Types
AdamParams Objects¶
class AdamParams(StrictBase)
learning_rate¶
Learning rate for the optimizer
beta1¶
Coefficient used for computing running averages of gradient
beta2¶
Coefficient used for computing running averages of gradient square
eps¶
Term added to the denominator to improve numerical stability
weight_decay¶
Weight decay for the optimizer. Uses decoupled weight decay.
grad_clip_norm¶
Gradient clip norm for the optimizer. 0.0 means no clipping.
SupportedModel Objects¶
class SupportedModel(BaseModel)
Information about a model supported by the server.
model_name¶
The name of the supported model.
GetServerCapabilitiesResponse Objects¶
class GetServerCapabilitiesResponse(BaseModel)
Response containing the server's supported models and capabilities.
supported_models¶
List of models available on the server.
OptimStepResponse Objects¶
class OptimStepResponse(BaseModel)
metrics¶
Optimization step metrics as key-value pairs
ModelInput Objects¶
class ModelInput(StrictBase)
chunks¶
Sequence of input chunks (formerly TokenSequence)
from_ints¶
def from_ints(cls, tokens: List[int]) -> "ModelInput"
Create a ModelInput from a list of ints (tokens).
to_ints¶
def to_ints() -> List[int]
Convert the ModelInput to a list of ints (tokens) Throws exception if there are any non-token chunks
length¶
def length() -> int
Return the total context length used by this ModelInput.
empty¶
def empty(cls) -> "ModelInput"
Create an empty ModelInput.
append¶
def append(chunk: ModelInputChunk) -> "ModelInput"
Add a new chunk, return a new ModelInput.
append_int¶
def append_int(token: int) -> "ModelInput"
Add a new token, return a new ModelInput.
WeightsInfoResponse Objects¶
class WeightsInfoResponse(BaseModel)
Minimal information for loading public checkpoints.
Checkpoint Objects¶
class Checkpoint(BaseModel)
checkpoint_id¶
The checkpoint ID
checkpoint_type¶
The type of checkpoint (training or sampler)
time¶
The time when the checkpoint was created
tinker_path¶
The checkpoint path
size_bytes¶
The size of the checkpoint in bytes
public¶
Whether the checkpoint is publicly accessible
ParsedCheckpointTinkerPath Objects¶
class ParsedCheckpointTinkerPath(BaseModel)
tinker_path¶
The checkpoint path
training_run_id¶
The training run ID
checkpoint_type¶
The type of checkpoint (training or sampler)
checkpoint_id¶
The checkpoint ID
from_tinker_path¶
def from_tinker_path(cls, tinker_path: str) -> "ParsedCheckpointTinkerPath"
Parse a checkpoint path to an instance of ParsedCheckpointTinkerPath
CheckpointArchiveUrlResponse Objects¶
class CheckpointArchiveUrlResponse(BaseModel)
url¶
Signed URL to download the checkpoint archive
expires¶
Unix timestamp when the signed URL expires, if available
SampledSequence Objects¶
class SampledSequence(BaseModel)
stop_reason¶
Reason why sampling stopped
tokens¶
List of generated token IDs
logprobs¶
Log probabilities for each token (optional)
TryAgainResponse Objects¶
class TryAgainResponse(BaseModel)
request_id¶
Request ID that is still pending
LoadWeightsRequest Objects¶
class LoadWeightsRequest(StrictBase)
path¶
A checkpoint URI for model weights at a specific step. Older deployments return
tinker://...; newer deployments may return logits://....
optimizer¶
Whether to load optimizer state along with model weights
TelemetrySendRequest Objects¶
class TelemetrySendRequest(StrictBase)
platform¶
Host platform name
sdk_version¶
SDK version string
ImageAssetPointerChunk Objects¶
class ImageAssetPointerChunk(StrictBase)
format¶
Image format
location¶
Path or URL to the image asset
expected_tokens¶
Expected number of tokens this image represents. This is only advisory: the backend will compute the number of tokens from the image, and we can fail requests quickly if the tokens does not match expected_tokens.
CheckpointsListResponse Objects¶
class CheckpointsListResponse(BaseModel)
checkpoints¶
List of available model checkpoints for the model
cursor¶
Pagination cursor information (None for unpaginated responses)
GenericEvent Objects¶
class GenericEvent(BaseModel)
event¶
Telemetry event type
event_name¶
Low-cardinality event name
severity¶
Log severity level
event_data¶
Arbitrary structured JSON payload
EncodedTextChunk Objects¶
class EncodedTextChunk(StrictBase)
tokens¶
Array of token IDs
ForwardBackwardInput Objects¶
class ForwardBackwardInput(StrictBase)
data¶
Array of input data for the forward/backward pass
loss_fn¶
Fully qualified function path for the loss function
loss_fn_config¶
Optional configuration parameters for the loss function (e.g., PPO clip thresholds, DPO beta)
SessionStartEvent Objects¶
class SessionStartEvent(BaseModel)
event¶
Telemetry event type
severity¶
Log severity level
TrainingRunsResponse Objects¶
class TrainingRunsResponse(BaseModel)
training_runs¶
List of training runs
cursor¶
Pagination cursor information
SaveWeightsResponse Objects¶
class SaveWeightsResponse(BaseModel)
path¶
A checkpoint URI for model weights at a specific step
SampleRequest Objects¶
class SampleRequest(StrictBase)
num_samples¶
Number of samples to generate
base_model¶
Optional base model name to sample from.
Is inferred from model_path, if provided. If sampling against a base model, this is required.
model_path¶
Optional checkpoint path (tinker://... or logits://...) to your model
weights or LoRA weights.
If not provided, samples against the base model.
sampling_session_id¶
Optional sampling session ID to use instead of model_path/base_model.
If provided along with seq_id, the model configuration will be loaded from the sampling session. This is useful for multi-turn conversations.
seq_id¶
Sequence ID within the sampling session.
Required when sampling_session_id is provided. Used to generate deterministic request IDs for the sampling request.
prompt_logprobs¶
If set to true, computes and returns logprobs on the prompt tokens.
Defaults to false.
topk_prompt_logprobs¶
If set to a positive integer, returns the top-k logprobs for each prompt token.
ForwardBackwardOutput Objects¶
class ForwardBackwardOutput(BaseModel)
loss_fn_output_type¶
The class name of the loss function output records (e.g., 'TorchLossReturn', 'ArrayRecord').
loss_fn_outputs¶
Dictionary mapping field names to tensor data
metrics¶
Training metrics as key-value pairs
SampleResponse Objects¶
class SampleResponse(BaseModel)
prompt_logprobs¶
If prompt_logprobs was set to true in the request, logprobs are computed for
every token in the prompt. The prompt_logprobs response contains a float32
value for every token in the prompt.
topk_prompt_logprobs¶
If topk_prompt_logprobs was set to a positive integer k in the request,
the top-k logprobs are computed for every token in the prompt. The
topk_prompt_logprobs response contains, for every token in the prompt,
a list of up to k (token_id, logprob) tuples.
CreateSamplingSessionResponse Objects¶
class CreateSamplingSessionResponse(BaseModel)
sampling_session_id¶
The generated sampling session ID
ModelData Objects¶
class ModelData(BaseModel)
Metadata about a model's architecture and configuration.
arch¶
The model architecture identifier.
model_name¶
The human-readable model name.
tokenizer_id¶
The identifier of the tokenizer used by this model.
GetInfoResponse Objects¶
class GetInfoResponse(BaseModel)
Response containing information about a training client's model.
type¶
Response type identifier.
model_data¶
Detailed metadata about the model.
model_id¶
Unique identifier for the model.
is_lora¶
Whether this is a LoRA fine-tuned model.
lora_rank¶
The rank of the LoRA adaptation, if applicable.
model_name¶
The name of the model.
Cursor Objects¶
class Cursor(BaseModel)
offset¶
The offset used for pagination
limit¶
The maximum number of items requested
total_count¶
The total number of items available
CreateModelRequest Objects¶
class CreateModelRequest(StrictBase)
base_model¶
The name of the base model to fine-tune (e.g., 'Qwen/Qwen3.5-4B').
user_metadata¶
Optional metadata about this model/training run, set by the end-user.
lora_config¶
LoRA configuration
Datum Objects¶
class Datum(StrictBase)
loss_fn_inputs¶
Dictionary mapping field names to tensor data
convert_tensors¶
def convert_tensors(cls, data: Any) -> Any
Convert torch.Tensor and numpy arrays to TensorData in loss_fn_inputs during construction.
TrainingRun Objects¶
class TrainingRun(BaseModel)
training_run_id¶
The unique identifier for the training run
base_model¶
The base model name this model is derived from
model_owner¶
The owner/creator of this model
is_lora¶
Whether this model uses LoRA (Low-Rank Adaptation)
corrupted¶
Whether the model is in a corrupted state
lora_rank¶
The LoRA rank if this is a LoRA model, null otherwise
last_request_time¶
The timestamp of the last request made to this model
last_checkpoint¶
The most recent training checkpoint, if available
last_sampler_checkpoint¶
The most recent sampler checkpoint, if available
user_metadata¶
Optional metadata about this training run, set by the end-user
SessionEndEvent Objects¶
class SessionEndEvent(BaseModel)
duration¶
ISO 8601 duration string
event¶
Telemetry event type
severity¶
Log severity level
TelemetryBatch Objects¶
class TelemetryBatch(BaseModel)
platform¶
Host platform name
sdk_version¶
SDK version string
UnhandledExceptionEvent Objects¶
class UnhandledExceptionEvent(BaseModel)
event¶
Telemetry event type
severity¶
Log severity level
traceback¶
Optional Python traceback string
ImageChunk Objects¶
class ImageChunk(StrictBase)
data¶
Image data as bytes
format¶
Image format
expected_tokens¶
Expected number of tokens this image represents. This is only advisory: the backend will compute the number of tokens from the image, and we can fail requests quickly if the tokens does not match expected_tokens.
validate_data¶
def validate_data(cls, value: Union[bytes, str]) -> bytes
Deserialize base64 string to bytes if needed.
serialize_data¶
def serialize_data(value: bytes) -> str
Serialize bytes to base64 string for JSON.
SaveWeightsRequest Objects¶
class SaveWeightsRequest(StrictBase)
path¶
A file/directory name for the weights
LoraConfig Objects¶
class LoraConfig(StrictBase)
rank¶
LoRA rank (dimension of low-rank matrices)
seed¶
Seed used for initialization of LoRA weights.
Useful if you need deterministic or reproducible initialization of weights.
train_unembed¶
Whether to add lora to the unembedding layer
train_mlp¶
Whether to add loras to the MLP layers (including MoE layers)
train_attn¶
Whether to add loras to the attention layers
CreateSamplingSessionRequest Objects¶
class CreateSamplingSessionRequest(StrictBase)
session_id¶
The session ID to create the sampling session within
sampling_session_seq_id¶
Sequence ID for the sampling session within the session
base_model¶
Optional base model name to sample from.
Is inferred from model_path, if provided. If sampling against a base model, this is required.
model_path¶
Optional checkpoint path (tinker://... or logits://...) to your model
weights or LoRA weights.
If not provided, samples against the base model.
FutureRetrieveRequest Objects¶
class FutureRetrieveRequest(StrictBase)
request_id¶
The ID of the request to retrieve
TensorData Objects¶
class TensorData(StrictBase)
data¶
Flattened tensor data as array of numbers.
shape¶
Optional.
The shape of the tensor (see PyTorch tensor.shape). The shape of a
one-dimensional list of length N is (N,). Can usually be inferred if not
provided, and is generally inferred as a 1D tensor.
to_numpy¶
def to_numpy() -> npt.NDArray[Any]
Convert TensorData to numpy array.
to_torch¶
def to_torch() -> "torch.Tensor"
Convert TensorData to torch tensor.
SaveWeightsForSamplerRequest Objects¶
class SaveWeightsForSamplerRequest(StrictBase)
path¶
A file/directory name for the weights
SamplingParams Objects¶
class SamplingParams(BaseModel)
max_tokens¶
Maximum number of tokens to generate
seed¶
Random seed for reproducible generation
stop¶
Stop sequences for generation
temperature¶
Sampling temperature
top_k¶
Top-k sampling parameter (-1 for no limit)
top_p¶
Nucleus sampling probability
SaveWeightsForSamplerResponseInternal Objects¶
class SaveWeightsForSamplerResponseInternal(BaseModel)
path¶
A checkpoint URI for model weights for sampling at a specific step
sampling_session_id¶
The generated sampling session ID
SaveWeightsForSamplerResponse Objects¶
class SaveWeightsForSamplerResponse(BaseModel)
path¶
A checkpoint URI for model weights for sampling at a specific step
LoadWeightsResponse Objects¶
class LoadWeightsResponse(BaseModel)
path¶
A checkpoint URI for model weights at a specific step