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: stop the SWA config probe from adding a phantom base model and refetching every load (#10607)

Chaitanya Laxman095e66476cda
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 5755366e2fa..0f4bc5cc686 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -2522,7 +2522,7 @@ _BOOTSTRAP_SWA_DEFAULTS: dict[str, int] = {
}
# Process-wide cache backed by JSON on disk. Values are int period or
-# list[bool] mask. Lazy-loaded.
+# list[bool] mask; `__missed_repos__` lists casefolded repos with no SWA field. Lazy-loaded.
_SWA_CACHE: Optional[dict] = None
_SWA_CACHE_LOCK = threading.Lock()
@@ -2704,25 +2704,56 @@ def _swa_entry_from_layer_types(lt) -> Optional[object]:
return None
+_SWA_CONFIRMED_MISS = False
+_SWA_MISSED_REPOS_KEY = "__missed_repos__"
+
+
+def _swa_missed_repos(cache: dict) -> set:
+ missed = cache.get(_SWA_MISSED_REPOS_KEY)
+ if not isinstance(missed, list):
+ return set()
+ return {item.casefold() for item in missed if isinstance(item, str)}
+
+
+def _remember_swa_repo_miss(cache: dict, repo_id: str) -> None:
+ folded = repo_id.casefold()
+ with _SWA_CACHE_LOCK:
+ missed = _swa_missed_repos(cache)
+ if folded in missed:
+ return
+ missed.add(folded)
+ cache[_SWA_MISSED_REPOS_KEY] = sorted(missed)
+ _save_swa_cache(cache)
+
+
def _fetch_swa_entry_from_hf(repo_id: str) -> Optional[object]:
try:
- from huggingface_hub import hf_hub_download
+ from huggingface_hub import hf_hub_download, try_to_load_from_cache
from utils.hf_cache_settings import active_hf_hub_cache
from utils.hf_probe import hf_file_definitely_absent
- # Avoid caching the expected 404 for GGUF repos without config.json.
- if hf_file_definitely_absent(repo_id, "config.json"):
- return None
- cfg_path = call_hub_with_anonymous_retry(
- hf_hub_download,
- None,
- repo_id,
- "config.json",
- repo_type = "model",
- cache_dir = active_hf_hub_cache(),
- )
- with open(cfg_path, encoding = "utf-8-sig") as f:
- cfg = json.load(f)
+ cached = try_to_load_from_cache(repo_id, "config.json", cache_dir = active_hf_hub_cache())
+ if isinstance(cached, str):
+ with open(cached, encoding = "utf-8-sig") as f:
+ cfg = json.load(f)
+ else:
+ # Only a confirmed 404 is a miss; timeouts, 429 and gated stay None so a later load retries.
+ if hf_file_definitely_absent(repo_id, "config.json"):
+ return _SWA_CONFIRMED_MISS
+ # Throwaway cache: a config-only repo in the user's cache lists as a phantom model (#10047).
+ with tempfile.TemporaryDirectory(
+ prefix = "unsloth-swa-", ignore_cleanup_errors = True
+ ) as tmp:
+ cfg_path = call_hub_with_anonymous_retry(
+ hf_hub_download,
+ None,
+ repo_id,
+ "config.json",
+ repo_type = "model",
+ cache_dir = tmp,
+ )
+ with open(cfg_path, encoding = "utf-8-sig") as f:
+ cfg = json.load(f)
except Exception:
return None
@@ -2730,7 +2761,8 @@ def _fetch_swa_entry_from_hf(repo_id: str) -> Optional[object]:
period = src.get("sliding_window_pattern")
if isinstance(period, int) and period > 0:
return period
- return _swa_entry_from_layer_types(src.get("layer_types"))
+ entry = _swa_entry_from_layer_types(src.get("layer_types"))
+ return entry if entry is not None else _SWA_CONFIRMED_MISS
def _arch_aliases(arch: str) -> tuple:
@@ -2842,12 +2874,21 @@ def _resolve_swa_pattern(
_persist(entry)
return _entry_to_mask(entry)
- # Tier 3: live HF fetch (result persistently cached)
+ # Tier 3: live HF fetch. Misses are keyed per repo, not arch: a new GGUF may name a new source.
