""")
)
- except Exception:
- # Fallback: Colab's built-in helper.
+ return True
+ except Exception as e:
+ logger.info(f"HTML iframe embed failed ({e}).")
+ return False
+
+
+def _show_and_embed(
+ port: int,
+ *,
+ cloudflare_url: "str | None" = None,
+ colab_login: "tuple[str, str] | None" = None,
+ cloudflare_requested: bool = False,
+):
+ """Render the Unsloth ready card + iframe for *port*.
+
+ Prefer Colab's ``serve_kernel_port_as_iframe`` on real Colab; raw HTML iframe is the
+ fallback. Cloudflare cards stay clickable.
+ """
+ url = get_colab_url(port)
+ logger.info(f"🌐 Unsloth Studio URL: {url}")
+ if cloudflare_url:
+ logger.info(f"🔗 Shareable Cloudflare link: {cloudflare_url}")
+
+ _warn_colab_cloudflare_missing(
+ use_cloudflare = cloudflare_requested,
+ cloudflare_url = cloudflare_url,
+ )
+
+ if cloudflare_url:
try:
- from google.colab import output as colab_output
- colab_output.serve_kernel_port_as_iframe(port, height = 900, width = "100%")
- except ImportError:
- pass
+ from IPython.display import HTML, display
+ display(HTML(_shareable_link_html(cloudflare_url)))
+ except Exception as e:
+ logger.info(f"Could not render Cloudflare link card ({e}).")
+
+ if colab_login:
+ try:
+ _show_colab_login_credentials(*colab_login)
+ except Exception as e:
+ logger.info(f"Could not render Colab login card ({e}).")
+
+ try:
+ show_link(
+ port,
+ _url = url,
+ has_cloudflare_link = bool(cloudflare_url),
+ cloudflare_requested = cloudflare_requested,
+ )
+ except Exception as e:
+ logger.info(f"Could not render Unsloth link card ({e}).")
+
+ # On Colab with a working tunnel, skip the in-cell proxy embed (often blank).
+ if _is_colab_runtime() and cloudflare_url:
+ return
+
+ # Real Colab: kernel helper needs only the port (works when eval_js failed).
+ if _is_colab_runtime():
+ if _embed_kernel_port_iframe(port):
+ return
+ _embed_html_iframe(url, port)
-def start(port: int = 8888, *, cloudflare: bool = False):
+def start(port: int = 8888, *, cloudflare: "bool | None" = None):
"""Start Unsloth Studio in Colab and display the URL.
Args:
port: Port to bind/serve on.
- cloudflare: Opt in to a shareable Cloudflare HTTPS link reachable from any
- device (default OFF). It exposes Unsloth's login page beyond Colab, so it
- stays an explicit opt-in; the default shows only the in-tab proxy iframe.
+ cloudflare: Shareable Cloudflare HTTPS link. ``None`` (default) auto-enables on
+ real Colab because the in-cell proxy embed is often blank; pass ``False`` to
+ skip the tunnel or ``True`` to force it on other runtimes.
Usage:
- start() # Colab-proxy iframe only (default)
- start(cloudflare=True) # also open a shareable Cloudflare link
+ start() # Cloudflare link on Colab (auto); proxy iframe elsewhere
+ start(cloudflare=False) # Colab proxy iframe only (often blank on current Colab)
+ start(cloudflare=True) # force Cloudflare link on any runtime
"""
import time
logger.info("🦥 Starting Unsloth Studio...")
+ use_cloudflare = _colab_wants_cloudflare(cloudflare)
- # Fast path: Unsloth already running (cell re-run). Re-launching would collide on
- # the port, so just re-show the link and iframe.
+ # Fast path: already running (cell re-run); re-show link/iframe instead of rebinding the port.
if _is_studio_healthy(port):
logger.info(f" Unsloth is already running on port {port} — reusing existing server.")
# try/finally: tear the tunnel down even if interrupted mid-start/render.
try:
- cf_url = start_cloudflare_tunnel(port) if cloudflare else None
+ colab_login = _finalize_colab_admin_password() if use_cloudflare else None
+ cf_url = start_cloudflare_tunnel(port) if use_cloudflare else None
_publish_cloudflare_url(cf_url)
- _show_and_embed(port, cloudflare_url = cf_url)
+ _show_and_embed(
+ port,
+ cloudflare_url = cf_url,
+ colab_login = colab_login,
+ cloudflare_requested = use_cloudflare,
+ )
for _ in range(10000):
time.sleep(300)
print("=", end = "", flush = True)
@@ -313,7 +634,6 @@ def start(port: int = 8888, *, cloudflare: bool = False):
logger.info(" Loading backend...")
from run import run_server
- # Auto-detect frontend path
repo_root = Path(__file__).parent.parent
frontend_path = repo_root / "frontend" / "dist"
@@ -323,8 +643,7 @@ def start(port: int = 8888, *, cloudflare: bool = False):
logger.info(" Starting server...")
try:
- # cloudflare=False: this helper owns the tunnel (Colab's own
- # start(cloudflare=...) drives it), so pin it off explicitly.
+ # cloudflare=False: this helper owns the tunnel (via start(cloudflare=...)), so pin it off.
app = run_server(
host = "0.0.0.0",
port = port,
@@ -339,14 +658,12 @@ def start(port: int = 8888, *, cloudflare: bool = False):
logger.error(f"❌ Unsloth Studio failed to start: {exc}")
return
- # run_server auto-increments the port if in use; read back the bound port so the
- # proxy URL and iframe point at the right place.
+ # run_server may auto-increment the port; read back the bound port for the proxy URL/iframe.
actual_port: int = getattr(getattr(app, "state", None), "server_port", None) or port
logger.info(f" Server started on port {actual_port}!")
- # Poll health endpoint before showing the link — avoids the race where ready_event
- # fires but the process hasn't finished binding.
+ # Poll health before showing the link: avoids the race where ready_event fires pre-bind.
import urllib.request
server_ready = False
@@ -365,12 +682,17 @@ def start(port: int = 8888, *, cloudflare: bool = False):
)
return
- # Open the tunnel now the server is healthy, publish its URL for /api/health, and
- # tear it down on interrupt (try/finally) rather than orphan the process.
+ # Server healthy: finalize Colab auth, open the tunnel, publish URL, tear down on interrupt.
try:
- cf_url = start_cloudflare_tunnel(actual_port) if cloudflare else None
+ colab_login = _finalize_colab_admin_password() if use_cloudflare else None
+ cf_url = start_cloudflare_tunnel(actual_port) if use_cloudflare else None
_publish_cloudflare_url(cf_url)
- _show_and_embed(actual_port, cloudflare_url = cf_url)
+ _show_and_embed(
+ actual_port,
+ cloudflare_url = cf_url,
+ colab_login = colab_login,
+ cloudflare_requested = use_cloudflare,
+ )
# Keep kernel alive so the daemon server thread runs.
for _ in range(10000):
diff --git a/studio/backend/core/export/export.py b/studio/backend/core/export/export.py
index c8be50b08b..1b24c46e65 100644
--- a/studio/backend/core/export/export.py
+++ b/studio/backend/core/export/export.py
@@ -1048,6 +1048,21 @@ class ExportBackend:
"Use the safetensors adapter instead.",
None,
)
+ # llama.cpp's convert_lora_to_gguf.py has no concept of DoRA's
+ # lora_magnitude_vector tensors: it only reads the standard
+ # lora_A/lora_B delta, so exporting a DoRA adapter would silently
+ # drop the magnitude rescaling and produce a GGUF LoRA file that
+ # loads fine but no longer matches the trained model.
+ _peft_config = getattr(self.current_model, "peft_config", {}).get("default")
+ if getattr(_peft_config, "use_dora", False):
+ return (
+ False,
+ "GGUF LoRA export is not supported for DoRA adapters: the GGUF LoRA "
+ "format has no way to represent DoRA's magnitude vectors, so the "
+ "exported file would silently lose the DoRA behavior. Use the "
+ "safetensors adapter instead, or merge to a full GGUF model.",
+ None,
+ )
outtype = str(gguf_outtype).lower()
if outtype not in _GGUF_LORA_OUTTYPES:
return (
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 9648fb0218..c3599beb7e 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -307,6 +307,26 @@ def _native_linux_system_rocm_lib_dirs(binary_dir: str = "") -> "list[str]":
_DEFAULT_MAX_TOKENS_FLOOR = 32768
_DEFAULT_FIRST_TOKEN_TIMEOUT_S = 1200.0 # 20 min
+
+def _finalize_reasoning_only_cumulative(
+ cumulative: str, reasoning_text: str, finish_reason: Optional[str], promote_reasoning_only: bool
+) -> str:
+ """Close a live thinking block and promote it only after a clean stop.
+
+ Local inference streams cumulative snapshots. Replacing ``
...`` with
+ bare reasoning at EOF makes the final snapshot shorter, so suffix-based
+ route consumers drop the intended fallback. Keep the snapshot append-only.
+ A length-truncated thought is not a final answer, so close it without
+ promotion and let the client surface the ``length`` terminal state. Raw
+ consumers that do not split reasoning from visible content can disable the
+ fallback to avoid returning the same reasoning twice.
+ """
+ visible_fallback = (
+ reasoning_text if promote_reasoning_only and finish_reason != "length" else ""
+ )
+ return cumulative + "" + visible_fallback
+
+
# Only large streamed tool payloads get an early provisional card; render_html
# is exempt because it needs immediate artifact feedback.
_PROVISIONAL_ARGS_MIN_CHARS = 256
@@ -2781,6 +2801,7 @@ class LlamaCppBackend:
"found": False,
"mtp_token": None,
"supports_mtp": False,
+ "mtp_probe_inconclusive": True,
"ngram_mod_flavor": None,
"supports_ngram_mod": False,
"spec_draft_n_max_flag": None,
@@ -2813,6 +2834,9 @@ class LlamaCppBackend:
supports_no_cache_prompt = False
supports_metrics = False
supports_slot_save = False
+ saw_spec_type = False
+ probe_ok = False
+ help_text = ""
try:
probe_env = cls._llama_server_env_for_binary(bin_path)
result = subprocess.run(
@@ -2824,6 +2848,7 @@ class LlamaCppBackend:
check = False,
env = probe_env,
)
+ probe_ok = result.returncode == 0
help_text = (result.stdout or "") + "\n" + (result.stderr or "")
# Split into per-flag blocks (each --flag line + its indented
# continuation), so the "argument has been removed" description
@@ -2868,17 +2893,19 @@ class LlamaCppBackend:
return False
return "argument has been removed" not in desc
- # MTP token from the --spec-type line.
- spec_line = ""
- for line in help_text.splitlines():
- if "--spec-type" in line:
- spec_line = line
- break
- # PR #22673 used draft-mtp; later renamed to mtp.
- if "draft-mtp" in spec_line:
- mtp_token = "draft-mtp"
- elif re.search(r"[|,\[]mtp[|,\]]", spec_line):
- mtp_token = "mtp"
+ # MTP token from the full --spec-type help block (decl + indented
+ # continuation). First-line-only probing missed builds putting the
+ # enum on the next line (#7302). Prefer draft-mtp (PR #22673) over mtp.
+ spec_help = blocks.get("--spec-type") or ""
+ if not spec_help:
+ # Fallback: join --spec-type lines, avoiding incidental "mtp" in --help.
+ spec_help = "\n".join(
+ line for line in help_text.splitlines() if "--spec-type" in line
+ )
+ mtp_token = cls._mtp_token_from_spec_help(spec_help)
+ # Only a resolved --spec-type block confirms missing MTP; empty/crash
+ # leaves saw_spec_type False so supports_mtp fails open.
+ saw_spec_type = bool(spec_help.strip()) and "--spec-type" in spec_help
# ngram-mod flag flavor. Post-rename builds advertise both new
# args (real) and legacy ones (stubs); pre-rename builds only
@@ -2914,11 +2941,29 @@ class LlamaCppBackend:
supports_slot_save = _is_real("--slot-save-path")
except (OSError, subprocess.SubprocessError) as exc:
logger.debug(f"llama-server --help probe failed: {exc}")
+ saw_spec_type = False
+ probe_ok = False
+ help_text = ""
+
+ help_nonempty = bool(help_text.strip())
+ # Confirmed only when a successful --help lists a --spec-type block with
+ # mtp/draft-mtp; nonempty --help without it is a definitive pre-spec
+ # binary; failed/empty probes stay inconclusive (#7302).
+ if saw_spec_type and probe_ok:
+ supports_mtp = mtp_token is not None
+ mtp_probe_inconclusive = False
+ elif help_nonempty and probe_ok:
+ supports_mtp = False
+ mtp_probe_inconclusive = False
+ else:
+ supports_mtp = False
+ mtp_probe_inconclusive = True
info = {
"found": True,
"mtp_token": mtp_token,
- "supports_mtp": mtp_token is not None,
+ "supports_mtp": supports_mtp,
+ "mtp_probe_inconclusive": mtp_probe_inconclusive,
"ngram_mod_flavor": ngram_mod_flavor,
"supports_ngram_mod": ngram_mod_flavor is not None,
"spec_draft_n_max_flag": spec_draft_n_max_flag,
@@ -2934,6 +2979,21 @@ class LlamaCppBackend:
cls._capability_cache[cache_key] = info
return info
+ @staticmethod
+ def _mtp_token_from_spec_help(spec_help: str) -> Optional[str]:
+ """Extract ``draft-mtp`` / ``mtp`` from a ``--spec-type`` help snippet.
+
+ Prefers ``draft-mtp`` (llama.cpp PR #22673) over the later bare ``mtp``
+ rename. Returns ``None`` when neither token appears as an enum value.
+ """
+ text = spec_help or ""
+ if "draft-mtp" in text:
+ return "draft-mtp"
+ # Bare `mtp` enum token (`|mtp|`, `,mtp,`, ...), not a substring.
+ if re.search(r"(? str:
+ """User-facing error when the text-only --mmproj strip retry also fails.
+
+ Confirmed projector-format mismatches keep the historical wording.
+ Bare signal crashes (common on some ROCm/driver paths) must not be
+ reported as "Vision projector incompatible" — that misled #7302.
+ """
+ if projector_confirmed:
+ return (
+ "Vision projector incompatible with this llama.cpp "
+ "build, and the text-only retry also failed: " + detail
+ )
+ return (
+ "Vision model failed to start (llama-server crashed with "
+ "--mmproj), and the text-only retry also failed: " + detail
+ )
+
@staticmethod
def _output_has_nonprojector_diagnostic(output: str) -> bool:
"""True when the output already names a concrete non-projector cause (out
@@ -8453,23 +8531,29 @@ class LlamaCppBackend:
self._kill_process()
# The #6415 split-axis abort is latched earlier (first spawn).
# Skip if a cancel/unload is pending (mirrors the MTP guard).
+ _projector_msg = self._is_projector_incompatibility(out)
+ _signal_mmproj_guess = self._is_signal_crash(
+ _crash_rc
+ ) and not self._output_has_nonprojector_diagnostic(out)
if (
launched_with_mmproj
and not self._cancel_event.is_set()
- and (
- self._is_projector_incompatibility(out)
- or (
- self._is_signal_crash(_crash_rc)
- and not self._output_has_nonprojector_diagnostic(out)
- )
- )
+ and (_projector_msg or _signal_mmproj_guess)
):
- logger.warning(
- "llama-server could not load this model's vision "
- "projector (--mmproj). The installed llama.cpp build is "
- "likely too old for it. Loading text-only for this "
- "session; run 'unsloth studio update' to enable vision."
- )
+ if _projector_msg:
+ logger.warning(
+ "llama-server could not load this model's vision "
+ "projector (--mmproj). The installed llama.cpp build is "
+ "likely too old for it. Loading text-only for this "
+ "session; run 'unsloth studio update' to enable vision."
+ )
+ else:
+ logger.warning(
+ "llama-server crashed while loading this model's vision "
+ "projector (--mmproj). Retrying text-only for this "
+ "session; if this persists, run 'unsloth studio update' "
+ "or check GPU/driver logs."
+ )
cmd = self._strip_mmproj_args(_last_spawn_cmd)
# This retry bypasses _spawn_and_wait, so refresh the
# launched-argv snapshot itself -- the zero-offload
@@ -8497,14 +8581,16 @@ class LlamaCppBackend:
"(e.g. ROCR_VISIBLE_DEVICES=0 exposes only the first "
"GPU) before launching Unsloth Studio."
)
+ _retry_detail = self._classify_llama_start_failure(
+ "\n".join(self._stdout_lines[-50:]),
+ gguf_path,
+ self._model_identifier,
+ _retry_rc,
+ )
raise RuntimeError(
- "Vision projector incompatible with this llama.cpp "
- "build, and the text-only retry also failed: "
- + self._classify_llama_start_failure(
- "\n".join(self._stdout_lines[-50:]),
- gguf_path,
- self._model_identifier,
- _retry_rc,
+ self._mmproj_retry_failure_message(
+ projector_confirmed = _projector_msg,
+ detail = _retry_detail,
)
)
else:
@@ -8734,18 +8820,29 @@ class LlamaCppBackend:
caps = self.probe_server_capabilities(binary)
mtp_token = caps.get("mtp_token") if caps else None
if not mtp_token:
- logger.warning(
- "Requested MTP speculative decoding but "
- "llama-server lacks --spec-type mtp/draft-mtp; "
- "run `unsloth studio update`. Loading without "
- "speculative decoding."
- )
+ inconclusive = bool(caps.get("mtp_probe_inconclusive")) if caps else True
+ if inconclusive:
+ logger.info(
+ "Requested MTP speculative decoding but llama-server MTP "
+ "capability probe was inconclusive; loading without "
+ "speculative decoding."
+ )
+ else:
+ logger.warning(
+ "Requested MTP speculative decoding but "
+ "llama-server lacks --spec-type mtp/draft-mtp; "
+ "run `unsloth studio update`. Loading without "
+ "speculative decoding."
+ )
# Override an inherited LLAMA_ARG_SPEC_TYPE=draft-mtp (CLI wins
# over env) so the child matches the binary-capability gate and
# the no-MTP budget, like the sibling no-head/non-MTP fallbacks.
flags.append("--spec-default")
self._speculative_type = "default"
- self._spec_fallback_reason = "binary_no_mtp"
+ if inconclusive:
+ self._spec_fallback_reason = None
+ else:
+ self._spec_fallback_reason = "binary_no_mtp"
return False
draft_n_max = _resolved_draft_n_max()
n_max_flag = caps.get("spec_draft_n_max_flag") or "--spec-draft-n-max"
@@ -10544,6 +10641,7 @@ class LlamaCppBackend:
reasoning_effort: Optional[str] = None,
preserve_thinking: Optional[bool] = None,
seed: Optional[int] = None,
+ promote_reasoning_only: bool = True,
_allow_respawn_retry: bool = True,
) -> Generator[Union[str, dict], None, None]:
"""
@@ -10626,7 +10724,12 @@ class LlamaCppBackend:
# model put its whole reply in reasoning
# (e.g. Qwen3 always-think). Show it as
# the main response, not a thinking block.
- cumulative = reasoning_text
+ cumulative = _finalize_reasoning_only_cumulative(
+ cumulative,
+ reasoning_text,
+ _metadata_finish_reason,
+ promote_reasoning_only,
+ )
yield cumulative
_stream_done = True
break # exit inner while
@@ -10723,6 +10826,7 @@ class LlamaCppBackend:
reasoning_effort = reasoning_effort,
preserve_thinking = preserve_thinking,
seed = seed,
+ promote_reasoning_only = promote_reasoning_only,
_allow_respawn_retry = False,
)
return
@@ -10764,6 +10868,7 @@ class LlamaCppBackend:
confirm_tool_calls: bool = False,
bypass_permissions: bool = False,
permission_mode: Optional[str] = None,
+ promote_reasoning_only: bool = True,
) -> Generator[dict, None, None]:
"""
Agentic loop: let the model call tools, execute them, and continue.
@@ -11106,7 +11211,12 @@ class LlamaCppBackend:
),
}
else:
- cumulative_display = reasoning_accum
+ cumulative_display = _finalize_reasoning_only_cumulative(
+ cumulative_display,
+ reasoning_accum,
+ _iter_finish_reason,
+ promote_reasoning_only,
+ )
if not _suppress_visible_output:
yield {
"type": "content",
@@ -11570,7 +11680,12 @@ class LlamaCppBackend:
if _reasoning_started_at is not None and not _reasoning_summary_emitted:
_reasoning_summary_emitted = True
yield _reasoning_summary_event(_reasoning_started_at)
- cumulative_display = reasoning_accum
+ cumulative_display = _finalize_reasoning_only_cumulative(
+ cumulative_display,
+ reasoning_accum,
+ _iter_finish_reason,
+ promote_reasoning_only,
+ )
if not _suppress_visible_output:
yield {
"type": "content",
@@ -12134,7 +12249,12 @@ class LlamaCppBackend:
"text": _strip_tool_markup(cumulative, final = True),
}
else:
- cumulative = reasoning_text
+ cumulative = _finalize_reasoning_only_cumulative(
+ cumulative,
+ reasoning_text,
+ _metadata_finish_reason,
+ promote_reasoning_only,
+ )
yield {"type": "content", "text": cumulative}
_stream_done = True
break # exit inner while
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 53cba522ed..b858fe6f17 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -891,6 +891,7 @@ class UnslothTrainer:
use_gradient_checkpointing: str = "unsloth",
use_rslora: bool = False,
use_loftq: bool = False,
+ use_dora: bool = False,
modules_to_save: list = None,
) -> bool:
"""
@@ -993,6 +994,7 @@ class UnslothTrainer:
use_gradient_checkpointing = use_gradient_checkpointing,
random_state = 3407,
use_rslora = use_rslora,
+ use_dora = use_dora,
loftq_config = {"loftq_bits": 4, "loftq_iter": 1} if use_loftq else None,
)
# Audio VLM models support VLM-style layer selection
@@ -1023,6 +1025,7 @@ class UnslothTrainer:
use_gradient_checkpointing = use_gradient_checkpointing,
random_state = 3407,
use_rslora = use_rslora,
+ use_dora = use_dora,
loftq_config = {"loftq_bits": 4, "loftq_iter": 1} if use_loftq else None,
task_type = None,
)
@@ -1042,6 +1045,7 @@ class UnslothTrainer:
use_gradient_checkpointing = use_gradient_checkpointing,
random_state = 3407,
use_rslora = use_rslora,
+ use_dora = use_dora,
loftq_config = {"loftq_bits": 4, "loftq_iter": 1} if use_loftq else None,
)
@@ -1067,6 +1071,7 @@ class UnslothTrainer:
use_gradient_checkpointing = use_gradient_checkpointing,
random_state = 3407,
use_rslora = use_rslora,
+ use_dora = use_dora,
loftq_config = {"loftq_bits": 4, "loftq_iter": 1} if use_loftq else None,
modules_to_save = modules_to_save,
)
@@ -1087,6 +1092,7 @@ class UnslothTrainer:
use_gradient_checkpointing = use_gradient_checkpointing,
random_state = 3407,
use_rslora = use_rslora,
+ use_dora = use_dora,
loftq_config = {"loftq_bits": 4, "loftq_iter": 1} if use_loftq else None,
modules_to_save = modules_to_save,
)
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index b87585ffe3..8592dabbfe 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -196,6 +196,7 @@ def _build_training_worker_config(values: dict[str, Any]) -> dict[str, Any]:
"gradient_checkpointing": values.get("gradient_checkpointing", "unsloth"),
"use_rslora": values.get("use_rslora", False),
"use_loftq": values.get("use_loftq", False),
+ "use_dora": values.get("use_dora", False),
"train_on_completions": values.get("train_on_completions", False),
"finetune_vision_layers": values.get("finetune_vision_layers", True),
"finetune_language_layers": values.get("finetune_language_layers", True),
@@ -455,6 +456,7 @@ class _MLXTrainerAdapter:
use_gradient_checkpointing: Union[str, bool] = "unsloth",
use_rslora: bool = False,
use_loftq: bool = False,
+ use_dora: bool = False,
) -> bool:
self._peft_config = {
"use_lora": bool(use_lora),
@@ -465,6 +467,7 @@ class _MLXTrainerAdapter:
"gradient_checkpointing": use_gradient_checkpointing,
"use_rslora": bool(use_rslora),
"use_loftq": bool(use_loftq),
+ "use_dora": bool(use_dora),
"finetune_vision_layers": bool(finetune_vision_layers),
"finetune_language_layers": bool(finetune_language_layers),
"finetune_attention_modules": bool(finetune_attention_modules),
@@ -572,6 +575,7 @@ class _MLXTrainerAdapter:
"gradient_checkpointing": "unsloth",
"use_rslora": False,
"use_loftq": False,
+ "use_dora": False,
"finetune_vision_layers": True,
"finetune_language_layers": True,
"finetune_attention_modules": True,
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 81010fb5df..03327d3320 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1547,6 +1547,10 @@ def _run_mlx_training(event_queue, stop_queue, config):
message = "LoftQ is not supported for MLX training yet."
_send("error", error = message)
raise NotImplementedError(message)
+ if config.get("use_dora"):
+ message = "DoRA is not supported for MLX training yet."
+ _send("error", error = message)
+ raise NotImplementedError(message)
if config.get("is_embedding"):
message = "Embedding model training is not supported for MLX training yet."
_send("error", error = message)
@@ -3186,6 +3190,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
use_gradient_checkpointing = config.get("gradient_checkpointing", "unsloth"),
use_rslora = config.get("use_rslora", False),
use_loftq = config.get("use_loftq", False),
+ use_dora = config.get("use_dora", False),
)
elif use_lora:
_send_status(event_queue, "Configuring LoRA adapters...")
