Skip to content

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.

Remove unreferenced private helpers, constants and an if False block (#16569)

Daniel Hanc093e0278996
diff --git a/studio/backend/core/inference/diffusion_engine_router.py b/studio/backend/core/inference/diffusion_engine_router.py
index ea302d77b0c..d0ca6530f88 100644
--- a/studio/backend/core/inference/diffusion_engine_router.py
+++ b/studio/backend/core/inference/diffusion_engine_router.py
@@ -420,10 +420,6 @@ def mlx_mode() -> str:
return MLX_OFF
-def _mlx_requested() -> bool:
- return mlx_mode() != MLX_OFF
-
-
def _mlx_host_refusal() -> Optional[str]:
if sys.platform != "darwin":
return "the MLX engine runs on macOS only"
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 99c19386711..a178dddda5b 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -1422,17 +1422,6 @@ def _has_unclosed_code_fence(text: str) -> bool:
return active_char is not None
-def _has_unclosed_markup_block(text: str) -> bool:
- """Counted rather than matched, so a closed block followed by an open one still counts as unclosed."""
- opens_html = len(re.findall(r"<html\b", text, re.IGNORECASE))
- closes_html = len(re.findall(r"</html\s*>", text, re.IGNORECASE))
- if opens_html > closes_html:
- return True
- opens_svg = len(re.findall(r"<svg\b", text, re.IGNORECASE))
- closes_svg = len(re.findall(r"</svg\s*>", text, re.IGNORECASE))
- return opens_svg > closes_svg
-
-
# "First, I'll create an <html></html> skeleton" is a plan, not a page.
_EMPTY_MARKUP_SKELETON = re.compile(
r"<(html|svg)\b[^>]*>\s*</\1\s*>",
diff --git a/studio/backend/core/inference/mlx_diffusion_convert.py b/studio/backend/core/inference/mlx_diffusion_convert.py
index 317938671b7..8188e8a1d15 100644
--- a/studio/backend/core/inference/mlx_diffusion_convert.py
+++ b/studio/backend/core/inference/mlx_diffusion_convert.py
@@ -67,7 +67,6 @@ _BLOCK_BITS = {
"Q6_K": 8,
}
_QK_K = 256
-_DENSE_TYPES = frozenset({"BF16", "F16", "F32"})
_PAD_TOKENS = ("cap_pad_token", "x_pad_token")
_QKV_PARTS = ("to_q", "to_k", "to_v")
diff --git a/studio/backend/core/inference/sd_cpp_backend.py b/studio/backend/core/inference/sd_cpp_backend.py
index 838a9755e5e..a8c7f3240ec 100644
--- a/studio/backend/core/inference/sd_cpp_backend.py
+++ b/studio/backend/core/inference/sd_cpp_backend.py
@@ -1635,12 +1635,6 @@ _RUNTIME_FINGERPRINT_MEMO: Optional[dict] = None
_HOST_FINGERPRINT_MEMO: Optional[dict] = None
-def _reset_host_fingerprint() -> None:
- global _HOST_FINGERPRINT_MEMO, _RUNTIME_FINGERPRINT_MEMO
- _HOST_FINGERPRINT_MEMO = None
- _RUNTIME_FINGERPRINT_MEMO = None
-
-
def _runtime_fingerprint() -> dict:
global _RUNTIME_FINGERPRINT_MEMO
if _RUNTIME_FINGERPRINT_MEMO is not None:
diff --git a/studio/backend/core/inference/tools.py b/studio/backend/core/inference/tools.py
index fa8be61e064..330a1c42a13 100644
--- a/studio/backend/core/inference/tools.py
+++ b/studio/backend/core/inference/tools.py
@@ -2131,11 +2131,6 @@ _AUTO_EXEC_MCP_VERB_ONLY_RE = re.compile(
r"powershell|pwsh|terminal|subprocess|interpreter)(?:[_\-]|$)",
re.IGNORECASE,
)
-_AUTO_EXEC_MCP_RUNTIME_NOUN_RE = re.compile(
- r"(?:^|[_\-])(?:python[0-9.]*|node|nodejs|deno|bun|ruby|perl|php|code|"
- r"script|repl|sandbox|notebook)(?:[_\-]|$)",
- re.IGNORECASE,
-)
