Train a decision model from a plain language model by danielhanchen · Pull Request #12772 · unslothai/unsloth

danielhanchen · 2026-10-05T19:07:10Z

Builds on #12585 (decision model fine-tuning for Laya and Clef). Until that merges, this PR also shows its commits; the new work is the last six commits, starting at "Turn a plain language model into a Clef-style decision model".

#12585 fine-tunes existing decision models. This PR turns a plain language model, for example unsloth/Qwen3.5-2B or unsloth/Qwen3.5-0.8B, into a Jev style decision model: one prefill over the state and the questions, then a scored distribution over every option of every question, with no generation.

Usage

from unsloth import FastDecisionModel, DecisionTrainer
from unsloth.models import decision_datasets as dd

model, tokenizer = FastDecisionModel.from_pretrained(
    "unsloth/Qwen3.5-2B", max_seq_length = 4096, decision_head = "clef",
)
model = FastDecisionModel.get_peft_model(model, r = 16, lora_alpha = 16)
rows = dd.build_decision_mixture(["banking77", "clinc150", "mnli", "boolq", "mmlu", "typed_decisions"], n_rows = 12000)
items, report = FastDecisionModel.build_dataset(rows, tokenizer, model)
train, held = FastDecisionModel.split_holdout(items)
DecisionTrainer(model = model, args = args, train_dataset = train, processing_class = tokenizer).train()
FastDecisionModel.calibrate(model, tokenizer, held)
model.save_pretrained_merged("qwen35_2b_decision", tokenizer)

Or end to end: python scripts/train_decision_from_lm.py --model unsloth/Qwen3.5-2B --rows 12000 --out outputs/qwen35_2b_decision.

What it adds

decision_head = "clef" in FastDecisionModel.from_pretrained: loads the backbone with the same rules as Clef in Studio: fine-tune Laya and Cloudflare Clef decision models and serve them #12585 (Unsloth's Qwen3.5 fast path, dynamic 4-bit, the float32 path for float16 requests) and attaches a fresh joint schema head sized from the backbone. Text-only backbones load through AutoModelForCausalLM.
Training, calibration and saving reuse DecisionTrainer, calibrate and save_pretrained_merged from Studio: fine-tune Laya and Cloudflare Clef decision models and serve them #12585 unchanged. freeze_backbone / unfreeze_backbone allow an optional head-only warm-up.
A saved model uses Cloudflare's Clef layout and ships their joint_schema_model.py, so it also loads with Cloudflare's load_release_model. The test converts a tiny Qwen3.5, saves it and checks that Cloudflare's loader picks the same option as ours on every question.
unsloth/models/decision_datasets.py: converters from public datasets (intent: banking77, clinc150; NLI: MNLI, SNLI, WANLI; BoolQ, AG News, SST-5, MMLU, CommonsenseQA, ARC, prompt injection, xLAM tool calls, typed-decisions) into the Decision API row format. Mixing adds schema augmentation (renamed option codes, subsampled and shuffled options, derived questions) so the head learns to read the schema rather than memorise label strings, plus n-gram decontamination against held-out sets. BFCL and When2Call are evaluation only.

Results

One B200, LoRA r16, 4,096 tokens, accuracy on held-out test decisions (typed-decisions 2,000, banking77 and clinc150 500 each):

Model	Training data	Steps	Peak VRAM	typed-decisions	banking77	clinc150
Qwen3.5-2B, untrained head				33.3%	1.4%	1.6%
Qwen3.5-2B	typed-decisions only, 1,200 rows	210	7.6 GB	78.4%	58.8%	62.2%
Qwen3.5-2B	13-source mixture, 13,200 rows	810	7.8 GB	76.2%	77.0%	79.0%
Qwen3.5-0.8B	13-source mixture, 13,200 rows	810	4.3 GB	73.7%	74.4%	76.6%

For reference, Cloudflare's Clef-flash scores 70.9% on typed-decisions before fine-tuning and 77.8% after QLoRA in #12585. A head-only warm-up of 150 steps before LoRA made no difference (75.8% / 76.2% / 78.6%), so it is off by default.

