gradio-app--gradio/gradio/profiling.py

from __future__ import annotations

import contextvars
import os
import time
from collections import deque
from contextlib import asynccontextmanager, contextmanager
from dataclasses import dataclass, field
from functools import wraps
from typing import Any

PROFILING_ENABLED = os.environ.get("GRADIO_PROFILING", "").strip() in ("1", "true")


@dataclass
class RequestTrace:
    event_id: str | None = None
    fn_name: str | None = None
    session_hash: str | None = None
    timestamp: float = field(default_factory=time.time)

    queue_wait_ms: float = 0.0
    preprocess_ms: float = 0.0
    fn_call_ms: float = 0.0
    postprocess_ms: float = 0.0
    streaming_diff_ms: float = 0.0
    total_ms: float = 0.0
    n_iterations: int = 0
    upload_ms: float = 0.0
    preprocess_move_to_cache_ms: float = 0.0
    preprocess_format_image_ms: float = 0.0
    postprocess_save_img_array_to_cache_ms: float = 0.0
    preprocess_audio_from_file_ms: float = 0.0
    postprocess_save_audio_to_cache_ms: float = 0.0
    preprocess_video_ms: float = 0.0
    postprocess_video_convert_video_to_playable_mp4_ms: float = 0.0
    postprocess_update_state_in_config_ms: float = 0.0
    postprocess_move_to_cache_ms: float = 0.0
    postprocess_video_ms: float = 0.0
    postprocess_save_pil_to_cache_ms: float = 0.0
    postprocess_save_bytes_to_cache_ms: float = 0.0
    save_file_to_cache_ms: float = 0.0

    def set_phase(self, name: str, duration_ms: float):
        attr = f"{name}_ms"
        if hasattr(self, attr):
            # Accumulate across generator iterations
            setattr(self, attr, getattr(self, attr) + duration_ms)
        if name == "total":
            self.n_iterations += 1

    def to_dict(self) -> dict[str, Any]:
        return {
            "event_id": self.event_id,
            "fn_name": self.fn_name,
            "session_hash": self.session_hash,
            "timestamp": self.timestamp,
            "queue_wait_ms": self.queue_wait_ms,
            "preprocess_ms": self.preprocess_ms,
            "fn_call_ms": self.fn_call_ms,
            "postprocess_ms": self.postprocess_ms,
            "streaming_diff_ms": self.streaming_diff_ms,
            "total_ms": self.total_ms,
            "n_iterations": self.n_iterations,
            "preprocess_move_to_cache_ms": self.preprocess_move_to_cache_ms,
            "preprocess_format_image_ms": self.preprocess_format_image_ms,
            "postprocess_save_img_array_to_cache_ms": self.postprocess_save_img_array_to_cache_ms,
            "preprocess_audio_from_file_ms": self.preprocess_audio_from_file_ms,
            "postprocess_save_audio_to_cache_ms": self.postprocess_save_audio_to_cache_ms,
            "preprocess_video_ms": self.preprocess_video_ms,
            "postprocess_video_convert_video_to_playable_mp4_ms": self.postprocess_video_convert_video_to_playable_mp4_ms,
            "postprocess_update_state_in_config_ms": self.postprocess_update_state_in_config_ms,
            "postprocess_move_to_cache_ms": self.postprocess_move_to_cache_ms,
            "postprocess_video_ms": self.postprocess_video_ms,
            "postprocess_save_pil_to_cache_ms": self.postprocess_save_pil_to_cache_ms,
            "postprocess_save_bytes_to_cache_ms": self.postprocess_save_bytes_to_cache_ms,
            "save_file_to_cache_ms": self.save_file_to_cache_ms,
        }


_current_trace: contextvars.ContextVar[RequestTrace | None] = contextvars.ContextVar(
    "_current_trace", default=None
)


def get_current_trace() -> RequestTrace | None:
    return _current_trace.get()


def set_current_trace(trace: RequestTrace) -> contextvars.Token:
    return _current_trace.set(trace)


@asynccontextmanager
async def trace_phase(name: str):
    """Async context manager that records timing for a named phase into the current trace."""
    trace = _current_trace.get()
    if trace is None:
        yield
        return
    start = time.monotonic()
    try:
        yield
    finally:
        duration_ms = (time.monotonic() - start) * 1000
        trace.set_phase(name, duration_ms)


@contextmanager
def trace_phase_sync(name: str):
    """Context manager that records timing for a named phase into the current trace."""
    trace = _current_trace.get()
    if trace is None:
        yield
        return
    start = time.monotonic()
    try:
        yield
    finally:
        duration_ms = (time.monotonic() - start) * 1000
        trace.set_phase(name, duration_ms)


def traced(phase):
    if not PROFILING_ENABLED:
        return lambda f: f

    def _factory(f):
        @wraps(f)
        async def wrapper(*args, **kwargs):
            async with trace_phase(phase):
                return await f(*args, **kwargs)

        return wrapper

    return _factory


def traced_sync(phase):
    if not PROFILING_ENABLED:
        return lambda f: f

    def _factory(f):
        @wraps(f)
        def wrapper(*args, **kwargs):
            with trace_phase_sync(phase):
                return f(*args, **kwargs)

        return wrapper

    return _factory


class TraceCollector:
    def __init__(self, maxlen: int = 100_000):
        self._traces: deque[RequestTrace] = deque(maxlen=maxlen)

    def add(self, trace: RequestTrace):
        self._traces.append(trace)

    def get_all(self, last_n: int | None = None) -> list[dict[str, Any]]:
        traces = list(self._traces)
        if last_n is not None:
            traces = traces[-last_n:]
        return [t.to_dict() for t in traces]

    def get_summary(self) -> dict[str, Any]:
        if not self._traces:
            return {"count": 0, "phases": {}}

        import numpy as np

        prediction_traces = [
            t for t in self._traces if t.fn_name != "gradio_file_upload"
        ]
        upload_traces = [t for t in self._traces if t.fn_name == "gradio_file_upload"]

        phases = [
            "queue_wait",
            "preprocess",
            "fn_call",
            "postprocess",
            "streaming_diff",
            "total",
        ]

        def _percentiles(arr):
            return {
                "p50": float(np.percentile(arr, 50)),
                "p90": float(np.percentile(arr, 90)),
                "p95": float(np.percentile(arr, 95)),
                "p99": float(np.percentile(arr, 99)),
                "mean": float(np.mean(arr)),
                "min": float(np.min(arr)),
                "max": float(np.max(arr)),
            }

        result: dict[str, Any] = {
            "count": len(prediction_traces),
            "phases": {},
        }
        for phase in phases:
            values = [getattr(t, f"{phase}_ms") for t in prediction_traces]
            if values:
                result["phases"][phase] = _percentiles(np.array(values))
            else:
                result["phases"][phase] = {
                    "p50": 0.0,
                    "p90": 0.0,
                    "p95": 0.0,
                    "p99": 0.0,
                    "mean": 0.0,
                    "min": 0.0,
                    "max": 0.0,
                }

        if upload_traces:
            upload_values = [t.upload_ms for t in upload_traces]
            result["upload"] = {
                "count": len(upload_traces),
                **_percentiles(np.array(upload_values)),
            }

        return result

    def clear(self):
        self._traces.clear()


# Global collector instance
collector = TraceCollector()


if not PROFILING_ENABLED:
    # Replace with no-ops for zero overhead

    @asynccontextmanager
    async def trace_phase(name: str):  # noqa: ARG001
        yield

    @contextmanager
    def trace_phase_sync(name: str):  # noqa: ARG001
        yield