mirror of
https://github.com/meta-llama/llama-stack.git
synced 2025-07-29 07:14:20 +00:00
cleanup
This commit is contained in:
Ashwin Bharambe
parent
6d6c07b882
commit
f6b2b2fb39
13 changed files with 20 additions and 2286 deletions
23
README.md
23
README.md
|
|
@ -2,9 +2,26 @@ This repo contains the API specifications for various parts of the Llama Stack.
|
|||
The Stack consists of toolchain-apis and agentic-apis.
|
||||
|
||||
The tool chain apis that are covered --
|
||||
- chat_completion
|
||||
- batch inference
|
||||
- fine tuning
|
||||
- inference / batch inference
|
||||
- post training
|
||||
- reward model scoring
|
||||
- synthetic data generation
|
||||
|
||||
|
||||
### Generate OpenAPI specs
|
||||
|
||||
Set up virtual environment
|
||||
|
||||
```
|
||||
python3.9 -m venv ~/.venv/toolchain/
|
||||
source ~/.venv/toolchain/bin/activate
|
||||
|
||||
with-proxy pip3 install -r requirements.txt
|
||||
|
||||
```
|
||||
|
||||
Run the generate.sh script
|
||||
|
||||
```
|
||||
cd source && sh generate.sh
|
||||
```
|
||||
|
|
|
|||
|
|
@ -1,167 +0,0 @@
|
|||
openapi: 3.0.0
|
||||
info:
|
||||
title: Batch Inference API
|
||||
version: 0.0.1
|
||||
paths:
|
||||
/batch_inference/submit_job:
|
||||
post:
|
||||
summary: Submit a batch inference job
|
||||
description: |
|
||||
This endpoint allows clients to submit a batch inference job using a model and a prompt file.
|
||||
requestBody:
|
||||
required: true
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
model:
|
||||
type: string
|
||||
description: "The model identifier to be used for inference."
|
||||
prompt_file_path:
|
||||
$ref: '#/components/schemas/Path'
|
||||
description: "Path to a JSONL file where each line is a JSON-encoded list of messages."
|
||||
options:
|
||||
$ref: '#/components/schemas/Options'
|
||||
num_generations:
|
||||
type: integer
|
||||
description: "Number of generations to produce."
|
||||
responses:
|
||||
'200':
|
||||
description: Batch inference job successfully submitted
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/BatchInference'
|
||||
|
||||
/batch_inference/job_status:
|
||||
get:
|
||||
summary: Get status for an already submitted job
|
||||
description: |
|
||||
Retrieve the status and details of a previously submitted batch inference job using its unique job ID.
|
||||
parameters:
|
||||
- in: query
|
||||
name: job_id
|
||||
schema:
|
||||
type: string
|
||||
required: true
|
||||
description: "Unique identifier for the batch inference job."
|
||||
responses:
|
||||
'200':
|
||||
description: Batch inference job status retrieved successfully
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/BatchInference'
|
||||
|
||||
components:
|
||||
schemas:
|
||||
Message:
|
||||
type: object
|
||||
properties:
|
||||
role:
|
||||
type: string
|
||||
text:
|
||||
type: string
|
||||
attachments:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/MediaAttachment'
|
||||
eot:
|
||||
type: boolean
|
||||
description: "End of transmission flag."
|
||||
tool_call:
|
||||
type: boolean
|
||||
description: "Indicates if it's a tool call - builtin, custom, or ipython."
|
||||
is_complete:
|
||||
type: boolean
|
||||
description: "For streaming, indicates if the message is complete."
|
||||
is_header_complete:
|
||||
type: boolean
|
||||
description: "For streaming, indicates if the header of the message is complete."
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
description: "Additional metadata as JSON."
|
||||
|
||||
MediaAttachment:
|
||||
type: object
|
||||
properties:
|
||||
attachment_type:
|
||||
$ref: '#/components/schemas/MediaAttachmentType'
|
||||
data_type:
|
||||
$ref: '#/components/schemas/MediaAttachmentDataType'
|
||||
data:
|
||||
type: string
|
||||
|
||||
MediaAttachmentType:
|
||||
type: string
|
||||
enum:
|
||||
- image
|
||||
- video
|
||||
- audio
|
||||
- text
|
||||
description: "Type of media attachment."
|
||||
|
||||
MediaAttachmentDataType:
|
||||
type: string
|
||||
enum:
|
||||
- raw_bytes
|
||||
- filepath
|
||||
- uri
|
||||
description: "Data type of the media attachment."
|
||||
|
||||
BatchInference:
|
||||
type: object
|
||||
properties:
|
||||
job_id:
|
||||
type: string
|
||||
description: "ID provided by the API for the job."
|
||||
created:
|
||||
type: string
|
||||
format: date-time
|
||||
description: "Timestamp when the job was created."
