Adds ``--compile_opts coord_descent`` that sets
``torch._inductor.config.coordinate_descent_tuning = True`` alone.
Tested; it regresses 4bit at bs=48 from 3383 → 2147 tok/s so
``coord_descent`` alone is not worth shipping. Keeping the option
in the bisection bench so the regression stays reproducible.
The default ``graph_bs = [1, 2, 4, 8] + range(16, max_bs+1, 16)``
matches the dense FlexInference pattern and is optimal for
power-of-2 batch sizes. For workloads that routinely hit
intermediate sizes (e.g. bs=24, 40, 56), capturing dedicated
buckets improves replay efficiency ~5-10% on those sizes
(they otherwise round up to the next power-of-2 bucket).
Measurements on Qwen3-30B-A3B 4bit, bs=24/40/56 with
compile_walker=True:
| bs | default ladder (round-up) | UNSLOTH_FLEX_GRAPH_BS=8,16,24,32,40,48,56,64 |
|----|------------------------------:|---------------------------------------------:|
| 24 | (rounds to 32, ~1800 tok/s) | 1836 tok/s |
| 40 | (rounds to 48, ~2750 tok/s) | 2772 tok/s |
| 56 | (rounds to 64, ~2700 tok/s) | 3797 tok/s |
The fine ladder costs a few extra capture calls at startup and
slightly larger CUDA graph pool footprint. Power-of-2 bucket
throughput is marginally lower with the fine ladder (memory pool
is shared across more captures), so it's opt-in via env var.
Invalid values fall back to the default ladder with a log.
Optional ``FlexMoEInference(compile_walker=True)`` / env var
``UNSLOTH_FLEX_COMPILE_WALKER=1`` wraps the decode walker
(``call_moe_model_with_flex_kwargs``) with
``torch.compile(fullgraph=False, dynamic=False)`` before the CUDA
graph capture kicks in. Inductor fuses the layernorm + residual +
router pointwise ops, and the compiled kernels end up recorded
inside the captured graph. Net: ~2x decode tok/s on the grouped_mm
path with no change in VRAM, no correctness regression, and no
user-facing API change unless the flag is set.
Numbers on Qwen3-30B-A3B-Instruct-2507, B200, 128 new tokens,
bs sweep, median of 2 timed rounds after 1 warmup:
| precision | bs | baseline (v2) | + compile_walker | speedup |
|-----------|---:|--------------:|-----------------:|--------:|
| 4bit | 16 | 699 | 1347.8 | 1.93x |
| 4bit | 32 | 1243 | 2423.9 | 1.95x |
| 4bit | 48 | 1735 | 3383.2 | 1.95x |
| 4bit | 64 | 1523 | 2981.9 | 1.96x |
| bf16 | 16 | — | 1401.1 | — |
| bf16 | 32 | — | 2864.0 | — |
| bf16 | 48 | — | **3911.1** | — |
Peak throughput: 3911 tok/s at bf16 bs=48 — 39x the pure-HF naive
baseline on the same workload (101.1 tok/s with
``AutoModelForCausalLM`` + eager attn, left-padded, no unsloth).
At 4bit bs=48, 51x the pure-HF naive baseline (66.7 tok/s).
GRPO 5-step validation (Qwen3_MoE_GRPO.py --backend flex
--max_steps 5 on DAPO-Math-17k):
| precision | baseline (v2) | + compile_walker | speedup |
|-----------|--------------:|-----------------:|--------:|
| 4bit | 548.3s | 434.5s | 1.26x |
| bf16 | 451.8s | **407.4s** | 1.10x |
Peak VRAM unchanged (130-133 GB). Loss / KL stable on both, no
NaN, rewards pegged at -7.5 (base-model artifact; orthogonal).
Parity (greedy 32 tokens × 3 prompts at bf16 and 4bit via
``FLEX_MOE_COMPILE_WALKER=1 tests/flex_moe_parity.py``):
flex-captured with the compile wrap matches flex-captured without
the compile wrap on 6/6 prompts with no gibberish, and matches pure
``transformers.AutoModelForCausalLM`` 32/32 on 5 of 6 (prompt ×
precision) pairs (the one divergence is a tie-break logit boundary
on an open-ended continuation — both coherent English).
Bisection of a few torch.compile flag sets against the default at
bs=32 4bit (max_batch_size=32):
| config | tok/s |
|------------------------------------------------------------|-------:|
| default (``torch.compile(fullgraph=False, dynamic=False)``)| 1581.5 |
| + max_autotune + coord_descent + aggressive_fusion | 1704.9 |
| + ``freezing=True`` | 935.7 |
``freezing=True`` is a regression on this path; shipping with the
default config only. The other flags are +7.8% at this size but
at large bs (48+) the max_autotune variant timed out during
compile (>40 min) so the default stays the ship-target for now.
Other attention backends don't help on B200 today:
- pure HF with ``attn_implementation="sdpa"``: cuDNN Frontend error
("No valid execution plans built") on sm_100 + torch 2.11.
- ``flash_attention_2`` 2.8.3: works, but kernels compiled for
sm_80/sm_90 only — slower than eager on B200 (46.7 / 67.9 tok/s
vs eager 66.7 / 101.1 at 4bit / bf16).
- ``flash_attention_3``: ``no kernel image for execution on the
device`` — sm_100 kernels not yet in flash_attn_interface.
- FA4 / ``flash_attention_4``: works standalone but transformers'
integration hard-codes ``flash_attn_with_kvcache = None`` for it,
so it can't service decode. Prefill-only, out of scope here.
New tests:
- ``tests/flex_moe_micro_bench.py``: tight probe that loads the
model once, sweeps batch sizes, prints a sample completion per
bucket (catches gibberish early). Supports ``--compile_mode
{off, walker, walker_fullgraph}`` and ``--compile_opts
{stock, unsloth_O3, inference_freeze}``.
- ``tests/flex_moe_bench.py``: add ``--backend hf_naive`` which
imports pure ``transformers`` (no ``import unsloth``) for the
reference HF baseline, with ``HF_ATTN_IMPL`` env var to switch
between eager / sdpa / flash_attention_{2,3,4}.
- ``tests/flex_moe_parity.py``: add ``FLEX_MOE_COMPILE_WALKER=1``
env var to exercise the compile wrap through the parity harness.
Qwen3-30B-A3B decode on bs=8/64-new-tokens jumps from 55.6 tok/s
(eager) to 374.3 tok/s at 4bit / 421.6 tok/s at bf16 — 4-7x over pure
Hugging Face generate (66.7 / 101.1 tok/s) on B200. Output is bit-exact
against eager decode on all 32 greedy tokens across 3 prompts.
FlexMoEInference.capture_decode_cudagraph now mirrors the dense
FlexInference capture pattern: reserve a scratch page per batch_idx
slot, allocate static input_ids / batch_idx / outputs buffers, capture
one graph per bucket in [1, 2, 4, 8] + range(16, max_bs+1, 16), share
the CUDA memory pool across buckets, erase scratch pages at the end.
_decode_step replays on subsequent calls: zero graph buffers (except
outputs), copy live input_ids / batch_idx into the pinned buffers,
graph.replay(), slice outputs. Matches flex_qwen3_llama.py:725-801
verbatim.
Capture is gated on select_moe_backend() == "grouped_mm". Other
backends (unsloth_triton, native_torch) print a warning and leave
cudagraph_captured = False so generate() silently falls back to eager.
The legacy NotImplementedError rationale — torch.where + Python
for-loop over experts — only applied to native_torch; grouped_mm's
bincount + cumsum + argsort + torch._grouped_mm + index_add_ path is
fully capture-friendly on H100/B200.
FlexEngine.__init__ drops the blanket capture_cudagraph=False force
for qwen3_moe arch since capture correctness is now backend-aware.
tests/flex_moe_bench.py gains a hf_naive backend that imports pure
transformers (no unsloth patches) for a clean "fast inference vs naive
HF" comparison: flex 374.3 vs 66.7 tok/s at 4bit = 5.6x, flex 421.6
vs 101.1 tok/s at bf16 = 4.2x.
tests/flex_moe_parity.py adds greedy-decode parity across three
prompts with flex-captured / flex-eager / HF-pure. Results at bf16:
flex-captured matches flex-eager 32/32 on all three prompts (CUDA
graph is numerically identical to eager), and matches HF pure
transformers 32/32 on 2/3 prompts with the remaining divergence at a
tie-break logit boundary.
Pairs with an unsloth-zoo PR dropping @torch.compiler.disable on the
sparse MoE block and replacing torch.bincount with a capture-safe
scatter_add_ — without those, capture trips on a CPU→CUDA scalar copy
inside bincount(minlength=python_int).
Adds tests/flex_moe_bench.py: 2-round median decode throughput bench
comparing flex (FlexMoEInference) and HF generate on the same
(n_prompts, max_new_tokens, precision) workload. Writes
async_task_outputs/qwen3_moe_grpo_bench/bench_decode_{backend}_{precision}.json.
Also defensively unpacks self.mlp(hidden_states) in
Qwen3MoeDecoderLayer_fast_forward's training branch:
unsloth_zoo.temporary_patches.qwen3_moe.sparse_moe_block_forward
returns a plain tensor for transformers 5.x stacked experts, but the
decoder wrapper unpacked a 2-tuple. The inference branch was already
fixed in the previous commit; the training branch hit the same
ValueError under plain HF generate (no _flag_for_generation).
Bench numbers (Qwen3-30B-A3B, 4bit, rank 16 LoRA, bs=8, 64 new tokens,
B200):
| backend | median tok/s | peak VRAM (GB) | median wall (s) |
|---------|--------------|----------------|------------------|
| HF | 80.5 | 57.2 | 6.36 |
| Flex | 55.6 | 116.0 | 9.21 |
Flex is correctness-complete but not yet performance-competitive on
MoE decode at bs=8. Two structural reasons:
- MoE decode runs eager (forward_moe_backend uses bincount + Python
expert loops which are not CUDA-graph capturable), so flex loses
its main dense-model advantage.
- FlexEngine deep-copies the HF model for the rollout copy, doubling
weight residency. For Qwen3-30B-A3B at bf16 that is ~60 GB extra.
The pristine-base third copy is skipped for Qwen3 MoE (see the
first commit of this series) but the inference deep-copy remains.
Follow-ups (not blockers for correctness):
- torch.compile(dynamic=True) on call_moe_model_with_flex_kwargs to
recover some of the CUDA-graph throughput without requiring graph
capture.
- Evaluate flex's scaling vs HF generate at bs=32 / bs=64, where
paged-KV reuse should dominate per-prompt cost.
- A quantised-only inference copy (4bit forward, fp32 LoRA injection)
so the flex path fits inside 2x 4bit weight residency (~34 GB)
instead of the current post-dequantisation footprint.
Four pre-existing integration blockers hit while wiring Qwen3 30B A3B
MoE + GRPO end-to-end. All four must be fixed for the smoke to run; none
depend on the FlexMoEInference work itself but they surface because
fast_inference=True is the first path where the full MoE + GRPO chain
gets exercised on current transformers 5.x.
1. unsloth/import_fixes.py + unsloth/__init__.py: port fix_trl_vllm_ascend
from #5129 onto this branch so `from trl import GRPOConfig,
GRPOTrainer` works without installing vllm_ascend. transformers 4.48
changed _is_package_available to a tuple, and TRL's module-level
_*_available caches remain truthy on "not installed" hosts, which
then triggers an unconditional `import vllm_ascend` on GRPO import.
2. unsloth/models/qwen3_moe.py FastQwen3MoeModel.pre_patch: do NOT
overwrite Qwen3MoeSparseMoeBlock.forward with the legacy
Qwen3MoeSparseMoeBlock_fast_forward. That fast_forward expects a
flat self.gate_proj attribute which no longer exists on transformers
5.x stacked-expert MoE blocks (self.gate + self.experts).
unsloth_zoo's patch_qwen3_moe installs the correct
sparse_moe_block_forward at TEMPORARY_PATCHES init time; the
override here stomps on that with a broken function and causes
AttributeError('gate_proj') during training.
3. unsloth/models/qwen3_moe.py Qwen3MoeDecoderLayer_fast_forward (inference
path): replace the direct call to Qwen3MoeSparseMoeBlock_fast_forward
with self.mlp(...) so the class-level (unsloth_zoo-patched) forward
runs instead of the broken legacy path. Unpack the (hidden_states,
router_logits) tuple defensively in case a downstream patch returns
a plain tensor.
4. unsloth/models/llama.py LlamaModel_fast_forward_inference_custom:
delegate MoE MLP to decoder_layer.mlp(X) and skip
mlp_fast_forward_inference when the block has no gate_proj /
up_proj / down_proj attribute. Without this, every
model.generate(...) call on a Qwen3 MoE model (e.g. TRL GRPO's
use_vllm=False rollout) crashes in fast_swiglu_inference trying to
read Qwen3MoeSparseMoeBlock.gate_proj.
Smoke-B parity (max_steps=20, 4bit, num_generations=2, seed 3407,
Qwen3-30B-A3B, DAPO-Math-17k, GPUs 2 and 3):
| metric | naive (fast_inference=False) | flex (fast_inference=True) |
|---------------------|------------------------------|----------------------------|
| train_runtime (s) | 2745.97 | 2746.63 |
| peak VRAM (GB) | 66.4 | 136.1 |
| loss step 1 / step 20 | 1192 / 4.65e-06 | 3.41 / 4.01e-06 |
| KL step 1 / step 20 | 1.19e6 / 0.00465 | 3.41e3 / 0.00401 |
Both runs exit cleanly, both converge KL from a huge initial spike
down to ~0.004 by step 20, both see identical reward saturation at
-7.5 (base model needs the SFT format-priming step the reference
Qwen3_(4B)-GRPO.ipynb does before GRPO; our smoke skipped that for
speed). Shapes track each other; absolute magnitudes differ because
the flex deep-copy is allocated from a different stream and the PEFT
adapter is initialised in a different order.
