From 0176ec0a5b227122ec680bdb228660439e0f6c48 Mon Sep 17 00:00:00 2001
From: Vuong Hoang <vh@phasefinal.com>
Date: Wed, 30 Sep 2026 09:02:45 -0700
Subject: [PATCH] =?UTF-8?q?scriberr:=20fit=20GPU=201=20beside=20intern-dec?=
 =?UTF-8?q?ision=20=E2=80=94=20120=20s=20Parakeet=20slices=20+=20expandabl?=
 =?UTF-8?q?e=5Fsegments?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Measured Parakeet peak on a 35-min file (n=3 each, deterministic): 300 s
9,384 MiB, 120 s 6,510, 60 s 5,976, 10 s 5,634 (fixed floor); with
expandable_segments 120 s 5,496 and 60 s 5,502. Verified 5,496 under the
recreated container's own env. Transcripts: 95.8% word-sequence similarity
vs 300 s, diffs mostly casing/punctuation.
---
 stacks/scriberr/.env.example |  6 ++++++
 stacks/scriberr/compose.yaml | 17 +++++++++++++++++
 2 files changed, 23 insertions(+)

diff --git a/stacks/scriberr/.env.example b/stacks/scriberr/.env.example
index 43490d1..5e24d73 100644
--- a/stacks/scriberr/.env.example
+++ b/stacks/scriberr/.env.example
@@ -42,3 +42,9 @@ SCRIBERR_SECURE_COOKIES=false
 # ⚠ That key also reaches PAID passthrough models (GLM, Kimi) on a shared
 #   tab — keep the configured model on a free local seat.
 # SCRIBERR_OPENAI_API_KEY=
+
+# GPU 1 memory budget (2026-09-30). Parakeet slice length in seconds and the
+# torch allocator mode; see compose.yaml for the measurements. Defaults apply
+# when unset; override only with a re-measured peak.
+# SCRIBERR_PARAKEET_CHUNK_SECS=120
+# SCRIBERR_PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
diff --git a/stacks/scriberr/compose.yaml b/stacks/scriberr/compose.yaml
index e46809b..1b85341 100644
--- a/stacks/scriberr/compose.yaml
+++ b/stacks/scriberr/compose.yaml
@@ -70,6 +70,23 @@ services:
       # and takes out the Parakeet + Sortformer backends (WhisperX survives).
       # `copy` trades a little disk and time for it actually working.
       - UV_LINK_MODE=${SCRIBERR_UV_LINK_MODE:-copy}
+      # ── GPU 1 memory budget (2026-09-30, Prime: Scriberr shares GPU 1 with
+      # intern-decision, which holds ~9.7 GB resting / 10.3 GB peak). ──────────
+      # Parakeet's buffered path cuts audio into slices of this many seconds
+      # (Scriberr reads it in parakeet_adapter.go for BOTH the "is this long
+      # audio" threshold and --chunk-len; upstream default 300). Measured peak
+      # GPU memory on a 35-min file, n=3 each, deterministic:
+      #   300 s 9,384 MiB · 120 s 6,510 · 60 s 5,976 · 10 s 5,634 (fixed floor)
+      #   120 s + expandable_segments 5,496 · 60 s + expandable_segments 5,502
+      # The floor, not the slice, dominates below ~120 s; expandable_segments is
+      # what removes the fragmentation on top of it. 120 s + expandable fits
+      # beside intern-decision even at both peaks (295 MiB spare). Transcripts
+      # change slightly: 95.8% word-sequence similarity vs 300 s (7,599 vs
+      # 7,645 words); the diffs are mostly casing/punctuation spread through
+      # the file, ~40 words at the 17 cuts. A-vs-A at 300 s: identical.
+      # Scriberr passes os.Environ() to the uv subprocess, so both reach NeMo.
+      - PARAKEET_CHUNK_THRESHOLD_SECS=${SCRIBERR_PARAKEET_CHUNK_SECS:-120}
+      - PYTORCH_CUDA_ALLOC_CONF=${SCRIBERR_PYTORCH_CUDA_ALLOC_CONF:-expandable_segments:True}
     deploy:
       resources:
         reservations: