Hweh-6M

Hweh-6M

Summary

Task: Weather Forecasting
Inputs: 72 hours time-series
Outputs: 12h multivariate forecast
Params: 6M
Framework: PyTorch

Author: Paul Courneya (Harley-ml)

Description

Hweh-6M is a 6-million-parameter LSTM model trained to predict the next 12 hours of weather, including temperature, humidity, pressure, precipitation, and more, using the previous 72 hours of weather context.
We recommend using this model as a backup to a weather API or for offline forecasting when internet access is unavailable.
However, this model was primarily trained to serve as a teacher for Hweh-446k.

We would also like to give a shoutout to Open-Meteo for providing a free-to-use weather forecasting API.

Why “Hweh”?

In Proto-Indo-European, the root *h₂weh₁- means “to blow.” We chose it as the name for a weather forecasting model because of its connection to wind and air.

Architecture

The model uses a multitask LSTM setup:

Training

We trained Hweh-6M on 4.06 million rows of weather data from 82 locations for one epoch, using a batch size of 16 and gradient accumulation of 5. Training ran for 6 hours on an RTX 2060 6GB GPU.

Input Features

1. temperature_2m_norm
2. relative_humidity_2m_norm
3. apparent_temperature_norm
4. precipitation_log_norm
5. sea_level_pressure_norm
6. surface_pressure_norm
7. cloud_cover_total_norm
8. visibility_norm
9. wind_speed_10m_norm
10. wind_direction_10m_sin
11. wind_direction_10m_cos
12. hour_sin
13. hour_cos
14. day_of_year_sin
15. day_of_year_cos
16. weather_code_onehot_clear
17. weather_code_onehot_cloudy
18. weather_code_onehot_fog
19. weather_code_onehot_drizzle
20. weather_code_onehot_rain
21. weather_code_onehot_snow
22. weather_code_onehot_thunderstorm

Output Features

1. y_temp_c: continuous regression
2. y_humidity: continuous regression
3. y_apparent_temperature: continuous regression
4. y_precipitation_mm: continuous regression
5. y_sea_level_pressure_hpa: continuous regression
6. y_surface_pressure_hpa: continuous regression
7. y_cloud_cover_total: continuous regression
8. y_wind_speed_10m: continuous regression
9. y_wind_direction_sin: continuous regression
10. y_wind_direction_cos: continuous regression
11. y_rain_prob: binary classification
12. y_weather_class: multiclass classification

Training Results

Training & Evaluation Metrics

Regression Error Metrics (MAE)

Generation Examples

City=Seattle

{
  "city": "Seattle",
  "location_id": "1",
  "model_location_id": 0,
  "data_source": "open-meteo forecast api (past-hours context only)",
  "requested_at_utc": "2026-05-08T10:50:59.201916+00:00",
  "context": {
    "hours": 72,
    "start_utc": "2026-05-05T10:00:00+00:00",
    "end_utc": "2026-05-08T09:00:00+00:00",
    "start_local": "[REDACTED]",
    "end_local": "[REDACTED]"
  },
  "model": {
    "encoder_type": "lstm",
    "seq_len": 72,
    "input_dim": 22,
    "num_weather_classes": 7
  },
  "forecast": [
    {
      "lead_hours": 1,
      "target_utc": "2026-05-08T10:00:00+00:00",
      "target_local": "[REDACTED]",
      "temperature_2m_c": 10.081672668457031,
      "relative_humidity_2m_pct": 83.57363891601562,
      "apparent_temperature_c": 8.784963607788086,
      "precipitation_mm": 0.0014822654193267226,
      "pressure_msl_hpa": 1019.1531372070312,
      "surface_pressure_hpa": 1010.2232666015625,
      "cloud_cover_pct": 23.58745574951172,
      "wind_speed_10m_kmh": 5.772420883178711,
      "rain_probability": 0.0017376710893586278,
      "weather_class": 0,
      "weather_class_name": "class_0",
      "weather_class_probabilities": {
        "class_0": 0.570536732673645,
        "class_1": 0.40933191776275635,
        "class_2": 0.019639208912849426,
        "class_3": 0.00027382391272112727,
        "class_4": 0.00020224291074555367,
        "class_5": 1.215106385643594e-05,
        "class_6": 3.905456196662271e-06
      }
    },
    {
      "lead_hours": 2,
      "target_utc": "2026-05-08T11:00:00+00:00",
      "target_local": "[REDACTED]",
      "temperature_2m_c": 10.004827499389648,
      "relative_humidity_2m_pct": 84.41477966308594,
      "apparent_temperature_c": 8.68188762664795,
      "precipitation_mm": 0.00013178338122088462,
      "pressure_msl_hpa": 1018.954345703125,
      "surface_pressure_hpa": 1010.0797729492188,
      "cloud_cover_pct": 30.078432083129883,
      "wind_speed_10m_kmh": 5.812187194824219,
      "rain_probability": 0.004383997060358524,
      "weather_class": 1,
      "weather_class_name": "class_1",
      "weather_class_probabilities": {
        "class_0": 0.4759916663169861,
        "class_1": 0.4977237284183502,
        "class_2": 0.024964570999145508,
        "class_3": 0.0006970795802772045,
        "class_4": 0.0005866154097020626,
        "class_5": 3.2389318221248686e-05,
        "class_6": 3.920148628822062e-06
      }
    },
    {
      "lead_hours": 3,
      "target_utc": "2026-05-08T12:00:00+00:00",
      "target_local": "[REDACTED]"
    }
  ],
  "sanity": {
    "sequence_shape": [
      72,
      22
    ],
    "finite_features": true
  }
}

