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 one GGUF row per build in repos whose model name contains NVFP4 (#14733)

Daniel Han9d45fad70be3
diff --git a/studio/backend/hub/services/models/gguf_variants.py b/studio/backend/hub/services/models/gguf_variants.py
index 784c8fcea2a..2c073a5abc9 100644
--- a/studio/backend/hub/services/models/gguf_variants.py
+++ b/studio/backend/hub/services/models/gguf_variants.py
@@ -871,7 +871,9 @@ _CACHE_SPLIT_RE = re.compile(r"-(\d{3,})-of-(\d{3,})(?=\.gguf$)", re.IGNORECASE)
# Only enumerates candidates for the resolver to adjudicate; drift costs a candidate, not a verdict.
_KNOWN_QUANT_RE = re.compile(
r"(UD-)?"
- r"((?:MX|NV)FP[0-9]+(?:_[A-Z0-9]+)*"
+ r"(MXFP[0-9]+(?:_[A-Z0-9]+)*"
+ r"|NVFP[0-9]+(?:_[A-Z0-9]+)*"
+ r"(?=(?:-[0-9]+(?:[.][0-9]+)?bpw)?(?:-[0-9]{3,}-of-[0-9]{3,})?(?:[.]gguf)?$)"
r"|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?"
r"|P?TQ[0-9]+_[0-9]+"
r"|Q[0-9]+_K_[A-Z]+"
diff --git a/studio/backend/tests/test_gguf_delete_keeps_sibling_builds.py b/studio/backend/tests/test_gguf_delete_keeps_sibling_builds.py
index 7948fe409de..d1e08460d31 100644
--- a/studio/backend/tests/test_gguf_delete_keeps_sibling_builds.py
+++ b/studio/backend/tests/test_gguf_delete_keeps_sibling_builds.py
@@ -70,9 +70,10 @@ def test_a_custom_folder_lists_every_build_at_one_quant(tmp_path):
from utils.models.model_config import _find_local_gguf_by_variant
# custom folders have no remote listing to constrain their keys, so every build keeps a row (#13926).
- folder = tmp_path / "esatapedico" / "Qwen3.8-27B-NVFP4-MTP-GGUF"
+ # NVFP4 ends each name, so it is the quant all three share (mid-name it is part of the model name).
+ folder = tmp_path / "esatapedico" / "Qwen3.8-27B-MTP-NVFP4-GGUF"
folder.mkdir(parents = True)
- names = [f"Qwen3.8-27B-NVFP4-MTP-{build}.gguf" for build in ("HIGHEST", "LOW", "MID-HIGH_enru")]
+ names = [f"Qwen3.8-27B-MTP-{build}-NVFP4.gguf" for build in ("HIGHEST", "LOW", "MID-HIGH_enru")]
for name in names:
(folder / name).write_bytes(b"x" * 64)
variants, _ = list_local_gguf_variants(str(folder))
diff --git a/studio/backend/tests/test_gguf_nvfp4_model_name_variants.py b/studio/backend/tests/test_gguf_nvfp4_model_name_variants.py
new file mode 100644
index 00000000000..cbaa201e231
--- /dev/null
+++ b/studio/backend/tests/test_gguf_nvfp4_model_name_variants.py
@@ -0,0 +1,190 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""``NVFP4`` inside a community model name is not the quant of every file.
