gateway.openapi_Gateway

CRUD APIs for training shape. Create Training Shape

post/v1/accounts/{account_id}/trainingShapes

Request

  • Base URL: https://api.fireworks.ai
  • URL: https://api.fireworks.ai/v1/accounts/{account_id}/trainingShapes
  • Auth: HTTP bearer

Path parameters

account_idstring required

The Account Id

Query parameters

trainingShapeIdstring

The ID of the training shape. If not specified, a random ID will be generated. Must follow AIP-122 segment format and start with a letter.

Request body

namestring
displayNamestring

Human-readable display name of the training shape. e.g. "Llama3 70B H200 BF16" Must be fewer than 64 characters long.

descriptionstring

The description of the training shape. Must be fewer than 1000 characters long.

createTimestring date-time

The creation time of the training shape.

updateTimestring date-time

The update time for the training shape.

baseModelstring required
deploymentShapeVersionstring
trainerImageTagstring required

The validated trainer runtime image tag used for numerics verification.

trainerMode'TRAINER_MODE_UNSPECIFIED' | 'POLICY_TRAINER' | 'FORWARD_ONLY' | 'LORA_TRAINER'

Trainer execution mode used for validated launch-profile matching.

nodeCountinteger

Node count validated for the launch profile.

modelTypestring

The model type of the base model (e.g. "llama", "qwen").

parameterCountstring int64

The parameter count of the base model, used for billion-parameter bucket matching.

acceleratorType'ACCELERATOR_TYPE_UNSPECIFIED' | 'NVIDIA_A100_80GB' | 'NVIDIA_H100_80GB' | 'AMD_MI300X_192GB' | 'NVIDIA_A10G_24GB' | 'NVIDIA_A100_40GB' | 'NVIDIA_L4_24GB' | 'NVIDIA_H200_141GB' | 'NVIDIA_B200_180GB' | 'AMD_MI325X_256GB' | 'AMD_MI350X_288GB' | 'NVIDIA_B300_288GB' | 'NVIDIA_GB200' | 'NVIDIA_GB300'
acceleratorCountinteger

Total number of accelerators used by the job.

baseModelWeightPrecision'WEIGHT_PRECISION_UNSPECIFIED' | 'BFLOAT16' | 'INT8' | 'NF4' | 'FP8' | 'FP4_FP8'

The weight precision for model training/inference.

  • BFLOAT16: no quantization applied
  • INT8: enable 8-bit quantization with LLM.int8()
  • NF4: enable 4-bit quantization with LLM.nf4()
  • FP8: base model quantization in FP8
  • FP4_FP8: base model linear module quantization in FP4, mixed with experts and some special keys/layers in FP8.
maxSupportedContextLengthinteger

Capacity limits validated for this shape.

minTotalAcceleratorCountinteger

Minimum total accelerator count (world size, across all data-parallel replicas) this shape has been validated at. Written from shape-validation CI evidence; do not hand-edit downward.

Response

A successful response.

namestring
displayNamestring

Human-readable display name of the training shape. e.g. "Llama3 70B H200 BF16" Must be fewer than 64 characters long.

descriptionstring

The description of the training shape. Must be fewer than 1000 characters long.

createTimestring date-time

The creation time of the training shape.

updateTimestring date-time

The update time for the training shape.

baseModelstring required
deploymentShapeVersionstring
trainerImageTagstring required

The validated trainer runtime image tag used for numerics verification.

trainerMode'TRAINER_MODE_UNSPECIFIED' | 'POLICY_TRAINER' | 'FORWARD_ONLY' | 'LORA_TRAINER'

Trainer execution mode used for validated launch-profile matching.

nodeCountinteger

Node count validated for the launch profile.

modelTypestring

The model type of the base model (e.g. "llama", "qwen").

parameterCountstring int64

The parameter count of the base model, used for billion-parameter bucket matching.

acceleratorType'ACCELERATOR_TYPE_UNSPECIFIED' | 'NVIDIA_A100_80GB' | 'NVIDIA_H100_80GB' | 'AMD_MI300X_192GB' | 'NVIDIA_A10G_24GB' | 'NVIDIA_A100_40GB' | 'NVIDIA_L4_24GB' | 'NVIDIA_H200_141GB' | 'NVIDIA_B200_180GB' | 'AMD_MI325X_256GB' | 'AMD_MI350X_288GB' | 'NVIDIA_B300_288GB' | 'NVIDIA_GB200' | 'NVIDIA_GB300'
acceleratorCountinteger

Total number of accelerators used by the job.

baseModelWeightPrecision'WEIGHT_PRECISION_UNSPECIFIED' | 'BFLOAT16' | 'INT8' | 'NF4' | 'FP8' | 'FP4_FP8'

The weight precision for model training/inference.

  • BFLOAT16: no quantization applied
  • INT8: enable 8-bit quantization with LLM.int8()
  • NF4: enable 4-bit quantization with LLM.nf4()
  • FP8: base model quantization in FP8
  • FP4_FP8: base model linear module quantization in FP4, mixed with experts and some special keys/layers in FP8.
maxSupportedContextLengthinteger

Capacity limits validated for this shape.

minTotalAcceleratorCountinteger

Minimum total accelerator count (world size, across all data-parallel replicas) this shape has been validated at. Written from shape-validation CI evidence; do not hand-edit downward.

Changes

No recorded changes to this operation across all 1 revision of this API.