lutflow 0.1.0

GPU Financial Firewall — budget enforcement for AI inference

Lutflow — GPU Financial Firewall

Budget enforcement for AI inference workloads. Track costs, enforce budgets, kill runaway GPU spending in real time.

pip install lutflow

5-Line Quickstart

from lutflow import Client
import openai

client = Client(tenant_id="acme", budget_usd=10.00)
wrapped = client.wrap(openai.OpenAI())
response = wrapped.chat.completions.create(model="gpt-4o", messages=[{"role": "user", "content": "Hello!"}])

Lutflow tracks every token, calculates cost, and terminates the process if the budget is exceeded.

CLI

# GPU price lookup (powered by TinyFish)
lutflow lookup --gpu nvidia-l4

# Deploy with budget enforcement
lutflow deploy --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 --budget 0.10

# Live dashboard
lutflow watch

# Check deployment status
lutflow status <deployment_id>

# Manual kill
lutflow kill <deployment_id> --reason overspend

Installation

pip install lutflow            # Core + CLI
pip install lutflow[openai]    # With OpenAI wrapper
pip install lutflow[anthropic] # With Anthropic wrapper
pip install lutflow[gemini]    # With Google Gemini wrapper
pip install lutflow[all]       # All providers + Kafka transport

SDK Usage

Token-Based Pricing (API Providers)

from lutflow import Client, BudgetStrategy

client = Client(
    tenant_id="my-tenant",
    budget_usd=10.00,
    on_budget_exceeded=BudgetStrategy.RAISE_ERROR,
)

wrapped = client.wrap(openai.OpenAI())
response = wrapped.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}],
)

print(f"Spent: ${client.accumulated_cost_usd:.4f}")
print(f"Remaining: ${client.remaining_budget_usd:.4f}")

Cloud Mode (with Lutflow API)

from lutflow import Client, BudgetStrategy

client = Client(
    tenant_id="my-tenant",
    api_key="lut_xxx",  # Enables cloud reporting
    budget_usd=10.00,
    on_budget_exceeded=BudgetStrategy.RAISE_ERROR,
)

GPU-Time Pricing (Self-Hosted Models)

from lutflow import Client, PricingMode

client = Client(
    tenant_id="my-tenant",
    budget_usd=5.00,
    pricing_mode=PricingMode.GPU_TIME,
    gpu_type="nvidia-l4",
)

client.start_gpu_timer()
# ... run inference ...
cost = client.stop_gpu_timer()

GPU Price Lookup (TinyFish)

from lutflow.tinyfish import TinyFishClient

tf = TinyFishClient()  # Uses TINYFISH_API_KEY env var
result = tf.fetch_all_prices(gpu_type="nvidia-l4")

best = result.cheapest
print(f"Cheapest: {best.provider} at ${best.spot_usd_per_hour:.2f}/hr")

Budget Strategies

Strategy	Behavior
RAISE_ERROR	Raises BudgetExceededError (default)
WARN_ONLY	Logs warning, continues execution
CALLBACK	Calls a custom function you provide
SELF_KILL	Sends SIGKILL to the process

Supported Providers

• OpenAI — GPT-4o, GPT-4, GPT-3.5, embeddings
• Anthropic — Claude 3.5, Claude 3 Opus/Sonnet/Haiku
• Google Gemini — Gemini 2.0 Flash, Gemini 1.5 Pro/Flash
• Self-hosted — vLLM, TGI, BentoML (GPU-time pricing)

Powered by

Component	Role
TinyFish	Live GPU price discovery (Lookup layer)
Confluent Kafka	Telemetry stream (lut.telemetry, lut.pcpo-features)
Apache Flink	Real-time cost aggregation (TUMBLE 10s windows)
Confluent Tableflow	Delta Lake sync for PCPO training
Databricks	PCPO model training (XGBoost v2, 21 features)
Neo4j AuraDB	DSPM knowledge graph
Google Cloud	GKE cluster, eBPF enforcement

Lutflow is a Confluent AI Accelerator Cohort 3 participant and TinyFish Accelerator participant — the first Latin American team in both programs.

Environment Variables

Variable	Description
LUTFLOW_ENDPOINT	API endpoint (default: https://api.lutflow.dev)
TINYFISH_API_KEY	TinyFish API key for live GPU pricing

Links

• Website: lutflow.dev
• Docs: pypi.org/project/lutflow
• GitHub: github.com/Lutflow/lutflow

License

Business Source License 1.1 (BSL 1.1)

• Free for: internal use, development, testing, evaluation, non-commercial use
• Commercial license required for: offering GPU cost management as a service
• Change Date: March 29, 2030 (converts to Apache 2.0)

See LICENSE for full terms. For commercial licensing: [e-mail]