Tests

tests/test_decision_from_lm.py (tiny Qwen3 and Qwen3.5: load, train, calibrate, save, reload, Cloudflare loader parity) and tests/test_decision_datasets.py (every converter on fixture rows, augmentation keeps the gold answer, decontamination) pass on GPU and CPU.

…ady the toy test

…he repo root

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

…memory test on a real CUDA device

…ey engage

Clef and Clef-flash repos (config.json plus joint_head.*) are detected as decision models next to Laya. Training goes through FastDecisionModel with QLoRA allowed and the Qwen3.5 fast-path kernel setup the worker skips for Laya. The Decision API lists clef and clef-flash and Clef fine-tunes, and serves them from a spawn child that loads the backbone through Unsloth, so unloading returns the GPU memory. Clef waits for the GPU while a run trains instead of loading on CPU.

… the loss here

…ier loss options

… on CPU when causal_conv1d is installed

…pile it

… route test

…les, honour a cancel during calibration

# Conflicts: # studio/backend/core/training/worker.py

…ckets

… accepts it

…ure into its head

…at16

…-op accumulation wrapper

…un, and give its asserts messages

…n-from-lm

… a precision flag

danielhanchen · 2026-10-07T11:04:54Z

@codex review

danielhanchen · 2026-10-07T11:05:07Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8e4eb8cb97

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+        if load_in_4bit and kwargs.get("quantization_config") is None:
+            kwargs["quantization_config"] = _clef_bnb_config(dtype)


Disable the explicit 4-bit config for full finetuning

When callers combine load_in_4bit=True with full_finetuning=True, this installs an explicit bitsandbytes configuration. FastModel.from_pretrained disables its load_in_4bit flag for full finetuning, but the explicit configuration remains and still quantizes the backbone, so the requested full finetune may either fail or train only the remaining non-quantized parameters. Match the existing Clef loader by adding not full_finetuning to this condition.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: full finetuning no longer installs the 4-bit config, matching _load_clef.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+    max_len = int(max_seq_length or CLEF_MAX_LEN)
+    if revision is not None:
+        kwargs["revision"] = revision
+    backbone, processor, fast = _load_backbone(


Honor the requested revision and offline mode

The new loading path never forwards local_files_only to the actual backbone load, so local_files_only=True can still contact the Hub on both GPU and CPU. In addition, revision is placed in kwargs here but the CPU branch ignores kwargs, causing CPU callers to load the repository default rather than the requested revision. Forward these controls to both the FastModel and Transformers config/model/tokenizer loads so offline use and reproducible checkpoint selection work as advertised.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: revision and local_files_only reach FastModel and every CPU config, tokenizer and model load.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+        if new_id in questions:
+            new_id = qid
+        questions[new_id], gold[new_id] = question, answer


Preserve questions when renamed IDs collide

For a multi-question row where a later original ID is one of the generated IDs (for example q1), an earlier question can first be renamed to that ID. When the later question is processed, this collision fallback assigns qid, which is the same already-occupied key, and the following assignment silently overwrites the earlier question and gold label. Check candidates against all original IDs or generate a genuinely unused fallback instead of reusing an occupied qid.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: a rename never takes another question's original id, and collisions get a fresh id.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+            ),
+        )
+
+    torch.cuda.reset_peak_memory_stats()


Guard CUDA memory reset on CPU runs

When this script is run on a CPU-only PyTorch installation, the unconditional torch.cuda.reset_peak_memory_stats() call raises before training starts. The rest of the script explicitly supports the CPU case when choosing precision and reporting memory/device details, so this call should likewise be guarded by torch.cuda.is_available().

Useful? React with 👍 / 👎.

