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.

Studio: remove private helpers and constants that nothing uses (#15696)

Daniel Han1f55d8131c7d
diff --git a/studio/backend/core/inference/chat_template_helpers.py b/studio/backend/core/inference/chat_template_helpers.py
index 115f6dc2dd3..355bc9f8c27 100644
--- a/studio/backend/core/inference/chat_template_helpers.py
+++ b/studio/backend/core/inference/chat_template_helpers.py
@@ -22,7 +22,6 @@ _UNPARSED = object()
_THINK_OPEN = "<think>"
_THINK_CLOSE = "</think>"
-_GEMMA_CHANNEL_START = "<|channel>"
_GEMMA_THOUGHT_OPEN = "<|channel>thought"
_GEMMA_THOUGHT_CLOSE = "<channel|>"
_GEMMA_TEMPLATE_OPENERS = (
diff --git a/studio/backend/core/inference/context_window.py b/studio/backend/core/inference/context_window.py
index e4b7d362e26..3ae005da261 100644
--- a/studio/backend/core/inference/context_window.py
+++ b/studio/backend/core/inference/context_window.py
@@ -412,16 +412,6 @@ _ARG_COMPACTION_AGGREGATE_LEAF_FLOOR = 256
_ARG_COMPACTION_TOTAL_FLOOR_CHARS = 1024
-def _largest_leaf(value: Any) -> int:
- if isinstance(value, str):
- return len(value)
- if isinstance(value, dict):
- return max((_largest_leaf(item) for item in value.values()), default = 0)
- if isinstance(value, list):
- return max((_largest_leaf(item) for item in value), default = 0)
- return 0
-
-
def _total_leaves(value: Any) -> int:
if isinstance(value, str):
return len(value)
diff --git a/studio/backend/core/inference/diffusion_attention.py b/studio/backend/core/inference/diffusion_attention.py
index 8d0ece10061..b03cf918b9a 100644
--- a/studio/backend/core/inference/diffusion_attention.py
+++ b/studio/backend/core/inference/diffusion_attention.py
@@ -1831,7 +1831,6 @@ def _cudnn_attention_supported() -> bool:
_SAGE_PROBE_CACHE: dict[tuple[str, str, int], str] = {}
_SAGE_MAX_HEAD_DIM = 128 # sageattn raises above 128
-_SAGE_DTYPE_NAMES = ("float16", "bfloat16", "fp16", "bf16", "half")
# SageAttention 2 lands near cos 0.9999 / rel-L1 0.01-0.03 vs fp32 here; a wrong-arch or broken build is far outside.
_SAGE_MIN_COSINE = 0.99
_SAGE_MAX_REL_L1 = 0.08
diff --git a/studio/backend/core/inference/diffusion_cache.py b/studio/backend/core/inference/diffusion_cache.py
index cb72d21b42c..3e9b69fa243 100644
--- a/studio/backend/core/inference/diffusion_cache.py
+++ b/studio/backend/core/inference/diffusion_cache.py
@@ -31,7 +31,6 @@ AUTO_STEP_CACHE_TIER = "max"
_REDUCTION_FILTER_OPTION = "test_configs.force_filter_reduction_configs"
-_TIER_DEFAULT = "default"
_TIER_MAX = "max"
diff --git a/studio/backend/core/inference/diffusion_cast_cache.py b/studio/backend/core/inference/diffusion_cast_cache.py
index cadbcb9faca..60dfb32f0e1 100644
--- a/studio/backend/core/inference/diffusion_cast_cache.py
+++ b/studio/backend/core/inference/diffusion_cast_cache.py
@@ -163,10 +163,6 @@ def _cast_bytes(shards: list) -> Optional[int]:
return total if seen else None
-def _all_wide(shards: list) -> bool:
- return _cast_bytes(shards) is not None
-
-
def _fit_budget(cands: list, logger: Any = None) -> list:
"""A set over the LRU budget would evict its own entries every load. Greedy largest-first, so it is deterministic."""