City=Nuuk

{
  "city": "Nuuk",
  "location_id": "83",
  "model_location_id": 0,
  "data_source": "open-meteo forecast api (past-hours context only)",
  "requested_at_utc": "2026-05-08T10:59:51.127984+00:00",
  "context": {
    "hours": 72,
    "start_utc": "2026-05-05T10:00:00+00:00",
    "end_utc": "2026-05-08T09:00:00+00:00",
    "start_local": "2026-05-05T09:00:00-01:00",
    "end_local": "2026-05-08T08:00:00-01:00"
  },
  "model": {
    "encoder_type": "lstm",
    "seq_len": 72,
    "input_dim": 22,
    "num_weather_classes": 7
  },
  "forecast": [
    {
      "lead_hours": 1,
      "target_utc": "2026-05-08T10:00:00+00:00",
      "target_local": "2026-05-08T09:00:00-01:00",
      "temperature_2m_c": 3.745473861694336,
      "relative_humidity_2m_pct": 87.24557495117188,
      "apparent_temperature_c": -1.1178970336914062,
      "precipitation_mm": 0.9192219972610474,
      "pressure_msl_hpa": 999.2293090820312,
      "surface_pressure_hpa": 984.041015625,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 22.042539596557617,
      "rain_probability": 0.9964759945869446,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 9.087142825592309e-05,
        "class_1": 0.011110298335552216,
        "class_2": 0.0009554459829814732,
        "class_3": 0.28936928510665894,
        "class_4": 0.19745060801506042,
        "class_5": 0.5009360313415527,
        "class_6": 8.744438673602417e-05
      }
    },
    {
      "lead_hours": 2,
      "target_utc": "2026-05-08T11:00:00+00:00",
      "target_local": "2026-05-08T10:00:00-01:00",
      "temperature_2m_c": 3.6379480361938477,
      "relative_humidity_2m_pct": 87.90388488769531,
      "apparent_temperature_c": -1.197652816772461,
      "precipitation_mm": 0.8211548924446106,
      "pressure_msl_hpa": 998.41796875,
      "surface_pressure_hpa": 983.3368530273438,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 21.754901885986328,
      "rain_probability": 0.9918462634086609,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.0002011256874538958,
        "class_1": 0.021235737949609756,
        "class_2": 0.0012973389821127057,
        "class_3": 0.20620499551296234,
        "class_4": 0.25035062432289124,
        "class_5": 0.5206562280654907,
        "class_6": 5.390366641222499e-05
      }
    },
    {
      "lead_hours": 3,
      "target_utc": "2026-05-08T12:00:00+00:00",
      "target_local": "2026-05-08T11:00:00-01:00",
      "temperature_2m_c": 3.482311248779297,
      "relative_humidity_2m_pct": 88.61299896240234,
      "apparent_temperature_c": -1.3543472290039062,
      "precipitation_mm": 0.7267112731933594,
      "pressure_msl_hpa": 997.7637939453125,
      "surface_pressure_hpa": 982.8118286132812,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 21.31927490234375,
      "rain_probability": 0.9851851463317871,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.0003230531292501837,
        "class_1": 0.030711255967617035,
        "class_2": 0.0014986724127084017,
        "class_3": 0.17889709770679474,
        "class_4": 0.2378082126379013,
        "class_5": 0.5507404208183289,
        "class_6": 2.121348188666161e-05
      }
    },
    {
      "lead_hours": 4,
      "target_utc": "2026-05-08T13:00:00+00:00",
      "target_local": "2026-05-08T12:00:00-01:00",
      "temperature_2m_c": 3.324540138244629,
      "relative_humidity_2m_pct": 89.35970306396484,
      "apparent_temperature_c": -1.5628299713134766,
      "precipitation_mm": 0.6503503322601318,
      "pressure_msl_hpa": 997.3221435546875,
      "surface_pressure_hpa": 982.2531127929688,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 20.908214569091797,
