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: pick the right model size when the file names are lowercase (#11722)

Nilay856d80e5274e
diff --git a/studio/backend/core/inference/openai_auto_download.py b/studio/backend/core/inference/openai_auto_download.py
index 33e4b61f1c6..f91d81661ff 100644
--- a/studio/backend/core/inference/openai_auto_download.py
+++ b/studio/backend/core/inference/openai_auto_download.py
@@ -876,10 +876,10 @@ def preferred_quant(labels) -> Optional[str]:
"""
from utils.models.model_config import _pick_best_gguf
- # _pick_best_gguf ranks filenames and matches upper-case tokens, so feed "<LABEL>.gguf"
+ # _pick_best_gguf ranks filenames, so feed "<LABEL>.gguf"
synthetic: dict[str, str] = {}
for name in labels:
- synthetic.setdefault(f"{name.upper()}.gguf", name)
+ synthetic.setdefault(f"{name}.gguf", name)
best = _pick_best_gguf(list(synthetic))
return synthetic.get(best) if best else None
diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.py
index fecf589e976..f946fa09188 100644
--- a/studio/backend/hub/tests/test_model_services.py
+++ b/studio/backend/hub/tests/test_model_services.py
@@ -508,6 +508,12 @@ def test_big_endian_detection_ignores_model_name_be_token():
)
+def test_pick_best_gguf_prefers_an_unlisted_quant_only_over_full_precision():
+ assert gguf.pick_best_gguf(["model-bf16.gguf", "model-Q3_K.gguf"]) == "model-Q3_K.gguf"
+ assert gguf.pick_best_gguf(["model-F32.gguf", "model-bf16.gguf"]) == "model-bf16.gguf"
+ assert gguf.pick_best_gguf(["model-APEX.gguf", "model-Q4_0.gguf"]) == "model-APEX.gguf"
+
+
def test_custom_inventory_filters_mtp_companions_at_registered_root(tmp_path, monkeypatch):
root = tmp_path / "MTP"
root.mkdir()
diff --git a/studio/backend/hub/utils/gguf.py b/studio/backend/hub/utils/gguf.py
index 767ebec886c..37bd07e69f6 100644
--- a/studio/backend/hub/utils/gguf.py
+++ b/studio/backend/hub/utils/gguf.py
@@ -269,8 +269,20 @@ def pick_best_gguf(filenames: list[str]) -> Optional[str]:
by_quant.setdefault(extract_quant_label(name).upper(), name)
for quant in GGUF_QUANT_PREFERENCE:
filename = by_quant.get(quant.upper())
- if filename is not None:
- return filename
+ if filename is None:
+ continue
+ if quant in _FLOAT_PRECISION_QUANTS:
+ # The list leaves out quants such as Q4_0, Q3_K and TQ1_0; the first listed of them still beats full precision.
+ filename = next(
+ (
+ name
+ for name in by_quant.values()
+ if (token := extract_quant_token(name))
+ and token.upper() not in _FLOAT_PRECISION_QUANTS
+ ),
+ filename,
+ )
+ return filename
return gguf_files[0]
diff --git a/studio/backend/tests/test_gguf_variant_preflight.py b/studio/backend/tests/test_gguf_variant_preflight.py
index 539be0f498d..7a5e4e2f74e 100644
--- a/studio/backend/tests/test_gguf_variant_preflight.py
+++ b/studio/backend/tests/test_gguf_variant_preflight.py
@@ -249,3 +249,37 @@ def test_nothing_is_carried_when_no_cached_copy_verifies(remote_gguf_repo):
"""A first-time download has no cached path to carry."""