budget = _cache.max_cache_bytes()
diff --git a/studio/backend/core/inference/diffusion_layer_split_plan.py b/studio/backend/core/inference/diffusion_layer_split_plan.py
index 0ef16a29ea4..4c12ad6c215 100644
--- a/studio/backend/core/inference/diffusion_layer_split_plan.py
+++ b/studio/backend/core/inference/diffusion_layer_split_plan.py
@@ -385,14 +385,6 @@ def plan_layer_split(
)
-_BLOCK_LIST_ATTRS: tuple[str, ...] = (
- "transformer_blocks",
- "single_transformer_blocks",
- "blocks",
- "layers",
-)
-
-
def split_blocks(module: Any) -> Optional[list[tuple[str, Any]]]:
"""The denoiser's repeated blocks in forward order as ``(qualified name, module)`` (the engine's own list, so a
plan's indices name the blocks it places); None when the engine cannot cut this denoiser."""
diff --git a/studio/backend/core/inference/diffusion_lora.py b/studio/backend/core/inference/diffusion_lora.py
index 05b9e396014..9033b08d7fb 100644
--- a/studio/backend/core/inference/diffusion_lora.py
+++ b/studio/backend/core/inference/diffusion_lora.py
@@ -21,9 +21,6 @@ from utils.paths.storage_roots import account_path
from .diffusion_families import DIFFUSION_CANCELLED_MSG
from utils.paths.path_utils import is_appledouble_metadata
-# Accepted LoRA formats. safetensors + gguf only (.pt is pickled -> excluded for safety).
-_NATIVE_EXTS = (".safetensors", ".gguf")
-_DIFFUSERS_EXTS = (".safetensors",)
_ALL_EXTS = (".safetensors", ".gguf")
# ``kind`` in a ``<stem>.json`` sidecar marking the weight beside it as an image LoRA, not model weights.
LORA_SIDECAR_KIND = "diffusion-lora"
diff --git a/studio/backend/core/inference/diffusion_prequant_cache.py b/studio/backend/core/inference/diffusion_prequant_cache.py
index 22edf9d11e2..9cd4bccf69f 100644
--- a/studio/backend/core/inference/diffusion_prequant_cache.py
+++ b/studio/backend/core/inference/diffusion_prequant_cache.py
@@ -292,12 +292,6 @@ def source_shards(
return None
-def _source_shards_in(base: str, component: str, cache_dir: Optional[str]) -> Optional[list[str]]:
- for found in _shard_sets_in(base, component, cache_dir):
- return found
- return None
-
-
def _shard_sets_in(base: str, component: str, cache_dir: Optional[str]):
candidates = [str(base)]
try:
@@ -589,7 +583,6 @@ _REMOTE_MEMO: dict[tuple, tuple] = {}
_GONE = "gone"
