Developer documentation

One Triangle Prefill SDK

A dependency-free Python client for mapped prefill and target decode in one request. Use qwen3-8b or qwen3.8-27b. You never manually transfer a cache.

Install and call Request API access

01 / Quickstart

Install and configure

Python 3.9 or newer is required. Install the currently published wheel directly:

python -m pip install https://617b515ef135455fa22568d7c9d3af84-0.lambdaspaces.com/one-triangle/sdk/one_triangle-0.1.0-py3-none-any.whl

Set credentials in the process environment. API access details are supplied by One Triangle.

export ONE_TRIANGLE_API_KEY="<your-api-key>"
export ONE_TRIANGLE_BASE_URL="<your-gateway-url>"

api_key, base_url, and timeout can also be passed directly. Direct arguments override environment variables.

02 / Blocking calls

Call either supported model

prefill blocks until mapped prefill and target decode are complete. max_tokens defaults to 256.

Qwen3 8B

import one_triangle

result = one_triangle.prefill(
    model="qwen3-8b",
    context=context,
    max_tokens=256,
)

print(result.decode())

Qwen3.8 27B

import one_triangle

result = one_triangle.prefill(
    model="qwen3.8-27b",
    context=context,
    max_tokens=256,
)

print(result.decode())

03 / Streaming

Stream decoded 8B deltas

qwen3-8b supports streaming decoded text. qwen3.8-27b currently returns a complete blocking result.

import one_triangle

for delta in one_triangle.prefill_stream(
    model="qwen3-8b",
    context=context,
    max_tokens=256,
):
    print(delta, end="", flush=True)

04 / Result fields

Inspect the complete response

decode() / text: The decoded text.
usage: Upstream token and accounting metadata.
prefill: Upstream mapped-prefill metadata.
model: The response model, or the requested alias.
request_id: The request identifier when the gateway supplies one.
raw: The complete gateway response object.

05 / HTTP API

Call `/v1/prefill` directly

The JSON body contains exactly model, context, and max_tokens. Unknown or missing fields are rejected.

curl "$ONE_TRIANGLE_BASE_URL/v1/prefill" \
  -H "Authorization: Bearer $ONE_TRIANGLE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3-8b","context":"Long source context...","max_tokens":256}'

06 / Errors

Handle gateway failures

HTTP, gateway, transport, and invalid-response failures raise one_triangle.PrefillError. Gateway responses may provide status_code and body.

import one_triangle

try:
    result = one_triangle.prefill(model="qwen3-8b", context=context)
except one_triangle.PrefillError as error:
    print(error)
    print(error.status_code)
    print(error.body)

07 / Reference

Supported aliases and limits

Alias	Blocking	Streaming
qwen3-8b	Supported	Supported
qwen3.8-27b	Supported	Not currently available

max_tokens defaults to 256.
Customers call one prefill operation and never manually transfer a cache.
The SDK accepts decoded text from text, output, choices[0].text, or choices[0].message.content.

Request API access