axe-strada-knivesysl-27b

qwen3.8-27b converted to the knivesysl fp6 format (e2m3, 128-wide pow2 block scales, qmma fragment layout) with the mtp speculative-decoding head preserved.

this is the weight file for the knivesysl inference engine — a bare-cuda sm120 engine. it is not loadable by transformers, vllm or sglang; the layout is the tensor-core fragment layout the engine's hand-written mma.sync ... kind::mxf8f6f4.block_scale kernels read directly.

qwen3.8-27b (bf16) --> convert_qwen_tqf.py --> model.tqf --> libforward_qwen.so

format	fp6 e2m3, 128-wide pow2 block scales
size	22.6 gb
tower	64 layers, 5120 hidden, 16 full-attn + 48 gated-deltanet
context	262144 native
mtp head	included (bf16 projections)
target	sm120 only (rtx 5090 / rtx pro 6000)

use

git clone https://github.com/knowrohit/knivesysl && cd knivesysl
cmake -B build-qwen -DCMAKE_CUDA_ARCHITECTURES=120
cmake --build build-qwen --target knivesysl-forward-qwen -j

huggingface-cli download srswti/axe-strada-knivesysl-27b --local-dir ./model
tools/serve.sh daemon        # TQF and MODEL_DIR are auto-discovered

the tokenizer and chat template are included here, so --model-dir can point at this same directory.

conversion

TQ_EMIT_MTP=1 TQ_GPU_PACK=1 python3 tools/convert_qwen_tqf.py <qwen3.8-27b> \
    -o model.tqf --block-scaled always --block-layout qmma-e2m3 --block-scale-policy pow2

inspect_tqf.py reports block_scaled_e2m3: True, has_mtp_section: True, flags 0x53d.

provenance

derivative of Qwen/Qwen3.8-27B (apache-2.0). weights are requantized from the bf16 checkpoint; the vision tower is dropped (text tower only). no fine-tuning, no distillation — numerics only.