You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Gemma-2-9B (ZeroFold QTensor Edge Optimization)

This repository provides the compiled weights and evaluation artifacts for Gemma-2-9B compiled via the ZeroFold QTensor Engine (Patent Pending: AU 2026907153).

By applying low-rank subspace decomposition with Subspace Residual Bridging, the compiler preserves high-frequency spectral energy across attention and feed-forward pathways while operating within strict physical VRAM envelopes.

Empirical Benchmark Suite

1. MBPP (Mostly Basic Python Problems)

Evaluated via the standard lm-evaluation-harness (v0.4.12) under greedy decoding ($\tau=0.0$) using 3-shot in-context demonstrations across the complete 500-problem test split.

Metric	Official DeepMind Baseline	Gemma-2-9B QTensor (Ours)	Delta
Pass@1 (n=500)	52.40%	56.00%	+3.60%
Standard Error	—	±0.0222	—
VRAM Consumption	>18.0 GB (Uncompressed)	15.93 GB Peak	Consumer 16GB Fit
Exit Code	—	0 (Deterministic)	—

Key Observation: Beyond memory boundary enforcement, the Subspace Residual Bridge functions as a structural regularizer—eliminating overparameterized noise in attention projections to yield higher task accuracy over baseline uncompressed weights.

2. OpenAI HumanEval

Evaluated via the official OpenAI human_eval.evaluation.evaluate_functional_correctness execution engine with sandboxed execution timeout (3.0s) under strict greedy decoding ($\tau=0.0$) across the complete 164-problem benchmark.

Metric	Metric Value	Operational Significance
Certified Pass@1	59.15% (97 / 164 passed)	Highest functional correctness in ZeroFold benchmark portfolio
Portfolio Standing	🏆 Suite Champion	+1.22% vs Llama 3.1 8B (57.93%), +4.27% vs Qwen 2.5-Coder-7B (54.88%)
AST Syntactic Validity	100.0% (0 SyntaxErrors)	Flawless Python AST syntax generation across all 164 programs
Degenerate Loop Rate	0.0%	Zero unigram repetition loops / zero low-entropy attractor lockups
Compilation Speed	< 9.0s / layer	NVIDIA GeForce RTX 5080 with zero host-to-device paging overhead

Hardware Telemetry & Runtime Environment

Accelerator: NVIDIA GeForce RTX 5080 (16GB GDDR7)
CUDA Runtime: 13.2 | Windows 11
Dedicated VRAM Peak: 15,926 MiB / 16,303 MiB (97.7% physical boundary adherence)
Post-Run State: 665 MiB (zero memory leakage, clean driver release)
Execution Throughput: ~32.4s / problem (batch size = 1)
Total MBPP Runtime: 04:30:17 (Exit Code: 0)

⚠️ Critical Runtime Execution Notice: SDK Required

The QTensor Engine restructures linear tensor projections via Subspace Residual Bridging (Australian Provisional Patent #2026907153) to preserve high-frequency spectral energy inside bounded hardware envelopes.

Because tensor routing paths are factored on-the-fly, standard Hugging Face AutoModelForCausalLM and PyTorch linear implementations cannot evaluate these weights natively. Attempting to load this model with unpatched linear layers will result in logit collapse or runtime exceptions.

Enterprise Diligence & Evaluation Workflow

To verify execution and benchmark telemetry on your local silicon:

Request Weight Access: Submit an access request via this Hugging Face repository. (Please ensure your Hugging Face profile lists your verified organizational affiliation).
Request Runtime SDK Access: Access to the closed-source qtensor-runtime Python wheel (.whl) and private GitHub evaluation repository is granted under standard bilateral evaluation/NDA terms.
- Enterprise & Institutional Inquiries: Submit your request to the repository maintainer with your organization name, target hardware specification, and evaluation intent.

Verified Client Execution Example (via `qtensor-runtime`):

from qtensor import QTensorModel

# Streams directly into GPU VRAM with zero host-RAM overhead
model = QTensorModel.from_pretrained(
    "trentzap/gemma-2-9b-it-qtensor", # or trentzap/qwen-2.5-coder-7b-qtensor
    device="cuda:0"
)

output = model.generate(
    "def fibonacci(n: int) -> list[int]:\n    '''Return first n Fibonacci numbers.'''\n",
    max_new_tokens=64,
)
print(output)

Verification & Reproducibility Dossier

Evaluation logs and execution telemetry artifacts are archived in the repository:

12_gemma_9b_qtensor_mbpp_full.json (Aggregated metrics, sample responses, test execution results for MBPP)
qtensor_humaneval_samples-sanitized_eval_results.json (Certified evaluation execution records for HumanEval)
Telemetry proof card asset: gemma_9b_mbpp_proof_card.png

Runtime Integration

These compiled weights are intended for execution via the public QTensor SDK:

pip install git+https://github.com/Trentzap1/qtensor-engine.git

from qtensor import QTensorModel

# Load compiled Gemma 2 9B Enterprise runtime weights
model = QTensorModel.from_pretrained("trentzap/gemma-2-9b-it-qtensor")
output = model.generate("def quicksort(arr: list[int]) -> list[int]:", max_tokens=128)
print(output)

Technical Inquiries & Diligence Access

For corporate evaluation, institutional venture capital technical diligence packs, encrypted reproduction logs, or architecture deep-dives under NDA, contact the ZeroFold team:

Lead Architect: Trent Parsons (LinkedIn)
Product: qtensor.com.au
Engineering Practice: zerofold.com.au · zerofold.co/contact

Downloads last month: -

Safetensors

Model size

9B params

Tensor type

BF16