# Code/command runners run on the MCP server outside the sandbox; whole segments only.
_AUTO_EXEC_MCP_TOOL_RE = re.compile(
r"(?:^|[_\-])(?:"
@@ -3302,20 +3297,6 @@ def _references_studio_credential(text: str) -> bool:
# Any run of text that could be a path argument, used to read one token at a time.
_PATH_TOKEN_RE = re.compile(r"[^\s'\"()\[\]{},;|&<>]+")
-# Only traversal is worth resolving: a relative path without `..` stays inside the sandbox.
-_TRAVERSAL_TOKEN_RE = re.compile(r"[^\s'\"()\[\]{},;|&<>]*\.\.[^\s'\"()\[\]{},;|&<>]*")
-# Used only after a `cd`, where `auth/auth.db` stops meaning "inside the sandbox".
-_RELATIVE_PATH_TOKEN_RE = re.compile(r"[^\s'\"()\[\]{},;|&<>]*[/\\][^\s'\"()\[\]{},;|&<>]*")
-# Case-insensitive: the shell is `cmd /c` on Windows without a trusted bash.
-_CD_TARGET_RE = re.compile(
- # `builtin cd`, `command cd`, `! cd` and `time cd` move like a bare `cd`.
- r"(?:^\s*|[;&|({\n]\s*|\b(?:then|do|else|if|elif|while|until)\s+)"
- # Bash allows assignment words before a special builtin, so `X=1 cd ../..` moves like a bare `cd`.
- r"(?:(?:builtin|command|exec|nohup)\s+|time\s+(?:-p\s+)?|!\s*|[A-Za-z_]\w*=[^\s;&|()]*\s+)*"
- r"(?:cd|pushd)\s+"
- r"(?:(?:-[LPe@]+|--|/d)\s+)*([^\s;&|)]+)",
- re.IGNORECASE,
-)
# `cd -` and `popd` return; matched in one pass so order is kept.
_DIRECTORY_MOVE_RE = re.compile(
r"(?:^\s*|[;&|({\n]\s*|\b(?:then|do|else|if|elif|while|until)\s+)"
@@ -6609,8 +6590,6 @@ _HIGH_RISK_FORWARDING_COMMANDS = frozenset(
"valgrind",
}
)
-# Tracers and profilers run the rest of the line as a child.
-_TRACER_LAUNCHERS = frozenset({"strace", "ltrace", "ktrace", "dtruss", "perf", "valgrind"})
_EXEC_FLAG_FORWARDING_COMMANDS = frozenset({"find", "fd"})
_EXEC_FORWARD_FLAGS = frozenset(
{"-exec", "-execdir", "-ok", "-okdir", "--exec", "--exec-batch", "-x", "-X"}
@@ -6718,7 +6697,6 @@ _INLINE_CODE_INTERPRETERS = frozenset(
"php",
}
)
-_INLINE_CODE_FLAGS = frozenset({"-c", "-e", "-E", "-r", "--eval", "--exec"})
# Flags are per interpreter (`python -E` is not eval). Value: (exact flags, cluster letters).
_INLINE_CODE_FLAG_SPEC = {
"python": (frozenset({"-c"}), "c"),
@@ -10067,15 +10045,6 @@ def live_project_owns(
return False
-def _project_exists(project_id: str) -> bool:
- """Whether a project of the user's is stored under this exact id."""