|
||||
status:
|
||||
type: string
|
||||
enum:
|
||||
- validating
|
||||
- running
|
||||
- completed
|
||||
- failed
|
||||
description: "Current status of the job."
|
||||
input_file_path:
|
||||
$ref: '#/components/schemas/Path'
|
||||
success_file_path:
|
||||
$ref: '#/components/schemas/Path'
|
||||
error_file_path:
|
||||
$ref: '#/components/schemas/Path'
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
description: "Additional metadata related to the job."
|
||||
|
||||
Options:
|
||||
type: object
|
||||
properties:
|
||||
logprobs:
|
||||
type: boolean
|
||||
max_tokens:
|
||||
type: integer
|
||||
temperature:
|
||||
type: number
|
||||
top_p:
|
||||
type: number
|
||||
|
||||
Path:
|
||||
type: object
|
||||
properties:
|
||||
value:
|
||||
type: string
|
||||
description: "The path value."
|
||||
type:
|
||||
type: string
|
||||
enum:
|
||||
- raw_bytes
|
||||
- filepath
|
||||
- uri
|
||||
description: "Data Type of the path."
|
||||
|
|
@ -1,140 +0,0 @@
|
|||
openapi: 3.0.0
|
||||
info:
|
||||
title: Chat Completion API
|
||||
version: 0.0.1
|
||||
paths:
|
||||
/chat_completion/:
|
||||
post:
|
||||
summary: Submit a chat completion request
|
||||
description: |
|
||||
This endpoint allows clients to submit a chat completion request.
|
||||
requestBody:
|
||||
required: true
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
messages:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/Message'
|
||||
model:
|
||||
type: string
|
||||
options:
|
||||
$ref: '#/components/schemas/Options'
|
||||
n_completions:
|
||||
type: integer
|
||||
responses:
|
||||
'200':
|
||||
description: Successful response
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
id:
|
||||
type: string
|
||||
candidates:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/Completion'
|
||||
model_called:
|
||||
type: string
|
||||
usage:
|
||||
$ref: '#/components/schemas/TokenUsage'
|
||||
components:
|
||||
schemas:
|
||||
Message:
|
||||
type: object
|
||||
properties:
|
||||
role:
|
||||
type: string
|
||||
text:
|
||||
type: string
|
||||
attachments:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/MediaAttachment'
|
||||
eot:
|
||||
type: boolean
|
||||
description: "End of transmission flag."
|
||||
tool_call:
|
||||
type: boolean
|
||||
description: "Indicates if it's a tool call - builtin, custom, or ipython."
|
||||
is_complete:
|
||||
type: boolean
|
||||
description: "For streaming, indicates if the message is complete."
|
||||
is_header_complete:
|
||||
type: boolean
|
||||
description: "For streaming, indicates if the header of the message is complete."
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
description: "Additional metadata as JSON."
|
||||
MediaAttachment:
|
||||
type: object
|
||||
properties:
|
||||
attachment_type:
|
||||
$ref: '#/components/schemas/MediaAttachmentType'
|
||||
data_type:
|
||||
$ref: '#/components/schemas/MediaAttachmentDataType'
|
||||
data:
|
||||
type: string
|
||||
MediaAttachmentType:
|
||||
type: string
|
||||
enum:
|
||||
- image
|
||||
- video
|
||||
- audio
|
||||
- text
|
||||
description: "Type of media attachment."
|
||||
MediaAttachmentDataType:
|
||||
type: string
|
||||
enum:
|
||||
- raw_bytes
|
||||
- filepath
|
||||
- uri
|
||||
description: "Data type of the media attachment."
|
||||
Completion:
|
||||
type: object
|
||||
properties:
|
||||
id:
|
||||
type: string
|
||||
message:
|
||||
$ref: '#/components/schemas/Message'
|
||||
tokens:
|
||||
type: array
|
||||
items:
|
||||
type: integer
|
||||
logprobs:
|
||||
type: array
|
||||
items:
|
||||
type: number
|
||||
finish_reason:
|
||||
type: string
|
||||
enum:
|
||||
- stop
|
||||
- safety
|
||||
- max-length
|
||||
description: "Reason for completion termination."
|
||||
Options:
|
||||
type: object
|
||||
properties:
|
||||
logprobs:
|
||||
type: boolean
|
||||
max_tokens:
|
||||
type: integer
|
||||
temperature:
|
||||
type: number
|
||||
top_p:
|
||||
type: number
|
||||
TokenUsage:
|
||||
type: object
|
||||
properties:
|
||||
input_tokens:
|
||||
type: integer
|
||||
output_tokens:
|
||||
type: integer
|
||||
total_tokens:
|
||||
type: integer
|
||||
266
fine_tuning.yaml
266
fine_tuning.yaml
|
|
@ -1,266 +0,0 @@
|
|||
openapi: 3.0.0
|
||||
info:
|
||||
title: Fine Tuning API
|
||||
version: 0.0.1
|
||||
description: API for managing fine tuning jobs for machine learning models.
|
||||
|
||||
paths:
|
||||
/fine_tuning/jobs/submit:
|
||||
post:
|
||||
summary: Submit a fine tuning job
|
||||
description: Submit a fine tuning job with the specified configuration.
|
||||
requestBody:
|
||||
required: true
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/Config'
|
||||
responses:
|
||||
200:
|
||||
description: Successfully submitted the fine tuning job.
