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: keep a cached quant On Device when only its companion is missing (#11236)

oobaboogab1b664e52ea4
diff --git a/studio/backend/hub/services/models/gguf_variants.py b/studio/backend/hub/services/models/gguf_variants.py
index 6524c1b7e56..7fc3f754f46 100644
--- a/studio/backend/hub/services/models/gguf_variants.py
+++ b/studio/backend/hub/services/models/gguf_variants.py
@@ -1728,6 +1728,29 @@ async def get_gguf_variants_answer(
),
)
+ # A complete main with an unstarted companion has no partial signal. Keep it visible
+ # unless the existing readiness check says it can load.
+ for variant in partial_scan_variants:
+ if variant.quant in partial_quants:
+ continue
+ requirement = requirements_by_quant.get(variant.quant.lower())
+ if requirement is None or _is_fully_downloaded(variant):
+ continue
+ if not (
+ _filenames_cached(requirement.main_filenames, requirement.main_size_bytes)
+ or _quant_and_companions_ready(variant.quant.lower(), variant.size_bytes, ())
+ ):
+ continue
+ partial_quants.add(variant.quant)
+ partial_quant_transports.setdefault(
+ variant.quant,
+ _partial_transport_for_variant(
+ repo_id,
+ variant.quant,
+ repo_cache_dir,
+ ),
+ )
+
local_blobs_by_quant = (
_local_main_gguf_blobs_by_quant(repo_id, repo_cache_dir)
if cache_reads_authorized
diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.py
index 19fe3ed327a..6be4399eff5 100644
--- a/studio/backend/hub/tests/test_model_services.py
+++ b/studio/backend/hub/tests/test_model_services.py
@@ -4466,7 +4466,8 @@ def test_cached_flash_next_quant_needs_managed_mtp_before_it_is_downloaded(
assert before.dependencies_resolved is True
assert before.variants[0].downloaded is False
assert before.variants[0].download_size_bytes == (130 if "projector" in cache_case else 120)
- assert before.variants[0].partial is (cache_case == "partial-mtp")
+ # Every case has complete weights but lacks its planned drafter.
+ assert before.variants[0].partial is True
if cache_case in {"alternate-projector", "stale-main"}:
assert before.variants[0].pending_drafter_filename is None
assert before.variants[0].pending_drafter_size_bytes == 0
@@ -4488,6 +4489,86 @@ def test_cached_flash_next_quant_needs_managed_mtp_before_it_is_downloaded(
assert after.variants[0].pending_drafter_size_bytes == 0
+def test_a_cached_quant_missing_only_its_projector_stays_listed(monkeypatch, tmp_path):
+ """Keep cached weights visible when their required projector is missing."""
+ with gguf_variants._VARIANT_HASH_LOCK:
+ gguf_variants._VARIANT_REQUIREMENT_CACHE.clear()
+
+ async def _run_inline(fn, *args, **kwargs):
+ return fn(*args, **kwargs)
+
+ repo_id = "Org/Vision-GGUF"
+ repo_dir = tmp_path / "cache" / "models--Org--Vision-GGUF"
+ snapshot = repo_dir / "snapshots" / "rev0"
+ snapshot.mkdir(parents = True)
+ (snapshot / "Model-UD-Q4_K_XL.gguf").write_bytes(b"m" * 100)
+ # Mirror the blob store so pricing credits the cached weights.
+ (repo_dir / "blobs").mkdir()
+ (repo_dir / "blobs" / "q4").write_bytes(b"m" * 100)
+ siblings = [
+ _sibling("Model-UD-Q4_K_XL.gguf", 100, "q4"),
+ _sibling("Model-Q8_0.gguf", 200, "q8"),
+ _sibling("mmproj-F16.gguf", 10, "projector"),
+ ]
+
+ monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state")
+ monkeypatch.setattr(gguf_variants.asyncio, "to_thread", _run_inline)
+ monkeypatch.setattr(gguf_variants, "_local_main_gguf_blobs_by_quant", lambda *_args: {})
+ monkeypatch.setattr(
+ gguf_variants,
+ "list_gguf_variants",
+ lambda *_args, **_kwargs: (
+ [
+ SimpleNamespace(
+ filename = "Model-UD-Q4_K_XL.gguf",
+ quant = "UD-Q4_K_XL",
+ display_label = None,
+ size_bytes = 100,
+ ),
+ SimpleNamespace(
+ filename = "Model-Q8_0.gguf",
+ quant = "Q8_0",
+ display_label = None,
+ size_bytes = 200,
+ ),
+ ],
+ True,
+ siblings,
+ ),
+ )
+ monkeypatch.setattr(
+ gguf_variants.download_registry,
+ "incomplete_blob_hashes",
+ lambda *_args, **_kwargs: set(),
+ )
+
+ # The picker pins cached Hub rows to a snapshot.
+ listed = asyncio.run(
+ gguf_variants.get_gguf_variants_response(repo_id, local_path = str(snapshot))
+ )
+ before = {v.quant: v for v in listed.variants}
+
+ assert before["UD-Q4_K_XL"].downloaded is False
+ assert before["UD-Q4_K_XL"].partial is True
+ assert before["UD-Q4_K_XL"].download_size_bytes == 110
+ assert before["UD-Q4_K_XL"].download_remaining_bytes == 10
+ assert before["Q8_0"].downloaded is False
+ assert before["Q8_0"].partial is False
+ # Match the picker's On Device filter.
+ assert [q for q, v in before.items() if v.downloaded or v.partial] == ["UD-Q4_K_XL"]
+
+ (snapshot / "mmproj-F16.gguf").write_bytes(b"p" * 10)
+ (repo_dir / "blobs" / "projector").write_bytes(b"p" * 10)
+ relisted = asyncio.run(
+ gguf_variants.get_gguf_variants_response(repo_id, local_path = str(snapshot))
+ )
+ after = {v.quant: v for v in relisted.variants}
+
+ assert after["UD-Q4_K_XL"].downloaded is True
+ assert after["UD-Q4_K_XL"].partial is False
+ assert after["Q8_0"].partial is False
+
+
def test_download_registry_repo_keys_are_case_insensitive():
registry = download_registry.DownloadRegistry()