# Shared hub_reachability verdict: after one failure, lookups answer "unknown" (local snapshot stands) for the TTL.
_UNREACHABLE = _reach._UNREACHABLE
-_connection_failed = _reach.connection_failed
_endpoint = _reach.endpoint
diff --git a/studio/backend/core/inference/diffusion_transformer_quant.py b/studio/backend/core/inference/diffusion_transformer_quant.py
index a1ba7dfce5b..de00fa9ead0 100644
--- a/studio/backend/core/inference/diffusion_transformer_quant.py
+++ b/studio/backend/core/inference/diffusion_transformer_quant.py
@@ -467,12 +467,6 @@ def _nvfp4_gate_failure(family) -> Optional[str]:
return None
-_TRAIN_DENY_NVFP4_REASON = (
- "nvfp4 is inference-only: the accuracy gate behind it measures a frozen forward at a "
- "per-layer policy, and no training run has been measured on 4-bit frozen linears"
-)
-
-
def _family_train_denied(
family,
scheme: str,
diff --git a/studio/backend/core/inference/diffusion_vae_fused.py b/studio/backend/core/inference/diffusion_vae_fused.py
index b378ff155c1..fb7ba9bdbc0 100644
--- a/studio/backend/core/inference/diffusion_vae_fused.py
+++ b/studio/backend/core/inference/diffusion_vae_fused.py
@@ -1157,7 +1157,6 @@ def uninstall(vae: Any) -> None:
_WAN_VAES = frozenset({"AutoencoderKLWan", "AutoencoderKLQwenImage", "AutoencoderKLQwenImage21"})
-_CACHE_T = 2
def _is_rms(norm: Any) -> bool:
diff --git a/studio/backend/core/inference/diffusion_vae_oom.py b/studio/backend/core/inference/diffusion_vae_oom.py
index b3cda3c55d7..344db8b5284 100644
--- a/studio/backend/core/inference/diffusion_vae_oom.py
+++ b/studio/backend/core/inference/diffusion_vae_oom.py
@@ -656,11 +656,6 @@ class DecodeSession:
return Change(f"moved:{name}", lambda: _return_to(comp, device), settle, defer)
- def _component_undo(self, name: str, comp: Any) -> tuple:
- """(undo, defer) for a component moved to the host; kept for callers recording a move by hand."""
- change = self._component_change(name, comp)
- return change.retry, lambda exc: change.defer(change.retry, exc)
-
def offload_vae(self) -> bool:
"""Step (c): the VAE's weights stream per layer from the host (diffusers leaf-level group offload)."""
if self.vae_offloaded or getattr(self.device, "type", None) != "cuda" or _hooked(self.vae):
diff --git a/studio/backend/core/inference/llama_server_args.py b/studio/backend/core/inference/llama_server_args.py
index d7200d1bf69..bb58c605e6b 100644
--- a/studio/backend/core/inference/llama_server_args.py
+++ b/studio/backend/core/inference/llama_server_args.py
@@ -584,9 +584,6 @@ MEMORY_ENV_VARS: tuple[str, ...] = (
"LLAMA_ARG_NO_DIO",
)
-_SHADOWING_FLAGS: frozenset[str] = (
- _CONTEXT_FLAGS | _CACHE_FLAGS | _SPEC_FLAGS | _TEMPLATE_FLAGS | _SPLIT_SHADOWING_FLAGS
-)
# Shadowing flags that take no value -- strip the flag only, not the next token.
_BOOLEAN_SHADOWING_FLAGS: frozenset[str] = frozenset(
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 7edcc9f9a24..414d4e71a98 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -213,7 +213,6 @@ _STOP_NOTICE_INTERVAL = 0.1
_DISPATCH_POLL_INTERVAL = 0.5
_DISPATCH_STOP_TIMEOUT = 5.0
_DISPATCH_IDLE_TIMEOUT = 30.0
-_DISPATCH_DRAIN_TIMEOUT = 5.0
_CANCELLED_ROWS_GRACE = 5.0
# Transformers TTS only; generous since _ensure_subprocess_alive already catches dead workers.
diff --git a/studio/backend/core/inference/passthrough_healing.py b/studio/backend/core/inference/passthrough_healing.py
index 773f5a9b1b7..7e4c255c75f 100644
--- a/studio/backend/core/inference/passthrough_healing.py
+++ b/studio/backend/core/inference/passthrough_healing.py
@@ -179,17 +179,6 @@ def _promote(
return promoted
-def _remove_spans(text: str, spans: list) -> str:
- """Text with the given non-overlapping, sorted (start, end) ranges removed."""