Four integration fixes wired up while bringing Qwen3-30B-A3B-Instruct-2507
green end-to-end on UNSLOTH_FAST_INFERENCE=1:
1. unsloth/models/llama.py patch_peft_model: transformers 5.x reports
model_type as "qwen3_moe" (with underscore); the PR's check was
"qwen3moe" and fell through to NotImplementedError.
2. unsloth/models/llama.py patch_peft_model dense MLP patching: the
fused gate/up/down LoRAMLP swap walks layer.mlp.gate_proj, which is
a Qwen3MoeSparseMoeBlock for MoE and has no gate_proj attribute.
Skip the swap when the MLP does not expose the dense trio; MoE
LoRA is wired through unsloth_zoo/moe_utils anyway.
3. unsloth/inference/flex_qwen3_llama.py flex attention forward:
bnb-4bit Linear compute produces fp32 k / v even under autocast,
which makes the paged KV index_put_ refuse the mixed dtype (bf16
cache, fp32 update). Cast k / v to self._paged_cache.k_cache.dtype
before update. Also benefits the dense path.
4. unsloth/inference/flex_engine.py bind_peft_model: for Qwen3 MoE the
ParamWrapper keeps LoRA un-merged on the stacked expert tensors, so
the training model's expert weights ARE the pristine source. Skip
the pristine-base deep-copy for arch=="qwen3_moe" and point
refresh_moe_lora_merge_from_pristine at the training base directly.
Avoids a third 30-60 GB residency on 30B-A3B.
5. unsloth/inference/flex_moe.py call_moe_model_with_flex_kwargs: lock
activations to the embed dtype across layernorm + MoE MLP; RMSNorm
+ bnb-4bit compute promote activations to fp32 along the MoE path
under autocast. Also force-restore Qwen3MoeSparseMoeBlock.forward
to the stock or unsloth_zoo version if FastQwen3MoeModel.pre_patch
clobbered it with a legacy Qwen3MoeSparseMoeBlock_fast_forward that
expects a flat self.gate_proj (which does not exist on transformers
5.x stacked-expert MoE blocks).
Adds tests/flex_moe_smoke.py: generates 32 tokens twice (cold + warm),
records first-call / warm-call tokens/s, peak VRAM, arch, impl. Writes
async_task_outputs/qwen3_moe_grpo_bench/smoke_A_{4bit,bf16}.json.
Measured on a single B200 (sm_100), Qwen3-30B-A3B-Instruct-2507 +
LoRA rank 16 + grouped_mm MoE backend:
| precision | t_load (s) | peak VRAM (GB) | cold tok/s | warm tok/s |
|-----------|------------|----------------|------------|------------|
| 4bit | 27.0 | 123.4 | 3.7 | 6.9 |
| bf16 | 33.5 | 125.9 | 3.7 | 6.7 |
Both completions coherent ("the lazy dog. ...").
FastLanguageModel.from_pretrained(model_name=\"unsloth/Qwen3-30B-A3B-...\",
fast_inference=True) was silently routing Qwen3MoeForCausalLM to the
dense FlexInference path because _detect_arch matched the \"qwen3\"
substring first. That path works for attention but drops all MoE LoRA
adapters at rollout time: refresh_lora_merge_from_pristine walks
named_modules for LoraLayer instances and calls
base_model.get_submodule(name).weight.data, which does not see the
stacked nn.Parameter tensors on Qwen3MoeExperts.gate_up_proj /
down_proj. Decode capture also breaks on bincount + Python expert
loops inside forward_moe_backend.
Adds unsloth/inference/flex_moe.py:
- call_moe_model_with_flex_kwargs: Qwen3 MoE decoder walker. Identical
to the dense walker for the attention half. Unpacks the mlp(...)
return for both stock HF (plain tensor) and Unsloth's patched
Qwen3MoeSparseMoeBlock_fast_forward (tuple of
(hidden_states, router_logits)).
- FlexMoEInference: API-compatible with FlexInference so the arch
dispatch is a one-line change. cudagraph_captured is permanently
False; capture_decode_cudagraph raises NotImplementedError so a stray
capture_cudagraph=True fails loudly rather than producing silently
wrong output.
- refresh_moe_lora_merge_from_pristine: batched torch.baddbmm LoRA
fuse over stacked 3D expert tensors. Handles both standard
(E, 2*I, H) and transposed (E, H, 2*I) orientations via a runtime
shape check against the flat lora_A / lora_B shapes. In-place write
so flex prefill and paged-KV replay see refreshed values.
Wires the new class into the engine:
- flex_engine._detect_arch: check \"qwen3moe\" / \"qwen3_moe\" BEFORE
the dense \"qwen3\" substring (Qwen3MoeForCausalLM contains both).
- flex_engine.FlexEngine.__init__: route arch==\"qwen3_moe\" to
FlexMoEInference and force self.capture_cudagraph = False so the
MoE expert loops are never captured.
- inference.__init__: export FlexMoEInference.
- models/loader.py: uncomment the qwen3_moe branch so FastQwen3MoeModel
applies training-side patches before FlexEngine wraps the model.
Flex attention / paged KV / block-mask / sampling-param shim / vLLM
shim / sleep-wake are reused verbatim.
Follow-up to 35231d4f (initial flex backend wiring).
Sleep/wake:
- sleep_mode.py adds kv_cache_pool / weight_pool context managers
backed by vLLM's CuMemAllocator. Activated when UNSLOTH_VLLM_STANDBY=1
and vLLM is importable; no-op otherwise so TRL's unconditional
sleep / wake_up calls stay valid.
- FlexEngine routes the inference deep-copy + per-layer PagedKVCache
through the pools. Captured CUDA graphs survive a sleep -> wake
round-trip because cuMem keeps the GPU virtual addresses stable.
- 4-bit single-copy path drops only the KV cache; level 2 warns and
falls back to level 1.
- tests/flex_sleep_mode_smoke.py covers sleep / wake memory deltas,
captured-graph survival, and the no-op path.
Lazy batch sizing:
- FlexEngine's max_batch_size drives fixed-shape page tables, the
input_pos_buffer, the block_mask_logical build, and the CUDA-graph
bucket list at __init__ time. There is no post-init resize, so
picking it at from_pretrained time forces over- or under-shoot.
- build_flex_engine() defers construction until the GRPO rollout shape
is known. install_flex_sentinel() attaches a _LazyFlexEngineSentinel
to model.vllm_engine so hasattr(model, "vllm_engine") keeps working
between from_pretrained and the first build; fast_generate triggers
a floor build on first call.
- rl.py injects _build_flex_from_args(model, args) before both
self.llm = model.vllm_engine rewrite sites (pre-TRL-0.18
sampling_params prefix and >=0.18 colocate LLM replacement). Sizes
the engine from max(pdbs * spg, pdbs * spg * ngen) derived from the
GRPO args. No-op on non-flex models, so the injection is safe for
every TRL backend.
- Precedence: user's max_batch_size kwarg is a floor; the GRPO target
overrides only when strictly larger, with a warning naming both.
- First-build only: post-build growth raises RuntimeError pointing the
user back to max_batch_size= in from_pretrained. The pristine
inference deep-copy is consumed on first build and gemma4 shell
extraction mutates its module tree, so a safe rebuild would require
a second deep-copy.
- tests/flex_lazy_batch_smoke.py (unit, stubbed FlexEngine) covers
default, GRPO bump, user-floor, and post-build-refused cases.
tests/flex_lazy_live_smoke.py exercises sentinel + build against a
live Qwen3-0.6B-Base.
All eight files introduced on this branch now carry the SPDX AGPLv3
header used by the MoE kernels. flex_paged_attention.py keeps its
BSD 3-Clause attribution to attention-gym alongside the new header.
Setting UNSLOTH_FAST_INFERENCE=1 makes
FastLanguageModel.from_pretrained(..., fast_inference=True) return the
flex attention + paged KV + CUDA-graph backend instead of vLLM for
Qwen3, Llama-3, and Gemma-4-E2B-it. Default "0" keeps the existing vLLM
path.
The flex engine lives in a new unsloth.inference subpackage so the
FastLanguageModel wiring and any external caller can reach it without
going through scripts/. The user-facing GRPO notebooks for Qwen3,
Llama-3, and Gemma-4 run end-to-end with no code edits, just the env
var.
New subpackage
- unsloth/inference/flex_paged_attention.py: paged KV cache and
block-mask helpers.
- unsloth/inference/flex_qwen3_llama.py: FlexInference for Qwen3 and
Llama-3.2. tokenize() skips pre-populated input_ids so the engine
can accept vLLM-style token-id prompts.
- unsloth/inference/flex_gemma4.py: FlexGemma4Inference. Text-shell
extraction from Gemma4ForConditionalGeneration, per-layer-type
sliding-window masks, KV sharing for layers 15-34.
- unsloth/inference/flex_engine.py: FlexEngine with the vLLM LLM
surface. Arch dispatch, deep-copy of the HF model for colocate
rollout so training forward stays intact, lazy pristine-base copy
on first bind_peft_model, torch.amp.autocast("cuda", dtype=...)
wrapping for the whole generate path, auto-tune for Triton block
sizes per GPU capability and head_dim band. FA4 default OFF
because the sm_100 FLASH backend crashes on short prompts; users
can opt in with fa4_prefill=True.
- unsloth/inference/vllm_shim.py: LoRARequest, RequestOutput,
CompletionOutput dataclasses plus save_lora / load_lora that
mirror unsloth_zoo.vllm_utils line for line so the TRL GRPO patch
sees the same attribute shape.
Wiring
- unsloth/models/loader.py: UNSLOTH_FAST_INFERENCE=1 bypasses the
vLLM import guard in FastLanguageModel.from_pretrained (L356) and
FastModel.from_pretrained (L982).
- unsloth/models/llama.py: the elif not fast_inference load branch
now also handles the flex case, pops max_batch_size from kwargs
(FlexEngine-only), snapshots the model into an inference copy
BEFORE Unsloth's post-patch so flex-attention patching lands on a
clean HF layout, and after the tokenizer loads constructs
FlexEngine and attaches .vllm_engine / .fast_generate /
.fast_generate_batches. FastLlamaModel.get_peft_model calls
engine.bind_peft_model(peft) right after patch_peft_fast_inference
so state_dict() reads the training LoRA tensors.
- unsloth/models/vision.py: same pattern for FastBaseModel. Gemma-4
is allowed through the vision-model gate when the flex backend is
selected; the engine extracts the text shell from
Gemma4ForConditionalGeneration internally.
- unsloth/models/_utils.py: patch_peft_fast_inference picks
unsloth.inference.vllm_shim.save_lora / load_lora when the engine
is a FlexEngine, so the flex path never imports
vllm.lora.request. fast_inference_setup skips patch_vllm() when
UNSLOTH_FAST_INFERENCE=1.
Smoke tests (B200 sm_100)
- Qwen3-4B-Base bf16 (no LoRA and with get_peft_model LoRA)
- Qwen3-4B-Base fp16
- Llama-3.2-3B-Instruct bf16
- gemma-4-E2B-it bf16
all load through FastLanguageModel + UNSLOTH_FAST_INFERENCE=1 and
generate coherent completions.
Batched steady-state throughput via FastLanguageModel (autocast ON,
n_prompts=8 max_new_tokens=64 max_batch_size=16 max_seq_length=1024):
| Model | Flex CLI | Integration | Delta |
|----------------------------|-----------|-------------|-------|
| Qwen3-4B-Base | 819 tok/s | 595 tok/s | -27% |
| Llama-3.2-3B-Instruct | 1163 | 1357 | +17% |
Qwen3's regression is outside the 10% budget; most likely cause is
q_norm / k_norm being promoted to fp32 under autocast. Tracked for
a follow-up; the integration is functional and GRPO-usable today.
Note
The flex engine source files live here in unsloth/inference/. The
CLI benchmark drivers in scripts/benchmarks/ land in a separate PR
(#5108) which also carries the Gemma-4 and Llama-3 regression runs.
This PR depends only on what is already on main.
Known follow-ups
- FA4 on sm_100 needs a more robust enable rule before it can be
default-on.
- Memory: 2x base-model VRAM without LoRA, 3x with LoRA.
- FlexEngine.sleep() / .wake_up() are no-op stubs; real CPU-offload
parity with UNSLOTH_VLLM_STANDBY is a separate PR.
* Studio: forward standard OpenAI tools / tool_choice on /v1/responses
Mirrors the /v1/chat/completions client-side tool pass-through from #5099
so clients (OpenAI Codex CLI, OpenAI Python SDK, ...) that target the
Responses API receive structured function_call output items instead of
plain text with tool-call tokens leaking into content.
- ResponsesRequest: type tools/tool_choice properly, add parallel_tool_calls;
accept function_call and function_call_output input items for multi-turn
- Translate flat Responses tool / tool_choice shape to the nested Chat
Completions shape before forwarding to llama-server
- _normalise_responses_input: map function_call_output -> role="tool",
function_call -> assistant tool_calls (preserving call_id)
- Non-streaming: map returned tool_calls -> top-level function_call
output items keyed by call_id
- Streaming: emit response.output_item.added (function_call),
response.function_call_arguments.delta/.done, and response.output_item.done
per tool call while keeping the text message at output_index 0
- Pytest coverage: tools/tool_choice translation, multi-turn input mapping,
non-streaming tool_calls mapping, response round-trip
* Studio: merge system messages and close inner stream on /v1/responses
Fixes two issues surfacing when OpenAI Codex CLI drives /v1/responses
against a GGUF with a strict chat template (gpt-oss harmony, Qwen3, ...).
1. "System message must be at the beginning" upstream errors
Codex sends `instructions` AND a `role:"developer"` message in `input`,
producing two separate system-role messages. Strict templates raise
when a second system message exists or when one appears after a user
turn. _normalise_responses_input now hoists all instructions / system /
developer content into a single merged system message at the top of
the Chat Completions message list.
