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.

fix: context override from JIT API load (#11524)

Jyotirmoy Roydef4d4791252
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index f20fbbd52df..737184b51ab 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -7662,6 +7662,16 @@ def _public_model_identifier(requested: str, resolved: str) -> str:
return requested if is_ollama_manifest_ref(requested) else resolved
+def _canonical_model_identity(model_id: Optional[str]) -> str:
+ """Case-folded *model_id*, reading an HF cache snapshot path as its ``org/name`` repo id.
+ Other local paths stay distinct: two files sharing a stem are two models."""
+ from core.inference.model_ids import hf_cache_repo_id
+
+ if not model_id:
+ return ""
+ return (hf_cache_repo_id(model_id) or str(model_id)).lower()
+
+
def _as_ollama_manifest_request(request):
"""*request* with a materialized Ollama ``.gguf`` link rewritten to the tag that produced it."""
from hub.services.models.ollama import ollama_manifest_ref_for_path, ollama_model_ref_files
@@ -14786,7 +14796,11 @@ def _resolve_inherited_extra_args(
resolved_variant = (config.gguf_variant or "").lower()
request_variant = (request.gguf_variant or "").lower()
stored_variant = (source[1] or "").lower() if source else ""
- same_model = bool(source and source[0] and source[0].lower() == model_identifier.lower())
+ # A picker load records the snapshot path, an API auto-switch loads the repo id.
+ stored_identity = _canonical_model_identity(source[0]) if source else ""
+ same_model = bool(
+ stored_identity and stored_identity == _canonical_model_identity(model_identifier)
+ )
if request.gguf_variant:
variant_mismatch = request_variant != stored_variant
else:
diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index aba099e0886..916020adaa3 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -1867,10 +1867,11 @@ def update_openai_auto_switch_override(
) -> ModelOverridesResponse:
from core.inference.llama_server_args import (
drop_managed_flags,
+ parse_ctx_override,
strip_shadowing_flags,
validate_extra_args,
)
- from utils.openai_auto_switch_settings import get_model_override
+ from utils.openai_auto_switch_settings import MAX_SEQ_LENGTH_CEILING, get_model_override
try:
if payload.fill_absent_fields and payload.remove is True:
@@ -2056,12 +2057,27 @@ def update_openai_auto_switch_override(
):
_kept_reasoning_budget = -1
_kept_reasoning_budget_message = ""
+ # A -c sent with this save is what its load runs at (llama.cpp takes the last -c); store it as
+ # the context or auto-switch strips it as stale (#11511). Carried-over flags and fills keep that rule.
+ max_seq_length = payload.max_seq_length
+ custom_context_length = payload.custom_context_length
+ if payload.llama_extra_args is not None and not payload.fill_absent_fields:
+ try:
+ explicit_ctx = parse_ctx_override(extra_args)
+ except ValueError:
+ explicit_ctx = None
+ # Past the stored ceiling the field would be dropped, leaving the flag unchecked.
+ if explicit_ctx and explicit_ctx <= MAX_SEQ_LENGTH_CEILING:
+ if max_seq_length is not None:
+ max_seq_length = explicit_ctx
+ if custom_context_length is not None:
+ custom_context_length = explicit_ctx
set_model_override(
target_id,
llama_extra_args = extra_args,
keep_empty_extra_args = keep_empty,
- max_seq_length = payload.max_seq_length,
- custom_context_length = payload.custom_context_length,
+ max_seq_length = max_seq_length,
+ custom_context_length = custom_context_length,
kv_cache_dtype = payload.kv_cache_dtype,
mlx_kv_bits = payload.mlx_kv_bits,
speculative_type = payload.speculative_type,
diff --git a/studio/backend/tests/test_gguf_reload_inheritance.py b/studio/backend/tests/test_gguf_reload_inheritance.py
index d9351ff7397..8b4fdd0fa57 100644
--- a/studio/backend/tests/test_gguf_reload_inheritance.py
+++ b/studio/backend/tests/test_gguf_reload_inheritance.py
@@ -548,3 +548,244 @@ def test_an_inherited_env_budget_does_not_force_a_reload():
)
is True
)
+
+
+# ── Canonical identity normalization ────────────────────────────
+
+
+def test_canonical_model_identity_resolves_snapshot_to_repo():
+ from routes.inference import _canonical_model_identity
+ snapshot = "/models--unsloth--Qwen3.8-27B-GGUF/snapshots/abc123"
+ assert _canonical_model_identity(snapshot) == "unsloth/qwen3.8-27b-gguf"
+
+
+def test_canonical_model_identity_folds_repo_id_case():
+ from routes.inference import _canonical_model_identity
+ assert _canonical_model_identity("unsloth/Qwen3.8-27B-GGUF") == "unsloth/qwen3.8-27b-gguf"
+
+
+def test_canonical_model_identity_reads_windows_snapshot_path():
+ from routes.inference import _canonical_model_identity
+ snapshot = r"C:\hub\models--unsloth--Qwen3.8-27B-GGUF\snapshots\abc123"
+ assert _canonical_model_identity(snapshot) == "unsloth/qwen3.8-27b-gguf"
+
+
+def test_canonical_model_identity_empty_and_none():
+ from routes.inference import _canonical_model_identity
+ assert _canonical_model_identity("") == ""
+ assert _canonical_model_identity(None) == ""
+
+
+def test_canonical_model_identity_keeps_other_local_paths():
+ from routes.inference import _canonical_model_identity
+ assert _canonical_model_identity("/models/my-model.gguf") == "/models/my-model.gguf"
+ # A models-- folder outside the cache layout is not a repo id.
