diff --git a/scripts/run-swift15-bf16.sh b/scripts/run-swift15-bf16.sh index f1a8ba3..cb6ac9d 100755 --- a/scripts/run-swift15-bf16.sh +++ b/scripts/run-swift15-bf16.sh @@ -46,23 +46,29 @@ ARGS=(--model /models/rad/qwen4exp-swift15-bf16.rad case "$MODE" in kld) # 1.0.13: the engine that will serve the candidate — the reference is - # recorded on the same engine that scores it. The ratio-derived KV pool - # (1.95 GiB) is SHORT by 217 MiB at 32x16k; the pair is explicit: 16 slots - # need 2.17 fixed + 1.87 GiB of fp8 KV, 4608 MiB carries the margin. + # recorded on the same engine that scores it. The weights slab DERIVES as + # card - headroom - kv, and the disk tier refuses under 9.57 GiB (4.21 of + # layer buffers + 5.36 fixed), so kv is explicit SMALL: recording is + # prefill-only, self-batched at MNBT 32768 = two 16k prompts at a time; + # 4 slots x 16k fp8 (29 KiB/pos, 1.7 GiB) + headroom fits 3072. IMAGE="stilldeadcode/radiance@$PIN_1013" mkdir -p "$KLD_REF_DIR" ARGS+=(--kld-record /data/kld-ref --kld-corpus /data/kld.jsonl - --vram-kv-mib 4608 - --max-num-seqs 16 --max-num-batched-tokens 32768) + --vram-kv-mib 3072 + --max-num-seqs 4 --max-num-batched-tokens 32768) ;; calib) # 1.2.0: the calibration tap (RADIANCE_CALIB_DIR + RADIANCE_CALIB_TOKENS) # lives in the 1.2.0 qwen4exp plugin; 1.0.13 has only the DIR half with an # unparameterisable threshold. Grams are bf16 activations — engine-invariant. + # Same KV ceiling as kld: the disk-tier weights floor (9.57 GiB) plus + # ~18 GiB of fixed per-rank claims cap the pool near 3.5 GiB on a 32 GiB + # card. 4 slots x 16k fp8 fits; prefill rate is disk-bound, so fewer + # resident slots only queues pushes, it does not slow the record. IMAGE="stilldeadcode/radiance@$PIN_120" mkdir -p "$CALIB_DIR" - ARGS+=(--port 7770 --vram-kv-mib 9984 - --max-num-seqs 64 --max-num-batched-tokens 65536 + ARGS+=(--port 7770 --vram-kv-mib 3072 + --max-num-seqs 4 --max-num-batched-tokens 32768 --prefix-cache-dir /data/calib-kvcache) ;; probe)