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/FineTuningJob'
|
||||
|
||||
/fine_tuning/jobs/status:
|
||||
get:
|
||||
summary: Gets last N fine tuning jobs
|
||||
description: Retrieve the status of the last N fine tuning jobs based on the provided job ID.
|
||||
parameters:
|
||||
- in: query
|
||||
name: job_id
|
||||
schema:
|
||||
type: string
|
||||
required: true
|
||||
description: The ID of the job to retrieve status for.
|
||||
responses:
|
||||
200:
|
||||
description: Successfully retrieved the job status.
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/FineTuningJob'
|
||||
|
||||
/fine_tuning/jobs/cancel:
|
||||
post:
|
||||
summary: Cancel provided job
|
||||
description: Cancel the fine tuning job with the specified job ID.
|
||||
requestBody:
|
||||
required: true
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
job_id:
|
||||
type: string
|
||||
responses:
|
||||
200:
|
||||
description: Successfully cancelled the fine tuning job.
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/FineTuningJob'
|
||||
|
||||
/fine_tuning/jobs/tail:
|
||||
get:
|
||||
summary: Tail logs of a particular job
|
||||
description: Stream the logs of a particular job in real-time. This endpoint supports streaming responses.
|
||||
parameters:
|
||||
- in: query
|
||||
name: job_id
|
||||
schema:
|
||||
type: string
|
||||
required: true
|
||||
description: The ID of the job to tail logs for.
|
||||
responses:
|
||||
200:
|
||||
description: Streaming logs in real-time.
|
||||
content:
|
||||
application/x-ndjson:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
logs:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/Log'
|
||||
headers:
|
||||
Content-Type:
|
||||
schema:
|
||||
type: string
|
||||
default: 'application/x-ndjson'
|
||||
Transfer-Encoding:
|
||||
schema:
|
||||
type: string
|
||||
default: 'chunked'
|
||||
|
||||
components:
|
||||
schemas:
|
||||
Message:
|
||||
# keep in sync with /chat_completion
|
||||
|
||||
TrainingDataItem:
|
||||
type: object
|
||||
properties:
|
||||
dialog:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/Message'
|
||||
keep_loss:
|
||||
type: array
|
||||
items:
|
||||
type: boolean
|
||||
|
||||
WandBLogger:
|
||||
type: object
|
||||
properties:
|
||||
project:
|
||||
type: string
|
||||
description: The project name in WandB where logs will be stored.
|
||||
|
||||
DiskLogger:
|
||||
type: object
|
||||
properties:
|
||||
filename:
|
||||
type: string
|
||||
description: The filename where logs will be stored on disk.
|
||||
|
||||
FullFineTuneOptions:
|
||||
type: object
|
||||
properties:
|
||||
enable_activation_checkpointing:
|
||||
type: boolean
|
||||
default: true
|
||||
memory_efficient_fsdp_wrap:
|
||||
type: boolean
|
||||
default: true
|
||||
fsdp_cpu_offload:
|
||||
type: boolean
|
||||
default: true
|
||||
|
||||
LoraFineTuneOptions:
|
||||
type: object
|
||||
properties:
|
||||
lora_attn_modules:
|
||||
type: array
|
||||
items:
|
||||
type: string
|
||||
apply_lora_to_mlp:
|
||||
type: boolean
|
||||
default: false
|
||||
apply_lora_to_output:
|
||||
type: boolean
|
||||
default: false
|
||||
lora_rank:
|
||||
type: integer
|
||||
lora_alpha:
|
||||
type: integer
|
||||
|
||||
FineTuningOptions:
|
||||
type: object
|
||||
properties:
|
||||
n_epochs:
|
||||
type: integer
|
||||
batch_size:
|
||||
type: integer
|
||||
lr:
|
||||
type: number
|
||||
format: float
|
||||
gradient_accumulation_steps:
|
||||
type: integer
|
||||
seed:
|
||||
type: integer
|
||||
shuffle:
|
||||
type: boolean
|
||||
custom_training_options:
|
||||
oneOf:
|
||||
- $ref: '#/components/schemas/FullFineTuneOptions'
|
||||
- $ref: '#/components/schemas/LoraFineTuneOptions'
|
||||
discriminator:
|
||||
propertyName: finetuning_type
|
||||
extras:
|
||||
# json to put other config overrides that are required by torchtune
|
||||
type: object
|
||||
additionalProperties: true
|
||||
|
||||
Config:
|
||||
type: object
|
||||
properties:
|
||||
model:
|
||||
type: string
|
||||
description: The model identifier that you want to fine tune.