- try:
- from storage.studio_db import get_chat_project
- return get_chat_project(project_id) is not None
- except Exception: # noqa: BLE001 - a storage hiccup must not delete files
- return True
-
-
def _is_managed_project_workdir(session_id: "str | None", workdir: str) -> bool:
"""Compares with the sandbox path, not the working-folder setting, which can change before the chmod."""
if not session_id or not session_id.startswith(_PROJECT_SESSION_PREFIX):
diff --git a/studio/backend/core/inference/video_minimax_h3_ref_encode.py b/studio/backend/core/inference/video_minimax_h3_ref_encode.py
index 8099b2f555c..393ba91233b 100644
--- a/studio/backend/core/inference/video_minimax_h3_ref_encode.py
+++ b/studio/backend/core/inference/video_minimax_h3_ref_encode.py
@@ -13,7 +13,6 @@ import threading
from typing import Any, Optional
H3_CHUNKED_REF_ENCODE_ENV = "UNSLOTH_H3_CHUNKED_REF_ENCODE"
-_ENCODERS_MODULE = "diffusers.modular_pipelines.minimax_h3.encoders"
_FUNC = "encode_vae_condition"
# sha256[:16] of inspect.getsource of the mirrored functions (diffusers 0.42.1).
diff --git a/studio/backend/core/rag/office_formats.py b/studio/backend/core/rag/office_formats.py
index 79810f164d7..f866fa4933c 100644
--- a/studio/backend/core/rag/office_formats.py
+++ b/studio/backend/core/rag/office_formats.py
@@ -406,10 +406,6 @@ def _apply_section(value: float, plan) -> str | None:
return prefix + text + percent + suffix
-def _format_section(value: float, tokens: list[tuple[str, str]]) -> str | None:
- return _apply_section(value, _section_plan(tokens))
-
-
def _number_plan(code: str):
"""Each section of a numeric format read once (with whether it has tokens), or None if unsupported."""
tokens = _format_tokens(code) if code else None
@@ -435,11 +431,6 @@ def _apply_number(value: float, plan) -> str | None:
return None if text is None else ("-" + text if value < 0 else text)
-def _format_number(value: float, code: str) -> str | None:
- """``value`` as a numeric format displays it, or None for formats left as stored."""
- return _apply_number(value, _number_plan(code))
-
-
_NUMBER_ERRORS = (ArithmeticError, ValueError, decimal.InvalidOperation)
diff --git a/studio/backend/core/training/multi_lora_trainer.py b/studio/backend/core/training/multi_lora_trainer.py
index 1f66b3ca171..2022a9a2bad 100644
--- a/studio/backend/core/training/multi_lora_trainer.py
+++ b/studio/backend/core/training/multi_lora_trainer.py
@@ -23,7 +23,6 @@ from loggers import get_logger
logger = get_logger(__name__)
_DEFAULT_TARGETS = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
-_OPTIMIZERS = ("adamw_8bit", "adamw_torch", "adamw_torch_fused")
@dataclass
diff --git a/studio/backend/routes/auth.py b/studio/backend/routes/auth.py
index dc27ba2e40c..035c91e60a3 100644
--- a/studio/backend/routes/auth.py
+++ b/studio/backend/routes/auth.py
@@ -135,7 +135,6 @@ _LOGIN_BUCKETS_LOCK = threading.Lock()
