From a04c9d7d02747a795fed969ff88109169103fcc4 Mon Sep 17 00:00:00 2001
From: Frank <frank@dankware>
Date: Sun, 30 Aug 2026 00:04:47 -0700
Subject: [PATCH] =?UTF-8?q?frank-default-large:=20CTX=5FSIZE=2032768?=
 =?UTF-8?q?=E2=86=92401408,=20PARALLEL=201=E2=86=922=20(untested=20experim?=
 =?UTF-8?q?ent)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 scripts/frank-default-large.env | 24 ++++++++----------------
 scripts/serve-frank-large.sh    |  4 ++--
 2 files changed, 10 insertions(+), 18 deletions(-)

diff --git a/scripts/frank-default-large.env b/scripts/frank-default-large.env
index 6ba6896..93d1ef4 100644
--- a/scripts/frank-default-large.env
+++ b/scripts/frank-default-large.env
@@ -58,21 +58,18 @@ ALIAS=frank-default-large
 BIND_ADDR=0.0.0.0
 LISTEN_PORT=7770
 
-# Serving: 32768-token context on ONE slot (parallel 1), batch 4096 /
+# Serving: 401408-token context, TWO slots (parallel 2), batch 4096 /
 # ubatch 4096, 16+16 threads, f16 KV, 16 context checkpoints, 8 GiB
 # prompt cache. --fit off: auto-fit segfaults on this model's
 # hyper-connection/indexer config; -ngl 999 puts every layer on the
 # SVM-backed GPU pool.
 #
-# CTX_SIZE is capped at 32768 (the author's own quick-start). This model is
-# fully GPU-resident (87.06 GiB) — the PLE n-gram table no longer streams
-# from RAM like STRIX_LEAN did, so there is no room for a huge context
-# buffer: HIP heap is 119.3 GiB, and the context buffer alone is 50.8 GiB
-# at 131072 and 103 GiB at 262144 (verified live:
-# `cudaMalloc … failed: out of memory` on first graph alloc). At 32768 the
-# context buffer drops to ~13 GiB (~100 GiB total) and loads clean.
-CTX_SIZE=32768
-PARALLEL=1
+# CTX_SIZE=401408 / PARALLEL=2: untested experiment. At 32768×1 the
+# context buffer was ~13 GiB and total ~100 GiB in the 119.3 GiB HIP
+# heap. Doubling slots × 12× context could OOM — verify load.
+CTX_SIZE=401408
+PARALLEL=2
+
 BATCH_SIZE=4096
 UBATCH_SIZE=4096
 NUM_THREADS=16
@@ -106,13 +103,8 @@ HSA_OVERRIDE_GFX_VERSION=11.5.1
 
 VERBOSITY=4
 
+
 # Speculative decoding: ngram-mod only. The MTP draft head is retained
-# (agentionai Qwen3.8-Flash-Next-MTP-ROCmFP4-FAST-GGUF, 2.28 GiB) but does
-# NOT load on build-qwen38 (loader requires blk.0.hc_attn_norm.weight the
-# head lacks — see KNOWN CONSTRAINTS). If a compatible binary appears, set
-# SPEC_TYPE=draft-mtp,ngram-mod (n-max 3 measured best at 32.4 t/s / 0.612
-# acceptance; adaptive p-split measured worse, stays off). SPEC_TYPE=none
-# disables speculative decoding entirely.
 DRAFT_MODEL_DIR=/home/dank/models/mtp-Qwen3.8-Flash-Next-FAST
 DRAFT_MODEL_FILE=Qwen3.8-Flash-Next-MTP-ROCmFP4-FAST.gguf
 SPEC_TYPE=ngram-mod
diff --git a/scripts/serve-frank-large.sh b/scripts/serve-frank-large.sh
index a136c1d..751982b 100755
--- a/scripts/serve-frank-large.sh
+++ b/scripts/serve-frank-large.sh
@@ -45,8 +45,8 @@ fi
 : "${ALIAS:=frank-default-large}"
 : "${BIND_ADDR:=0.0.0.0}"
 : "${LISTEN_PORT:=7770}"
-: "${CTX_SIZE:=32768}"
-: "${PARALLEL:=1}"
+: "${CTX_SIZE:=401408}"
+: "${PARALLEL:=2}"
 : "${BATCH_SIZE:=4096}"
 : "${UBATCH_SIZE:=4096}"
 : "${NUM_THREADS:=16}"