Fixed in 084b6e6: the peak memory reset only runs when CUDA is available.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+    if kind == "score" and isinstance(label, int):
+        levels = question["criteria"]
+        threshold = rng.randint(1, len(levels) - 1)


Skip derived thresholds for single-level scores

A score question with one criterion is accepted by the decision-schema validation, but when derived-question augmentation is selected this calls randint(1, 0) and aborts the entire augmentation. This can occur with typed/custom decision rows under the default probabilistic augmentation; return no derived question when len(levels) < 2.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: a one-level score gets no derived threshold question.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+    out = {k: v for k, v in row.items() if k != "_variants"}
+    if max_options is not None:
+        rng = random.Random(seed)
+        for qid, question in list(out["questions"].items()):
+            if question["type"] == "choice" and isinstance(question.get("criteria"), dict):
+                out["questions"][qid], out["gold"][qid] = _subsample_options(
+                    question, out["gold"][qid], rng, max_options


Copy nested mappings before canonical subsampling

When max_options is supplied, out["questions"] and out["gold"] still reference the input row's mappings, so these assignments permanently replace entries in row as well. A caller that later reuses the converted row for another evaluation or augmentation therefore sees the already-subsampled option set. Copy the nested question and gold mappings before modifying them.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: canonical_row copies the questions and gold mappings before subsampling.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+        share = max(1, round(n_rows * weights[name] / total))
+        picks = (
+            rng.sample(pool, share)
+            if share <= len(pool)
+            else [rng.choice(pool) for _ in range(share)]


Allocate mixture remainders instead of returning too few rows

Independent rounding can make the source shares sum to less than n_rows, and the final slice cannot restore the missing rows. For example, the documented default of 12,000 rows across 13 equal nonempty pools allocates 923 each and returns 11,999 rows; smaller requests such as 5 rows across 2 pools return only 4. Allocate the rounding remainder across sources so the function fulfills its requested row count whenever any pool is available.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: shares are allocated by largest remainder, so the mixture returns exactly n_rows.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+    per_source = max(1, args.rows // max(1, len(sources)))
+    for name in sources:
+        try:
+            pools[name] = dd.load_source(name, "train", limit = per_source, seed = args.seed)


Recompute source limits after unavailable datasets are skipped

The per-source limit is calculated from every requested source before gated or unavailable sources are removed. If one default source such as gated xLAM fails, each remaining pool is still capped at rows / 13, while build_decision_mixture recomputes larger shares across 12 pools and fills the difference by sampling those truncated pools with replacement. This injects avoidable duplicate training examples; load enough rows based on the surviving sources or avoid pre-truncating before failures are known.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: unavailable sources are dropped first, then the per-source limit is computed from the survivors.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+    if args.typed_decisions:
+        rng = random.Random(args.seed)
+        rows += [dd.augment_row(r, rng) for r in dd.load_source("typed_decisions", "train")]
+        random.Random(args.seed).shuffle(rows)


Decontaminate typed-decisions before adding it to training

With the default --typed-decisions setting, these training rows are appended only after build_decision_mixture has applied decontaminate_against. Consequently, any typed-decisions training state that overlaps its test split or the BANKING77/CLINC150 evaluation states remains in training and can leak benchmark content into the reported tuned metrics. Apply the same decontaminator to this source before augmentation and concatenation.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: typed-decisions training rows go through the same decontaminator as the mixture.

chatgpt-codex-connector · 2026-10-07T11:14:21Z

+            return load_lm_as_decision_model(
+                model_name,
+                max_seq_length = max_seq_length,


Forward subfolder when converting a plain language model

When decision_head is set, the public subfolder argument is consumed by FastDecisionModel.from_pretrained but omitted from this delegated call. A plain LM stored in a repository or local root subfolder is therefore loaded from the root instead, producing either a missing-config error or the wrong model. Pass the requested subfolder through to the underlying backbone loader.

Useful? React with 👍 / 👎.

Fixed in 084b6e6: a subfolder is resolved (local or Hub snapshot) before the plain LM loads.

