicker or disappear during multi-tool runs.
* Delay wasActiveRef reset to bridge inter-iteration tool gaps
The backend emits a status-clear event between tool iterations,
which was resetting wasActiveRef immediately and causing the next
tool to be re-debounced (300ms hidden gap between consecutive tools
in the same turn). Now the ref reset is delayed by 500ms so a
follow-up tool within the same agentic turn shows the badge
immediately, while a genuinely new turn still gets the debounce.
* Use thread lifecycle to track tool-run boundaries
Replace the 500ms wall-clock timeout with the actual thread.isRunning
state to determine when wasActiveRef should reset. This properly
handles all cases:
- Consecutive tools within the same run stay visible without flicker
- The badge hides only when the thread run actually ends
- New turns always get a fresh 300ms debounce on the first tool
- No heuristic timeout that can misfire on slow or fast inference
* Consolidate wasActiveRef reset into single effect
Removes the separate isThreadRunning effect to avoid a race where
the ref resets before the tool-status effect reads it (when
isThreadRunning flips to false before setToolStatus(null) from
the adapter's finally block). Now wasActiveRef resets only when
both toolStatus is null AND the thread run has ended, eliminating
any flicker on the last tool of a run.
* Simplify debounce: use visible state instead of ref tracking
Drop wasActiveRef entirely and use the visible state as the
debounce gate. When the badge is not yet on screen, debounce
for 300ms before showing. When already visible from a prior tool,
keep showing immediately. This correctly handles all cases:
- All fast tools (<300ms) are suppressed, not just the first
- Consecutive tools after the badge is shown stay visible
- Badge persists across inter-iteration clears while thread runs
- New turns get a fresh debounce after visible resets
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Simplify tool-call dedup: drop hashlib, inline helpers
The duplicate tool-call detector only compares calls within a single
request from the same JSON parser, so dict key order is guaranteed
identical for identical calls (Python 3.7+ insertion-ordered dicts).
- Replace hashlib.md5(json.dumps(...)) with name + str(args)
- Inline _tool_call_key, _is_duplicate_call, _record_tool_call
since each was a one-liner used once
- Remove unused hashlib import
* Remove tool_calling_benchmark_results.md from repo
* Replace html2text with builtin HTML-to-Markdown converter
Drop the external html2text (GPL-3.0) dependency and its regex
fallback. Add _html_to_md.py (~190 lines, stdlib only) using
html.parser.HTMLParser that handles headings, links, bold/italic,
lists, tables, blockquotes, code blocks, and entity decoding.
Strips script/style/head tags entirely.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use json.dumps(sort_keys=True) for tool-call dedup key
str(dict) is sensitive to insertion order, so semantically identical
calls with different key ordering would bypass duplicate detection.
Switch to json.dumps with sort_keys=True for a canonical representation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Revert dedup key to str(arguments)
json.dumps(sort_keys=True) is unnecessary here -- the arguments dict
always comes from the same JSON parser within a single request, so
key insertion order is deterministic (Python 3.7+). str() is faster
and sufficient for consecutive-call dedup.
* Address review comments on _html_to_md.py
- Remove "hr" from _BLOCK_TAGS so the dedicated hr handler is reachable
- Prefix all newlines with ">" inside blockquotes (multi-line support)
- Emit full  for images instead of alt text only
- Replace newlines with spaces inside table cells
- Track header cells per-row (_row_has_th) instead of last-cell-only
- Strip trailing tabs in addition to spaces in cleanup regex
* Fix blockquote rendering, truncated-HTML buffer flush, and dedup key canonicalization
_html_to_md.py:
- Rewrite blockquote handling with stack-based buffer approach so nested
blockquotes, pre blocks inside blockquotes, and multi-paragraph quotes
all render correctly with proper "> " prefix on every line.
- Add flush_pending() to recover content from truncated HTML where closing
tags are missing (common when _fetch_page_text caps the download size).
Flushes open <a>, <td>, <pre>, and blockquote buffers.
- Skip <img> tags to match prior html2text ignore_images=True behavior
and avoid data-URI amplification consuming the output budget.
- Collapse all whitespace (including newlines) in non-pre content per
standard HTML whitespace rules: \s+ -> single space.
- Escape pipe characters in table cell content to prevent column breakage.
- Emit separator row after the first row for tables without <th> headers.
- Guard against IndexError on _ol_counter for orphan <li> elements.
- Normalize CRLF line endings before parsing.
llama_cpp.py:
- Restore canonical dedup key with json.dumps(sort_keys=True) so that
semantically identical tool calls with different JSON key order are
correctly detected as duplicates.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix table optional end tags, inline code whitespace, and link text normalization
_html_to_md.py:
- Extract _finish_cell() and _finish_row() helpers to handle HTML tables
that omit optional </td>, </th>, or </tr> end tags. This is valid HTML
and common on real web pages -- previously the parser would silently
drop earlier cells and entire rows.
- Call _finish_cell()/_finish_row() from handle_starttag for <tr>/<td>/<th>,
handle_endtag for </tr>/<td>/<th>/<table>, and flush_pending() so all
three paths (normal close, implicit close, truncated HTML) use the same
row-finalization logic including header separator emission.
- Add _in_inline_code flag so handle_data() preserves literal whitespace
inside <code> spans instead of collapsing it. Source like
<code>pip install unsloth</code> now correctly renders as
`pip install unsloth` rather than `pip install unsloth`.
- Extract _finish_link() helper that normalizes accumulated link text with
\s+ -> single space before building the Markdown link. Prevents block-
level content inside <a> tags (e.g. <a><div>one</div><div>two</div></a>)
from producing multiline [one\n\ntwo](href) link labels.
- Empty blockquotes now produce no output instead of a stray ">".
- Remove unused _bq_depth field (all routing uses _bq_stack).
- Flush open cells and rows in handle_endtag("table") for robustness.
* Support <ol start=N>, <dl>/<dt>/<dd>, and preserve code block whitespace
_html_to_md.py:
- Honor <ol start="N"> attribute so ordered lists preserve their original
numbering instead of always restarting from 1. Important for docs/tutorials
that continue numbering across sections.
- Add dl, dt, dd to _BLOCK_TAGS so definition lists (common on MDN, Python
docs, Django docs) produce separated text instead of concatenated blobs.
- Rewrite _cleanup() to be fence-aware: content inside fenced code blocks
is now preserved verbatim (intentional blank lines in <pre> content are
no longer collapsed). Outside code blocks, blank runs are limited to one
and trailing whitespace is stripped.
- Fix _prefix_blockquote() to strip trailing whitespace before collapsing
blank lines, preventing the "\n\n \n\n" pattern from sneaking through.
* Suppress whitespace-only text nodes between table structural elements
Indented HTML tables (nearly all real-world pages) produce whitespace
text nodes between <table>, <tr>, </tr> etc. that land in the output
as leading spaces before table rows, breaking Markdown table alignment.
Skip whitespace-only text nodes when inside a table but not inside a
cell, so indentation from source HTML does not leak into the output.