|
||||
data:
|
||||
type: string
|
||||
format: uri
|
||||
description: Path to the JSONL file with each row representing a TrainingDataItem.
|
||||
validation_data:
|
||||
type: string
|
||||
format: uri
|
||||
description: Path to the JSONL file used for validation metrics.
|
||||
fine_tuning_options:
|
||||
$ref: '#/components/schemas/FineTuningOptions'
|
||||
logger:
|
||||
oneOf:
|
||||
- $ref: '#/components/schemas/DiskLogger'
|
||||
- $ref: '#/components/schemas/WandBLogger'
|
||||
discriminator:
|
||||
propertyName: log_type
|
||||
overrides:
|
||||
# eg. --nproc_per_node 4 instead of default that we need to pass through to torchrun
|
||||
# when running locally
|
||||
type: string
|
||||
description: Custom override options for the fine tuning process.
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
|
||||
FineTuningJob:
|
||||
type: object
|
||||
properties:
|
||||
job_id:
|
||||
type: string
|
||||
description: Unique identifier for the fine tuning job.
|
||||
created:
|
||||
type: string
|
||||
format: date-time
|
||||
description: The creation date and time of the job.
|
||||
finished_at:
|
||||
type: string
|
||||
format: date-time
|
||||
description: The completion date and time of the job.
|
||||
status:
|
||||
type: string
|
||||
enum: [validation, queued, running, failed, success, cancelled]
|
||||
description: The current status of the job.
|
||||
error_path:
|
||||
type: string
|
||||
format: uri
|
||||
description: Path to the error log file.
|
||||
checkpoints:
|
||||
type: array
|
||||
items:
|
||||
type: string
|
||||
format: uri
|
||||
description: List of paths to checkpoint files for various epochs.
|
||||
logs:
|
||||
type: string
|
||||
format: uri
|
||||
description: Path to the logs, either local or a WandB URI.
|
||||
input_config:
|
||||
$ref: '#/components/schemas/Config'
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
|
||||
Log:
|
||||
type: object
|
||||
properties:
|
||||
message:
|
||||
type: string
|
||||
description: The log message.
|
||||
timestamp:
|
||||
type: string
|
||||
format: date-time
|
||||
description: The timestamp of the log message.
|
||||
1018
openapi/spec.yaml
1018
openapi/spec.yaml
File diff suppressed because it is too large
Load diff
|
|
@ -1,162 +0,0 @@
|
|||
openapi: 3.0.0
|
||||
info:
|
||||
title: Reward Model Service API
|
||||
version: 0.0.1
|
||||
paths:
|
||||
/reward_model_scoring/:
|
||||
post:
|
||||
summary: Score a prompt-response pair using a reward model
|
||||
description: |
|
||||
This endpoint scores a given prompt-response pair using a specified reward model and scoring function.
|
||||
requestBody:
|
||||
required: true
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
prompt_with_response:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/Message'
|
||||
description: "Prompt and response joined as a list of messages."
|
||||
reward_model:
|
||||
type: string
|
||||
description: "Identifier for the reward model to be used."
|
||||
scoring_function:
|
||||
$ref: '#/components/schemas/ScoringFunction'
|
||||
options:
|
||||
$ref: '#/components/schemas/Options'
|
||||
responses:
|
||||
'200':
|
||||
description: Scoring completed successfully
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
id:
|
||||
type: string
|
||||
logprob:
|
||||
type: number
|
||||
format: float
|
||||
score:
|
||||
type: number
|
||||
format: float
|
||||
|
||||
/reward_model_scoring/submit_job/:
|
||||
post:
|
||||
summary: Batch scoring using reward models
|
||||
description: |
|
||||
Submit a batch job for scoring multiple prompt-response pairs using a reward model.
|
||||
requestBody:
|
||||
required: true
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
reward_model:
|
||||
type: string
|
||||
description: "Identifier for the reward model to be used."
|
||||
prompt_with_response_path:
|
||||
$ref: '#/components/schemas/Path'
|
||||
description: "Path to a JSONL file where each line is a List[Message] and custom_id."
|
||||
scoring_function:
|
||||
$ref: '#/components/schemas/ScoringFunction'
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
description: "Metadata to carry forward in the response."
|
||||
responses:
|
||||
'200':
|
||||
description: Batch scoring job successfully submitted
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/BatchRewardModelScoringJob'
|
||||
|
||||
/reward_model_scoring/submit_job/job_status:
|
||||
get:
|
||||
summary: Get status for an already submitted job
|
||||
description: |
|
||||
Retrieve the status and details of a previously submitted batch scoring job using its unique job ID.
|
||||
parameters:
|
||||
- in: query
|
||||
name: job_id
|
||||
schema:
|
||||
type: string
|
||||
required: true
|
||||
description: "Unique identifier for the batch scoring job."
