Mirror _check_token_patterns: first match in dict order wins.

True if the model landed on a GPU, False if only CPU buffers landed despite GPU intent, None when there's no signal. Delegates to the shared

Stub llama-server under a supported install layout.

If *model_name* points to a LoRA adapter, return its base model. Checks ``adapter_config.json`` locally first. Only calls the heavier ``get_b

Handle a streaming Responses API call, emitting named SSE events. For GGUF models the request goes directly to llama-server's ``/v1/chat/comp

Run Python code in a sandboxed interpreter.

Send a response to the parent process.

Valid capability token for ``ref`` under the patched test secret.

Convert Anthropic messages + system to OpenAI-format message dicts. User messages with ``image`` blocks are emitted as OpenAI multimodal cont

Find the mmproj GGUF for a model. ``path``: directory or a .gguf file. ``search_root``: optional ancestor to also walk (snapshot layouts wher

Reset any cached generation state to prevent hanging after errors

Run a bash fragment. 60s default tolerates slow CI shell startup.

Terminate the subprocess and cancel any in-flight download.

True if we have enough GGUF metadata to estimate KV cache size.

Drive ``stream_chat_completion`` against a mocked Anthropic response and return the SSE lines. Pass ``captured_body`` to also capture the outg

Classify reasoning/tool capabilities via the GGUF classifier so flags match across backends. gpt-oss is overridden: Harmony routes reasoning and

Import `studio` without triggering server start; backend imports are lazy inside run().

Classify a ROCm device as unified-memory (APU) or discrete. Returns ``(gcn_arch, is_unified)``: - ``gcn_arch``: canonical arch string (e.g. `

Print a single step line in the column format.

Return (path, src) for the first existing candidate path, else None.

Run a bash fragment, return stdout.

Map an exception to a generic, client-safe message (never raw ``str(error)``, which can leak paths). Log the real exception server-side.

Hard fail in STRICT mode, info-warn otherwise.

Return True if newly written, False if already present.

True if the effective --spec-type selects MTP (mtp/draft-mtp), so the budget must reserve for it.

Install a minimal stub backend on routes.inference. Generation methods record which path the route entered, then yield one content event so t

Stub core.inference + routes.inference modules so the lazy imports inside training_vram resolve to fakes (avoids importing torch-heavy backends).

Build a fake device-properties object with the given attributes.

Attach a one-shot forward pre-hook recording whether a forward ran before trainer.train(), so prepare_for_training_mode can drop a torch.compile g

Anthropic-compatible Messages API endpoint. Translates Anthropic message format to internal OpenAI format, runs through the existing age

Remove unwanted characters, normalize whitespace

Check if a local path is or contains a GGUF model file. Handles a direct .gguf path or a directory of .gguf files. Skips mmproj files (pass t

Resolve a LoRA adapter's base model for a LOCAL dir OR a REMOTE HF repo. ``get_base_model_from_lora`` only reads a local adapter directory (it re

Grouped GEMM for MoE MLPs. The implementation offers a number of fusions specific to MoE: - `permute_x`: fuse the permutation of hidden

Machine-readable detail for the frontend. ``error_kind`` splits a user-approvable prompt (``remote_code_consent_required``) from a CRITICAL ha

Emit message_start and open the first text content block.

Build authentication headers using the provider's registry config.

Backend with just the dims the compute-buffer estimate reads.