From a2743083309156c5e729c2e55aa7b2bee529e3f4 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sun, 5 Jul 2026 05:25:03 +0000
Subject: [PATCH] Check the cancel event between predownload files and probe
 local dirs for LTX-2.3

A warm-cache predownload sweep never consults the cancel event (each cached
file returns instantly), so an unload during it was ignored until a cold
file hit the network. The 2.3 detection also only probed bare-file local
repos; resolve directory repos through the same child resolver the loader
uses so their base pull is scoped too.
---
 studio/backend/core/inference/video.py     | 20 +++++++++++++-
 studio/backend/tests/test_video_backend.py | 31 ++++++++++++++++++++++
 2 files changed, 50 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/inference/video.py b/studio/backend/core/inference/video.py
index 4bd7751f7c..3b64e37dca 100644
--- a/studio/backend/core/inference/video.py
+++ b/studio/backend/core/inference/video.py
@@ -331,8 +331,21 @@ class VideoBackend:
 
                 probe = checkpoint_local
                 if probe is None:
+                    # Local repos: a bare file, or a directory whose child the same
+                    # resolver load_pipeline uses picks out. Unresolvable here means
+                    # load_pipeline will surface the real error; keep the wide pull.
                     root = Path(kwargs["repo_id"]).expanduser()
-                    probe = root if root.is_file() else None
+                    if root.is_file():
+                        probe = root
+                    elif root.is_dir():
+                        try:
+                            probe = self._resolve_checkpoint_path(
+                                kwargs["repo_id"],
+                                kwargs.get("gguf_filename"),
+                                kwargs.get("hf_token"),
+                            )
+                        except Exception:  # noqa: BLE001 -- surfaced by load_pipeline
+                            probe = None
                 ltx23 = probe is not None and is_ltx23_checkpoint(probe)
                 if ltx23:
                     expected = self._estimate_download_bytes(
@@ -463,6 +476,11 @@ class VideoBackend:
 
             snapshot_root: Optional[Path] = None
             for name, _ in files:
+                # Explicit per-file check: a fully-cached file returns without ever
+                # consulting the event, so a warm-cache sweep would otherwise run to
+                # completion after an unload already cancelled this load.
+                if self._cancel_event.is_set():
+                    raise RuntimeError(VIDEO_CANCELLED_MSG)
                 local = Path(
                     hf_hub_download_with_xet_fallback(
                         base, name, hf_token, cancel_event = self._cancel_event
diff --git a/studio/backend/tests/test_video_backend.py b/studio/backend/tests/test_video_backend.py
index c34cee5bf9..209983e4f6 100644
--- a/studio/backend/tests/test_video_backend.py
+++ b/studio/backend/tests/test_video_backend.py
@@ -528,3 +528,34 @@ def test_base_download_files_ltx23_keeps_only_shared_components():
     assert not any(
         n.startswith(("vae/", "connectors/", "latent_upsampler/", "transformer/")) for n in names
     )
+
+
+def test_predownload_base_honors_cancel_between_files(monkeypatch):
+    # A warm-cache sweep returns each file instantly without consulting the event,
+    # so the loop must check it explicitly or an unload mid-predownload is ignored.
+    backend = VideoBackend()
+    backend._cancel_event.set()
+    calls: list = []
+    monkeypatch.setattr(
+        "utils.hf_xet_fallback.hf_hub_download_with_xet_fallback",
+        lambda repo, fn, tok, **kw: (calls.append(fn), f"/cache/{fn}")[1],
+    )
+
+    class _Api:
+        def __init__(self, token = None):
+            pass
+
+        def model_info(self, repo, files_metadata = True):
+            return types.SimpleNamespace(
+                siblings = [
+                    _sibling("model_index.json", 1),
+                    _sibling("vae/diffusion_pytorch_model.safetensors", 2),
+                ]
+            )
+
+    import huggingface_hub
+
+    monkeypatch.setattr(huggingface_hub, "HfApi", _Api)
+    with pytest.raises(RuntimeError, match = "cancelled"):
+        backend._predownload_base("base/repo", None, "pipeline")
+    assert calls == []