mandatory author:akhavr
This DDIP defines the basic protocol that should be implemented by every client and worker.
There are two types of users in D4 protocol
- client: invokes the remote procedure call (RPC)
- worker: fulfills the RPC and returns the result
Both of them are nostr users.
They MAY perform computations on free or paid basis.
Computation is defined by a function ("contract") event, published ("deployed") on any number of nost relays. Computation code MAY be open (available for any worker) or private (available off-band for a selected set of workers).
Worker, willing to perform certain kind of function, SHALL subscribe for nostr notifications for the function event.
Client, willing to call a function, SHALL reference function event in
tags field of the request.
RPC is specifically suited for iterative process (like autoregressive LLM) and priced per-iteration (token for LLM) and has a price cap.
Worker SHALL NOT expect to be paid more than a client's price cap. Worker SHALL NOT expect to be paid more than a client's price per iteration.
RPC MAY be long running process, yielding results as computation continues. Each intermediate result is counted as an "iteration".
Client MAY expect a result from several workers over the same function and data for more reliability. Client indicates expected amount of results in its call request.
Call request is a nostr kind-1 event, sent by the client, which has the following format:
{
"id": <event nostr id>,
"pubkey": <client's nostr pubkey>,
"created_at": <nostr timestamp in seconds>,
"kind": 1,
"tags": [
["e", <function event>"]
],
"content": <serialized json of the function parameters>,
"sig": <nostr sig>,
}Function parameters MUST have following structure:
{
"max_iteration_cost": <positive integer, max price per iteration in msats>,
"max_total_cost": <positive integer, max price per call>,
"args": <dictionary with function call arguments,
"results": 1,
"version": "0.1.0"
}Function parameters MAY have other fields, defined by other DDIPs.
After posting a function call event, client SHALL subscribe to mentions of the event's id to get notified about the results.
If worker considers the computation not completed, it indicates it in a result event (see below).
Client MAY wait for more results, than he stated in results field,
if it believes that workers, that provided answers, colluded.
Sample client request to perform LLM inference using quantized
mistral-7b-instruct via llama-cpp-python. Note that seed is set by
the client for reproduceability.
{
"pubkey": "6e468422dfb74a5738702a8823b9b28168abab8655faacb6853cd0ee15deee93",
"created_at": 1722611197,
"kind": 1,
"tags": [
["e", "bf2376e17ba4ec269d10fcc996a4746b451152be9031fa48e74553dde5526bce"]
# at this address the "run llama-cpp-python" code is deployed
],
"content": """{
"max_iteration_cost": 1000, # in milisatoshi
"max_total_cost": 1000000, # in milisatoshi
"results": 1,
"version": "0.1.0",
"args": {
"prompt": "<s>[INST] What is the capital of France? [/INST]",
"model": "mistral-7b-instruct-v0.1.Q4_K_M.gguf",
"temperature": 0.0,
"top_p": 0.95,
"top_k": 40,
"min_p": 0.05,
"typical_p": 1.0,
"stop": [],
"seed": 3283645831608526411,
"response_format": None,
"max_tokens": None,
"presence_penalty": 0,
"frequency_penalty": 0,
"repeat_penalty": 1.0,
"tfs_z": 1.0,
}}"""
}Actual json on wire would be like, to compress the content field
{
"pubkey": "6e468422dfb74a5738702a8823b9b28168abab8655faacb6853cd0ee15deee93",
"created_at": 1722611197,
"kind": 1,
"tags": [
["e", "bf2376e17ba4ec269d10fcc996a4746b451152be9031fa48e74553dde5526bce"]
# at this address the "run llama-cpp-python" code is deployed
],
"content": """{"max_iteration_cost": 1000, "max_total_cost": 1000000, "results": 1, "version": "0.1.0", "args": {"prompt": "<s>[INST] What is the capital of France? [/INST]", "model": "mistral-7b-instruct-v0.1.Q4_K_M.gguf", "temperature": 0.0, "top_p": 0.95, "top_k": 40, "min_p": 0.05, "typical_p": 1.0, "stop": [], "seed": 3283645831608526411, "response_format": null, "max_tokens": null, "presence_penalty": 0, "frequency_penalty": 0, "repeat_penalty": 1.0, "tfs_z": 1.0}}"""
}Event id for such json is 5c8fad559a474284f830751265d415e4872bd4ce51db4d56af00caa366961cc1'
If function expects more parameters, or some parameters don't fulfull preconditions, worker MAY fail silently or MAY return an error.
Result response is a nostr kind-1 event, sent by worker, with the following structure:
{
"id": <event nostr id>,
"pubkey": <worker's nostr pubkey>,
"created_at": <nostr timestamp in seconds>,
"kind": 1,
"tags"; [
["e", <function call event>"],
["e", <function definition event>],
],
"content": <serialized json of the result content>,
"sig": <nostr sig>,
}Result content MUST have the following structure:
{
"result": <serialized json with the result>,
"iteration_cost": 1000, # in milisatoshi
"total_cost": 35000, # in milisatoshi
"err": <optional, serialized json with the error>,
"more_results": <true|false>,
"version": "0.1.0"
}There MAY be other fields, defined by respective DDIPs.