      "rain_probability": 0.9797365069389343,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.0005368430283851922,
        "class_1": 0.036228444427251816,
        "class_2": 0.00266513810493052,
        "class_3": 0.1584056168794632,
        "class_4": 0.2387750893831253,
        "class_5": 0.5633592009544373,
        "class_6": 2.9700731829507276e-05
      }
    },
    {
      "lead_hours": 5,
      "target_utc": "2026-05-08T14:00:00+00:00",
      "target_local": "2026-05-08T13:00:00-01:00",
      "temperature_2m_c": 3.088955879211426,
      "relative_humidity_2m_pct": 90.08441162109375,
      "apparent_temperature_c": -1.7932510375976562,
      "precipitation_mm": 0.5726789832115173,
      "pressure_msl_hpa": 997.1259155273438,
      "surface_pressure_hpa": 982.1145629882812,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 20.37297821044922,
      "rain_probability": 0.9752851724624634,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.0007767326897010207,
        "class_1": 0.04325678199529648,
        "class_2": 0.0034333993680775166,
        "class_3": 0.15728847682476044,
        "class_4": 0.24588856101036072,
        "class_5": 0.5493185520172119,
        "class_6": 3.7467838410520926e-05
      }
    },
    {
      "lead_hours": 6,
      "target_utc": "2026-05-08T15:00:00+00:00",
      "target_local": "2026-05-08T14:00:00-01:00",
      "temperature_2m_c": 2.8550186157226562,
      "relative_humidity_2m_pct": 90.83024597167969,
      "apparent_temperature_c": -1.9607181549072266,
      "precipitation_mm": 0.4950953722000122,
      "pressure_msl_hpa": 997.0792236328125,
      "surface_pressure_hpa": 981.837646484375,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 19.884090423583984,
      "rain_probability": 0.9711479544639587,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.001066686469130218,
        "class_1": 0.05448324978351593,
        "class_2": 0.003781423671171069,
        "class_3": 0.15267883241176605,
        "class_4": 0.23838046193122864,
        "class_5": 0.5495800971984863,
        "class_6": 2.9315853680600412e-05
      }
    },
    {
      "lead_hours": 7,
      "target_utc": "2026-05-08T16:00:00+00:00",
      "target_local": "2026-05-08T15:00:00-01:00",
      "temperature_2m_c": 2.6384010314941406,
      "relative_humidity_2m_pct": 91.38716888427734,
      "apparent_temperature_c": -2.114431381225586,
      "precipitation_mm": 0.43851515650749207,
      "pressure_msl_hpa": 997.214111328125,
      "surface_pressure_hpa": 981.5133666992188,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 19.454288482666016,
      "rain_probability": 0.9665488600730896,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.0014152604853734374,
        "class_1": 0.059757016599178314,
        "class_2": 0.0037699665408581495,
        "class_3": 0.1557641476392746,
        "class_4": 0.22861963510513306,
        "class_5": 0.550651490688324,
        "class_6": 2.2469554096460342e-05
      }
    },
    {
      "lead_hours": 8,
      "target_utc": "2026-05-08T17:00:00+00:00",
      "target_local": "2026-05-08T16:00:00-01:00",
      "temperature_2m_c": 2.4830856323242188,
      "relative_humidity_2m_pct": 91.72871398925781,
      "apparent_temperature_c": -2.212369918823242,
      "precipitation_mm": 0.38016656041145325,
      "pressure_msl_hpa": 997.3843994140625,
      "surface_pressure_hpa": 981.6067504882812,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 19.01665496826172,
      "rain_probability": 0.9600462913513184,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.001761714811436832,
        "class_1": 0.06388058513402939,
        "class_2": 0.005221065599471331,
        "class_3": 0.13114923238754272,
        "class_4": 0.21121768653392792,
