Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: let unsloth start vibe start with smaller context windows (#15863)

Nilaydb7c2e44e06f
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 104a8e44aed..ee4c0f66d15 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -11,6 +11,7 @@ import functools
import hashlib
import http.client
import json
+import math
import os
import re
import shlex
@@ -5941,8 +5942,12 @@ def _vibe_env(
entry["temperature"] = float(temperature)
window = model.get("context_length") or model.get("max_context_length")
if window:
- # Vibe compacts at 200k tokens by default, past most local windows.
- entry["auto_compact_threshold"] = max(1, int(int(window) * 0.9))
+ window = int(window)
+ entry["max_context_length"] = window
+ # Vibe's 200k default is past most local windows; Vibe 3 needs min(24k, window / 4) free.
+ entry["auto_compact_threshold"] = min(
+ int(window * 0.9), window - min(24_000, math.ceil(window / 4))
+ )
provider = {
"name": _VIBE_PROVIDER,
"api_base": f"{base}/v1",
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index 463dd03f9b6..e087a5ea6da 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -9,6 +9,7 @@ import http.client
import http.server
import io
import json
+import math
import os
import re
import shlex
@@ -8387,7 +8388,8 @@ def test_vibe_env(tmp_path):
"name": MODEL["id"],
"provider": start._VIBE_PROVIDER,
"alias": start._VIBE_MODEL_ALIAS,
- "auto_compact_threshold": int(MODEL["context_length"] * 0.9),
+ "max_context_length": MODEL["context_length"],
+ "auto_compact_threshold": MODEL["context_length"] - 24_000,
}
]
assert env["VIBE_ACTIVE_MODEL"] == start._VIBE_MODEL_ALIAS
@@ -8400,11 +8402,21 @@ def test_vibe_env_carries_temperature_and_odd_model_ids():
_, models = _vibe_settings(start._vibe_env(BASE, model, {"temperature": 0.7}))
assert models[0]["name"] == model["id"]
assert models[0]["temperature"] == 0.7
- assert models[0]["auto_compact_threshold"] == int(32768 * 0.9)
+ assert models[0]["max_context_length"] == 32768
+ assert models[0]["auto_compact_threshold"] == 32768 * 3 // 4
_, models = _vibe_settings(start._vibe_env(BASE, {"id": "m"}))
assert set(models[0]) == {"name", "provider", "alias"}
[email protected]("window", [16384, 40960, 131072, 262144])
+def test_vibe_env_compaction_threshold_fits_the_window(window):
+ _, models = _vibe_settings(start._vibe_env(BASE, {"id": "m", "context_length": window}))
+ assert models[0]["max_context_length"] == window
+ low = max(min(50_000, window // 2), 10_000)
+ high = window - min(24_000, math.ceil(window / 4))
+ assert low <= models[0]["auto_compact_threshold"] <= high
+
+
def test_connect_vibe_no_launch_uses_env_only(fake_studio, tmp_path):
result = CliRunner().invoke(start.start_app, ["vibe", "--no-launch", "--yolo"])
assert result.exit_code == 0, result.output
@@ -8425,7 +8437,8 @@ def test_vibe_launch_keeps_user_home(fake_studio, monkeypatch):
@pytest.mark.skipif(shutil.which("vibe") is None, reason = "needs the mistral-vibe CLI")
-def test_vibe_real_cli_reaches_unsloth(monkeypatch, tmp_path):
[email protected]("window", [40960, MODEL["context_length"]])
+def test_vibe_real_cli_reaches_unsloth(monkeypatch, tmp_path, window):
requests = []
class Handler(http.server.BaseHTTPRequestHandler):
@@ -8435,18 +8448,28 @@ def test_vibe_real_cli_reaches_unsloth(monkeypatch, tmp_path):
def do_POST(self):
body = json.loads(self.rfile.read(int(self.headers["Content-Length"])))
requests.append((self.headers.get("Authorization"), body))
- chunk = {
- "choices": [{"index": 0, "delta": {"content": "pong"}, "finish_reason": "stop"}]
- }
+ chunks = [
+ {"choices": [{"index": 0, "delta": {"content": "pong"}, "finish_reason": "stop"}]},
+ # Vibe 3 refuses a stream without usage; Studio sends it for include_usage.
+ {
+ "choices": [],
+ "usage": {"prompt_tokens": 5, "completion_tokens": 1, "total_tokens": 6},
+ },
+ ]
self.send_response(200)
self.send_header("Content-Type", "text/event-stream")
self.end_headers()
- self.wfile.write(f"data: {json.dumps(chunk)}\n\ndata: [DONE]\n\n".encode())
+ self.wfile.write("".join(f"data: {json.dumps(c)}\n\n" for c in chunks).encode())
+ self.wfile.write(b"data: [DONE]\n\n")
server = http.server.ThreadingHTTPServer(("127.0.0.1", 0), Handler)
threading.Thread(target = server.serve_forever, daemon = True).start()
base = f"http://127.0.0.1:{server.server_address[1]}"
- monkeypatch.setattr(start, "_connect", lambda *args, **kwargs: (base, "sk-unsloth-e2e", MODEL))
+ monkeypatch.setattr(
+ start,
+ "_connect",
+ lambda *args, **kwargs: (base, "sk-unsloth-e2e", {**MODEL, "context_length": window}),
+ )
monkeypatch.setenv("VIBE_HOME", str(tmp_path / "vibe"))
monkeypatch.chdir(tmp_path)
try: