Sampling
Generate completions
Generate one or more completions from the specified sampler checkpoint. The sampler-weights save request must have completed before sampling. Poll the returned request identifier for the result.
Header parameters
Foundry-Features
enum
required
A feature flag opt-in required when using preview operations or modifying persisted preview resources.
Path parameters
sessionId
string
required
Identifier of the fine-tuning session containing the sampler checkpoint.
Query parameters
checkpoint_id
string
required
Identifier of the completed sampler checkpoint to use for generation.
api-version
string
required
The API version to use for this operation.
Request Body
Prompt and generation parameters for the requested completions.
num_samples
integer
Number of independent completions to generate. If omitted, the service defaults to 1.
prompt
SessionFineTuning.FineTuningModelInput
required
Full model input as an ordered sequence of token-ID and image chunks.
chunks
SessionFineTuning.FineTuningInputChunk[]
required
Chunks forming the complete model input. At most 64 image chunks are allowed per example; the service enforces this limit at runtime.
sampling_params
SessionFineTuning.FineTuningSamplingParams
required
Token-generation sampling parameters.
max_tokens
integer
required
Maximum number of tokens to generate. Must be positive.
temperature
number
Softmax temperature. If omitted, the service defaults to 1.0.
top_p
number
Nucleus (top-p) probability mass. If omitted, the service defaults to 1.0.
top_k
integer
Top-k candidates. If omitted, the service defaults to -1. A value of -1 disables top-k filtering.
seed
integer
RNG seed for reproducible samples. If omitted, the service chooses a random seed.
stop_criteria
SessionFineTuning.FineTuningStopCriteria
Stop criteria for generation — either a list of token IDs or a list of strings, but never both.
stop_token_ids
integer[]
Stop when any of these token IDs is produced.
stop_strings
string[]
Stop when any of these strings appears in the output.
response_format
object | null
Optional structured-output format passed to the sampling provider.
sampling_session_id
string
Sampling session identifier from a prior sampler-weights save request.
seq_id
integer
Training step index; must match the seq_id used when saving sampler weights.
prompt_logprobs
boolean
Whether to return prompt-token log-probabilities in the overall sample result. If omitted, the service defaults to false.
topk_prompt_logprobs
integer
Number of top-k log-probabilities per prompt token. If omitted, the service defaults to 0. Zero disables this output; maximum 20.
Responses
The request has succeeded.
Body
request_id
string
required
Request identifier to pass to the session's request-status endpoint.
status
enum
required
Indicates that the request has not yet completed.
Request
POST /fine_tuning/sessions/{sessionId}/sample HTTP/1.1Response
Status: 200
Content-Type: application/jsonStatus: 4XX
Content-Type: application/jsonStatus: 5XX
Content-Type: application/json