|
||||
responses:
|
||||
'200':
|
||||
description: Batch scoring job status retrieved successfully
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/BatchRewardModelScoringJob'
|
||||
|
||||
components:
|
||||
schemas:
|
||||
Message:
|
||||
# reuse from /chat_completion
|
||||
|
||||
Options:
|
||||
type: object
|
||||
properties:
|
||||
logprobs:
|
||||
type: boolean
|
||||
max_tokens:
|
||||
type: integer
|
||||
temperature:
|
||||
type: number
|
||||
top_p:
|
||||
type: number
|
||||
# TODO: Add/Remove more reward model specific params
|
||||
|
||||
ScoringFunction:
|
||||
type: object
|
||||
properties:
|
||||
name:
|
||||
type: string
|
||||
params:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
|
||||
Path:
|
||||
type: object
|
||||
properties:
|
||||
value:
|
||||
type: string
|
||||
type:
|
||||
type: string
|
||||
enum:
|
||||
- raw_bytes
|
||||
- filepath
|
||||
- uri
|
||||
|
||||
BatchRewardModelScoringJob:
|
||||
type: object
|
||||
properties:
|
||||
job_id:
|
||||
type: string
|
||||
created:
|
||||
type: string
|
||||
format: date-time
|
||||
status:
|
||||
type: string
|
||||
enum:
|
||||
- validating
|
||||
- running
|
||||
- completed
|
||||
- failed
|
||||
input_file_path:
|
||||
$ref: '#/components/schemas/Path'
|
||||
success_file_path:
|
||||
$ref: '#/components/schemas/Path'
|
||||
error_file_path:
|
||||
$ref: '#/components/schemas/Path'
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
description: "Metadata carried forward from the job submission."
|
||||
|
|
@ -1,65 +0,0 @@
|
|||
== Schema ==
|
||||
Message:
|
||||
role: str
|
||||
text: str
|
||||
attachements: List[MediaAttachment]
|
||||
eot: bool
|
||||
tool_call: bool # if it's a tool call - builtin or custom or ipython
|
||||
# for streaming
|
||||
is_complete: bool
|
||||
is_header_complete: bool
|
||||
metadata: json
|
||||
|
||||
MediaAttachment:
|
||||
attachement_type: MediaAttachmentType
|
||||
data_type: MediaAttachmentDataType
|
||||
data: str
|
||||
|
||||
MediaAttachmentType: # enum [image, video, audio, text(or file)]
|
||||
MediaAttachmentDataType: # enum [raw_bytes, filepath, uri]
|
||||
|
||||
BatchInference:
|
||||
job_id: str # id provided by the api
|
||||
created: string # format - date-time
|
||||
status: string # enum (validating, running, completed, failed)
|
||||
input_file_path: Path # jsonl style file where each
|
||||
success_file_path: Path
|
||||
error_file_path: Path
|
||||
metadata: json
|
||||
|
||||
Options:
|
||||
logprobs: bool
|
||||
max_tokens: int
|
||||
temperature: float
|
||||
top_p: float
|
||||
|
||||
Path:
|
||||
value: string
|
||||
type: string # enum [raw_bytes, filepath, uri]
|
||||
|
||||
== Callsites ==
|
||||
|
||||
callsite:
|
||||
/batch_inference/submit_job
|
||||
request_type:
|
||||
post
|
||||
description:
|
||||
Submit a batch inference job
|
||||
request:
|
||||
model: str
|
||||
prompt_file_path: Path # jsonl style file where each line is a json encoded List[Message]
|
||||
options: Options
|
||||
num_generations: int
|
||||
response:
|
||||
batch_inference_job: BatchInference
|
||||
|
||||
callsite:
|
||||
/batch_inference/job_status
|
||||
request_type:
|
||||
get
|
||||
description:
|
||||
Get status for an already submitted job
|
||||
request:
|
||||
job_id: str # unique identifier for the job
|
||||
response:
