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.

Fused preference loss tests: build the CPU float64 model from transformers' own Llama (#15301)

Daniel Han7132f8a5f6d9
diff --git a/tests/test_fused_preference_loss.py b/tests/test_fused_preference_loss.py
index 1b17972ad6f..59448bb7ccb 100644
--- a/tests/test_fused_preference_loss.py
+++ b/tests/test_fused_preference_loss.py
@@ -26,7 +26,28 @@ import torch
transformers = pytest.importorskip("transformers")
trl = pytest.importorskip("trl")
-from transformers import LlamaConfig, LlamaForCausalLM
+from transformers import LlamaConfig
+
+
+def _pristine_llama_for_causal_lm():
+ """transformers' own LlamaForCausalLM, loaded fresh from its file. Once a GPU test in the same
+ process loads a model through unsloth, unsloth's patches (CUDA rotary caches, fast forwards) sit
+ on the shared Llama classes, and this file's float64 CPU model would run through them."""
+ import importlib.util
+ import sys
+
+ import transformers.models.llama.modeling_llama as installed
+
+ name = "transformers.models.llama._pristine_modeling_llama"
+ if name not in sys.modules:
+ spec = importlib.util.spec_from_file_location(name, installed.__file__)
+ module = importlib.util.module_from_spec(spec)
+ sys.modules[name] = module
+ spec.loader.exec_module(module)
+ return sys.modules[name].LlamaForCausalLM
+
+
+LlamaForCausalLM = _pristine_llama_for_causal_lm()
from transformers.modeling_outputs import CausalLMOutputWithPast
ROOT = pathlib.Path(__file__).resolve().parents[1]