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.

Fix pipeline denoiser downloads labeled as required assets (#16465)

Lee Jackson40bb5d6e547a
diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.py
index dbaa3b5974b..45ee9e6d04a 100644
--- a/studio/backend/core/inference/diffusion.py
+++ b/studio/backend/core/inference/diffusion.py
@@ -6958,6 +6958,7 @@ class DiffusionBackend:
scoped_files: dict[str, list[str]] = {}
scoped_gguf: dict[str, Optional[str]] = {}
missing_checkpoints: set[str] = set()
+ missing_denoisers: set[str] = set()
def add_missing_entry(
repo: str,
@@ -6966,6 +6967,7 @@ class DiffusionBackend:
*,
gguf: Optional[str] = None,
checkpoint: bool = False,
+ hosted_denoiser: bool = False,
) -> None:
"""Adds only files neither cache root resolves, probed at the plan's revision; one entry per repo, else two jobs fight over progress and cancel."""
@@ -7012,6 +7014,8 @@ class DiffusionBackend:
)
if checkpoint:
missing_checkpoints.add(repo)
+ if hosted_denoiser:
+ missing_denoisers.add(repo)
for entry in entries:
if entry["repo_id"] != repo:
continue
@@ -7023,6 +7027,7 @@ class DiffusionBackend:
)
entry["gguf_filename"] = scoped_gguf[repo]
entry["checkpoint"] = repo in missing_checkpoints
+ entry["hosted_denoiser"] = repo in missing_denoisers
return
entries.append(
{
@@ -7034,6 +7039,7 @@ class DiffusionBackend:
),
"gguf_filename": scoped_gguf[repo],
"checkpoint": repo in missing_checkpoints,
+ "hosted_denoiser": repo in missing_denoisers,
}
)
@@ -7073,7 +7079,14 @@ class DiffusionBackend:
if dit_prequant is not None:
# Staged, not just counted: leaving it out means a multi-GB inline pull under the load lock.
prequant_repo, prequant_file, prequant_size = dit_prequant
- add_missing_entry(prequant_repo, [prequant_file], {prequant_file: prequant_size})
+ # Part of a pipeline pick, replacing its dense denoiser; a GGUF pick keeps it a required asset.
+ add_missing_entry(
+ prequant_repo,
+ [prequant_file],
+ {prequant_file: prequant_size},
+ checkpoint = skip_transformer_weights,
+ hosted_denoiser = skip_transformer_weights,
+ )
for ref in component_overrides.files.values() if component_overrides else ():
if not ref.is_hub:
continue
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index e10bb664172..b90f0b169cc 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -5951,6 +5951,11 @@ class DiffusionDownloadPlanEntry(BaseModel):
"planner has to say so: a gated base is staged from its ungated mirror, so the entry's "
"repo id is not the id the caller picked and the role cannot be recovered downstream.",
)
+ hosted_denoiser: bool = Field(
+ False,
+ description = "This checkpoint entry holds a hosted denoiser replacing the pipeline's "
+ "dense weights, which a pinned snapshot of that pipeline does not contain.",
+ )
class DiffusionDownloadPlanResponse(BaseModel):
diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.py
index 9c0ebee53c4..44326f15b68 100644
--- a/studio/backend/tests/test_diffusion_backend.py
+++ b/studio/backend/tests/test_diffusion_backend.py
@@ -8814,6 +8814,11 @@ def test_download_plan_counts_the_hosted_prequant_in_the_required_footprint(monk
assert plan["required_bytes"] - baseline["required_bytes"] == 6 * GB
prequant = next(e for e in plan["entries"] if e["repo_id"] == "unsloth/Z-Image-Turbo-FP8")
assert prequant["checkpoint"] is False
+ assert prequant["hosted_denoiser"] is False
+ selected = [entry for entry in plan["entries"] if entry["checkpoint"]]
+ assert len(selected) == 1
+ assert selected[0]["repo_id"] == "unsloth/Z-Image-GGUF"
+ assert selected[0]["files"] == ["Z-Image-Turbo-Q4_K_M.gguf"]
def test_download_plan_counts_a_cached_lower_auto_prequant(monkeypatch):
diff --git a/studio/backend/tests/test_diffusion_pipeline_prequant.py b/studio/backend/tests/test_diffusion_pipeline_prequant.py
index 5c98bc6d75d..bce11404af5 100644
--- a/studio/backend/tests/test_diffusion_pipeline_prequant.py
+++ b/studio/backend/tests/test_diffusion_pipeline_prequant.py
@@ -22,6 +22,7 @@ from core.inference.diffusion_denoiser_prequant import PIPELINE_SEED_DECLINED
from core.inference.diffusion_device import DiffusionDeviceTarget
from core.inference.diffusion_families import detect_family_for_pick
from core.inference.diffusion_memory import DeviceMemory
+from models.inference import DiffusionDownloadPlanResponse
MIB = 1024 * 1024
@@ -211,7 +212,7 @@ def test_the_plan_counts_and_stages_the_hosted_checkpoint(monkeypatch, hub):
assert "transformer/config.json" in entries[Z_IMAGE_REPO]["files"]
assert plan["required_bytes"] == ALL_BYTES - DENOISER_BYTES + PREQUANT_BYTES
assert entries[Z_IMAGE_REPO]["checkpoint"] is True
- assert entries[PREQUANT_REPO]["checkpoint"] is False
+ assert entries[PREQUANT_REPO]["checkpoint"] is True
def test_an_unplanned_pipeline_pick_plans_exactly_as_before(monkeypatch, hub):
@@ -1958,3 +1959,28 @@ def test_a_large_variant_of_a_small_dit_family_quantises_under_auto(fake_runtime
_load(backend, _pipeline_prequant_planned = None, _pipeline_prequant_skipped = ())
assert Z_IMAGE_REPO in seen # the load hands the decider its resolved base
assert spy.quantised == ["auto"]
+
+
[email protected]("shared_repo", [False, True])
+def test_denoiser_is_a_checkpoint_and_encoder_only_is_an_asset(monkeypatch, hub, shared_repo):
+ backend = _plan_backend(monkeypatch, planned = "fp8")
+ encoder_repo = PREQUANT_REPO if shared_repo else "test/shared-encoder"
+ monkeypatch.setattr(
+ DiffusionBackend,
+ "_te_prequant_plan_files",
+ lambda *_a, **_k: {
+ "text_encoder": (encoder_repo, [("text_encoder/model.safetensors", 1234)])
+ },
+ )
+ plan = backend.download_plan(Z_IMAGE_REPO, model_kind = "pipeline")
+ entries = {entry["repo_id"]: entry for entry in plan["entries"]}
+ assert entries[PREQUANT_REPO]["checkpoint"] is True
+ assert entries[PREQUANT_REPO]["hosted_denoiser"] is True
+ assert entries[Z_IMAGE_REPO]["hosted_denoiser"] is False
+ # The route returns this model, which would drop an undeclared field.
