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: key learned peaks on the unsloth mirror a fresh install downloads (#13847)

Daniel Hanfc91656ff7b5
diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.py
index bb622cbfd55..10cf6379d5b 100644
--- a/studio/backend/core/inference/diffusion.py
+++ b/studio/backend/core/inference/diffusion.py
@@ -2624,15 +2624,41 @@ def _local_weights_digest(path: Any) -> Optional[str]:
def _peak_model_identity(
- repo_id: Any, gguf_filename: Optional[str], kind: Optional[str], base: Any
+ repo_id: Any,
+ gguf_filename: Optional[str],
+ kind: Optional[str],
+ base: Any,
+ fetch_base: Optional[str] = None,
) -> str:
"""The exact build for the learned-peak key: the denoiser source and its weights' revision / fingerprint, plus the
companion base and its revision for a GGUF / single-file load. ``unknown`` (no learned peak used or recorded)
- when a revision cannot be read."""
+ when a revision cannot be read. ``fetch_base``: the repo the load actually read ``base`` from."""
from .diffusion_component_sizes import checkpoint_fingerprint
-
+ from .diffusion_families import mirror_repo
+
+ def _fingerprint(
+ repo: Any,
+ filename: Optional[str] = None,
+ fetched: Optional[str] = None,
+ ) -> tuple[Any, Any]:
+ if fetched and str(fetched) != str(repo):
+ return fetched, checkpoint_fingerprint(fetched, filename)
+ revision = checkpoint_fingerprint(repo, filename)
+ if revision or filename or fetched:
+ return repo, revision
+ # No load decision: a fresh install may hold only the unsloth/ mirror.
+ mirror = mirror_repo(str(repo)) if repo else None
+ if mirror and mirror != str(repo):
+ mirrored = checkpoint_fingerprint(mirror)
+ if mirrored:
+ return mirror, mirrored
+ return repo, revision
+
+ pipeline = (kind or "pipeline") == "pipeline"
+ repo_id, revision = _fingerprint(
+ repo_id, gguf_filename, fetch_base if pipeline and str(base) == str(repo_id) else None
+ )
source = f"{repo_id}/{gguf_filename}" if gguf_filename else str(repo_id)
- revision = checkpoint_fingerprint(repo_id, gguf_filename)
if not revision:
return "unknown"
if str(revision).startswith("local:"):
@@ -2641,8 +2667,8 @@ def _peak_model_identity(
return "unknown"
revision = f"{revision}:{weights}"
identity = f"{source}@{revision}"
- if (kind or "pipeline") != "pipeline" and base and str(base) != str(repo_id):
- base_revision = checkpoint_fingerprint(base)
+ if not pipeline and base and str(base) != str(repo_id):
+ base, base_revision = _fingerprint(base, None, fetch_base)
if not base_revision:
return "unknown"
if str(base_revision).startswith("local:"):
@@ -9507,7 +9533,9 @@ class DiffusionBackend:
"; ".join(refined_plan.reasons),
)
plan = refined_plan
- peak_model = _peak_model_identity(repo_id, gguf_filename, kind, base)
+ peak_model = _peak_model_identity(
+ repo_id, gguf_filename, kind, base, fetch_base
+ )
if getattr(pipe, "_unsloth_loras", None):
peak_model = (
"unknown" # adapters baked in at load: a build the key cannot name
diff --git a/studio/backend/core/inference/video.py b/studio/backend/core/inference/video.py
index 7d65168f79f..dc707d3ace3 100644
--- a/studio/backend/core/inference/video.py
+++ b/studio/backend/core/inference/video.py
@@ -1431,6 +1431,9 @@ class _VideoLoadState:
# resolves a bare device. Mirrors the image backend.
placed_ordinal: Optional[int] = None
gguf_filename: Optional[str] = None
+ # The repo the load actually read ``base_repo`` from (the unsloth mirror on a fresh install): the learned-peak key
+ # names it. None: no load decision (the key falls back to the mirror when the upstream is not cached).
