Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Clear the duplicate-definition backlog and gate it at zero (#9769)
Rajarshi Dattab349936f09fa
diff --git a/.github/workflows/lint-ci.yml b/.github/workflows/lint-ci.ymlindex 269886ee65f..6c8886aab09 100644--- a/.github/workflows/lint-ci.yml+++ b/.github/workflows/lint-ci.yml@@ -292,11 +292,8 @@ jobs:# duplicate that was already in the file is printed and ignored while one# this PR introduced fails. Comparing added LINES instead would miss the# case where the new copy is inserted ABOVE the old one, since the finding- # is reported at the second, unchanged definition. There are 37 pre-existing- # findings on main today; they are a separate cleanup, not a reason to leave- # the gate off. Four of those are only visible once each control-flow branch is- # scanned on its own: a repeated `from unsloth import is_bfloat16_supported`- # inside one `if __name__ == "__main__":` block in four of the saving tests.+ # is reported at the second, unchanged definition. The whole-tree step below+ # catches pre-existing ones.if: github.event_name == 'pull_request'env:GH_TOKEN: ${{ github.token }}@@ -338,6 +335,13 @@ jobs:python scripts/lint_duplicate_definitions.py \--before "$MERGE_BASE" --after "$HEAD_SHA" "${CHANGED[@]}"+ - name: Duplicate definitions (whole tree)+ # The diff-scoped step passes pre-existing findings, so the backlog regrew+ # to 38 while it was the only gate. Runs on push too.+ run: |+ python scripts/lint_duplicate_definitions.py \+ unsloth unsloth_cli studio scripts tests cli.py unsloth-cli.py+- name: No silent llama-server parallel-slot downgrades# #7717 clamped --parallel to 1 whenever MTP resolved, costing batched# callers 4x throughput. Launch fewer slots only for a real capabilitydiff --git a/studio/backend/core/inference/tools.py b/studio/backend/core/inference/tools.pyindex 3a14f6edf2e..e7847441e52 100644--- a/studio/backend/core/inference/tools.py+++ b/studio/backend/core/inference/tools.py@@ -5204,8 +5204,6 @@ _MAX_PATH_SCAN_CHARS = 2048_MAX_TERMINAL_SCAN_CHARS = 4096# Each pass doubles resolved alias hops; leftover work after the cap fails closed._MAX_SHELL_ASSIGN_EXPAND_PASSES = 16-# A glob needs one of these to expand into anything but itself; used to skip the glob scans outright.-_GLOB_META_RE = re.compile(r"[?*\[]")# Where the memoised node list is parked on a parsed tree (see _tree_nodes)._TREE_NODES_ATTR = "_unsloth_walk_nodes"@@ -7448,9 +7446,6 @@ _OPENSSL_NETWORK_RE = re.compile(# -c/eval it runs an unscreened payload. Paired with the var-executed-as-command test so `echo "${a[@]}"` is left# alone._ARRAY_EXPANSION_RE = re.compile(r"\$\{\w+\[[@*]\]\}")-# A wrapper's bare duration/count argument (timeout 5 rm, timeout 1.5s rm) that precedes the real command, so it is-# not mistaken for the command itself.-_WRAPPER_DURATION_RE = re.compile(r"\d+(?:\.\d+)?[smhd]?$")# Non-shell interpreters running an inline program (python -c, node -e, php -r): the terminal path never screens that# program the way the python tool does. sh/bash -c are omitted, the hard-block already recursing into their payloads._INLINE_CODE_INTERPRETERS = frozenset(diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.pyindex 62f55d0a481..34586d1916a 100644--- a/studio/backend/tests/test_diffusion_backend.py+++ b/studio/backend/tests/test_diffusion_backend.py@@ -2613,7 +2613,7 @@ def test_load_unknown_family_raises():# load_progress state machine (no threads / network / real cache)-from core.inference.diffusion import _LoadingState, _LoadState # noqa: E402+from core.inference.diffusion import _LoadingState # noqa: E402def test_load_progress_idle_and_ready():diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.pyindex 3df8c045aab..075b4b3128e 100644--- a/studio/backend/tests/test_openai_tool_passthrough.py+++ b/studio/backend/tests/test_openai_tool_passthrough.py@@ -3794,12 +3794,6 @@ class TestFriendlyErrorHttpx:assert _friendly_error(RuntimeError("unrelated")) == "An internal error occurred"-from routes.inference import ( # noqa: E402- _drop_empty_assistant_sentinels,- _openai_messages_for_gguf_chat,-)--def _image_question_messages():"""One user turn holding a question and a tiny png."""return [diff --git a/tests/kaggle/test_launch_cleanup.py b/tests/kaggle/test_launch_cleanup.pyindex 33568b11d36..6994e5c8055 100644--- a/tests/kaggle/test_launch_cleanup.py+++ b/tests/kaggle/test_launch_cleanup.py@@ -894,24 +894,6 @@ def test_the_leaked_kernel_warning_does_not_strand_the_handler(tmp_path):assert _deletions(tmp_path), "no delete was attempted, so nothing could leak"-def _failing_kaggle(bin_dir: Path, record: Path) -> None:- """A `kaggle` that records the call and then refuses it, so nothing is ever- released and release() reaches its leaked-kernel warning."""- bin_dir.mkdir(parents = True, exist_ok = True)- shim = bin_dir / "kaggle"- shim.write_text(- textwrap.dedent(f"""\- #!{sys.executable}- import sys, pathlib- pathlib.Path({str(record)!r}).open("a").write(" ".join(sys.argv[1:]) + "\\n")- sys.stderr.write("500 Server Error\\n")- sys.exit(1)- """),- encoding = "utf-8",- )- shim.chmod(0o755)--def test_the_leaked_warning_is_still_a_github_annotation(tmp_path, monkeypatch, capsys):"""Routing it through the safe writer must not add the [launch] prefix.diff --git a/tests/python/test_compiler_watch_temp_scan.py b/tests/python/test_compiler_watch_temp_scan.pyindex c7424c8eaf1..6e42c994018 100644--- a/tests/python/test_compiler_watch_temp_scan.py+++ b/tests/python/test_compiler_watch_temp_scan.py@@ -33,8 +33,6 @@ import subprocessimport pytestfrom unsloth_pwsh_runner import run_pwsh-from unsloth_pwsh_runner import run_pwsh-REPO = pathlib.Path(__file__).resolve().parents[2]SCRIPT = REPO / ".github" / "scripts" / "Watch-ForCompiler.ps1"diff --git a/tests/saving/language_models/test_merge_model_perplexity_llama_3_2.py b/tests/saving/language_models/test_merge_model_perplexity_llama_3_2.pyindex 9500141f4a3..47a48641cb7 100644--- a/tests/saving/language_models/test_merge_model_perplexity_llama_3_2.py+++ b/tests/saving/language_models/test_merge_model_perplexity_llama_3_2.py@@ -159,8 +159,6 @@ if __name__ == "__main__":loftq_config = None,)- from unsloth import is_bfloat16_supported-trainer = SFTTrainer(model = model,tokenizer = tokenizer,diff --git a/tests/saving/language_models/test_merge_model_perplexity_mistral.py b/tests/saving/language_models/test_merge_model_perplexity_mistral.pyindex d8a6c7e4260..72ff725b6db 100644--- a/tests/saving/language_models/test_merge_model_perplexity_mistral.py+++ b/tests/saving/language_models/test_merge_model_perplexity_mistral.py@@ -217,8 +217,6 @@ if __name__ == "__main__":loftq_config = None,)- from unsloth import is_bfloat16_supported-trainer = SFTTrainer(model = model,tokenizer = tokenizer,diff --git a/tests/saving/language_models/test_merge_model_perplexity_phi_4.py b/tests/saving/language_models/test_merge_model_perplexity_phi_4.pyindex 26a99f46ec4..778974e41d1 100644--- a/tests/saving/language_models/test_merge_model_perplexity_phi_4.py+++ b/tests/saving/language_models/test_merge_model_perplexity_phi_4.py@@ -159,8 +159,6 @@ if __name__ == "__main__":loftq_config = None,)- from unsloth import is_bfloat16_supported-trainer = SFTTrainer(model = model,tokenizer = tokenizer,diff --git a/tests/saving/language_models/test_merged_model_perplexity_llama_3_1_8b.py b/tests/saving/language_models/test_merged_model_perplexity_llama_3_1_8b.pyindex a3474444404..9d8f7763ffa 100644--- a/tests/saving/language_models/test_merged_model_perplexity_llama_3_1_8b.py+++ b/tests/saving/language_models/test_merged_model_perplexity_llama_3_1_8b.py@@ -161,8 +161,6 @@ if __name__ == "__main__":loftq_config = None,)- from unsloth import is_bfloat16_supported-trainer = SFTTrainer(model = model,tokenizer = tokenizer,diff --git a/tests/saving/text_to_speech_models/test_csm.py b/tests/saving/text_to_speech_models/test_csm.pyindex 146346450a3..15e16c4afc8 100644--- a/tests/saving/text_to_speech_models/test_csm.py+++ b/tests/saving/text_to_speech_models/test_csm.py@@ -141,7 +141,6 @@ print(f"{'='*80}")from transformers import pipeline-import torchoutput_audio_path = "csm_audio.wav"try:diff --git a/tests/saving/text_to_speech_models/test_lasa.py b/tests/saving/text_to_speech_models/test_lasa.pyindex e1df3a2a469..118fbfef101 100644--- a/tests/saving/text_to_speech_models/test_lasa.py+++ b/tests/saving/text_to_speech_models/test_lasa.py@@ -143,7 +143,6 @@ print(f"{'='*80}")from transformers import pipeline-import torchoutput_audio_path = "lasa_audio.wav"input_text = "Hey there my name is Elise, <giggles> and I'm a speech generation model that can sound like a person."diff --git a/tests/saving/vision_models/test_save_merge_qwen2_5vl32B_model_ocr_benchmark.py b/tests/saving/vision_models/test_save_merge_qwen2_5vl32B_model_ocr_benchmark.pyindex 79e9c9cc897..f6458a1999d 100644--- a/tests/saving/vision_models/test_save_merge_qwen2_5vl32B_model_ocr_benchmark.py+++ b/tests/saving/vision_models/test_save_merge_qwen2_5vl32B_model_ocr_benchmark.py@@ -31,8 +31,6 @@ from tests.utils.cleanup_utils import safe_remove_directoryfrom tests.utils.ocr_eval import OCRModelEvaluator-from datasets import load_dataset-dataset = load_dataset("lbourdois/OCR-liboaccn-OPUS-MIT-5M-clean", "en", split = "train")train_dataset = dataset.select(range(2000))eval_dataset = dataset.select(range(2000, 2200))@@ -71,12 +69,9 @@ system_message = "You are an expert french ocr system."train_dataset = [format_data(sample) for sample in train_dataset]eval_dataset = [format_data(sample) for sample in eval_dataset]-import os-import torchfrom tqdm import tqdmimport pandas as pdfrom jiwer import wer, cer-from qwen_vl_utils import process_vision_infoocr_evaluator = OCRModelEvaluator()model_comparison_results = {}diff --git a/tests/saving/vision_models/test_save_merge_vision_model_ocr_benchmark.py b/tests/saving/vision_models/test_save_merge_vision_model_ocr_benchmark.pyindex d905a3df997..da7c455fa3c 100644--- a/tests/saving/vision_models/test_save_merge_vision_model_ocr_benchmark.py+++ b/tests/saving/vision_models/test_save_merge_vision_model_ocr_benchmark.py@@ -31,8 +31,6 @@ from tests.utils.cleanup_utils import safe_remove_directoryfrom tests.utils.ocr_eval import OCRModelEvaluator-from datasets import load_dataset-dataset = load_dataset("lbourdois/OCR-liboaccn-OPUS-MIT-5M-clean", "en", split = "train")train_dataset = dataset.select(range(2000))@@ -72,12 +70,9 @@ system_message = "You are an expert french ocr system."train_dataset = [format_data(sample) for sample in train_dataset]eval_dataset = [format_data(sample) for sample in eval_dataset]-import os-import torchfrom tqdm import tqdmimport pandas as pdfrom jiwer import wer, cer-from qwen_vl_utils import process_vision_infoocr_evaluator = OCRModelEvaluator()model_comparison_results = {}diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.pyindex 05bed762418..dc5b76fc255 100644--- a/tests/studio/install/test_rocm_support.py+++ b/tests/studio/install/test_rocm_support.py@@ -7447,8 +7447,6 @@ class TestHipSdkInstalledButDeviceInaccessible:# TEST: --rocm-gfx forwarding -- setup.sh/setup.ps1 forward their resolved gfx# arch to install_llama_prebuilt.py so the per-gfx prebuilt is picked.-_SETUP_SH_PATH = PACKAGE_ROOT / "studio" / "setup.sh"-class TestNormalizeForwardedGfx:"""A forwarded gfx string is reduced to a single clean gfx token."""@@ -7798,7 +7796,6 @@ def test_pick_rocm_gfx_target_same_arch_multi_gpu(monkeypatch):# TEST: WSL ROCDXG fixes -- drop-in persistence + system-HIP-before-bundle-_INSTALL_SH_PATH = PACKAGE_ROOT / "install.sh"_LLAMA_CPP_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "inference" / "llama_cpp.py"diff --git a/tests/studio/studiobench/instruments/coverage.py b/tests/studio/studiobench/instruments/coverage.pyindex 1eea1d12366..5df233aceab 100644--- a/tests/studio/studiobench/instruments/coverage.py+++ b/tests/studio/studiobench/instruments/coverage.py@@ -296,7 +296,6 @@ def ambiguity(snapshot: CoverageSnapshot, name: str) -> int:# boundary, enforced by `assert_integers_only`.import time # noqa: E402-from typing import Any # noqa: E402from ..analysis import assert_no_bare_zero, measured, merge, unmeasured # noqa: E402from . import register_instrument # noqa: E402diff --git a/tests/test_studio_install_workspace_guard.py b/tests/test_studio_install_workspace_guard.pyindex 48243845630..e89a114ef8c 100644--- a/tests/test_studio_install_workspace_guard.py+++ b/tests/test_studio_install_workspace_guard.py@@ -12,8 +12,6 @@ from pathlib import Pathimport pytestfrom unsloth_pwsh_runner import run_pwsh-from unsloth_pwsh_runner import run_pwsh-REPO_ROOT = Path(__file__).resolve().parents[1]INSTALL_SH = REPO_ROOT / "install.sh"INSTALL_PS1 = REPO_ROOT / "install.ps1"diff --git a/unsloth/dataprep/synthetic.py b/unsloth/dataprep/synthetic.pyindex ffb2e9a6229..96f713bb6ef 100644--- a/unsloth/dataprep/synthetic.py+++ b/unsloth/dataprep/synthetic.py@@ -44,7 +44,6 @@ if (import requestsimport torchimport gc-import timeimport refrom unsloth_zoo.log import loggerimport numpy as npdiff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.pyindex 7dde169e53b..433b3dcf129 100644--- a/unsloth/kernels/utils.py+++ b/unsloth/kernels/utils.py@@ -30,7 +30,6 @@ from ..device_type import (from ..bnb_availability import native_kernels_readyfrom .fp8 import weight_dequant, fp8_linear, can_use_fp8_rowwise_gemv, fp8_rowwise_gemvfrom .nvfp4 import NVFP4QuantState, nvfp4_dequantize, nvfp4_linear-import functools# torch.cuda.amp.custom_fwd is deprecated from 2.4.import torch@@ -65,7 +64,6 @@ elif DEVICE_TYPE == "npu":# tl.math.tanh is now libdevice.tanh.-import tritonimport triton.language as tlif Version(triton.__version__) >= Version("3.0.0"):diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.pyindex 1c703c5c4f0..a48f8a48f98 100644--- a/unsloth/models/_utils.py+++ b/unsloth/models/_utils.py@@ -3866,9 +3866,6 @@ def has_internet(return False-import psutil--def _get_statistics(statistics = None, force_download = True):# Basic stats on which environment is in use: a README.md is downloaded from HF, all data public, so broken envs can be detected. Disable with UNSLOTH_DISABLE_STATISTICS.n_cpus = psutil.cpu_count(logical = False)diff --git a/unsloth/models/llama.py b/unsloth/models/llama.pyindex c121af9bd9a..07c24bfaa1c 100644--- a/unsloth/models/llama.py+++ b/unsloth/models/llama.py@@ -169,7 +169,7 @@ def patch_saving_functions(*args, **kwargs):patch_saving_functions._unsloth_deferred_shim = True-import re, os, inspect, math, sys+import re, os, inspect, sysimport typestry:diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.pyindex b4609619831..a80a34bdea7 100644--- a/unsloth/models/mistral.py+++ b/unsloth/models/mistral.py@@ -55,7 +55,7 @@ try:except:MistralSdpaAttention = MistralAttentionMistralFlashAttention2 = MistralAttention-from unsloth_zoo.utils import Version, _get_dtype+from unsloth_zoo.utils import Versiondef MistralAttention_fast_forward(diff --git a/unsloth/models/vision.py b/unsloth/models/vision.pyindex a6cfffce87d..1f3e9a772fd 100644--- a/unsloth/models/vision.py+++ b/unsloth/models/vision.py@@ -201,7 +201,6 @@ from unsloth_zoo.patching_utils import patch_model_and_tokenizerfrom unsloth_zoo.training_utils import prepare_model_for_trainingfrom unsloth_zoo.utils import Version-from transformers import __version__ as transformers_versionimport typesimport functoolsdiff --git a/unsloth/save.py b/unsloth/save.pyindex 97367224498..4e656255344 100644--- a/unsloth/save.py+++ b/unsloth/save.py@@ -60,7 +60,6 @@ import subprocessimport tracebackimport psutilimport re-from transformers.models.llama.modeling_llama import loggerfrom .models.loader_utils import (get_model_name,_resolve_hub_repo_cached_file,@@ -5155,16 +5154,8 @@ def unsloth_convert_lora_to_ggml_and_save_locally(return _unsloth_save_lora_gguf(self, tokenizer, save_directory, outtype = outtype)-from .models.loader_utils import (- get_model_name,- _resolve_hub_repo_cached_file,- _tokenizer_cache_dir,- _tokenizer_wants_local_only,-)-# Imported lazily at the two call sites: an older zoo, before its bitsandbytes import became optional, would otherwise break `import unsloth` on a host without bnb.from unsloth_zoo.llama_cpp import (- install_llama_cpp,convert_to_gguf as _convert_to_gguf,)