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.

Compare Kaggle reference repo ids without case and prefetch the Qwen3 4bit repo under its new spelling (#12411)

Daniel Haneba095f0fd06
diff --git a/.github/scripts/kaggle_t4_ci/legs.py b/.github/scripts/kaggle_t4_ci/legs.py
index 1861f768f95..00056f5288b 100644
--- a/.github/scripts/kaggle_t4_ci/legs.py
+++ b/.github/scripts/kaggle_t4_ci/legs.py
@@ -466,7 +466,7 @@ PREFETCH_REPOS: tuple[str, ...] = (
"unsloth/Qwen3.5-2B",
# Default's model and the 4bit sibling `load_in_4bit=True` actually resolves to. Both, because FLOAT_TO_INT_MAPPER redirects at load time and warming only the name in the args warms a cache the leg never reads.
"unsloth/Qwen3-0.6B",
- "unsloth/qwen3-0.6b-unsloth-bnb-4bit",
+ "unsloth/Qwen3-0.6B-unsloth-bnb-4bit",
"unsloth/Qwen2.5-0.5B-Instruct",
# Last now rather than first: gptoss is admitted only once a card empties, around t~500 on the measured schedule, so it has the most slack of anything here.
"unsloth/gpt-oss-20b-unsloth-bnb-4bit",
@@ -475,8 +475,8 @@ PREFETCH_REPOS: tuple[str, ...] = (
# Declared name -> what actually gets loaded on an sm_75 card. Kept beside the list it corrects so the two cannot drift apart silently.
LOAD_REDIRECTS: dict[str, str] = {
"unsloth/gpt-oss-20b": "unsloth/gpt-oss-20b-unsloth-bnb-4bit",
- # CASE MATTERS HERE, and it is not a typo: two runs report `resolved_checkpoint: unsloth/qwen3-0.6b-unsloth-bnb-4bit` in lower case while gpt-oss above resolves with its capitals intact. The HF cache keys on the literal string, so prefetching the pretty spelling warms a directory the leg never reads and the download happens twice. Copied from the report rather than typed.
- "unsloth/Qwen3-0.6B": "unsloth/qwen3-0.6b-unsloth-bnb-4bit",
+ # CASE MATTERS HERE: the HF cache keys on the literal string, so prefetching another spelling warms a directory the leg never reads and the download happens twice. Copied from the report rather than typed. Runs before #8058 reported `unsloth/qwen3-0.6b-unsloth-bnb-4bit` in lower case; since #8058 the loader keeps the canonical case, and kernel unsloth-t4-ci-n361d0b65ec92-0d7d reports `resolved_checkpoint: unsloth/Qwen3-0.6B-unsloth-bnb-4bit`.
+ "unsloth/Qwen3-0.6B": "unsloth/Qwen3-0.6B-unsloth-bnb-4bit",
}
diff --git a/tests/kaggle/t4_smoke/run_t4_smoke.py b/tests/kaggle/t4_smoke/run_t4_smoke.py
index 3a8e9fc27b7..71aeee410eb 100644
--- a/tests/kaggle/t4_smoke/run_t4_smoke.py
+++ b/tests/kaggle/t4_smoke/run_t4_smoke.py
@@ -66,6 +66,7 @@ import hashlib
import json
import os
import platform
+import re
import subprocess
import tempfile
import sys
@@ -1397,6 +1398,21 @@ def reference_step_count(ref: dict):
# what is in the file, and canary_dataset.jsonl is inside the workflow's paths
# filter, so a run triggered BY editing it is exactly the run that would
# otherwise be band-checked against a trace of the old rows.
+# Hub repo ids among the reference-defining fields, compared without case.
+_REPO_ID_KEYS = frozenset({"model", "resolved_checkpoint"})
+_HUB_REPO_ID_RE = re.compile(r"[A-Za-z0-9][A-Za-z0-9._-]*/[A-Za-z0-9._-]+")
+
+
+def _is_hub_repo_id(value: Any) -> bool:
+ """``owner/name`` and not a directory here. A local checkpoint path keeps its case:
+ on a case-sensitive filesystem /models/Foo and /models/foo are different weights."""
