True if ``flag`` is Studio-managed. Normalises via ``_flag_name`` so `-np8` / `--parallel=8` classify like the canonical tokens.

Build and run a pip install command (uses uv when available, falls back to pip).

Lowest tier whose built-in parser loads the config; *floor* is the fail-safe. Escalates ``_PROBE_TIER_ORDER`` (prefixed with the ambient ``defaul

Invoke _ensure_cuda_torch under a fully mocked host; return the pip mock. cvd controls CUDA_VISIBLE_DEVICES: None removes it from the env, any st

Return a child-process env with the native path lease secret removed.

Parse `llama-server --help` for feature flags. Returns {found, mtp_token, supports_mtp, ngram_mod_flavor, supports_ngram_mod, spec_dra

Create ``$root/models--<repo>/snapshots/<sha>/<rel>`` for each entry.

Drive a stream with a custom set of SSE events and return raw lines.

Leftmost-position match; ties prefer the longer token.

Build a configured loop with a multi-turn fake generator. ``turns`` is a list of chunk-lists; iteration N yields chunks from ``turns[N]``.

Convert a ResponsesRequest's ``input`` into a Chat-format ``ChatMessage`` list. Handles the three input item shapes allowed by the Responses API:

Process one OpenAI streaming chat.completion.chunk.

Record the user's "allow"/"deny" decision and unblock the loop. Returns ``True`` if a pending call matched, ``False`` otherwise (e.g. a stale

Qwen3.6-27B-MTP-class backend (embedded head) with the MTP-math dims.

Build a ``\\ue200cite\\ue202<sid>[\\ue202<sid>...][\\ue202<loc>]\\ue201`` marker from one or more ``source_ids`` plus an optional ``locator``.

MTP-named GGUF backend that's already running with draft-mtp.

Print an in-place progress bar aligned to the step column layout.

Create a fake prebuilt install and return the llama-server path.

Format a single Anthropic SSE event.

Decide whether a ``trust_remote_code=True`` load may proceed, over every repo whose code the load would execute. A LoRA load runs adapter AND base

Return the detected device, auto-detecting if detect_hardware() hasn't run. Prefer calling detect_hardware() explicitly at startup.

Return (status_code, parsed_json_or_raw_bytes).

Check if path is a local filesystem path vs HuggingFace model identifier. Examples: True: /home/user/model, C:\\models, ./model, ~/m

Drive an agentic tool loop on top of a cumulative-text generator. ``single_turn(messages)`` must yield cumulative assistant text (each yield

Yield OpenAI-format SSE lines from the external provider. OpenAI-compatible providers forward lines verbatim. For Anthropic, the

Validate user-supplied llama-server args. Returns a flat list ready to extend the llama-server command; raises ``ValueError`` naming the offen

Extract quant label like Q4_K_M, IQ4_XS, BF16 from a GGUF filename. Examples: "gemma-3-4b-it-Q4_K_M.gguf" → "Q4_K_M"

``(tier, reason)`` from name substrings (order 510 > 550 > 530), or ``None``. Underscore aliases match (``Qwen3_5`` == ``Qwen3.5``); a dot-versio

Single-repo consent; thin wrapper over the for_targets form. ``trusted_org`` is accepted for backward compatibility but no longer changes the deci

Build an OpenAI-style error envelope. Returns ``{"error": {"message", "type", "param", "code"}}``. The ``param`` and ``code`` keys are always

torch stub: version.hip, cuda.is_available/device_count, get_device_properties(i).name.

Apply a subprocess event to local state. State updates happen inside self._lock; DB I/O happens after releasing it so status-polling

True iff every selected NVIDIA GPU is a datacenter/professional part. NVIDIA-only, fails open to False (consumer GeForce, ROCm, CPU and errors

Return parsed ``config.json`` for *model_name*, checking local files first. ``hf_token`` authenticates the raw fetch so gated/private repos resol

Schedule one background reload without MTP after a mid-generation death. MTP+tensor can crash the flash-attn kernel on a later request, after

Plan a ``--split-mode tensor`` load. Pure: no model or GPU needed. ``gpus`` is a list of ``(gpu_index, free_mib)``; ``model_size`` is the

True only if *name* is a genuine HF repo under a trusted org. Fails closed (local paths, malformed names, untrusted namespaces, Hub errors); never

Run the field validator without building a full TrainingStartRequest.

Per-device compute-graph buffer (bytes) from GGUF dims: a vocab-width output buffer + activation scratch. Context-independent; scales with

Terminate the subprocess if running.

Backend with a deterministic probe so the resolver is hermetic.

Return DLL dirs from pip-installed CUDA wheels under ``<prefix>/Lib/site-packages/`` so llama-server.exe can load ``cudart64_X.dll`` /

Mint an unguessable id for one pending tool-call confirmation.

Shortcut for simulating a worker log message.