@@ -3202,6 +3207,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
use_gradient_checkpointing = config.get("gradient_checkpointing", "unsloth"),
use_rslora = config.get("use_rslora", False),
use_loftq = config.get("use_loftq", False),
+ use_dora = config.get("use_dora", False),
)
else:
_send_status(event_queue, "Preparing model for full finetuning...")
@@ -3630,6 +3636,7 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
use_gradient_checkpointing = gradient_checkpointing,
random_state = config.get("random_seed", 3407),
use_rslora = config.get("use_rslora", False),
+ use_dora = config.get("use_dora", False),
loftq_config = {"loftq_bits": 4, "loftq_iter": 1}
if config.get("use_loftq")
else None,
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 1f5d02dc54..6490cc5a9e 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -445,7 +445,11 @@ def _run_llama_cpp_startup_probes(app: FastAPI) -> None:
import structlog as _structlog
_log = _structlog.get_logger(__name__)
- if _caps.get("found") and not _caps.get("supports_mtp"):
+ if (
+ _caps.get("found")
+ and not _caps.get("supports_mtp")
+ and not _caps.get("mtp_probe_inconclusive")
+ ):
_msg = (
"llama.cpp prebuilt lacks MTP support "
"(--spec-type mtp/draft-mtp). Run `unsloth studio update`. "
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 2663242187..1758efe515 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -70,7 +70,10 @@ class LoadRequest(BaseModel):
cache_type_kv: Optional[str] = Field(
None,
- description = "KV cache data type for both K and V (e.g. 'f16', 'bf16', 'q8_0', 'q4_1', 'q5_1')",
+ description = (
+ "KV cache data type for both K and V "
+ "(e.g. 'f16', 'bf16', 'q8_0', 'q4_0', 'q4_1', 'q5_0', 'q5_1', 'iq4_nl', 'f32')"
+ ),
)
gpu_ids: Optional[List[int]] = Field(
None,
@@ -442,7 +445,10 @@ class LoadResponse(BaseModel):
)
cache_type_kv: Optional[str] = Field(
None,
- description = "KV cache data type for K and V (e.g. 'f16', 'bf16', 'q8_0')",
+ description = (
+ "KV cache data type for K and V "
+ "(e.g. 'f16', 'bf16', 'q8_0', 'q4_0', 'q4_1', 'q5_0', 'q5_1', 'iq4_nl', 'f32')"
+ ),
)
chat_template: Optional[str] = Field(
None,
@@ -602,7 +608,11 @@ class InferenceStatusResponse(BaseModel):
)
cache_type_kv: Optional[str] = Field(
None,
- description = "KV cache quantization dtype (e.g. 'q8_0'), or None for default",
+ description = (
+ "KV cache quantization dtype "
+ "(e.g. 'f16', 'bf16', 'q8_0', 'q4_0', 'q4_1', 'q5_0', 'q5_1', 'iq4_nl', 'f32'), "
+ "or None for default"
+ ),
)
chat_template: Optional[str] = Field(
None, description = "Model's default chat template (Jinja2 source), if any"
diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py
index 6416b86069..0aca5da72c 100644
--- a/studio/backend/models/training.py
+++ b/studio/backend/models/training.py
@@ -470,6 +470,7 @@ class TrainingStartRequest(BaseModel):
gradient_checkpointing: str = Field("", description = "Gradient checkpointing setting")
use_rslora: bool = Field(False, description = "Use RSLoRA")
use_loftq: bool = Field(False, description = "Use LoftQ")
+ use_dora: bool = Field(False, description = "Use DoRA")
train_on_completions: bool = Field(False, description = "Train on completions only")
# Vision-specific LoRA parameters
@@ -545,6 +546,37 @@ class TrainingStartRequest(BaseModel):
raise ValueError("Either num_epochs or max_steps must be > 0; both cannot be 0.")
return self
+ @model_validator(mode = "after")
+ def _validate_lora_variant_flags(self) -> "TrainingStartRequest":
+ # The frontend only ever sends one of these and never under Full
+ # Finetuning, but a direct API/YAML/CLI caller can bypass that. Nothing
+ # downstream breaks (full finetune ignores them, MLX rejects use_dora/
+ # use_loftq outright), but reject early here for a clear error instead
+ # of a silently-ignored flag.
+ active = [
+ name
+ for name, enabled in (
+ ("use_rslora", self.use_rslora),
+ ("use_loftq", self.use_loftq),
+ ("use_dora", self.use_dora),
+ )
+ if enabled
+ ]
+ if len(active) > 1:
+ raise ValueError(
+ f"Only one LoRA variant may be enabled at a time; got {active}. "
+ "use_rslora, use_loftq, and use_dora are mutually exclusive."
+ )
+ # getattr, not self.training_type: model_construct() (used by tests that
+ # validate a single field in isolation) leaves required fields unset, and
+ # this is a mode="after" validator so it still runs on that partial instance.
+ if getattr(self, "training_type", None) == "Full Finetuning" and active:
+ raise ValueError(
+ f"{active[0]} requires an adapter method (LoRA/QLoRA or "
+ "Continued Pretraining); it has no effect under Full Finetuning."
+ )
+ return self
+
class TrainingJobResponse(BaseModel):
"""Immediate response when training is initiated"""
diff --git a/studio/backend/routes/chat_history.py b/studio/backend/routes/chat_history.py
index 24b6dfb36d..6a0d49b47d 100644
--- a/studio/backend/routes/chat_history.py
+++ b/studio/backend/routes/chat_history.py
@@ -160,11 +160,26 @@ class ChatInferenceSettings(BaseModel):
fastMode: Optional[bool] = None
+class ChatPresetLoadConfig(BaseModel):
+ model_config = ConfigDict(extra = "forbid")
+
+ customContextLength: Optional[int] = Field(default = None, gt = 0)
+ maxSeqLength: Optional[float] = None
+ kvCacheDtype: Optional[str] = None
+ speculativeType: Optional[str] = None
+ specDraftNMax: Optional[int] = Field(default = None, ge = 1, le = 16)
+ tensorParallel: Optional[bool] = None
+ gpuMemoryMode: Optional[Literal["manual"]] = None
+ gpuLayers: Optional[int] = None
+ nCpuMoe: Optional[int] = Field(default = None, ge = 0)
+
+
class ChatPreset(BaseModel):
model_config = ConfigDict(extra = "forbid")
name: str
params: ChatInferenceSettings
+ loadConfig: Optional[ChatPresetLoadConfig] = None
class ChatSettingsPayload(BaseModel):
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 04edfcd9d5..598412ad65 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -1794,7 +1794,16 @@ router = APIRouter()
studio_router = APIRouter()
-_ARTIFACT_PREVIEW_FRAME_ANCESTORS = "'self' tauri://localhost http://tauri.localhost"
+# Packaged desktop runs at tauri://localhost (macOS/Linux) or http://tauri.localhost
+# (Windows WebView2); the web build is same-origin ('self'). The `tauri dev` shell,
+# however, serves the frontend from the Vite dev origin (http://localhost:5173),
+# so the packaged allowlist alone leaves the preview blocked in dev with an
+# "ancestor violates frame-ancestors" error. This shell exposes no server resource
+# (it only renders postMessage'd HTML in a no-same-origin sandbox), so also allowing
+# any localhost/127.0.0.1 dev origin to frame it is safe and unblocks the dev shell.
+_ARTIFACT_PREVIEW_FRAME_ANCESTORS = (
+ "'self' tauri://localhost http://tauri.localhost http://localhost:* http://127.0.0.1:*"
+)
_ARTIFACT_PREVIEW_FRAME_STRICT_CSP = (
"default-src 'none'; "
"script-src 'unsafe-inline'; "
@@ -5843,10 +5852,15 @@ async def get_status(current_subject: str = Depends(get_current_subject)):
try:
_bin = type(llama_backend)._find_llama_server_binary()
_caps = type(llama_backend).probe_server_capabilities(_bin)
- _supports_mtp = bool(_caps.get("supports_mtp", False))
+ # Fail open on inconclusive probes: False means a definitive
+ # "binary lacks MTP" to API consumers.
+ _supports_mtp = bool(
+ _caps.get("supports_mtp", False)
+ or (_caps.get("found", False) and _caps.get("mtp_probe_inconclusive", False))
+ )
except Exception:
_bin = None
- _supports_mtp = True # fail open
+ _supports_mtp = False # no usable binary: MTP genuinely unavailable
try:
from utils.llama_cpp_freshness import check_prebuilt_freshness
_freshness = check_prebuilt_freshness(_bin)
@@ -13358,6 +13372,7 @@ async def anthropic_messages(
disable_parallel_tool_use = _disable_parallel,
bypass_permissions = bool(payload.bypass_permissions),
permission_mode = getattr(payload, "permission_mode", None),
+ promote_reasoning_only = False,
)
if payload.stream:
@@ -13397,6 +13412,7 @@ async def anthropic_messages(
max_tokens = payload.max_tokens,
stop = stop,
cancel_event = cancel_event,
+ promote_reasoning_only = False,
)
if payload.stream:
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 3c0d6ff4ba..ed83a12f48 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -2692,7 +2692,10 @@ async def get_kv_cache_estimate(
repo_id: str = Query(..., description = "HF repo ID or local path"),
quant: str = Query(..., description = "Quantization label (e.g. Q4_K_M)"),
n_ctx: int = Query(..., ge = 1, description = "Context length to size the KV cache for"),
- cache_type_kv: Optional[str] = Query(None, description = "KV cache dtype (e.g. q8_0)"),
+ cache_type_kv: Optional[str] = Query(
+ None,
+ description = "KV cache dtype (e.g. q8_0, q4_0, q5_0, iq4_nl, f32)",
+ ),
current_subject: str = Depends(get_current_subject),
):
"""Estimate KV cache + weight bytes for a downloaded GGUF at n_ctx.
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 5b10652fdd..8be4283415 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -332,6 +332,7 @@ async def start_training(
else "unsloth",
"use_rslora": request.use_rslora,
"use_loftq": request.use_loftq,
+ "use_dora": request.use_dora,
"train_on_completions": request.train_on_completions,
"finetune_vision_layers": request.finetune_vision_layers,
"finetune_language_layers": request.finetune_language_layers,
diff --git a/studio/backend/tests/test_anthropic_messages.py b/studio/backend/tests/test_anthropic_messages.py
index 9ccc3f44dd..621ac9aaca 100644
--- a/studio/backend/tests/test_anthropic_messages.py
+++ b/studio/backend/tests/test_anthropic_messages.py
@@ -68,16 +68,15 @@ def _emitter_client_text(events: list[str]) -> str:
def test_anthropic_emitter_closes_reasoning_only_think_block():
- # A reasoning-only reply streams
X live then shrinks to bare X at EOF.
- # This emitter diffs cumulative snapshots and drops the shrink, so without a
- # closing pass the client text would end on an unclosed . finish()
- # must balance it.
+ # Anthropic asks the GGUF generator not to promote reasoning into a duplicate
+ # visible fallback, so its final cumulative snapshot only balances the block.
emitter = AnthropicStreamEmitter()
events = emitter.start("msg_1", "m")
events += emitter.feed({"type": "content", "text": "The capital"})
events += emitter.feed({"type": "content", "text": "The capital of France is Paris."})
- # The generator's final bare-text shrink (dropped by the cumulative diff).
- events += emitter.feed({"type": "content", "text": "The capital of France is Paris."})
+ events += emitter.feed(
+ {"type": "content", "text": "The capital of France is Paris."}
+ )
events += emitter.finish()
assert _emitter_client_text(events) == "The capital of France is Paris."
@@ -1563,6 +1562,44 @@ class TestAnthropicMessagesToolRouting:
assert entry["context_length"] == 2048
assert monitor.active_count() == 0
+ @pytest.mark.parametrize("stream", [False, True])
+ @pytest.mark.parametrize("with_tools", [False, True])
+ def test_reasoning_only_output_is_not_duplicated(self, monkeypatch, stream, with_tools):
+ reasoning = "The capital of France is Paris."
