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 diffusion: do not let functools.wraps copy the prompt-cache transparent-wrapper marker (#17348)

Daniel Hancef96a9e89bc
diff --git a/studio/backend/core/inference/diffusion_prompt_cache.py b/studio/backend/core/inference/diffusion_prompt_cache.py
index 121a715aecd..bf976809f45 100644
--- a/studio/backend/core/inference/diffusion_prompt_cache.py
+++ b/studio/backend/core/inference/diffusion_prompt_cache.py
@@ -230,6 +230,16 @@ _MAX_FOLLOWED_METHODS = 64
TRANSPARENT_WRAPPER_ATTR = "_unsloth_prompt_cache_transparent"
+def mark_transparent_wrapper(fn: Any) -> Any:
+ """Self-reference, so a ``functools.wraps`` copy on an outer wrapper is not trusted."""
+ setattr(fn, TRANSPARENT_WRAPPER_ATTR, fn)
+ return fn
+
+
+def is_transparent_wrapper(fn: Any) -> bool:
+ return fn is not None and getattr(fn, TRANSPARENT_WRAPPER_ATTR, None) is fn
+
+
def _scoped_lora_key() -> bool:
return (os.environ.get(_ENV_SCOPED_LORA_KEY) or "").strip().lower() not in _FALSE_TOKENS
@@ -345,7 +355,7 @@ def _transparent_wrapper(cls: Any, attr: str, value: Any, instance: Any) -> bool
if not callable(value):
return False
for _ in range(8):
- if not getattr(value, TRANSPARENT_WRAPPER_ATTR, False):
+ if not is_transparent_wrapper(value):
break
value = getattr(value, "__wrapped__", None)
else:
diff --git a/studio/backend/core/inference/diffusion_qwenimage21_te.py b/studio/backend/core/inference/diffusion_qwenimage21_te.py
index 49e75348cfa..aa6673ef2ca 100644
--- a/studio/backend/core/inference/diffusion_qwenimage21_te.py
+++ b/studio/backend/core/inference/diffusion_qwenimage21_te.py
@@ -59,9 +59,9 @@ def skip_text_encoder_kv_cache(pipe: Any, logger: Any = None) -> bool:
handle.remove()
setattr(_get_qwen_prompt_embeds, _MARK, True)
- from .diffusion_prompt_cache import TRANSPARENT_WRAPPER_ATTR
+ from .diffusion_prompt_cache import mark_transparent_wrapper
- setattr(_get_qwen_prompt_embeds, TRANSPARENT_WRAPPER_ATTR, True)
+ mark_transparent_wrapper(_get_qwen_prompt_embeds)
pipe._get_qwen_prompt_embeds = _get_qwen_prompt_embeds
if logger is not None:
logger.info(
@@ -103,6 +103,10 @@ def select_final_hidden_state(pipe: Any, logger: Any = None) -> bool:
setattr(_get_qwen_prompt_embeds, _SELECT_MARK, True)
setattr(_get_qwen_prompt_embeds, _MARK, getattr(original, _MARK, False))
+ from .diffusion_prompt_cache import is_transparent_wrapper, mark_transparent_wrapper
+
+ if is_transparent_wrapper(original):
+ mark_transparent_wrapper(_get_qwen_prompt_embeds)
pipe._get_qwen_prompt_embeds = _get_qwen_prompt_embeds
if logger is not None:
logger.info(
diff --git a/studio/backend/tests/test_diffusion_prompt_cache.py b/studio/backend/tests/test_diffusion_prompt_cache.py
index 622e722f6cf..75e47dea114 100644
--- a/studio/backend/tests/test_diffusion_prompt_cache.py
+++ b/studio/backend/tests/test_diffusion_prompt_cache.py
@@ -1043,6 +1043,55 @@ def test_an_unmarked_wrapper_of_the_same_helper_keeps_the_full_key(monkeypatch):
assert not prompt_cache._encode_skips_denoisers(type(pipe), "encode_prompt", pipe)
+def test_a_wraps_copy_of_the_transparent_marker_keeps_the_full_key(monkeypatch):
+ # functools.wraps copies the no-K/V wrapper's marker onto this denoiser-reading wrapper
+ import functools
+
+ from core.inference.diffusion_qwenimage21_te import skip_text_encoder_kv_cache
+
+ monkeypatch.delenv(SCOPED_ENV, raising = False)
+ monkeypatch.delenv("UNSLOTH_QWEN21_TE_NO_KV", raising = False)
+ pipe = _Qwen21LikePipe({"transformer": ["style"]})
+ assert skip_text_encoder_kv_cache(pipe)
+ original = pipe._get_qwen_prompt_embeds
+
+ @functools.wraps(original)
+ def _get_qwen_prompt_embeds(prompt):
+ return original(prompt) * pipe.transformer.weight.detach().mean()
+
+ pipe._get_qwen_prompt_embeds = _get_qwen_prompt_embeds
+ assert not prompt_cache._encode_skips_denoisers(type(pipe), "encode_prompt", pipe)
+ prompt_cache.install(pipe)
+ out = []
+ for strength in (0.25, 0.75):
+ pipe._unsloth_loras = (("style", "/x/style.safetensors", strength),)
+ with torch.no_grad():
+ pipe.transformer.weight.fill_(strength)
+ out.append(float(pipe.encode_prompt("same")[0].flatten()[0]))
+ assert out[0] != out[1] and pipe.calls == 2
+
+
+class _Qwen21SelectablePipe(_Qwen21LikePipe):
+ def _get_qwen_prompt_embeds(self, prompt):
+ # select_final_hidden_state markers: norm.register_forward_hook outputs.hidden_states[-1]
+ return self.text_encoder(torch.full((1, 2), float(len(prompt))))
+
+
+def test_final_state_wrapper_over_the_no_kv_wrapper_keeps_the_scoped_key(monkeypatch):
+ from core.inference.diffusion_qwenimage21_te import (
+ select_final_hidden_state,
+ skip_text_encoder_kv_cache,
+ )
+
+ monkeypatch.delenv(SCOPED_ENV, raising = False)
+ monkeypatch.delenv("UNSLOTH_QWEN21_TE_NO_KV", raising = False)
+ monkeypatch.delenv("UNSLOTH_DIFFUSION_TE_SELECT_HIDDEN", raising = False)
+ pipe = _Qwen21SelectablePipe({"transformer": ["style"]})
+ assert skip_text_encoder_kv_cache(pipe)
+ assert select_final_hidden_state(pipe)
+ assert prompt_cache._encode_skips_denoisers(type(pipe), "encode_prompt", pipe)
+
+
def test_a_marked_wrapper_around_a_foreign_callable_keeps_the_full_key(monkeypatch):
monkeypatch.delenv(SCOPED_ENV, raising = False)
pipe = _Qwen21LikePipe({"transformer": ["style"]})