+ return (
+ isinstance(value, str)
+ and _HUB_REPO_ID_RE.fullmatch(value) is not None
+ and not os.path.exists(value)
+ )
+
+
REFERENCE_DEFINING_SETTINGS = (
"max_steps",
"dataset_digest",
@@ -1602,6 +1618,12 @@ def check_reference(
continue
observed_pairs.append((key, ref[key], observed))
for key, expected, observed in observed_pairs:
+ if key in _REPO_ID_KEYS and _is_hub_repo_id(expected) and _is_hub_repo_id(observed):
+ # Hub repo ids are case-insensitive: unsloth/qwen2.5-0.5b-instruct-unsloth-bnb-4bit and
+ # unsloth/Qwen2.5-0.5B-Instruct-unsloth-bnb-4bit are one repo, and which spelling
+ # the loader reports changed in #8058. The revision below still pins the weights.
+ if expected.casefold() == observed.casefold():
+ continue
if expected != observed:
verdict["config_differences"].append(
{"key": key, "reference": expected, "observed": observed}
diff --git a/tests/kaggle/test_prefetch_covers_the_wired_legs.py b/tests/kaggle/test_prefetch_covers_the_wired_legs.py
index 5c5eebd0538..0bbf5fda8f0 100644
--- a/tests/kaggle/test_prefetch_covers_the_wired_legs.py
+++ b/tests/kaggle/test_prefetch_covers_the_wired_legs.py
@@ -120,7 +120,7 @@ def test_every_redirect_target_is_prefetched_under_its_EXACT_name():
`models--unsloth--qwen3-0.6b-unsloth-bnb-4bit` and
`models--unsloth--Qwen3-0.6B-unsloth-bnb-4bit` are different directories, so
- prefetching the pretty spelling of a repo the loader asks for in lower case
+ prefetching a spelling other than the one the loader asks for
warms a cache nobody reads and the session downloads it twice -- at full
cost, with no error and nothing red. Two hardware reports give the exact
strings; this asserts the lists agree with them character for character.
@@ -142,10 +142,12 @@ def test_every_redirect_target_is_prefetched_under_its_EXACT_name():
)
-def test_the_lower_case_qwen3_redirect_is_not_tidied_away():
- """It looks like a typo and it is a measurement. Written down so the next
- reader corrects the capitals in the report, not in this file."""
- assert legs.LOAD_REDIRECTS["unsloth/Qwen3-0.6B"] == "unsloth/qwen3-0.6b-unsloth-bnb-4bit"
+def test_the_qwen3_redirect_matches_the_measured_case():
+ """A measurement, not a style choice. Before #8058 the loader reported this repo in
+ lower case; since #8058 it keeps the canonical capitals, and the kernel report cited in
+ legs.py shows exactly this string. If the loader's spelling moves again, correct it here
+ from a report, not by hand."""
+ assert legs.LOAD_REDIRECTS["unsloth/Qwen3-0.6B"] == "unsloth/Qwen3-0.6B-unsloth-bnb-4bit"
def test_a_blob_is_counted_once_not_once_per_symlink(tmp_path, monkeypatch):
diff --git a/tests/kaggle/test_t4_payload_assertions.py b/tests/kaggle/test_t4_payload_assertions.py
index 12eca6310f1..8e82743c85d 100644
--- a/tests/kaggle/test_t4_payload_assertions.py
+++ b/tests/kaggle/test_t4_payload_assertions.py
@@ -825,6 +825,96 @@ def test_the_model_revision_travels_with_the_reference(tmp_path):
assert reference_failures(moved, 0.10)
+def test_a_repo_id_that_differs_only_in_case_is_the_same_reference(tmp_path):
+ """The Hub treats repo ids case-insensitively, and #8058 changed the loader from reporting
+ the lower-case spelling to the canonical one. The committed reference still carries the
+ old spelling; that is the same checkpoint, and the revision pin is what moves with weights."""
