Skip to content
View as Markdown

tinker.types.ClientConfigResponse

class tinker.types.ClientConfigResponse(BaseModel)

Server-side feature flags resolved for this caller.

Uses BaseModel (extra="ignore") so new flags from the server are silently dropped until the SDK adds fields for them.

Fields:

  • pjwt_auth_enabled (bool)
  • credential_default_source (str)
  • sample_dispatch_bytes_semaphore_size (int)
  • inflight_response_bytes_semaphore_size (int)
  • parallel_fwdbwd_chunks (bool)
  • proto_compress_fwdbwd (bool) – When true, the SDK zstd-compresses the proto fwd/bwd request body and sets Content-Encoding: zstd. Real fwd/bwd payloads compress >10× — the API server decompresses transparently via an ASGI middleware.
  • fwdbwd_max_chunk_len (int) – Maximum number of datums per fwd/bwd chunk request.
  • fwdbwd_max_chunk_bytes_count (int) – Maximum estimated request payload bytes (serialized proto, before zstd compression) per fwd/bwd chunk request.
  • fwdbwd_dispatch_bytes_semaphore_size (int) – Maximum estimated fwd/bwd request payload bytes being dispatched at the same time (same pre-compression estimate as fwdbwd_max_chunk_bytes_count). Bounds client memory for serialized bodies and smooths submission bursts.
  • billing_exception_max_pause_duration_sec (int)
  • sample_no_retries (bool)
  • sample_enable_stuck_detection (bool) – When true, the SDK runs the retry handler's progress timeout check that raises APIConnectionError("...Requests appear to be stuck.") when no progress is made within RetryConfig.progress_timeout.
  • sample_max_concurrent_requests (int) – Maximum number of in-flight sampling requests a SamplingClient will allow (the size of its retry handler's semaphore). Always applied as the sampling RetryConfig.max_connections, overriding any value in a caller-provided retry_config.
  • use_pyqwest_transport (bool) – When true, the SDK builds its default httpx async client on top of the pyqwest (reqwest/hyper-based) transport adapter. Set to false server-side to force every client to fall back to httpx's default transport.
  • create_model_via_load_weights (bool) – When true, create_training_client_from_state* skips the weights_info and create_model round-trips and sends a single LoadWeightsRequest with session addressing (session_id + model_seq_id) that creates the model server-side, configured from the checkpoint's owning model. Requires a server with load-weights-first support.