_LOGIN_WINDOW_SECONDS = 60.0
_LOGIN_MAX_FAILS = 5
_LOGIN_IP_MAX_FAILS = 30
-_LOGIN_LOCKOUT_SECONDS = 60
# On overflow, reclaim expired buckets, then use sharded overflow rather than evict a hot bucket.
_LOGIN_MAX_BUCKETS = 4096
# Last full stale-sweep time; rate-limits the O(n) sweep under a burst of new IPs.
diff --git a/studio/backend/routes/data_recipe/jobs.py b/studio/backend/routes/data_recipe/jobs.py
index f8aa5f521e3..41e8ec35ce2 100644
--- a/studio/backend/routes/data_recipe/jobs.py
+++ b/studio/backend/routes/data_recipe/jobs.py
@@ -199,20 +199,6 @@ def _resolve_local_v1_endpoint(request: Request) -> str:
return f"http://{host}:{int(port)}/v1"
-def _request_has_desktop_access_token(request: Request) -> bool:
- auth_header = request.headers.get("authorization")
- if not auth_header:
- return False
-
- parts = auth_header.split(None, 1)
- if len(parts) != 2 or parts[0].lower() != "bearer":
- return False
-
- from auth.authentication import is_desktop_access_token
-
- return is_desktop_access_token(parts[1])
-
-
def _used_llm_model_aliases(recipe: dict[str, Any]) -> set[str]:
"""Return model_aliases actually referenced by an LLM column. Narrows the "Chat model loaded" gate
so orphan model_config nodes don't block unrelated runs. The ``llm-`` prefix matches
diff --git a/studio/backend/utils/hardware/gpu_query.py b/studio/backend/utils/hardware/gpu_query.py
index 1fa54e66c30..b18cb6167f5 100644
--- a/studio/backend/utils/hardware/gpu_query.py
+++ b/studio/backend/utils/hardware/gpu_query.py
@@ -134,10 +134,6 @@ def _query_fields(argv: Sequence[str]) -> list[str]:
return []
-def _nounits(argv: Sequence[str]) -> bool:
- return any(arg.startswith("--format=") and "nounits" in arg for arg in argv[1:])
-
-
@dataclass
class _Entry:
# None = the CLI's latest answer had no rows (failed or empty): never served, only orders writers.
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index a4be489cd7d..91743de6851 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -5168,25 +5168,6 @@ def _reconcile_cuda_integrated_memory(
)
-def _reconcile_primary_rocm_unified_memory(
- utilization: Dict[str, Any], parent_visible_spec: Dict[str, Any]
-) -> None:
- """Same fix as _reconcile_rocm_unified_memory for the flat primary-GPU dict."""
- numeric_ids = parent_visible_spec.get("numeric_ids")
- if numeric_ids is None:
- # No visibility env var set: torch ordinal 0 is the primary device.
- primary_idx = [0]
- elif len(numeric_ids) == 0:
- # Empty mask: no GPU visible. Querying torch device 0 would raise or return stale data, so bail rather than write bad values.
- return
- else:
- primary_idx = [int(numeric_ids[0])]
- torch_devices = _torch_get_per_device_info(primary_idx)
- if not torch_devices:
- return
- _apply_unified_memory_correction(utilization, torch_devices[0])
-
-
def _rocm_visibility_mask_active() -> bool:
"""True when any ROCm/CUDA visibility variable filters the device set."""
for var in (
diff --git a/studio/backend/utils/llama_cpp_freshness.py b/studio/backend/utils/llama_cpp_freshness.py
index 5f3133f1577..fd037d4478c 100644
--- a/studio/backend/utils/llama_cpp_freshness.py
+++ b/studio/backend/utils/llama_cpp_freshness.py
@@ -47,10 +47,6 @@ def read_install_marker(binary_path: Optional[str]) -> Optional[dict]:
)
-def _load_disk_cache(repo: str) -> Optional[tuple[float, Optional[str]]]:
- return _flow.load_disk_cache(repo, _cache_dir())
-
-
def _save_disk_cache(repo: str, latest_tag: Optional[str]) -> None:
_flow.save_disk_cache(
repo, latest_tag, _cache_dir(), log_message = "freshness cache write failed"
@@ -129,10 +125,6 @@ def update_download_size_bytes(
return None
-def _parse_installed_at(value: object) -> Optional[datetime]:
- return _flow.parse_installed_at(value)
-
-
def parse_base_build(tag: object) -> Optional[int]:
"""Numeric base build from a release tag. Handles both a plain ``bNNNN`` and a mix-build tag like ``b9596-mix-<sha>`` (anchored at the start, so the mix suffix does not defeat it). None for anything not starting with ``bNNNN``."""
if not isinstance(tag, str):
diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py
index d043bb11c8b..cc8e35f0dc8 100644
--- a/studio/backend/utils/llama_cpp_update.py
+++ b/studio/backend/utils/llama_cpp_update.py
@@ -61,7 +61,6 @@ class _LlamaPhaseError(RuntimeError):
self.reload_required = reload_required
-_JOB_IDLE = _flow.JOB_IDLE
_JOB_RUNNING = _flow.JOB_RUNNING
_JOB_SUCCESS = _flow.JOB_SUCCESS
_JOB_ERROR = _flow.JOB_ERROR
@@ -73,7 +72,6 @@ _job: dict = _flow.new_job()
_ALREADY_RUNNING_MESSAGE = "Another llama.cpp install is already running."