+ check_reference, ref, reference_failures = _shared_setup_2(tmp_path)
+ payload = json.loads(ref.read_text())
+ payload["model"] = "unsloth/qwen2.5-0.5b-instruct"
+ payload["resolved_checkpoint"] = "unsloth/qwen2.5-0.5b-instruct-unsloth-bnb-4bit"
+ payload["resolved_revision"] = "10413c288cb9629acdf60b3e0229f3ba75efe413"
+ ref.write_text(json.dumps(payload))
+ observed = [{"step": s, "loss": 1.0 / s, "grad_norm": 3.0} for s in (1, 2, 3)]
+
+ recased = check_reference(
+ observed,
+ ref,
+ 0.10,
+ 0.05,
+ max_steps = 3,
+ model = "unsloth/Qwen2.5-0.5B-Instruct",
+ resolved_checkpoint = "unsloth/Qwen2.5-0.5B-Instruct-unsloth-bnb-4bit",
+ resolved_revision = "10413c288cb9629acdf60b3e0229f3ba75efe413",
+ )
+ assert recased["status"] == "ok", recased
+ assert not reference_failures(recased, 0.10)
+
+ # Case is the only thing forgiven: another checkpoint, or a revision with different
+ # capitals, is still a different experiment.
+ other = check_reference(
+ observed,
+ ref,
+ 0.10,
+ 0.05,
+ max_steps = 3,
+ resolved_checkpoint = "unsloth/Qwen2.5-1.5B-Instruct-unsloth-bnb-4bit",
+ resolved_revision = "10413c288cb9629acdf60b3e0229f3ba75efe413",
+ )
+ assert other["status"] == "config_mismatch"
+ revision = check_reference(
+ observed,
+ ref,
+ 0.10,
+ 0.05,
+ max_steps = 3,
+ resolved_checkpoint = "unsloth/qwen2.5-0.5b-instruct-unsloth-bnb-4bit",
+ resolved_revision = "10413C288CB9629ACDF60B3E0229F3BA75EFE413",
+ )
+ assert revision["status"] == "config_mismatch"
+
+
+def test_a_local_checkpoint_path_keeps_its_case(tmp_path):
+ """Only Hub repo ids are case-insensitive. A local directory is a path, and on a
+ case-sensitive filesystem another capitalisation can hold other weights."""
+ check_reference, ref, reference_failures = _shared_setup_2(tmp_path)
+ upper = tmp_path / "models" / "Foo"
+ lower = tmp_path / "models" / "foo"
+ upper.mkdir(parents = True)
+ if lower.exists():
+ pytest.skip("case-insensitive filesystem: the two paths are one directory here")
+ lower.mkdir()
+ payload = json.loads(ref.read_text())
+ payload["resolved_checkpoint"] = str(upper)
+ ref.write_text(json.dumps(payload))
+ observed = [{"step": s, "loss": 1.0 / s, "grad_norm": 3.0} for s in (1, 2, 3)]
+
+ verdict = check_reference(
+ observed, ref, 0.10, 0.05, max_steps = 3, resolved_checkpoint = str(lower)
+ )
+ assert verdict["status"] == "config_mismatch", verdict
+ assert reference_failures(verdict, 0.10)
+
+ # A relative "owner/name" that exists here is a directory too, not a repo id.
+ rel_upper = tmp_path / "Owner" / "Model"
+ rel_lower = tmp_path / "owner" / "model"
+ rel_upper.mkdir(parents = True)
+ rel_lower.mkdir(parents = True)
+ payload["resolved_checkpoint"] = "Owner/Model"
+ ref.write_text(json.dumps(payload))
+ import os
+
+ cwd = os.getcwd()
+ os.chdir(tmp_path)
+ try:
+ relative = check_reference(
+ observed, ref, 0.10, 0.05, max_steps = 3, resolved_checkpoint = "owner/model"
+ )
+ finally:
+ os.chdir(cwd)
+ assert relative["status"] == "config_mismatch", relative
+
+
def test_a_reference_with_no_recorded_revision_does_not_refuse(tmp_path):
"""The committed file predates this; unknown is unknown, not a mismatch."""
check_reference, ref = _shared_setup_3(tmp_path)