assert ModelConfig.from_identifier(REPO, gguf_variant = "Q8_0").gguf_verified is None
assert ModelConfig.from_identifier(REPO).gguf_verified is None
+
+
+QWEN_LOWERCASE_VARIANTS = [
+ GgufVariantInfo(
+ filename = f"qwen2.5-7b-instruct-{quant}-00001-of-0000{shards}.gguf",
+ quant = quant,
+ size_bytes = size,
+ )
+ for quant, shards, size in [
+ ("fp16", 4, 15_237_853_600),
+ ("q8_0", 3, 8_098_525_600),
+ ("q6_k", 2, 6_254_197_600),
+ ("q5_k_m", 2, 5_444_831_600),
+ ("q4_k_m", 2, 4_683_073_600),
+ ("q2_k", 1, 3_015_940_000),
+ ]
+]
+
+
+def test_no_variant_autoselects_a_lowercase_quant_over_the_first_listed(
+ remote_gguf_repo, monkeypatch
+):
+ monkeypatch.setattr(
+ mc,
+ "list_gguf_variants",
+ lambda identifier, hf_token = None: (list(QWEN_LOWERCASE_VARIANTS), False),
+ )
+ filenames = [v.filename for v in QWEN_LOWERCASE_VARIANTS]
+ assert mc._pick_best_gguf(filenames) == "qwen2.5-7b-instruct-q4_k_m-00001-of-00002.gguf"
+ assert ModelConfig.from_identifier("Qwen/Qwen2.5-7B-Instruct-GGUF").gguf_variant == "q4_k_m"
+
+
+def test_no_variant_autoselect_matches_whole_quant_labels():
+ assert mc._pick_best_gguf(["m-Q6_K_P.gguf", "m-IQ4_XS.gguf"]) == "m-IQ4_XS.gguf"
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index c189176f737..c993e64531f 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -2660,66 +2660,10 @@ def detect_gguf_model(path: str, model_root: Optional[str] = None) -> Optional[s
return None
-# Preferred GGUF quant levels, descending. UD (Unsloth Dynamic) variants beat standard
-# quants on quality per bit; ordered by size/quality tradeoff, not raw quality.
-_GGUF_QUANT_PREFERENCE = [
- # UD variants (best quality per bit) -- Q4 is the sweet spot
- "UD-Q4_K_XL",
- "UD-Q4_K_L",
- "UD-Q5_K_XL",
- "UD-Q3_K_XL",
- "UD-Q6_K_XL",
- "UD-Q6_K_S",
- "UD-Q8_K_XL",
- "UD-Q2_K_XL",
- "UD-IQ4_NL",
- "UD-IQ4_XS",
- "UD-IQ3_S",
- "UD-IQ3_XXS",
- "UD-IQ2_M",
- "UD-IQ2_XXS",
- "UD-IQ1_M",
- "UD-IQ1_S",
- # Standard quants (fallback for non-Unsloth repos)
- "Q4_K_M",
- "Q4_K_S",
- "Q5_K_M",
- "Q5_K_S",
- "Q6_K",
- "Q8_0",
- "Q3_K_M",
- "Q3_K_L",
- "Q3_K_S",
- "Q2_K",
- "Q2_K_L",
- "IQ4_NL",
- "IQ4_XS",
- "IQ3_M",
- "IQ3_XXS",
- "IQ2_M",
- "IQ1_M",
- "F16",
- "BF16",
- "F32",
-]
-
-
def _pick_best_gguf(filenames: list[str]) -> Optional[str]:
- """Pick the best GGUF file: quant levels in _GGUF_QUANT_PREFERENCE order, else first .gguf."""
- from hub.utils.gguf import drop_shadowed_appledouble_names
-
- # See hub.utils.gguf.pick_best_gguf: the first matching name wins.
- filenames = drop_shadowed_appledouble_names(list(filenames))
- gguf_files = [f for f in filenames if f.lower().endswith(".gguf")]
- if not gguf_files:
- return None
-
- for quant in _GGUF_QUANT_PREFERENCE:
- for f in gguf_files:
- if quant in f:
- return f
-
- return gguf_files[0]
+ """Pick the best GGUF file by the model picker's ranking, so a bare repo id loads its default."""
+ from hub.utils.gguf import pick_best_gguf
+ return pick_best_gguf(filenames)
@dataclass
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index aadce25fe3a..d7e654e7826 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -2056,7 +2056,7 @@ def _resolve_model(
else:
payload["gguf_variant"] = load.gguf_variant
elif attach_public_id is not None and status_snapshot.get("is_gguf"):
- # Re-send the running quant: a repo id carries none, so from_identifier would auto-pick (_GGUF_QUANT_PREFERENCE, UD-Q4_K_XL first) and changing only the context would evict a chosen Q8_0 to download a different quant. Skip a .gguf path, which loads as itself; the server gates on the same suffix.
+ # Re-send the running quant: a repo id carries none, so from_identifier would auto-pick (GGUF_QUANT_PREFERENCE, UD-Q4_K_XL first) and changing only the context would evict a chosen Q8_0 to download a different quant. Skip a .gguf path, which loads as itself; the server gates on the same suffix.
resident_variant = status_snapshot.get("gguf_variant")
if resident_variant and not str(requested).lower().endswith(".gguf"):
payload["gguf_variant"] = resident_variant