_utcnow = _flow.utcnow
-_is_under = _flow.is_under
_is_external_link = _flow.is_external_link
_rocm_install_args = _flow.rocm_install_args
diff --git a/studio/backend/utils/security/remote_code_scan.py b/studio/backend/utils/security/remote_code_scan.py
index 02991f53b95..e9b586e748e 100644
--- a/studio/backend/utils/security/remote_code_scan.py
+++ b/studio/backend/utils/security/remote_code_scan.py
@@ -673,11 +673,6 @@ def _auto_map_refs(cfg: dict) -> set:
return out
-def _auto_map_py(cfg: dict) -> set[str]:
- """Own-repo ``.py`` filenames referenced by auto_map (external refs excluded)."""
- return {fn for repo, fn in _auto_map_refs(cfg) if repo is None}
-
-
def external_auto_map_repos(
model_name: str,
hf_token: Optional[str] = None,
diff --git a/studio/backend/utils/whisper_cpp_freshness.py b/studio/backend/utils/whisper_cpp_freshness.py
index 6f417d4f9d3..124e633d05b 100644
--- a/studio/backend/utils/whisper_cpp_freshness.py
+++ b/studio/backend/utils/whisper_cpp_freshness.py
@@ -49,10 +49,6 @@ def read_install_marker(binary_path: Optional[str]) -> Optional[dict]:
return marker
-def _load_disk_cache(repo: str) -> Optional[tuple[float, Optional[str]]]:
- return _flow.load_disk_cache(repo, _cache_dir())
-
-
def _save_disk_cache(repo: str, latest_tag: Optional[str]) -> None:
_flow.save_disk_cache(
repo, latest_tag, _cache_dir(), log_message = "whisper freshness cache write failed"
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index 723ab0515ec..b3c12421e4a 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -420,7 +420,6 @@ def fix_xformers_performance_issue():
# flash-attn 4 makes flash_attn a namespace package without flash_attn_interface, which breaks
# `import xformers.ops`; import xformers once with flash_attn hidden from find_spec.
_FLASH_ATTN_INTERFACE_NAME = "flash_attn_interface"
-_FLASH_ATTN_INTERFACE_MODULE = "flash_attn." + _FLASH_ATTN_INTERFACE_NAME
_FA4_NAMESPACE_WARNED = [False]
diff --git a/unsloth/kernels/fast_lora.py b/unsloth/kernels/fast_lora.py
index 4e66f235541..e15f5f7c4bd 100644
--- a/unsloth/kernels/fast_lora.py
+++ b/unsloth/kernels/fast_lora.py
@@ -128,13 +128,6 @@ def _lora_dropout_p(
return p if p < 1 else None
-def _lora_B_bias(proj):
- adapter = _active_lora_adapter(proj)
- if adapter is None:
- return None
- return proj.lora_B[adapter].bias
-
-
def lora_bias_blocks_fused(proj):
return not FUSED_LORA_BIAS and _has_active_lora_bias(proj)
diff --git a/unsloth/kernels/moe/grouped_gemm/kernels/autotuning.py b/unsloth/kernels/moe/grouped_gemm/kernels/autotuning.py
index cb05f502ad9..3ff335338d5 100644
--- a/unsloth/kernels/moe/grouped_gemm/kernels/autotuning.py
+++ b/unsloth/kernels/moe/grouped_gemm/kernels/autotuning.py
@@ -46,16 +46,6 @@ def convert_args_to_list(args):
return [val_to_list(arg) for arg in args]
-def _triton_supports_tma():
- """Check if current Triton version supports TMA API."""