|
||||
batch_inference_job: BatchInference
|
||||
|
|
@ -1,72 +0,0 @@
|
|||
# Simple bullet form for ease of read and iteration
|
||||
# Use LLMs to translate this to a OpenAPI spec.
|
||||
|
||||
== Schema ==
|
||||
|
||||
Message:
|
||||
role: str
|
||||
text: str
|
||||
attachements: List[MediaAttachment]
|
||||
eot: bool
|
||||
tool_call: bool # if it's a tool call - builtin or custom or ipython
|
||||
# for streaming
|
||||
is_complete: bool
|
||||
is_header_complete: bool
|
||||
metadata: json
|
||||
|
||||
MediaAttachment:
|
||||
attachement_type: MediaAttachmentType
|
||||
data_type: MediaAttachmentDataType
|
||||
data: str
|
||||
|
||||
MediaAttachmentType: # enum [image, video, audio, text(or file)]
|
||||
MediaAttachmentDataType: # enum [raw_bytes, filepath, uri]
|
||||
|
||||
Completion:
|
||||
id: str
|
||||
message: Message
|
||||
tokens: List[int]
|
||||
logprobs: List[floats]
|
||||
finish_reason: str # Enum (stop, safety, max-length, etc)
|
||||
|
||||
Options:
|
||||
logprobs: bool
|
||||
max_tokens: int
|
||||
temperature: float
|
||||
top_p: float
|
||||
#TODO: Get more options from metagen
|
||||
|
||||
TokenUsage:
|
||||
input_tokens: int
|
||||
output_tokens: int
|
||||
total_tokens: int
|
||||
|
||||
== Callsite ==
|
||||
|
||||
callsite:
|
||||
chat_completion/
|
||||
request_type:
|
||||
post
|
||||
description:
|
||||
submit a chat completion request
|
||||
request:
|
||||
messages: List[Message]
|
||||
model: str
|
||||
options: Options
|
||||
n_complections: int
|
||||
# TODO: how to handle tooling control if any ?
|
||||
# Add `tools` and `tool_choice` --
|
||||
# for eg. "auto": use model's guess
|
||||
# how to force to use particular tool
|
||||
# how to disbale inbuilt tools
|
||||
# tools: List[Tool]
|
||||
# tool_choice: Any
|
||||
response:
|
||||
id: str
|
||||
candidates: List[Completion] # a list to account for when n_completions > 1
|
||||
model_called: str # info on that model that produced this result
|
||||
usage: TokenUsage
|
||||
|
||||
# TODO
|
||||
# callsite:
|
||||
# chat_completion_stream/
|
||||
|
|
@ -1,134 +0,0 @@
|
|||
# Fine Tuning APIs
|
||||
== Schema ==
|
||||
|
||||
TrainingDataItem:
|
||||
dialog: List[Message]
|
||||
keep_loss: List[bool]
|
||||
|
||||
|
||||
WandBLogger:
|
||||
project: str
|
||||
|
||||
DiskLogger:
|
||||
# log_dir will be pre-configured in environment
|
||||
filename: str
|
||||
|
||||
FullFineTuneOptions:
|
||||
enable_activation_checkpointing: True
|
||||
memory_efficient_fsdp_wrap: True
|
||||
fsdp_cpu_offload: True
|
||||
|
||||
LoraFineTuneOptions:
|
||||
lora_attn_modules: ['q_proj', 'v_proj']
|
||||
apply_lora_to_mlp: False
|
||||
apply_lora_to_output: False
|
||||
lora_rank: 8
|
||||
lora_alpha: 16
|
||||
|
||||
FineTuningOptions:
|
||||
n_epochs: int
|
||||
batch_size: int
|
||||
lr: float
|
||||
gradient_accumulation_steps: int
|
||||
seed: int
|
||||
shuffle: bool
|
||||
|
||||
# Unions in OpenAPI with a reference field that can help disambiguate
|
||||
custom_training_options:
|
||||
discriminator:
|
||||
propertyName: fine_tuning_type
|
||||
mapping:
|
||||
fft: FullFineTuneOptions
|
||||
lora: LoraFineTuneOptions
|
||||
|
||||
# other options that can be passed in
|
||||
extras: json
|
||||
|
||||
Config:
|
||||
model: str # model that you want to fine tune
|
||||
data: Path # jsonl with each row representing a TrainingDataItem
|
||||
validation_data: Path # same as data but to get validation metrics on
|
||||
|
||||
# fine tuning args
|
||||
fine_tuning_options: FineTuningOptions
|
||||
|
||||
# metric logging
|
||||
logger:
|
||||
discriminator:
|
||||
propertyName: log_type
|
||||
mapping:
|
||||
disk: DiskLogger
|
||||
wandb: WandBLogger
|
||||
|
||||
# Override options
|
||||
# eg. --nproc_per_node 4 insted of defaults,
|
||||
# this might be impl specific and can allow for various customizations
|
||||
overrides: str
|
||||
metadata: json # to carry over to job details
|
||||
|
||||
FineTuningJob:
|
||||
job_id: str
|
||||
created: str # format date-time
|
||||
finished_at: str # format date-time
|
||||
status: str # enum - validation, queued, running, failed, success, cancelled
|
||||
error_path: Path # error logging
|
||||
checkpoints: List[Path] # checkpoints for various epochs
|
||||
logs: Path # local path / wandb uri
|
||||
input_config: Config # config used to submit this job
|
||||
metadata: json # carried over rom user provided input
|
||||
|
||||
Log:
|
||||
message: string # The log message.