+ served = DiffusionDownloadPlanResponse(**plan).entries
+ assert {e.repo_id for e in served if e.hosted_denoiser} == {PREQUANT_REPO}
+ assert entries[encoder_repo]["checkpoint"] is shared_repo
+ assert entries[encoder_repo]["hosted_denoiser"] is shared_repo
+ assert "text_encoder/model.safetensors" in entries[encoder_repo]["files"]
+ assert entries[PREQUANT_REPO]["bytes"] == PREQUANT_BYTES + (1234 if shared_repo else 0)
diff --git a/studio/frontend/src/features/images/api.ts b/studio/frontend/src/features/images/api.ts
index ced87ea63ed..05bce080484 100644
--- a/studio/frontend/src/features/images/api.ts
+++ b/studio/frontend/src/features/images/api.ts
@@ -448,6 +448,7 @@ export interface DiffusionDownloadPlan {
gguf_filename: string | null;
/** Only the planner knows: a gated pick is staged from an ungated mirror under another repo id. */
checkpoint?: boolean;
+ hosted_denoiser?: boolean;
}[];
total_bytes: number;
required_bytes?: number;
diff --git a/studio/frontend/src/lib/diffusion-pipeline-load-target.ts b/studio/frontend/src/lib/diffusion-pipeline-load-target.ts
index 3482e2b9438..d1bc9a71939 100644
--- a/studio/frontend/src/lib/diffusion-pipeline-load-target.ts
+++ b/studio/frontend/src/lib/diffusion-pipeline-load-target.ts
@@ -14,7 +14,7 @@ export function diffusionPipelineLoadTarget(
: { repoId: model, source: meta.source, onDevice: meta.source === "local" };
}
-/** A pinned snapshot is already on disk (and its revision may move), so only companions download. */
+/** A pinned snapshot is already on disk (and its revision may move), so only companions and a hosted denoiser, which it lacks, download. */
export function diffusionStagingEntries(
entries: readonly {
repo_id: string;
@@ -22,6 +22,7 @@ export function diffusionStagingEntries(
bytes: number;
gguf_filename: string | null;
checkpoint?: boolean;
+ hosted_denoiser?: boolean;
}[],
repoId: string,
opts: { filename?: string; displayRepoId?: string },
@@ -36,6 +37,7 @@ export function diffusionStagingEntries(
// `??`: false is an answer; the fallback is only for older backends.
checkpoint:
e.checkpoint ?? (opts.filename ? e.files.includes(opts.filename) : e.repo_id === planRepoId),
+ hostedDenoiser: e.hosted_denoiser,
}))
- .filter((e) => planRepoId === repoId || !e.checkpoint);
+ .filter((e) => planRepoId === repoId || !e.checkpoint || e.hostedDenoiser);
}
diff --git a/studio/frontend/tests/family-override.test.ts b/studio/frontend/tests/family-override.test.ts
index 888b49d778f..8e3af4f3b7e 100644
--- a/studio/frontend/tests/family-override.test.ts
+++ b/studio/frontend/tests/family-override.test.ts
@@ -102,6 +102,15 @@ test("a pinned cache snapshot loads by path, plans by its Hub id, and stages onl
diffusionStagingEntries(plan, "MiniMaxAI/MiniMax-H3", {}).map((e) => [e.repoId, e.checkpoint]),
[["MiniMaxAI/MiniMax-H3", true], ["external/quant", false]],
);
+
+ const withDenoiser = [
+ { ...entry("MiniMaxAI/MiniMax-H3", true), hosted_denoiser: false },
+ { ...entry("unsloth/fp8", true), hosted_denoiser: true },
+ ];
+ assert.deepEqual(
+ diffusionStagingEntries(withDenoiser, pinned.repoId, { displayRepoId: pinned.displayRepoId }).map((e) => [e.repoId, e.checkpoint]),
+ [["unsloth/fp8", true]],
+ );
});
for (const page of ["features/images/images-page.tsx", "features/video/video-page.tsx"]) {