scripts/run-qwen38-bf16.sh

@@ -39,37 +39,45 @@
 # speculation, fp8 KV, a 32,768-token window, the PLE n-gram table pinned to RAM,
 # and the production chat template rather than the one baked into the container —
 # every scored container is asked the same tokenised questions the reference
-# answered. The window is 32,768 because this corpus is long: measured on
+# answered. The window is 32,768: kld.jsonl is 197 records with a longest of
-# Dankserver 2026-10-27, kld.jsonl is 197 records with a longest of 38,263
+# 38,263 characters (longest prompt FIELD 36,855, median record 27,228). Its
-# characters (longest prompt FIELD 36,855, median record 27,228), which is ~9,600
+# token length — 10,554 for that longest record, counted by the engine itself —
-# tokens at the pessimistic chars/4 and ~7,700 at the ratio this repo actually
+# is in the funding paragraph below; the kldref gate re-checks the file against
-# measured on the K3 corpus (380,143 chars = 76,568 tokens, chars/4.97). A 16,384
+# the window before it boots anything.
-# window cleared that by 1.7x; 32,768 clears every record of both corpora by at
-# least 2.6x, and the kldref gate below re-checks it against the file on disk.
 #
 # NO --vram-kv-mib: it was REMOVED at 1.3.0 and naming it is a boot refusal, not
 # a no-op — verified on the pinned image 2026-10-27: "E --vram-kv-mib was
 # removed: the card is split between the weights and the cache from what the
-# model needs." The budget resolver derives the per-card split from
+# model needs." The resolver derives the split from --host-pool-mib,
-# --host-pool-mib and --gpu-headroom-mib, and those two are deliberately NOT
+# --gpu-headroom-mib and the activation arena; the rig inherits the first two from
-# production's 49,152 / 1,536 (frank-archon.env:77): the pool decides WHICH
+# frank-archon.env, the file that owns them. On THIS payload the host pool is not
-# experts stay resident, so it buys speed and never logits — the shape the gate
+# the cache's supplier: the boot lands in vram_budget.cpp's `no lender` branch,
-# has to hold identical is the kv dtype, the slot count, the step size, the spec
+# where the experts keep their plan-fixed 25.29 GiB and the cache gets the residue.
-# depth and the window. It is also the only figure that fits: a 335 GiB bf16
+# Measured twice — 12,288 and 49,152 MiB of host pool both printed the same
-# container cannot be booted against a 49 GiB resident set on a 192 GiB machine,
+# "kv cache 271.25 MiB ... 7,436"; only --gpu-headroom-mib moved anything
-# and a prefill-only pass does not care that the disk tier paces at ~3 GiB/s.
+# (768 -> 96 grew vram_weights 29.41 -> 30.06 GiB). The supplier the source names
+# is the arena: vram_budget.cpp:117, "lower --max-num-batched-tokens (the arena
+# scales with it)" — which is what the step size below now does.
+#
+# The pool, not the window, decides what the recorder accepts: it refuses any
+# document longer than the pool, so the 32,768-token window was decorative while the
+# cache held 7,436 — "E kld: document 196 is 8507 tokens and the KV pool holds
+# 7436, so it can never be scored" on all 197 records, exit 1. The longest record
+# is 10,554 tokens by the engine's own count (chars/3.6 for this file, not the
+# chars/4.97 this repo measured on the K3 corpus).
 #
 # The step size is ONE value for the recording and every scoring pass:
-# FRANK_QWEN38_MNBT, default 2,048. The gate's whole claim is that the reference,
+# FRANK_QWEN38_MNBT, default 1,024. The gate's whole claim is that the reference,
 # the served iq4r and every arm rode one shape, so the ceiling is set by the
 # TIGHTEST payload the gate boots rather than by the bf16 reference: the quantised
 # expert FILE tiers (the served iq4r, the arms) stream only at MNBT<=2048 and
-# refuse the widened form above it, and 2,048 is what production holds anyway.
+# refuse the widened form above it. 2,048 was still the wrong default on this
-# Recording the reference at 32,768 would therefore measure a shape the payloads
+# payload, because the arena the engine charges for a step comes straight out of
-# cannot reproduce; at 2,048 the 50,432 positions cost more steps, which is still
+# the cache — 1.08 GiB at a 2,048-token step, 504.39 MiB at a 1,024-token one, and
-# one drained window. Calibration keeps the wide 65,536 — grams are not logits,
+# the experts keep their share either way. At 1,024 the cache is 866 MiB: 24,300
-# nothing downstream is compared against a step size, and that is where a big
+# tokens, which scores every one of the 197 documents (longest 11,970).
-# step pays (a disk-tier step costs the same at 1 token as at 64k).
+# Calibration keeps the wide 65,536 — grams are not logits, nothing downstream is
+# compared against a step size, and that is where a big step pays.
 #
 # usage: scripts/run-qwen38-bf16.sh kldref|calib|score RAD OUT [--dry-run] [--image REF]
 #   (RAD and OUT are score's two positionals: RAD is a container under the models
@@ -86,8 +94,9 @@
 #        bf16 container, kldref only), FRANK_QWEN38_CAMPAIGN (holds kld-ref/ and
 #        calib/), FRANK_QWEN38_KLD_CORPUS, FRANK_QWEN38_TPL (the template copy
 #        the container reads through /models), FRANK_QWEN38_LOG_DIR, and
-#        FRANK_QWEN38_MNBT — the step size shared by kldref and score (2,048 is
+#        FRANK_QWEN38_MNBT — the step size shared by kldref and score (default
-#        the file tier's ceiling and production's own value, four slots and all),
+#        1,024: the file tier's ceiling is 2,048 and production runs there, but on
+#        this payload the arena charged for a wider step is taken from the cache),
 #        and FRANK_QWEN38_MAX_MODEL_LEN — the window every mode boots at (32,768;
 #        see the corpus census above, and raise it together with the corpus).
 # kldref refuses (exit 2) when the longest record of the corpus it is about to
@@ -148,12 +157,19 @@ KLD_REF_DIR="$CAMPAIGN/kld-ref"
 CALIB_DIR="$CAMPAIGN/calib"
 C=qwen38-bf16
 SERVICE=frank-archon
+# The card split, taken from the file that owns it instead of a private pair:
+# --host-pool-mib also funds the KV cache (see the header), so hardcoding it
+# here once starved the recorder. The `:?` guards keep an empty env fatal.
+HOST_POOL_MIB="${FRANK_ARCHON_HOST_POOL_MIB:?FRANK_ARCHON_HOST_POOL_MIB must be set}"
+GPU_HEADROOM_MIB="${FRANK_ARCHON_GPU_HEADROOM_MIB:?FRANK_ARCHON_GPU_HEADROOM_MIB must be set}"
 
