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: use the browser's timezone for today's date in chat (#11852)

Nilay6a169935db65
diff --git a/studio/backend/core/inference/chat_generation_runs.py b/studio/backend/core/inference/chat_generation_runs.py
index 81b0bfe0d53..5d6bdae9b29 100644
--- a/studio/backend/core/inference/chat_generation_runs.py
+++ b/studio/backend/core/inference/chat_generation_runs.py
@@ -66,7 +66,12 @@ class _SSEDecoder:
return values
-def _background_request(app: Any, run_id: str, cancel_event: threading.Event) -> Request:
+def _background_request(
+ app: Any,
+ run_id: str,
+ cancel_event: threading.Event,
+ timezone_headers: dict[str, str] | None = None,
+) -> Request:
scope = {
"type": "http",
"asgi": {"version": "3.0", "spec_version": "2.3"},
@@ -80,6 +85,10 @@ def _background_request(app: Any, run_id: str, cancel_event: threading.Event) ->
(b"x-unsloth-generation-run", run_id.encode("ascii", "ignore")),
# Durable runs replay their event log to the UI, which needs the Unsloth control frames (see routes.inference).
(b"x-unsloth-events", b"1"),
+ *(
+ (name.encode("latin-1"), value.encode("latin-1"))
+ for name, value in (timezone_headers or {}).items()
+ ),
],
"client": ("127.0.0.1", 0),
"server": ("127.0.0.1", 0),
@@ -739,13 +748,15 @@ class ChatGenerationSupervisor:
from routes.inference import produce_openai_chat_completions
- payload = ChatCompletionRequest.model_validate(run["requestPayload"])
+ request_payload = dict(run["requestPayload"])
+ timezone_headers = request_payload.pop(db.TIMEZONE_HEADERS_FIELD, None)
+ payload = ChatCompletionRequest.model_validate(request_payload)
# Switching, idle reload and auto-download all happen in the call below, and llama.cpp's first-token
# budget only starts after it. One touch afterwards cannot cover a preparation longer than the lease
# itself.
response = await produce_openai_chat_completions(
payload,
- _background_request(self.app, run_id, cancel_event),
+ _background_request(self.app, run_id, cancel_event, timezone_headers),
owner,
cancel_on_disconnect = False,
)
diff --git a/studio/backend/routes/chat_generation_runs.py b/studio/backend/routes/chat_generation_runs.py
index d2bdd9a0661..775f74a50b4 100644
--- a/studio/backend/routes/chat_generation_runs.py
+++ b/studio/backend/routes/chat_generation_runs.py
@@ -27,6 +27,10 @@ from core.inference.llama_keepwarm import inference_lifecycle_gate
from models.inference import ChatCompletionRequest
from storage import chat_generation_runs_db as db
from utils.api_errors import safe_validation_errors
+from utils.current_date_prompt_settings import (
+ CURRENT_DATE_TIMEZONE_HEADER,
+ CURRENT_DATE_TIMEZONE_OFFSET_HEADER,
+)
router = APIRouter()
_EVENT_WAIT_EXECUTOR = ThreadPoolExecutor(
@@ -243,6 +247,14 @@ def _sanitize_request(payload: CreateChatGenerationRun) -> dict[str, Any]:
return sanitized
+def _timezone_headers(request: Request) -> dict[str, str]:
+ return {
+ name: value
+ for name in (CURRENT_DATE_TIMEZONE_HEADER, CURRENT_DATE_TIMEZONE_OFFSET_HEADER)
+ if 0 < len(value := request.headers.get(name, "").strip()) <= 64
+ }
+
+
def _require_run(run_id: str) -> dict[str, Any]:
run = db.get_run(run_id)
if run is None:
@@ -299,6 +311,8 @@ async def create_chat_generation_run(
current_subject: str = Depends(get_current_subject),
):
sanitized = _sanitize_request(payload)
+ if timezone_headers := _timezone_headers(request):
+ sanitized[db.TIMEZONE_HEADERS_FIELD] = timezone_headers
# Serialize the off-loop commit with model lifecycle work, so a run is registered either before the gate opens or
# after an unload/swap, never mid-swap.
async with inference_lifecycle_gate():
diff --git a/studio/backend/storage/chat_generation_runs_db.py b/studio/backend/storage/chat_generation_runs_db.py
index 98868b7058b..818cfc7d398 100644
--- a/studio/backend/storage/chat_generation_runs_db.py
+++ b/studio/backend/storage/chat_generation_runs_db.py
@@ -341,6 +341,11 @@ def _loads(value: str | None, fallback: Any) -> Any:
return fallback
+# Kept with the run for the producer but outside its identity, so a retried create from a browser whose offset moved
+# (DST) still matches the committed run.