        "class_5": 0.5867400765419006,
        "class_6": 2.95877835014835e-05
      }
    },
    {
      "lead_hours": 9,
      "target_utc": "2026-05-08T18:00:00+00:00",
      "target_local": "2026-05-08T17:00:00-01:00",
      "temperature_2m_c": 2.3713502883911133,
      "relative_humidity_2m_pct": 91.92076110839844,
      "apparent_temperature_c": -2.2490768432617188,
      "precipitation_mm": 0.3401757478713989,
      "pressure_msl_hpa": 997.632568359375,
      "surface_pressure_hpa": 981.5086059570312,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 18.69092559814453,
      "rain_probability": 0.9514879584312439,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.002726545324549079,
        "class_1": 0.08241794258356094,
        "class_2": 0.007858957163989544,
        "class_3": 0.14714762568473816,
        "class_4": 0.2103891372680664,
        "class_5": 0.5494091510772705,
        "class_6": 5.06224823766388e-05
      }
    },
    {
      "lead_hours": 10,
      "target_utc": "2026-05-08T19:00:00+00:00",
      "target_local": "2026-05-08T18:00:00-01:00",
      "temperature_2m_c": 2.334117889404297,
      "relative_humidity_2m_pct": 91.97515869140625,
      "apparent_temperature_c": -2.216022491455078,
      "precipitation_mm": 0.29920822381973267,
      "pressure_msl_hpa": 997.88671875,
      "surface_pressure_hpa": 981.637451171875,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 18.297332763671875,
      "rain_probability": 0.9422094821929932,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.0030131975654512644,
        "class_1": 0.0824747234582901,
        "class_2": 0.00828808918595314,
        "class_3": 0.1445200890302658,
        "class_4": 0.1672952026128769,
        "class_5": 0.5943665504455566,
        "class_6": 4.211435225442983e-05
      }
    },
    {
      "lead_hours": 11,
      "target_utc": "2026-05-08T20:00:00+00:00",
      "target_local": "2026-05-08T19:00:00-01:00",
      "temperature_2m_c": 2.399325370788574,
      "relative_humidity_2m_pct": 91.36509704589844,
      "apparent_temperature_c": -2.106843948364258,
      "precipitation_mm": 0.2678143382072449,
      "pressure_msl_hpa": 998.099853515625,
      "surface_pressure_hpa": 981.798583984375,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 17.996307373046875,
      "rain_probability": 0.9368607401847839,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.004000477027148008,
        "class_1": 0.09458598494529724,
        "class_2": 0.01128436904400587,
        "class_3": 0.14569725096225739,
        "class_4": 0.17799843847751617,
        "class_5": 0.5663700699806213,
        "class_6": 6.342450069496408e-05
      }
    },
    {
      "lead_hours": 12,
      "target_utc": "2026-05-08T21:00:00+00:00",
      "target_local": "2026-05-08T20:00:00-01:00",
      "temperature_2m_c": 2.473114013671875,
      "relative_humidity_2m_pct": 90.60014343261719,
      "apparent_temperature_c": -1.94183349609375,
      "precipitation_mm": 0.23492039740085602,
      "pressure_msl_hpa": 998.2453002929688,
      "surface_pressure_hpa": 981.8583374023438,
      "cloud_cover_pct": 100.0,
      "wind_speed_10m_kmh": 17.61905860900879,
      "rain_probability": 0.9265281558036804,
      "weather_class": 5,
      "weather_class_name": "class_5",
      "weather_class_probabilities": {
        "class_0": 0.004572723060846329,
        "class_1": 0.10134521126747131,
        "class_2": 0.016071945428848267,
        "class_3": 0.12629590928554535,
        "class_4": 0.11316182464361191,
        "class_5": 0.6384702920913696,
        "class_6": 8.211767271859571e-05
      }
    }
  ],
  "sanity": {
    "sequence_shape": [
      72,
      22
    ],
    "finite_features": true
  }
}