- import triton.language as tl
-
- # Both the old experimental and the new stable API names.
- return hasattr(tl, "make_tensor_descriptor") or hasattr(
- tl, "_experimental_make_tensor_descriptor"
- )
-
-
# TMA disabled for now: incompatible with permute_x/permute_y in the MoE grouped GEMM passes.
_TRITON_HAS_TMA = False
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 0de8d497a03..ff0ecc46928 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -2018,10 +2018,6 @@ def _get_flash_attention_disable_reason(config):
return None
-def _is_flash_attention_disabled(config):
- return _get_flash_attention_disable_reason(config) is not None
-
-
def _is_flash_attention_requested(attn_implementation):
return isinstance(attn_implementation, str) and attn_implementation.startswith(
"flash_attention"
@@ -4908,20 +4904,6 @@ try:
"```\n"
)
- if False:
- raise ImportError(
- "Unsloth: If you are in Colab, we updated the top cell install instructions - please change it to below "
- "then press Disconnect Runtime and then Restart it.\n"
- "\n"
- "%%capture\n"
- "# Installs Unsloth, Xformers (Flash Attention) and all other packages!\n"
- '!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"\n'
- '!pip install --no-deps "xformers<=0.0.27" trl peft accelerate bitsandbytes\n'
- "\n"
- f"Otherwise in local machines, your xformers version of {xformers_version} is too new.\n"
- 'Please downgrade xformers via `pip install --force-reinstall "xformers<=0.0.27"'
- )
-
if Version(torch_version) < Version("2.2.0") and Version(xformers_version) >= Version("0.0.24"):
raise ImportError(
f"Unsloth: You have torch = {torch_version} but xformers = {xformers_version}.\n"
@@ -5308,10 +5290,6 @@ def _download_statistics(names):
# Do not retry without a time limit: it would freeze offline.
-def _get_statistics(statistics = None, force_download = True):
- _download_statistics([_statistics_environment(statistics)])
-
-
def get_statistics(local_files_only = False):
# Basic stats on which environment is in use, and whether HuggingFace is down: a README.md is downloaded from HF, all data public. Disable with UNSLOTH_DISABLE_STATISTICS.
import os
diff --git a/unsloth/parallel_capabilities.py b/unsloth/parallel_capabilities.py
index 51caafc48cd..0fdeef28cfa 100644
--- a/unsloth/parallel_capabilities.py
+++ b/unsloth/parallel_capabilities.py
@@ -656,7 +656,6 @@ _CP_DP = {k: v for k, v in _CP_TP.items() if k != "block_swap"}
_CP_DP.update(tp = (1,), moe = (False,))
_FSDP2_CP = _GLOO + "test_cp_with_native_fsdp2_matches_one_process"
_HYBRID_FSDP2_CP = _GLOO + "test_hybrid_qwen3_5_cp_with_native_fsdp2_matches_one_process"
-_MOE_CP = _GLOO + "test_moe_without_aux_loss_under_cp_matches_one_process"
# Packed rows under generic Ulysses (HF attention, DDP over cp).
_PACKED_CP = {
"trainer": ("sft",),
diff --git a/unsloth/utils/prefix_grouper_kernel.py b/unsloth/utils/prefix_grouper_kernel.py
index 44bb264a344..cdc8fa165c7 100644
--- a/unsloth/utils/prefix_grouper_kernel.py
+++ b/unsloth/utils/prefix_grouper_kernel.py
@@ -77,8 +77,6 @@ def _kernel_options_for_dtype(dtype, one_stage = False):
return options
-_FLEX_KERNEL_OPTIONS = {"BLOCK_M": _FP32_BLOCK_M, "BLOCK_N": _FP32_BLOCK_N}
-
# The compiled backward asserts unless T is a multiple of 128; pad tokens are fully masked and sliced off.
_PAD_MULTIPLE = 128
_PAD_GROUP = -99 # sentinel group id / suffix id for pad tokens