+ assert _canonical_model_identity("/models--org--name/x.gguf") == "/models--org--name/x.gguf"
+
+
+# ── _resolve_inherited_extra_args: snapshot vs repo id ──────────
+
+
+def test_inheritance_works_when_stored_is_snapshot_and_load_is_repo():
+ """Storing extras under a snapshot-path identifier and resolving on a JIT
+ auto-switch load of the same repo id must still inherit them.
+ """
+ from routes.inference import _resolve_inherited_extra_args
+
+ class FakeRequest:
+ model_path = "unsloth/Qwen3.8-27B-GGUF"
+ gguf_variant = None
+ llama_extra_args = None
+ model_fields_set = set()
+
+ class FakeConfig:
+ is_gguf = True
+ gguf_variant = "ud-q6_k_l"
+ identifier = "unsloth/Qwen3.8-27B-GGUF"
+
+ class FakeBackend:
+ extra_args = ["--rope-scaling", "yarn", "-c", "300000"]
+ extra_args_source = ("/models--unsloth--Qwen3.8-27B-GGUF/snapshots/abc123", "ud-q6_k_l")
+
+ import routes.inference as _routes
+
+ _get_llama_cpp_backend_orig = _routes.get_llama_cpp_backend
+ _routes.get_llama_cpp_backend = lambda: FakeBackend()
+ try:
+ result = _resolve_inherited_extra_args(
+ FakeRequest(), FakeConfig(), "unsloth/Qwen3.8-27B-GGUF", None
+ )
+ assert result == [
+ "--rope-scaling",
+ "yarn",
+ "-c",
+ "300000",
+ ], f"Expected inherited extras but got: {result}"
+ finally:
+ _routes.get_llama_cpp_backend = _get_llama_cpp_backend_orig
+
+
+def test_inheritance_works_when_case_differs():
+ """A stored snapshot/repo with different casing than the load model_identifier
+ still inherits extras for the same Hub GGUF + same variant."""
+ from routes.inference import _resolve_inherited_extra_args
+
+ class FakeRequest:
+ model_path = "unsloth/Qwen3.8-27B-GGUF"
+ gguf_variant = None
+ llama_extra_args = None
+ model_fields_set = set()
+
+ class FakeConfig:
+ is_gguf = True
+ gguf_variant = "ud-q6_k_l"
+ identifier = "unsloth/Qwen3.8-27B-GGUF"
+
+ class FakeBackend:
+ extra_args = ["--rope-scaling", "yarn", "-c", "300000"]
+ extra_args_source = ("/models--Unsloth--Qwen3.8-27B-GGUF/snapshots/abc123", "ud-q6_k_l")
+
+ import routes.inference as _routes
+
+ _get_llama_cpp_backend_orig = _routes.get_llama_cpp_backend
+ _routes.get_llama_cpp_backend = lambda: FakeBackend()
+ try:
+ result = _resolve_inherited_extra_args(
+ FakeRequest(), FakeConfig(), "unsloth/qwen3.8-27b-gguf", None
+ )
+ assert result == [
+ "--rope-scaling",
+ "yarn",
+ "-c",
+ "300000",
+ ], f"Expected inherited extras for case-variant match, got: {result}"
+ finally:
+ _routes.get_llama_cpp_backend = _get_llama_cpp_backend_orig
+
+
+def test_inheritance_refuses_different_repo():
+ """Different repo ids must not inherit extras."""