- pieces = []
- pos = 0
- for start, end in spans:
- pieces.append(text[pos:start])
- pos = end
- pieces.append(text[pos:])
- return "".join(pieces)
-
-
def heal_openai_message_events(
msg: dict,
allowed_tools: set,
diff --git a/studio/backend/core/training/diffusion_checkpoint.py b/studio/backend/core/training/diffusion_checkpoint.py
index 6b07bf5691b..15a2e568868 100644
--- a/studio/backend/core/training/diffusion_checkpoint.py
+++ b/studio/backend/core/training/diffusion_checkpoint.py
@@ -117,8 +117,6 @@ _OPTIONAL_IDENTITY_FIELDS = frozenset(
)
# Bundle side only: an incoming None is the route's unresolved pre-trainer identity ("cannot tell").
_LEGACY_IDENTITY_DEFAULTS: dict[str, str] = {"bucketing": "off"}
-# What source_revision() returns when it cannot resolve a revision offline.
-_UNRESOLVED_REVISION = "unresolved"
def _revision_is_comparable(value: Any) -> bool:
diff --git a/studio/backend/hub/services/models/account_access.py b/studio/backend/hub/services/models/account_access.py
index ebf0550fcc4..5174fe71fb5 100644
--- a/studio/backend/hub/services/models/account_access.py
+++ b/studio/backend/hub/services/models/account_access.py
@@ -588,13 +588,6 @@ def _public_key(repo_id: str, repo_type: str) -> tuple[str, str, str]:
return (hf_constants.ENDPOINT.rstrip("/"), repo_type, repo_id.lower())
-def _public_verdict(repo_id: str, repo_type: str) -> bool | None:
- key = _public_key(repo_id, repo_type)
- with _public_lock:
- cached = _public_repos.get(key)
- return cached[1] if cached is not None and cached[0] > time.monotonic() else None
-
-
def _source_speaks_for_the_cache() -> bool:
from utils.hub_settings import MODELSCOPE, active_source
return active_source() != MODELSCOPE
diff --git a/studio/backend/hub/services/models/gguf_variants.py b/studio/backend/hub/services/models/gguf_variants.py
index 5acfd685588..2d62475827d 100644
--- a/studio/backend/hub/services/models/gguf_variants.py
+++ b/studio/backend/hub/services/models/gguf_variants.py
@@ -783,9 +783,6 @@ def _direct_gguf_split_is_whole(path: Path) -> bool:
return found >= set(range(1, total + 1))
-# Cache scan's looser grammar (hub.utils.inventory_scan._GGUF_SPLIT_RE), not the five digits above.
-_CACHE_SPLIT_RE = re.compile(r"-(\d{3,})-of-(\d{3,})(?=\.gguf$)", re.IGNORECASE)
-
# Only enumerates candidates for the resolver to adjudicate; drift costs a candidate, not a verdict.
_KNOWN_QUANT_RE = re.compile(
r"(UD-)?"
diff --git a/studio/backend/hub/utils/gguf_header_delta.py b/studio/backend/hub/utils/gguf_header_delta.py
index d15e355a6b9..42953611328 100644
--- a/studio/backend/hub/utils/gguf_header_delta.py
+++ b/studio/backend/hub/utils/gguf_header_delta.py
@@ -37,7 +37,6 @@ _MAX_HEADER_BYTES = 256 << 20
_COPY_CHUNK = 16 << 20
_DISK_MARGIN = 256 << 20
_METADATA_TIMEOUT = 10.0
-_RANGE_TIMEOUT = 60.0
# "image-diffusion-unsupported" is an image / video family this install cannot run yet; still rebuilt.
_MEDIA_TASKS = frozenset({"text-to-image", "text-to-video", "image-diffusion-unsupported"})
diff --git a/studio/backend/routes/data_recipe/seed.py b/studio/backend/routes/data_recipe/seed.py
index bf83cc30a29..bc035f1ef28 100644
--- a/studio/backend/routes/data_recipe/seed.py
+++ b/studio/backend/routes/data_recipe/seed.py
@@ -133,11 +133,6 @@ def _list_hf_repo_files(*, dataset_name: str, token: HfTokenArg) -> list[str]:
return []
-def _list_hf_data_files(*, dataset_name: str, token: HfTokenArg) -> list[str]:
- repo_files = _list_hf_repo_files(dataset_name = dataset_name, token = token)
- return [file for file in repo_files if file.lower().endswith(DATA_EXTS)]
-
-
def _list_hf_dataset_configs(*, dataset_name: str, token: HfTokenArg) -> list[dict[str, Any]]:
"""Read configs from the card's configs block; config names need not match folder names."""
