- Run GGUF load_model in asyncio.to_thread so the event loop stays free
for progress polling during download (was blocking all requests).
- Extract download phase out of the lock in LlamaCppBackend.load_model
so unload_model/cancel can take effect immediately during download.
- Fix "downloaded" badge for split GGUFs: check total cached bytes
across all shards vs expected size, not just first shard existence.
- Respect CUDA_VISIBLE_DEVICES in /api/system GPU reporting so the
frontend GGUF fit estimation uses actual available VRAM.
- Sort tight variants (need CPU offload) smallest-first instead of
largest-first -- closer to GPU budget = faster inference.
- Fix cancel: use refs instead of React state for abort controller and
toast ID so both cancel buttons (text + toast) work reliably. Make
cancel synchronous (fire-and-forget unload) for instant UI response.
Check abortCtrl.signal.aborted after loadModel returns to prevent
ghost model state. Skip rollback and suppress errors on cancel.
- Dynamic top 4 GGUF models fetched from HF API sorted by downloads,
prepended to the default recommended list.
- Remove turnAnchor="top" for auto-scroll to bottom during generation.
- Set default toast duration to 10s (was infinite for loading toasts).
- Deduplicate cached GGUF repos using scan_cache_dir API (fixes
Qwen/X-GGUF vs qwen/x-gguf duplicates from lowercased HF cache).
- Pre-compile repo_id validation regex to silence CodeQL ReDoS warning.
- Change welcome text and default suggestion text.
_get_gpu_free_memory was filtering by CUDA_VISIBLE_DEVICES, so with
CUDA_VISIBLE_DEVICES='0' set by the training env, llama-server only
saw 1 GPU and used --fit for CPU offloading instead of spreading
across all 8 GPUs.
Since llama-server manages its own GPU allocation (the _select_gpus
method picks GPUs and sets CUDA_VISIBLE_DEVICES for the subprocess),
the query must see ALL physical GPUs to make the right decision.
1. Progress endpoint now takes a variant parameter and only counts
.gguf files matching that variant (not all files in the repo cache,
which would include previously downloaded variants)
2. Tracks .incomplete files in HF blobs dir for in-progress single-shard
downloads, capping at 99% until the file is fully committed
3. Fixed loading text: "Loading model..." for cached, "Downloading
model..." for new downloads, with appropriate descriptions
4. Wording: "Downloading and loading model. Large models can take a
while." instead of "This may include downloading."
1. Loading text: shows "Loading model..." for cached models,
"Downloading model..." for new downloads. Toast description
adapts accordingly.
2. Download progress: polls /api/models/gguf-download-progress every
2s during downloads, updating the toast with percentage and GB
downloaded. Progress is estimated by checking the HF cache folder
size against the expected total bytes.
3. Passes isDownloaded and expectedBytes through the full chain from
variant click to selectModel for accurate UI state.
1. n_gpu_layers kwarg: accept (and ignore) in load_model signature
so callers like llm_assist.py don't get TypeError
2. mmproj exclusion: filter out mmproj files in _find_smallest_fitting_variant
so fallback doesn't pick a tiny vision projection as the "model"
3. Shard preservation after fallback: re-discover shards for the
fallback variant instead of resetting to empty list, so split
GGUFs download all shards
4. Orphan cleanup safety: only kill llama-server processes whose
cmdline contains ".unsloth/", avoiding termination of unrelated
llama-server instances on the same machine
5. Path expression sanitization: validate repo_id format before using
it in cache directory lookups
The variant filename includes a subfolder prefix (e.g.
UD-Q4_K_XL/Kimi-K2.5-UD-Q4_K_XL-00001-of-00013.gguf) but rglob
returns just the filename. Use Path.name for the comparison.
HF cache dirs use the exact case from the repo_id at download time
(e.g. models--unsloth--kimi-k2.5-gguf) which may differ from the
canonical HF repo_id (unsloth/Kimi-K2.5-GGUF). Use case-insensitive
matching to find the cache directory.
- Backend: /gguf-variants now checks HF cache for each variant's file
and returns a downloaded flag per variant
- Frontend: downloaded variants sort before non-downloaded (after
recommended), and show a green "downloaded" badge
- Sort order: recommended -> downloaded+fits -> downloaded+tight ->
fits -> tight -> OOM
1. Interruptible downloads: load_model now checks a cancel event
between shard downloads. unload_model sets the event so cancel
stops the download at the next shard boundary.
2. /api/models/cached-gguf endpoint: scans the HF cache for
already-downloaded GGUF repos with their total size and cache path.
3. "Downloaded" section in Hub model picker: shows cached GGUF repos
at the top (before Recommended) so users can quickly re-load
previously downloaded models without re-downloading.
The unload endpoint checked is_loaded (requires healthy=True), but
during initial loading the server is not yet healthy. Cancel had no
effect because the unload route fell through to the Unsloth backend.
Fix: add is_active property (process exists, loading or loaded) and
check it in the unload route so cancel kills llama-server even during
the download/loading phase.
Also: toast cancel button now properly triggers the backend unload.
When the studio process is killed (SIGTERM/SIGKILL), atexit handlers
may not run in the subprocess orchestrator, leaving llama-server
processes orphaned and holding GPU memory. This caused OOM errors when
trying to load a new model after a studio restart.
On init, LlamaCppBackend now runs pgrep to find and SIGKILL any stale
llama-server processes before starting fresh.
Two fixes for accurate GGUF OOM detection:
1. /api/system now uses nvidia-smi to enumerate all physical GPUs
instead of torch.cuda which only sees CUDA_VISIBLE_DEVICES. This
matches llama-server which can use all GPUs regardless of the env
var. Falls back to torch-based detection if nvidia-smi unavailable.
2. Frontend GGUF OOM check now uses 70% of total GPU memory as the
budget, matching the PR's _select_gpus logic (30% reserved for KV
cache and compute buffers). Previously used checkVramFit's 100%
threshold which was too generous.
Move the sort logic from the backend to the frontend GgufVariantExpander
component where GPU VRAM info is available. The backend now does a simple
size-descending sort. The frontend pins the recommended variant at the
top, pushes OOM variants to the bottom, and sorts the rest by file size
descending (largest/best quality first).
The variants list was returned in HuggingFace file listing order (alphabetical),
making the dropdown confusing (e.g. BF16 before Q4_0). Now sorted as:
1. Recommended variant (from _pick_best_gguf) pinned at top
2. Other UD (Unsloth Dynamic) variants sorted by disk size ascending
3. Non-UD variants sorted by disk size ascending
If the requested port (default 8000) is already in use, auto-
increment and try the next port, up to 20 attempts. Prints a
message like "Port 8000 is in use, using port 8001 instead".
Previously, if port 8000 was busy, uvicorn would fail with
"[Errno 98] address already in use" and the studio would not
start. Now it gracefully finds the next free port.
Uses socket.bind() to check availability before starting uvicorn.
Cross-platform (Linux, macOS, Windows).
Reorder _GGUF_QUANT_PREFERENCE so all UD (Unsloth Dynamic) variants
come before standard quants. UD-Q4_K_XL is the default (best
size/quality tradeoff), followed by other UD quants in decreasing
preference order.
For repos without UD variants (e.g., bartowski), falls through to
standard quants starting with Q4_K_M.
Verified with:
- unsloth/Qwen3.5-35B-A3B-GGUF -> UD-Q4_K_XL
- bartowski/Qwen_Qwen3.5-35B-A3B-GGUF -> Q4_K_M
- unsloth/DeepSeek-V3.2-GGUF -> UD-Q4_K_XL (9 shards)
- unsloth/Llama-3.2-1B-Instruct-GGUF -> UD-Q4_K_XL
The smallest-fitting-variant fallback now groups split GGUF shards
by their variant prefix and sums all shard sizes per variant.