2. "async generator ignored GeneratorExit" / "Attempted to exit cancel
scope in a different task"
_responses_stream consumed the inner chat-completions body_iterator
without an explicit aclose() in a finally block. On client disconnect
(Codex frequently cancels mid-stream), Python 3.13 finalized the inner
async generator on a different task, tripping anyio's cancel-scope
check. Mirrored the same try/finally + aclose pattern used by the
/v1/messages, /v1/chat/completions, and /v1/completions passthroughs.
Tests: hoisting of instructions + developer, developer mid-conversation,
multiple system messages in input, no-system passthrough.
* Studio: accept Codex multi-turn shapes and fix cross-task stream close on /v1/responses
Two issues observed driving /v1/responses from OpenAI Codex CLI against a
GGUF backend.
1. 422 on every turn after the first
Codex replays prior assistant turns with
`content:[{"type":"output_text","text":...,"annotations":[],"logprobs":[]}]`
and carries forward `reasoning` items (o-series / gpt-5) between turns.
Our `ResponsesContentPart` union only accepted input_text / input_image,
and `ResponsesInputItem` only message / function_call / function_call_output,
so Pydantic failed the whole list and FastAPI returned
`"Input should be a valid string"` against the `str` branch of the
outer union.
- Add `ResponsesOutputTextPart` for assistant-replay content.
- Add `ResponsesUnknownContentPart` and `ResponsesUnknownInputItem`
as permissive catch-alls (drop during normalisation).
- Wire an explicit `Discriminator` so dispatch is deterministic and
the fallthrough reaches the catch-all instead of misreporting via
the outer `Union[str, list[...]]`.
- `_normalise_responses_input` now accepts output_text parts, flattens
single-part assistant text to a plain string (keeps legacy chat
templates happy), and silently drops reasoning / unknown items.
2. "async generator ignored GeneratorExit" / cross-task cancel scope
`_responses_stream` awaited `openai_chat_completions` in the parent
route-handler task, which opens the httpx client for the inner
passthrough on *that* task. The outer `StreamingResponse` then iterates
in a child task, so the asyncgen GC finalises the inner httpcore byte
stream on the child task, tripping anyio's "Attempted to exit cancel
scope in a different task". Move the `await` inside `event_generator`
so the httpx lifecycle stays within the single streaming child task,
and surface any HTTPException as a `response.failed` SSE frame.
Tests: assistant output_text replay, reasoning-item tolerance, unknown
content-part tolerance, end-to-end Codex-shape payload (developer + user +
reasoning + function_call + function_call_output + assistant output_text +
user), and single-part assistant flattening to plain string.
* Studio: call llama-server directly from streaming /v1/responses
The previous fix (running the inner await inside event_generator) was not
enough. Wrapping the existing `openai_chat_completions` pass-through still
stacks two async generators: when the outer generator is closed, the
innermost `HTTP11ConnectionByteStream.__aiter__` in httpcore doesn't
receive GeneratorExit before Python's asyncgen GC finalises it in a
sibling task, tripping "Attempted to exit cancel scope in a different
task" and "async generator ignored GeneratorExit" — the same Python 3.13
+ httpcore 1.0.x interaction already seen in PRs #4956, #4981, #5099.
Cure both pass-throughs had: a single same-task httpx lifecycle with
explicit `aiter_lines().aclose()` BEFORE `resp.aclose()` / `client.aclose()`
in the generator's finally block.
Apply it at the Responses layer by dropping the wrapper entirely for GGUF:
open httpx, consume `resp.aiter_lines()`, parse `chat.completion.chunk`,
emit Responses SSE events, close everything in finally — all in the
single StreamingResponse child task. Non-GGUF streaming is rejected with
a 400 (wrapping the transformers backend would re-introduce the
double-layer pattern and isn't a Codex-compatible path today anyway).
Also surfaces upstream httpx.RequestError / non-200 as a
`response.failed` SSE frame rather than a dropped stream now that the
request is dispatched after SSE headers have gone out.
* Studio: silence benign httpcore asyncgen GC warnings on Python 3.13
The streaming pass-throughs (/v1/chat/completions, /v1/messages,
/v1/responses, /v1/completions) all use the proven #4981 / #5099 pattern
— single-task httpx lifecycle with explicit aiter_lines().aclose() ahead
of resp.aclose() / client.aclose() in the generator's finally block.
That handles our own iterators correctly.
The residual noise ("async generator ignored GeneratorExit" /
"Attempted to exit cancel scope in a different task") comes from an
innermost HTTP11ConnectionByteStream.__aiter__ that httpcore creates
internally inside its pool. We hold no reference to it, so we cannot
aclose it ourselves. Python 3.13's asyncgen GC hook finalises it on the
finaliser task, its aclose path enters an anyio CancelScope shield, and
Python flags the cross-task exit. The response has already been
delivered with a 200 by then — it is purely log noise, not a functional
failure. Same interaction seen in modelcontextprotocol/python-sdk #831,
agno #3556, chainlit #2361, langchain-mcp-adapters #254.
Install a targeted sys.unraisablehook that swallows this specific tuple
— RuntimeError mentioning "cancel scope" or "GeneratorExit" plus an
object repr referencing HTTP11ConnectionByteStream — and defers to the
default hook for every other unraisable. Idempotent; guarded by a
sentinel attribute so repeated imports don't stack filters.
* Chatbox, scroll, and menu fixes
- Fixed chatbox auto-expand height for multi-line text on the compare page
- Fixed chatbox UI to be consistent across compare and new chat
- Fixed scrolling being enabled on pages with no content, which also triggered the scroll-to-bottom button
- Fixed scroll-to-bottom button to only appear after scrolling up a reasonable amount instead of instantly
- Added shutdown studio button to the menu for easier access
- Fixed pop-up menu width to match the user button width
(cherry picked from commit cd4e390dfa84fe311fae79a781b96cc0ef5970a9)
* fix: correct compare scroll viewport and clean up chat composer UI polish
* Dark theme refactor and sidebar/chat UI refinements
- Complete refactoring of dark theme
- Replaced square rounded-corner user profile image with a circular bordered one
- Replaced user profile icon with 'U' initial and renamed label from 'Studio' to 'User'
- Chat bubbles now have a pointy top-right edge
- Sidebar menu tab line color selection is now consistent across all menus
- Tab-selection color animation now also applies to recent chats
- Removed 'Compare' menu autoselect when a compare chat conversation is selected
- Fixed UI consistency in Compare to match New Chat
- Removed sidebar animation and tab line, replaced with rounded selection for consistency
- Further adjustments to sidebar UI
- Further adjustments to compare chat UI
* Fixed sidebar collapse/expand for recent chats and recent runs not being clickable
* Chatbox, scroll, and menu fixes
- Fixed chatbox auto-expand height for multi-line text on the compare page
- Fixed chatbox UI to be consistent across compare and new chat
- Fixed scrolling being enabled on pages with no content, which also triggered the scroll-to-bottom button
- Fixed scroll-to-bottom button to only appear after scrolling up a reasonable amount instead of instantly
- Added shutdown studio button to the menu for easier access
- Fixed pop-up menu width to match the user button width
* Sidebar, fonts, and chat UI refinements
- Replaced logo PNG with real font text for 'unsloth' and 'BETA' label
- Added Hellix font and applied it across menus and UI elements
- Lighter scrollbar in the sidebar compared to other areas of the app
- Adjusted chat font and chat bubble styling
- Adjusted app menu design to stay consistent with the sidebar
- Adjusted text style for 'New Chat' and repositioned content/chatbox
- Adjusted model selector and top area UI
- Fixed footer text from 'LLM's' to 'LLMs'
- Fixed active selection border color incorrectly appearing on page refresh and during general navigation
- Logo now defaults to 'New Chat' when clicked
* Sidebar, model selector, and mobile UI fixes
- Further adjustments to sidebar UI and logo
- Changed right bar icon
- Model selector adjustments
- Collapsed sidebar now matches the content area background
- Adjusted Hellix font spacing across pages
- Fixed sidebar icon overlap on mobile screens
* Adjust sidebar icons
* Adjust sidebar icons
* Fixed compare chat UI and scrolling issues
* Fixed inference settings icon behavior and context info positioning
- Fixed top right inference settings icon to move into sidepanel during expand/collapse, matching left sidebar behavior
- Adjusted context information element positioning
* Fix: textarea overflow in system prompt editor
* Code block redesign, font, and chat bubble adjustments
- Redesigned code block colors and theme
- Changed code block font to Fira Code
- Fixed scrollbar disappearing when expanding/collapsing tool calls in chats
- Adjusted chat bubble background color
* Fix chat bubble background color in dark theme
* fix: restore textarea auto-sizing and scope prompt editor sizing
* fix: add explicit textarea field sizing for prompt editor overflow
* fix: generate chat nonce on click instead of render
* fix: respect training lock on logo navigation
* Refactor compare page dual chat scrolling behavior
* Revert "Refactor compare page dual chat scrolling behavior"
This reverts commit d056ec09f2.
---------
Co-authored-by: sneakr <hauzin@hotmail.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
* export: update GGUF quant list and ordering
* gguf: add Q2_K_L quantize flags for output and embeddings
* export: add live console logs for LoRA export flow
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: stream q2_k_l quantize logs and include subprocess error details
* fix: route Q2_K_L preset to q2_k ftype with q8_0 output+embeddings
---------
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>
Trashing a thread mid-stream used to delete the Dexie rows while the
model kept generating, because the sidebar has no access to the
@assistant-ui aui context. Expose per-thread cancelRun() through the
chat runtime store and call it from deleteChatItem so trash behaves
like Stop → Trash. Covers compare pairs by cancelling each paired
thread.
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* fix(studio): forward OpenAI tools/tool_choice to llama-server (#4999)
Studio's /v1/chat/completions silently stripped standard OpenAI `tools`
and `tool_choice` fields, so clients using standard function calling
(opencode, Claude Code, Cursor, Continue, ...) never got structured
tool_calls back. Adds a client-side pass-through path mirroring the
existing Anthropic /v1/messages flow: when `tools` is present without
Studio's `enable_tools` shorthand, the request is forwarded to
llama-server verbatim so the client sees native id, finish_reason
("tool_calls"), delta.tool_calls, and accurate usage tokens.
Also wires Anthropic tool_choice forwarding: /v1/messages previously
accepted tool_choice on the request model but silently dropped it with
a warning. Translate the four Anthropic shapes to OpenAI format and
forward them so agentic clients can actually enforce tool use.
- ChatCompletionRequest: add tools, tool_choice, stop; extra="allow"
- ChatMessage: accept role="tool", optional tool_call_id / tool_calls /
name; content is now optional (assistant with only tool_calls)
- routes/inference.py: _openai_passthrough_stream /
_openai_passthrough_non_streaming helpers, routing branch in
openai_chat_completions, vision+tools via content-parts injection
- _build_passthrough_payload: tool_choice parameter (default "auto")
- anthropic_compat: anthropic_tool_choice_to_openai() translator
- tests/test_openai_tool_passthrough.py: Pydantic + translator unit tests
- tests/test_studio_api.py: 5 new E2E tests (non-stream, stream,
multi-turn, OpenAI SDK, Anthropic tool_choice=any regression)
* fix(studio): surface httpx transport errors from OpenAI passthrough
When the managed llama-server subprocess crashes mid-request, the
async pass-through helpers in routes/inference.py used to return a
bare 500 (non-streaming) or an "An internal error occurred" SSE chunk
(streaming) because _friendly_error only recognized the sync path's
"Lost connection to llama-server" substring -- httpx transport
failures (ConnectError / ReadError / RemoteProtocolError /
ReadTimeout) stringify differently and fell through to the generic
case.
- _friendly_error: map any httpx.RequestError subclass to the same
"Lost connection to the model server" message the sync chat path
emits. Placed before the substring heuristics so the streaming path
automatically picks it up via its existing except Exception catch.
- _openai_passthrough_non_streaming: wrap the httpx.AsyncClient.post
in a try/except httpx.RequestError and re-raise as HTTPException
502 with the friendly detail.
- tests/test_openai_tool_passthrough.py: new TestFriendlyErrorHttpx
class pinning the mapping for ConnectError, ReadError,
RemoteProtocolError, ReadTimeout, and confirming non-httpx paths
(context-size heuristic, generic fallback) are unchanged.
* fix(studio): close aiter_bytes/aiter_lines explicitly in passthroughs
The httpcore asyncgen cleanup fix in 5cedd9a5 is incomplete on Python
3.13 + httpcore 1.0.x: it switched to manual client/response lifecycle
but still used anonymous `async for raw_line in resp.aiter_lines():`
patterns in all three streaming paths. Python's async for does NOT
auto-close the iterator on break/return, so the aiter_lines /
aiter_bytes async generator remains alive, reachable only from the
surrounding coroutine frame. Once `_stream()` returns the frame is
GC'd and the orphaned asyncgen is finalized on a LATER GC pass in a
DIFFERENT asyncio task, where httpcore's
HTTP11ConnectionByteStream.aclose() enters anyio.CancelScope.__exit__
with a mismatched task and prints "Exception ignored in: <async
generator>" / "async generator ignored GeneratorExit" / "Attempted
to exit cancel scope in a different task" to the server log.
User observed this on /v1/messages after successful (status 200)
requests, with the traceback pointing at HTTP11ConnectionByteStream
.__aiter__ / .aclose inside httpcore.
Fix: save resp.aiter_lines() / resp.aiter_bytes() as a variable and
explicitly `await iter.aclose()` in the finally block BEFORE
resp.aclose() / client.aclose(). This closes the asyncgen inside the
current task's event loop, so the internal httpcore byte stream is
cleaned up before Python's asyncgen GC hook has anything orphaned to
finalize. Each aclose is wrapped in try/except Exception so nested
anyio cleanup noise can't bubble out.