+
+ def _gen_plain(**kwargs):
+ assert kwargs["promote_reasoning_only"] is False
+ yield f"{reasoning}"
+ yield f"{reasoning}"
+
+ def _gen_tools(**kwargs):
+ assert kwargs["promote_reasoning_only"] is False
+ yield {"type": "content", "text": f"{reasoning}"}
+ yield {"type": "content", "text": f"{reasoning}"}
+
+ _mock_backend(
+ monkeypatch,
+ generate_chat_completion = _gen_plain,
+ generate_chat_completion_with_tools = _gen_tools,
+ )
+ payload_fields = {"stream": stream}
+ if with_tools:
+ payload_fields.update(
+ {
+ "enable_tools": True,
+ "tools": [{"type": "web_search_20250305", "name": "web_search"}],
+ }
+ )
+ payload = _basic_payload(**payload_fields)
+
+ response = _drive(anthropic_messages(payload, request = self._Request(), current_subject = "t"))
+ if stream:
+ body = self._sse_blob(self._consume_response(response))
+ assert body.count(reasoning) == 1
+ else:
+ body = json.loads(response.body)
+ assert body["content"][0]["text"] == f"{reasoning}"
+
def test_tool_use_non_streaming_records_api_monitor_reply(self, monkeypatch):
import routes.inference as inf_mod
diff --git a/studio/backend/tests/test_chat_history_routes.py b/studio/backend/tests/test_chat_history_routes.py
index a60ac700bf..896bf1a6cd 100644
--- a/studio/backend/tests/test_chat_history_routes.py
+++ b/studio/backend/tests/test_chat_history_routes.py
@@ -91,6 +91,28 @@ def test_chat_settings_payload_accepts_fast_mode_presets():
assert dumped["customPresets"][0]["params"]["fastMode"] is True
+def test_chat_settings_payload_accepts_preset_load_config():
+ payload = chat_history.ChatSettingsPayload.model_validate(
+ {
+ "customPresets": [
+ {
+ "name": "GGUF preset",
+ "params": {"temperature": 0.7, "maxTokens": 512},
+ "loadConfig": {
+ "customContextLength": 256,
+ "kvCacheDtype": "q8_0",
+ "tensorParallel": False,
+ },
+ },
+ ],
+ }
+ )
+
+ dumped = payload.model_dump(exclude_unset = True)
+ assert dumped["customPresets"][0]["loadConfig"]["customContextLength"] == 256
+ assert dumped["customPresets"][0]["loadConfig"]["kvCacheDtype"] == "q8_0"
+
+
def test_chat_settings_payload_accepts_nudge_tool_calls():
# extra="forbid" 400s PUT /api/chat/settings on unknown keys, so the
# frontend's persisted nudgeToolCalls needs a payload field (like
diff --git a/studio/backend/tests/test_colab_embed.py b/studio/backend/tests/test_colab_embed.py
new file mode 100644
index 0000000000..dae0c7dae0
--- /dev/null
+++ b/studio/backend/tests/test_colab_embed.py
@@ -0,0 +1,479 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression coverage for Colab iframe embedding (#7344)."""
+
+import sys
+import types
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+import colab
+
+
+def _mock_google_colab_modules(colab_mod):
+ """Mock ``google`` and ``google.colab`` for environments without Google packages."""
+ google_mod = types.ModuleType("google")
+ google_mod.colab = colab_mod
+ return {"google": google_mod, "google.colab": colab_mod}
+
+
+def test_short_colab_url_truncates_proxy_host():
+ url = "https://8888-gpu-a100-s-kkb-usc1f0-9hzedjcxrlu8-f.us-central1-0.prod.colab.dev/"
+ assert colab._short_colab_url(url, 8888) == "https://8888-gpu-..."
+
+
+def test_short_colab_url_falls_back_on_unexpected_shape():
+ assert colab._short_colab_url("https://example.com", 8888) == "https://example.com"
+
+
+def test_is_colab_proxy_url_requires_https_proxy():
+ assert colab._is_colab_proxy_url("https://8888-test.prod.colab.dev/", 8888) is True
+ assert colab._is_colab_proxy_url("http://localhost:8888", 8888) is False
+ assert colab._is_colab_proxy_url("http://127.0.0.1:8888", 8888) is False
+
+
+def test_ready_card_html_does_not_open_colab_proxy_in_new_tab():
+ """Colab proxy hosts 404 as top-level tabs (#7349 reporter); never window.open them."""
+ html = colab._ready_card_html("https://8888-test.prod.colab.dev/", 8888)
+ assert "window.open" not in html
+ assert 'href="https://8888-test.prod.colab.dev/"' not in html
+ assert "start(cloudflare=True)" in html
+
+
+def test_ready_card_html_points_to_cloudflare_when_link_ready(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ html = colab._ready_card_html(
+ "https://8888-test.prod.colab.dev/",
+ 8888,
+ has_cloudflare_link = True,
+ )
+ assert "Cloudflare link above" in html
+
+
+def test_ready_card_html_warns_when_cloudflare_tunnel_missing(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ html = colab._ready_card_html(
+ "https://8888-test.prod.colab.dev/",
+ 8888,
+ cloudflare_requested = True,
+ )
+ assert "Could not open a Cloudflare tunnel" in html
+
+
+def test_warn_colab_cloudflare_missing_logs_on_colab_without_tunnel(monkeypatch):
+ warnings: list[str] = []
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(colab.logger, "warning", lambda msg, **kwargs: warnings.append(msg))
+ colab._warn_colab_cloudflare_missing(use_cloudflare = True, cloudflare_url = None)
+ assert warnings
+ assert "Cloudflare tunnel unavailable" in warnings[0]
+
+
+def test_warn_colab_cloudflare_missing_skips_when_tunnel_ready(monkeypatch, caplog):
+ import logging
+
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ with caplog.at_level(logging.WARNING):
+ colab._warn_colab_cloudflare_missing(
+ use_cloudflare = True,
+ cloudflare_url = "https://share.trycloudflare.com",
+ )
+ assert "Cloudflare tunnel unavailable" not in caplog.text
+
+
+def test_is_colab_runtime_uses_backend_colab_detector(monkeypatch):
+ fake_main = types.ModuleType("main")
+ fake_main._IS_COLAB = True
+ monkeypatch.setitem(sys.modules, "main", fake_main)
+ assert colab._is_colab_runtime() is True
+ fake_main._IS_COLAB = False
+ assert colab._is_colab_runtime() is False
+
+
+def test_ready_card_html_uses_cloudflare_hint_on_colab_runtime_localhost(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ html = colab._ready_card_html("http://localhost:8888", 8888)
+ assert "window.open" not in html
+ assert "start(cloudflare=True)" in html
+
+
+def test_ready_card_html_keeps_open_button_for_localhost_outside_colab(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: False)
+ html = colab._ready_card_html("http://localhost:8888", 8888)
+ assert "window.open" in html
+ assert 'href="http://localhost:8888"' in html
+ assert "Open Unsloth Studio" in html
+
+
+def test_embed_kernel_port_iframe_uses_colab_helper(monkeypatch):
+ colab_output = MagicMock()
+ google_colab = SimpleNamespace(output = colab_output)
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ with patch.dict("sys.modules", _mock_google_colab_modules(google_colab)):
+ assert colab._embed_kernel_port_iframe(8888) is True
+ colab_output.serve_kernel_port_as_iframe.assert_called_once_with(
+ 8888,
+ height = colab._COLAB_IFRAME_HEIGHT,
+ width = "100%",
+ )
+
+
+def test_embed_kernel_port_iframe_returns_false_without_colab():
+ with patch.dict("sys.modules", _mock_google_colab_modules(None)):
+ assert colab._embed_kernel_port_iframe(8888) is False
+
+
+def test_embed_kernel_port_iframe_skips_colabtools_without_runtime(monkeypatch):
+ """colabtools can queue JS without appending an iframe; only trust the helper on Colab."""
+ colab_output = MagicMock()
+ google_colab = SimpleNamespace(output = colab_output)
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: False)
+ with patch.dict("sys.modules", _mock_google_colab_modules(google_colab)):
+ assert colab._embed_kernel_port_iframe(8888) is False
+ colab_output.serve_kernel_port_as_iframe.assert_not_called()
+
+
+def test_show_and_embed_prefers_kernel_port_iframe(monkeypatch):
+ calls: list[str] = []
+
+ monkeypatch.setattr(colab, "get_colab_url", lambda port: f"https://{port}-test.prod.colab.dev/")
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(
+ colab,
+ "show_link",
+ lambda port,
+ *,
+ _url = None,
+ has_cloudflare_link = False,
+ cloudflare_requested = False: calls.append("show_link"),
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_kernel_port_iframe",
+ lambda port: calls.append("kernel_iframe") or True,
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_html_iframe",
+ lambda url, port: calls.append("html_iframe") or True,
+ )
+
+ colab._show_and_embed(8888)
+
+ assert calls == ["show_link", "kernel_iframe"]
+
+
+def test_show_and_embed_falls_back_to_html_iframe(monkeypatch):
+ calls: list[str] = []
+
+ monkeypatch.setattr(colab, "get_colab_url", lambda port: f"https://{port}-test.prod.colab.dev/")
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: False)
+ monkeypatch.setattr(
+ colab,
+ "show_link",
+ lambda port, *, _url = None, has_cloudflare_link = False: None,
+ )
+ monkeypatch.setattr(colab, "_embed_kernel_port_iframe", lambda port: False)
+ monkeypatch.setattr(
+ colab,
+ "_embed_html_iframe",
+ lambda url, port: calls.append((url, port)) or True,
+ )
+
+ colab._show_and_embed(8888)
+
+ assert calls == [("https://8888-test.prod.colab.dev/", 8888)]
+
+
+def test_colab_wants_cloudflare_auto_enables_on_runtime(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ assert colab._colab_wants_cloudflare(None) is True
+ assert colab._colab_wants_cloudflare(True) is True
+ assert colab._colab_wants_cloudflare(False) is False
+
+
+def test_colab_wants_cloudflare_defaults_off_outside_runtime(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: False)
+ assert colab._colab_wants_cloudflare(None) is False
+ assert colab._colab_wants_cloudflare(True) is True
+
+
+def test_finalize_colab_admin_password_skips_outside_runtime(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: False)
+ assert colab._finalize_colab_admin_password() is None
+
+
+def test_finalize_colab_admin_password_clears_bootstrap_gate(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(colab, "_load_colab_login_credentials", lambda: None)
+ stored: list[tuple[str, str]] = []
+ monkeypatch.setattr(
+ colab,
+ "_store_colab_login_credentials",
+ lambda username, password: stored.append((username, password)),
+ )
+
+ storage = SimpleNamespace(
+ DEFAULT_ADMIN_USERNAME = "unsloth",
+ ensure_default_admin = MagicMock(),
+ get_bootstrap_password = MagicMock(return_value = "alpha-beta-gamma"),
+ generate_bootstrap_password = MagicMock(return_value = "alpha-beta-gamma"),
+ requires_password_change = MagicMock(return_value = True),
+ update_password = MagicMock(return_value = True),
+ )
+ auth_pkg = types.ModuleType("auth")
+ auth_pkg.storage = storage
+ with patch.dict("sys.modules", {"auth": auth_pkg, "auth.storage": storage}):
+ result = colab._finalize_colab_admin_password()
+
+ assert result == ("unsloth", "alpha-beta-gamma")
+ storage.ensure_default_admin.assert_called_once()
+ storage.update_password.assert_called_once_with("unsloth", "alpha-beta-gamma")
+ assert stored == [("unsloth", "alpha-beta-gamma")]
+
+
+def test_start_skips_finalize_when_cloudflare_disabled(monkeypatch):
+ import time
+
+ finalize_calls: list[str] = []
+ monkeypatch.setattr(colab, "_is_studio_healthy", lambda port: True)
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(
+ colab,
+ "_finalize_colab_admin_password",
+ lambda: finalize_calls.append("finalize") or ("unsloth", "secret"),
+ )
+ monkeypatch.setattr(
+ colab, "start_cloudflare_tunnel", lambda port: "https://share.trycloudflare.com"
+ )
+ monkeypatch.setattr(colab, "_publish_cloudflare_url", lambda url: None)
+ monkeypatch.setattr(colab, "_show_and_embed", lambda port, **kwargs: None)
+ monkeypatch.setattr(colab, "_stop_cloudflare_tunnel", lambda: None)
+ monkeypatch.setattr(time, "sleep", lambda _: (_ for _ in ()).throw(KeyboardInterrupt))
+
+ colab.start(cloudflare = False)
+
+ assert finalize_calls == []
+
+
+def test_finalize_colab_admin_password_redisplay_on_rerun(monkeypatch):
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(
+ colab,
+ "_load_colab_login_credentials",
+ lambda: ("unsloth", "saved-pass"),
+ )
+ monkeypatch.setattr(colab, "_colab_credentials_still_valid", lambda username, password: True)
+
+ storage = SimpleNamespace(
+ DEFAULT_ADMIN_USERNAME = "unsloth",
+ ensure_default_admin = MagicMock(),
+ get_bootstrap_password = MagicMock(),
+ generate_bootstrap_password = MagicMock(),
+ requires_password_change = MagicMock(return_value = False),
+ update_password = MagicMock(),
+ )
+ auth_pkg = types.ModuleType("auth")
+ auth_pkg.storage = storage
+ with patch.dict("sys.modules", {"auth": auth_pkg, "auth.storage": storage}):
+ result = colab._finalize_colab_admin_password()
+
+ assert result == ("unsloth", "saved-pass")
+ storage.update_password.assert_not_called()
+
+
+def test_finalize_colab_admin_password_drops_stale_cached_credentials(monkeypatch):
+ """After an in-app password change the cached first-run password no longer
+ authenticates, so it must not be redisplayed (#7349 Codex review)."""