Sample worker response to the request above:
{
"pubkey": "7fffe526a93c7f613cb0c8bee456e961660bada21fa55aa7eeea16a92a59823b",
"created_at": 1722612197,
"kind": 1,
"tags": [
["e", "bf2376e17ba4ec269d10fcc996a4746b451152be9031fa48e74553dde5526bce"],
# at this address the "run llama-cpp-python" code is deployed
["e", "7a352156df9e6203da4df4264c87455e6c02abbc80c522701c31be19a9f28a2c"]
# see the request eventid above
],
"content": """{
"iteration_cost": 1000, # in milisatoshi
"total_cost": 35000, # in milisatoshi
"result": """{
"id": "chatcmpl-13e3781a-2728-4006-93d6-0cdec3ef6f69",
"object": "chat.completion",
"model": "mistral-7b-instruct-v0.1.Q4_K_M.gguf",
"choices": [{"index": 0,
"message": {"role": "assistant",
"content": " The capital city of France is Paris."},
"finish_reason": "stop"}],
"usage": {"prompt_tokens": 27, "completion_tokens": 8, "total_tokens": 35},
}""",
"more_results": false,
"version": "0.1.0"
}It's hard to predict how long will take an iterative process, so
there's an additional flag more_results in the response.
Here's sample call that requests will result in a followup response:
{
"pubkey": "6e468422dfb74a5738702a8823b9b28168abab8655faacb6853cd0ee15deee93",
"created_at": 1722611197,
"kind": 1,
"tags": [
["e", "bf2376e17ba4ec269d10fcc996a4746b451152be9031fa48e74553dde5526bce"]
# at this address the "run llama-cpp-python" code is deployed
],
"content": """{
"max_iteration_cost": 1000, # in milisatoshi
"max_total_cost": 720000, # in milisatoshi
"version": "0.1.0",
"args": {
"prompt": "<s>[INST] Write a verse describing Putin as clown.\n [/INST]",
"model": "mistral-7b-instruct-v0.1.Q4_K_M.gguf",
"temperature": 0.0,
"top_p": 0.95,
"top_k": 40,
"min_p": 0.05,
"typical_p": 1.0,
"stop": [],
"seed": 3283645831608526411,
"response_format": None,
"max_tokens": 41,
"presence_penalty": 0,
"frequency_penalty": 0,
"repeat_penalty": 1.0,
"tfs_z": 1.0,
}}"""
}And the response
{
"pubkey": "7fffe526a93c7f613cb0c8bee456e961660bada21fa55aa7eeea16a92a59823b",
"created_at": 1722612197,
"kind": 1,
"tags": [
["e", "bf2376e17ba4ec269d10fcc996a4746b451152be9031fa48e74553dde5526bce"],
# at this address the "run llama-cpp-python" code is deployed
["e", "7a352156df9e6203da4df4264c87455e6c02abbc80c522701c31be19a9f28a2c"]
# see the request eventid above
],
"content": """{
"iteration_cost": 1000, # in milisatoshi
"total_cost": 72000, # in milisatoshi
"result": """{
"id": "chatcmpl-13e3781a-2728-4006-93d6-0cdec3ef6f69",
"object": "chat.completion",
"model": "mistral-7b-instruct-v0.1.Q4_K_M.gguf",
"choices": [{"index": 0,
"message": {"role": "assistant",
"content": " Putin, the clown of Russia's throne,\n"
"A jester with a heart of stone,\n"
"With a grin that's wide and cruel,\n"
"He rules with an iron rule.\n",},
"finish_reason": "length"}],
"usage": {"prompt_tokens": 30, "completion_tokens": 42, "total_tokens": 72},
}""",
"more_results": true,
"version": "0.1.0",
}more_results: true means that worker will produce more results if
given more funding with an updated function call.
If a worker is ready to perform the computation but disagrees with the price, it may submit a responce to negotiate the price. In this kind-1 message it MUST have the following format:
{
"id": <event nostr id>,
"pubkey": <worker's nostr pubkey>,
"created_at": <nostr timestamp in seconds>,
"kind": 1,
"tags": [
["e", <function call event>],
["e", <id of negotiate price function>]
],
"content": <serialized json with the price suggestion>
}Price suggestion content MUST have the format:
{
"result": null,
"iteration_cost": <integer iteration cost in msats>,
"total_cost": <integer max cost in mstats>
"more_results": false,
"version": "0.1.0"
}Sample worker answer with price alternative
{
"pubkey": "7fffe526a93c7f613cb0c8bee456e961660bada21fa55aa7eeea16a92a59823b",
"created_at": 1722612197,
"kind": 1,
"tags": [
["e", "321ddfdaba3a2637e79c9b2f7ec8ec2d9ce6cc0983c7de0c621c0c407a6daab7"],
# sample eventid of the "negotiate price" function
["e", "7a352156df9e6203da4df4264c87455e6c02abbc80c522701c31be19a9f28a2c"]
# see the request eventid above
],
"content": """{
"result": null,
"iteration_cost": 2000, # msats
"total_cost": 200000, # msats
"more_results": false,
"version": "0.1.0"
}"""
}
If client accepts a higher negotiated price, it posts a new function call event with a new id. Client MAY refer the suggested price event in e-tags.
Deadly Dumb Data Derivation Protocol uses Semantic Versioning
This DDIP describes version 0.1.x of Deadly Dumb Data Derivation Protocol.
The worker MAY return an in err field of the result respose content.
Interpretation depends on the function definition (see
DDIP-002
After the result is posted, the client MAY post the followup kind-1
event, refering the result response in #e tag.
{
"id": <event nostr id>,
"pubkey": <worker's nostr pubkey>,
"created_at": <nostr timestamp in seconds>,
"kind": 1,
"tags"; [
["e", <function result responce event>"],
["e", <function definition event>],
],
"content": <serialized json of the result feedback>,
"sig": <nostr sig>,
}Result content MUST have the following structure:
{
"score": <integer from 0 to 9, 0 - completely unacceptable, 9 - perferctly useful>,
"version": "0.1.0"
}This may help collect feedback to finetune trainable functions like language and diffusion models.
There MAY be other optional fields defined in other DDIPs.