* Revert dedup key to str(arguments) with explanatory comment
json.dumps(sort_keys=True) is unnecessary overhead here: arguments
always comes from json.loads on model output within a single request,
so dict insertion order is deterministic in Python 3.7+. A repeated
call from the model produces the same JSON, which parses to the same
dict repr. str() avoids re-serialization on every tool call.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* studio: improve GGUF tool calling accuracy and reliability
- Add URL fetching to web_search tool so models can read full page
content instead of only getting search snippets. Uses html2text for
clean markdown conversion with regex fallback.
- Inject current date and behavioral guidance (URL fetch workflow,
no repeated queries, use code for data processing) into the
tool-use system prompt.
- Append error recovery nudge to tool results that indicate failure,
helping small models avoid looping on the same broken call.
- Strip leaked <tool_call> XML from assistant messages in conversation
history and from the outgoing SSE stream.
- Raise default max tool iterations from 10 to 25 across backend,
model schema, and frontend defaults.
- Increase _MAX_PAGE_CHARS from 4k to 16k so fetched pages contain
enough content for the model to extract useful information.
- Add "IMPORTANT: These are only short snippets" hint to search
results so models know to fetch full pages when needed.
Tested with Qwen3.5-4B-GGUF (UD-Q4_K_XL), 10 runs before/after:
- XML leaks in responses: 10/10 -> 0/10
- URL fetch usage: 0 -> 4/10 runs
- Runs producing actual correct answers: 0/10 -> 2/10
- Average tool calls per query: 5.5 -> 3.8 (more efficient)
- Average response time: 12.3s -> 9.8s
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add tool calling benchmark results across model sizes and quants
Tested 16 configurations (4 models x 2 quants x 2 KV cache types)
with 10 runs each on NVIDIA B200.
Best config: 27B UD-Q4_K_XL + bf16 KV -- 6/10 runs found all 4
correct songs, 0 XML leaks, 131s average response time.
* Add duplicate tool-call detection and final-answer synthesis
When the model repeats the exact same tool call (same name + arguments)
twice in a row, skip execution and return a redirect message telling it
to try a different approach. This prevents the 8x-repeated-query loops
observed on 27B and 35B models.
When the tool iteration cap (25) is reached, inject a "provide your
final answer now" message before the final streaming pass. This lets
the model synthesize a useful answer from everything it gathered
instead of being silently cut off.
Tested on Qwen3.5-27B UD-Q4_K_XL (10 runs):
- Repeated query runs: 4/10 -> 2/10
- Cap hits: 1/10 -> 0/10
- All 4/4 accuracy: 5/10 -> 7/10
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix CodeQL alert: handle whitespace in script/style closing tags
The regex fallback for HTML stripping did not match closing tags
with whitespace before the angle bracket (e.g. </script >).
Use \s* before > in both script and style patterns.
* Address reviewer findings: SSRF, timeout crash, XML regex, dedup
- SSRF: resolve hostname via getaddrinfo and reject private, loopback,
link-local, multicast, and reserved addresses before fetching
- Timeout: handle timeout=None (unlimited mode) in URL fetch path
by defaulting to 60s instead of crashing on min(None, 60)
- Download cap: read at most max_chars*4+1 bytes instead of the
full response body before truncating
- XML regex: match both <tool_call> and <function=...> markup in
the history/stream cleanup (inference.py)
- CodeQL: use [^>]* in closing script/style tags to handle any
whitespace or attributes before >
- Dedup: track whether each tool call failed so retries after
transient errors are allowed; only block consecutive identical
calls that both succeeded
- Final-answer synthesis: guard on max_tool_iterations > 0 so
callers who disable tools do not get a false "used all calls" turn
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix redirect SSRF, SSE streaming regression, dedup off-by-one
- SSRF redirect bypass: disable auto-redirect in urllib, manually
follow up to 5 hops with host validation at each step. Prevents
public URLs from redirecting to loopback/private targets.
- SSE streaming: track prev_text on the raw cumulative and strip
XML from the delta only, so completed tool_call tags do not cause
the cumulative to shrink and drop trailing real text.
- Dedup off-by-one: check the immediately previous call (window=1)
instead of requiring 2 matching history entries, so the second
identical successful call is blocked rather than the third.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix redirect HTTPError handling and tighten error prefixes
- Redirect fix: urllib raises HTTPError (not a normal response) when
the redirect handler returns None. Catch HTTPError for 3xx codes
and extract the Location header from the exception object.
- Error prefixes: remove overly broad "No " prefix that matched
"No results found." (a valid empty-search outcome, not an error).
Replace with specific prefixes like "Blocked:", "No query provided",
"Failed to resolve". This ensures empty search results are correctly
classified as non-errors for duplicate-call tracking.
* Fix SSE cross-chunk XML leaks, cleanup review findings
- SSE streaming: sanitize the full cumulative text before diffing
against the previous sanitized snapshot, so XML tags that span
chunk boundaries are stripped correctly. The previous delta-based
approach leaked split tags.
- DRAINING fallback: use _strip_tool_markup() helper instead of a
manual regex that only handled <tool_call> but not <function=...>.
- Move hashlib import, _TOOL_XML_RE compile, and datetime import to
module level per style guide.
- Remove unused _hit_tool_cap variable.
* Fix DNS rebinding, charset detection, HTTPError handling, dedup double-record
- DNS rebinding: resolve hostname once via getaddrinfo, pin the
returned IP, rewrite the URL to connect to the pinned IP with
a Host header. Each redirect hop re-resolves and re-validates.
Closes the TOCTOU window between validation and connection.
- Charset: use resp.headers.get_content_charset() instead of
hardcoding utf-8, so pages with other encodings decode correctly.
- HTTPError: return descriptive "HTTP {code} {reason}" instead of
re-raising into a generic "Search failed" message.
- Dedup: remove redundant _record_tool_call in the duplicate branch;
the single call at the end of the loop handles all 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>
* fix: auto-retry stalled HF downloads with HF_HUB_DISABLE_XET=1
The heartbeat thread now monitors the HF Hub cache directory for
file-size growth. If no bytes are written for 3 minutes, it sends a
"stall" message to the orchestrator, which kills the subprocess and
retries with HF_HUB_DISABLE_XET=1 (falling back from Xet to standard
HTTPS). If the retry also stalls, it errors out with a clear message.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: include transport type (xet/https) in heartbeat and stall log messages
Makes it clear in backend logs whether the download is using xet or
https transport, and which transport stalled — helpful for debugging.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: monitor HF Hub .tmp dir to avoid false stall detections
huggingface_hub downloads into .tmp/ before atomically moving to
blobs/. Without monitoring .tmp, a large shard actively downloading
for several minutes would show zero blob growth and trigger a false
stall.
* fix: scope HF cache size check to specific model being loaded
Instead of scanning every models--*/blobs directory (O(N) with cached
models), only check the specific model's blobs dir plus the global
.tmp dir. Much faster on systems with many cached models.
* Fix false stall detection on cached/local models and cleanup issues
- Only fire stall if download activity was observed (cache size changed
at least once). Previously, any model load taking >180s would trigger
a false stall, even for already-cached or local models where no
download is happening.
- Return -1 from _get_hf_cache_size on exception to distinguish
"unable to measure" from "genuinely zero bytes". Skip stall logic
when measurement fails.
- Add _shutdown_subprocess before raising on terminal stall path to
prevent leaking a stuck subprocess.
