Views
No views yet
model.safetensors, config.json, tokenizer.json, ... — the materialized
clustered checkpoint (standard Qwen3ForCausalLM; load with
AutoModelForCausalLM.from_pretrained). Tokenizer is bundled.activation_phase_lut/policy.json — the FP8 activation phase-LUT (one FP8
format per activation row, chosen from e0m7, e1m6, e2m5, e3m4, e4m3, e5m2)
for the mixed-activation arm (--act-composition-policy phase_lut). The fixed
arm uses plain FP8 E4M3 and does not need this file.scripts/hpc/run_k8_kv_scheme_seeds.sh sweeps two activation arms
(fixed E4M3 / mixed FP8 phase-LUT) x three KV error schemes
(attn_only / maxval / proxy) x seeds {1, 42, 99} at KV budget 2 (ladder16),
then scripts/hpc/aggregate_k8_kv_scheme_seeds.py reports mean +/- std over
seeds for wiki2 ppl, c4 ppl, KL/token, and MC accuracy. Qwen/Qwen3-4B is the
KL teacher only; the tokenizer comes from this checkpoint.