+ from routes.inference import _resolve_inherited_extra_args
+
+ class FakeRequest:
+ model_path = "different/repo"
+ gguf_variant = None
+ llama_extra_args = None
+ model_fields_set = set()
+
+ class FakeConfig:
+ is_gguf = True
+ gguf_variant = None
+ identifier = "different/repo"
+
+ class FakeBackend:
+ extra_args = ["--rope-scaling", "yarn"]
+ extra_args_source = ("/models--unsloth--Qwen3.8-27B-GGUF/snapshots/abc123", "ud-q6_k_l")
+
+ import routes.inference as _routes
+
+ _get_llama_cpp_backend_orig = _routes.get_llama_cpp_backend
+ _routes.get_llama_cpp_backend = lambda: FakeBackend()
+ try:
+ result = _resolve_inherited_extra_args(FakeRequest(), FakeConfig(), "different/repo", None)
+ assert result == [], f"Expected empty extras for cross-model, got: {result}"
+ finally:
+ _routes.get_llama_cpp_backend = _get_llama_cpp_backend_orig
+
+
+def test_inheritance_refuses_different_variant():
+ """Different variants must not inherit extras."""
+ from routes.inference import _resolve_inherited_extra_args
+
+ class FakeRequest:
+ model_path = "unsloth/Qwen3.8-27B-GGUF"
+ gguf_variant = "Q4_K_M"
+ llama_extra_args = None
+ model_fields_set = set()
+
+ class FakeConfig:
+ is_gguf = True
+ gguf_variant = "Q4_K_M"
+ identifier = "unsloth/Qwen3.8-27B-GGUF"
+
+ class FakeBackend:
+ extra_args = ["--rope-scaling", "yarn"]
+ extra_args_source = ("/models--unsloth--Qwen3.8-27B-GGUF/snapshots/abc123", "q6_k_l")
+
+ import routes.inference as _routes
+
+ _get_llama_cpp_backend_orig = _routes.get_llama_cpp_backend
+ _routes.get_llama_cpp_backend = lambda: FakeBackend()
+ try:
+ result = _resolve_inherited_extra_args(
+ FakeRequest(), FakeConfig(), "unsloth/Qwen3.8-27B-GGUF", None
+ )
+ assert result == [], f"Expected empty extras for variant mismatch, got: {result}"
+ finally:
+ _routes.get_llama_cpp_backend = _get_llama_cpp_backend_orig
+
+
+def test_explicit_empty_extras_still_clears():
+ """An explicit [] still clears extras regardless of identity match."""
+ from routes.inference import _resolve_inherited_extra_args
+
+ class FakeRequest:
+ model_path = "unsloth/Qwen3.8-27B-GGUF"
+ gguf_variant = None
+ llama_extra_args = []
+ model_fields_set = set()
+
+ class FakeConfig:
+ is_gguf = True
+ gguf_variant = None
+ identifier = "unsloth/Qwen3.8-27B-GGUF"
+
+ class FakeBackend:
+ extra_args = ["--rope-scaling", "yarn"]
+ extra_args_source = ("unsloth/Qwen3.8-27B-GGUF", "ud-q6_k_l")
+
+ import routes.inference as _routes
+
+ _get_llama_cpp_backend_orig = _routes.get_llama_cpp_backend
+ _routes.get_llama_cpp_backend = lambda: FakeBackend()
+ try:
+ result = _resolve_inherited_extra_args(
+ FakeRequest(), FakeConfig(), "unsloth/Qwen3.8-27B-GGUF", None
+ )
+ assert result is None # early return because llama_extra_args is not None (it's [])
+ finally:
+ _routes.get_llama_cpp_backend = _get_llama_cpp_backend_orig
+
+
+def test_missing_extra_args_source_refuses_without_raising():
+ """Stored extras with no recorded source are not inherited, and the refusal log
+ must not index into the missing source."""
