This repository contains a specialized LiteRT-LM conversion of google/gemma-4-31B-it. It is engineered for local-first DPM//BENCH experiments, specifically targeting long-horizon incident narratives and red-team traces.

Objective: Package the base instruction model into a runtime format for deterministic projection memory experiments, ensuring that append-only event logs map to a consistent structured memory surface.

🛠️ Conversion Architecture

The conversion utilizes the latest LiteRT-LM stack, requiring specific flags to handle the Gemma 4 per-layer embedding structure.

View Conversion Script

bash

python -m litert_torch.generative.export_hf \
    --model /path/to/google/gemma-4-31B-it \
    --output_dir /path/to/out/gemma-4-31B-it-litert-lm \
    --externalize_embedder True \
    --single_token_embedder True \
    --experimental_lightweight_conversion True \
    --bundle_litert_lm True \
    --task text_generation

Critical Flags for Compatibility:

--externalize_embedder True: Essential for per-layer embedding paths.
--experimental_lightweight_conversion True: Prevents runtime artifact corruption.
--bundle_litert_lm True: Packages tokenizer and templates into the .litertlm artifact.

💻 Infrastructure Requirements

Requirement	Specification	Context
RAM	128 GB+	Minimum for 31B conversion overhead
Disk Space	500 GB	Workspace for intermediate FlatBuffer assets
Storage Type	NVMe SSD	Crucial for large model serialization
Inference	Apple Silicon / GPU	31B is unsuitable for fast CPU-only DPM

🔍 Validation Protocol

For a successful DPM//BENCH run, the artifact must maintain byte-stability. Ensure the following conditions are met:

Integrity: LiteRT-LM binary successfully parses the .litertlm bundle.
Determinism: At temp 0 and a fixed seed, repeated projection calls must yield identical memory-surface bytes.
Format: JSON-only prompts must satisfy schema constraints under high-compression DPM tests.

⚠️ Implementation Boundaries

Intended Use: Security incident summarization, telemetry trace compression, and blue-team event reasoning.
Non-Intended Use: This is not a standalone decision-making system. It is a projection tool. All outputs require human review and replay-validation in high-stakes environments.

Base Model: google/gemma-4-31B-it

macmacmacmac/gemma-4-31B-it-litert-lm

google/gemma-4-31B-it LiteRT-LM Optimized

🟢 Overview

🛠️ Conversion Architecture

💻 Infrastructure Requirements

🔍 Validation Protocol

⚠️ Implementation Boundaries

macmacmacmac/
gemma-4-31B-it-litert-lm

google/gemma-4-31B-it
_{LiteRT-LM Optimized}