ThakiCloud/Qwen3-Coder-30B-A3B-W4A16

docs: 카드 메타데이터 보강 (base_model_relation/library_name/datasets/language)

Measure the Hopper case directly: W4A16 loses to FP8 there too (0.81-0.84x)

Answer the open kernel-tuning hedge with the four-way NVFP4 vs FP8 result; note the FP8 size discrepancy

card: NVFP4 sibling — 4-bit slowness is W4A16's, not 4-bit's

add model.safetensors

add tokenizer.json

add config.json

add qwen3coder_tool_parser.py

add chat_template.jinja

add tokenizer_config.json

add generation_config.json

add recipe.yaml

card: measured results before weights

initial commit