unsloth/studio/backend/loggers/handlers.py
Daniel Han 9a966adf51
Studio: trim serving-log noise and surface llama-server engine stats (#6377)
* Studio: trim serving-log noise and surface llama-server engine stats

Studio prints one structured line per HTTP request, so the SPA's polling and
per-invalidation fan-out bury the lines that matter.

- Dedup identical successful GETs within a short window (default 300ms,
  UNSLOTH_STUDIO_ACCESS_LOG_DEDUP_MS) so a burst logs once. The dedup key
  includes the query string, so distinct query-driven GETs are not collapsed.
  Runs after the response is sent, so it adds no request latency; mutations,
  non-2xx, and loading polls are untouched.
- Collapse pure-liveness polls (/api/health, /api/auth/status,
  /api/inference/status, /api/inference/monitor) to a longer heartbeat
  (default 10s, UNSLOTH_STUDIO_ACCESS_LOG_POLL_DEDUP_MS). The API monitor
  console polls /monitor every 1.5s while open.
- Translate llama-server's Prometheus /metrics into a periodic vLLM-style
  engine_stats line (generation/prompt throughput and requests in flight) from
  a daemon poller, gated on UNSLOTH_STUDIO_ENGINE_STATS. Throughput uses
  llama-server's predicted_tokens_seconds / prompt_tokens_seconds gauges, with
  a tokens_predicted_total / prompt_tokens_total counter-delta fallback; it does
  not use n_decode_total (which counts llama_decode() calls, not tokens). No KV
  field is emitted, since llama.cpp does not expose kv_cache_usage_ratio.
  --metrics is added only when probe_server_capabilities reports the binary
  supports it, so older/custom binaries still load. The poller keeps retrying
  through transient scrape failures (stop() drives shutdown) and a malformed
  sample cannot crash its thread.
- api_monitor.append_reply: once the preview cap is reached, skip the per-chunk
  re-concat (avoids O(n^2) on long generations) while still recording the "..."
  truncation marker for a reply that lands exactly on the cap.
- unsloth studio --verbose and unsloth studio run --verbose both restore every
  per-request log; --verbose before a subcommand is rejected with guidance
  (matching --secure / --parallel). run --verbose still forwards --log-verbose
  to llama-server, preserving the pre-existing pass-through verbosity.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
2026-06-17 05:37:57 -07:00