+TIMEZONE_HEADERS_FIELD = "timezone_headers"
+
+
def canonical_request(
*,
thread_id: str,
@@ -359,7 +364,11 @@ def canonical_request(
"threadId": thread_id,
"userMessageId": user_message_id,
"assistantMessageId": assistant_message_id,
- "requestPayload": request_payload,
+ "requestPayload": {
+ key: value
+ for key, value in request_payload.items()
+ if key != TIMEZONE_HEADERS_FIELD
+ },
},
sort_keys = True,
separators = (",", ":"),
diff --git a/studio/backend/tests/test_account_deactivated_account_fences.py b/studio/backend/tests/test_account_deactivated_account_fences.py
index f79d421583a..a719360ab6e 100644
--- a/studio/backend/tests/test_account_deactivated_account_fences.py
+++ b/studio/backend/tests/test_account_deactivated_account_fences.py
@@ -144,7 +144,8 @@ async def test_owner_run_cannot_take_a_deactivated_accounts_supervisor_slot(
start = lambda run_id, **identity: started.append(run_id),
)
request = SimpleNamespace(
- app = SimpleNamespace(state = SimpleNamespace(chat_generation_supervisor = supervisor))
+ app = SimpleNamespace(state = SimpleNamespace(chat_generation_supervisor = supervisor)),
+ headers = {},
)
committed: list = []
diff --git a/studio/backend/tests/test_chat_generation_supervisor.py b/studio/backend/tests/test_chat_generation_supervisor.py
index a01d2afb88b..c81daac2f7e 100644
--- a/studio/backend/tests/test_chat_generation_supervisor.py
+++ b/studio/backend/tests/test_chat_generation_supervisor.py
@@ -5,10 +5,12 @@ import asyncio
import json
import threading
import time
+from datetime import date, datetime, timezone
from types import SimpleNamespace
from unittest.mock import AsyncMock
import pytest
+from starlette.requests import Request
from core.inference import llama_keepwarm
from core.inference.chat_generation_runs import (
@@ -22,6 +24,7 @@ from routes import inference
from state import active_generations
from storage import chat_generation_runs_db as runs_db
from storage import studio_db
+from utils.current_date_prompt_settings import _request_local_date
@pytest.fixture
@@ -72,7 +75,8 @@ def _create_payload(content = "Hello"):
def _route_request(supervisor):
return SimpleNamespace(
- app = SimpleNamespace(state = SimpleNamespace(chat_generation_supervisor = supervisor))
+ app = SimpleNamespace(state = SimpleNamespace(chat_generation_supervisor = supervisor)),
+ headers = {},
)
@@ -197,6 +201,67 @@ async def test_background_producer_persists_chunks_and_completes(durable_run, mo
] == chunks
+async def test_producer_dates_the_prompt_in_the_browser_timezone(monkeypatch):
+ studio_db.upsert_chat_thread(
+ {"id": "thread-1", "title": "Chat", "modelType": "base", "modelId": "local", "createdAt": 1}
+ )
+ studio_db.upsert_chat_message(
+ {"id": "user-1", "threadId": "thread-1", "role": "user", "content": [], "createdAt": 2}
+ )
+ browser = Request(
+ {
+ "type": "http",
+ "method": "POST",
+ "path": "/api/inference/chat-runs",
+ "query_string": b"",
+ "headers": [
+ (b"x-unsloth-timezone", b"Pacific/Kiritimati"),
+ (b"x-unsloth-timezone-offset-minutes", b"-840"),
+ ],
+ "app": SimpleNamespace(state = SimpleNamespace()),
+ }
+ )
+ await run_routes.create_chat_generation_run(_create_payload(), browser, "alice")
+ instant = datetime(2026, 9, 24, 10, 5, tzinfo = timezone.utc)
+ observed = []
+
+ async def body():
+ yield "data: [DONE]\n\n"
+
+ async def fake(payload, request, _subject, *, cancel_on_disconnect):
+ observed.append((sorted(payload.model_extra), _request_local_date(request, instant)))
+ return SimpleNamespace(status_code = 200, body_iterator = body())
+
+ monkeypatch.setattr(inference, "produce_openai_chat_completions", fake)
+ supervisor = ChatGenerationSupervisor(SimpleNamespace(state = SimpleNamespace()))
+ await supervisor._produce("run-1")
+ assert observed == [(["generation_run_id"], date(2026, 9, 25))]
+
+
+async def test_a_create_retried_across_a_dst_change_returns_the_committed_run():
+ studio_db.upsert_chat_thread(
+ {"id": "thread-1", "title": "Chat", "modelType": "base", "modelId": "local", "createdAt": 1}
+ )
+ studio_db.upsert_chat_message(
+ {"id": "user-1", "threadId": "thread-1", "role": "user", "content": [], "createdAt": 2}
+ )
+
+ def browser(offset):
+ request = _route_request(None)
+ request.headers = {
+ "x-unsloth-timezone": "America/Los_Angeles",
+ "x-unsloth-timezone-offset-minutes": offset,
+ }
+ return request
+
+ first = await run_routes.create_chat_generation_run(_create_payload(), browser("420"), "alice")
+ retry = await run_routes.create_chat_generation_run(_create_payload(), browser("480"), "alice")
+ assert (first["created"], retry["created"]) == (True, False)
+ assert retry["requestPayload"]["timezone_headers"]["x-unsloth-timezone-offset-minutes"] == "420"
+
+
@pytest.mark.asyncio
async def test_a_prefill_reporting_only_progress_renews_the_lease(durable_run, monkeypatch):
"""A 250K prefill outruns the 1200s lease before its first token, and the