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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team.
"""Moonshine's LM head is `proj_out`; untied on moonshine-streaming, so 4-bit loads quantized it and the
fused CE loss failed reading it as a plain weight (`mat1 and mat2 shapes cannot be multiplied`)."""
import ast
from pathlib import Path
import pytest
VISION_PATH = Path(__file__).resolve().parents[2] / "unsloth" / "models" / "vision.py"
def _architecture_skip_modules():
tree = ast.parse(VISION_PATH.read_text(encoding = "utf-8"))
node = next(
n
for n in tree.body
if isinstance(n, ast.FunctionDef) and n.name == "_architecture_skip_modules"
)
ns = {}
exec(compile(ast.Module([node], []), str(VISION_PATH), "exec"), ns)
return ns["_architecture_skip_modules"]
def test_moonshine_keeps_its_lm_head_unquantized():
skip = _architecture_skip_modules()
for model_types in (["moonshine_streaming", "moonshine_streaming_encoder"], ["moonshine"]):
assert "proj_out" in skip(model_types), model_types
for model_types in (["whisper"], ["llama"], ["gemma3n"], [], None):
assert "proj_out" not in skip(model_types), model_types
def test_only_the_head_matches_the_transformers_skip_matcher():
torch = pytest.importorskip("torch")
transformers = pytest.importorskip("transformers")
utils = pytest.importorskip("transformers.quantizers.quantizers_utils")
should_convert = getattr(utils, "should_convert_module", None)
config_class = getattr(transformers, "MoonshineStreamingConfig", None)
if should_convert is None or config_class is None:
pytest.skip("needs transformers 5 with Moonshine Streaming")
config = config_class(
encoder_config = {
"hidden_size": 32,
"intermediate_size": 64,
"num_hidden_layers": 1,
"num_attention_heads": 2,
"num_key_value_heads": 2,
},
hidden_size = 32,
intermediate_size = 64,
num_hidden_layers = 1,
num_attention_heads = 2,
num_key_value_heads = 2,
vocab_size = 128,
)
with torch.device("meta"):
model = transformers.AutoModelForSpeechSeq2Seq.from_config(config)
skip = _architecture_skip_modules()(["moonshine_streaming"])
kept = [
n
for n, m in model.named_modules()
if isinstance(m, torch.nn.Linear) and not should_convert(n, skip)
]
assert kept == ["proj_out"]