 # ONE step size for both logits modes. The file tier streams only at <=2,048 and
 # the gate scores file-tier payloads (the served iq4r, the arms) against the
-# reference, so the reference is recorded at the ceiling the payloads can meet —
+# reference, so the reference is recorded at a step the payloads can meet — never
-# never two numbers that a KL between them would have to forgive.
+# two numbers that a KL between them would have to forgive. 1,024 rather than the
-MNBT="${FRANK_QWEN38_MNBT:-2048}"
+# 2,048 ceiling: on this payload the arena charged for the step is the cache's only
+# remaining supplier, and 1,024 is the rung that funds every corpus document.
+MNBT="${FRANK_QWEN38_MNBT:-1024}"
 if ! [[ "$MNBT" =~ ^[0-9]+$ ]]; then
     echo "REFUSED: FRANK_QWEN38_MNBT must be a token count (got: $MNBT)" >&2
     exit 2
@@ -201,7 +217,7 @@ fi
 # The measuring shape. Identical for kldref and score by construction — this
 # array is the only place the campaign's KL numbers are defined.
 ARGS=(--tp 2 --placement expert_tiered --weights-disk-tier
-      --host-pool-mib 12288 --gpu-headroom-mib 768
+      --host-pool-mib "$HOST_POOL_MIB" --gpu-headroom-mib "$GPU_HEADROOM_MIB"
       --deterministic --num-speculative-tokens 0
       --kv-cache-dtype fp8 --max-model-len "$MAX_MODEL_LEN"
       --ngram-placement ram
@@ -210,9 +226,10 @@ MOUNTS=(-v "$MODELS:/models:ro")
 TAPENV=()
 case "$MODE" in
   kldref)
-    # Recording is prefill-only over four slots and the shared step, which is
+    # Recording is prefill-only over four slots at the campaign's single step size
-    # production's own operating point (frank-archon.env: FRANK_ARCHON_MAX_SEQS=4
+    # — 1,024, not production's 2,048 (frank-archon.env: FRANK_ARCHON_MAX_SEQS=4,
-    # with FRANK_ARCHON_MAX_BATCHED_TOKENS=2048). Corpus read-only, record dir rw.
+    # FRANK_ARCHON_MAX_BATCHED_TOKENS=2048): the wider step costs the cache the
+    # room the corpus needs. Corpus read-only, record dir rw.
     ARGS+=(--model "$MODEL_IN"
            --kld-record /data/kld-ref --kld-corpus /data/kld.jsonl
            --max-num-seqs 4 --max-num-batched-tokens "$MNBT")

tests/qwen38_reference.bats

@@ -29,7 +29,7 @@ RUNNER="$REPO_ROOT/scripts/run-qwen38-bf16.sh"
 # The reference shape, spelled exactly as run-qwen38-bf16.sh must pass it: this
 # is the contract the recorded reference is measured at, so a silent edit here
 # is a silent edit to every KLD number in the campaign.
-SHAPE="--tp 2 --placement expert_tiered --weights-disk-tier --host-pool-mib 12288 --gpu-headroom-mib 768 --deterministic --num-speculative-tokens 0 --kv-cache-dtype fp8 --max-model-len 32768 --ngram-placement ram --override-chat-template /models/tpl.jinja"
+SHAPE="--tp 2 --placement expert_tiered --weights-disk-tier --host-pool-mib 49152 --gpu-headroom-mib 96 --deterministic --num-speculative-tokens 0 --kv-cache-dtype fp8 --max-model-len 32768 --ngram-placement ram --override-chat-template /models/tpl.jinja"
 
 # oneline — both dry runs print one shell-quoted argument a line; joining them
 # lets an option and its value be asserted as the pair they are.