Applied to all three streaming passthrough paths:
- _anthropic_passthrough_stream (/v1/messages client-side tool path)
- _openai_passthrough_stream (/v1/chat/completions client-side tool
path, new in this PR)
- openai_completions (/v1/completions bytes proxy from PR #4956)
* fix(studio): default ChatCompletionRequest.stream to false per OpenAI spec
OpenAI's /v1/chat/completions spec defaults `stream` to false, so
clients that omit the field (naive curl, minimal integrations) expect
a single JSON response back. Studio was defaulting to true, silently
switching those clients into SSE and breaking any parser that didn't
also handle streaming. ResponsesRequest and AnthropicMessagesRequest
already default to false correctly; only ChatCompletionRequest was
wrong.
Studio's own frontend always sets `stream` explicitly on every
chat-adapter / chat-api / runtime-provider call site, so the flip has
no UI impact. SDK users (OpenAI Python/JS SDK, opencode, Claude Code,
Cursor, Continue) also always pass `stream` explicitly, so they're
unaffected. The only clients feeling the change are raw-curl users
who were relying on the wrong default -- those get the correct OpenAI
behavior now.
Added a regression test pinning the default so it can't silently
flip back.
* fix(studio): reject images in OpenAI tool passthrough for text-only GGUFs
The new tool passthrough branch runs before _extract_content_parts,
skipping the existing not is_vision guard. Requests combining tools
with an image on a text-only tool-capable GGUF were forwarded to
llama-server, producing opaque upstream errors instead of the
pre-existing clear 400. Restore the guard inline at the dispatch
point, checking both legacy image_base64 and inline image_url parts.
* fix(studio): require tool_call_id on role=tool chat messages
Enforce the OpenAI spec rule that role="tool" messages must carry a
tool_call_id. Without it, upstream backends cannot associate a tool
result with the assistant's prior tool_calls entry and the request
fails in non-obvious ways through the passthrough path. Reject at the
request boundary with a 422 instead.
* fix(studio): harden OpenAI tool passthrough validation and error surfacing
Three related fixes called out by the PR review:
1. Preserve upstream status codes in the streaming passthrough. The
httpx request is now dispatched before the StreamingResponse is
constructed. Non-200 upstream responses and httpx RequestError
transport failures raise HTTPException with the real status
instead of being buried inside a 200 SSE error frame, so OpenAI
SDK clients see APIError/BadRequestError/... as expected.
2. Require non-empty content on user/system/tool messages. Per the
OpenAI spec, content may only be omitted on assistant messages
that carry tool_calls; enforce that at the request boundary so
malformed messages never reach the passthrough path.
3. Role-constrain tool-call metadata. tool_calls is only valid on
role=assistant, tool_call_id and name only on role=tool. Without
this, a user/system message with tool_calls would flip the
passthrough branch on and be forwarded to llama-server, surfacing
as an opaque upstream error.
* fix(studio): normalize image mode and passthrough JSON verbatim
Two Gemini-code-assist review findings on PR #5099:
1. Unconditionally convert decoded images to RGB before PNG encoding.
The prior code only handled RGBA, letting CMYK/I/F images crash
at img.save(format="PNG") and surface as opaque 400s. Applied to
both the passthrough helper and the non-passthrough GGUF path
that originally carried this pattern, keeping the two sites in
sync.
2. Return the upstream JSON body as raw bytes via Response rather
than parse-then-re-serialize with JSONResponse. Matches the
passthrough helper's "verbatim" contract and drops a redundant
round-trip.
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* unsloth gemma4 support files
* some fixes
* Fixing cache.empty() calls (#4813)
* Fixing cache.empty() calls
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Manan Shah <mananshah@Manans-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix/gemma4 mlx (#4816)
* Fixing cache.empty() calls
* fixing for mlx versions
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Manan Shah <mananshah@Manans-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* removed bidirectional check for 31b (#4839)
Co-authored-by: Manan17 <shahmanan170602@gmail.coml>
* Add Gemma 4 26B MoE support (MLX) (#4844)
* removed bidirectional check for 31b
* Change gemma4_text for moe
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Manan Shah <mananshah@Manans-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(gemma4): cast RoPE offset to int before mx.arange() (#4901)
* fix(gemma4): cast RoPE offset to int before mx.arange()
* fix(gemma4): use zero-based arange + offset to avoid CPU-GPU sync
* qwen3.6 patches for multi-turn chat
* qwen3.6 script
* removing unnecessary scripts
* displaying errors for not installed packages
---------
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: Manan Shah <mananshah@Manans-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Manan17 <shahmanan170602@gmail.coml>
Co-authored-by: Théophile Lafargue <138336683+eauchs@users.noreply.github.com>
* Add Qwen3.6 inference defaults for Studio
Add qwen3.6 family entry to inference_defaults.json with the
recommended sampling parameters from Qwen's documentation:
temperature=0.7, top_p=0.8, top_k=20, min_p=0.0,
presence_penalty=1.5, repetition_penalty=1.0.
Without this, Qwen3.6 models fall through to the generic qwen3
pattern which uses different defaults (temperature=0.6,
top_p=0.95, no presence_penalty).
* Add Qwen3.6-35B-A3B-GGUF to default model lists
* Add Qwen3.5/3.6 presence_penalty to thinking toggle and small-model disable logic
- Thinking toggle (on-load + button click) now sets presencePenalty: 1.5 for
Qwen3.5 and Qwen3.6 models (both thinking-ON and thinking-OFF states)
- Small-model thinking-disable check (<9B defaults to no-thinking) extended
from Qwen3.5-only to also cover Qwen3.6, in all 3 locations:
frontend on-load, frontend refresh, backend llama_cpp.py
* fix: multi-GPU inference crash for bnb 4-bit/8-bit models
When load_in_4bit or load_in_8bit is used with device_map="sequential"
and max_memory constraints that place weights across multiple GPUs (or
entirely on a non-default GPU like cuda:1), the bitsandbytes loading
path in transformers never calls dispatch_model. No AlignDevicesHook is
installed, and the first forward/generate call crashes with:
RuntimeError: Expected all tensors to be on the same device
This adds _attach_bnb_multidevice_hooks() which is called after
from_pretrained returns. It infers a device map from actual parameter
placements and calls dispatch_model(force_hooks=True) to install the
missing hooks. The function is a complete no-op for the common
single-GPU cuda:0 case.
Call sites: FastBaseModel.from_pretrained (vision.py) and
FastLlamaModel.from_pretrained (llama.py).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: align with PR #5053 final review improvements
- Add hook call to the bnb quantized loading branch in llama.py (the
primary load_in_4bit path), not just the non-fast-inference fallback
- Expand bnb detection: also check model.is_loaded_in_4bit,
model.is_loaded_in_8bit, model.quantization_method
- Pass explicit main_device and skip_keys to dispatch_model
- Use logger.info instead of print for the success message
- Use kwargs.get("load_in_8bit", False) at llama.py call sites
* [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>
* auth: default to chat
* settings: relaunch onboarding
* onboarding: return to launch page
* studio: stop auto guided tour
* ui: soften global radius
* cleanup: rename onboarding exit prop
* fix onboarding redirect safety
* Show real Unsloth version in settings
* [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): replace navbar navigation with collapsible sidebar
Add an app-wide sidebar with hover-expand and pin-to-dock behavior.
Navigation items (Studio, Recipes, Export, Chat) move from the center
pill navbar to the sidebar. Chat threads and recipes render as
collapsible sub-lists. Navbar simplified to logo + update + close.
- Extend SidebarProvider with pinned/hovered state model
- New AppSidebar with animated active indicator, sloth profile menu,
theme toggle, guided tour, back/forward navigation
- Chat page refactored to URL-driven view state via search params
- Extract reusable hooks for chat thread and recipe sidebar data
- Guard startViewTransition for browser compatibility
- Wrap chat deletions in Dexie transaction for data integrity
* feat(studio): move logo to sidebar and make navbar overlay
- Sidebar is now full-height with logo in SidebarHeader
- Collapsed sidebar shows sticker.png, expanded shows full logo
- Navbar is absolute-positioned overlay (no layout space)
- Main content extends to top, aligning with navbar controls
* feat(studio): full-height sidebar with recents, edge-to-edge nav buttons
- Sidebar outside max-w-7xl, pinned to left edge
- Remove sidebar rounding, menu buttons rounded-md
- Nav buttons flush to sidebar edges with no left rounding
- Replace collapsible recipes/chat with flat nav items
- Add Recents section with chat history (1 item when not on chat, full on chat)
- New Chat as first nav item with PencilEdit02Icon
- Cursor pointer on all sidebar buttons
- Navbar temporarily hidden for screenshots
* fix(studio): fix chat scroll, action bar hover, collapsible recents
- Fix sticky composer by removing `relative` override on viewport footer
- Action bar buttons only show on hover (autohide=always)
- Remove floating border/shadow from action bar
- Add scroll space above composer for last message actions
- Back/forward buttons use router history (stay in-app)
- Recents section collapsible with chevron on chat route
- Set html/body/#root height for proper h-full chain
* fix(studio): address review feedback, clean up unused code
- Unhide navbar (was left hidden from screenshot)
- Remove unused imports: SidebarMenuSub*, BubbleChatIcon, ColumnInsertIcon
- Remove unused vars: recipeItems, activeRecipeId, canCompare, recipesOpen
- Include compare query id in active sidebar selection
- Use store type for contextUsage instead of inline type
- Simplify noop in sidebar.tsx
- Remove empty className prop
* feat(studio): add mobile sidebar, recent runs section, and misc UX fixes
* feat(studio): scaffold settings feature module with dialog store
* feat(studio): add tri-state theme store for settings
* feat(chat): add clear-all-chats and export-chat-history utils
* feat(studio): add settings dialog shell with tab rail
* feat(studio): add appearance tab with theme and sidebar pin
* feat(studio): add settings general tab with hf token, auto-title, reset prefs
* feat(studio): add settings chat tab with export and clear
* feat(studio): add api keys tab with list and revoke flow
* feat(studio): add create-key form and reveal dialog
* feat(studio): add usage examples panel to api keys tab
* feat(studio): add settings about tab with update and shutdown
* feat(studio): add settings dropdown item and cmd-comma shortcut
* feat(studio): remove legacy api-keys route and chat-sheet preference rows
* fix(studio): settings dialog a11y + polish pass
* feat(studio): inline api key reveal card replacing nested dialog
* fix(studio): hide revoked keys from settings list
* refactor(studio): strip navbar and hoist training unload guard
* feat(studio): explicit sidebar toggle, remove hover-open and pin icons
* fix(studio): use SidebarRight01Icon for collapsed sidebar open toggle
* fix(studio): address code review findings for settings dialog
* feat(studio): collapsible navigate group with standalone new-chat and compare
* fix(studio): chat-only standalone actions, use ColumnInsertIcon for compare
* fix(studio): sidebar new-chat/compare state reset and icon-mode collapsible
* feat(studio): add compact logo assets for sidebar header
* Fixed sidebar design
* fix(studio): sidebar delete icon hover contrast and sizing
* feat(studio): route-gate sidebar recents (chats off /studio, runs on /studio)
* feat(studio): add chat search store
* feat(studio): add chat search index hook with snapshot-on-open
* feat(studio): add chat search command dialog with global shortcut
* feat(studio): wire chat search into sidebar
* fix(studio): trim hf token on save, add show/hide toggle, commit on close
* revert(studio): restore original sidebar/border colors, brighten sidebar
* feat(studio): forward overlayClassName through CommandDialog
* fix(studio): wrap search dialog in Command context, redesign as flat 635px card
* fix(studio): reserve right padding on recent items so delete icon stops overlapping title
* fix(studio): skip hf token unmount-commit during reset-prefs reload
* chore(studio): drop unused icon import and unreachable runs navigate branch
* fix(studio): chat search index filters archived before limit, batches message query, picks up reasoning text
* fix(studio): keep CommandEmpty in tree so empty state renders correctly
* fix(studio): cap system prompt and chat template textareas so they scroll instead of growing
* fix(studio): attach chat-compare tour anchor to sidebar compare button
* fix(studio): persist system theme explicitly so next-themes does not clobber on reload
* fix(studio): auto-switch to history tab when selecting a recent run from sidebar
* UI overhaul: chatbox, scrollbar, sidebar, and compare view
UI Changes:
- Redesigned the Compare UI with general cleanup
- Redesigned the Chatbox UI
- Reduced the width of the user chat bubble for improved readability
- Narrowed the user chat box across the content page
- Adjusted thinking-box text color to be slightly darker
- Removed faded text effect from chat messages
- Removed faded text effect from the thinking box
- Added a small LLM chat safety note at the bottom of the chatbox
- Restyled the scrollbar
Layout & Behavior:
- Reworked the scrollbar to span the full height of the page (no top/bottom padding) and remain persistently visible when content is scrollable, rather than only on hover
- Reworked the Configuration sidebar to span full height — removed rounded corners and borders, with the scrollbar adjusted to match the full top-to-bottom layout
- Adjusted the top menu and bottom chatbox content areas to work correctly with the new full-page scroll behavior
- Made chat content match the chatbox width, with content sliding slightly behind the chatbox when scrolling
- Aligned chat text width with the chatbox for visual consistency, including how far the text extends behind the chatbox
Fixes:
- Fixed the chatbox not auto-expanding when typing multi-line input while bottom-positioned during an active chat (previously only worked before a chat had started)
- Fixed positioning and design of the user chat hover menu buttons to match the assistant chat box — now displayed below the chat bubble instead of on the left side
* Fix user message layout in thread component
* swap code icon
* fix compare layout
* fix compare pane flex
* Sidebar improvements and fixes
- Added scrolling support to the sidebar so menus and recent chats no longer get hidden
- Recent chats are now always visible in the sidebar, not hidden when in Studio, Recipes, or Export
- Recent chat is now deselected when selecting other navigations
- Fixed sidebar glitch where browser resize could make the sidebar and expand button disappear completely
- Fixed glitch where the open-sidebar hover tooltip appeared above the logo when clicking expand sidebar
- Reduced sidebar width on mobile to around 2/3 of the screen (was too wide)
- Made the close-sidebar hover tooltip consistent with the rest of the design
- Removed sidebar collapse/expand animation
- Small adjustment to chat width
* Fix route scrolling, polling, and theme sync issues
* Fix Studio page scrolling
---------
Co-authored-by: sneakr <hauzin@hotmail.com>
* Studio: Ollama support, recommended folders, Custom Folders UX polish
Backend:
- Add _scan_ollama_dir that reads manifests/registry.ollama.ai/library/*
and creates .gguf symlinks under <ollama_dir>/.studio_links/ pointing
at the content-addressable blobs, so detect_gguf_model and llama-server
-m work unchanged for Ollama models
- Filter entries under .studio_links from the generic models/hf/lmstudio
scanners to avoid duplicate rows and leaked internal paths in the UI
- New GET /api/models/recommended-folders endpoint returning LM Studio
and Ollama model directories that currently exist on the machine
(OLLAMA_MODELS env var + standard paths, ~/.lmstudio/models, legacy
LM Studio cache), used by the Custom Folders quick-add chips
- detect_gguf_model now uses os.path.abspath instead of Path.resolve so
the readable symlink name is preserved as display_name (e.g.
qwen2.5-0.5b-Q4_K_M.gguf instead of sha256-abc...)
