-
Checkpoint: qwen-fp8 (tcclaviger RDNA4 fork) tuning campaign complete
released this
2026-09-29 23:29:35 -07:00 Tuned incumbent (measured 2026-09-29): orcarouter/Qwen3.8-Flash-Next-
Uncensored-FP8 on 2x R9700, TP=2, MTP depth 4, stock gather 8x16,
PREFETCH OFF (telemetry: 36% precision, ~50% wasted gather traffic),
TUNABLE=0, UTIL=0.98, KV 2 GiB, MAXLEN 32768.
Decode 40.57/37.44/37.57 tok/s (166/1087/4064 ctx); prefill 2116 tok/s.
256k ctx variant: MAXLEN=262144 KV_BYTES=4328521728 -> 36.48/34.13/35.02.
Image pinned by digest sha256:da402966... (local alias
qwen-fp8-runtime:tuned-20260929, ID 3322dbbbc96d).Revert to this state: git checkout qwen-fp8-tuned-20260929
Runbook: serverAppConfigs-docs/LLM/dankserver-qwen38-fp8-expert-offload-tuning.mdDownloads