zamba2-2.7b-instruct-hxq

Zamba2-2.7B-Instruct-HXQ

Important: Mainline llama.cpp does not support Zamba2 yet. The "Use this model" instructions Hugging Face auto-generates above (brew install, Ollama, Docker Model Runner, etc.) will fail with unknown model architecture. Use the fork build below, or the Python path via helix-substrate. Zamba2 support is in review at llama.cpp PR #21412.

Zamba2-2.7B-Instruct compressed with HXQ (HelixCode vector quantization).

Available as both HuggingFace safetensors (via helix-substrate) and native GGUF (via llama.cpp HXQ fork).

First hybrid Mamba2+Transformer architecture with HXQ runtime benchmarks.

Which Fork Branch?

This repo contains an HXQ-compressed GGUF. The fork has two branches — use the right one for your file:

For plain-quant Zamba2 GGUFs, use zamba2-support — it's on a newer mainline base and carries the latest Zamba2 fixes (including attention scaling corrections). For HXQ files, you need hxq-affine-type. A branch unification is planned.

Quick Start

Option 1: Native GGUF (llama.cpp fork)

# Build llama.cpp with HXQ + Zamba2 support
git clone -b hxq-affine-type https://github.com/echo313unfolding/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build -j --target llama-cli

# Run
./build/bin/llama-cli -m zamba2-2.7b-instruct-hxq-affine6.gguf \
  -ngl 99 -p "Explain the theory of relativity in simple terms:" -n 128

Known-good fork commit: 580e9a2 on hxq-affine-type branch.

CPU-only build: omit -DGGML_CUDA=ON. Metal (macOS): use -DGGML_METAL=ON instead.

Option 2: HuggingFace (Python)

pip install "helix-substrate[hf]"
# Fast Mamba SSM kernels (CUDA GPU only — won't install on CPU-only or macOS):
pip install mamba-ssm causal-conv1d
# If the install fails, see https://github.com/state-spaces/mamba/#installation
# CPU/macOS users: use the GGUF path above instead.

import helix_substrate  # registers the HXQ quantizer with HuggingFace
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("EchoLabs33/zamba2-2.7b-instruct-hxq")
tokenizer = AutoTokenizer.from_pretrained("EchoLabs33/zamba2-2.7b-instruct-hxq")

inputs = tokenizer("The capital of France is", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Both paths verified 2026-09-26: GGUF clean-room build + generation at 580e9a2 (CPU); Python path loads Zamba2ForCausalLM with 147 HelixLinear layers and generates coherent text.

Troubleshooting

GGUF Runtime Benchmark (RTX 3090, hxq-affine-type @ 580e9a2)

Benchmarked against standard GGUF K-quants on RTX 3090, full GPU offload (-ngl 99), using the hxq-affine-type branch at commit 580e9a2.

Decode Speed (tg128, 3 runs)

Perplexity (WikiText-2, 654 chunks, ctx=512)

Prefill (pp512, 3 runs)

Summary: HXQ_AF6 decodes faster than both Q6_K (+2.0%) and Q5_K_M (+5.7%) while being smaller than Q6_K (2.79 vs 2.93 GB). PPL is second-best, only 0.080 behind Q6_K. Prefill is within 2% across all formats (SSM-dominated, not matmul-dominated). This is the first HXQ runtime benchmark on a hybrid Mamba2+Transformer architecture.

Reproducibility

All claims are within-run comparisons using the same dataset, llama.cpp commit, and hardware. Do not compare these PPL numbers with numbers from other runs using different model variants, dataset files, or build configurations.

Note: Zamba2 absolute PPL (~22) is higher than Qwen (~8-10) on this dataset due to different tokenizer and training distribution. The important metric is relative ranking within this run.

Receipt with SHA256 artifact hashes, exact commands, and dataset provenance: hxq_runtime_3090_zamba2_2.7b_20260509

Safetensors Benchmark

Note: The safetensors PPL (5.68) and GGUF PPL (22.653) use different evaluation configurations (ctx=2048/stride=512 vs ctx=512/654 chunks) and different tokenization. They are not directly comparable.

Good to Know

• GPU and CPU supported — runs on any CUDA GPU or CPU via standard PyTorch. Native GGUF runs via llama.cpp.
• Hybrid architecture — 45 Mamba2 layers + 9 shared Transformer layers. SSM tensors (ssm_in, ssm_out, ssm_mix) and shared FFN all compressed.
• Fine-tunable via LoRA — compressed weights remain frozen, but LoRA adapters attach to each HelixLinear layer via HelixLinearSTE. See helix-substrate for training infrastructure.
• Requires helix-substrate for safetensors path — the quantizer is not built into transformers.
• Requires llama.cpp HXQ fork for GGUF path — standard llama.cpp does not have HXQ type support yet. The hxq-affine-type branch also includes Zamba2 architecture support.

What is HXQ?

HXQ is a weight compression codec based on vector quantization with per-group affine correction:

• Each weight matrix is replaced by a 256-entry codebook + uint8 index matrix + per-group affine scale/offset
• The compressed form is the executable — codebook[indices] * scale + offset during matmul, no decompression step
• Works on any nn.Linear regardless of architecture (Transformer, Mamba, MLP)
• No calibration data required — codebooks are fit from the weights alone via k-means
• 6.27 bits per weight in the GGUF affine-6 format

Companion Models

Same codec, multiple architectures:

Citation

@software{hxq_2026,
  title={HXQ: Vector Quantization with Per-Group Affine Correction for Neural Network Weight Compression},
  author={Echo Labs},
  year={2026},
  url={https://github.com/echo313unfolding/helix-substrate}
}

License

Apache 2.0 (inherited from Zyphra/Zamba2-2.7B-instruct).