- llama-server failure with a path under .studio_links or .cache/ollama
surfaces a friendlier message ("Some Ollama models do not work with
llama.cpp. Try a different model, or use this model directly through
Ollama instead.") instead of the generic validation error
Frontend:
- ListLabel supports an optional leading icon and collapse toggle; used
for Downloaded (download icon), Custom Folders (folder icon), and
Recommended (star icon)
- Custom Folders header gets folder icon on the left, and +, search,
and chevron buttons on the right; chevron uses ml-auto so it aligns
with the Downloaded and Recommended chevrons
- New recommended folder chips render below the registered scan folders
when there are unregistered well-known paths; one click adds them as
a scan folder
- Custom folder rows that are direct .gguf files (Ollama symlinks) load
immediately via onSelect instead of opening the GGUF variant expander
(which is for repos containing multiple quants, not single files)
- When loading a direct .gguf file path, send max_seq_length = 0 so the
backend uses the model's native context instead of the 4096 chat
default (qwen2.5:0.5b now loads at 32768 instead of 4096)
- New listRecommendedFolders() helper on the chat API
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: log silent exceptions and support read-only Ollama dirs
Replace silent except blocks in _scan_ollama_dir and the
recommended-folders endpoint with narrower exception types plus debug
or warning logs, so failures are diagnosable without hiding signal.
Add _ollama_links_dir helper that falls back to a per-ollama-dir hashed
namespace under Studio's own cache (~/.unsloth/studio/cache/ollama_links)
when the Ollama models directory is read-only. Common for system installs
at /usr/share/ollama/.ollama/models and /var/lib/ollama/.ollama/models
where the Studio process has read but not write access. Previously the
scanner returned an empty list in that case and Ollama models would
silently not appear.
The fallback preserves the .gguf suffix on symlink names so
detect_gguf_model keeps recognising them. The prior "raw sha256 blob
path" fallback would have missed the suffix check and failed to load.
* Address review: detect mmproj next to symlink target for vision GGUFs
Codex P1 on model_config.py:1012: when detect_gguf_model returns the
symlink path (to preserve readable display names), detect_mmproj_file
searched the symlink's parent directory instead of the target's. For
vision GGUFs surfaced via Ollama's .studio_links/ -- where the weight
file is symlinked but any mmproj sidecar lives next to the real blob
-- mmproj was no longer detected, so the model was misclassified as
text-only and llama-server would start without --mmproj.
detect_mmproj_file now adds the resolved target's parent to the scan
order when path is a symlink. Direct (non-symlink) .gguf paths are
unchanged, so LM Studio and HF cache layouts keep working exactly as
before. Verified with a fake layout reproducing the bug plus a
regression check on a non-symlink LM Studio model.
* Address review: support all Ollama namespaces and vision projector layers
- Iterate over all directories under registry.ollama.ai/ instead of
hardcoding the "library" namespace. Custom namespaces like
"mradermacher/llama3" now get scanned and include the namespace
prefix in display names, model IDs, and symlink names to avoid
collisions.
- Create companion -mmproj.gguf symlinks for Ollama vision models
that have an "application/vnd.ollama.image.projector" layer, so
detect_mmproj_file can find the projector alongside the model.
- Extract symlink creation into _make_symlink helper to reduce
duplication between model and projector paths.
* Address review: move imports to top level and add scan limit
- Move hashlib and json imports to the top of the file (PEP 8).
- Remove inline `import json as _json` and `import hashlib` from
function bodies, use the top-level imports directly.
- Add `limit` parameter to `_scan_ollama_dir()` with early exit
when the threshold is reached.
- Pass `_MAX_MODELS_PER_FOLDER` into the scanner so it stops
traversing once enough models are found.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: Windows fallback, all registry hosts, collision safety
_make_link (formerly _make_symlink):
- Falls back to os.link() hardlink when symlink_to() fails (Windows
without Developer Mode), then to shutil.copy2 as last resort
- Uses atomic os.replace via tmp file to avoid race window where the
.gguf path is missing during rescan
Scanner now handles all Ollama registry layouts:
- Uses rglob over manifests/ instead of hardcoding registry.ollama.ai
- Discovers hf.co/org/repo:tag and any other host, not just library/
- Filenames include a stable sha1 hash of the manifest path to prevent
collisions between models that normalize to the same stem
Per-model subdirectories under .studio_links/:
- Each model's links live in their own hash-keyed subdirectory
- detect_mmproj_file only sees the projector for that specific model,
not siblings from other Ollama models
Friendly Ollama error detection:
- Now also matches ollama_links/ (the read-only fallback cache path)
and model_identifier starting with "ollama/"
Recommended folders:
- Added os.access(R_OK | X_OK) check so unreadable system directories
like /var/lib/ollama/.ollama/models are not advertised as chips
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: filter ollama_links from generic scanners
The generic scanners (models_dir, hf_cache, lmstudio) already filter
out .studio_links to avoid duplicate Ollama entries, but missed the
ollama_links fallback cache directory used for read-only Ollama
installs. Add it to the filter.
* Address review: idempotent link creation and path-component filter
_make_link:
- Skip recreation when a valid link/copy already exists (samefile or
matching size check). Prevents blocking the model-list API with
multi-GB copies on repeated scans.
- Use uuid4 instead of os.getpid() for tmp file names to avoid race
conditions from concurrent scans.
- Log cleanup errors instead of silently swallowing them.
Path filter:
- Use os.sep-bounded checks instead of bare substring match to avoid
false positives on paths like "my.studio_links.backup/model.gguf".
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: drop copy fallback, targeted glob, robust path filter
_make_link:
- Drop shutil.copy2 fallback -- copying multi-GB GGUFs inside a sync
API request would block the backend. Log a warning and skip the
model when both symlink and hardlink fail.
Scanner:
- Replace rglob("*") with targeted glob patterns (*/*/* and */*/*/*)
to avoid traversing unrelated subdirectories in large custom folders.
Path filter:
- Use Path.parts membership check instead of os.sep substring matching
for robustness across platforms.
Scan limit:
- Skip _scan_ollama_dir when _generic already fills the per-folder cap.
* Address review: sha256, top-level uuid import, Path.absolute()
- Switch hashlib.sha1 to hashlib.sha256 for path hashing consistency.
- Move uuid import to the top of the file instead of inside _make_link.
- Replace os.path.abspath with Path.absolute() in detect_gguf_model
to match the pathlib style used throughout the codebase.
* Address review: fix stale comments (sha1, rglob, copy fallback)
Update three docstrings/comments that still referenced the old
implementation after recent changes:
- sha1 comment now says "not a security boundary" (no hash name)
- "rglob" -> "targeted glob patterns"
- "file copies as a last resort" -> removed (copy fallback was dropped)
* Address review: fix stale links, support all manifest depths, scope error
_make_link:
- Drop size-based idempotency shortcut that kept stale links after
ollama pull updates a tag to a same-sized blob. Only samefile()
is used now -- if the link doesn't point at the exact same inode,
it gets replaced.
Scanner:
- Revert targeted glob back to rglob so deeper OCI-style repo names
(5+ path segments) are not silently skipped.
Ollama error:
- Only show "Some Ollama models do not work with llama.cpp" when the
server output contains GGUF compatibility hints (key not found,
unknown architecture, failed to load). Unrelated failures like
OOM or missing binaries now show the generic error instead of
being misdiagnosed.
---------
Co-authored-by: Daniel Han <info@unsloth.ai>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* Fix review findings for PR #49
1. Sandbox fallback Jinja env in _VariantTokenizerProxy.apply_chat_template
(use SandboxedEnvironment, matching _derive_assistant_prefix_by_render)
2. Unwrap benign outer-If guards in _template_ends_with_toplevel_for so
templates like {% if messages %}{% for ... %}{% endfor %}{% endif %}
are still repairable (preserves Qwen3-Guard rejection via else-branch
and add_generation_prompt-name checks)
3. Preserve raw name_or_path in _VariantTokenizerProxy._source_path so
local-path detection works for dict/list variant tokenizers
4. Context-aware strict-mode messages: omit "will still load" and
"Set UNSLOTH_STRICT_CHAT_TEMPLATE=1" when already raising
* [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>
Older installers persisted the venv Scripts directory directly in the
User PATH registry. The shim approach from #4961 no longer writes that
entry, but on upgrade the old one survived and python.exe / pip.exe
from the unsloth venv continued winning resolution in every new shell.
Before creating the shim, read the current User PATH, filter out any
entry matching $VenvDir\Scripts (using the same symmetric raw+expanded
comparison as Add-ToUserPath), and write back if changed. No-op on
fresh installs where the legacy entry was never written.
Confirmed on a real Windows machine: `where.exe python` was returning
the venv interpreter first even after the shim PR merged.
Older installers persisted the venv Scripts directory directly in the
User PATH registry. The shim approach (added in this PR) no longer writes
that entry, but it also did not remove the old one. On upgrade, the
legacy entry survived and python.exe / pip.exe from the unsloth venv
continued winning resolution in every new shell, which is exactly the
hijack the shim was designed to prevent.
Before creating the shim, read the current User PATH, filter out any
entry matching $VenvDir\Scripts (using the same symmetric raw+expanded
comparison as Add-ToUserPath), and write back if changed. This runs
once per install and is a no-op on fresh installs where the legacy
entry was never written.
* Restrict flash attn to <=256 head dim. Consolidate attn impl checks
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Consolidate the changes into single function
* safeguard for dict instead of object
* [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>
* Chat-template repair: warn-by-default, AST classification, dict support
Follow-up hardening on top of PR #4426 (which fixed the #4150
RuntimeError for ChatML LoRA reloads).
Behavior changes:
- Warn-by-default instead of RuntimeError. When fix_chat_template cannot
repair a broken template, emit a warning and return the original.
Set UNSLOTH_STRICT_CHAT_TEMPLATE=1 to restore the pre-warn hard fail.
Fixes the UX where a missing `{% if add_generation_prompt %}` block on
a saved LoRA (typical after LlamaFactory / Axolotl re-serialize) would
block model loading entirely.
- Local path vs HF hub distinguished in the warning message. For local
paths the message points at the likely downstream tool; for HF IDs it
points at the upstream model maintainers. Previously both said "file a
bug report to the maintainers of <path>" even when <path> was the
user's own saves/ directory.
- Dict / list chat_template now handled. Hermes-3 ships with
{default, tool_use} and the previous code crashed with
AttributeError: 'dict' object has no attribute 'find' when entering
_fix_chat_template with a dict. Each variant is now fixed
independently; structure is preserved.
Internals:
- _find_end_position now matches all four Jinja whitespace-control
variants ({% %}, {%- %}, {% -%}, {%- -%}) and returns the rightmost
endfor/endif so multi-for templates aren't locked onto the first loop.
Previously {%- endfor -%} (both-side dash, used by Qwen3-Guard) was
silently bypassed.
- _has_add_generation_prompt_block uses Jinja AST via
jinja2.nodes.If/Name walks instead of substring matching, so
templates that hide the block behind comments or dash-style variants
are classified correctly.
- _template_ends_with_toplevel_for gates the GH#4150 ChatML repair on
the AST: only fires when the last structural top-level node is a For
(standard ChatML shape), ignoring trailing pure-whitespace output
nodes. Templates wrapped in an outer If (Qwen3-Guard) are now
explicitly skipped at the _fix_chat_template level as well, not just
at load_correct_tokenizer's name-based exemption.
- _validate_patched_template renders the patched template with and
without add_generation_prompt and confirms the patched output
responds to the flag by appending (not replacing) content. If
validation fails, the patch is discarded and we fall through to the
warn path.
Verified with an expanded regression suite in tests/:
- test_fix_chat_template_pr4426.py: 42/42 template-matrix cells
- test_load_correct_tokenizer_pr4426.py: 5/5 tokenizer loads
- test_chat_template_followups.py: 10/10 new follow-up tests
- test_mistral_pr4426.py: 5 Mistral variants byte-identical
- test_qwen_pr4426.py: 14 Qwen variants byte-identical
(Qwen1.5, Qwen2, Qwen2.5-Instruct/Coder/Math/VL, Qwen3,
Qwen3-Coder, QwQ, Qwen3-Guard-Gen)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard _validate_patched_template against read-only chat_template
If tokenizer.chat_template is a property or otherwise read-only, the
validation helper would crash with AttributeError when trying to
temporarily set the patched template. Catch the assignment failure and
return False (skip validation), and best-effort restore in the finally
block.