+ fetch_base: Optional[str] = None
# Content-detected checkpoint variant identifier (LTX-2.3 "ltx-2.3-22b-distilled" / "-dev"), read before the names.
variant_id: Optional[str] = None
# Resident MiniMax-H3 denoiser partition, if any.
@@ -3653,6 +3656,7 @@ def _calibration_revision(state: Any) -> Optional[str]:
getattr(state, "gguf_filename", None),
getattr(state, "kind", None),
getattr(state, "base_repo", None),
+ getattr(state, "fetch_base", None),
)
try:
pipe._unsloth_calibration_revision = revision or ""
@@ -9690,6 +9694,7 @@ class VideoBackend:
repo_id = repo_id,
display_repo_id = display_repo_id,
base_repo = base,
+ fetch_base = fetch_base,
device = device,
gpu_ordinal = target.ordinal,
placed_ordinal = placed_cuda_ordinal(target),
@@ -10937,6 +10942,7 @@ class VideoBackend:
repo_id = repo_id,
display_repo_id = display_repo_id,
base_repo = base,
+ fetch_base = fetch_base,
device = device,
gpu_ordinal = umem_target.ordinal,
placed_ordinal = placed_cuda_ordinal(umem_target),
diff --git a/studio/backend/tests/test_diffusion_learned_peaks.py b/studio/backend/tests/test_diffusion_learned_peaks.py
index a8c0308a702..a4016fede98 100644
--- a/studio/backend/tests/test_diffusion_learned_peaks.py
+++ b/studio/backend/tests/test_diffusion_learned_peaks.py
@@ -1209,7 +1209,109 @@ def test_peak_model_identity_names_the_whole_build(monkeypatch):
src = (__import__("pathlib").Path(dm.__file__).parent / "diffusion.py").read_text(
encoding = "utf-8"
)
- assert "peak_model = _peak_model_identity(repo_id, gguf_filename, kind, base)" in src
+ # whitespace-free: the formatter may wrap the call
+ assert "peak_model=_peak_model_identity(repo_id,gguf_filename,kind,base,fetch_base)" in "".join(
+ src.split()
+ )
+
+
+def test_peak_model_identity_reads_the_unsloth_mirror_on_a_fresh_install(monkeypatch):
+ """Studio fetches a mirrored vendor base from unsloth/: with only the mirror cached, the key must still name a
+ revision (the mirror's), or learned peaks are never recorded nor read on a default install."""
+ import core.inference.diffusion as diffusion
+ import core.inference.diffusion_component_sizes as sizes
+ import core.inference.diffusion_families as families
+
+ revisions = {
+ ("unsloth/FLUX.2-klein-4B", None): "m1",
+ ("unsloth/dit-GGUF", "dit-Q4.gguf"): "g1",
+ }
+ monkeypatch.setattr(
+ sizes, "checkpoint_fingerprint", lambda repo, filename = None: revisions.get((repo, filename))
+ )
+ monkeypatch.setattr(
+ families,
+ "mirror_repo",
+ lambda repo: {"black-forest-labs/flux.2-klein-4b": "unsloth/FLUX.2-klein-4B"}.get(
+ (repo or "").lower()
+ ),
+ )
+ assert (
+ diffusion._peak_model_identity("black-forest-labs/FLUX.2-klein-4B", None, "pipeline", None)
+ == "unsloth/FLUX.2-klein-4B@m1"
+ )
+ assert diffusion._peak_model_identity(
+ "unsloth/dit-GGUF", "dit-Q4.gguf", "gguf", "black-forest-labs/FLUX.2-klein-4B"
+ ) == ("unsloth/dit-GGUF/dit-Q4.gguf@g1+unsloth/FLUX.2-klein-4B@m1")
+ # No mirror and no cached revision: still unknown.
