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: compose the FP8 and NVFP4 compile cache tags in a fixed order (#15692)

Daniel Hand513620286a5
diff --git a/studio/backend/core/inference/diffusion_compile_config.py b/studio/backend/core/inference/diffusion_compile_config.py
index f801955ecd6..701ba603d6c 100644
--- a/studio/backend/core/inference/diffusion_compile_config.py
+++ b/studio/backend/core/inference/diffusion_compile_config.py
@@ -138,3 +138,16 @@ def _reset_for_tests() -> None:
with _LOCK:
_KNOBS.clear()
_generation += 1
+
+
+_STUDIO_CACHE_TAG_PREFIXES = ("unsloth_fp8_unit_scale=", "unsloth_nvfp4_act_quant=")
+
+
+def canonical_compile_cache_tag(current: str, prefix: str, tag: str | None) -> str:
+ """``prefix``'s token set to ``tag`` (None drops it); Studio's sorted so setter order never changes torch's cache key."""
+ parts = [p for p in current.split(";") if p and not p.startswith(prefix)]
+ if tag:
+ parts.append(prefix + tag)
+ foreign = [p for p in parts if not p.startswith(_STUDIO_CACHE_TAG_PREFIXES)]
+ owned = sorted(p for p in parts if p.startswith(_STUDIO_CACHE_TAG_PREFIXES))
+ return ";".join(foreign + owned)
diff --git a/studio/backend/core/inference/diffusion_fp8_unit_scale.py b/studio/backend/core/inference/diffusion_fp8_unit_scale.py
index 0b2c552198a..4c1381f9c9e 100644
--- a/studio/backend/core/inference/diffusion_fp8_unit_scale.py
+++ b/studio/backend/core/inference/diffusion_fp8_unit_scale.py
@@ -61,11 +61,10 @@ def _set_compile_cache_tag(tag: Optional[str]) -> None:
current = getattr(_cc, "cache_key_tag", None)
if not isinstance(current, str):
return
- parts = [p for p in current.split(";") if p and not p.startswith(_CACHE_TAG_PREFIX)]
- if tag:
- parts.append(_CACHE_TAG_PREFIX + tag)
+ from .diffusion_compile_config import canonical_compile_cache_tag
+
try:
- _cc.cache_key_tag = ";".join(parts)
+ _cc.cache_key_tag = canonical_compile_cache_tag(current, _CACHE_TAG_PREFIX, tag)
except Exception: # noqa: BLE001
pass
diff --git a/studio/backend/core/inference/diffusion_nvfp4_act_quant.py b/studio/backend/core/inference/diffusion_nvfp4_act_quant.py
index 102c068a686..fe9e462ff67 100644
--- a/studio/backend/core/inference/diffusion_nvfp4_act_quant.py
+++ b/studio/backend/core/inference/diffusion_nvfp4_act_quant.py
@@ -128,11 +128,10 @@ def _set_compile_cache_tag(tag: str | None) -> None:
current = getattr(_cc, "cache_key_tag", None)
if not isinstance(current, str):
return
- parts = [p for p in current.split(";") if p and not p.startswith(_CACHE_TAG_PREFIX)]
- if tag:
- parts.append(_CACHE_TAG_PREFIX + tag)
+ from .diffusion_compile_config import canonical_compile_cache_tag
+
try:
- _cc.cache_key_tag = ";".join(parts)
+ _cc.cache_key_tag = canonical_compile_cache_tag(current, _CACHE_TAG_PREFIX, tag)
except Exception: # noqa: BLE001
pass
diff --git a/studio/backend/tests/test_diffusion_cache_tag_order.py b/studio/backend/tests/test_diffusion_cache_tag_order.py
new file mode 100644
index 00000000000..55581db554d
--- /dev/null
+++ b/studio/backend/tests/test_diffusion_cache_tag_order.py
@@ -0,0 +1,57 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""FP8 + NVFP4 cache tags compose to one string whatever order they were set in (torch keys its caches on it)."""
+
+from __future__ import annotations
+
+import pytest
+
+pytest.importorskip("torch")
+import torch.compiler.config as cc # noqa: E402
+
+from core.inference import diffusion_fp8_unit_scale as fp8 # noqa: E402
+from core.inference import diffusion_nvfp4_act_quant as nvfp4 # noqa: E402
+
+
[email protected](autouse = True)
+def _restore_tag():
+ saved = cc.cache_key_tag
+ yield
+ cc.cache_key_tag = saved
+
+
+def _apply(steps, start = ""):
+ cc.cache_key_tag = start
+ for step in steps:
+ step()
+ return cc.cache_key_tag
+
+
+FP8 = lambda: fp8._set_compile_cache_tag("256") # noqa: E731
+NVFP4_ON = lambda: nvfp4._set_compile_cache_tag(nvfp4._TAG_ON) # noqa: E731
+
+
+def test_setter_order_does_not_change_the_tag():
+ assert _apply([FP8, NVFP4_ON]) == _apply([NVFP4_ON, FP8])
+
+
+def test_setting_again_does_not_change_the_tag():
+ once = _apply([FP8, NVFP4_ON])
+ assert _apply([FP8, NVFP4_ON, FP8]) == once
+ assert _apply([FP8, NVFP4_ON, NVFP4_ON, FP8, NVFP4_ON]) == once
+
+
+def test_values_and_removal_still_count():
+ on = _apply([FP8, NVFP4_ON])
+ off = _apply([FP8, lambda: nvfp4._set_compile_cache_tag(nvfp4._TAG_OFF)])
+ other_rows = _apply([lambda: fp8._set_compile_cache_tag("512"), NVFP4_ON])
+ assert len({on, off, other_rows}) == 3
+ removed = _apply([FP8, NVFP4_ON, lambda: fp8._set_compile_cache_tag(None)])
+ assert "unsloth_fp8_unit_scale=" not in removed and "unsloth_nvfp4_act_quant=1" in removed
+
+
+def test_foreign_tokens_are_kept_in_place():
+ tag = _apply([NVFP4_ON, FP8], start = "user_tag;other=2")
+ assert tag.split(";")[:2] == ["user_tag", "other=2"]
+ assert tag == _apply([FP8, NVFP4_ON], start = "user_tag;other=2")