+
+Teaching the quant parser ``NVFP4`` (for the Nemotron 3.5 Super ``NVFP4/`` folder) made it the
+first quant token of every file in repos whose model name contains ``-NVFP4-``, so all builds got
+the one variant key ``NVFP4`` and the picker kept a single row. NVFP now names the quant only last
+in the name, where llama.cpp naming puts the type; earlier it is part of the model name, and those
+files keep the keys they had before NVFP was a token. File names below are the real Hub listings.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+_backend_root = Path(__file__).resolve().parent.parent
+_tests_dir = Path(__file__).resolve().parent
+for _p in (_backend_root, _tests_dir):
+ if str(_p) not in sys.path:
+ sys.path.insert(0, str(_p))
+
+from hub.utils.gguf import ( # noqa: E402
+ extract_quant_token,
+ gguf_variant_key,
+ group_gguf_variant_files,
+ is_mtp_drafter_path,
+)
+
+
+# repo -> {file: key the picker had before NVFP was a quant token}
+COMMUNITY_REPOS = {
+ "esatapedico/Qwen3.8-27B-NVFP4-MTP-GGUF": {
+ f"Qwen3.8-27B-NVFP4-MTP-{build}.gguf": f"Qwen3.8-27B-NVFP4-MTP-{build}"
+ for build in (
+ "COMPACT-LOW",
+ "HIGH",
+ "HIGHEST",
+ "LOW",
+ "MEDIUM",
+ "MID-HIGH",
+ "ORIG",
+ "VERY-HIGH",
+ "VERY-LOW",
+ )
+ },
+ "michaelw9999/Qwen3.6-35B-A3B-NVFP4-MTP-GGUF": {
+ "Qwen3.6-35B-A3B-NVFP4-MTP-HQ.gguf": "Qwen3.6-35B-A3B-NVFP4-MTP-HQ",
+ "Qwen3.6-35B-A3B-NVFP4-MTP-TURBO.gguf": "Qwen3.6-35B-A3B-NVFP4-MTP-TURBO",
+ },
+ "llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-NVFP4-Experts-Only-GGUF": {
+ "Qwen3.6-35B-A3B-uncensored-heretic-NVFP4-Experts-Only-Q8_0.gguf": "Q8_0",
+ "Qwen3.6-35B-A3B-uncensored-heretic-NVFP4-Experts-Only.gguf": (
+ "Qwen3.6-35B-A3B-uncensored-heretic-NVFP4-Experts-Only"
+ ),
+ },
+ "utautako/Qwen3.8-27B-NVFP4-MTP-Q8attn-GGUF": {
+ "Qwen3.8-27B-NVFP4-MTP-Q6attn.gguf": "Qwen3.8-27B-NVFP4-MTP-Q6attn",
+ "Qwen3.8-27B-NVFP4-MTP-Q8attn.gguf": "Qwen3.8-27B-NVFP4-MTP-Q8attn",
+ },
+ "CompiledThoughts/Qwen3.8-Flash-Next-NVFP4-Q8_0": {
+ "Qwen3.8-Flash-Next-NVFP4-Q8_0.gguf": "Q8_0",
+ },
+}
+
+SUPER = "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B"
+
+# Names whose quant really is the NVFP / MXFP token: unchanged by this fix.
+QUANT_SUFFIX_KEYS = {
+ f"NVFP4/{SUPER}-NVFP4-00001-of-00003.gguf": "NVFP4",
+ f"NVFP4/{SUPER}-NVFP4-00003-of-00003.gguf": "NVFP4",
+ "Qwen3.6-27B-NVFP4.gguf": "NVFP4",
+ "Qwen3.6-27B-nvfp4.gguf": "nvfp4",
+ "Model-UD-NVFP4.gguf": "UD-NVFP4",
+ "Model-NVFP4_MOE.gguf": "NVFP4_MOE",
+ "Model-NVFP4-4.25bpw.gguf": "NVFP4-4.25bpw",
+ "NVFP4": "NVFP4",
+ "gpt-oss-20b-MXFP4.gguf": "MXFP4",
+ "Qwen3.6-35B-A3B-MXFP4_MOE.gguf": "MXFP4_MOE",
+ "Qwen3.6-35B-A3B-UD-MXFP4_MOE.gguf": "UD-MXFP4_MOE",
+ "MXFP4_MOE/Qwen3.6-35B-A3B-MXFP4_MOE-00002-of-00004.gguf": "MXFP4_MOE",
+ # MXFP was a token before NVFP and still matches anywhere, so stored MXFP keys stay put.
+ "DeepSeek-V4-Flash-Dspark-Abliterated-MXFP4-BF16.gguf": "MXFP4",
+ "Some-Model-MXFP4-MTP-HIGH.gguf": "MXFP4",
+}
+
+
[email protected]("repo", sorted(COMMUNITY_REPOS))
+def test_every_build_in_an_nvfp4_named_repo_keeps_its_own_row(repo):
+ expected = COMMUNITY_REPOS[repo]
+ rows = group_gguf_variant_files([(name, 1) for name in expected])
+ assert len(rows) == len(expected)
+ assert {key: filename for key, (filename, _size) in rows.items()} == {
+ key: name for name, key in expected.items()
+ }
+
+
+ "name,key",
+ [(name, key) for files in COMMUNITY_REPOS.values() for name, key in files.items()],
+)
+def test_nvfp4_inside_the_model_name_keeps_the_pre_nvfp_key(name, key):
+ assert gguf_variant_key(name) == key
+
+
[email protected]("name,key", sorted(QUANT_SUFFIX_KEYS.items()))
+def test_trailing_nvfp_and_mxfp_tokens_are_still_the_quant(name, key):
+ assert gguf_variant_key(name) == key
+ assert extract_quant_token(name) == key.removesuffix("-4.25bpw")
+
+
+def test_all_quant_label_mirrors_agree():
+ """The loader label, the loader key and the variant-service scan must read the same token."""
+ from core.inference import llama_cpp
+ from hub.services.models import gguf_variants
+ from utils.models import model_config
+
+ name = "Qwen3.6-35B-A3B-uncensored-heretic-NVFP4-Experts-Only-Q8_0.gguf"
+ assert model_config._extract_quant_label(name) == "Q8_0"
+ assert model_config._gguf_variant_key(name) == "Q8_0"
+ assert [m.group(2) for m in gguf_variants._KNOWN_QUANT_RE.finditer(name[:-5])] == ["Q8_0"]
+ assert [m.group(2) for m in llama_cpp._GGUF_KNOWN_QUANT_RE.finditer(name[:-5])] == ["Q8_0"]
+
+ bare = "Qwen3.8-27B-NVFP4-MTP-HIGH"
+ assert model_config._gguf_variant_key(f"{bare}.gguf") == bare
+ assert gguf_variants._KNOWN_QUANT_RE.search(bare) is None
+ assert llama_cpp._GGUF_KNOWN_QUANT_RE.search(bare) is None
+
+ super_file = f"NVFP4/{SUPER}-NVFP4-00001-of-00003.gguf"
+ assert model_config._extract_quant_label(super_file) == "NVFP4"
+ assert model_config._gguf_variant_key(super_file) == "NVFP4"
+
+
[email protected]("repo", sorted(COMMUNITY_REPOS))
+def test_each_key_loads_exactly_its_own_file(monkeypatch, repo):
+ from core.inference import llama_cpp
+
+ files = [".gitattributes", "README.md", *COMMUNITY_REPOS[repo]]
+ monkeypatch.setattr(
+ llama_cpp, "call_hub_with_anonymous_retry", lambda fn, token, repo_id: list(files)
+ )
+ for name, key in COMMUNITY_REPOS[repo].items():
+ assert llama_cpp._resolve_variant_gguf_files(repo, key) == (name, [])
+
+
[email protected]("repo", sorted(COMMUNITY_REPOS))
+def test_each_download_plan_holds_exactly_its_own_file(repo):
+ from hub.utils.gguf_plan import build_gguf_variant_plans
+
+ siblings = [
+ SimpleNamespace(rfilename = name, size = 1 << 20, lfs = None) for name in COMMUNITY_REPOS[repo]
+ ]
+ plans = build_gguf_variant_plans(siblings)
+ assert {key: set(plan.main_filenames) for key, plan in plans.items()} == {
+ key.lower(): {name} for name, key in COMMUNITY_REPOS[repo].items()
+ }
+
+
+def test_super_nvfp4_shards_group_into_one_row():
+ shards = [f"NVFP4/{SUPER}-NVFP4-0000{i}-of-00003.gguf" for i in (1, 2, 3)]
+ rows = group_gguf_variant_files([(name, 10) for name in shards])
+ assert rows == {"NVFP4": (shards[0], 30)}
+
+
+def test_nvfp4_suffixed_drafters_stay_drafters():
+ assert is_mtp_drafter_path(f"LiLiCorr/lilicorr-{SUPER}-NVFP4.gguf")