- Detect pre-existing HF_HUB_DISABLE_XET=1 in the parent environment
to avoid a redundant retry cycle when Xet is already disabled.
- Remove global .tmp directory scanning (not used by modern
huggingface_hub; in-progress downloads use .incomplete files in
blobs/ which are already captured by iterdir).
- Add f.is_file() guard in cache size calculation.
- Replace em dashes with ASCII dashes for Windows terminal compat.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden stall detection edge cases
- Guard -1 to valid value transition: when initial _get_hf_cache_size
returns -1 (error) and later recovers to a real value, do not count
that as download activity. Only set saw_download_activity when the
previous measurement was also valid (>= 0).
- Move os import to top-level in orchestrator.py instead of inline
import os as _os.
- Fix misleading comment about post-download protection.
* Use .incomplete files to detect active downloads for stall detection
Replace the saw_download_activity heuristic with direct .incomplete file
detection. huggingface_hub creates *.incomplete files in blobs/ during
active downloads and removes them on completion. This gives a reliable
signal for whether a download is actually in progress.
Benefits:
- Cached models: no .incomplete files -> no stall fired even after 180s
- Post-download init (quantization, GPU loading): .incomplete files gone
so stall timer resets, long init phases are not killed
- Pre-download hangs (XET handshake stall): .incomplete files are
created at download start, so zero-byte stalls are now detected
- No more false positives from -1 to valid measurement transitions
The _get_hf_download_state function now returns (total_bytes,
has_incomplete) tuple or None on error, replacing _get_hf_cache_size.
* Add debug logging to download state exception handler
Log the exception at debug level when _get_hf_download_state fails,
instead of silently returning None. Helps with troubleshooting cache
measurement issues.
* Watch both adapter and base model repos for LoRA stall detection
When loading a LoRA adapter, the actual download bottleneck is often
the base model, not the adapter itself. Update the heartbeat to watch
both mc.identifier and mc.base_model cache directories so stall
detection works for LoRA loads where the base model stalls on Xet.
Also update _get_hf_download_state to accept multiple model names and
skip names without "/" (local paths) since those do not have HF cache
directories.
* Fix model name filtering for official HF models without org prefix
Models like gpt2 and bert-base-uncased do not contain a slash but are
still valid HF Hub models with cache directories. Replace the "/" check
with a proper local-path detection that checks for path separators and
path-like prefixes instead.
Also fix the base_model watch list to not require "/" in the base model
name, so official models used as LoRA bases are also monitored.
* Fix local path detection that broke all org/model names on Linux
The os.path.sep check matched "/" in HF model IDs like "org/model" on
Linux, causing the stall detector to skip ALL standard HF models.
Replace with a check that only skips names starting with "/" (absolute
paths), "." (relative paths), "~" (home-relative), or containing "\"
(Windows paths). HF model IDs like "org/model" or "gpt2" pass through
correctly on all platforms.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): change default weight_decay from 0.01 to 0.001
The default weight decay across Studio was 0.01 but should be 0.001.
Updated the default in all backend fallbacks, the Pydantic model, the
frontend config, and every YAML preset/model-default config.
* fix(studio): auto-set learning rate based on training method
Default LR should be 2e-4 for LoRA/QLoRA and 2e-5 for full fine-tuning.
Frontend: track whether the user has manually edited the LR field via a
_learningRateManuallySet flag (same pattern as trainOnCompletions).
When switching training method and the user has not touched the LR,
auto-set it to the appropriate default. Reset the flag on model load.
Backend: change trainer.py start_training default from 5e-5 to 2e-4,
update default.yaml fallback from 5e-5 to 2e-4, and fix
full_finetune.yaml from 0.0002 (2e-4) to 2e-5.
* refactor(studio): centralize weight_decay and learning rate defaults
Create studio/backend/core/training/constants.py as the single source of
truth for DEFAULT_WEIGHT_DECAY (0.001), DEFAULT_LEARNING_RATE (2e-4),
DEFAULT_LEARNING_RATE_FULL (2e-5), and DEFAULT_LEARNING_RATE_STR ("2e-4").
All backend modules (trainer.py, training.py, worker.py, models/training.py)
now import from constants.py instead of hardcoding values.
On the frontend, add LR_DEFAULT_LORA and LR_DEFAULT_FULL to
config/training.ts and use them in the store instead of magic numbers.
A comment cross-references the backend constants file.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model-specific LR override, persist migration, and flag resets
- Preserve model-specific learning rates from YAML configs when the
async autoSelectTrainingMethod callback fires (fixes Qwen2.5-1.5B
getting 2e-4 instead of its configured 1e-5, etc.)
- Bump zustand persist version to 9 with migration so existing users
with weightDecay=0.01 get updated to 0.001
- Clear _learningRateManuallySet in reset() and applyConfigPatch()
for consistency with trainOnCompletions flag behavior
- Add DEFAULT_LEARNING_RATE_FULL_STR to constants.py
* Refine applyConfigPatch to only clear LR flag when patch includes LR
Only reset _learningRateManuallySet when the applied config patch
actually provides a learningRate value. This prevents unrelated config
patches from silently disarming the manual-edit guard, which would
cause a subsequent setTrainingMethod call to overwrite the user's
custom LR.
* Preserve model-specific LR when switching between qlora and lora
Only auto-switch the learning rate when the training category changes
(adapter <-> full fine-tuning). Switching between qlora and lora keeps
the current LR since both methods share the same learning rate range.
This preserves curated per-model defaults (e.g. 1e-5 for
Qwen2.5-1.5B-Instruct) when the user toggles between adapter methods.
* Remove constants.py, use YAML configs as the source of truth
The YAML config files (model-specific + default.yaml) are the intended
config layer for training defaults. The Python backend fallbacks now use
inline values that match the YAML configs, rather than importing from a
separate constants module. This keeps the config architecture simple:
YAML files are the single source of truth, and the inline Python
fallbacks are just safety nets that mirror them.
* fix(studio): preserve model-specific LR when switching training method
Stash YAML-provided learning rate and use it to restore the correct
value when switching between adapter and full fine-tune modes.
- qlora <-> lora no longer overwrites the model's LR
- full -> adapter restores the YAML LR instead of a hardcoded constant
- selecting a model while on full fine-tune uses LR_DEFAULT_FULL
instead of applying the YAML adapter LR
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Co-authored-by: Roland Tannous <rolandtannous@gravityq.ai>
* Revert to balanced for inference
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove unused for_inference parameter from get_device_map
Since inference and training both use "balanced" now, the for_inference
flag is dead code. Remove it from the function signature, the call site
in inference.py, and simplify the tests accordingly.
* Remove redundant TestDeviceMapForInference test class
TestGpuAutoSelection already covers the same multi-gpu and single-gpu
device_map assertions. The TestDeviceMapForInference class was left
over from when for_inference had distinct behavior.
* Remove redundant test_get_device_map_multi_gpu_uses_balanced
Its assertions ([0,1] -> balanced, [0] -> sequential) are already
covered by test_get_device_map_uses_explicit_gpu_selection.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
The 180s wall-clock timeout would kill model loads on slow connections
even when the download was actively progressing. Now the worker sends
heartbeat status messages every 30s during loading, and the orchestrator
resets its 300s deadline on each one — so it only times out when the
subprocess goes truly silent.
