slosizer: Profit-aware reserved LLM capacity planning

slosizer sizes reserved LLM capacity against throughput, latency SLOs, and an explicit economic objective.

It takes request traces, converts them into provider-specific capacity work, simulates queueing under bursty arrivals, and tells you how many reserved units you should buy plus how much slack capacity you are likely to carry.

The package is built for the extremely normal situation where:

• you know your request shape better than your vendor calculator does,
• you care about p95 or p99 latency, not just average throughput,
• and you do not want your capacity plan to be a sacred spreadsheet that nobody trusts.

What problem does this solve?

Reserved-capacity systems like GSU/PTU are fundamentally throughput constructs, but production teams usually care about latency SLOs, burst risk, and headroom.

slosizer gives you one place to:

1. load request logs into the format the planner expects,
2. convert requests into provider-specific capacity units (GSU/PTU),
3. plan capacity for either:
   • throughput: control overload probability or required-unit percentile,
   • latency: satisfy p95/p99 queue-aware latency targets,
   • hybrid: balance provisioned cost against paygo overflow,
   • profit: maximize expected gross value after provisioned and SLO-failure costs,
4. quantify:
   • spare capacity,
   • overload probability,
   • expected overflow,
   • optimization benefit.

How latency works

Total latency = model latency + queue delay.

Model latency is how long the LLM takes to process your request with no contention, estimated from token counts and provider throughput rates. Queue delay is waiting time caused by bursty arrivals: when requests arrive faster than capacity can serve them, a backlog forms.

The package simulates an FCFS queue against your request trace to estimate tail latencies (p95/p99). More reserved capacity = shorter queues = lower tail latency. The goal is finding the minimum capacity that keeps queue delay acceptable.

Two ways to start

Option 1: No data yet

Use the synthetic generator to explore capacity planning before you have real logs:

import slosizer as slz

trace = slz.make_synthetic_trace(seed=42)
profile = slz.vertex_profile("gemini-2.5-flash-lite")

result = slz.plan_capacity(
    trace,
    profile,
    slz.LatencyTarget(slz.LatencySLO(threshold_s=1.5, percentile=0.99, metric="e2e")),
)

Option 2: You have request logs

You need a typed Parquet table (or DataFrame) with at minimum these 3 columns:

Column	What it means
timestamp	When the request arrived (datetime or seconds)
input_tokens	Tokens in the prompt
output_tokens	Tokens in the response

That's it. The package normalizes timestamps and fills defaults for everything else.

import pandas as pd
import slosizer as slz

df = pd.read_parquet("requests.parquet")

trace = slz.from_dataframe(
    df,
    schema=slz.RequestSchema(
        time_col="timestamp",
        input_tokens_col="input_tokens",
        output_tokens_col="output_tokens",
    ),
    provider="vertex",
    model="gemini-2.5-flash-lite",
)

Quickstart

1) Create the environment with uv

uv sync --all-groups

2) Run the shipped synthetic demo

uv run python examples/quickstart.py

This writes:

• examples/output/comparison.parquet
• examples/output/latency_vs_capacity.png
• examples/output/required_units_distribution.png
• examples/output/scenario_benefit.png
• examples/output/percentile_tradeoff.png

3) Run the checks

uv run pytest -q
uv run ruff check src tests examples
uv run ruff format --check src tests examples
uv run deptry .
uv run vulture

Install and use it on your own trace

Minimal latency-oriented example

import pandas as pd
import slosizer as slz

df = pd.read_parquet("requests.parquet")

trace = slz.from_dataframe(
    df,
    schema=slz.RequestSchema(
        time_col="timestamp",
        class_col="route",
        input_tokens_col="prompt_tokens",
        cached_input_tokens_col="cached_prompt_tokens",
        output_tokens_col="completion_tokens",
        thinking_tokens_col="reasoning_tokens",
        max_output_tokens_col="max_output_tokens",
        latency_col="latency_s",
    ),
    provider="vertex",
    model="gemini-2.5-flash-lite",
)

profile = slz.vertex_profile("gemini-2.5-flash-lite")

result = slz.plan_capacity(
    trace,
    profile,
    slz.LatencyTarget(
        slz.LatencySLO(
            threshold_s=1.5,
            percentile=0.99,
            metric="e2e",
        )
    ),
)

print(result.recommended_units)
print(result.metrics)

Throughput-oriented example

import slosizer as slz

trace = slz.make_synthetic_trace(seed=42)
profile = slz.vertex_profile("gemini-2.5-flash-lite")

result = slz.plan_capacity(
    trace,
    profile,
    slz.ThroughputTarget(
        percentile=0.99,
        max_overload_probability=0.01,
        windows_s=(1.0, 5.0, 30.0),
    ),
)

print(result.recommended_units)
print(result.slack_summary)

Cost-optimal hybrid planning

This is the normal operating case. The request trace is the demand forecast. With fixed demand and a hard SLO, minimizing inference cost subject to the SLO maximizes profit. Users do not need to estimate request value or demand elasticity.

from datetime import date

import slosizer as slz

trace = slz.make_synthetic_trace(seed=42)
profile = slz.vertex_profile("gemini-2.5-flash")
pricing = slz.RateCard(
    provisioned=slz.ProvisionedPricing(cost_per_unit_hour=3.698630137),
    paygo=slz.PaygoPricing(
        input_cost_per_million=0.30,
        cached_input_cost_per_million=0.03,
        output_cost_per_million=2.50,
        thinking_cost_per_million=2.50,
    ),
    currency="USD",
    provider="vertex",
    model="gemini-2.5-flash",
    verified_on=date(2026, 8, 15),
    source="https://cloud.google.com/vertex-ai/generative-ai/pricing",
)

result = slz.plan_hybrid_capacity(
    trace,
    profile,
    pricing,
    slz.HybridTarget(
        strategy="cost_optimal",
        latency_slo=slz.LatencySLO(
            threshold_s=1.5,
            percentile=0.99,
        ),
    ),
    options=slz.PlanOptions(baseline_latency_model=slz.BaselineLatencyModel()),
)

print(f"Provision {result.provisioned_units} GSUs + paygo overflow")
print(
    f"Saves ${result.savings_vs_full_provision:.2f}/hr ({result.savings_percent:.0f}%)"
)

That public list rate was checked on 2026-08-15. Production analysis should use the effective rate on your invoice or contract and record its source and validity dates.

cost_optimal finds the cheapest provisioned and paygo blend. percentile_split provisions at a chosen workload percentile and sends the rest to paygo.

