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: fall back when a global forward hook is registered (#15226)

Shubhankar Tripathy3e12d46c2a0a
diff --git a/tests/test_fused_preference_loss.py b/tests/test_fused_preference_loss.py
index 59448bb7ccb..be51776cd3e 100644
--- a/tests/test_fused_preference_loss.py
+++ b/tests/test_fused_preference_loss.py
@@ -731,8 +731,11 @@ def test_rl_replacement_injects_dispatch_only_when_enabled(monkeypatch):
)
[email protected]("kind", ["forward_hook", "pre_hook", "hf_hook", "subclass"])
-def test_hooked_or_custom_head_falls_back(dpo_module, monkeypatch, kind):
+ "kind",
+ ["forward_hook", "pre_hook", "global_forward_hook", "global_pre_hook", "hf_hook", "subclass"],
+)
+def test_hooked_or_custom_head_falls_back(dpo_module, monkeypatch, request, kind):
"""The kernel reads the head's weight directly, so a head whose forward does more is left to TRL."""
monkeypatch.setattr(dpo_module, "_logit_transforms", lambda unwrapped: (0.0, 0.0, 0.0))
batch = _orpo_batch(seed = 4)
@@ -742,6 +745,14 @@ def test_hooked_or_custom_head_falls_back(dpo_module, monkeypatch, kind):
head.register_forward_hook(lambda m, i, o: o * 2)
elif kind == "pre_hook":
head.register_forward_pre_hook(lambda m, i: (i[0] * 2,))
+ elif kind == "global_forward_hook":
+ from torch.nn.modules.module import register_module_forward_hook
+ handle = register_module_forward_hook(lambda m, i, o: o * 2 if m is head else None)
+ request.addfinalizer(handle.remove)
+ elif kind == "global_pre_hook":
+ from torch.nn.modules.module import register_module_forward_pre_hook
+ handle = register_module_forward_pre_hook(lambda m, i: (i[0] * 2,) if m is head else None)
+ request.addfinalizer(handle.remove)
elif kind == "hf_hook":
head._hf_hook = object()
else:
diff --git a/unsloth/models/dpo.py b/unsloth/models/dpo.py
index d5dec1f2666..4242f398391 100644
--- a/unsloth/models/dpo.py
+++ b/unsloth/models/dpo.py
@@ -82,11 +82,13 @@ def _uses_fsdp1(trainer, model):
return cache[model]
-def _global_module_backward_hooks():
- # torch.nn.modules.module.register_module_full_backward_(pre_)hook applies to every module.
+def _global_module_hooks():
+ # torch's register_module_{forward,full_backward}(_pre)_hook tables apply to the head too.
from torch.nn.modules import module as _module
return bool(
- getattr(_module, "_global_backward_hooks", None)
+ getattr(_module, "_global_forward_hooks", None)
+ or getattr(_module, "_global_forward_pre_hooks", None)
+ or getattr(_module, "_global_backward_hooks", None)
or getattr(_module, "_global_backward_pre_hooks", None)
)
@@ -120,7 +122,7 @@ def _output_head(trainer, model):
or head._forward_pre_hooks
or head._backward_hooks
or getattr(head, "_backward_pre_hooks", None)
- or _global_module_backward_hooks()
+ or _global_module_hooks()
):
return None, None
weight = getattr(head, "weight", None)