For example, DeepSeek-V3.2 UD-Q4_K_XL has 9 shards totaling
379.8 GB. The previous code treated each shard as a separate
"variant" and would have incorrectly selected a single 50 GB shard
as fitting, ignoring the other 8 shards needed.
Tested with unsloth/DeepSeek-V3.2-GGUF (237 GGUF files, 27
variants from 150 GB to 1.25 TB). Correctly groups and sorts
all variants by total size.
Two changes for GGUF variant selection:
1. Default variant preference now starts with UD-Q4_K_XL (Unsloth
Dynamic quantization) which provides better quality per bit than
standard Q4_K_M. Also added UD-Q2_K_XL, UD-IQ2_M, UD-IQ1_M,
UD-IQ1_S as small fallback options.
2. If the selected variant doesn't fit on disk, automatically fall
back to the smallest GGUF variant in the repo that does fit.
Queries all GGUF file sizes via get_paths_info() and picks the
smallest one under the free disk space limit. If nothing fits,
raises a clear error.
This means users with limited disk space won't get a download
error -- they'll get a smaller quantization instead.
Query file sizes from HuggingFace via get_paths_info() before
downloading, and compare against free disk space on the cache
partition. Raises a clear error if there is not enough space,
instead of failing mid-download.
Uses get_paths_info() instead of repo_info() because xet-stored
repos return size=None from repo_info().siblings, but
get_paths_info() returns the actual file sizes.
If the size check fails for any reason (network error, API change),
it logs a warning and continues with the download anyway.
Set HF_HOME, HF_HUB_CACHE, HF_XET_CACHE, UV_CACHE_DIR, and
VLLM_CACHE_ROOT to a unified location under ~/.unsloth/studio/cache/
on startup. This keeps all model downloads, datasets, and caches
in one place instead of scattered across ~/.cache/huggingface,
~/.cache/uv, etc.
Layout:
~/.unsloth/studio/cache/
huggingface/ (HF_HOME)
hub/ (HF_HUB_CACHE -- model/dataset downloads)
xet/ (HF_XET_CACHE -- xet blob store)
uv/ (UV_CACHE_DIR -- uv package cache)
vllm/ (VLLM_CACHE_ROOT -- vllm compiled kernels)
Only sets variables that are not already in the environment, so
user overrides (e.g. HF_HOME=/data/models) are respected.
Cross-platform: uses Path.home() which resolves correctly on
Linux (~), macOS (~), and Windows (C:\Users\<user>).
If CUDA_VISIBLE_DEVICES is already set in the environment (e.g.,
by the user or a wrapper script), only consider those GPUs when
selecting devices for llama-server. nvidia-smi reports all physical
GPUs regardless of CUDA_VISIBLE_DEVICES, so we filter its output
to match the allowed set.
Without this, the GPU selector could pick a GPU outside the user's
allowed set, overriding their restriction.
Automatically select the best GPU(s) for a GGUF model based on
file size and available VRAM, instead of relying on hardcoded
-ngl -1 or letting llama-server guess.
Logic:
1. Measure total GGUF file size (including split shards)
2. Query free memory per GPU via nvidia-smi
3. If the model fits in 70% of the most-free GPU's memory,
pin to that single GPU (CUDA_VISIBLE_DEVICES=X, no --fit)
4. If it needs multiple GPUs, pick the N most-free GPUs
(CUDA_VISIBLE_DEVICES=X,Y, no --fit)
5. If it's too large for all GPUs combined, omit
CUDA_VISIBLE_DEVICES and use --fit on to let llama-server
handle partial offloading
The 70% threshold accounts for KV cache and compute buffers
that sit on top of the model weights.
Removed the -ngl parameter (was hardcoded to -1). llama-server's
default of "auto" handles layer offloading correctly, especially
with --fit on for oversized models.