+ from routes.inference import _resolve_inherited_extra_args
+ import routes.inference as _routes
+
+ class FakeRequest:
+ model_path = "unsloth/Qwen3.8-27B-GGUF"
+ gguf_variant = None
+ llama_extra_args = None
+ model_fields_set = set()
+
+ class FakeConfig:
+ is_gguf = True
+ gguf_variant = "ud-q6_k_l"
+ identifier = "unsloth/Qwen3.8-27B-GGUF"
+
+ original = _routes.get_llama_cpp_backend
+ try:
+ for source in (None, (None, "ud-q6_k_l"), ("", None)):
+
+ class FakeBackend:
+ extra_args = ["--rope-scaling", "yarn"]
+ extra_args_source = source
+
+ _routes.get_llama_cpp_backend = lambda: FakeBackend()
+ result = _resolve_inherited_extra_args(
+ FakeRequest(), FakeConfig(), "unsloth/Qwen3.8-27B-GGUF", None
+ )
+ assert result == [], (source, result)
+ finally:
+ _routes.get_llama_cpp_backend = original
diff --git a/studio/backend/tests/test_openai_auto_switch.py b/studio/backend/tests/test_openai_auto_switch.py
index 378af06c27e..6e9eaa95c6b 100644
--- a/studio/backend/tests/test_openai_auto_switch.py
+++ b/studio/backend/tests/test_openai_auto_switch.py
@@ -7192,6 +7192,76 @@ def test_a_matching_explicit_ctx_flag_survives_auto_switch(monkeypatch):
assert request.llama_extra_args == ["--ctx-size", "100352"]
+def test_a_ctx_flag_saved_from_the_picker_reaches_an_api_load(monkeypatch):
+ """#11511: the picker saves its slider context beside a typed -c. Its own load runs at the
+ -c (llama.cpp takes the last one), so an API auto-switch of the same row must too, instead
+ of stripping the flag as a stale shadow of the slider."""
+ _mock_override_store(monkeypatch)
+ saved = _put(
+ "unsloth/B-GGUF:Q4_K_M",
+ llama_extra_args = ["-c", "300000", "--rope-scaling", "yarn"],
+ custom_context_length = 262144,
+ )
+ entry = saved.overrides["unsloth/B-GGUF:Q4_K_M"]
+ assert entry["custom_context_length"] == 300000
+ assert entry["llama_extra_args"] == ["-c", "300000", "--rope-scaling", "yarn"]
+
+ backend, rec = _wired(
+ monkeypatch, _FakeBackend(None), ("unsloth/B-GGUF", "Q4_K_M", "unsloth/B-GGUF")
+ )
+ _run_hook("unsloth/B-GGUF")
+ request = rec.calls[0]
+ assert request.max_seq_length == 300000
+ assert request.llama_extra_args == ["-c", "300000", "--rope-scaling", "yarn"]
+
+
+ "extra_args, fields, expected",
+ [
+ # llama.cpp's last -c wins, in either spelling.
+ (
+ ["--ctx-size", "8192", "-c", "65536"],
+ {"max_seq_length": 4096, "custom_context_length": 4096},
+ {"max_seq_length": 65536, "custom_context_length": 65536},
+ ),
+ # -c 0 asks llama.cpp for the model's own context: nothing to record.
+ (["-c", "0"], {"custom_context_length": 4096}, {"custom_context_length": 4096}),
+ # No context field sent: none is invented, the flag stays the only control.
+ (["-c", "65536"], {"kv_cache_dtype": "q8_0"}, {}),
+ # No -c: the slider value is stored as sent.
+ (["--top-k", "40"], {"custom_context_length": 4096}, {"custom_context_length": 4096}),
+ # Past the stored ceiling the slider value stays, so the flag is still checked on load.
+ (["-c", "99999999"], {"custom_context_length": 4096}, {"custom_context_length": 4096}),
+ ],
+)
+def test_a_saved_ctx_flag_sets_only_the_context_fields_sent(
+ monkeypatch, extra_args, fields, expected
+):
+ _mock_override_store(monkeypatch)
+ saved = _put("unsloth/B-GGUF:Q4_K_M", llama_extra_args = extra_args, **fields)
+ entry = saved.overrides["unsloth/B-GGUF:Q4_K_M"]
+ stored = {
+ key: entry[key] for key in ("max_seq_length", "custom_context_length") if key in entry
+ }
+ assert stored == expected
+
+
+def test_a_fill_keeps_the_sent_context_when_it_does_not_store_the_flag(monkeypatch):
+ """A fill (the localStorage migration) keeps a stored row's flags, so a -c in its payload is
+ not what any load will run with and must not rewrite the context."""
+ _mock_override_store(monkeypatch)
+ _put("unsloth/B-GGUF:Q4_K_M", llama_extra_args = ["--top-k", "7"])
+ saved = _put(
+ "unsloth/B-GGUF:Q4_K_M",
+ llama_extra_args = ["-c", "65536"],
+ custom_context_length = 4096,
+ fill_absent_fields = True,
+ )
+ entry = saved.overrides["unsloth/B-GGUF:Q4_K_M"]
+ assert entry["llama_extra_args"] == ["--top-k", "7"]
+ assert entry["custom_context_length"] == 4096
+
+
@pytest.mark.parametrize(
"stored_max_seq_length",
["100352", "not-a-number", 100352.0, True, [100352], {"v": 100352}],