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(
+ colab,
+ "_load_colab_login_credentials",
+ lambda: ("unsloth", "stale-pass"),
+ )
+ monkeypatch.setattr(colab, "_colab_credentials_still_valid", lambda username, password: False)
+ cleared: list[bool] = []
+ monkeypatch.setattr(colab, "_clear_colab_login_credentials", lambda: cleared.append(True))
+
+ storage = SimpleNamespace(
+ DEFAULT_ADMIN_USERNAME = "unsloth",
+ ensure_default_admin = MagicMock(),
+ get_bootstrap_password = MagicMock(),
+ generate_bootstrap_password = MagicMock(),
+ requires_password_change = MagicMock(return_value = False),
+ update_password = MagicMock(),
+ )
+ auth_pkg = types.ModuleType("auth")
+ auth_pkg.storage = storage
+ with patch.dict("sys.modules", {"auth": auth_pkg, "auth.storage": storage}):
+ result = colab._finalize_colab_admin_password()
+
+ assert result is None
+ assert cleared == [True]
+ storage.update_password.assert_not_called()
+
+
+def test_colab_credentials_still_valid_matches_stored_hash(monkeypatch):
+ from auth.hashing import hash_password
+
+ salt, pwd_hash = hash_password("right-pass")
+ storage = SimpleNamespace(
+ get_user_and_secret = MagicMock(return_value = (salt, pwd_hash, "jwt", False)),
+ )
+ with patch.dict("sys.modules", {"auth.storage": storage}):
+ assert colab._colab_credentials_still_valid("unsloth", "right-pass") is True
+ assert colab._colab_credentials_still_valid("unsloth", "wrong-pass") is False
+
+
+def test_colab_credentials_still_valid_false_when_user_missing(monkeypatch):
+ storage = SimpleNamespace(get_user_and_secret = MagicMock(return_value = None))
+ with patch.dict("sys.modules", {"auth.storage": storage}):
+ assert colab._colab_credentials_still_valid("unsloth", "any") is False
+
+
+def test_colab_login_html_includes_credentials():
+ html = colab._colab_login_html("unsloth", "alpha-beta-gamma-delta")
+ assert "unsloth" in html
+ assert "alpha-beta-gamma-delta" in html
+
+
+def test_show_and_embed_renders_cloudflare_before_colab_login(monkeypatch):
+ displayed: list[str] = []
+ ipython_display = SimpleNamespace(
+ HTML = lambda html: SimpleNamespace(html = html),
+ display = lambda html: displayed.append(html.html),
+ )
+
+ monkeypatch.setattr(colab, "get_colab_url", lambda port: "https://8888-test.prod.colab.dev/")
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(
+ colab,
+ "show_link",
+ lambda port, *, _url = None, has_cloudflare_link = False, cloudflare_requested = False: None,
+ )
+ monkeypatch.setattr(colab, "_embed_kernel_port_iframe", lambda port: True)
+ with patch.dict("sys.modules", {"IPython.display": ipython_display}):
+ colab._show_and_embed(
+ 8888,
+ cloudflare_url = "https://share.trycloudflare.com",
+ colab_login = ("unsloth", "secret-pass"),
+ )
+
+ assert len(displayed) == 2
+ assert "share.trycloudflare.com" in displayed[0]
+ assert "secret-pass" in displayed[1]
+
+
+def test_show_and_embed_skips_iframe_on_colab_when_cloudflare_ready(monkeypatch):
+ calls: list[str] = []
+
+ monkeypatch.setattr(colab, "get_colab_url", lambda port: f"https://{port}-test.prod.colab.dev/")
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(
+ colab,
+ "show_link",
+ lambda port, *, _url = None, has_cloudflare_link = False, cloudflare_requested = False: None,
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_kernel_port_iframe",
+ lambda port: calls.append("kernel_iframe") or True,
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_html_iframe",
+ lambda url, port: calls.append("html_iframe") or True,
+ )
+
+ colab._show_and_embed(8888, cloudflare_url = "https://share.trycloudflare.com")
+
+ assert calls == []
+
+
+def test_show_and_embed_uses_kernel_helper_on_colab_runtime_despite_localhost(monkeypatch):
+ calls: list[str] = []
+
+ monkeypatch.setattr(colab, "get_colab_url", lambda port: f"http://localhost:{port}")
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+
+ monkeypatch.setattr(
+ colab,
+ "show_link",
+ lambda port,
+ *,
+ _url = None,
+ has_cloudflare_link = False,
+ cloudflare_requested = False: calls.append("show_link"),
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_kernel_port_iframe",
+ lambda port: calls.append("kernel_iframe") or True,
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_html_iframe",
+ lambda url, port: calls.append("html_iframe") or True,
+ )
+
+ colab._show_and_embed(8888)
+
+ assert calls == ["show_link", "kernel_iframe"]
+
+
+def test_show_and_embed_skips_kernel_helper_for_localhost_outside_colab(monkeypatch):
+ calls: list[str] = []
+
+ monkeypatch.setattr(colab, "get_colab_url", lambda port: f"http://localhost:{port}")
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: False)
+
+ monkeypatch.setattr(
+ colab,
+ "show_link",
+ lambda port,
+ *,
+ _url = None,
+ has_cloudflare_link = False,
+ cloudflare_requested = False: calls.append("show_link"),
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_kernel_port_iframe",
+ lambda port: calls.append("kernel_iframe") or True,
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_html_iframe",
+ lambda url, port: calls.append("html_iframe") or True,
+ )
+
+ colab._show_and_embed(8888)
+
+ assert calls == ["show_link", "html_iframe"]
+
+
+def test_show_and_embed_still_embeds_when_show_link_fails(monkeypatch):
+ calls: list[str] = []
+
+ monkeypatch.setattr(colab, "get_colab_url", lambda port: f"https://{port}-test.prod.colab.dev/")
+ monkeypatch.setattr(colab, "_is_colab_runtime", lambda: True)
+ monkeypatch.setattr(
+ colab,
+ "show_link",
+ lambda port, *, _url = None: (_ for _ in ()).throw(RuntimeError("no display")),
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_kernel_port_iframe",
+ lambda port: calls.append("kernel_iframe") or True,
+ )
+ monkeypatch.setattr(
+ colab,
+ "_embed_html_iframe",
+ lambda url, port: calls.append("html_iframe") or True,
+ )
+
+ colab._show_and_embed(8888)
+
+ assert calls == ["kernel_iframe"]
diff --git a/studio/backend/tests/test_file_security.py b/studio/backend/tests/test_file_security.py
index b4c8f5d242..e02c33a0f1 100644
--- a/studio/backend/tests/test_file_security.py
+++ b/studio/backend/tests/test_file_security.py
@@ -165,6 +165,23 @@ def test_skips_local_path():
assert "local" in d.reason
+def test_scans_inactive_hf_cache_snapshot_path(tmp_path):
+ # An inactive HF cache loads by snapshot path; the gate must recover the repo id +
+ # commit from models--org--repo/snapshots/ and scan that exact commit, not exempt
+ # it and not fall back to the default branch (an older commit may hold a dropped pickle).
+ snapshot = tmp_path / "models--evil--repo" / "snapshots" / "deadbeef"
+ snapshot.mkdir(parents = True)
+ status = {
+ "scansDone": True,
+ "filesWithIssues": [{"path": "pytorch_model.bin", "level": "unsafe"}],
+ }
+ with _patch_status(status) as model_info:
+ d = evaluate_file_security(str(snapshot))
+ assert d.blocked is True
+ assert model_info.call_args.args[0] == "evil/repo"
+ assert model_info.call_args.kwargs["revision"] == "deadbeef"
+
+
def test_remote_gguf_named_repo_is_still_scanned():
# Only LOCAL paths skip the Hub scan, so a remote .gguf repo is still scanned and a
# poisoned pickle smuggled into it is blocked.
diff --git a/studio/backend/tests/test_gguf_load_cache_reuse.py b/studio/backend/tests/test_gguf_load_cache_reuse.py
index 6d1fac980b..dfd6fc6034 100644
--- a/studio/backend/tests/test_gguf_load_cache_reuse.py
+++ b/studio/backend/tests/test_gguf_load_cache_reuse.py
@@ -789,9 +789,12 @@ class TestLoadHubDownloadExclusion:
# The gguf_load_in_flight marker must be entered before the hub-download
# guard and the unload so a concurrent load can't race the download
- # manager. The llama_extra_args inheritance that used to sit between the
- # marker and the guard now runs in _guard_chat_load_against_training, ahead
- # of the GGUF branch, so it is no longer a landmark inside this slice.
+ # manager. The llama_extra_args inheritance moved out of the branch into
+ # _resolve_inherited_extra_args, which must still run BEFORE it: the
+ # inherited value (e.g. a carried --no-mmproj) shapes the guard's
+ # require_mmproj. Anchor on the call form so the assertion pins the
+ # endpoint's call site, not the function definition.
+ assert source.index("= _resolve_inherited_extra_args(") < source.index("if config.is_gguf:")
assert (
gguf_branch.index("enter_context(gguf_load_in_flight")
< gguf_branch.index("_hub_download_blocks_gguf_load")
diff --git a/studio/backend/tests/test_llama_cpp_mmproj_fallback.py b/studio/backend/tests/test_llama_cpp_mmproj_fallback.py
index 049058e511..4332a440a5 100644
--- a/studio/backend/tests/test_llama_cpp_mmproj_fallback.py
+++ b/studio/backend/tests/test_llama_cpp_mmproj_fallback.py
@@ -335,3 +335,24 @@ class TestRetryContract:
def test_external_kill_skips_flash_attn_retry(self):
# SIGKILL (-9, OOM killer) is not a program fault: no FA-off retry.
assert _signal_crash(-9) is False
+
+
+class TestMmprojRetryFailureMessage:
+ """#7302: bare mmproj crashes must not be reported as projector-format."""
+
+ def test_confirmed_projector_keeps_historical_wording(self):
+ msg = LlamaCppBackend._mmproj_retry_failure_message(
+ projector_confirmed = True,
+ detail = "llama-server failed to start",
+ )
+ assert msg.startswith("Vision projector incompatible with this llama.cpp")
+ assert "llama-server failed to start" in msg
+
+ def test_bare_crash_does_not_claim_projector_incompatibility(self):
+ msg = LlamaCppBackend._mmproj_retry_failure_message(
+ projector_confirmed = False,
+ detail = "llama-server failed to start. Check that the GGUF file is valid",
+ )
+ assert "Vision projector incompatible" not in msg
+ assert "crashed with --mmproj" in msg
+ assert "GGUF file is valid" in msg
diff --git a/studio/backend/tests/test_llama_cpp_mtp_detection.py b/studio/backend/tests/test_llama_cpp_mtp_detection.py
index 1d15647967..27c1b17a85 100644
--- a/studio/backend/tests/test_llama_cpp_mtp_detection.py
+++ b/studio/backend/tests/test_llama_cpp_mtp_detection.py
@@ -637,7 +637,9 @@ def test_probe_server_capabilities_uses_binary_library_env(tmp_path, monkeypatch
def fake_run(cmd, **kwargs):
captured["cmd"] = cmd
captured["env"] = kwargs.get("env")
- return _types.SimpleNamespace(stdout = "--spec-type none,mtp,ngram-simple\n", stderr = "")
+ return _types.SimpleNamespace(
+ stdout = "--spec-type none,mtp,ngram-simple\n", stderr = "", returncode = 0
+ )
monkeypatch.setattr("core.inference.llama_cpp.subprocess.run", fake_run)
@@ -678,6 +680,95 @@ def test_probe_server_capabilities_reports_outdated_binary(tmp_path):
assert caps["found"] is True
assert caps["mtp_token"] is None
assert caps["supports_mtp"] is False
+ assert caps["mtp_probe_inconclusive"] is False
+
+
+@_NEEDS_BASH
+def test_probe_server_capabilities_reads_mtp_from_multiline_help(tmp_path):
+ # Enum on the indented line: first-line-only probing falsely reported
+ # "lacks MTP" (#7302).
+ fake = _make_fake_llama_server(
+ tmp_path / "llama-server",
+ "--spec-type TYPE\n"
+ " speculative decoding type\n"
+ " (none,draft-simple,draft-mtp,ngram-mod)\n",
+ )
+ _clear_caps_cache()
+ caps = LlamaCppBackend.probe_server_capabilities(str(fake))
+ assert caps["mtp_token"] == "draft-mtp"
+ assert caps["supports_mtp"] is True
+ assert caps["mtp_probe_inconclusive"] is False
+
+
+@_NEEDS_BASH
+def test_probe_server_capabilities_empty_help_fails_open(tmp_path):
+ # --help prints nothing: must not claim the prebuilt lacks MTP (#7302).