…and mixture sizes - 4-bit plus full finetuning no longer installs a bitsandbytes config - revision, local_files_only and subfolder reach the plain LM load (CPU too) - a renamed question never overwrites another; one-level scores get no threshold - canonical_row copies before subsampling; the mixture returns n_rows exactly - the training script sizes sources after unavailable ones drop out, decontaminates typed-decisions, and runs on CPU-only torch

danielhanchen · 2026-10-07T12:41:27Z

@codex review

danielhanchen · 2026-10-07T12:41:40Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 084b6e69c8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-10-07T12:51:24Z

+            dd.canonical_row(r) for r in dd.load_source(name, "test", limit = args.eval_rows, seed = 0)
+        ]
+    eval_texts = [
+        json.dumps(r["state"], ensure_ascii = False) for rows in evals.values() for r in rows


Canonicalize evaluation states before decontamination

When a typed-decisions state is a multi-key mapping with fewer than 13 tokens and its insertion order is not sorted, serializing it here preserves insertion order, whereas _words serializes the corresponding training mapping with sort_keys=True. Identical logical states therefore produce different fallback shingles and the train/test duplicate survives, contaminating the tuned metric. The earlier typed-decisions decontamination fix still misses this representation mismatch; pass the structured state directly or canonicalize both sides identically.

Useful? React with 👍 / 👎.

Fixed in #12949 (ccf40aa).

chatgpt-codex-connector · 2026-10-07T12:51:24Z

+    pools = {
+        name: dd.load_source(name, "train", limit = per_source, seed = args.seed) for name in pools
+    }


Decontaminate before truncating source pools

When decontamination removes even one of these per_source rows, build_decision_mixture fills the source's assigned share by sampling the truncated remainder with replacement, despite additional clean examples being available in the underlying dataset; converter-filtered rows cause the same problem. The unavailable-source fix still pre-truncates before this later filtering, introducing avoidable duplicates and potentially placing augmentations of the same raw row in both training and calibration holdout.

Useful? React with 👍 / 👎.

Fixed in #12949 (ccf40aa).

chatgpt-codex-connector · 2026-10-07T12:51:25Z

+    if not folder.is_dir():
+        from huggingface_hub import snapshot_download
+        folder = Path(
+            snapshot_download(str(head_init), token = token, allow_patterns = list(HEAD_FILES))


Forward Hub controls when loading head_init

When head_init is a Hub repository, this download ignores both the supplied revision and local_files_only=True, so an offline load can still contact the network and a warm-start head can come from the repository default rather than the requested revision. The earlier forwarding fix covers the backbone loads, but this separate head download remains uncontrolled; pass the same Hub controls into snapshot_download.

Useful? React with 👍 / 👎.

Fixed in #12949 (ccf40aa): local_files_only is forwarded. The backbone revision is not, since head_init can be a different repository.

chatgpt-codex-connector · 2026-10-07T12:51:25Z

+            from transformers import AutoModelForCausalLM as AutoClass
+            AutoProcessor = AutoTokenizer


Support causal backbones without a .model attribute

When the CPU path loads a valid causal architecture such as GPT2LMHeadModel, this generic AutoModelForCausalLM call and the output-embedding checks succeed, but the resulting decision model fails on its first forward because ClefDecisionModel.forward unconditionally accesses backbone.model; GPT-2, GPT-NeoX, BLOOM, and similar classes expose their transformer under other attributes. Either resolve the base model through the Transformers base-model interface or reject unsupported architectures during loading.

Useful? React with 👍 / 👎.

Fixed in #12876: forward falls back to backbone.base_model when there is no .model.

chatgpt-codex-connector · 2026-10-07T12:51:25Z

+            share = max(
+                1,


Exclude zero-weight sources from the mixture

When callers use a zero weight to disable a source, max(1, ...) still loads at least one row and the later allocation also grants that source at least one row. This can make a supposedly disabled gated source such as xLAM fail the entire build, or silently include its examples in the mixture; if every weight is zero, the denominator also becomes zero. Skip zero-weight entries and reject negative or all-zero weight maps before loading.

Useful? React with 👍 / 👎.

Fixed in #12949 (ccf40aa).

chatgpt-codex-connector · 2026-10-07T12:51:25Z

+    out = t.train()
+    steps = t.state.global_step
+    losses = [h["loss"] for h in t.state.log_history if "loss" in h]
+    calibration = FastDecisionModel.calibrate(model, processor, holdout)


Guard calibration when the holdout is empty

When a small requested mixture or heavy filtering leaves fewer decisions than the holdout target can accommodate, split_holdout rounds its target down and may return an empty holdout because it only keeps whole rows. This unconditional calibration then evaluates empty lists, producing NaN accuracy, ECE, and loss (and a non-standard JSON result) while saving an uncalibrated model; either ensure at least one row is held out or skip calibration with an explicit result.