|
||||
timestamp: string # format: date-time
|
||||
|
||||
== Callsites ==
|
||||
|
||||
callsite:
|
||||
/fine_tuning/jobs/submit
|
||||
request_type:
|
||||
post
|
||||
description:
|
||||
Submit a fine tuning job
|
||||
request:
|
||||
config: Config
|
||||
response:
|
||||
fine_tuning_job: FineTuningJob
|
||||
|
||||
|
||||
callsite:
|
||||
/fine_tuning/jobs/status
|
||||
request_type:
|
||||
get
|
||||
description:
|
||||
Gets last N fine tuning jobs
|
||||
request:
|
||||
job_id: str
|
||||
response:
|
||||
fine_tuning_job: FineTuningJob
|
||||
|
||||
|
||||
callsite:
|
||||
/fine_tuning/jobs/cancel
|
||||
request_type:
|
||||
post
|
||||
description:
|
||||
Cancel provided job
|
||||
request:
|
||||
job_id: str
|
||||
response:
|
||||
fine_tuning_job: FineTuningJob
|
||||
|
||||
|
||||
callsite:
|
||||
/fine_tuning/jobs/tail
|
||||
request_type:
|
||||
get
|
||||
description:
|
||||
Tail logs of a particular job
|
||||
request:
|
||||
job_id: str
|
||||
response:
|
||||
logs: List[Log]
|
||||
streaming:
|
||||
enabled: True
|
||||
chunkSize: 1024
|
||||
|
|
@ -1,70 +0,0 @@
|
|||
# Reward Model Service
|
||||
== Schema ==
|
||||
Message:
|
||||
# Same as /chat_completion
|
||||
|
||||
Options:
|
||||
logprobs: bool
|
||||
max_tokens: int
|
||||
temperature: float
|
||||
top_p: float
|
||||
#TODO: Figure out what other reward model specific params
|
||||
|
||||
ScoringFunction:
|
||||
name: str
|
||||
params: json
|
||||
|
||||
BatchRewardModelScoringJob:
|
||||
job_id: str
|
||||
created: str # format date-time
|
||||
status: string # enum (validating, running, completed, failed)
|
||||
input_file_path: Path
|
||||
success_file_path: Path # jsonl where each row has {custom_id: <from input>, logprob: <float>, score: float}
|
||||
error_file_path: Path # jsonl where each row has {custom_id: <from input>, error: <error_info>}
|
||||
metadata: json # carry forward from job submission api
|
||||
|
||||
|
||||
== Callsites ==
|
||||
|
||||
callsite:
|
||||
reward_model_scoring/
|
||||
request_type:
|
||||
post
|
||||
description:
|
||||
Score a prompt-response pair using a reward model
|
||||
request:
|
||||
prompt_with_response: List[Message] # prompt and response joined as a List[Message]
|
||||
reward_model: str
|
||||
scoring_function: ScoringFunction
|
||||
options: Options
|
||||
response:
|
||||
id: str
|
||||
logprob: float
|
||||
score: float
|
||||
|
||||
|
||||
callsite:
|
||||
reward_model_scoring/submit_job/
|
||||
request_type:
|
||||
post
|
||||
description:
|
||||
Batch scoring using reward models
|
||||
request:
|
||||
reward_model: str
|
||||
prompt_with_response_path: Path # jsonl file where each line is a List[Message] and custom_id
|
||||
scoring_function: ScoringFunction
|
||||
metadata: json # anything to carry forward over in the response
|
||||
response:
|
||||
batch_reward_model_scoring_job: BatchRewardModelScoringJob
|
||||
|
||||
|
||||
callsite:
|
||||
/reward_model_scoring/submit_job/job_status
|
||||
request_type:
|
||||
get
|
||||
description:
|
||||
Get status for an already submitted job
|
||||
request:
|
||||
job_id: str # unique identifier for the job
|
||||
response:
|
||||
batch_reward_model_scoring_job: BatchRewardModelScoringJob
|
||||
|
|
@ -1,58 +0,0 @@
|
|||
# Synthetic Data Generation API
|
||||
== Schema ==
|
||||
|
||||
FilteringFunction:
|
||||
name: str
|
||||
params: json
|
||||
|
||||
SyntheticDataPoint:
|
||||
custom_id: str
|
||||
index: int
|
||||
prompt: List[Message]
|
||||
response: Message
|
||||
logprob: float
|
||||
score: float
|
||||
|
||||
SyntheticDataGenerationJob:
|
||||
job_id: str # id provided by the api
|
||||
created: string # format - date-time
|
||||
status: string # enum (validating, running, completed, failed)
|
||||
input_file_path: Path # jsonl style file where each row contains custom_id and message_list
|
||||
success_file_path: Path # jsonl each line is SyntheticDataPoint
|
||||
error_file_path: Path # custom_ids where we failed with some info
|
||||
metadata: json
|
||||
|
||||
== Callsites ==
|
||||
|
||||
callsite:
|
||||
/synthetic_data_gen/submit_job
|
||||
request_type:
|
||||
post
|
||||
description:
|
||||
Submit a job to generate synthetic data using llm + reward model scoring + filtering
|
||||
request:
|
||||
# batch inference params
|
||||
model: str
|
||||
prompt_file_path: Path # jsonl style file where each line is a json encoded List[Message] + custom_id
|
||||
options: Options
|
||||
