Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: copy generated SNAC codes to the host once per decode (#15666)

Daniel Han0b618dfd1e42
diff --git a/studio/backend/core/inference/audio_codecs.py b/studio/backend/core/inference/audio_codecs.py
index 9d1f0969545..cc7010486a8 100644
--- a/studio/backend/core/inference/audio_codecs.py
+++ b/studio/backend/core/inference/audio_codecs.py
@@ -249,7 +249,8 @@ class AudioCodecManager:
if len(row) == 0:
raise ValueError("No valid audio codes found after START_OF_SPEECH token")
- codes = [t.item() - 128266 for t in row]
+ # One host copy; .tolist() yields the same Python scalars per element as .item()
+ codes = [t - 128266 for t in row.tolist()]
# Redistribute into 3 SNAC layers (7 codes per frame → 1+2+4)
layer_1, layer_2, layer_3 = [], [], []
diff --git a/studio/backend/tests/test_snac_decode_codes.py b/studio/backend/tests/test_snac_decode_codes.py
new file mode 100644
index 00000000000..0f118707b2c
--- /dev/null
+++ b/studio/backend/tests/test_snac_decode_codes.py
@@ -0,0 +1,117 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""decode_snac copies the codes to the host once and builds the same three SNAC layers."""
+
+import sys
+from pathlib import Path
+
+import pytest
+import torch
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+from core.inference.audio_codecs import AudioCodecManager # noqa: E402
+
+
+def _reference_layers(generated_ids):
+ """The previous per-token .item() unpacking, kept verbatim as the oracle."""
+ token_indices = (generated_ids == 128257).nonzero(as_tuple = True)
+ if len(token_indices[1]) > 0:
+ cropped = generated_ids[:, token_indices[1][-1] + 1 :]
+ else:
+ cropped = generated_ids
+ row = cropped[0]
+ row = row[row != 128258]
+ row = row[: (len(row) // 7) * 7]
+ codes = [t.item() - 128266 for t in row]
+ layer_1, layer_2, layer_3 = [], [], []
+ for i in range(len(codes) // 7):
+ layer_1.append(codes[7 * i])
+ layer_2.append(codes[7 * i + 1] - 4096)
+ layer_3.append(codes[7 * i + 2] - 8192)
+ layer_3.append(codes[7 * i + 3] - 12288)
+ layer_2.append(codes[7 * i + 4] - 16384)
+ layer_3.append(codes[7 * i + 5] - 20480)
+ layer_3.append(codes[7 * i + 6] - 24576)
+ return [torch.tensor(layer).unsqueeze(0) for layer in [layer_1, layer_2, layer_3]]
+
+
+class _RecordingSnac:
+ def __init__(self):
+ self.codes = None
+
+ def decode(self, codes):
+ self.codes = codes
+ return torch.zeros(1, 1, 16)
+
+
+def _frames(rng, n_frames):
+ offsets = 128266 + 4096 * torch.arange(7)
+ return (torch.randint(0, 4096, (n_frames, 7), generator = rng) + offsets).reshape(-1)
+
+
+def _cases():
+ rng = torch.Generator().manual_seed(0)
+ body = _frames(rng, 40)
+ m, eos = torch.tensor([128257]), torch.tensor([128258])
+ yield "marker", torch.cat([torch.tensor([1, 2, 128257]), body]).unsqueeze(0)
+ yield "two_markers_last_wins", torch.cat([m, _frames(rng, 3), m, body]).unsqueeze(0)
+ yield "no_marker", body.unsqueeze(0)
+ yield (
+ "eos_inside_and_tail",
+ torch.cat([m, body[:20], eos, body[20:], eos, torch.tensor([5, 6])]).unsqueeze(0),
+ )
+ yield (
+ "out_of_range",
+ torch.cat(
+ [m, torch.tensor([0, 1, 2, 3, 4, 5, 6, 10**9, -3, 128258, 7, 8, 9, 10])]
+ ).unsqueeze(0),
+ )
+ yield "int32", torch.cat([m, body]).unsqueeze(0).to(torch.int32)
+ yield "float32", torch.cat([m, body]).unsqueeze(0).to(torch.float32)
+ yield "batch_of_two", torch.stack([torch.cat([m, body]), torch.cat([m, body.flip(0)])])
+ yield "non_contiguous", torch.cat([m, body, body]).reshape(1, -1)[:, ::2]
+
+
[email protected]("name,generated", list(_cases()), ids = [c[0] for c in _cases()])
+def test_layers_match_the_per_token_path_without_item(name, generated, monkeypatch):
+ mgr = AudioCodecManager()
+ mgr._snac_model = _RecordingSnac()
+ calls = []
+ real_item = torch.Tensor.item
+ monkeypatch.setattr(torch.Tensor, "item", lambda self: calls.append(1) or real_item(self))
+ _, sr = mgr.decode_snac(generated, "cpu")
+ monkeypatch.setattr(torch.Tensor, "item", real_item)
+
+ assert calls == []
+ assert sr == 24000
+ expected = _reference_layers(generated)
+ assert len(mgr._snac_model.codes) == 3
+ for got, exp in zip(mgr._snac_model.codes, expected):
+ assert got.dtype == exp.dtype and got.device.type == "cpu"
+ assert torch.equal(got, exp)
+
+
+ "generated",
+ [
+ torch.tensor([[128257, 1, 2, 3]]),
+ torch.tensor([[128257, 128258, 128258]]),
+ torch.tensor([[5, 6, 128257]]),
+ ],
+)
+def test_too_few_codes_raise(generated):
+ mgr = AudioCodecManager()
+ mgr._snac_model = _RecordingSnac()
+ with pytest.raises(ValueError, match = "No valid audio codes found after START_OF_SPEECH token"):
+ mgr.decode_snac(generated, "cpu")
+
+
+def test_decode_entry_point_passes_the_same_codes():
+ token_ids = [128257] + _frames(torch.Generator().manual_seed(1), 5).tolist()
+ mgr = AudioCodecManager()
+ mgr._snac_model = _RecordingSnac()
+ mgr.decode("snac", "cpu", token_ids = token_ids)
+ for got, exp in zip(mgr._snac_model.codes, _reference_layers(torch.tensor([token_ids]))):
+ assert torch.equal(got, exp) and got.dtype == exp.dtype