Useful? React with 👍 / 👎.

Fixed in #12949 (ccf40aa).

danielhanchen · 2026-10-07T13:18:04Z

Clef head trained from a plain LM with scripts/train_decision_from_lm.py (B200, LoRA r=64, 1 epoch unless noted). Test sets decontaminated from training. Accuracy is base -> tuned.

Model	Training data	typed-decisions	BANKING77	CLINC150	Holdout acc / ECE	Time	Peak mem
Qwen3.5-0.8B	13.2k-row mix	36% -> 73%	7% -> 74%	19% -> 76%	78% / 3%	24.7 min	4.3 GB
Qwen3.5-2B	13.2k-row mix	33% -> 76%	1% -> 77%	1% -> 79%	83% / 5%	22.9 min	7.8 GB
Qwen3.5-2B	mix + 150 head warmup steps	33% -> 75%	1% -> 76%	1% -> 78%	83% / 4%	25.9 min	7.8 GB
Qwen3.5-2B	typed-decisions only, 3 epochs	33% -> 78%	1% -> 58%	1% -> 62%	81% / 19%	9.6 min	7.6 GB

Mix = 12 sources (ag_news, arc, banking77, boolq, clinc150, commonsense_qa, mmlu, mnli, prompt_injections, snli, sst5, wanli) + typed-decisions. Test rows: 2000 typed-decisions, 500 BANKING77, 500 CLINC150.

…, unslothai#12939 (unslothai#12949) * Decision models: follow-up fixes for mixture weights, offline head download, GGUF-only no-torch serving and GPU order * Keep Vulkan device order and reject non-finite mixture weights * Follow HIP visibility order for ROCm and keep the Clef fallback reason until Laya serves * Report the fallback reason only for a resident Clef, and describe the Qwen3.5 precision path without a weight dtype * Resolve UUID masks when choosing a GPU and detect GPUs through llama.cpp on no-torch installs * Let ROCm fall back to CUDA_VISIBLE_DEVICES when choosing a GPU * Skip the shebang llama-server stub tests on Windows * Ask llama-server for its GPUs on no-torch installs, cached, and fix the idle unload test race * Key the no-torch GPU probe cache by llama-server path and mtime * Use a standalone ROCR_VISIBLE_DEVICES mask as the ROCm device order

NilayYadav and others added 30 commits October 3, 2026 04:15

Studio: fine-tune Laya decision models and serve them

8fc2e67

Check decision checkpoints at their subfolder, refuse DoRA/LoftQ, ste…

6fad6e5

…ady the toy test

Turn gradient checkpointing back on in FastDecisionModel.for_training

f7c5a51

Merge decision LoRA on a CPU copy, check decision subfolders before t…

95ed4bf

…he repo root

Log decision runs to TensorBoard when it is enabled

79527a5

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

1df8988

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

Block a decision run whose model config failed to load, gate the GPU …

9e4b676

…memory test on a real CUDA device

Use the AGPL-3.0 SPDX header on the new decision model files

8461ac0

Merge main into fine-tune-laya-decision-models-v2

2a60d6a

Record the vendored laya import in the risky loader baseline

c5cb7cb

Train Cloudflare Clef decision models with FastDecisionModel

1570c18

Install the Qwen3.5 fast-path kernels for Clef decision runs, test th…

3464797

…ey engage

Keep Clef off fp16 autocast and fp16 weights

28658e5

Vendor Cloudflare's Clef joint_schema_model.py unmodified

6189a56

Drop the decision trainer's backward rescale: accelerate never scales…

2280117

… the loss here

Run the decision toy task on CPU, where its dropout stream is fixed

dfedea2

Fold Clef calibration into the saved head; add label smoothing and Br…

81d1db4

…ier loss options

Build Clef's encoding and head on the vendored joint_schema_model.py

ce25b7a

Check Clef reloads against the folded temperature; skip Clef forwards…

5771e27

… on CPU when causal_conv1d is installed

Batch the Clef head, pool its fp32 norm in chunks, checkpoint and com…

7bf566e

…pile it

Record QLoRA in Clef run metadata, stub decision_layout in the export…

9444be7

… route test

Decision runs: hold out from training when eval files are training fi…

c346bf1

…les, honour a cancel during calibration

Merge branch 'clef-support' into clef-studio

335027c

# Conflicts: # studio/backend/core/training/worker.py

Serve Clef with the trainer's temperature rule, incl. option-count bu…

3b23ffd

…ckets

Fetch a decision checkpoint by its layout so Unsloth's download check…

46b6206

… accepts it

Save decision models crash-safely and fold Clef's calibrated temperat…

6f95bcd

…ure into its head

Load Clef in Unsloth's dynamic 4-bit and train it in float16 and bflo…

0288172

…at16

Add Cloudflare's Clef recipe options and record accuracy, drop the no…

26c0b46

…-op accumulation wrapper

Run the toy learning test on CPU, where its step budget was tuned

ef20dbc

danielhanchen added 5 commits October 6, 2026 05:25

Bound the Clef reload check by the logit scale, which varies run to r…

93a0eea

…un, and give its asserts messages

Merge remote-tracking branch 'refs/remotes/local/r12585' into decisio…

169d43d

…n-from-lm

Test that a plain LM decision model with float32 norms trains without…

988feee

… a precision flag

Bound the plain LM reload check by its measured spread across runs

89d5683

Score the plain LM reload check under the autocast serving uses

2334202

This was referenced Oct 6, 2026

Clef: train under the autocast it evaluates in, and bound the reload check by logit scale NilayYadav/unsloth#443

Open

Train any text or vision LLM as a Clef decision model from Studio, with FastDecisionModel.predict and adapter saves #12876

Merged

danielhanchen added the conflicts# label Oct 7, 2026

danielhanchen added 2 commits October 7, 2026 03:26

Merge remote-tracking branch 'origin/main' into decision-from-lm

dabcde6

Trim comments in the decision-from-LM code and tests

8e4eb8c

unslothai deleted a comment from chatgpt-codex-connector Bot Oct 7, 2026

shimmyshimmer mentioned this pull request Oct 7, 2026

[staging CI] unslothai/unsloth#12772 shimmyshimmer/unsloth-staging-4#1132

Closed

danielhanchen added non-studio# and removed >3000# affects UI# conflicts# labels Oct 7, 2026

chatgpt-codex-connector Bot reviewed Oct 7, 2026

View reviewed changes

danielhanchen merged commit 6cbf311 into main Oct 7, 2026

danielhanchen deleted the decision-models-from-lm branch October 7, 2026 12:50

chatgpt-codex-connector Bot reviewed Oct 7, 2026

View reviewed changes

danielhanchen mentioned this pull request Oct 7, 2026

Decision models: follow-up fixes after #12772, #12876, #12939 #12949

Merged

Lyxot mentioned this pull request Oct 7, 2026

feat(mlx): train, score and save decision models on Apple Silicon unslothai/unsloth-zoo#1615

Merged

danielhanchen mentioned this pull request Oct 8, 2026

feat(mlx): train, score and save decision models on Apple Silicon danielhanchen/unsloth-zoo-staging-review#131

Closed

		if load_in_4bit and kwargs.get("quantization_config") is None:
		kwargs["quantization_config"] = _clef_bnb_config(dtype)

		from transformers import AutoModelForCausalLM as AutoClass
		AutoProcessor = AutoTokenizer

Conversation

danielhanchen commented Oct 5, 2026

Usage

What it adds

Results

Tests

Uh oh!

danielhanchen commented Oct 7, 2026

Uh oh!

danielhanchen commented Oct 7, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

danielhanchen commented Oct 7, 2026

Uh oh!

danielhanchen commented Oct 7, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Oct 7, 2026

Choose a reason for hiding this comment

danielhanchen commented Oct 7, 2026 •

edited

Loading