176 lines
6.1 KiB
Python

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""Structured logging handlers and middleware.
LoggingMiddleware (request/response logging with timing),
filter_sensitive_data (structlog processor for sanitization), and
get_logger (factory for structured loggers).
"""
import os
import re
import time
import structlog
from starlette.types import ASGIApp, Message, Receive, Scope, Send
from utils.native_path_leases import redact_native_paths
logger = structlog.get_logger(__name__)
def _env_int(name: str, default: int) -> int:
try:
raw = (os.environ.get(name) or "").strip()
return int(raw) if raw else default
except ValueError:
return default
# Drop duplicate successful-GET access logs repeated within the window: the SPA
# fans one cache invalidation into many identical list fetches; only the first
# informs. Loading polls, mutations, and errors are unaffected. 0 = log all.
_ACCESS_LOG_DEDUP_MS = _env_int("UNSLOTH_STUDIO_ACCESS_LOG_DEDUP_MS", 300)
# Pure-liveness/UI polls whose access line carries no signal beyond "client still
# polling" (state changes are logged by their own modules). Collapsed to a longer
# heartbeat instead of one line per poll; first hit and any error still log. 0 = off.
_QUIET_POLL_DEDUP_MS = _env_int("UNSLOTH_STUDIO_ACCESS_LOG_POLL_DEDUP_MS", 10000)
_QUIET_POLL_PATHS = {
"/api/health",
"/api/auth/status",
"/api/inference/status",
"/api/inference/monitor",
}
_DEDUP_MAP_MAX = 4096
_NATIVE_PATH_LEASE_RE = re.compile(
r"(?i)(\b(?:native_path_lease|nativePathLease)[\"']?\s*[:=]\s*[\"']?)[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+"
)
_EXCLUDED_PATHS = {
"/api/train/status",
"/api/train/metrics",
"/api/train/hardware",
"/api/system",
}
_EXCLUDED_SUFFIXES = (
".png",
".jpg",
".jpeg",
".svg",
".ico",
".woff",
".woff2",
".ttf",
)
class LoggingMiddleware:
"""ASGI request logger that avoids BaseHTTPMiddleware streaming wrappers."""
def __init__(self, app: ASGIApp) -> None:
self.app = app
# (method, path, query, status_code) -> monotonic ts of the last EMITTED log.
self._last_log: dict[tuple[str, str, bytes, int], float] = {}
def _is_redundant_repeat(
self, method: str, path: str, query: bytes, status_code: int, now: float
) -> bool:
"""True if an identical GET/2xx log fired < window ago. The query string
is part of the identity, so distinct query-driven GETs are not collapsed.
Mutations and non-2xx are never deduped. Quiet-poll paths use a longer
heartbeat window. Stamps only on emit, so steady polls still log."""
if method != "GET" or not (200 <= status_code < 300):
return False
window_ms = _QUIET_POLL_DEDUP_MS if path in _QUIET_POLL_PATHS else _ACCESS_LOG_DEDUP_MS
if window_ms <= 0:
return False
key = (method, path, query, status_code)
last = self._last_log.get(key)
if last is not None and (now - last) * 1000.0 < window_ms:
return True
self._last_log[key] = now
if len(self._last_log) > _DEDUP_MAP_MAX:
cutoff = now - (max(_ACCESS_LOG_DEDUP_MS, _QUIET_POLL_DEDUP_MS) / 1000.0)
self._last_log = {k: v for k, v in self._last_log.items() if v >= cutoff}
return False
async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
if scope["type"] != "http":
await self.app(scope, receive, send)
return
path = scope["path"]
excluded = (
path in _EXCLUDED_PATHS
or path.startswith("/assets/")
or path.endswith(_EXCLUDED_SUFFIXES)
)
start_time = time.perf_counter()
status_code = 500
async def send_wrapper(message: Message) -> None:
nonlocal status_code
if message["type"] == "http.response.start":
status_code = message["status"]
await send(message)
try:
await self.app(scope, receive, send_wrapper)
except Exception as exc:
logger.error(
"request_failed",
path = path,
method = scope["method"],
status_code = status_code,
error = str(exc),
process_time_ms = round((time.perf_counter() - start_time) * 1000, 2),
exc_info = True,
)
raise
else:
end_time = time.perf_counter()
if not excluded and not self._is_redundant_repeat(
scope["method"], path, scope.get("query_string", b""), status_code, end_time
):
logger.info(
"request_completed",
method = scope["method"],
path = path,
status_code = status_code,
process_time_ms = round((end_time - start_time) * 1000, 2),
)
def filter_sensitive_data(logger, method_name, event_dict):
"""Structlog processor to redact native path leases from logs."""
def filter_value(value):
if isinstance(value, str):
try:
value = redact_native_paths(value)
except Exception:
pass
value = _NATIVE_PATH_LEASE_RE.sub(r"\1<redacted native path lease>", value)
return value
elif isinstance(value, dict):
return {
k: "<redacted native path lease>"
if str(k).replace("_", "").lower() == "nativepathlease"
else filter_value(v)
for k, v in value.items()
}
elif isinstance(value, list):
return [filter_value(item) for item in value]
return value
return {
k: "<redacted native path lease>"
if str(k).replace("_", "").lower() == "nativepathlease"
else filter_value(v)
for k, v in event_dict.items()
}
def get_logger(name: str) -> structlog.BoundLogger:
"""Get a bound structured logger for a module (name is usually __name__)."""
return structlog.get_logger(name)