+ assert diffusion._peak_model_identity("org/unmirrored", None, "pipeline", None) == "unknown"
+
+
+def test_peak_model_identity_keys_on_the_repo_the_load_fetched(monkeypatch):
+ """An upstream cache holding only model_index.json while the load pulled the weights from the mirror: the key
+ names the mirror the load read, not the upstream id whose index happens to be cached (and vice versa)."""
+ import core.inference.diffusion as diffusion
+ import core.inference.diffusion_component_sizes as sizes
+ import core.inference.diffusion_families as families
+
+ revisions = {
+ ("black-forest-labs/FLUX.2-klein-4B", None): "u1",
+ ("unsloth/FLUX.2-klein-4B", None): "m1",
+ ("unsloth/dit-GGUF", "dit-Q4.gguf"): "g1",
+ }
+ monkeypatch.setattr(
+ sizes, "checkpoint_fingerprint", lambda repo, filename = None: revisions.get((repo, filename))
+ )
+ monkeypatch.setattr(
+ families,
+ "mirror_repo",
+ lambda repo: {"black-forest-labs/flux.2-klein-4b": "unsloth/FLUX.2-klein-4B"}.get(
+ (repo or "").lower()
+ ),
+ )
+ up, mirror = "black-forest-labs/FLUX.2-klein-4B", "unsloth/FLUX.2-klein-4B"
+ assert diffusion._peak_model_identity(up, None, "pipeline", up, mirror) == f"{mirror}@m1"
+ assert diffusion._peak_model_identity(up, None, "pipeline", up, up) == f"{up}@u1"
+ assert (
+ diffusion._peak_model_identity("unsloth/dit-GGUF", "dit-Q4.gguf", "gguf", up, mirror)
+ == f"unsloth/dit-GGUF/dit-Q4.gguf@g1+{mirror}@m1"
+ )
+ # The load read the upstream bytes but only the mirror's revision is cached: no key, never the mirror's.
+ del revisions[(up, None)]
+ assert diffusion._peak_model_identity(up, None, "pipeline", up, up) == "unknown"
+ assert diffusion._peak_model_identity(up, None, "pipeline", up, mirror) == f"{mirror}@m1"
+
+
+def test_video_peak_key_names_the_repo_the_load_fetched(monkeypatch):
+ """An upstream video cache holding only model_index.json while the load staged the complete mirror: the video
+ calibration key names the mirror the weights came from, not the partial upstream snapshot."""
+ import core.inference.diffusion_component_sizes as sizes
+ import core.inference.video as video
+
+ up, mirror = "Wan-AI/Wan2.2-TI2V-5B-Diffusers", "unsloth/Wan2.2-TI2V-5B-Diffusers"
+ revisions = {(up, None): "u1", (mirror, None): "m1"}
+ monkeypatch.setattr(
+ sizes, "checkpoint_fingerprint", lambda repo, filename = None: revisions.get((repo, filename))
+ )
+
+ def state(fetch_base):
+ return types.SimpleNamespace(
+ pipe = types.SimpleNamespace(),
+ repo_id = up,
+ base_repo = up,
+ kind = "pipeline",
+ gguf_filename = None,
+ fetch_base = fetch_base,
+ )
+
+ assert video._calibration_revision(state(mirror)) == f"{mirror}@m1"
+ assert video._calibration_revision(state(up)) == f"{up}@u1"
+ # no load decision: the upstream's own revision, as before
+ assert video._calibration_revision(state(None)) == f"{up}@u1"
+ src = "".join((__import__("pathlib").Path(video.__file__)).read_text(encoding = "utf-8").split())
+ assert "fetch_base" in video._VideoLoadState.__dataclass_fields__
+ # both torch loaders (load_pipeline, the H3 modular pipeline) hand their fetch repo to the state
+ assert src.count("base_repo=base,fetch_base=fetch_base,") == 2
def test_phase_table_seed_needs_its_measured_configuration(klein_pipe):