Siddachan/qwen2.5-3b-chpe-raw

Qwen2.5-3B-Instruct: CHPE Contiguous Raw Weights (ARM Neoverse-N1)

This repository contains the zero-overhead, strictly aligned raw weight archives for Qwen2.5-3B-Instruct, formatted specifically for the Christopher Hamil Prediction Engine (CHPE) on 64-bit ARM microarchitectures.

Key Performance Results (ARM Neoverse-N1 @ 3.00 GHz, 16 vCPUs)

Architecture	Model	Precision	Token Latency	Generation Speed	Delta vs Llama.cpp	Verification
CHPE (Hamil)	Qwen2.5-3B-Instruct	FP16	$106.25\text{ ms}$	$9.138\text{ tok/s}$	$-16.57%$	argmax=50994, 0 NaN/Inf
CHPE (Hamil)	Qwen2.5-3B-Instruct	BF16	$126.94\text{ ms}$	$7.810\text{ tok/s}$	$-0.32%$	argmax=50994, 0 NaN/Inf
Llama.cpp	Qwen2.5-3B-Instruct	Full FP	$127.35\text{ ms}$	$7.785\text{ tok/s}$	Baseline ($0.0%$)	Benchmark ref

• Official OpenBenchmarking Run: 2609164-NE-CHPELAMB82

---

File Manifest & Checksums

All archives are raw binary memory-mappable blocks aligned to $16,384\text{ bytes}$ ($256 \times 64\text{B}$ cache lines):

• Qwen2.5-3B-Instruct.fp16.raw.chpe: $6,174,363,648\text{ bytes}$ (IEEE 754 half-precision float)
• Qwen2.5-3B-Instruct.bf16.raw.chpe: $6,174,363,648\text{ bytes}$ (Brain Floating Point 16-bit)
• tokenizer.json: Fast HF Tokenizer definition for Qwen2.5 vocabulary ($151,936\text{ tokens}$)
• manifest.json: Structural tensor offsets and streaming SHA256 signatures

---

Microarchitectural Invariants & Formal Proofs

The memory alignment, tensor map projections, and threadpool barriers have been mechanically validated via automated theorem provers:

• Z3 SMT2: Theorem SAT, $0$ bounds violations.
• Vampire First-Order Prover: 15/15 refutation goals proved valid.
• Leo-III Higher-Order Prover: 14/14 goals proved valid.
• Energy-Based Model (EBM): Reconstruction loss $E = 0.0000$.
• Citation Key: 098ad4dba5ecddbe

---

Usage with CHPE

# Clone the private engine repository
git clone https://github.com/christopherlhamil-creator/chpe-qwen-engine.git
cd chpe-qwen-engine

# Build with Zig 0.17
zig build -Doptimize=ReleaseFast

# Fetch weights using authenticated huggingface_hub
python3 scripts/fetch_weights.py --precision fp16

# Run inference
./zig-out/bin/chpe_qwen3b \
    --weights models/Qwen2.5-3B-Instruct.fp16.raw.chpe \
    --threads 16 \
    --warmup 5 \
    --steps 20