Tested on 8x B200:
- 1B model (0.75 GB): picks 1 GPU, no --fit
- 27B model (17 GB): picks 1 GPU, no --fit
- 405B model (230 GB): picks 2 GPUs, no --fit
- 2TB model: all GPUs, --fit on
Refactor command building (deduplicate HF/local paths) and add
flags for better performance:
- --parallel 1: studio is single-user, so only 1 inference slot
is needed. The previous auto-detect picked 4 slots, wasting
VRAM on 3 unused KV caches.
- --flash-attn on: force flash attention for faster inference.
Default is "auto" which may not always enable it.
- --fit on: auto-adjust parameters to fit in available device
memory. Already the default but now explicit.
Also cleaned up the duplicated command building for HF vs local
mode into a single block.
Remove the hard max_tokens=2048 default and le=4096 cap for GGUF
chat completions. When max_tokens is not set (None), the field is
omitted from the llama-server payload entirely, letting the model
generate until it produces an EOS token or hits the context limit.
This is critical for thinking/reasoning models (Qwen3.5, DeepSeek-R1,
etc.) where the thinking phase alone can consume 1000+ tokens before
the actual answer. With the previous 2048 default, simple questions
like "What is 2+2?" used all tokens on thinking and produced empty
visible responses.
Changes:
- llama_cpp.py: max_tokens default None, only include in payload
when explicitly set
- models/inference.py: default None, remove le=4096 cap
- routes/inference.py: pass max_tokens directly, no "or 2048" fallback
llama-server handles omitted max_tokens gracefully (generates until
EOS or context limit). The context size (-c flag, default 4096) acts
as the hard upper bound.
llama-server sends thinking/reasoning tokens as "reasoning_content"
in the SSE delta (separate from "content"). The studio was only
reading delta.content, so all reasoning tokens from models like
Qwen3.5, Qwen3-Thinking, DeepSeek-R1, etc. were silently dropped.
This caused "replies with nothing" for thinking models: the model
would spend its entire token budget on reasoning, produce zero
content tokens, and the user would see an empty response.
Fix: read reasoning_content from the delta and wrap it in
<think>...</think> tags. The frontend already has full support
for these tags (parse-assistant-content.ts splits them into
reasoning parts, reasoning.tsx renders a collapsible "Thinking..."
indicator).
Verified with Qwen3.5-27B-GGUF (UD-Q4_K_XL):
- Before: "What is 2+2?" -> empty response (all tokens in reasoning)
- After: shows collapsible thinking + answer "4"
* fix: resolve compare mode deadlock, cancel_event poisoning, and add dispatcher-based IPC optimization
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* revert to 2048 tokens
* refactor: extract dispatcher timeout values into named constants
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: guard dispatcher shutdown against active compare mailboxes
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* miscallenous studio
* chore: upload dataset misc
* chore: redudancy studio cleanup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: adress the pr comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: adress comments about recipes
* [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: quotation marks
* diceware passphrase generation
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Roland Tannous <rolandtannous@gravityq.ai>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(seed): disable remote code execution for seed inspect loads
* fix(test): use __file__-relative path in seed test
The test used a CWD-relative path (`studio/backend/routes/...`) which
only resolved when pytest was invoked from the repo root. Use
`Path(__file__).resolve()` so the test passes regardless of CWD.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Test <test@test.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix: disable remote code loading for ai-assist model hint lookup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The studio was disabling flex attention entirely on Blackwell+ GPUs
(sm_120 and above) by setting UNSLOTH_ENABLE_FLEX_ATTENTION=0 at
startup. This was a workaround for the flex_attention backward kernel
exceeding shared memory limits on these GPUs.
The root cause is now fixed in unsloth-zoo (PR #542) which patches the
backward kernel config selection to generate safe fallback configs that
fit within the GPU's shared memory limit. With that fix, flex attention
works correctly on Blackwell GPUs and provides a ~1.3x speedup over
the SDPA fallback.
* Update CODEOWNERS for studio and cli
* [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>