* [WIP] balanced device map for studio
* gpus as a request parameter
* API for multi GPU stuff
* return multi gpu util in new API
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use balanced_low0 instead of balanced
* Use balanced_low0 instead of balanced
* Fix device_map typo, UUID parsing crash, set() filter bug, and broken tests
- balanced_low0 -> balanced_low_0 (transformers/accelerate rejects the old string)
- get_parent_visible_gpu_ids() now handles UUID/MIG CUDA_VISIBLE_DEVICES
gracefully instead of crashing on int() parse
- _get_backend_visible_gpu_info() set() or None bug: empty set is falsy so
CUDA_VISIBLE_DEVICES=-1 would disable filtering and report all GPUs
- test_gpu_selection.py: add missing get_visible_gpu_utilization import and
add required job_id arg to start_training() calls
* Smart GPU determinism using estimates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* disallow gpu selection for gguf for now
* cleanup
* Slightly larger baseline
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Treat empty list as auto
* Verbose logging/debug
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cleanup and revert unnecessary deletions
* Cleanup excessive logs and guard against disk/cpu offload
* auth for visibility API. cleanup redundant imports. Adjust QLoRA estimate
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* support for non cuda gpus
* Fix multi-GPU auto-selection memory accounting
The multi_gpu_factor was applied uniformly to all GPUs including the
first one, which unfairly penalizes single-GPU capacity when
transitioning to multi-GPU. This created a discontinuity where a model
that barely fits 1 GPU would suddenly require 2 GPUs because the first
GPU's free memory was discounted by 20%.
Now the first GPU keeps its full free memory, and only additional GPUs
have an overhead factor (0.85) applied to account for inter-GPU
communication and sharding overhead. This gives more accurate
auto-selection and avoids unnecessary multi-GPU for models that
comfortably fit on one device.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add sandbox tests for multi-GPU selection logic
24 tests covering model size estimation, memory requirements, automatic
GPU selection, device map generation, GPU ID validation, and multi-GPU
overhead accounting. All tests use mocks so they run without GPUs on
Linux, macOS, and Windows.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix reviewer findings: 4bit inference estimate, fallback, GGUF gpu_ids, retry
1. 4-bit inference now uses reduced memory estimate (model_size/3 + buffer)
instead of the FP16 1.3x multiplier. This prevents over-sharding
quantized models across unnecessary GPUs.
2. When model size estimation fails, auto_select_gpu_ids now falls back to
all visible GPUs instead of returning None (which could default to
single-GPU loading for an unknown-size model).
3. GGUF inference route now treats gpu_ids=[] as auto-selection (same as
None) instead of rejecting it as an unsupported explicit request.
4. Training retry path for "could not get source code" now preserves the
gpu_ids parameter so the retry lands on the same GPUs.
5. Updated sandbox tests to cover the new 4-bit inference estimate branch.
* Remove accidentally added unsloth-zoo submodule
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix UUID/MIG visibility and update test expectations
1. nvidia.py: When CUDA_VISIBLE_DEVICES uses UUID/MIG tokens, the
visibility APIs now return "unresolved" with empty device lists instead
of exposing all physical GPUs. This prevents the UI from showing GPUs
that the backend process cannot actually use.
2. test_gpu_selection.py: Updated test expectations to match the new
multi-GPU overhead accounting (first GPU at full capacity, 0.85x for
additional GPUs) and 4-bit inference memory estimation formula.
All 60 tests now pass.
* Add CPU/disk offload guard to audio inference path
The audio model loading branch returned before the common
get_offloaded_device_map_entries() check, so audio models loaded with a
multi-GPU device_map that spilled layers to CPU/disk would be accepted
instead of rejected. Now audio loads also verify no modules are offloaded.
* Improve VRAM requirement estimates
* Replace balanced_low_0 with balanced
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* refine calculations for slightly easier nums
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* adjust estimates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use nums instead of obj to avoid seralisation error
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden nvidia-smi parsing and fix fallback GPU list
1. nvidia.py: Wrap int() casts for GPU index and memory in try/except
so MIG slices, N/A values, or unexpected nvidia-smi output skip the
unparseable row instead of aborting the entire GPU list.
2. nvidia.py: Handle GPU names containing commas by using the last
field as memory instead of a fixed positional index.
3. hardware.py: fallback_all now uses gpu_candidates (GPUs with verified
VRAM data) instead of raw devices list, which could include GPUs
with null VRAM that were excluded from the ranking.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* cleanup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* consolidate raise_if_offload
* Improve MoE support. Guard against nvidia-smi failures
* Improve MoE support. Guard against nvidia-smi failures
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix shared-expert LoRA undercount, torch VRAM fallback, and apply_gpu_ids edge case
1. vram_estimation.py: compute_lora_params now includes shared experts
(n_shared_experts) alongside routed experts when computing MoE LoRA
adapter parameters. Previously only n_experts were counted, causing
the estimator to undercount adapter, optimizer, and gradient memory
for DeepSeek/GLM-style models with shared experts.
2. hardware.py: _torch_get_per_device_info now uses mem_get_info (which
reports system-wide VRAM usage) instead of memory_allocated (which
only reports this process's PyTorch allocations). This prevents
auto-selection from treating a GPU as mostly free when another
process is consuming VRAM. Falls back to memory_allocated when
mem_get_info is unavailable.
3. hardware.py: apply_gpu_ids([]) now returns early instead of setting
CUDA_VISIBLE_DEVICES="" which would disable CUDA entirely. Empty
list inherits the parent visibility, same as None.
4. hardware.py: Upgraded fallback_all GPU selection log from debug to
warning so operators are notified when the model likely will not fit
in available VRAM.
* Guard nvidia-smi subprocess calls against OSError and TimeoutExpired
get_visible_gpu_utilization and get_backend_visible_gpu_info now catch
OSError (nvidia-smi not found) and TimeoutExpired internally instead
of relying on callers to wrap every invocation. Returns the standard
available=False sentinel on failure so the torch-based fallback in
hardware.py can take over.
* Guard get_primary_gpu_utilization and reset GPU caches between tests
1. nvidia.py: get_primary_gpu_utilization now catches OSError and
TimeoutExpired internally, matching the pattern already used in
get_visible_gpu_utilization and get_backend_visible_gpu_info. All
three nvidia-smi callers are now self-contained.
2. test_gpu_selection.py: Added _GpuCacheResetMixin that resets the
module-level _physical_gpu_count and _visible_gpu_count caches in
tearDown. Applied to all test classes that exercise GPU selection,
device map, or visibility functions. This prevents stale cache
values from leaking between tests and causing flaky results on
machines with real GPUs.
* Fix nvidia-smi fallback regression and physical GPU count validation
1. hardware.py: get_gpu_utilization, get_visible_gpu_utilization, and
get_backend_visible_gpu_info now check result.get("available") before
returning the nvidia-smi result. When nvidia-smi is unavailable or
returns no data (e.g., containers without nvidia-smi, UUID/MIG masks),
the functions fall through to the torch-based fallback instead of
returning an empty result. This fixes a regression where the internal
exception handling in nvidia.py prevented the caller's except block
from triggering the fallback.