try:
diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index a9b9239cf23..e993ce2b234 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -3277,14 +3277,6 @@ def _gguf_conversion_name_matches(hit_name: str, base: str) -> bool:
return name == base or name in {f"{base}-gguf", f"{base}_gguf", f"{base}.gguf"}
-def _gguf_files_for_pick(names: list[str], picked: str) -> Optional[list[str]]:
- """The complete downloadable file family for a picked GGUF. llama-server opens split siblings implicitly,
- so a single selected shard is not a usable plan, and incomplete published families are rejected.
- Deferred to the loader so the plan offered here and the transfer name one set."""
- from core.rag.embed_llama_server import LlamaServerBackend
- return LlamaServerBackend._split_family(names, picked)
-
-
def _pick_downloadable_gguf(names: list[str]) -> Optional[list[str]]:
"""Pick the loader's preferred GGUF, skipping torn split families."""
from core.rag.embed_llama_server import LlamaServerBackend
diff --git a/studio/backend/storage/chat_generation_runs_db.py b/studio/backend/storage/chat_generation_runs_db.py
index 376eab8b469..39099d7a58b 100644
--- a/studio/backend/storage/chat_generation_runs_db.py
+++ b/studio/backend/storage/chat_generation_runs_db.py
@@ -149,17 +149,6 @@ def _unregister_locked(entry: _PoolEntry | None) -> None:
_pool_registry[:] = surviving
-def _discard_pooled() -> None:
- """Drop this thread's cached connection, for when it has errored or its database has gone."""
- entry = getattr(_pool, "entry", None)
- if entry is None:
- return
- _pool.entry = None
- with _pool_lock:
- _unregister_locked(entry)
- entry.release()
-
-
def _discard_all_pooled() -> None:
"""Closes every pooled connection, idle ones at once; one in use is closed by its borrower on return, since yanking it would fail that query."""
global _pool_generation
diff --git a/studio/backend/utils/worker_stderr.py b/studio/backend/utils/worker_stderr.py
index b689a87aad1..f35878a17c5 100644
--- a/studio/backend/utils/worker_stderr.py
+++ b/studio/backend/utils/worker_stderr.py
@@ -66,12 +66,6 @@ def _utf8_edge_bounds(data: bytes, *, trim_end: bool = True) -> "tuple[int, int]
return start, end
-def _without_partial_utf8_edges(data: bytes, *, trim_end: bool = True) -> bytes:
- """``trim_end`` is off at EOF, where a trailing 0xC0+ byte is a complete cp1252 character, not a severed one."""
- start, end = _utf8_edge_bounds(data, trim_end = trim_end)
- return data[start:end]
-
-
def decode_worker_stderr(data: bytes, *, ends_at_eof: bool = True) -> str:
"""UTF-8 before cp1252: cp1252 has a meaning for almost every byte, so the stricter encoding goes first."""
text: str | None = None
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index c9b7b82e695..fa5a9d73a66 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -1787,20 +1787,6 @@ def _write_private_text(path: Path, text: str) -> None:
handle.write(text)
-def _read_yaml_object(path: Path) -> Optional[dict]:
- import yaml
-
- if not path.exists():
- return {}
- try:
- data = yaml.safe_load(path.read_text(encoding = "utf-8"))
- except (yaml.YAMLError, OSError):
- return None
- if data is None:
- return {}
- return data if isinstance(data, dict) else None
-
-
def _read_json_object(path: Path) -> Optional[dict]:
# {} when missing, None when it cannot be parsed as an object, so the caller leaves a user-managed file untouched rather than clobbering it.
if not path.exists():