+ fake = tmp_path / "llama-server"
+ fake.write_text("#!/usr/bin/env bash\nexit 0\n")
+ fake.chmod(0o755)
+ _clear_caps_cache()
+ caps = LlamaCppBackend.probe_server_capabilities(str(fake))
+ assert caps["found"] is True
+ assert caps["mtp_token"] is None
+ assert caps["supports_mtp"] is False
+ assert caps["mtp_probe_inconclusive"] is True
+
+
+@_NEEDS_BASH
+def test_probe_server_capabilities_no_spec_type_is_definitive(tmp_path):
+ # Nonempty --help without --spec-type: pre-spec binary, not inconclusive.
+ fake = _make_fake_llama_server(
+ tmp_path / "llama-server",
+ "--gpu-layers N\n GPU layers to offload\n",
+ )
+ _clear_caps_cache()
+ caps = LlamaCppBackend.probe_server_capabilities(str(fake))
+ assert caps["found"] is True
+ assert caps["mtp_token"] is None
+ assert caps["supports_mtp"] is False
+ assert caps["mtp_probe_inconclusive"] is False
+
+
+@_NEEDS_BASH
+def test_probe_server_capabilities_failed_help_with_output_is_inconclusive(tmp_path):
+ fake = tmp_path / "llama-server"
+ fake.write_text(
+ "#!/usr/bin/env bash\n"
+ 'if [ "$1" = "--help" ]; then\n'
+ " echo 'illegal instruction'\n"
+ " exit 1\n"
+ "fi\n"
+ )
+ fake.chmod(0o755)
+ _clear_caps_cache()
+ caps = LlamaCppBackend.probe_server_capabilities(str(fake))
+ assert caps["found"] is True
+ assert caps["supports_mtp"] is False
+ assert caps["mtp_probe_inconclusive"] is True
+
+
+@_NEEDS_BASH
+def test_probe_server_capabilities_crash_on_help_fails_open(tmp_path):
+ fake = tmp_path / "llama-server"
+ fake.write_text("#!/usr/bin/env bash\nkill -SEGV $$\n")
+ fake.chmod(0o755)
+ _clear_caps_cache()
+ caps = LlamaCppBackend.probe_server_capabilities(str(fake))
+ assert caps["found"] is True
+ assert caps["mtp_token"] is None
+ assert caps["supports_mtp"] is False
+ assert caps["mtp_probe_inconclusive"] is True
+
+
+def test_mtp_token_from_spec_help_prefers_draft_mtp():
+ assert (
+ LlamaCppBackend._mtp_token_from_spec_help("--spec-type none,draft-mtp,mtp,ngram-mod")
+ == "draft-mtp"
+ )
+ assert LlamaCppBackend._mtp_token_from_spec_help("--spec-type [none|mtp|ngram-cache]") == "mtp"
+ assert LlamaCppBackend._mtp_token_from_spec_help("--spec-type none,ngram-mod") is None
+ # No incidental substring matches.
+ assert LlamaCppBackend._mtp_token_from_spec_help("prompt cache") is None
def test_probe_server_capabilities_handles_missing_binary():
@@ -685,6 +776,7 @@ def test_probe_server_capabilities_handles_missing_binary():
caps = LlamaCppBackend.probe_server_capabilities("/no/such/llama-server")
assert caps["found"] is False
assert caps["supports_mtp"] is False
+ assert caps["mtp_probe_inconclusive"] is True
assert caps["supports_cache_ram"] is False
assert caps["supports_ctx_checkpoints"] is False
assert caps["supports_no_cache_prompt"] is False
@@ -1176,12 +1268,14 @@ def _resolver_backend(
*,
ngram_supported = True,
mtp_token = "draft-mtp",
+ mtp_probe_inconclusive = False,
):
"""Backend with a deterministic probe so the resolver is hermetic."""
fake = {
"found": True,
"mtp_token": mtp_token,
"supports_mtp": bool(mtp_token),
+ "mtp_probe_inconclusive": mtp_probe_inconclusive,
"ngram_mod_flavor": "new" if ngram_supported else None,
"supports_ngram_mod": bool(ngram_supported),
"spec_draft_n_max_flag": "--spec-draft-n-max",
@@ -1879,6 +1973,24 @@ def test_spec_fallback_reason_set_when_binary_lacks_mtp(monkeypatch):
assert backend.spec_fallback_reason == "binary_no_mtp"
+def test_spec_fallback_reason_none_when_mtp_probe_inconclusive(monkeypatch):
+ backend = _resolver_backend(
+ monkeypatch,
+ mtp_token = None,
+ mtp_probe_inconclusive = True,
+ )
+ backend._build_speculative_flags(
+ speculative_type = "mtp",
+ spec_draft_n_max = None,
+ extra_args = None,
+ model_identifier = _MTP_MODEL,
+ model_path = None,
+ gpus = True,
+ binary = "/fake/llama-server",
+ )
+ assert backend.spec_fallback_reason is None
+
+
def test_spec_fallback_reason_none_when_mtp_engages(monkeypatch):
backend = _resolver_backend(monkeypatch)
backend._build_speculative_flags(
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index e99e227d40..cf9fde7118 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -37,6 +37,24 @@ def _done() -> str:
return "data: [DONE]\n"
+def _finish(reason: str) -> str:
+ return (
+ "data: "
+ + json.dumps(
+ {
+ "choices": [
+ {
+ "index": 0,
+ "delta": {},
+ "finish_reason": reason,
+ }
+ ]
+ }
+ )
+ + "\n"
+ )
+
+
def _make_backend(monkeypatch, streams: list[list[str]], payloads: list[dict]):
backend = LlamaCppBackend.__new__(LlamaCppBackend)
backend._process = object()
@@ -299,9 +317,8 @@ def test_reasoning_streams_incrementally_with_tools(monkeypatch):
def test_reasoning_only_reply_matches_no_tool_path_with_tools(monkeypatch):
# A reasoning-only turn (whole answer in reasoning_content, no content, no
# tool) with a tool active streams the reasoning live, then resolves to the
- # bare reasoning text -- identical to the no-tool generate_chat_completion
- # path -- so the non-streaming drain still returns it as `content`, not an
- # empty answer.
+ # same text on the visible channel. The final cumulative snapshot stays
+ # append-only so route suffix extraction cannot drop that fallback.
stream = [
_sse({"reasoning_content": "The capital of France is Paris."}),
_done(),
@@ -321,8 +338,49 @@ def test_reasoning_only_reply_matches_no_tool_path_with_tools(monkeypatch):
content_texts = [e["text"] for e in events if e["type"] == "content"]
# Reasoning streamed live during BUFFERING (the fix).
assert content_texts[0] == "The capital of France is Paris."
- # Resolves to bare reasoning, matching the no-tool sibling.
- assert content_texts[-1] == "The capital of France is Paris."
+ assert content_texts[-1] == (
+ "The capital of France is Paris.The capital of France is Paris."
+ )
+
+
+def _assert_reasoning_only_raw_consumer_gets_one_balanced_think_block(monkeypatch, with_tools):
+ stream = [
+ _sse({"reasoning_content": "The capital of France is Paris."}),
+ _done(),
+ ]
+ backend = _make_backend(monkeypatch, [stream], [])
+
+ if with_tools:
+ items = list(
+ backend.generate_chat_completion_with_tools(
+ messages = [{"role": "user", "content": "capital of France?"}],
+ tools = [{"type": "function", "function": {"name": "web_search"}}],
+ max_tool_iterations = 1,
+ promote_reasoning_only = False,
+ )
+ )
+ cumulatives = [item["text"] for item in items if item.get("type") == "content"]
+ else:
+ items = list(
+ backend.generate_chat_completion(
+ messages = [{"role": "user", "content": "capital of France?"}],
+ promote_reasoning_only = False,
+ )
+ )
+ cumulatives = [item for item in items if isinstance(item, str)]
+
+ assert cumulatives[-1] == "The capital of France is Paris."
+ assert all(
+ current.startswith(previous) for previous, current in zip([""] + cumulatives, cumulatives)
+ )
+
+
+def test_reasoning_only_raw_consumer_without_tools_gets_one_balanced_think_block(monkeypatch):
+ _assert_reasoning_only_raw_consumer_gets_one_balanced_think_block(monkeypatch, False)
+
+
+def test_reasoning_only_raw_consumer_with_tools_gets_one_balanced_think_block(monkeypatch):
+ _assert_reasoning_only_raw_consumer_gets_one_balanced_think_block(monkeypatch, True)
def test_reasoning_before_structured_tool_closes_think_block(monkeypatch):
@@ -392,8 +450,8 @@ def _replay_route_reasoning_extractor(cumulatives: list[str]) -> tuple[str, str]
def test_reasoning_only_route_output_matches_no_tool_path(monkeypatch):
# Parity contract: a reasoning-only reply must reach the client identically
# whether tools are on or off. Both generators stream live then
- # resolve to the bare reasoning text; the route's suffix-diff + extractor
- # must therefore produce the same (visible, reasoning) split for both.
+ # append a balanced close plus visible fallback; the route's suffix-diff +
+ # extractor must therefore produce the same split for both.
stream = [
_sse({"reasoning_content": "The capital"}),
_sse({"reasoning_content": " of France is Paris."}),
@@ -430,10 +488,37 @@ def test_reasoning_only_route_output_matches_no_tool_path(monkeypatch):
no_tool_out = _replay_route_reasoning_extractor(no_tool_cumulatives)
assert tool_out == no_tool_out
# Pin the shared contract so a change to either path shows up here.
- _visible, reasoning = tool_out
+ visible, reasoning = tool_out
+ assert visible == "The capital of France is Paris."
assert reasoning == "The capital of France is Paris."
+def test_length_truncated_reasoning_stays_append_only_without_visible_promotion(monkeypatch):
+ stream = [
+ _sse({"reasoning_content": "The proof begins by assuming finitely many primes."}),
+ _finish("length"),
+ _done(),
+ ]
+ backend = _make_backend(monkeypatch, [stream], [])
+
+ items = list(
+ backend.generate_chat_completion(
+ messages = [{"role": "user", "content": "Prove infinitely many primes"}],
+ max_tokens = 16,
+ )
+ )
+ cumulatives = [item for item in items if isinstance(item, str)]
+
+ assert all(
+ current.startswith(previous) for previous, current in zip([""] + cumulatives, cumulatives)
+ )
+ assert cumulatives[-1] == ("The proof begins by assuming finitely many primes.")
+ visible, reasoning = _replay_route_reasoning_extractor(cumulatives)
+ assert visible == ""
+ assert reasoning == "The proof begins by assuming finitely many primes."
+ assert items[-1]["finish_reason"] == "length"
+
+
def test_reasoning_before_bare_json_tool_closes_think_block(monkeypatch):
# _drain_silently sibling of the structured-tool close: a bare-JSON tool call
# with a live reasoning prefix must also close before draining, and
diff --git a/studio/backend/tests/test_training_config_popover_source.py b/studio/backend/tests/test_training_config_popover_source.py
index 4263b012eb..452a3a1ea8 100644
--- a/studio/backend/tests/test_training_config_popover_source.py
+++ b/studio/backend/tests/test_training_config_popover_source.py
@@ -105,5 +105,6 @@ def test_shared_mapper_matches_backend_config_keys():
"lora_dropout",
"use_rslora",
"use_loftq",
+ "use_dora",
):
assert key in src, f"run-config mapper lost backend key {key}"
diff --git a/studio/backend/utils/security/file_security.py b/studio/backend/utils/security/file_security.py
index 892f7862a9..3e12c15096 100644
--- a/studio/backend/utils/security/file_security.py
+++ b/studio/backend/utils/security/file_security.py
@@ -114,6 +114,28 @@ def _file_suffix(path: str) -> str:
return "." + base.rsplit(".", 1)[1].lower() if "." in base else ""
+def _hf_cache_snapshot_ref(local_path: str) -> Optional[tuple]:
+ """``(repo_id, revision)`` for an HF-cache snapshot path, else None. An inactive Studio
+ cache loads by its snapshot path but keeps the ``models--org--repo/snapshots/``
+ layout, so the gate recovers its provenance and scans that exact commit instead of
+ exempting it (an older cached commit can hold a pickle since dropped from the branch)."""
+ try:
+ path = Path(local_path).resolve(strict = False)
+ except (OSError, ValueError):
+ return None
+ for parent in path.parents:
+ if parent.name != "snapshots":
+ continue
+ encoded = parent.parent.name
+ if not encoded.startswith("models--"):
+ return None
+ repo_id = encoded.removeprefix("models--").replace("--", "/")
+ if not repo_id:
+ return None
+ return repo_id, path.relative_to(parent).parts[0] # dir under snapshots/
+ return None
+
+
def _load_relative_path(norm: str, load_subdirs) -> str:
"""``norm`` relative to a ``from_pretrained`` load root. Some loads read from a
snapshot SUBDIRECTORY (Spark-TTS / BiCodec load ``/LLM``), where a file
@@ -141,13 +163,14 @@ def _indexed_shard_paths(
model_name: str,
hf_token: Optional[str],
load_subdirs = (),
+ revision: Optional[str] = None,
):
"""Repo-relative weight paths a load could fetch via weight-index files. Returns a
set (empty when the repo ships no index files -- a definitive "nothing sharded"), or
None when the lookup was inconclusive (transient error) so the caller treats a
flagged subdir pickle conservatively. Reads only small JSON indexes, never weights.