* Replace regex separator inference with render-diff; broaden repair to non-ChatML templates
The previous `_infer_assistant_separator` was a four-tier regex heuristic that
only worked on ChatML-shaped templates and forced a hard `<|im_start|>` /
`<|im_end|>` presence gate on Case 2 repair. This meant a Llama-3, Gemma, or
Phi-3 template stripped of its generation-prompt block by a downstream tool
(LlamaFactory, Axolotl, etc.) would still warn-and-return even though the
structural shape is identical to the ChatML case the PR already handles.
This replaces the regex with `_derive_assistant_prefix_by_render`: render the
template with two dialogs that differ only in assistant content, then
`os.path.commonprefix` on the tails captures the exact assistant-turn prefix
the template emits. The template itself is ground truth, so non-ChatML shapes
work as long as the assistant block is a literal the template emits once per
message.
Three guards keep the derivation safe:
A. both assistant renders extend the base render (no reordering);
B. the divergence point is exactly the content-insertion site (sentinel
follows the common prefix);
C. a user-role cross-check: if a render with a user sentinel also emits
the same prefix, role has no effect on output and we reject. A render
failure on [user, user] (e.g. Gemma's `raise_exception` alternation
check) is evidence that role matters; we accept.
Sentinels differ at character 0 so `commonprefix` cannot absorb them, and
trailing whitespace/comments after the last `{% endfor %}` are stripped
before probing (they would appear in base but not after the appended
assistant turn and break Guard A).
`_fix_chat_template` and `_repair_string_template` now thread an
`is_sharegpt` kwarg; `_fix_chat_template` retries once with
`is_sharegpt=True` if the first probe returns None (dual-probe fallback
for dict/list callers).
The ChatML `<|im_start|>` / `<|im_end|>` hard gate in Case 2 is dropped.
`_infer_assistant_separator` is deleted.
Verified via:
- tests/test_fix_chat_template_pr4426.py: 51/51 cells (new Llama-3,
Gemma, Phi-3 broken-template rows all repair FIX-OK)
- tests/test_load_correct_tokenizer_pr4426.py: 5/5
- tests/test_chat_template_followups.py: 18/18 (T11-T18 cover
non-ChatML repair + probe failure modes)
- tests/test_mistral_pr4426.py: 5/5 byte-identical
- tests/test_qwen_pr4426.py: 14/14 byte-identical (Qwen3-Guard AST
gate still rejects)
- tests/hermes3_lora_pr4426.py reload: patched template ends with
`<|im_start|>assistant\n`, inference returns sensible output.
- temp/sim/battery.py: 79/79 followup; vs baseline: 0 regressions,
9 improvements.
- Spot-check probe on real stripped tokenizers (Hermes-3, Phi-4,
Llama-3.2-1B, Gemma-3-1B): all derive the expected prefix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address reviewer findings: variant routing, positive-gate detection, comment-safe end scan
Resolves three reviewer findings on PR #5049 (`fix/chat-template-followups`):
Finding #1 [10/10]: dict/list variants now route through
`_fix_chat_template_for_tokenizer` via a new `_VariantTokenizerProxy`
adapter. Previously the dict/list branches called `_fix_chat_template`
directly, silently bypassing the warn/strict (`UNSLOTH_STRICT_CHAT_TEMPLATE`)
contract, the `no == yes` diagnostic, broken-existing-block detection,
and `_validate_patched_template` guard. The proxy swaps
`base.chat_template` to the variant string before each
`apply_chat_template` call so tokenizer globals (`bos_token`, custom
filters, `raise_exception`) remain available; if the base is read-only
it falls back to isolated Jinja rendering.
Finding #2 [1/10]: `_has_add_generation_prompt_block` now requires the
`If` body to contain at least one `Output` node (a new
`_if_body_emits_content` helper walks descendants). This distinguishes a
real generation-prompt block from a header guard like
`{% if not add_generation_prompt is defined %}{% set ... %}{% endif %}`
(body contains only `Assign`) which references the name but emits
nothing. Also dropped a now-redundant `"add_generation_prompt" not in
scrubbed` guard in `_fix_chat_template` Case 2 so header-guarded
templates still get repaired.
Finding #4 [1/10]: `_find_end_position` now replaces Jinja comments with
equal-length whitespace before scanning for `{% endfor %}` / `{% endif %}`
tokens. This prevents a trailing comment containing those tokens from
being picked as the real end tag. Positions in the padded string map 1:1
to positions in the original template.
Tests:
- tests/test_chat_template_followups.py: 21/21 (T19 strict-mode
dict variant, T20 header-guard repair, T21 comment-endfor trap
added; T4/T5 stubs updated with a working apply_chat_template
that routes through Jinja).
- tests/test_fix_chat_template_pr4426.py: 51/51 cells unchanged.
- tests/test_load_correct_tokenizer_pr4426.py: 5/5.
- tests/test_mistral_pr4426.py: 5/5 byte-identical.
- tests/test_qwen_pr4426.py: 14/14 byte-identical.
- temp/sim/battery.py: 79/79 followup; 0 regressions vs baseline.
- Phase 3 Hermes-3 broken-LoRA reload: inference still returns
`'The answer to the equation 2+2 is 4.'`.
- Spot-checks on Hermes-3 / Phi-4 / Llama-3.2-1B / Gemma-3-1B real
stripped templates: probe still derives the expected prefix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in chat-template helpers
Pure comment minimization across `_find_end_position`,
`_has_add_generation_prompt_block`, `_if_body_emits_content`,
`_derive_assistant_prefix_by_render`, `_fix_chat_template` Case 2,
and `_VariantTokenizerProxy`. No behavior change; same intent,
fewer lines. All 21 follow-up tests and the 51-cell Phase 1 matrix
still pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Sandbox probe, fix is_sharegpt validator mismatch, reject negated gates
Three real bugs from the 10-agent Opus review:
1. Probe now uses `jinja2.sandbox.SandboxedEnvironment` instead of bare
`jinja2.Environment`. The probe renders at model-load time (before
the user calls `apply_chat_template`), so it was a new eager
code-execution surface that the base HF tokenizer loading does not
have. SandboxedEnvironment blocks attribute-chain exploits at
negligible cost.
2. `_repair_string_template` now tries validation with both
`is_sharegpt=False` and `is_sharegpt=True`. Previously, when
`_fix_chat_template` internally fell back to the other schema via
its dual-probe, the outer validation still used the caller's
original `is_sharegpt` -- rendering with the wrong message keys and
spuriously dropping a valid repair.
3. `_has_add_generation_prompt_block` now skips `If` nodes whose test
is a `Not` expression. A negated gate like
`{% if not add_generation_prompt %}{{ x }}{% endif %}` fires when
agp=False, so its emitting body is not a generation block -- but the
old code counted any Name reference regardless of polarity.
Cleanup: removed unused `self._label`, added `\r` escape in
generation-block literal, switched variant labels to `!r` formatting,
removed redundant `import os as _os`.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix jinja2.sandbox import and sandbox proxy fallback
Two critical findings from the 20-reviewer pass:
1. [20/20] The proxy read-only fallback used bare `jinja2.Environment`,
not sandboxed. All 20 reviewers independently reproduced marker-file
creation via `cycler.__init__.__globals__['os'].system(...)` during
`fix_chat_template()`. Fixed: fallback now uses
`from jinja2.sandbox import SandboxedEnvironment`.
2. [14/20] The render-diff probe did `import jinja2` then referenced
`jinja2.sandbox.SandboxedEnvironment`. `jinja2.sandbox` is a
submodule that is NOT auto-imported by `import jinja2` on Jinja 3.1.6.
This caused `AttributeError` (swallowed by `except Exception`),
making the entire Case 2 repair path silently return None in a clean
process. The 6 reviewers who saw it work had `jinja2.sandbox`
pre-imported by an earlier module in their process. Fixed: both the
probe and the proxy fallback now use
`from jinja2.sandbox import SandboxedEnvironment`.
* [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>
Reduce inline comments from ~160 lines to ~25 across both files.
Keep one-line summaries of the "why"; drop multi-paragraph rationale
blocks that repeated information already captured in commit messages
and PR discussion.
* fix: replacing SetEnvironmentVariable with direct registry API
* apply reviews
* Use CreateSubKey for HKCU\Environment
* Store PATH backup under HKCU\Software\Unsloth
* Fix $backupKey registry handle leak in PATH backup block
Wrap $backupKey operations in try/finally so the handle is closed even
if GetValue or SetValue throws. The Add-ToUserPath helper already uses
this pattern for its registry key -- the backup block was the only
place missing it.
* Isolate WM_SETTINGCHANGE broadcast from PATH write error handling
Wrap the broadcast dummy-variable calls in their own try/catch so a
broadcast failure does not mask a successful registry PATH write.
Previously, if SetEnvironmentVariable threw after SetValue already
committed the new PATH, Add-ToUserPath would return $false and the
caller would skip Refresh-SessionPath.
* PATH helper polish: venv precedence, quoted entries, raw/expanded dedup
Three small follow-ups surfaced by a 10-reviewer pass against the rebased
PR head. None fix a regression vs main; each strictly improves the new
helpers.
Refresh-SessionPath / Refresh-Environment:
- Move $env:Path to the front of the merge so an activated venv keeps
precedence over machine/user PATH after a refresh. Pre-PR dropped
process-only entries entirely; post-PR kept them but at the back.
- Dedup on both raw and expanded forms so %USERPROFILE%\foo and the
already-expanded C:\Users\me\foo do not both survive.
Add-ToUserPath:
- Trim whitespace and surrounding double-quotes from each compared entry
so quoted PATH entries like "C:\Program Files\CMake\bin" deduplicate
against an unquoted directory of the same path.
* Back up User PATH inside Add-ToUserPath, before first mutation
Previously only studio/setup.ps1 took a one-time PATH backup, at script
top (line ~547). install.ps1 (the irm | iex entry point) had no backup,
so users who installed via that path had no recovery surface if anything
clobbered their PATH. The PR description's "one-time backup before any
modifications" promise only held for the studio installer flow.
Move the backup into Add-ToUserPath itself: just before the first actual
SetValue mutation, write the pristine raw PATH to
HKCU\Software\Unsloth\PathBackup if no backup already exists. This:
- Covers both entry points (install.ps1 and studio/setup.ps1).
- Captures the TRUE pristine PATH even when install.ps1 runs first and
studio/setup.ps1 runs afterwards (the script-top backup in setup.ps1
would otherwise see an already-modified PATH).
- Is idempotent: once a backup exists, subsequent calls preserve it.
- Skips when nothing would mutate (dedup match) or PATH is empty.
The script-top backup in studio/setup.ps1 is kept for defense in depth.
* Refresh PATH: venv-aware merge order
Reconcile two competing concerns about Refresh-SessionPath /
Refresh-Environment surfaced by separate review rounds:
- venv at the back -> activated venv loses precedence to system Python
- process at the front -> stale shims (old node, old python, etc.)
still on $env:Path can beat a freshly installed tool
New merge order:
1. Activated venv Scripts dir, only if $env:VIRTUAL_ENV is set
2. Machine PATH freshly read from registry
3. User PATH freshly read from registry
4. Current $env:Path as fallback
This way an explicitly-activated venv keeps priority while a tool the
script just installed wins over any stale entry that was already on
the inherited shell PATH. When no venv is active, fresh registry
entries take precedence as expected.
* Append to User PATH by default, close $envKey in finally
Add-ToUserPath gains a -Position Append|Prepend parameter defaulting to
Append so installing unsloth no longer prepends the bundled venv Scripts
directory ahead of the user's existing python / pip on new shells. The
four current call sites (install.ps1 launcher, studio/setup.ps1 CMake,
nvcc, Python user Scripts) all take the Append default because each one
that needs in-session precedence already does an inline $env:Path prepend
independently. This matches rustup / cargo / nvm / pyenv / uv behavior.
Also wrap the script-top $envKey.GetValue in a try/finally so the
registry handle is released even if the read throws. Matches the pattern
already used for $backupKey five lines below.
* Prepend cmake, nvcc, Python Scripts; keep venv Scripts appended
The previous commit switched Add-ToUserPath to append by default so that
installing unsloth would not silently hijack the user's system python /
pip. That was correct for the venv Scripts dir (which contains python.exe
and pip.exe alongside unsloth.exe), but wrong for the three studio/setup
call sites. Those persist cmake, the driver-compatible nvcc, and the
Python user Scripts dir for future shells, and in all three cases an
older tool already earlier in the user PATH would keep winning after the
install finished. The nvcc case is especially load-bearing: setup selects
a driver-compatible CUDA toolkit, then llama.cpp builds against whatever
wins PATH resolution, so a stale older nvcc produces broken builds.
Pass -Position 'Prepend' explicitly at the three setup.ps1 call sites
(cmake at line 754, nvcc bin at line 1025, Python user Scripts at line
1191). None of those directories holds python.exe, so prepending them
does not re-introduce the original hijack problem. Leave the install.ps1
venv Scripts call on the default Append with a comment explaining why.
* Symmetric dedup, Prepend reorders duplicates, unsloth shim dir
Address three separate findings surfaced by review:
1. Dedup asymmetry (Gemini high-priority): the existing dedup expanded
registry entries via ExpandEnvironmentVariables but did NOT expand the
new directory. Passing "%USERPROFILE%\foo" when "C:\Users\me\foo" was
already in PATH produced a duplicate. Expand both sides so the check
is symmetric.
2. -Position Prepend no-op on existing duplicates: the dedup loop
returned $false as soon as it saw a match, regardless of position.
That left a late-position duplicate in place instead of moving it to
the front, so "prepend the newly selected cmake/nvcc" did not always
beat an older copy earlier in PATH. Partition entries into kept and
dropped lists, then reinsert a single copy at the requested position.