Advanced economic planning

Use plan_profit_capacity() when you need an absolute profit estimate or want to price SLO misses instead of treating the SLO as a hard constraint. This requires expected gross value per request and, for a priced SLO, a defensible cost per miss.

import slosizer as slz

trace = slz.make_synthetic_trace(seed=42)
profile = slz.vertex_profile("gemini-2.5-flash")
pricing = slz.RateCard(
    provisioned=slz.ProvisionedPricing(cost_per_unit_hour=3.698630137),
    provider="vertex",
    model="gemini-2.5-flash",
)

result = slz.plan_profit_capacity(
    trace,
    profile,
    pricing,
    slz.ProfitTarget(
        latency_slo=slz.LatencySLO(threshold_s=1.5, percentile=0.99),
        slo_policy="priced",
        value_per_request=0.05,
        slo_violation_cost_per_request=0.01,
    ),
    options=slz.PlanOptions(baseline_latency_model=slz.BaselineLatencyModel()),
)

print(result.expected_profit_hourly)
print(result.candidate_plans)

business_value is gross contribution before inference and SLO costs. It changes the value assigned to a request, not demand. The package does not estimate demand effects. If a model is expected to receive different traffic, supply a different forecast trace in an optional ProfitScenario. Most users should not need this API.

With one trace and a hard SLO, business value changes reported profit but not recommended capacity. Use the hybrid planner for that case.

headroom_factor is rejected for cost_optimal because adding capacity after the search would no longer be cost optimal. Model uncertainty with workload scenarios instead.

Azure PTU example

Azure support is calibration-first: you seed a profile from the Azure calculator and benchmark results, then use the same planning machinery.

import slosizer as slz

profile = slz.azure_profile(
    "gpt-5.2",
    throughput_per_unit=3400 / 60,
    purchase_increment=5,
    min_units=15,
    input_weight=1.0,
    cached_input_weight=0.0,
    output_weight=8.0,
    thinking_weight=8.0,
    deployment_type="data_zone_provisioned",
)

Optional fields for better planning

The 3-column minimum works, but you get more accurate capacity estimates with:

Column	Why it helps
cached_input_tokens	Cached tokens cost less capacity
thinking_tokens	Reasoning models use extra tokens
max_output_tokens	Helps estimate worst-case latency
class_name	Separate capacity needs by request type
latency_s	Calibrate model latency estimates
request_model / response_model	Distinguish the requested alias from the model that actually served
service_tier	Separate provisioned, standard, priority, batch, and other routes
business_value	Optional absolute-profit or priced-SLO analysis

See docs/data-requirements.md for full details.

Example input files:

• examples/input/synthetic_request_trace_baseline.parquet
• examples/input/synthetic_request_trace_optimized.parquet

Built-in provider support

Vertex GSU

The package ships a reviewed, versioned TOML catalog for text-capable Vertex Provisioned Throughput models. Call available_vertex_profiles() for the current set; this avoids duplicating a model list in documentation. Provider facts live in src/slosizer/data/vertex.toml, not in optimizer code. Add a new model by updating the catalog or load your own with load_capacity_profiles().

Azure PTU

Azure PTU support is user-calibrated on purpose. The package gives you the same planning engine, but you provide the model-specific PTU profile from your calculator + benchmark loop.

See docs/provider-adapters.md.

Synthetic demo: what it shows

The repo ships with a fake but bursty workload containing three classes:

• chat
• rag
• reasoning

The optimized variant simulates:

• tighter prompts,
• more caching,
• shorter outputs,
• lower thinking-token budgets.

That lets you inspect two things immediately:

1. Optimization can reduce reserved-capacity needs.
2. Planning for stricter percentiles usually increases slack capacity.

Snapshot of the current synthetic outputs

The demo writes the complete current result table to examples/output/comparison.parquet and prints the same values to the terminal. That generated table is the result source; the documentation does not maintain a second copy of the numbers.

These numbers are synthetic. They are there to show the mechanics, not to cosplay as your production traffic.

Output plots

Latency vs provisioned capacity

Distribution of required reserved units

Optimization benefit

Slack trade-off

Repo map

• docs/formalization.md: generic throughput/latency model
• docs/economics-and-data.md: profit objective, SLO policy, and storage boundaries
• docs/data-requirements.md: what columns you need and why
• docs/provider-adapters.md: how GSU/PTU adaptation works
• docs/examples.md: the synthetic walkthrough
• examples/quickstart.py: reproducible demo script

Caveats

• The queue model is intentionally simple: FCFS fluid queueing, not a perfect service simulator.
• Built-in Vertex profiles are text-centric. Multimodal traffic needs more columns and weights.
• Azure PTU math is workload-sensitive, so the package does not fake vendor-authoritative PTU values for you.
• If you do not have a latency column, the package falls back to a simple token-based baseline latency model. That is a starting point, not gospel.

Name

The package name is slosizer because "how many units do I need, and how much empty air am I buying to hit p99?" is the real question under all the vendor jargon.