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: a numbered shard without its index is not a complete snapshot (#16418)

Daniel Hane02bc9eaa784
diff --git a/studio/backend/core/inference/diffusion_small_host.py b/studio/backend/core/inference/diffusion_small_host.py
index 73b909da39e..d47005e70a0 100644
--- a/studio/backend/core/inference/diffusion_small_host.py
+++ b/studio/backend/core/inference/diffusion_small_host.py
@@ -7,6 +7,7 @@ from __future__ import annotations
import json
import os
+import re
import sys
import uuid
from dataclasses import dataclass, field
@@ -268,8 +269,7 @@ def snapshot_weights_complete(
skip: Any = (),
variant: Optional[str] = None,
) -> bool:
- """False when a ``model_index.json`` component lacks weights or index-listed shards; True without an index.
- ``skip`` names caller-supplied components; ``variant`` checks only that layout's files."""
+ """Check pipeline weights for the requested variant, excluding caller-supplied components."""
try:
root = Path(snapshot)
index = json.loads((root / "model_index.json").read_text(encoding = "utf-8"))
@@ -293,6 +293,7 @@ def snapshot_weights_complete(
]
if not files and (sub / "config.json").is_file():
return False # a model folder without weights; tokenizer / scheduler folders have no config.json
+ indexed_shards = set()
for shard_index in shard_indexes:
try:
shards = set(
@@ -304,6 +305,14 @@ def snapshot_weights_complete(
return False
if any(not (sub / shard).is_file() for shard in shards):
return False
+ indexed_shards.update(shards)
+ # All numbered shards still need the index the loader opens.
+ if any(
+ re.search(r"-\d{5}-of-\d{5}(?:\.[^.]+)?\.(?:safetensors|bin)$", f.name)
+ and f.name not in indexed_shards
+ for f in files
+ ):
+ return False
return True
diff --git a/studio/backend/tests/test_diffusion_small_host.py b/studio/backend/tests/test_diffusion_small_host.py
index b3841a27994..5d8aa506cf9 100644
--- a/studio/backend/tests/test_diffusion_small_host.py
+++ b/studio/backend/tests/test_diffusion_small_host.py
@@ -580,7 +580,9 @@ def _header_only(
(sub / filename).write_bytes(struct.pack("<Q", len(raw)) + raw)
[email protected]("cached", ["configs_only", "vae_only", "one_dit_shard"])
+ "cached", ["configs_only", "vae_only", "one_dit_shard", "one_dit_shard_no_index"]
+)
def test_a_partially_cached_snapshot_downloads_before_the_decision(tmp_path, monkeypatch, cached):
"""An interrupted download leaves some weights cached; sizing them alone hides the large denoiser."""
from core.inference import diffusion as dif
@@ -604,10 +606,16 @@ def test_a_partially_cached_snapshot_downloads_before_the_decision(tmp_path, mon
_header_only(full, "vae", 161)
if cached == "vae_only":
_header_only(partial, "vae", 161)
- if cached == "one_dit_shard":
+ if cached in ("one_dit_shard", "one_dit_shard_no_index"):
_header_only(
partial, "transformer", 4096, "diffusion_pytorch_model-00001-of-00005.safetensors"
)
+ if cached == "one_dit_shard_no_index":
+ _header_only(partial, "text_encoder", 1000, "model.safetensors")
+ _header_only(partial, "vae", 161)
+ for name in ("transformer", "text_encoder", "vae"):
+ (partial / name / "config.json").write_text("{}")
+ if cached == "one_dit_shard":
(partial / "transformer" / "diffusion_pytorch_model.safetensors.index.json").write_text(
json.dumps(
{
@@ -640,6 +648,46 @@ def test_a_partially_cached_snapshot_downloads_before_the_decision(tmp_path, mon
assert d.engaged and d.int8_spill
+ "stem,extension,variant",
+ [
+ ("diffusion_pytorch_model", "safetensors", None),
+ ("model", "safetensors", None),
+ ("pytorch_model", "bin", None),
+ ("diffusion_pytorch_model", "safetensors", "fp16"),
+ ],
+)
+def test_numbered_shards_require_an_index_covering_them(tmp_path, stem, extension, variant):
+ (tmp_path / "model_index.json").write_text(
+ json.dumps({"transformer": ["diffusers", "X"], "vae": ["diffusers", "Y"]})
+ )
+ mark = f".{variant}" if variant else ""
+ _header_only(tmp_path, "vae", 1, f"diffusion_pytorch_model{mark}.safetensors")
+ sub = tmp_path / "transformer"
+ sub.mkdir()
+ (sub / "config.json").write_text("{}")
+ names = [f"{stem}{mark}-{i:05d}-of-00002.{extension}" for i in (1, 2)]
+ (sub / names[0]).touch()
+ assert not sh.snapshot_weights_complete(tmp_path, variant = variant)
+ assert sh.snapshot_weights_complete(tmp_path, skip = ("transformer",), variant = variant)
+ (sub / names[1]).touch()
+ assert not sh.snapshot_weights_complete(tmp_path, variant = variant)
+ shard_index = sub / f"{stem}.{extension}.index{mark}.json"
+ orphan = f"{stem}{mark}.{extension}"
+ (sub / orphan).touch()
+ for content in (
+ "{",
+ "{}",
+ json.dumps({"weight_map": {"other": orphan}}),
+ ):
+ shard_index.write_text(content)
+ assert not sh.snapshot_weights_complete(tmp_path, variant = variant)
+ shard_index.write_text(json.dumps({"weight_map": {f"w{i}": n for i, n in enumerate(names)}}))
+ assert sh.snapshot_weights_complete(tmp_path, variant = variant)
+ (sub / names[1]).unlink()
+ assert not sh.snapshot_weights_complete(tmp_path, variant = variant)
+
+
def test_a_complete_snapshot_with_a_supplied_component_skips_the_download(tmp_path):
index = {
"transformer": ["diffusers", "X"],