Append still returns $false on any match so user-curated orderings
are not reshuffled. Prepend also returns $false when the only copy
is already at position 0 so we preserve the user's casing.
3. Stop adding the venv Scripts dir to User PATH entirely. That dir
holds python.exe and pip.exe alongside unsloth.exe, so neither
Prepend nor Append worked: prepend hijacked the user's system python
and pip, append made the freshly-installed unsloth.exe lose to any
older unsloth.exe earlier on PATH. Replace the Scripts-dir PATH add
with a dedicated shim directory that contains only unsloth.cmd, and
prepend that dir. The shim calls the venv's unsloth.exe by absolute
path so future pip upgrades inside the venv propagate automatically.
* Shim via hardlink, Append user Scripts, drop venv sysconfig fallback
Three follow-ups to the c0ab1ab shim commit, targeting concerns raised in
the second 20-reviewer pass:
1. Shim uses unsloth.exe (hardlink, copy fallback) instead of unsloth.cmd.
The batch-file approach had three distinct regressions:
- cmd.exe expanded %...% sequences inside user arguments, so prompts
like "What does 50% mean?" got mangled before reaching the CLI
- Git Bash / MSYS2 / POSIX-style shells on Windows do not resolve
bare-name lookups to .cmd files, so `unsloth` stopped working there
- Set-Content -Encoding ASCII replaced non-ASCII profile characters
with '?', so installs under C:\Users\Jörg\... wrote a broken shim
A hardlink (fallback: copy) of unsloth.exe is a native Windows
executable with no shell indirection. PATHEXT picks .exe before .cmd
in cmd.exe and PowerShell, Git Bash honors .exe natively, subprocess
callers hit it directly, and a hardlink stays in sync with the venv
on pip upgrades because both names point at the same inode.
2. studio/setup.ps1 Python user Scripts dir is added with default Append
instead of -Position Prepend. That directory holds every pip-installed
user console script (pip, pytest, huggingface-cli, and so on), not
just unsloth, so reordering it silently changed resolution order for
unrelated tools. The new install.ps1 shim at PATH position 0 already
guarantees `unsloth` resolves to the freshly installed copy, so the
Python user Scripts entry only needs to be present, not at the front.
3. The sysconfig lookup in studio/setup.ps1 no longer falls back to
sysconfig.get_path('scripts') when the nt_user scheme dir does not
exist. When setup.ps1 is invoked from an activated venv (a flow the
linked issue actually hits) that fallback returns the venv's Scripts
directory, which would then be added to the persisted User PATH and
re-introduce the python / pip hijack the shim dir is meant to avoid.
Stick strictly to the nt_user scheme; skip the block if it does not
exist on disk.
* Do not crash installer when unsloth.exe shim is locked
The shim update sequence at install.ps1:1095 did a bare Remove-Item /
New-Item HardLink / Copy-Item. Under the script's $ErrorActionPreference
a locked target (most commonly 'unsloth studio' still running while the
user re-invokes the installer) turns the Remove-Item failure into a
terminating error that aborts the install with no actionable message.
The existing shim is perfectly usable in that state, so there is no
reason to abort. Wrap the whole remove/link/copy sequence in a try/catch
that logs the probable cause (Studio still running), points at the fix
(close Studio and re-run), and lets the installer finish with the old
launcher still serving the command.
Also only emit the "added unsloth launcher to PATH" step line when the
launcher was actually (re)created AND the PATH entry was newly added --
previously the message fired even when the shim refresh silently failed,
which was confusing.
* Guard shim PATH entry on existence, use NullString for broadcast delete
Two follow-ups surfaced by the latest review pass:
1. Do not add the shim directory to User PATH when the launcher was not
actually created. Antivirus blocking unsloth.exe, a disk-full volume,
or restrictive filesystem permissions can make both the hardlink and
the copy fallback fail on a fresh install. In that case the existing
sequence would report "added unsloth launcher to PATH" warnings but
still prepend the empty $ShimDir to User PATH -- the user sees an
install that claims success but then cannot resolve `unsloth` in a
new shell. Gate Add-ToUserPath on Test-Path $ShimExe so the PATH
entry is only persisted when the launcher is really there.
2. Pass [NullString]::Value instead of $null to the broadcast-delete
call in Add-ToUserPath. On PowerShell 7.5 and later (running on .NET
9), a bare $null going into [Environment]::SetEnvironmentVariable
can be coerced to an empty string rather than a true .NET null,
which sets the dummy UnslothPathRefresh_XXXXXXXX variable to "" in
HKCU\Environment instead of deleting it. The leaked variable is
visible in System Properties and accumulates one entry per install
run. [NullString]::Value is a PowerShell-specific sentinel that
crosses the interop boundary as a real null and works on both PS 5.1
and PS 7.x. See PowerShell/PowerShell#24637 for the underlying issue.
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Fixes#4150.
Pre-PR, `_fix_chat_template` only patched templates where a trailing `{{ ... }}` expression followed the last `{% endfor %}`. ChatML templates (Hermes, Magnum, Phi-4, etc.) that end cleanly at `{% endfor %}` with no generation-prompt block were left unchanged, so the outer `fix_chat_template` raised:
```
RuntimeError: Unsloth: The tokenizer `...` does not have a
{% if add_generation_prompt %} for generation purposes.
```
This commonly shows up when a downstream tool (LlamaFactory, Axolotl) re-serializes the tokenizer during LoRA save and strips the generation-prompt block.
This PR adds a second branch to `_fix_chat_template` that fires when:
- the content after the last `{% endfor %}` is empty modulo Jinja `{# ... #}` comments,
- the scrubbed template contains `<|im_start|>` and `<|im_end|>`,
- and the scrubbed template does not already mention `add_generation_prompt`.
The assistant-turn separator is inferred from the template itself (preferring an explicit `'<|im_start|>assistant<sep>'` literal, then the unique `message['role'] + '<sep>'` from role concatenations, then `<|im_sep|>` for Phi-4-mini mixed-separator templates, then `\n`), so Phi-4-style templates are not silently corrupted with the wrong separator.
Verified against the existing chat-template corpus:
- Hermes-3, Magnum-v2, Phi-4-mini, Phi-4 multi-sep, ChatML with trailing whitespace, ChatML with trailing Jinja comment, dot-access `message.role`, split-literal `'<|im_start|>assistant'`: all repaired with the correct assistant prefix.
- Already-fixed ChatML templates: idempotent NOP.
- Trap templates with `<|im_start|>` only inside a Jinja comment: correctly not rewritten.
- Llama-3, Gemma-3, Qwen2.5 (non-ChatML): byte-identical.
- Mistral family (5 models including Mistral-Nemo, Mistral-Small-24B, Mixtral): byte-identical, protected both by the structural guard (no ChatML tokens) and the existing name-based exemption in `load_correct_tokenizer`.
- Qwen family (14 models including Qwen2.5, Qwen3, Qwen3-Coder, QwQ, VL, Math, Qwen3-Guard): byte-identical.
End-to-end reproduction: Hermes-3 LoRA SFT, save with stripped chat_template, reload. Pre-PR code path raises the RuntimeError above. Post-PR reload loads cleanly, patches the template at load time, and `apply_chat_template(add_generation_prompt=True)` produces the correct `<|im_start|>assistant\n` prefix.
* fix pass attn implementation
* [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: add folder browser modal for Custom Folders
The Custom Folders row in the model picker currently only accepts a
typed path. On a remote-served Studio (Colab, shared workstation) that
means the user has to guess or paste the exact server-side absolute
path. A native browser folder picker can't solve this: HTML
`<input type="file" webkitdirectory>` hides the absolute path for
security, and the File System Access API (Chrome/Edge only) returns
handles rather than strings, neither of which the server can act on.
This PR adds a small in-app directory browser that lists paths on the
server and hands the chosen string back to the existing
`POST /api/models/scan-folders` flow.
## Backend
* New endpoint `GET /api/models/browse-folders`:
* `path` query param (expands `~`, accepts relative or absolute; empty
defaults to the user's home directory).
* `show_hidden` boolean to include dotfiles/dotdirs.
* Returns `{current, parent, entries[], suggestions[]}`. `parent` is
null at the filesystem root.
* Immediate subdirectories only (no recursion); files are never
returned.
* `entries[].has_models` is a cheap hint: the directory looks like it
holds models if it is named `models--*` (HF hub cache layout) or
one of the first 64 children is a .gguf/.safetensors/config.json/
adapter_config.json or another `models--*` subfolder.
* Sort order: model-bearing dirs, then plain, then hidden; case-
insensitive alphabetical within each bucket.
* Suggestions auto-populate from HOME, the HF cache root, and any
already-registered scan folders, deduplicated.
* Error surface: 404 for missing path, 400 for non-directory, 403 on
permission errors. Auth-required like the other models routes.
* New Pydantic schemas `BrowseEntry` and `BrowseFoldersResponse` in
`studio/backend/models/models.py`.
## Frontend
* New `FolderBrowser` component
(`studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx`)
using the existing `Dialog` primitive. Features:
* Clickable breadcrumb with a `..` row for parent navigation.
* Quick-pick chips for the server-provided suggestions.
* `Show hidden` checkbox.
* In-flight fetch cancellation via AbortController so rapid
navigation doesn't flash stale results.
* Badges model-bearing directories inline.
* `chat-api.ts` gains `browseFolders(path?, showHidden?)` and matching
types.
* `pickers.tsx` adds a folder-magnifier icon next to the existing `Add`
button. Opening the browser seeds it with whatever the user has
already typed; confirming fills the text input, leaving the existing
validation and save flow unchanged.
## What it does NOT change
* The existing text-input flow still works; the browser is additive.
* No new permissions or escalation; the endpoint reads only directories
the server process is already allowed to read.
* No model scanning or filesystem mutation happens from the browser
itself -- it just returns basenames for render.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: cap folder-browser entries and expose truncated flag
Pointing the folder browser at a huge directory (``/usr/lib``,
``/proc``, or a synthetic tree with thousands of subfolders) previously
walked the whole listing and stat-probed every child via
``_looks_like_model_dir``. That is both a DoS shape for the server
process and a large-payload surprise for the client.
Introduce a hard cap of 2000 subdirectory entries and a
``truncated: bool`` field on the response. The frontend renders a small
hint below the list when it fires, prompting the user to narrow the
path. Below-cap directories are unchanged.
Verified end-to-end against the live backend with a synthetic tree of
2050 directories: response lands at 2000 entries, ``truncated=true``,
listing finishes in sub-second time (versus tens of seconds if we were
stat-storming).
* Studio: suggest LM Studio / Ollama dirs + 2-level model probe
Three improvements to the folder-browser, driven by actually dropping
an LM Studio-style install (publisher/model/weights.gguf) into the
sandbox and walking the UX:
## 1. Quick-pick chips for other local-LLM tools
`well_known_model_dirs()` (new) returns paths commonly used by
adjacent tools. Only paths that exist are returned so the UI never
shows dead chips.
* LM Studio current + legacy roots + user-configured
`downloadsFolder` from its `settings.json` (reuses the existing
`lmstudio_model_dirs()` helper).
* Ollama: `$OLLAMA_MODELS` env override, then `~/.ollama/models`,
`/usr/share/ollama/.ollama/models`, and `/var/lib/ollama/.ollama/models`
(the systemd-service install path surfaced in the upstream "where is
everything?" issue).
* Generic user-choice locations: `~/models`, `~/Models`.
Dedup is stable across all sources.
## 2. Two-level model-bearing probe
LM Studio and Ollama both use `root/publisher/model/weights.gguf`.
The previous `has_models` heuristic only probed one level, so the
publisher dir (whose immediate children are model dirs, not weight
files) was always marked as non-model-bearing. Pulled the direct-
signal logic into `_has_direct_model_signal` and added a grandchild
probe so the classic layout is now recognised.
Still O(PROBE^2) worst-case, still returns immediately for
`models--*` names (HF cache layout) and for any direct weight file.
## 3. model_files_here hint on response body
A leaf model dir (just GGUFs, no subdirs) previously rendered as
`(empty directory)` in the modal, confusing users into thinking the
folder wasn't scannable. Added a `model_files_here` count on the
response (capped at 200) and a small hint row in the modal: `N model
files in this folder. Click "Use this folder" to scan it.`
## Verification
Simulated an LM Studio install by downloading the real 84 MB
`unsloth/SmolLM2-135M-Instruct-Q2_K.gguf` into
`~/.lmstudio/models/unsloth/SmolLM2-135M-Instruct-GGUF/`. Confirmed
end-to-end:
* Home listing suggests `~/.lmstudio/models` as a chip.
* Browsing `~/.lmstudio/models` flags `unsloth` (publisher) as
`has_models=true` via the 2-level probe.
* Browsing the publisher flags `SmolLM2-135M-Instruct-GGUF` (model
dir) as `has_models=true`.
* Browsing the model dir returns empty entries but
`model_files_here=1`, and the frontend renders a hint telling the
user it is a valid target.
* Studio: one-click scan-folder add + prominent remove + plain search icon
Three small Custom Folders UX fixes after real-use walkthrough:
* **One-click add from the folder browser**. Confirming `Use this
folder` now submits the path directly to
`POST /api/models/scan-folders` instead of just populating the text
input. `handleAddFolder` takes an optional explicit path so the
submit lands in the same tick as `setFolderInput`, avoiding a
state-flush race. The typed-path + `Add` button flow is unchanged.
* **Prominent remove X on scan folders**. The per-folder delete
button was `text-muted-foreground/40` and hidden entirely on
desktop until hovered (`md:opacity-0 md:group-hover:opacity-100`).
Dropped the hover-only cloak, bumped color to `text-foreground/70`,
added a red hover/focus background, and sized the icon up from
`size-2.5` to `size-3`. Always visible on every viewport.
