From 927d59f8055b177d052871a37e5a210d0a3eedb2 Mon Sep 17 00:00:00 2001
From: javis-bot <claude@tkrmagid.kr>
Date: Sun, 14 Jun 2026 02:22:36 +0900
Subject: [PATCH] perf: run MeloTTS on the GPU (cu128 torch) + warm CUDA at
 startup

CPU MeloTTS serialised under concurrent load (whisper STT + bot) and blew
voice-reply TTS to 7-8s. Install the Blackwell-verified cu128 torch in the
melo venv, select the GPU via MELO_DEVICE=cuda, and do a throwaway synth at
worker startup so the one-off CUDA kernel-init (~5s) doesn't land on the
user's first reply. Measured: ~0.3s/sentence on GPU vs ~1.2-2.6s on CPU.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 bridge/melo_worker.py | 14 ++++++++++++++
 docker-compose.yml    |  3 +++
 docker/setup-melo.sh  | 17 ++++++++++-------
 3 files changed, 27 insertions(+), 7 deletions(-)

diff --git a/bridge/melo_worker.py b/bridge/melo_worker.py
index 073183e..4da17eb 100644
--- a/bridge/melo_worker.py
+++ b/bridge/melo_worker.py
@@ -66,6 +66,20 @@ def _ensure_model() -> None:
             speaker_id = spk_map[LANGUAGE] if LANGUAGE in spk_map else spk_map[keys[0]]
             _model = model
             _speaker_id = speaker_id
+            # Warm the GPU once at load: the first CUDA synth pays a one-off
+            # kernel-init cost (~5s) that would otherwise land on the user's
+            # first reply. A throwaway synth here moves it to startup. No-op
+            # cost on CPU.
+            try:
+                with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as _wt:
+                    _wp = _wt.name
+                model.tts_to_file("워밍업", speaker_id, _wp, speed=SPEED)
+                try:
+                    os.unlink(_wp)
+                except OSError:
+                    pass
+            except Exception as _we:  # pragma: no cover
+                print(f"[melo-worker] warmup synth skipped: {_we}", flush=True)
             print(
                 f"[melo-worker] ready (lang={LANGUAGE} speed={SPEED} "
                 f"device={DEVICE} speakers={list(spk_map.keys())})",
diff --git a/docker-compose.yml b/docker-compose.yml
index f7b0916..4be1b75 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -67,6 +67,9 @@ services:
       WHISPER_MODEL: ${WHISPER_MODEL:-medium}
       WHISPER_DEVICE: ${WHISPER_DEVICE:-cuda}
       WHISPER_COMPUTE_TYPE: ${WHISPER_COMPUTE_TYPE:-float16}
+      # MeloTTS on the GPU (cu128 torch baked by docker/setup-melo.sh). CPU synth
+      # serialised under load and pushed TTS to 7-8s; GPU does ~0.3s/sentence.
+      MELO_DEVICE: ${MELO_DEVICE:-cuda}
       # Optional single-language lock for replies (empty = user's own language).
       OUTPUT_LANGUAGE: ${OUTPUT_LANGUAGE:-}
       # Drop the pre-loop planner LLM call to cut voice-reply latency on small
diff --git a/docker/setup-melo.sh b/docker/setup-melo.sh
index 1a363d0..5202bd7 100755
--- a/docker/setup-melo.sh
+++ b/docker/setup-melo.sh
@@ -9,8 +9,11 @@
 #   - It isolates the heavy torch/transformers stack from the slim bridge env,
 #     which pins numpy<2 for faster-whisper.
 #
-# torch is pinned to the CPU build: TTS runs on CPU so the GPU stays reserved
-# for Ollama + Whisper, and we avoid pulling multi-GB CUDA wheels.
+# torch is the CUDA (cu128) build so MeloTTS runs on the GPU alongside Ollama +
+# Whisper. CPU synth serialised under concurrent load (whisper STT + bot) and
+# blew TTS up to 7-8s per reply; on the GPU a sentence synthesises in ~0.3s.
+# cu128 is the Blackwell (sm_120) wheel verified on this host's RTX 5050.
+# The worker selects the device via MELO_DEVICE=cuda (compose).
 # ============================================================================
 set -euxo pipefail
 
@@ -29,11 +32,11 @@ rm -rf /var/lib/apt/lists/*
 python3.11 -m venv /opt/melo
 /opt/melo/bin/pip install --no-cache-dir --upgrade pip wheel setuptools
 
-# CPU-only torch first, so MeloTTS's unpinned `torch` dep is already satisfied
-# and pip does not pull the CUDA build. Pinned for reproducible rebuilds (these
-# are the versions the CPU index resolved when this layer was verified).
-/opt/melo/bin/pip install --no-cache-dir torch==2.12.0 torchaudio==2.11.0 \
-    --index-url https://download.pytorch.org/whl/cpu
+# CUDA (cu128) torch first, so MeloTTS's unpinned `torch` dep is already
+# satisfied with the GPU build. Pinned to the Blackwell-verified versions
+# (2.11.0+cu128) for reproducible rebuilds.
+/opt/melo/bin/pip install --no-cache-dir torch==2.11.0+cu128 torchaudio==2.11.0+cu128 \
+    --index-url https://download.pytorch.org/whl/cu128
 
 # MeloTTS from GitHub. The PyPI sdist is broken (its setup.py reads a
 # requirements.txt that is not shipped in the sdist), so install from the repo.