Limitations
TurboPress is a research prototype. Known limitations, stated plainly:
- Model families. Tested on Llama / Qwen / Mistral-family causal LMs. The
decoder-layer locator handles
model.layers,transformer.h, andmodel.decoder.layers; other architectures are unsupported. - Dimensions. Each linear's
in_featuresmust have a power-of-2 factor for the block Hadamard rotation. Layers without one are unsupported. - Storage, not yet runtime. The artifact stores weights at the true low
bit-width, but
run_quantized.pydecodes them to fp16 in memory to run on stocktransformers. There is no packed-bit inference kernel yet, so inference VRAM is fp16-sized. Export to GGUF/FP8 for serving. - Evaluation scale. Reported KL / top-1 / perplexity use a few thousand tokens of WikiText-2 (or a public-domain fallback). These are consistent, reproducible relative metrics — not a substitute for full downstream benchmark suites.
- Certificates are unsigned.
quantization_config.jsonrecords settings and measured metrics but is not yet cryptographically signed. Re-verify withturbopress validate. - Scale of validation. Method comparisons here are on small models (≤~0.6B). Behavior at 7B+ is expected to follow the same trends but is not yet measured end-to-end in this repo.
- Codebook assumes near-Gaussian rotated weights. The data-free codebook is
optimized for
N(0,1); the rotation makes rotated weights approximately Gaussian, but heavy-tailed layers can deviate.