if allow_network:
+ seen = set()
+ missed = _swa_missed_repos(cache)
for repo_id in source_repo_candidates:
if not repo_id:
continue
+ folded = repo_id.casefold()
+ if folded in seen or folded in missed:
+ continue
+ seen.add(folded)
entry = _fetch_swa_entry_from_hf(repo_id)
+ if entry is _SWA_CONFIRMED_MISS:
+ _remember_swa_repo_miss(cache, folded)
+ continue
if entry is not None:
_persist(entry)
return _entry_to_mask(entry)
diff --git a/studio/backend/tests/test_kv_cache_estimation.py b/studio/backend/tests/test_kv_cache_estimation.py
index 53553624edc..42738e65c1c 100644
--- a/studio/backend/tests/test_kv_cache_estimation.py
+++ b/studio/backend/tests/test_kv_cache_estimation.py
@@ -613,6 +613,117 @@ class TestDynamicSwaResolver:
assert b._sliding_window_pattern is None
assert not (tmp_path / "swa_cache.json").exists()
+ def test_remembered_miss_skips_hf_refetch(self, monkeypatch, tmp_path):
+ self._isolate_cache(monkeypatch, tmp_path)
+ calls = []
+ monkeypatch.setattr(
+ lc,
+ "_fetch_swa_entry_from_hf",
+ lambda repo_id: calls.append(repo_id) or lc._SWA_CONFIRMED_MISS,
+ )
+ monkeypatch.setattr(lc, "_resolve_swa_entry_from_transformers", lambda arch: None)
+ general = {"general.source.huggingface.repository": "vendor/does-not-exist"}
+ first = _backend_from_gguf("newmodel", _SWA_FIELDS, general = general)
+ assert first._sliding_window_pattern is None
+ assert calls == ["vendor/does-not-exist"]
+ with open(tmp_path / "swa_cache.json") as f:
+ assert json.load(f) == {"__missed_repos__": ["vendor/does-not-exist"]}
+ monkeypatch.setattr(lc, "_SWA_CACHE", None)
+ second = _backend_from_gguf("newmodel", _SWA_FIELDS, general = general)
+ assert second._sliding_window_pattern is None
+ assert calls == ["vendor/does-not-exist"]
+
+ def test_transient_hf_error_is_retried(self, monkeypatch, tmp_path):
+ self._isolate_cache(monkeypatch, tmp_path)
+ calls = []
+ monkeypatch.setattr(
+ lc, "_fetch_swa_entry_from_hf", lambda repo_id: calls.append(repo_id) or None
+ )
+ monkeypatch.setattr(lc, "_resolve_swa_entry_from_transformers", lambda arch: None)
+ general = {"general.source.huggingface.repository": "vendor/newmodel-base"}
+ first = _backend_from_gguf("newmodel", _SWA_FIELDS, general = general)
+ assert first._sliding_window_pattern is None
+ assert not (tmp_path / "swa_cache.json").exists()
+ monkeypatch.setattr(lc, "_SWA_CACHE", None)
+ monkeypatch.setattr(
+ lc, "_fetch_swa_entry_from_hf", lambda repo_id: calls.append(repo_id) or 4
+ )
+ second = _backend_from_gguf("newmodel", _SWA_FIELDS, general = general)
+ assert second._sliding_window_pattern == [(i + 1) % 4 != 0 for i in range(12)]
+ assert calls == ["vendor/newmodel-base", "vendor/newmodel-base"]
+
+ def test_repo_miss_does_not_block_another_candidate(self, monkeypatch, tmp_path):
+ self._isolate_cache(monkeypatch, tmp_path)
+ calls = []
+
+ def fake_fetch(repo_id):
+ calls.append(repo_id)
+ return 4 if repo_id == "vendor/newmodel-base" else lc._SWA_CONFIRMED_MISS
+
+ monkeypatch.setattr(lc, "_fetch_swa_entry_from_hf", fake_fetch)
+ monkeypatch.setattr(lc, "_resolve_swa_entry_from_transformers", lambda arch: None)
+ b = _backend_from_gguf(
+ "newmodel",
+ _SWA_FIELDS,
+ general = {
+ "general.source.huggingface.repository": "quanter/newmodel-GGUF",
+ "general.base_model.0.repo_url": "https://huggingface.co/vendor/newmodel-base",
+ },
+ )
+ assert b._sliding_window_pattern == [(i + 1) % 4 != 0 for i in range(12)]
+ assert calls == ["quanter/newmodel-GGUF", "vendor/newmodel-base"]
+
+ def test_casefold_duplicate_repos_are_fetched_once(self, monkeypatch, tmp_path):
+ self._isolate_cache(monkeypatch, tmp_path)
+ calls = []
+ monkeypatch.setattr(
+ lc, "_fetch_swa_entry_from_hf", lambda repo_id: calls.append(repo_id) or None
+ )
+ monkeypatch.setattr(lc, "_resolve_swa_entry_from_transformers", lambda arch: None)
+ b = _backend_from_gguf(
+ "newmodel",
+ _SWA_FIELDS,
+ general = {
+ "general.source.huggingface.repository": "DeepSeek-AI/Flash",
+ "general.organization": "deepseek-ai",
+ "general.basename": "Flash",
+ },
+ )
+ assert b._sliding_window_pattern is None
+ assert calls == ["DeepSeek-AI/Flash"]
+
+ def test_hf_fetch_leaves_no_repo_in_the_model_cache(self, monkeypatch, tmp_path):
+ import huggingface_hub
+ import utils.hf_probe
+
+ active = tmp_path / "hub"
+ active.mkdir()
+ monkeypatch.setattr("utils.hf_cache_settings.active_hf_hub_cache", lambda: str(active))
+ monkeypatch.setattr(utils.hf_probe, "hf_file_definitely_absent", lambda *a, **k: False)
+
+ def fake_download(repo_id, filename, *, cache_dir, **kwargs):
+ path = Path(cache_dir) / f"models--{repo_id.replace('/', '--')}" / filename
+ path.parent.mkdir(parents = True)
+ path.write_text(json.dumps({"sliding_window": 128}))
+ return str(path)
+
+ monkeypatch.setattr(huggingface_hub, "hf_hub_download", fake_download)
+ assert (
+ lc._fetch_swa_entry_from_hf("deepseek-ai/DeepSeek-V4-Flash") is lc._SWA_CONFIRMED_MISS
+ )
+ assert list(active.iterdir()) == []
+
+ def test_hf_fetch_reads_an_already_cached_config_offline(self, monkeypatch, tmp_path):
+ import huggingface_hub
+
+ cfg = tmp_path / "config.json"
+ cfg.write_text(json.dumps({"sliding_window_pattern": 6}))
+ monkeypatch.setattr(huggingface_hub, "try_to_load_from_cache", lambda *a, **k: str(cfg))
+ monkeypatch.setattr(
+ huggingface_hub, "hf_hub_download", lambda *a, **k: pytest.fail("network fetch")
+ )
+ assert lc._fetch_swa_entry_from_hf("org/model") == 6
+
class TestTransformersIntrospection:
"""Tier 2.5: default-init the matching Config; on failure, parse via inspect."""