Indexes are looked up at the root and each ``load_subdirs`` root, with ``weight_map``
- entries re-prefixed to repo-relative paths.
+ entries re-prefixed to repo-relative paths. ``revision`` scopes to a cached commit.
"""
import json
@@ -166,6 +189,7 @@ def _indexed_shard_paths(
index_path = hf_hub_download(
model_name,
prefix + filename,
+ revision = revision,
token = hf_token or None,
cache_dir = active_hf_hub_cache(),
)
@@ -260,9 +284,14 @@ def _load_scan_target(model_name: str, load_subdirs: tuple) -> tuple:
return model_name, load_subdirs
-def _fetch_security_status(model_name: str, hf_token: Optional[str]):
+def _fetch_security_status(
+ model_name: str,
+ hf_token: Optional[str],
+ revision: Optional[str] = None,
+):
"""``security_repo_status`` (a dict) or None if unavailable. Hub metadata only;
retries once on a transient error, then returns None so the caller fails open.
+ ``revision`` scopes the scan to a specific cached commit (else the default branch).
"""
from huggingface_hub import model_info as hf_model_info
@@ -272,6 +301,7 @@ def _fetch_security_status(model_name: str, hf_token: Optional[str]):
try:
info = hf_model_info(
model_name,
+ revision = revision,
token = token_arg,
securityStatus = True,
timeout = timeout,
@@ -485,12 +515,17 @@ def evaluate_file_security(
# fails open): the Spark-TTS "/LLM" alias is really unsloth/ from LLM/.
model_name, load_subdirs = _load_scan_target(model_name, tuple(load_subdirs))
- # Local paths (including a local .gguf) have no Hub scan. A remote ref is scanned
- # even if named "*.gguf", so a repo cannot dodge the scan via its name.
+ # Local paths have no Hub scan, EXCEPT an HF-cache snapshot whose canonical path
+ # encodes a repo id + commit: scan that exact commit so an inactive-cache load can't
+ # dodge the gate. A remote ref is scanned even if named "*.gguf" (name can't dodge it).
+ snapshot_revision = None
try:
from utils.paths import is_local_path
if is_local_path(model_name):
- return FileSecurityDecision(model_name, False, reason = "local path; no Hub scan")
+ cache_ref = _hf_cache_snapshot_ref(model_name)
+ if cache_ref is None:
+ return FileSecurityDecision(model_name, False, reason = "local path; no Hub scan")
+ model_name, snapshot_revision = cache_ref
except Exception:
# Cannot classify the path -> do not block on that account.
return FileSecurityDecision(model_name, False, reason = "path check failed; not blocked")
@@ -499,7 +534,7 @@ def evaluate_file_security(
if local_only_load:
return _evaluate_local_only(model_name)
- status = _fetch_security_status(model_name, hf_token)
+ status = _fetch_security_status(model_name, hf_token, revision = snapshot_revision)
if not isinstance(status, dict):
return FileSecurityDecision(
model_name, False, reason = "scan unavailable; allowed (fail-open)"
@@ -536,7 +571,9 @@ def evaluate_file_security(
maybe_shard.append({"path": path, "level": level, "norm": norm})
if maybe_shard:
- indexed = _indexed_shard_paths(model_name, hf_token, load_subdirs)
+ indexed = _indexed_shard_paths(
+ model_name, hf_token, load_subdirs, revision = snapshot_revision
+ )
for m in maybe_shard:
# Block if a root index lists this shard, or if the lookup was inconclusive
# (transient error -> stay conservative). A definitive "no index / not listed"
diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index e6c89b9cd7..9232defd70 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -524,8 +524,10 @@ export function AppProvider({ children }: AppProviderProps) {
visibleToasts={2}
expand={true}
closeButton={true}
- // Clear the chat header buttons on the right.
- offset={{ top: 12, right: 64 }}
+ // Clear the chat header buttons on the right. On desktop, also drop
+ // below the ~34px custom window titlebar so toasts don't cover the
+ // minimize / maximize / close controls.
+ offset={{ top: isTauri ? 46 : 12, right: 64 }}
/>
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 2054384e5a..5976b0af96 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -89,6 +89,7 @@ import {
import { resolveLoadMaxSeqLength } from "../presets/preset-policy";
import {
generateAudio,
+ GenerationLengthError,
listCachedGguf,
listCachedModels,
listGgufVariants,
@@ -4100,7 +4101,15 @@ export function createOpenAIStreamAdapter(
);
if (!abortSignal.aborted) {
const msg = err instanceof Error ? err.message : String(err);
- if (err instanceof StreamInterruptedError) {
+ if (err instanceof GenerationLengthError) {
+ toast.error("Response ran out of tokens", {
+ description:
+ "The model used the full Max Tokens budget while thinking " +
+ "and did not produce a final answer. Increase Max Tokens in " +
+ "chat Settings or turn off thinking, then retry.",
+ duration: 8000,
+ });
+ } else if (err instanceof StreamInterruptedError) {
// Connection dropped mid-turn: surface it explicitly (the rethrow
// below also marks the message with an inline error + Retry).
toast.error("Response interrupted", {
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index ffaf099f29..4d123e98ab 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -50,6 +50,21 @@ export class StreamInterruptedError extends Error {
}
}
+/**
+ * Thrown when a reasoning model consumes its output budget before emitting any
+ * standard content. Keeping this distinct from a dropped connection lets the
+ * chat UI explain why a completed stream contains only a thinking panel.
+ */
+export class GenerationLengthError extends Error {
+ constructor() {
+ super(
+ "The model reached the Max Tokens limit before producing a final answer. " +
+ "Increase Max Tokens or disable thinking, then retry.",
+ );
+ this.name = "GenerationLengthError";
+ }
+}
+
export function notifyChatHistoryUpdated(): void {
if (typeof window !== "undefined") {
window.dispatchEvent(new Event(CHAT_HISTORY_UPDATED_EVENT));
@@ -982,6 +997,61 @@ function parseSseEvent(rawEvent: string): string[] {
return dataLines;
}
+function hasNonWhitespaceText(value: unknown): boolean {
+ if (typeof value === "string") {
+ return value.trim().length > 0;
+ }
+ if (Array.isArray(value)) {
+ return value.some((item) => hasNonWhitespaceText(item));
+ }
+ if (!value || typeof value !== "object") {
+ return false;
+ }
+ const record = value as Record;
+ return ["thinking", "text", "content", "reasoning", "summary"].some(
+ (key) => key in record && hasNonWhitespaceText(record[key]),
+ );
+}
+
+function classifyStructuredDeltaContent(content: unknown): {
+ hasAssistantContent: boolean;
+ hasReasoningContent: boolean;
+} {
+ if (typeof content === "string") {
+ return {
+ hasAssistantContent: hasNonWhitespaceText(content),
+ hasReasoningContent: false,
+ };
+ }
+ if (!Array.isArray(content)) {
+ return {
+ hasAssistantContent: false,
+ hasReasoningContent: false,
+ };
+ }
+
+ let hasAssistantContent = false;
+ let hasReasoningContent = false;
+ for (const part of content) {
+ if (typeof part === "string") {
+ hasAssistantContent ||= hasNonWhitespaceText(part);
+ continue;
+ }
+ if (!part || typeof part !== "object") {
+ continue;
+ }
+ const record = part as Record;
+ if (record.type === "thinking" || record.type === "reasoning") {
+ hasReasoningContent ||= hasNonWhitespaceText(record);
+ } else if (record.type === "text" || record.type === "output_text") {
+ const text =
+ typeof record.text === "string" ? record.text : record.content;
+ hasAssistantContent ||= hasNonWhitespaceText(text);
+ }
+ }
+ return { hasAssistantContent, hasReasoningContent };
+}
+
export async function* streamChatCompletions(
payload: OpenAIChatCompletionsRequest,
signal: AbortSignal,
@@ -1009,6 +1079,19 @@ export async function* streamChatCompletions(
// EOF without `[DONE]` or a finish_reason chunk means the stream was cut
// mid-generation: surface as interrupted, not silent success.
let sawTerminalSignal = false;
+ let terminalFinishReason: string | null = null;
+ let sawAssistantContent = false;
+ let sawReasoningContent = false;
+
+ const throwIfReasoningOnlyLength = () => {
+ if (
+ terminalFinishReason === "length" &&
+ sawReasoningContent &&
+ !sawAssistantContent
+ ) {
+ throw new GenerationLengthError();
+ }
+ };
try {
while (true) {
@@ -1018,6 +1101,7 @@ export async function* streamChatCompletions(
if (!sawTerminalSignal) {
throw new StreamInterruptedError();
}
+ throwIfReasoningOnlyLength();
break;
}
@@ -1039,6 +1123,7 @@ export async function* streamChatCompletions(
if (dataText === "[DONE]") {
completed = true;
sawTerminalSignal = true;
+ throwIfReasoningOnlyLength();
return;
}
@@ -1094,11 +1179,31 @@ export async function* streamChatCompletions(
}
// finish_reason is a valid terminal signal for providers that close
// the stream without an explicit [DONE] sentinel.
- const finishReason = (
+ const parsedChoices = (
parsed as {
- choices?: Array<{ finish_reason?: string | null }>;
+ choices?: Array<{
+ delta?: Record;
+ finish_reason?: string | null;
+ }>;
}
- ).choices?.[0]?.finish_reason;
+ ).choices;
+ for (const choice of parsedChoices ?? []) {
+ const delta = choice.delta;
+ if (delta) {
+ const contentState = classifyStructuredDeltaContent(delta.content);
+ sawAssistantContent ||= contentState.hasAssistantContent;
+ sawReasoningContent ||= contentState.hasReasoningContent;
+ const reasoning =
+ delta.reasoning_content ??
+ delta.reasoning ??
+ delta.reasoning_details;
+ sawReasoningContent ||= hasNonWhitespaceText(reasoning);
+ }
+ if (choice.finish_reason) {
+ terminalFinishReason = choice.finish_reason;
+ }
+ }
+ const finishReason = parsedChoices?.[0]?.finish_reason;
if (finishReason) {
sawTerminalSignal = true;
}
diff --git a/studio/frontend/src/features/chat/api/chat-settings-api.ts b/studio/frontend/src/features/chat/api/chat-settings-api.ts
index 1e00357ea4..ab13130525 100644
--- a/studio/frontend/src/features/chat/api/chat-settings-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-settings-api.ts
@@ -13,6 +13,7 @@ export type PersistedInferenceParams = Partial<
export interface PersistedChatPreset {
name: string;
params: PersistedInferenceParams;
+ loadConfig?: Record;
}
export interface PersistedChatSettings {
diff --git a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
index 0f4c83e0db..1955c3aca1 100644
--- a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
+++ b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
@@ -12,6 +12,8 @@ import {
import { MascotImg } from "@/components/mascot-img";
import { Button } from "@/components/ui/button";
import { copyToClipboard } from "@/lib/copy-to-clipboard";
+import { downloadFile, isDownloadCancelled } from "@/lib/native-files";
+import { toast } from "@/lib/toast";
import { cn } from "@/lib/utils";
import { CopyIcon, EyeIcon, Maximize2Icon, XIcon } from "lucide-react";
import { Download01Icon } from "@hugeicons/core-free-icons";
@@ -91,18 +93,6 @@ function ArtifactGeneratingPanel() {
);
}
-function downloadTextFile(filename: string, text: string): void {
- const blob = new Blob([text], { type: "text/html;charset=utf-8" });
- const url = URL.createObjectURL(blob);
- const anchor = document.createElement("a");
- anchor.href = url;
- anchor.download = filename;
- document.body.appendChild(anchor);
- anchor.click();
- document.body.removeChild(anchor);
- window.setTimeout(() => URL.revokeObjectURL(url), 0);
-}
-
export function ArtifactSurface({
artifact,
variant,
@@ -205,7 +195,7 @@ export function ArtifactSurface({
className={cn(
"relative flex min-h-0 flex-col bg-background",
variant === "panel"
- ? "artifact-panel-shell mx-2 mt-[72px] mb-8 h-[calc(100%_-_104px)] overflow-visible rounded-[28px] border-t border-border/70 bg-card/95"
+ ? "artifact-panel-shell mx-2 mt-[90px] mb-8 h-[calc(100%_-_122px)] overflow-visible rounded-[28px] border-t border-border/70 bg-card/95"
: "h-[min(92vh,900px)] w-[min(96vw,1200px)] overflow-hidden rounded-2xl border border-border shadow-xl",
)}
aria-label={`${artifact.title} canvas`}
@@ -265,7 +255,19 @@ export function ArtifactSurface({
size="icon"
className="size-8"
disabled={isLoadingArtifact || !hasArtifactCode}
- onClick={() => downloadTextFile(filename, artifact.code)}
+ onClick={() => {
+ // Route through the native save dialog on desktop; the plain
+ // blob-anchor download is silently dropped by the Tauri WebView2.