num_generations: int
|
||||
# reward model scoring params
|
||||
reward_model: str
|
||||
scoring_function: ScoringFunction
|
||||
# filtering params
|
||||
filtering_function: FilteringFunction
|
||||
metadata: json
|
||||
|
||||
response:
|
||||
synth_data_gen_job: SyntheticDataGenerationJob
|
||||
|
||||
callsite:
|
||||
/synthetic_data_gen/job_status
|
||||
request_type:
|
||||
get
|
||||
description:
|
||||
Get status for an already submitted job
|
||||
request:
|
||||
job_id: str # unique identifier for the job
|
||||
response:
|
||||
synth_data_gen_job: SyntheticDataGenerationJob
|
||||
|
|
@ -1,131 +0,0 @@
|
|||
openapi: 3.0.0
|
||||
info:
|
||||
title: Synthetic Data Generation API
|
||||
version: 0.0.1
|
||||
paths:
|
||||
/synthetic_data_gen/submit_job:
|
||||
post:
|
||||
summary: Submit a job to generate synthetic data
|
||||
description: Submit a job to generate synthetic data using llm + reward model scoring + filtering
|
||||
requestBody:
|
||||
required: true
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
type: object
|
||||
properties:
|
||||
model:
|
||||
type: string
|
||||
description: Model used for batch inference
|
||||
prompt_file_path:
|
||||
type: string
|
||||
format: path
|
||||
description: Path to the JSONL file containing message_lists and custom IDs
|
||||
options:
|
||||
$ref: '#/components/schemas/Options'
|
||||
num_generations:
|
||||
type: integer
|
||||
description: Number of generations to produce
|
||||
reward_model:
|
||||
type: string
|
||||
description: Model used for scoring
|
||||
scoring_function:
|
||||
$ref: '#/components/schemas/ScoringFunction'
|
||||
filtering_function:
|
||||
$ref: '#/components/schemas/FilteringFunction'
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
description: Additional metadata for the job
|
||||
responses:
|
||||
'200':
|
||||
description: Job successfully submitted
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/SyntheticDataGenerationJob'
|
||||
|
||||
/synthetic_data_gen/job_status:
|
||||
get:
|
||||
summary: Get job status
|
||||
description: Get status for an already submitted job
|
||||
parameters:
|
||||
- in: query
|
||||
name: job_id
|
||||
schema:
|
||||
type: string
|
||||
required: true
|
||||
description: Unique identifier for the job
|
||||
responses:
|
||||
'200':
|
||||
description: Job status retrieved successfully
|
||||
content:
|
||||
application/json:
|
||||
schema:
|
||||
$ref: '#/components/schemas/SyntheticDataGenerationJob'
|
||||
components:
|
||||
schemas:
|
||||
FilteringFunction:
|
||||
type: object
|
||||
properties:
|
||||
name:
|
||||
type: string
|
||||
description: Name of the filtering function
|
||||
params:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
description: JSON object containing parameters for the filtering function
|
||||
SyntheticDataPoint:
|
||||
type: object
|
||||
properties:
|
||||
custom_id:
|
||||
type: string
|
||||
description: Custom identifier for the data point
|
||||
index:
|
||||
type: integer
|
||||
description: Index of the data point
|
||||
prompt:
|
||||
type: array
|
||||
items:
|
||||
$ref: '#/components/schemas/Message'
|
||||
description: List of messages used as prompt
|
||||
response:
|
||||
$ref: '#/components/schemas/Message'
|
||||
logprob:
|
||||
type: number
|
||||
format: float
|
||||
description: Log probability of the response
|
||||
score:
|
||||
type: number
|
||||
format: float
|
||||
description: Score of the response based on the reward model
|
||||
SyntheticDataGenerationJob:
|
||||
type: object
|
||||
properties:
|
||||
job_id:
|
||||
type: string
|
||||
description: ID provided by the API
|
||||
created:
|
||||
type: string
|
||||
format: date-time
|
||||
description: Timestamp when the job was created
|
||||
status:
|
||||
type: string
|
||||
enum: [validating, running, completed, failed]
|
||||
description: Current status of the job
|
||||
input_file_path:
|
||||
type: string
|
||||
format: path
|
||||
description: Path to the input JSONL file
|
||||
success_file_path:
|
||||
type: string
|
||||
format: path
|
||||
description: Path to the JSONL file containing successful results
|
||||
error_file_path:
|
||||
type: string
|
||||
format: path
|
||||
description: Path to the JSONL file containing errors
|
||||
metadata:
|
||||
type: object
|
||||
additionalProperties: true
|
||||
description: Additional metadata about the job
|
||||
Loading…
Add table
Add a link
Reference in a new issue