Note

In observed outputs, the model is often within 0.3°C of the actual value.

Furthermore, you can pass locations that are not present in the model’s location embedding table. We’ve observed that the model can generalize to out-of-distribution (OOD) cities, with an estimated accuracy drop of only about 2–5%. However, this figure is an estimate and does not reflect a true ground-truth measurement.

Use Cases

Intended for:

1. Backup to API
2. Offline forecasting if you have the data
3. Research
4. Or more simply, for fun

Not intended for:

1. Safety-critical forecasting (aviation, emergency response)
2. Replacing meteorological or API services

Limitations

1. The model is not perfectly accurate and will produce approximate forecasts rather than exact real-world weather conditions.
2. Prediction accuracy decreases as the forecast horizon increases up to 12 hours.
3. Performance may degrade on unseen or underrepresented geographic regions and climate types.
4. The model does not enforce physical laws of atmospheric dynamics and may produce physically inconsistent outputs.
5. Forecast quality is sensitive to the quality and completeness of input weather data.
6. Rare or extreme weather events are underrepresented in training data and may be poorly predicted.
7. Weather class outputs are simplified and do not capture fine-grained meteorological distinctions.

Inference

#!/usr/bin/env python3
from __future__ import annotations

import json
import time
from pathlib import Path
from typing import Any

import numpy as np
import pandas as pd
import requests
import torch
from transformers import AutoConfig, AutoModel
from zoneinfo import ZoneInfo

# ----------------------------
# Change these values here
# ----------------------------
MODEL_ID = r"Harley-ml/Hweh-6M"   # HF repo id or local path
CITY = "New York"
SEQUENCE_META_PATH = "Harley-ml/Hweh-6M/weather_sequences.metadata.json"
CONTEXT_HOURS = 72
FORECAST_HOURS = 12
DEVICE = None  # "cpu", "cuda", "cuda:0", or None for auto

API_BASE_URL = "https://api.open-meteo.com/v1/forecast"
MAX_RETRIES = 6
REQUEST_TIMEOUT_S = 60

HOURLY_VARS = [
    "temperature_2m",
    "relative_humidity_2m",
    "apparent_temperature",
    "precipitation",
    "weather_code",
    "pressure_msl",
    "surface_pressure",
    "cloud_cover",
    "visibility",
    "wind_speed_10m",
    "wind_direction_10m",
]

