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.

Make the weights_at_root call-site and diffusion autoopt parity tests host independent (#13582)

Daniel Han2a4d4df8930e
diff --git a/studio/backend/tests/test_diffusion_autoopt_parity.py b/studio/backend/tests/test_diffusion_autoopt_parity.py
index fd3b39b1c60..a4547f8e599 100644
--- a/studio/backend/tests/test_diffusion_autoopt_parity.py
+++ b/studio/backend/tests/test_diffusion_autoopt_parity.py
@@ -37,6 +37,7 @@ _ENV_CLEAR = (
"UNSLOTH_DIFFUSION_OFFLOAD_CUDA_GRAPH",
"TORCHDYNAMO_DISABLE",
"UNSLOTH_DIFFUSION_AUTOOPT_DISCOVER",
+ "UNSLOTH_DIFFUSION_GGUF_COMPILE_DEQUANT",
)
@@ -157,8 +158,13 @@ def _engaged(result: Any) -> bool:
def _teardown() -> None:
"""Restore every process-global patch a case may have installed (old and new teardown entry points)."""
- from core.inference import diffusion_arch_patches, diffusion_eager_patches
+ from core.inference import (
+ diffusion_arch_patches,
+ diffusion_eager_patches,
+ diffusion_gguf_compile,
+ )
+ diffusion_gguf_compile.uninstall_all()
diffusion_eager_patches.uninstall_patches()
diffusion_arch_patches.uninstall_arch_patches()
for mod_name in (
@@ -446,6 +452,14 @@ def snapshot(entry: Callable) -> dict:
"_HELPERS",
{"patch": _patch_function, "restore": _restore_original},
)
+ # diffusers' GGUF utils need the optional `gguf` package; without a stand-in every gguf=1 case is host-dependent.
+ from core.inference import diffusion_gguf_compile
+
+ fake_gguf_utils = types.ModuleType("diffusers.quantizers.gguf.utils")
+ fake_gguf_utils.dequantize_gguf_tensor = lambda tensor: tensor
+ patch.setattr(diffusion_gguf_compile, "_gguf_utils", lambda: fake_gguf_utils)
+ patch.setattr(diffusion_gguf_compile, "_compiled_dequant_installed", False)
+ patch.setattr(diffusion_gguf_compile, "_compile_error", None)
saved_env = {k: os.environ.pop(k, None) for k in _ENV_CLEAR}
patch.setattr(torch.cuda, "is_available", lambda: True)
patch.setattr(torch.cuda, "get_device_capability", lambda *a, **k: (9, 0))
diff --git a/tests/test_st_subfolder_weights_are_fetched.py b/tests/test_st_subfolder_weights_are_fetched.py
index ec5cedb7786..cde4e955696 100644
--- a/tests/test_st_subfolder_weights_are_fetched.py
+++ b/tests/test_st_subfolder_weights_are_fetched.py
@@ -231,25 +231,51 @@ def test_an_older_unsloth_zoo_degrades_instead_of_crashing(modules_json):
def test_both_weights_at_root_call_sites_go_through_the_check():
- """`weights_at_root = True` is passed from exactly two places (vision.py and
- llama.py), both reaching the prune through maybe_prefetch_hf_snapshot, so one
- carve-out covers both. A third call site, or an inlined copy of the patterns,
- fails here instead of leaving half the loaders pruning ST weights."""
+ """`weights_at_root = True` is passed only from vision.py and llama.py, and every
+ such call reaches the prune through maybe_prefetch_hf_snapshot (directly, or via
+ resolve_vllm_model_path for the vLLM path, #13486), so one carve-out covers them
+ all. A call into anything else, or an inlined copy of the patterns, fails here
+ instead of leaving some loader pruning ST weights."""
+ # AST, not grep: the names also appear in docstring prose.
+ import ast
+
+ routed = {"maybe_prefetch_hf_snapshot", "resolve_vllm_model_path"}
root = Path(U.__file__).resolve().parents[1]
sites = []
for p in root.rglob("*.py"):
if "tests" in p.parts:
continue
- for n, line in enumerate(p.read_text(encoding = "utf-8").splitlines(), 1):
- if "weights_at_root = True" in line:
- sites.append(f"{p.name}:{n}")
- assert sorted(s.split(":")[0] for s in sites) == ["llama.py", "vision.py"], sites
-
- # AST, not grep: the name also appears in prose inside a docstring, and a text count would police the documentation
- # instead of the code.
- import ast
+ for n in ast.walk(ast.parse(p.read_text(encoding = "utf-8"))):
+ if not isinstance(n, ast.Call):
+ continue
+ for k in n.keywords:
+ if (
+ k.arg == "weights_at_root"
+ and isinstance(k.value, ast.Constant)
+ and k.value.value is True
+ ):
+ callee = getattr(n.func, "id", getattr(n.func, "attr", None))
+ sites.append((p.name, n.lineno, callee))
+ assert {s[0] for s in sites} == {"llama.py", "vision.py"}, sites
+ assert all(s[2] in routed for s in sites), sites
+ # Each loader's normal and vLLM paths, so losing the keyword on either one fails too.
+ expected = {(f, callee) for f in ("llama.py", "vision.py") for callee in routed}
+ assert expected <= {(s[0], s[2]) for s in sites}, sites
tree = ast.parse(Path(U.__file__).read_text(encoding = "utf-8"))
+ resolver = next(
+ n
+ for n in ast.walk(tree)
+ if isinstance(n, ast.FunctionDef) and n.name == "resolve_vllm_model_path"
+ )
+ # `**kwargs` forwarded, or the resolver could drop weights_at_root on the way.
+ assert any(
+ isinstance(c, ast.Call)
+ and getattr(c.func, "id", None) == "maybe_prefetch_hf_snapshot"
+ and any(k.arg is None for k in c.keywords)
+ for c in ast.walk(resolver)
+ ), "resolve_vllm_model_path no longer forwards its kwargs to maybe_prefetch_hf_snapshot"
+
loads = [
n
for n in ast.walk(tree)