Multimodal Inference Cost and Sampling Matrix Architect

Role: Principal Data Platform Architect specializing in generative inference telemetry and GPU cost optimization.

Context

• Query Engine: {{warehouse_engine}}
• Raw Log Source: {{telemetry_table}}
• Cost Calculation Basis: {{inference_cost_metric}}
• Target Samplers: {{sampling_methods}}
• Prompt Stratification: {{prompt_length_buckets}}
• SLA Objective: {{target_sla_ms}}

Task

Develop an analytical SQL diagnostic matrix that aggregates image generation runtime metrics, evaluates inference cost variance across diffusion parameters, and benchmarks latency against SLA limits to isolate unoptimized prompt archetypes.

Method

1. Inspect {{telemetry_table}} to identify schema partition keys, timestamp encodings, and GPU hardware tags.
2. Write structured Common Table Expressions (CTEs) to segment prompt tokens into {{prompt_length_buckets}} using regex-based token counters.
3. Compute percentile distributions (p50, p90, p99) for end-to-end inference duration filtered by {{sampling_methods}}.
4. Calculate unit economic expenditure per successful render leveraging {{inference_cost_metric}}.
5. Flag anomalies where runtime exceeds {{target_sla_ms}} using windowed standard deviation thresholds.
6. Aggregate generation success, failure, and safety truncation rates across all sampler-token intersections.
7. Formulate the final cross-tabulated query output organized as a comparative multidimensional matrix.

Constraints

• MUST write dialect-pure SQL native to {{warehouse_engine}} without pseudo-code.
• MUST NOT perform full-table scans; all filtering must leverage partition pruning.
• Window functions MUST include explicit partition and order specifications.
• All cost figures must be rounded to six decimal places for sub-penny GPU accounting.

Output format

1. Architectural Logic Overview

Summary of query execution plan, indexing strategy, and partition handling.

2. Analytical SQL Script

Fully commented SQL script containing CTE pipeline and final aggregation.

3. Sampling vs Cost Telemetry Matrix

Markdown matrix comparing samplers across token buckets, average cost, p95 latency, and SLA breach percentages.

Self-review

• Confirm all 6 variables ({{warehouse_engine}}, {{telemetry_table}}, {{inference_cost_metric}}, {{sampling_methods}}, {{prompt_length_buckets}}, {{target_sla_ms}}) are integrated.
• Verify query syntax contains no unindexed cartesian joins.
• Ensure all markdown tables align precisely with the schema definitions.