WEATHER_CODE_BUCKETS = 7
TEMP_SCALE = 50.0
HUMIDITY_SCALE = 100.0
WIND_SCALE = 100.0

# ----------------------------
# City metadata (82 locations)
# ----------------------------
CITY_SPECS: dict[str, dict[str, Any]] = {
    "Seattle": {"location_id": "1", "latitude": 47.6062, "longitude": -122.3321, "continent": "North America", "climate_tag": "temperate_oceanic", "elevation": 56},
    "Portland": {"location_id": "2", "latitude": 45.5152, "longitude": -122.6784, "continent": "North America", "climate_tag": "temperate_oceanic", "elevation": 15},
    "San Francisco": {"location_id": "3", "latitude": 37.7749, "longitude": -122.4194, "continent": "North America", "climate_tag": "foggy_mediterranean", "elevation": 16},
    "Los Angeles": {"location_id": "4", "latitude": 34.0522, "longitude": -118.2437, "continent": "North America", "climate_tag": "sunny_mediterranean", "elevation": 71},
    "Denver": {"location_id": "5", "latitude": 39.7392, "longitude": -104.9903, "continent": "North America", "climate_tag": "semi_arid_highland", "elevation": 1609},
    "Chicago": {"location_id": "6", "latitude": 41.8781, "longitude": -87.6298, "continent": "North America", "climate_tag": "humid_continental", "elevation": 181},
    "Dallas": {"location_id": "7", "latitude": 32.7767, "longitude": -96.7970, "continent": "North America", "climate_tag": "hot_subhumid", "elevation": 131},
    "Atlanta": {"location_id": "8", "latitude": 33.7490, "longitude": -84.3880, "continent": "North America", "climate_tag": "humid_subtropical", "elevation": 320},
    "New York": {"location_id": "9", "latitude": 40.7128, "longitude": -74.0060, "continent": "North America", "climate_tag": "humid_subtropical", "elevation": 10},
    "Miami": {"location_id": "10", "latitude": 25.7617, "longitude": -80.1918, "continent": "North America", "climate_tag": "tropical_humid", "elevation": 2},
    "Phoenix": {"location_id": "11", "latitude": 33.4484, "longitude": -112.0740, "continent": "North America", "climate_tag": "hot_arid", "elevation": 331},
    "Salt Lake City": {"location_id": "12", "latitude": 40.7608, "longitude": -111.8910, "continent": "North America", "climate_tag": "semi_arid", "elevation": 1288},
    "Anchorage": {"location_id": "13", "latitude": 61.2181, "longitude": -149.9003, "continent": "North America", "climate_tag": "subarctic_snowy", "elevation": 31},
    "Minneapolis": {"location_id": "14", "latitude": 44.9778, "longitude": -93.2650, "continent": "North America", "climate_tag": "cold_snowy", "elevation": 264},
    "Toronto": {"location_id": "15", "latitude": 43.6532, "longitude": -79.3832, "continent": "North America", "climate_tag": "humid_continental", "elevation": 76},
    "Montreal": {"location_id": "16", "latitude": 45.5017, "longitude": -73.5673, "continent": "North America", "climate_tag": "cold_snowy", "elevation": 233},
    "Vancouver": {"location_id": "17", "latitude": 49.2827, "longitude": -123.1207, "continent": "North America", "climate_tag": "temperate_oceanic", "elevation": 70},
    "Mexico City": {"location_id": "18", "latitude": 19.4326, "longitude": -99.1332, "continent": "North America", "climate_tag": "highland_subtropical", "elevation": 2240},
    "Havana": {"location_id": "19", "latitude": 23.1136, "longitude": -82.3666, "continent": "North America", "climate_tag": "tropical_humid", "elevation": 59},
    "San Juan": {"location_id": "20", "latitude": 18.4655, "longitude": -66.1057, "continent": "North America", "climate_tag": "tropical_humid", "elevation": 8},

    "Lima": {"location_id": "21", "latitude": -12.0464, "longitude": -77.0428, "continent": "South America", "climate_tag": "coastal_arid", "elevation": 154},
    "Santiago": {"location_id": "22", "latitude": -33.4489, "longitude": -70.6693, "continent": "South America", "climate_tag": "mediterranean", "elevation": 520},
    "Buenos Aires": {"location_id": "23", "latitude": -34.6037, "longitude": -58.3816, "continent": "South America", "climate_tag": "humid_subtropical", "elevation": 25},
    "Bogotá": {"location_id": "24", "latitude": 4.7110, "longitude": -74.0721, "continent": "South America", "climate_tag": "highland_cool", "elevation": 2640},
    "Quito": {"location_id": "25", "latitude": -0.1807, "longitude": -78.4678, "continent": "South America", "climate_tag": "highland_equatorial", "elevation": 2850},
    "Caracas": {"location_id": "26", "latitude": 10.4806, "longitude": -66.9036, "continent": "South America", "climate_tag": "tropical_humid", "elevation": 900},
    "Rio de Janeiro": {"location_id": "27", "latitude": -22.9068, "longitude": -43.1729, "continent": "South America", "climate_tag": "tropical_humid", "elevation": 5},
    "São Paulo": {"location_id": "28", "latitude": -23.5505, "longitude": -46.6333, "continent": "South America", "climate_tag": "humid_subtropical", "elevation": 760},
    "La Paz": {"location_id": "29", "latitude": -16.4897, "longitude": -68.1193, "continent": "South America", "climate_tag": "highland_cold", "elevation": 3640},
    "Cusco": {"location_id": "30", "latitude": -13.5319, "longitude": -71.9675, "continent": "South America", "climate_tag": "highland_cool", "elevation": 3399},
    "Montevideo": {"location_id": "31", "latitude": -34.9011, "longitude": -56.1645, "continent": "South America", "climate_tag": "temperate_oceanic", "elevation": 43},
    "Asunción": {"location_id": "32", "latitude": -25.2637, "longitude": -57.5759, "continent": "South America", "climate_tag": "humid_subtropical", "elevation": 43},
    "Manaus": {"location_id": "33", "latitude": -3.1190, "longitude": -60.0217, "continent": "South America", "climate_tag": "tropical_humid", "elevation": 92},
    "Recife": {"location_id": "34", "latitude": -8.0476, "longitude": -34.8770, "continent": "South America", "climate_tag": "tropical_coastal", "elevation": 4},
    "Punta Arenas": {"location_id": "35", "latitude": -53.1638, "longitude": -70.9171, "continent": "South America", "climate_tag": "cold_windy", "elevation": 34},