2. hardware.py: resolve_requested_gpu_ids now separates negative-ID
validation from physical upper-bound validation. The physical count
check is only enforced when it is plausibly a true physical count
(i.e., higher than the largest parent-visible ID), since
torch.cuda.device_count() under CUDA_VISIBLE_DEVICES returns the
visible count, not the physical total. The parent-visible-set check
remains authoritative in all cases. This prevents valid physical IDs
like [2, 3] from being rejected as "out of range" when nvidia-smi is
unavailable and CUDA_VISIBLE_DEVICES="2,3" makes torch report only
2 devices.
* Fix UUID/MIG torch fallback to enumerate devices by ordinal
When CUDA_VISIBLE_DEVICES uses UUID or MIG identifiers,
get_parent_visible_gpu_ids() returns [] because the tokens are
non-numeric. The torch fallback in get_visible_gpu_utilization() and
get_backend_visible_gpu_info() previously passed that empty list to
_torch_get_per_device_info(), getting nothing back.
Now both functions detect the empty-list case and fall back to
enumerating torch-visible ordinals (0..device_count-1) with
index_kind="relative". This means the UI and auto-selection still
see real device data in Kubernetes, MIG, and Slurm-style UUID
environments where nvidia-smi output cannot be mapped to physical
indices.
Updated test_uuid_parent_visibility to verify the new torch fallback
path returns available=True with relative ordinals.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add type hint for gpu_ids parameter in InferenceOrchestrator.load_model
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Fixes#4670
Separates the GGUF context slider ceiling from the currently active context length so lowering context via Chat Settings no longer locks the slider max to the reduced value.
- Backend: adds `max_context_length` to GGUF load/status responses, computed from the largest VRAM/KV-fit cap across all usable GPU subsets
- Frontend: stores `ggufMaxContextLength` and uses it for Context Length slider/input bounds; hydrates from both `/api/inference/load` and `/api/inference/status`
- Defaults UI ceiling to native context for CPU-only and fallback paths
- Seeds `effective_ctx` and `max_available_ctx` before GPU probing to prevent `UnboundLocalError` on probe failure
- Property fallback uses native `_context_length`, not effective `context_length`
* feat(studio): add HF/local model selection UI for GGUF export
* fix(studio):fix selector ring clipping
* fix(studio): export page trust_remote_code control and label styling
* fix(studio): accept hf_token in load_checkpoint orchestrator method
The route was passing hf_token to load_checkpoint() but the method
didn't accept it, causing a TypeError on every /api/export/load-checkpoint
request.
* fix(studio): clear HF model selection when input is edited
Previously selectedSourceModel was only cleared when the input became
empty, so editing to a different repo ID after selecting a model would
silently keep the old selection.
---------
Co-authored-by: Roland Tannous <rolandtannous@gravityq.ai>
* fix: default HF cache to standard platform path instead of legacy Unsloth cache
* feat: show LM Studio and local models in chat Fine-tuned tab
* feat: show LM Studio models in Hub models tab
* fix: fetch local models after auth refresh completes
* Revert "fix: fetch local models after auth refresh completes"
This reverts commit cfd61f0ac7.
* fix: increase llama-server health check timeout to 600s for large models
* feat: expandable GGUF variant picker for LM Studio local models
* fix: show GGUF variant label for locally loaded LM Studio models
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: show publisher name in LM Studio model labels
* fix: set model_id for loose GGUF files in LM Studio publisher dirs
* fix: show publisher prefix in Fine-tuned tab LM Studio models
* fix: only use model_id for lmstudio source models
* fix: only show LM Studio models in Hub tab on Mac/chat-only mode
* fix: respect XDG_CACHE_HOME, handle Windows paths in isLocalPath, refresh LM Studio on remount
- _setup_cache_env now reads XDG_CACHE_HOME (falls back to ~/.cache)
instead of hard-coding ~/.cache/huggingface. This follows the standard
HF cache resolution chain and respects distro/container overrides.
- isLocalPath in GgufVariantExpander uses a regex that covers Windows
drive letters (C:\, D:/), UNC paths (\\server\share), relative paths
(./, ../), and tilde (~/) -- not just startsWith("/").
- HubModelPicker.useEffect now calls listLocalModels() before the
alreadyCached early-return gate so LM Studio models are always
refreshed on remount. Also seeds useState from _lmStudioCache for
instant display on re-open.
* fix: add comment explaining isLocalPath regex for Windows/cross-platform paths
* fix: prioritize unsloth publisher in LM Studio model list
* fix: scope unsloth-first sort to LM Studio models on all platforms
* fix: add missing _lmStudioCache module-level declaration
* fix: prioritize unsloth publisher before timestamp sort in LM Studio group
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Detect always-on reasoning models and show Think button as locked-on
Models with hardcoded <think>/<think> tags or reasoning_content in
their chat template (e.g. distilled reasoning models) always produce
thinking output regardless of any toggle. Previously these models
were not detected as reasoning-capable at all, so the Think button
was grayed out even though the model was actively reasoning.
Backend:
- Detect <think>/<think> and reasoning_content in GGUF chat templates
as a fallback when enable_thinking is not present
- Add reasoning_always_on flag to LoadResponse and InferenceStatusResponse
- Pass the flag through all GGUF load and status response paths
Frontend:
- Add reasoningAlwaysOn to the chat runtime store and API types
- When reasoning_always_on is true, show the Think button as lit
(active) but not clickable, with a tooltip explaining the model
always uses thinking
- Force reasoningEnabled=true when the model always reasons
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use pointer-events-none instead of disabled for always-on Think button
The HTML disabled attribute was not fully blocking clicks on the Think
button for always-on reasoning models. Switch to pointer-events-none
CSS class which prevents all mouse interaction at the CSS level.
* Use a static span instead of disabled button for always-on Think
Replace the button element with a plain span when reasoning is
always on. This makes it physically impossible to toggle since
there is no clickable element at all, avoiding any CSS or
disabled-attribute edge cases.
* Simplify always-on Think button to stay lit and remain toggleable
Keep the Think button as a normal toggleable button but ensure it
shows as lit when reasoning_always_on is true. The model always
reasons regardless of the toggle state so there is no need to
block interaction.
---------
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix inference failing for transformers 5.x models (trust_remote_code)
The training worker in core/training/worker.py auto-enables
trust_remote_code for unsloth/* models that need transformers 5.x
(e.g. NVIDIA-Nemotron-3-Nano-4B). The inference worker did not have
the same logic, so loading these models for chat would fail with
"No config file found" while training worked fine.
Add the same auto-detection to the inference worker so
trust_remote_code is set automatically when needed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix: only kill studio-managed llama-server processes, not user's own servers
_kill_orphaned_servers() checked for "unsloth" anywhere in the process
cmdline, which matched the user's own llama-server when serving models
from unsloth/ HF repos (the model path in -m contains "unsloth"). This
caused the user's server to get SIGKILLed on Studio startup, destroying
their prompt cache and forcing full model re-loads.
Narrow the check to only match processes whose binary path lives under
~/.unsloth/llama.cpp/ (the Studio install directory).
* Address review: cover env var paths, move Path.home() inside try block
- Also check LLAMA_SERVER_PATH and UNSLOTH_LLAMA_CPP_PATH so orphans
from custom install locations are still cleaned up.
- Move studio_dirs construction inside the try/except so a Path.home()
failure (containers without HOME) does not crash the constructor.
* Address reviewer feedback: proper path ancestry, /proc/pid/exe, legacy paths
Changes based on 10-reviewer consensus:
- Use Path.is_relative_to() instead of substring matching to prevent
false positives on sibling paths like ~/.unsloth/llama.cpp-backup/.
- Use /proc/<pid>/exe (symlink to real binary) instead of parsing the
first cmdline token, which breaks on paths with spaces. Falls back
to cmdline parsing on non-Linux or when /proc is unavailable.
- Add legacy in-tree install paths (project_root/llama.cpp/ and
project_root/bin/) so orphans from older setup.sh are still cleaned.
- Treat LLAMA_SERVER_PATH as an exact binary match rather than widening
it to its parent directory, which could match unrelated servers in
shared locations like /usr/local/bin/.
- Keep everything inside the try/except so Path.home() failures in
containers do not crash the constructor.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: add Linux platform guard and log cleanup errors
- Guard pgrep fallback with sys.platform check so it does not crash
on Windows/macOS when psutil is unavailable.
- Replace silent except-pass with logger.warning for observability.
* [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>
* Guard against late tool_calls after visible content, filter incomplete fragments
1. If visible content was already emitted (_last_emitted is non-empty)
when delta.tool_calls arrives, ignore the tool_calls instead of
reclassifying the turn as a tool call. llama-server never
interleaves content and tool_calls (they are mutually exclusive),
but this guard is defensive for other OpenAI-compatible backends.
2. Filter out incomplete structured tool_calls fragments before
execution. Entries with empty function.name (from truncation by
max_tokens, disconnect, or interruption) are skipped instead of
being passed to execute_tool().
* [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: account for KV cache in GGUF GPU fit check and auto-cap context length
The GPU fit check only compared GGUF file size against free VRAM,
ignoring KV cache memory. Models with large native context lengths
(e.g. Qwen3.5-9B at 262k) would pass the fit check since the GGUF
is only 5.6 GB, but the KV cache at 262k context needs ~40 GB at
f16. This caused llama-server to silently fall back to CPU inference.
Changes:
- Parse block_count, head_count_kv, head_count, and embedding_length
from GGUF metadata alongside context_length
- Add KV cache VRAM estimation based on architecture params and the
selected cache quantization type (f16, q8_0, q4_0, etc.)
- Auto-reduce context length to the maximum that fits in available
GPU VRAM when the native context would exceed it
- Include estimated KV cache size in the _select_gpus total so the
fit decision reflects actual runtime memory, not just file size
For the reported scenario (Qwen3.5-9B on RTX 3090 with 22415 MiB
free), context is auto-reduced from 262144 to ~63k with f16 KV cache,
keeping the model fully on GPU. With q4_0 KV cache quantization the
context can reach ~226k.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: resolve 6 bugs in KV cache VRAM estimation and add test harness
- Fix q8_0 BPE constant: 1.125 -> 34/32 (1.0625) to match llama.cpp block size
- Fix _fit_context_to_vram returning min_ctx when weights exceed budget
(should return requested_ctx unchanged, let --fit handle it)
- Fix binary search inflating below-2048 requests (lo=min_ctx=2048 > hi)
- Fix n_ctx=0 regressing to 4096 when metadata unavailable (preserve sentinel)
- Fix multi-GPU auto-cap using single-GPU budget instead of aggregate
- Fix _context_length being overwritten with capped effective value
Add tests/test_gguf_kv_vram.py: 43 cross-platform pytest tests covering
pure logic, integration (monkeypatched load_model), and real GGUF parsing.
Runs in an isolated uv venv with only pytest -- no GPU/torch/structlog needed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: complete _effective_context_length lifecycle
- Initialize _effective_context_length in __init__ (prevents AttributeError)
- Reset _effective_context_length in unload_model (prevents stale values)
- Update context_length property to return effective (capped) value for
the UI/API, falling back to native _context_length if not set
* fix: multi-GPU selection tries smallest subset first
The previous approach summed all GPUs' memory to cap context, then
selected GPUs afterward. This was overly optimistic for heterogeneous
setups (e.g., 48 GiB + 4 GiB): the context was inflated by the tiny
GPU's contribution, then both GPUs were dragged in.
Now we try GPU subsets from smallest (1 GPU) to largest, capping
context for each. We pick the smallest subset where the model+KV
fits. This prefers single-GPU when possible (simpler, no tensor
split overhead) and avoids pulling in GPUs that barely help.
Add tests: test_multi_gpu_prefers_fewer_gpus,
test_multi_gpu_heterogeneous.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: prefer fewer GPUs over higher context in GPU selection
Multi-GPU inference is slower due to tensor-split overhead, so we
should prefer fewer GPUs with reduced context over more GPUs with
full context. Now the loop stops at the first GPU subset where the
model fits, rather than continuing to find subsets that allow higher
context. Only if the model can't fit on N GPUs do we try N+1.
This preserves the original behavior: use multi-GPU only when the
model doesn't fit on a single GPU.
* fix: make _kill_orphaned_servers cross-platform via psutil
Replace pgrep + os.kill(SIGKILL) with psutil.process_iter() and
proc.kill(), which work on Linux, macOS, and Windows. Build an
allowlist of install roots matching _find_llama_server_binary so
only studio-managed servers are killed.
* fix: skip KV estimation loop when effective context is unknown
When n_ctx=0 and GGUF metadata lacks context_length, effective_ctx
stays 0. _estimate_kv_cache_bytes(0) returns 0, so a GPU could be
selected with no KV headroom. Guard the loop with effective_ctx > 0
to fall back to file-size-only GPU selection in this case.
* chore: temporarily remove test harness (will add back separately)
* refactor: deduplicate UINT32/UINT64 handling in GGUF parser
Replace duplicated if/elif chains for vtype 4 and 10 with a single
block using setattr. No behavioral change.
* fix: honor explicit n_ctx by using multi-GPU before capping
When the user explicitly sets n_ctx, try to fit the full requested
context using _select_gpus (which adds GPUs as needed). Only cap
context if it doesn't fit on any GPU combination.
When n_ctx=0 (auto/native context), keep the existing behavior:
prefer fewer GPUs with reduced context, since multi-GPU is slower
and the user didn't ask for a specific context length.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: context_length property returns native value for frontend slider
The frontend uses context_length as the slider max. Returning the
capped effective value prevented users from requesting higher context
on reload (e.g., after switching to q4_0 KV cache). Revert to
returning the native GGUF metadata value -- the backend auto-caps
at load time regardless.
* revert: context_length returns effective (capped) value
The UI slider should show what the server is actually running at,
not the theoretical maximum. Revert to returning the effective
context length.
* fix: raise minimum context floor from 2048 to 4096
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix ~1.2s TTFT penalty when tools are enabled in Studio
When users enable web search, Python execution, or terminal tools,
every message gets a ~1.2s delay before any text appears -- even when
the model does not call any tool. This happens because
generate_chat_completion_with_tools() does a non-streaming detection
pass (stream: False) first, waits for the complete response, then
checks for tool calls. For the ~90% of messages that don't trigger a
tool call, this blocking wait is entirely wasted.
