From 41055307d47ed94d12a1d7b70ee9ebef42cfc698 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Rafa=C5=82=20Leszko?= <rafal@livepeer.org>
Date: Mon, 15 Jun 2026 14:15:59 +0000
Subject: [PATCH] feat(trt): register 30s decoder profile for the loop-focused
 workflow

Acts on the latent-size experiment finding (loop window of 20-30s is
viable). Adds a 30.0 TRT engine profile so the loop-focused / short-
source path runs on a decoder opt-tuned for a 750-frame (30s) window
instead of borrowing the 60s engine.

- The walk-window path (acestep/streaming/session.py) already resolves
  decoder + vae_decode at walk_window_s; with this profile,
  walk_window_s=30 now picks a real 30s decoder. Sources <=30s also pick
  it directly.
- Decoder-only profile: it reuses the existing 60s VAE engines. The walk
  path keeps vae_encode at full-song size, and the 60s vae_decode accepts
  <=1500-frame inputs, so 750-frame decodes need no dedicated engine. A
  sub-60s VAE engine is also not buildable today (TRT optimization-profile
  failure on Blackwell, see docs/TRT.md).
- Build: python -m acestep.engine.trt.build --all --decoder-only --duration 30
  Engine binaries live under ~/.daydream-scope (not in the repo), same as
  the 60s/120s/240s engines.

Validated on a 5090: 30s decoder loads via TRT, generates a 750-frame
latent, decodes OK, LoRA refit works. vs the same 750-frame window on the
60s engine: ~12% faster warm generate (0.121s vs 0.137s), ~0.25 GB less
process VRAM. resolve(): <=30s -> 30s engine, 31-60s -> 60s (unchanged).

Quality at a 30s loop window is the ear test this unblocks (set
walk_window=true, walk_window_s=30 in web/public/config.json).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 acestep/paths.py | 13 +++++++++++++
 docs/TRT.md      | 19 +++++++++++++++++++
 2 files changed, 32 insertions(+)

diff --git a/acestep/paths.py b/acestep/paths.py
index 618bdca8..35c08d71 100644
--- a/acestep/paths.py
+++ b/acestep/paths.py
@@ -312,6 +312,19 @@ def trt_engine_path(engine_name: str) -> Path:
 # scripts/benchmarks/vram_60s_vs_240s_results.md. Pick the smallest profile
 # that fits the audio (see `select_trt_engines` and `available_trt_engines`).
 _TRT_ENGINE_PROFILES: dict[float, dict[str, str]] = {
+    # 30s decoder for the loop-focused / short-source workflow: a smaller
+    # seq_max (750 frames vs 1500) reserves less DiT activation workspace
+    # (VRAM headroom) and is opt-tuned at 750 for lower param-update
+    # latency. The VAE engines are intentionally the 60s builds — the
+    # walk-window path keeps vae_encode sized to the full song, and the
+    # 60s vae_decode already accepts ≤1500 frames, so 750-frame decodes
+    # need no dedicated engine. Build with
+    # ``python -m acestep.engine.trt.build --all --duration 30 --decoder-only``.
+    30.0: {
+        "decoder": "spectral_decoder_mixed_refit_b8_30s",
+        "vae_encode": "vae_encode_fp16_60s",
+        "vae_decode": "vae_decode_fp16_60s",
+    },
     60.0: {
         "decoder": "spectral_decoder_mixed_refit_b8_60s",
         "vae_encode": "vae_encode_fp16_60s",
diff --git a/docs/TRT.md b/docs/TRT.md
index 522c8a08..eb4cfa0e 100644
--- a/docs/TRT.md
+++ b/docs/TRT.md
@@ -176,6 +176,25 @@ Build only decoder engines:
 uv run python -m acestep.engine.trt.build --all --decoder-only --duration 60
 ```
 
+Build the 30s loop-window decoder (decoder-only):
+
+```powershell
+uv run python -m acestep.engine.trt.build --all --decoder-only --duration 30
+```
+
+The `30.0` profile in `acestep/paths.py` serves the loop-focused /
+short-source workflow (`walk_window_s = 30`, or any source ≤ 30s). Its
+`seq_max` is 750 frames (vs 1500 for 60s), so the DiT context reserves
+less activation workspace and is opt-tuned at 750 for lower
+param-update latency. It is **decoder-only**: the profile reuses the
+existing 60s VAE engines, because the walk-window path keeps
+`vae_encode` sized to the full song and the 60s `vae_decode` already
+accepts ≤ 1500-frame inputs. (A dedicated sub-60s VAE engine is not
+needed and currently fails the TRT optimization-profile check on
+Blackwell — see `RTX 5090 / Blackwell VAE Builds` above.) Measured on a
+5090 vs running the same 750-frame window on the 60s engine: ~12% faster
+warm generate (0.121s vs 0.137s) and ~0.25 GB less process VRAM.
+
 Build only VAE engines:
 
 ```powershell