+ assert is_mtp_drafter_path("Bonsai-8B-dspark-NVFP4.gguf")
+ assert not is_mtp_drafter_path("Qwen3.8-27B-dspark-NVFP4-MTP-HIGH.gguf")
+
+
+def test_a_custom_folder_of_nvfp4_named_builds_lists_each_by_its_own_name(tmp_path):
+ from hub.utils.gguf import list_local_gguf_variants
+ from routes.models import _resolve_quant_gguf
+ from utils.models.model_config import _find_local_gguf_by_variant
+
+ folder = tmp_path / "esatapedico" / "Qwen3.8-27B-NVFP4-MTP-GGUF"
+ folder.mkdir(parents = True)
+ names = [f"Qwen3.8-27B-NVFP4-MTP-{build}.gguf" for build in ("HIGHEST", "LOW", "MID-HIGH")]
+ for name in names:
+ (folder / name).write_bytes(b"x" * 64)
+ variants, _ = list_local_gguf_variants(str(folder))
+ listed = {v.quant: v.filename for v in variants}
+ assert listed == {name[: -len(".gguf")]: name for name in names}
+ for quant, filename in listed.items():
+ assert Path(_find_local_gguf_by_variant(str(folder), quant)).name == filename
+ assert Path(_resolve_quant_gguf(str(folder), quant, True)[0]).name == filename
diff --git a/studio/backend/utils/gguf_names.py b/studio/backend/utils/gguf_names.py
index bcf9038cd7d..5c3ead81fc6 100644
--- a/studio/backend/utils/gguf_names.py
+++ b/studio/backend/utils/gguf_names.py
@@ -8,9 +8,13 @@ from __future__ import annotations
import re
from typing import Optional
+# NVFP is the quant only last in the name (Model-NVFP4.gguf, NVFP4/); mid-name it is part of a
+# community model name (Qwen3.8-27B-NVFP4-MTP-HIGH). MXFP predates NVFP, so it matches anywhere.
GGUF_QUANT_RE = re.compile(
r"(UD-)?"
- r"((?:MX|NV)FP[0-9]+(?:_[A-Z0-9]+)*"
+ r"(MXFP[0-9]+(?:_[A-Z0-9]+)*"
+ r"|NVFP[0-9]+(?:_[A-Z0-9]+)*"
+ r"(?=(?:-[0-9]+(?:[.][0-9]+)?bpw)?(?:-[0-9]{3,}-of-[0-9]{3,})?(?:[.]gguf)?$)"
r"|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?"
r"|P?TQ[0-9]+_[0-9]+"
r"|Q[0-9]+_K_[A-Z]+"
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 6e8a6ccce87..34bb764a00d 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -2854,7 +2854,9 @@ def _extract_quant_label(filename: str) -> str:
stem = re.sub(r"-\d{3,}-of-\d{3,}", "", basename.rsplit(".", 1)[0])
quant_re = (
r"(UD-)?" # Optional UD- prefix (Ultra Discrete)
- r"((?:MX|NV)FP[0-9]+(?:_[A-Z0-9]+)*" # MXFP variants: MXFP4, MXFP4_MOE
+ r"(MXFP[0-9]+(?:_[A-Z0-9]+)*" # MXFP variants: MXFP4, MXFP4_MOE
+ r"|NVFP[0-9]+(?:_[A-Z0-9]+)*" # NVFP4, last in the name only
+ r"(?=(?:-[0-9]+(?:[.][0-9]+)?bpw)?(?:-[0-9]{3,}-of-[0-9]{3,})?(?:[.]gguf)?$)"
r"|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?" # IQ variants: IQ4_XS, IQ4_NL, IQ1_S
r"|P?TQ[0-9]+_[0-9]+" # Ternary quant: TQ1_0, TQ2_0
r"|Q[0-9]+_K_[A-Z]+" # K-quant: Q4_K_M, Q3_K_S
diff --git a/studio/frontend/src/features/model-picker/model-config/model-identity.ts b/studio/frontend/src/features/model-picker/model-config/model-identity.ts
index 02e250d4f17..7c561def0c5 100644
--- a/studio/frontend/src/features/model-picker/model-config/model-identity.ts
+++ b/studio/frontend/src/features/model-picker/model-config/model-identity.ts
@@ -111,9 +111,9 @@ export function ggufVariantFromStorageKey(key: string): string | null {
// Mirrors split_quant_suffix in openai_auto_switch_settings.py. The bpw modifier
// ("IQ4_XS-3.53bpw") is optional: the backend label helpers disagree.
const BPW_SUFFIX = /-[0-9]+(?:\.[0-9]+)?bpw$/i;
-// One source for the anchored test and the scan below. Mirrors _GGUF_QUANT_RE in gguf.py.
+// One source for the anchored test and the scan below. Mirrors GGUF_QUANT_RE in utils/gguf_names.py.
const QUANT_TOKEN_SOURCE =
- "(UD-)?((?:MX|NV)FP[0-9]+(?:_[A-Z0-9]+)*|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?|P?TQ[0-9]+_[0-9]+|Q[0-9]+_K_[A-Z]+|P?Q[0-9]+_(?:[0-9]+(?:_G[0-9]+)?|G[0-9]+)|Q[0-9]+_K|BF16|F16|F32)";
+ "(UD-)?(MXFP[0-9]+(?:_[A-Z0-9]+)*|NVFP[0-9]+(?:_[A-Z0-9]+)*(?=(?:-[0-9]+(?:[.][0-9]+)?bpw)?(?:-[0-9]{3,}-of-[0-9]{3,})?(?:[.]gguf)?$)|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?|P?TQ[0-9]+_[0-9]+|Q[0-9]+_K_[A-Z]+|P?Q[0-9]+_(?:[0-9]+(?:_G[0-9]+)?|G[0-9]+)|Q[0-9]+_K|BF16|F16|F32)";
const KNOWN_QUANT = new RegExp(`^${QUANT_TOKEN_SOURCE}$`, "i");
const QUANT_TOKEN = new RegExp(QUANT_TOKEN_SOURCE, "gi");
const MAX_QUANT_SUFFIX_LEN = 64;
diff --git a/studio/frontend/tests/gguf-nvfp4-quant-label.test.ts b/studio/frontend/tests/gguf-nvfp4-quant-label.test.ts
new file mode 100644
index 00000000000..401339dc955
--- /dev/null
+++ b/studio/frontend/tests/gguf-nvfp4-quant-label.test.ts
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { registerBundlerResolver } from "./helpers/kit.ts";
+
+registerBundlerResolver();
+const { ggufQuantLabel, splitQuantSuffix } = await import(
+ "../src/features/model-picker/model-config/model-identity.ts"
+);
+
+test("NVFP4 inside a community model name is not the file's quant", () => {
+ // Mirrors utils.gguf_names: NVFP names the quant only last in the name.
+ assert.equal(
+ ggufQuantLabel("Qwen3.8-27B-NVFP4-MTP-HIGH.gguf"),
+ "Qwen3.8-27B-NVFP4-MTP-HIGH",
+ );
+ assert.equal(
+ ggufQuantLabel(
+ "Qwen3.6-35B-A3B-uncensored-heretic-NVFP4-Experts-Only-Q8_0.gguf",
+ ),
+ "Q8_0",
+ );
+ assert.equal(ggufQuantLabel("Qwen3.8-Flash-Next-NVFP4-Q8_0.gguf"), "Q8_0");
+});
+
+test("a trailing NVFP or MXFP token is still the quant", () => {
+ assert.equal(
+ ggufQuantLabel("NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-00001-of-00003.gguf"),
+ "NVFP4",
+ );
+ assert.equal(ggufQuantLabel("Model-NVFP4.gguf"), "NVFP4");
+ assert.equal(ggufQuantLabel("Qwen3.6-35B-A3B-UD-MXFP4_MOE.gguf"), "UD-MXFP4_MOE");
+ assert.equal(ggufQuantLabel("Some-Model-MXFP4-MTP-HIGH.gguf"), "MXFP4");
+ assert.deepEqual(splitQuantSuffix("unsloth/Model-GGUF:NVFP4"), [
+ "unsloth/Model-GGUF",
+ "NVFP4",
+ ]);
+});