Root cause: the detection pass payload uses stream: False, forcing
llama-server to generate the entire response before returning any
tokens.
Fix: replace the non-streaming detection pass with a streaming pass
(stream: True) and a speculative buffer state machine that detects
tool signals in the first 1-2 SSE chunks:
- BUFFERING: accumulate content tokens, check first chars for tool
signal prefixes (<tool_call>, <function=)
- STREAMING: no tool detected, yield tokens to caller immediately
- DRAINING: tool signal found, silently accumulate rest of stream
Three detection paths:
1. Structured delta.tool_calls -- detected instantly, transition to
DRAINING, accumulate fragments, assemble at stream end.
2. XML tool markup in content -- buffer holds up to 32 chars checking
for <tool_call> or <function= prefix, then transitions to DRAINING.
3. No tool signal -- first non-whitespace, non-XML char triggers
immediate transition to STREAMING (fast path, ~90% of requests).
Safety net: after any stream ends in STREAMING state, check accumulated
content for XML tool signals. Handles rare "content before tool call"
edge case.
Additional supporting changes:
- Add headers parameter to _stream_with_retry for auth forwarding
- Share _strip_tool_markup and regex patterns between the detection
pass and the final streaming pass (removes duplication)
- Remove the iteration==0 non-streaming content shortcut (no longer
needed since all iterations stream directly)
- Keep the final streaming pass as fallback for max_tool_iterations
exhaustion
Benchmarked on Qwen3.5-4B Q4_K_XL:
- No tools: TTFT ~112ms (unchanged)
- Tools enabled, no call: TTFT ~112ms (was ~1207ms)
- Decode TPS: 226 (unchanged in all cases)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add unit tests for streaming tool detection state machine
16 tests covering every tool call parsing path:
- Plain text (no tool call) streaming
- Structured delta.tool_calls detection and fragment assembly
- XML <tool_call>JSON</tool_call> detection via buffer
- XML <function=name> tag detection via buffer
- Whitespace before tool XML
- Safety net (content then tool XML)
- Parallel multi-tool calls
- Reasoning token bypass (thinking models)
- Reasoning then tool call
- Empty response handling
- Buffer prefix timeout (HTML not mistaken for tool)
- Non-XML first char instant streaming
- False positive rejection (<tool_tip> vs <tool_call>)
- Arguments split across multiple chunks
- auto_heal_tool_calls=False respects the flag
- Metrics accumulation across tool iterations
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix reasoning-only BUFFERING, pre-tool content emission, and code duplication
Addresses review feedback on the streaming tool detection:
1. Reasoning tokens are no longer yielded during BUFFERING/DRAINING
states. The consumer in routes/inference.py tracks prev_text across
tool iterations without resetting it, so yielding reasoning during
a detection pass that resolves to a tool call would corrupt the
delta computation for subsequent iterations. Reasoning is now
silently accumulated during detection (matching the old non-streaming
behavior) and flushed together with content when the buffer resolves
to STREAMING.
2. Handle reasoning-only responses in the BUFFERING resolver. When a
thinking model emits only reasoning_content with no content tokens,
the stream ends while still in BUFFERING state. The resolver now
detects this case and yields reasoning as plain text (without
<think> wrapper), matching the final streaming pass behavior for
models like Qwen3 in always-think mode.
3. Replace duplicated re.sub calls for stripping tool markup with
the existing _strip_tool_markup(content_text, final=True) helper,
removing ~40 lines of redundant regex code.
4. Update tests: adjust reasoning test expectations to match the new
behavior (reasoning batched with content, not streamed individually
during BUFFERING). Add test_reasoning_only_no_content for the
reasoning-only edge case. 17/17 tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address remaining reviewer findings: late tool_call IDs and XML speculation
1. Late-arriving tool_calls.id: when a provider sends the real ID on a
later delta chunk (after the initial one with index and function
name), the accumulator now updates the ID instead of keeping the
synthetic "call_{idx}" placeholder. (P2, 2/10 reviewers)
2. XML speculation respects auto_heal_tool_calls: when auto_heal is
explicitly disabled, _TOOL_XML_SIGNALS is empty so the BUFFERING
state never speculatively holds content for XML prefix detection.
Content starting with literal "<tool_call>" or "<function=" text
flows straight through without delay. (P2, 1/10 reviewers)
Skipped: finish_reason="tool_calls" without delta.tool_calls fallback
(P1, 1/10 reviewers). llama-server always sends delta.tool_calls
fragments in streaming mode. A non-streaming fallback for this edge
case would add complexity for a scenario that does not occur in
practice with the supported backend.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check request.is_disconnected() every 20 tokens instead of every token
The disconnect check is an async round-trip that adds overhead on every
loop iteration. Since the cancel watcher in llama_cpp.py already
handles connection teardown (closes the streaming response on cancel),
this route-layer check is a secondary safety net that does not need to
run on every single token.
Check every 20 tokens across all 4 streaming paths:
- gguf_tool_stream (tool-enabled GGUF)
- gguf_stream_chunks (standard GGUF)
- audio_input_generate (audio/whisper input)
- generic backend stream (non-GGUF fallback)
* Fix safety net, DRAINING metadata, and test import path
1. Safety net no longer retroactively executes tools after visible
content was already emitted to the user. Once _last_emitted is
non-empty, the stream is committed to normal content mode.
Retroactive tool execution after visible output would violate the
streaming contract and corrupt the route-layer cumulative delta
tracker (prev_text). The tool XML is still stripped by
_strip_tool_markup so the user sees clean content.
2. DRAINING false-positive path now merges accumulated metrics from
prior tool iterations instead of dropping them. Uses the same
merge formula as the STREAMING path.
3. Test import path fixed to use repo root instead of hardcoded
sibling directory. Works in clean checkouts and CI.
4. Renamed test_content_then_tool_xml_safety_net to
test_content_then_tool_xml_no_retroactive_execution to reflect
the corrected behavior.
17/17 tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Redact --api-key value from llama-server startup log
When UNSLOTH_DIRECT_STREAM=1, the generated bearer token was logged
verbatim in the startup command. Replace the secret with <redacted>
before logging.
* Remove test file temporarily
* Revert disconnect throttle, reset prev_text on tool_start, restore XML safety net
Addresses all P1 findings from reviewer round 3 (10 reviewers):
1. Revert disconnect check to every iteration (was every 20th).
All 10 reviewers flagged this as a correctness regression for
short streams and sparse tool event loops. The cancel watcher in
llama_cpp.py is the primary mechanism but the route-layer check
must remain per-iteration for completeness. [10/10]
2. Reset prev_text on tool_start in gguf_tool_stream. When a tool
cycle begins after visible content was already streamed, the
route-layer cumulative delta tracker (prev_text) must be reset
so the post-tool synthesis response is not truncated or dropped.
[9/10]
3. Remove the _last_emitted gate from the XML safety net. The gate
was added to prevent retroactive tool execution after visible
content, but with prev_text now reset on tool_start (#2), the
root cause is fixed and the safety net can correctly handle
content-then-tool-XML responses (matching pre-PR behavior).
