# Precision - Preserving Subtensor Inference for a 22B Joint Audio - Video Generator under a Fixed GPU Working Set

DevFeed: [Precision - Preserving Subtensor Inference for a 22B Joint Audio - Video Generator under a Fixed GPU Working Set](<https://devfeed.tech/articles/precision-preserving-subtensor-inference-for-a-22b-joint-audio-video-generator-under-a-fixed-gpu-working-set-121273.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.08023>)

Author: Hanyue Shen, Xinling Liao

Published: 2026-11-06T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>)

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [multimodal-ai](<https://devfeed.tech/topics/multimodal-ai.md>), [quantization](<https://devfeed.tech/topics/quantization.md>)

Tags: [3](<https://devfeed.tech/tags/3.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [all](<https://devfeed.tech/tags/all.md>), [arena](<https://devfeed.tech/tags/arena.md>), [arithmetic](<https://devfeed.tech/tags/arithmetic.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attention](<https://devfeed.tech/tags/attention.md>), [audio](<https://devfeed.tech/tags/audio.md>), [bias](<https://devfeed.tech/tags/bias.md>), [block](<https://devfeed.tech/tags/block.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [budget](<https://devfeed.tech/tags/budget.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [generator](<https://devfeed.tech/tags/generator.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [inference](<https://devfeed.tech/tags/inference.md>), [precision](<https://devfeed.tech/tags/precision.md>), [set](<https://devfeed.tech/tags/set.md>), [under](<https://devfeed.tech/tags/under.md>), [video](<https://devfeed.tech/tags/video.md>)

## AI overview

Precision-Preserving Subtensor Inference (PSI) runs the BF16 LTX-2.5 22B joint audio-video transformer with an explicit GPU-memory budget by streaming weights in sub-blocks and managing intermediate state separately in a fixed arena. Across nine matched RTX A6000 cases, modified PSI reduced sampled transformer-process peak memory from 8,578 MiB with official MATH to 360 MiB, with mean time for eleven transformer invocations increasing from 247.11 to 349.00 seconds; the managed arena was 80 MiB. Both native generation and a complete-block FLASH path succeeded under a 3 GiB available-memory cap, and complete audiovisual decoding succeeded in the A6000 comparison and a separate DGX Spark service suite. Numerical fidelity remains distinct from memory and transport correctness: FFN reuse cut mean block time by 4.0% without changing BF16 outputs, while targeted bias and attention changes improved local and one-token comparisons but increased mean endpoint error across full-prompt cases.

## Source excerpt

arXiv:2611.08023v1 Announce Type: new Abstract: Complete-block offloading reduces model residency but requires each transferred block to fit the GPU budget. We present Precision-Preserving Subtensor Inference (PSI), a fixed-arena runtime that combines sub-block weight streaming with separately managed intermediate state for the BF16 LTX-2.5 22B joint audio-video transformer. The contribution is a construction whose managed GPU footprint is determined by reusable pages, live state, and bounded workspace, together with controls that separate resource use, transport correctness, and numerical fidelity. Its partitioned operators preserve the dense real-arithmetic construction under stated assumptions. Across nine matched full-prompt cases on RTX A6000, modified PSI reduces sampled transformer-process peak memory from 8,578 to 360 MiB relative to official MATH, while mean time for the eleven transformer invocations rises from 247.11 to 349.00 s. Its managed arena is 80 MiB. Complete native generation and a complete-block FLASH path both succeed under a fixed 3 GiB available-memory cap. The matched A6000 comparison and separate DGX Spark service suite both achieve 100% success in complete audiovisual decoding. FFN reuse lowers mean block time by 4.0% with unchanged BF16 outputs; a fixed-arithmetic one-token transport control preserves all saved states bitwise. Targeted bias and attention changes improve local and one-token numerical comparisons, while mean endpoint error increases across nine full-prompt cases. The construction enables BF16 audiovisual inference with an explicit transformer-memory budget and makes numerical fidelity a separate, testable property of the implementation.