* **Plain search icon for the Browse button**. `FolderSearchIcon`
replaced with `Search01Icon` so it reads as a simple "find a
folder" action alongside the existing `Add01Icon`.
* Studio: align Custom Folders + and X buttons on the same right edge
The Custom Folders header used `px-2.5` with a `p-0.5` icon button,
while each folder row used `px-3` with a `p-1` button. That put the
X icon 4px further from the right edge than the +. Normalised both
rows to `px-2.5` with `p-1` so the two icons share a column.
* Studio: empty-state button opens the folder browser directly
The first-run empty state for Custom Folders was a text link reading
"+ Add a folder to scan for local models" whose click toggled the
text input. That's the wrong default: a user hitting the empty state
usually doesn't know what absolute path to type, which is exactly
what the folder browser is for.
* Reword to "Browse for a models folder" with a search-icon
affordance so the label matches what the click does.
* Click opens the folder browser modal directly. The typed-path +
Add button flow is still available via the + icon in the
section header, so users who know their path keep that option.
* Slightly bump the muted foreground opacity (70 -> hover:foreground)
so the button reads as a primary empty-state action rather than a
throwaway hint.
* Studio: Custom Folders header gets a dedicated search + add button pair
The Custom Folders section header had a single toggle button that
flipped between + and X. That put the folder-browser entry point
behind the separate empty-state link. Cleaner layout: two buttons in
the header, search first, then add.
* Search icon (left) opens the folder browser modal directly.
* Plus icon (right) toggles the text-path input (unchanged).
* The first-run empty-state link is removed -- the two header icons
cover both flows on every state.
Both buttons share the same padding / icon size so they line up with
each other and with the per-folder remove X.
* Studio: sandbox folder browser + bound caps + UX recoveries
PR review fixes for the Custom Folders folder browser. Closes the
high-severity CodeQL path-traversal alert and addresses the codex /
gemini P2 findings.
Backend (studio/backend/routes/models.py):
* New _build_browse_allowlist + _is_path_inside_allowlist sandbox.
browse_folders now refuses any target that doesn't resolve under
HOME, HF cache, Studio dirs, registered scan folders, or the
well-known third-party model dirs. realpath() is used so symlink
traversal cannot escape the sandbox. Also gates the parent crumb
so the up-row hides instead of 403'ing.
* _BROWSE_ENTRY_CAP now bounds *visited* iterdir entries, not
*appended* entries. Dirs full of files (or hidden subdirs when
show_hidden is False) used to defeat the cap.
* _count_model_files gets the same visited-count fix.
* PermissionError no longer swallowed silently inside the
enumeration / counter loops -- now logged at debug.
Frontend (folder-browser.tsx, pickers.tsx, chat-api.ts):
* splitBreadcrumb stops mangling literal backslashes inside POSIX
filenames; only Windows-style absolute paths trigger separator
normalization. The Windows drive crumb value is now C:/ (drive
root) instead of C: (drive-relative CWD-on-C).
* browseFolders accepts and forwards an AbortSignal so cancelled
navigations actually cancel the in-flight backend enumeration.
* On initial-path fetch error, FolderBrowser now falls back to HOME
instead of leaving the modal as an empty dead end.
* When the auto-add path (one-click "Use this folder") fails, the
failure now surfaces via toast in addition to the inline
paragraph (which is hidden when the typed-input panel is closed).
* Studio: rebuild browse target from trusted root for CodeQL clean dataflow
CodeQL's py/path-injection rule kept flagging the post-validation
filesystem operations because the sandbox check lived inside a
helper function (_is_path_inside_allowlist) and CodeQL only does
intra-procedural taint tracking by default. The user-derived
``target`` was still flowing into ``target.exists`` /
``target.is_dir`` / ``target.iterdir``.
The fix: after resolving the user-supplied ``candidate_path``,
locate the matching trusted root from the allowlist and rebuild
``target`` by appending each individually-validated segment to
that trusted root. Each segment is rejected if it isn't a single
safe path component (no separators, no ``..``, no empty/dot).
The downstream filesystem ops now operate on a Path constructed
entirely from ``allowed_roots`` (trusted) plus those validated
segments, so CodeQL's dataflow no longer sees a tainted source.
Behavior is unchanged for all valid inputs -- only the
construction of ``target`` is restructured. Live + unit tests
all pass (58 selected, 7 deselected for Playwright env).
* Studio: walk browse paths from trusted roots for CodeQL
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Ubuntu <ubuntu@h100-8-cheapest.us-east5-a.c.unsloth.internal>
* Reapply "updated models template mappers. added lfm2.5vl450m to transformers 5…" (#4945)
This reverts commit 33503ea248.
* Add missing gemma-4-31B-it bnb-4bit mapper entry and LFM2.5 upstream namespace for PR #4950
- Add unsloth/gemma-4-31B-it-unsloth-bnb-4bit to __INT_TO_FLOAT_MAPPER so
the int-to-float resolution works for this model (already listed in
TEMPLATE_TO_MODEL_MAPPER but had no mapper entry).
- Add LiquidAI/LFM2.5-1.2B-Instruct to lfm-2.5 TEMPLATE_TO_MODEL_MAPPER
entry so the canonical upstream namespace is mapped consistently with lfm-2.
* Add missing gemma-4-31B-it bnb-4bit Ollama mapping and lfm-2.5 chat template alias
- Add unsloth/gemma-4-31B-it-unsloth-bnb-4bit to OLLAMA_TEMPLATE_TO_MODEL_MAPPER
so Ollama export works for this model (E2B-it and E4B-it bnb-4bit variants were
already present, 31B-it was inconsistently omitted)
- Register CHAT_TEMPLATES["lfm-2.5"] as alias of the lfm-2 template to prevent
KeyError when Studio resolves LFM2.5 models through MODEL_TO_TEMPLATE_MAPPER
* Add missing LFM2 bnb-4bit INT_TO_FLOAT_MAPPER entry
unsloth/LFM2-1.2B-unsloth-bnb-4bit is referenced in model_mappings.py
but had no mapper.py entry, so model resolution would fail when users
load that variant with load_in_4bit=False or when the float name is
used with load_in_4bit=True.
* Fix review findings for PR #16
1. ollama_template_mappers.py: Restore dropped Gemma-4 base model IDs
(E2B, E4B, 31B, 26B-A4B) and add missing google/ upstream IDs to
the gemma4 Ollama mapper for consistency with other gemma entries.
2. mapper.py: Remove self-mapping non-bnb-4bit entries from
__INT_TO_FLOAT_MAPPER that were polluting FLOAT_TO_INT_MAPPER with
lowercase 16-bit names, causing load_in_4bit=True to return bad
model names. Add direct MAP_TO_UNSLOTH_16bit entries to preserve
the google->unsloth 16-bit redirects.
3. mapper.py: Add LFM2.5 MAP_TO_UNSLOTH_16bit redirect so
LiquidAI/LFM2.5-1.2B-Instruct resolves to its unsloth mirror.
* Add review tests for PR #4950
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove top-level test files
These test_*.py files were added at the repo root rather than under tests/.
Removing them from this PR; the production mapper changes remain.
* Add gemma-4-26B-A4B-it mapping
Adds unsloth/gemma-4-26B-A4B-it to __INT_TO_FLOAT_MAPPER as a 2-tuple so
google/gemma-4-26B-A4B-it routes to unsloth/gemma-4-26B-A4B-it across
INT_TO_FLOAT_MAPPER, FLOAT_TO_INT_MAPPER, and MAP_TO_UNSLOTH_16bit.
The 26B-A4B (MoE) model has no bnb-4bit variant, so the key uses the
plain unsloth name rather than the -unsloth-bnb-4bit suffix.
Removes the now-redundant standalone _add_with_lower call for the -it
variant; the 16bit mapping is registered via the dict loop.
* Add unsloth-bnb-4bit mappings for gemma-4 base (non-it) models
Adds E2B, E4B, 31B base unsloth-bnb-4bit entries to __INT_TO_FLOAT_MAPPER.
The 26B-A4B (MoE) base has no bnb-4bit variant on HF, so it stays on the
standalone _add_with_lower line for the 16bit-only routing.
Removes the redundant _add_with_lower lines for E2B, E4B, 31B base since
the dict loop now registers the same google->unsloth route through the
2-tuple entries, plus full FLOAT_TO_INT and INT_TO_FLOAT coverage.
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* feat: Add cactus QAT scheme support
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test(qat): add tests for cactus QAT scheme and fix missing import
* Fix cactus QAT scheme: correct MappingType import, tighten PerGroup filter
- Drop the broken `from torchao.dtypes import MappingType` import. `MappingType`
lives in `torchao.quantization` (and `torchao.quantization.quant_primitives`);
it is not exported from `torchao.dtypes` in any supported torchao release
(verified on 0.14, 0.16, 0.17). The previous code raised `ImportError` on
every cactus call and was masked as a misleading 'torchao not found' error.
- Since `IntxWeightOnlyConfig` already defaults `mapping_type` to
`MappingType.SYMMETRIC`, drop the explicit kwarg entirely and remove the
import. Behavior is unchanged.
- Introduce a named `group_size = 32` constant (matches the int4 / fp8-int4
pattern in the surrounding branches) and add a `% group_size == 0`
divisibility guard to the filter. `PerGroup(32)` requires
`in_features % 32 == 0` at `quantize_()` time, otherwise torchao raises
`ValueError: in_features (N) % group_size (32) must be == 0`. The old
`in_features >= 32` filter would admit non-aligned widths (e.g. 33, 48, 65,
127) and crash `_prepare_model_for_qat` for those shapes.
* Warn when cactus QAT skips non-divisible Linear layers
Multiple reviewers flagged that the divisibility guard added in the
previous commit can silently leave Linear layers in full precision when
their in_features is not a multiple of 32. For currently supported
Unsloth models (Qwen, Llama, Gemma, Mistral, Phi) every Linear width is
already a multiple of 32/64/128 so this never triggers, but surfacing
the coverage gap is cheap and avoids users assuming 100% QAT coverage
when they bring a custom model with unusual shapes.
Emit a UserWarning listing up to the first 8 skipped layers whenever
the cactus filter excludes any Linear due to the modulo guard. This
keeps the lenient silent-skip behavior (consistent with int4 /
fp8-int4), but stops making it silent.
---------
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: Add support for OLMo-3 model in mapping and tests
* Update unsloth/models/mapper.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update tests/test_get_model_name.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Fix casing, add Think variants, and align version gate for OLMo-3 PR 4678
Mapper: switch slugs from OLMo-3 to canonical Olmo-3 mixed case, drop the
non-existent unsloth/Olmo-3-7B-Instruct-bnb-4bit dead alias, and add the
already-published Olmo-3-7B-Think and Olmo-3-32B-Think Unsloth mirrors.
Loader: change the olmo3 transformers version gate from Version("4.57.0")
to Version("4.57.0.dev0") so nightly/source builds that already contain
olmo3 are not blocked, matching the OLMo-2, Gemma 3 and Cohere patterns.
* Use canonical Olmo-3 casing and cover Think variants in OLMo-3 tests
Mirrors the mapper.py fixes on pr-4678-code: HuggingFace canonical slugs
for the OLMo-3 family use mixed-case Olmo-3 (not OLMo-3 like OLMo-2), and
Unsloth already hosts Olmo-3-7B-Think and Olmo-3-32B-Think mirrors, so
the resolution matrix now covers all three published Olmo-3 families.
---------
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: refresh Downloaded GGUF list and recurse into variant subdirs
Two fixes for the model picker's "Downloaded" section.
Frontend (`pickers.tsx`):
* `HubModelPicker`'s mount effect short-circuited the cached-gguf and
cached-models refetch whenever the module-level cache already had
entries (`if (alreadyCached) return;`). After downloading a new repo
in the same session, reopening the picker rendered the stale cache
and the new repo never appeared in "Downloaded" until a full page
reload. The early return is removed so the lists are always refreshed
on mount; the module cache still drives the initial render so there
is no spinner flash when we already had data.
Backend (`utils/models/model_config.py`):
* `list_local_gguf_variants` and `_find_local_gguf_by_variant` used a
non-recursive `Path.glob("*.gguf")`. Some HF GGUF repos (e.g.
`unsloth/gemma-4-26B-A4B-it-GGUF`) place the largest quants under a
variant-named subdirectory such as `BF16/...gguf`, which the
top-level glob missed. Both helpers now use `rglob` and the variant
filename is stored as a path relative to the scan root so the
locator can still find the file.
The flat-layout case (variants directly in the snapshot root) is
unchanged: verified against `unsloth/gemma-4-E2B-it-GGUF` which still
returns its UD-Q4_K_XL variant correctly.
* Studio: emit posix-style relative filenames for local GGUF subdirs
`list_local_gguf_variants` was doing `str(f.relative_to(p))`, which on
Windows produces backslash-separated paths like `BF16\foo.gguf`. The
remote `list_gguf_variants` (HF API path) always returns forward-slash
filenames such as `BF16/foo.gguf`, so the two would diverge on Windows.
Switch to `.as_posix()` so the local and remote variant filenames stay
identical across Linux, macOS, and Windows. Verified by simulating with
`PureWindowsPath` in the test suite.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: detect mmproj at snapshot root for nested-variant layouts
When _find_local_gguf_by_variant returns a weight file inside a
quant-named subdir (e.g. snapshot/BF16/foo.gguf), detect_mmproj_file
was scanning only the immediate parent and missing the mmproj file
sitting at the snapshot root. The model was then loaded without
--mmproj, silently breaking vision support for repos that ship
nested variants.
detect_mmproj_file now takes an optional search_root and walks up
from the weight file to that root, in order, so the mmproj at the
snapshot root is picked up. Sibling quant subdirs are not scanned,
so an unrelated variant's mmproj does not leak in.
Also apply the suggested micro-optimization on relative_to in
list_local_gguf_variants -- only build the posix path when storing
the first file for a quant.
* [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>
danielhanchen