[8/10]
* Use None instead of {} for empty auth headers in TTS methods
* Include accumulated metrics in STREAMING metadata check
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* feat(studio): editable context length with Apply/Reset for GGUF model settings
Previously the Context Length field was read-only and the backend
hardcoded `-c 0`, ignoring custom values entirely. KV Cache Dtype also
triggered an immediate model reload with no way to cancel.
Backend:
- llama_cpp.py: pass the actual n_ctx value to `-c` instead of always 0
- models/inference.py: relax max_seq_length to 0..1048576 (0 = model
default) so GGUF models with large context windows are supported
Frontend:
- chat-runtime-store: add customContextLength and loadedKvCacheDtype
state fields for dirty tracking
- chat-settings-sheet: make Context Length an editable number input,
stop KV Cache Dtype from auto-reloading, show Apply/Reset buttons
when either setting has been changed
- use-chat-model-runtime: send customContextLength as max_seq_length
in the load request, reset after successful load
* fix: preserve maxSeqLength for non-GGUF models in load request
customContextLength ?? 0 sent max_seq_length=0 for non-GGUF models,
breaking the finetuning/inference path that needs the slider value.
Now uses a three-way branch:
- customContextLength set: use it (user edited GGUF context)
- GGUF without custom: 0 (model's native context)
- Non-GGUF: maxSeqLength from the sampling slider
* fix: keep max_seq_length default at 4096 for non-GGUF callers
Only relax the bounds (ge=0 for GGUF's "model default" mode,
le=1048576 for large context windows). The default stays at 4096
so API callers that omit max_seq_length still get a sane value
for non-GGUF models.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): rename trust remote code toggle and hide when no model selected
- Rename "Trust remote code" to "Enable custom code"
- Shorten subtitle to "Only enable if sure"
- Hide the toggle when no model is loaded (already hidden for GGUFs)
* fix: restore ge=128 for max_seq_length validation
Keep the minimum at 128 so the API rejects nonsensical values.
GGUF path now sends the model's native context length (from
ggufContextLength) instead of 0 when the user has not customized it.
The upper bound stays at 1048576 for large-context GGUF models.
* feat(studio): replace Context Length input with slider
Use a ParamSlider (512 to model's native context, step 512) instead
of a small number input. Shows "Max" when at the model's native
context length. Consistent with the other slider controls in the
settings panel.
* feat(studio): add editable number input alongside Context Length slider
The slider and number input stay synced -- dragging the slider updates
the number, typing a number moves the slider. The input also accepts
values beyond the slider range for power users who need custom context
lengths larger than the model default.
* fix(studio): widen context length input and use 1024 step for slider
Make the number input wider (100px) so large values like 262144 are
fully visible. Change slider step from 512 to 1024 and min from 512
to 1024.
* fix(studio): context length number input increments by 1024
* fix(studio): cap context length input at model's native max
Adds max attribute and clamps typed/incremented values so the context
length cannot exceed the GGUF model's reported context window.
* fix(studio): point "What's new" link to changelog page
Changed from /blog to /docs/new/changelog.
* fix(studio): preserve custom context length after Apply, remove stale subtitle
- After a reload with a custom context length, keep the user's value
in the UI instead of snapping back to the model's native max.
ggufContextLength always reports the model's native metadata value
regardless of what -c was passed, so we need to preserve
customContextLength when it differs from native.
- Remove "Reload to apply." from KV Cache Dtype subtitle since the
Apply/Reset buttons now handle this.
* feat(studio): auto-enable Search and Code tools when model supports them
Previously toolsEnabled and codeToolsEnabled stayed false after loading
a model even if it reported supports_tools=true. Now both toggles are
automatically enabled when the loaded model supports tool calling,
matching the existing behavior for reasoning.
* fix(studio): auto-enable tools in autoLoadSmallestModel path
The suggestion cards trigger autoLoadSmallestModel which bypasses
selectModel entirely. It was hardcoding toolsEnabled: false and
codeToolsEnabled: false even when the model supports tool calling.
Now both are set from the load response, matching the selectModel
behavior. Also sets kvCacheDtype/loadedKvCacheDtype for dirty
tracking consistency.
* fix(studio): re-read tool flags after auto-loading model
The runtime state was captured once at the start of the chat adapter's
run(), before autoLoadSmallestModel() executes. After auto-load enables
tools in the store, the request was still built with the stale snapshot
that had toolsEnabled=false. Now re-reads the store after auto-load so
the first message includes tools.
* fix(studio): re-read entire runtime state after auto-load, not just tools
The runtime snapshot (including params.checkpoint, model id, and all
tool/reasoning flags) was captured once before auto-load. After
autoLoadSmallestModel sets the checkpoint and enables tools, the
request was still built with stale params (empty checkpoint, tools
disabled). Now re-reads the full store state after auto-load so the
first message has the correct model, tools, and reasoning flags.
* feat(studio): add Hugging Face token field in Preferences
Adds a password input under Configuration > Preferences for users to
enter their HF token. The token is persisted in localStorage and
passed to all model validate/load/download calls, replacing the
previously hardcoded null. This enables downloading gated and private
models.
* fix(studio): use model native context for GGUF auto-load, show friendly errors
The auto-load paths and selectModel for GGUF were sending
max_seq_length=4096 which now actually limits the context window
(since we fixed the backend to respect n_ctx). Changed to send 0
for GGUF, which means "use model's native context size".
Also replaced generic "An internal error occurred" messages with
user-friendly descriptions for known errors like context size
exceeded and lost connections.
LoadRequest validation changed to ge=0 to allow the GGUF "model
default" signal. The frontend slider still enforces min=128 for
non-GGUF models.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): filter out FP8 models from model search results
Hide models matching *-FP8-* or *FP8-Dynamic* from both the
recommended list and HF search results. These models are not
yet supported in the inference UI.
---------
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The prebuilt llama.cpp binary (cuda13-newer) links against
libcudart.so.13 and libcublas.so.13. When torch is installed via pip,
these libraries live in the venv's site-packages under
nvidia/cu13/lib/, not in /usr/local/cuda/.
The existing LD_LIBRARY_PATH logic only searched /usr/local/cuda*
paths (which have CUDA 12.x), so the CUDA backend failed to load
silently and llama-server fell back to CPU -- even with -ngl -1.
This adds a glob scan of the venv's nvidia package directories
(cu*, cudnn, nvjitlink) to LD_LIBRARY_PATH before launching
llama-server, matching where pip puts the CUDA runtime.
Tested on Colab with RTX PRO 6000 Blackwell (CUDA 13.0, pip torch):
before -- 3 MiB GPU, 0% util, CPU inference
after -- 13317 MiB GPU, 77% util, full GPU inference
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
When _select_gpus determines that a GGUF model fits on the selected
GPU(s), the code sets CUDA_VISIBLE_DEVICES but never passes -ngl
(number of GPU layers) to llama-server. Without -ngl or --fit,
llama-server defaults to 0 GPU layers and runs entirely on CPU.
This adds -ngl -1 (offload all layers) in the elif branch where
gpu_indices is set and use_fit is False, so models that fit in VRAM
actually use the GPU for inference.
Co-authored-by: Daniel Han <danielhanchen@users.noreply.github.com>
Daniel Han