+ void downloadFile(
+ artifact.code,
+ filename,
+ "text/html;charset=utf-8",
+ ).catch((err) => {
+ if (!isDownloadCancelled(err)) {
+ toast.error("Failed to save canvas HTML");
+ }
+ });
+ }}
aria-label="Download canvas HTML"
>
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index c241607e28..7452cf3447 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -2289,7 +2289,7 @@ export function ChatPage({
} else if (outcome === "conflict") {
toast.info("Resume this download from Models", {
description:
- "An earlier partial download used a different transport. Open the Models tab to resume or restart it.",
+ "An earlier partial download used a different transport. Open the Model hub tab to resume or restart it.",
});
} else if (outcome === "busy") {
toast.info("Download already in progress", {
@@ -2410,7 +2410,7 @@ export function ChatPage({
// surface's onComplete auto-loads, mirroring the "started" branch.
toast.info("Resume this download from Models", {
description:
- "An earlier partial download used a different transport. Open the Models tab to resume or restart it.",
+ "An earlier partial download used a different transport. Open the Model hub tab to resume or restart it.",
});
return;
}
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index 672ab2b2d3..7b310c50d4 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -75,6 +75,12 @@ import {
isSamePresetConfig,
toPresetParams,
} from "./presets/preset-policy";
+import {
+ applyPresetLoadConfig,
+ capturePresetLoadConfig,
+ formatPresetLoadConfigSummary,
+ isSamePresetLoadConfig,
+} from "./presets/preset-load-config";
import {
type ProviderCapabilities,
getExternalMaxOutputTokens,
@@ -385,6 +391,12 @@ export function ChatSettingsPanel({
(s) => s.ggufMaxContextLength,
);
const customContextLength = useChatRuntimeStore((s) => s.customContextLength);
+ const kvCacheDtype = useChatRuntimeStore((s) => s.kvCacheDtype);
+ const gpuMemoryMode = useChatRuntimeStore((s) => s.gpuMemoryMode);
+ const gpuLayers = useChatRuntimeStore((s) => s.gpuLayers);
+ const nCpuMoe = useChatRuntimeStore((s) => s.nCpuMoe);
+ const tensorParallel = useChatRuntimeStore((s) => s.tensorParallel);
+ const specDraftNMax = useChatRuntimeStore((s) => s.specDraftNMax);
const speculativeType = useChatRuntimeStore((s) => s.speculativeType);
const specFallbackReason = useChatRuntimeStore((s) => s.specFallbackReason);
const mtpUpdatable =
@@ -469,11 +481,50 @@ export function ChatSettingsPanel({
if (activePresetDefinition == null) {
return false;
}
- if (activePresetDefinition.name === "Default") {
- return activePresetSource === "modified";
- }
- return !isSamePresetConfig(activePresetDefinition.params, params);
- }, [activePresetDefinition, activePresetSource, params]);
+ const samplingChanged =
+ activePresetDefinition.name === "Default"
+ ? activePresetSource === "modified"
+ : !isSamePresetConfig(activePresetDefinition.params, params);
+ const currentLoadConfig = capturePresetLoadConfig();
+ const loadChanged = !isSamePresetLoadConfig(
+ activePresetDefinition.loadConfig,
+ currentLoadConfig,
+ );
+ return samplingChanged || loadChanged;
+ }, [
+ activePresetDefinition,
+ activePresetSource,
+ params,
+ customContextLength,
+ ggufContextLength,
+ kvCacheDtype,
+ gpuMemoryMode,
+ gpuLayers,
+ nCpuMoe,
+ tensorParallel,
+ speculativeType,
+ specDraftNMax,
+ params.maxSeqLength,
+ ]);
+ const activePresetLoadSummary = useMemo(
+ () => formatPresetLoadConfigSummary(activePresetDefinition?.loadConfig),
+ [activePresetDefinition],
+ );
+ const currentLoadSummary = useMemo(
+ () => formatPresetLoadConfigSummary(capturePresetLoadConfig()),
+ [
+ customContextLength,
+ ggufContextLength,
+ kvCacheDtype,
+ gpuMemoryMode,
+ gpuLayers,
+ nCpuMoe,
+ tensorParallel,
+ speculativeType,
+ specDraftNMax,
+ params.maxSeqLength,
+ ],
+ );
const presetSaveState = useMemo(
() =>
getPresetSaveState({
@@ -549,8 +600,14 @@ export function ChatSettingsPanel({
onParamsChange({
...applyPresetParams(params, p.params),
});
+ if (p.loadConfig) {
+ applyPresetLoadConfig(p.loadConfig);
+ }
setActivePreset(name);
setActivePresetSource(getPresetSource(name));
+ if (p.loadConfig && params.checkpoint) {
+ toast.info("Reload the model to apply load settings from this preset.");
+ }
}
}
@@ -571,9 +628,14 @@ export function ChatSettingsPanel({
? getBuiltinVariantName(trimmed, usedNames)
: trimmed;
const next = customPresets.filter((p) => p.name !== saveName);
+ const loadConfig = capturePresetLoadConfig();
const merged = [
...next,
- { name: saveName, params: toPresetParams(params) },
+ {
+ name: saveName,
+ params: toPresetParams(params),
+ ...(loadConfig ? { loadConfig } : {}),
+ },
];
setCustomPresets(merged);
setActivePreset(saveName);
@@ -598,8 +660,11 @@ export function ChatSettingsPanel({
if (activePreset === name) {
if (fallbackPreset) {
onParamsChange({
- ...applyPresetParams(params, fallbackPreset.params),
+ ... applyPresetParams(params, fallbackPreset.params),
});
+ if (fallbackPreset.loadConfig) {
+ applyPresetLoadConfig(fallbackPreset.loadConfig);
+ }
setActivePreset(fallbackPreset.name);
setActivePresetSource("builtin-default");
}
@@ -901,6 +966,23 @@ export function ChatSettingsPanel({
Delete
+ Saving a preset also stores current load settings (context length,
+ KV cache dtype, speculative decoding, GPU layers).
+ {currentLoadSummary ? (
+ <>
+ {" "}
+ Active now: {currentLoadSummary}.
+
+ ) : null}
+ {activePresetLoadSummary &&
+ activePresetLoadSummary !== currentLoadSummary ? (
+ <>
+ {" "}
+ Saved in preset: {activePresetLoadSummary}.
+
+ ) : null}
+
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index 082f6d6ede..36eff27334 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -1274,12 +1274,19 @@ export function useChatModelRuntime() {
prog.expected_bytes,
dlSamples,
);
- setLoadProgress({
- percent: pct,
- label: progressLabel,
- phase: "downloading",
- });
- if (loadToastDismissedRef.current) return;
+ // loadProgress state is only read by the dismissed-toast inline
+ // status. Writing it while the toast is visible re-renders the
+ // whole chat page every poll — cheap in Chrome, janky in the
+ // desktop WebView2 (laggy typing). Feed the toast directly and
+ // only touch state when the inline view is actually live.
+ if (loadToastDismissedRef.current) {
+ setLoadProgress({
+ percent: pct,
+ label: progressLabel,
+ phase: "downloading",
+ });
+ return;
+ }
toast(null, {
id: toastId,
...modelLoadToastOptions(
@@ -1301,19 +1308,23 @@ export function useChatModelRuntime() {
const est = estimate(dlSamples, prog.downloaded_bytes, 0);
const rateSuffix =
est.stable ? ` • ${formatRate(est.rate)}` : "";
- setLoadProgress({
- percent: null,
- label: `${dlGb.toFixed(1)} GB downloaded${rateSuffix}`,
- phase: "downloading",
- });
+ // Inline-status-only state; skip the chat-page re-render unless it's shown.
+ if (loadToastDismissedRef.current) {
+ setLoadProgress({
+ percent: null,
+ label: `${dlGb.toFixed(1)} GB downloaded${rateSuffix}`,
+ phase: "downloading",
+ });
+ }
} else if (prog.progress >= 1 && hasShownProgress) {
downloadComplete = true;
- setLoadProgress({
- percent: 100,
- label: "Download complete",
- phase: "starting",
- });
- if (!loadToastDismissedRef.current) {
+ if (loadToastDismissedRef.current) {
+ setLoadProgress({
+ percent: 100,
+ label: "Download complete",
+ phase: "starting",
+ });
+ } else {
toast(null, {
id: toastId,
...modelLoadToastOptions(
@@ -1367,12 +1378,17 @@ export function useChatModelRuntime() {
formatEta(est.eta) !== "--" ? ` • ${formatEta(est.eta)} left` : ""
}`
: base;
- setLoadProgress({
- percent: pct,
- label,
- phase: "starting",
- });
- if (loadToastDismissedRef.current) return;
+ // Inline-status-only state (see pollDownload): while the toast is
+ // up, skip the state write so the chat page doesn't re-render every
+ // poll during "Starting model" — the desktop WebView2 typing-lag fix.
+ if (loadToastDismissedRef.current) {
+ setLoadProgress({
+ percent: pct,
+ label,
+ phase: "starting",
+ });
+ return;
+ }
toast(null, {
id: toastId,
...modelLoadToastOptions(
diff --git a/studio/frontend/src/features/chat/permission-mode-select.tsx b/studio/frontend/src/features/chat/permission-mode-select.tsx
index f3e1594795..2fafeab7d6 100644
--- a/studio/frontend/src/features/chat/permission-mode-select.tsx
+++ b/studio/frontend/src/features/chat/permission-mode-select.tsx
@@ -1,13 +1,8 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-import {
- ChevronDown,
- CircleAlert,
- CircleOff,
- Hand,
- ShieldCheck,
-} from "lucide-react";
+import { ChevronDown, CircleAlert, Hand, ShieldCheck } from "lucide-react";
+import type { ComponentType } from "react";
import { useState } from "react";
import {
@@ -29,6 +24,7 @@ import {
DropdownMenuTrigger,
} from "@/components/ui/dropdown-menu";
import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { SparklesGlyph } from "@/lib/sparkles-icon";
import { Tick02Icon } from "@/lib/tick-icon";
import { cn } from "@/lib/utils";
import { HugeiconsIcon } from "@hugeicons/react";
@@ -45,7 +41,7 @@ export const PERMISSION_MODE_OPTIONS: readonly {
value: PermissionMode;
label: string;
description: string;
- icon: typeof Hand;
+ icon: ComponentType<{ className?: string; strokeWidth?: number }>;
}[] = [
{
value: "ask",
@@ -63,7 +59,7 @@ export const PERMISSION_MODE_OPTIONS: readonly {
value: "off",
label: "Run automatically",
description: "Run tool calls without approval prompts inside the sandbox",
- icon: CircleOff,
+ icon: SparklesGlyph,
},
{
value: "full",
diff --git a/studio/frontend/src/features/chat/presets/preset-load-config.ts b/studio/frontend/src/features/chat/presets/preset-load-config.ts
new file mode 100644
index 0000000000..1083655cf2
--- /dev/null
+++ b/studio/frontend/src/features/chat/presets/preset-load-config.ts
@@ -0,0 +1,244 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+ applyPerModelConfigToRuntime,
+ currentRuntimePerModelConfig,
+ perModelConfigsEqual,
+} from "@/features/model-picker";
+import {
+ CONTEXT_LENGTH_MIN,
+ DEFAULT_PER_MODEL_CONFIG,
+ DEFAULT_MAX_SEQ_LENGTH,
+ KV_CACHE_DTYPES,
+ MTP_SPECULATIVE_TYPES,
+ SPECULATIVE_TYPES,
+ normalizeMaxSeqLength,
+ type PerModelConfig,
+} from "@/features/model-picker/model-config/per-model-config";
+import {
+ GPU_LAYERS_AUTO,
+ useChatRuntimeStore,
+ normalizeSpeculativeType,
+} from "../stores/chat-runtime-store";
+
+/** Load/runtime knobs saved in a chat preset (excludes per-model-only blobs). */
+export type PresetLoadConfig = Pick<
+ PerModelConfig,
+ | "customContextLength"
+ | "maxSeqLength"
+ | "kvCacheDtype"
+ | "speculativeType"
+ | "specDraftNMax"
+ | "tensorParallel"
+ | "gpuMemoryMode"
+ | "gpuLayers"
+ | "nCpuMoe"
+>;
+
+const VALID_KV_CACHE_DTYPES = new Set![\"Open](\"https://colab.research.google.com/assets/colab-badge.svg\"){kind=icon}
![](\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\")
![](\"https://github.com/unslothai/unsloth/raw/main/images/Discord)
![](\"https://github.com/unslothai/unsloth/blob/main/images/documentation%20green%20button.png?raw=true\")
![](\"https://github.com/unslothai/unsloth/raw/main/studio/frontend/public/studio%20github%20landscape%20colab%20display.png\")
![](\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\")
+ Unsloth Studio is Ready!
+