Qwen2.5-7B-Instruct compressed with HXQ (HelixCode vector quantization).
Available as both HuggingFace safetensors (via helix-substrate) and native GGUF (via llama.cpp HXQ fork).
GGUF Runtime Benchmark (RTX 3090 Ti)
Benchmarked against standard GGUF K-quants on RTX 3090 Ti, full GPU offload (-ngl 99), using the hxq-affine-type branch at commit 580e9a2.
Decode Speed (tg128, 3 runs)
Format
Size
bpw
tok/s
vs Q4
vs Q6
Q4_K_M
4.36 GB
4.5
127.33
100%
129%
Q5_K_M
5.07 GB
5.5
117.30
92.1%
119%
HXQ_AF6
5.56 GB
6.25
114.02
89.5%
115.7%
Q6_K
5.82 GB
6.56
98.56
77.4%
100%
Perplexity (WikiText-2, 50 chunks, ctx=512)
Format
bpw
PPL
vs Q4
HXQ_AF6
6.25
7.982
-0.333 (best)
Q6_K
6.56
8.116
-0.199
Q5_K_M
5.5
8.184
-0.131
Q4_K_M
4.5
8.315
baseline
Prefill (pp512, 3 runs)
Format
tok/s
vs Q4
Q5_K_M
2767
109%
Q4_K_M
2539
100%
HXQ_AF6
2293
90.3%
Q6_K
2259
89.0%
Summary: HXQ_AF6 has the lowest perplexity of all four formats and decodes 15.7% faster than Q6_K while being smaller (5.56 vs 5.82 GB). It trades ~10.5% decode speed vs Q4_K_M for better quality preservation at 6.25 bpw.
This pattern matches the 3B coder results, where HXQ also had the best PPL and fastest decode vs Q6_K.
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.
Receipt with SHA256 artifact hashes, exact commands, and dataset provenance: hxq_runtime_3090ti_qwen7b_instruct_20260509
Install and Run
Option 1: Native GGUF (llama.cpp)
bash
1# Build llama.cpp with HXQ support2git clone -b hxq-affine-type https://github.com/echo313unfolding/llama.cpp.git
3cd llama.cpp &&mkdir build &&cd build
4cmake .. -DGGML_CUDA=ON &&make -j$(nproc) llama-cli
56# Run7./bin/llama-cli -m qwen2.5-7b-instruct-hxq-affine6.gguf \8 -ngl 99 -p "Explain the theory of relativity in simple terms:" -n 128
Option 2: HuggingFace (Python)
pip install "helix-substrate[hf]"
python
1import helix_substrate # registers the HXQ quantizer with HuggingFace2from transformers import AutoModelForCausalLM, AutoTokenizer
34model = AutoModelForCausalLM.from_pretrained("EchoLabs33/qwen2.5-7b-instruct-hxq")5tokenizer = AutoTokenizer.from_pretrained("EchoLabs33/qwen2.5-7b-instruct-hxq")67inputs = tokenizer("Explain the theory of relativity in simple terms:", return_tensors="pt")8outputs = model.generate(**inputs, max_new_tokens=128)9print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Safetensors Benchmark
Dense (BF16)
HXQ (safetensors)
Size
14.2 GB
6.5 GB
Perplexity (WikiText-2, 2048 ctx)
6.949
7.388 (+6.34%)
Compression ratio
1x
2.2x
Compressed modules
0
196 HelixLinear layers
Note: The safetensors PPL (7.388) and GGUF PPL (7.982) use different evaluation configurations (ctx=2048/stride=512 vs ctx=512/50 chunks). 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.
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.
Tied embeddings — lm_head shares embed_tokens, stored at full precision.
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