    "London": {"location_id": "36", "latitude": 51.5074, "longitude": -0.1278, "continent": "Europe", "climate_tag": "temperate_oceanic", "elevation": 11},
    "Paris": {"location_id": "37", "latitude": 48.8566, "longitude": 2.3522, "continent": "Europe", "climate_tag": "temperate_oceanic", "elevation": 35},
    "Madrid": {"location_id": "38", "latitude": 40.4168, "longitude": -3.7038, "continent": "Europe", "climate_tag": "hot_summer_mediterranean", "elevation": 667},
    "Rome": {"location_id": "39", "latitude": 41.9028, "longitude": 12.4964, "continent": "Europe", "climate_tag": "hot_summer_mediterranean", "elevation": 21},
    "Berlin": {"location_id": "40", "latitude": 52.52, "longitude": 13.4050, "continent": "Europe", "climate_tag": "temperate_continental", "elevation": 34},
    "Stockholm": {"location_id": "41", "latitude": 59.3293, "longitude": 18.0686, "continent": "Europe", "climate_tag": "cold_marine", "elevation": 28},
    "Oslo": {"location_id": "42", "latitude": 59.9139, "longitude": 10.7522, "continent": "Europe", "climate_tag": "cold_snowy", "elevation": 23},
    "Helsinki": {"location_id": "43", "latitude": 60.1699, "longitude": 24.9384, "continent": "Europe", "climate_tag": "cold_snowy", "elevation": 25},
    "Reykjavik": {"location_id": "44", "latitude": 64.1466, "longitude": -21.9426, "continent": "Europe", "climate_tag": "cold_windy", "elevation": 12},
    "Kyiv": {"location_id": "45", "latitude": 50.4501, "longitude": 30.5234, "continent": "Europe", "climate_tag": "humid_continental", "elevation": 179},
    "Lisbon": {"location_id": "46", "latitude": 38.7223, "longitude": -9.1393, "continent": "Europe", "climate_tag": "sunny_mediterranean", "elevation": 7},
    "Athens": {"location_id": "47", "latitude": 37.9838, "longitude": 23.7275, "continent": "Europe", "climate_tag": "sunny_mediterranean", "elevation": 70},
    "Zurich": {"location_id": "48", "latitude": 47.3769, "longitude": 8.5417, "continent": "Europe", "climate_tag": "temperate_continental", "elevation": 408},
    "Dublin": {"location_id": "49", "latitude": 53.3498, "longitude": -6.2603, "continent": "Europe", "climate_tag": "temperate_oceanic", "elevation": 20},
    "Vienna": {"location_id": "50", "latitude": 48.2082, "longitude": 16.3738, "continent": "Europe", "climate_tag": "temperate_continental", "elevation": 171},

    "Dubai": {"location_id": "51", "latitude": 25.2048, "longitude": 55.2708, "continent": "Asia", "climate_tag": "hot_arid", "elevation": 16},
    "Riyadh": {"location_id": "52", "latitude": 24.7136, "longitude": 46.6753, "continent": "Asia", "climate_tag": "hot_arid", "elevation": 612},
    "Delhi": {"location_id": "53", "latitude": 28.7041, "longitude": 77.1025, "continent": "Asia", "climate_tag": "hot_semi_arid", "elevation": 216},
    "Mumbai": {"location_id": "54", "latitude": 19.0760, "longitude": 72.8777, "continent": "Asia", "climate_tag": "tropical_humid", "elevation": 14},
    "Bangkok": {"location_id": "55", "latitude": 13.7563, "longitude": 100.5018, "continent": "Asia", "climate_tag": "tropical_monsoon", "elevation": 2},
    "Singapore": {"location_id": "56", "latitude": 1.3521, "longitude": 103.8198, "continent": "Asia", "climate_tag": "tropical_humid", "elevation": 15},
    "Tokyo": {"location_id": "57", "latitude": 35.6762, "longitude": 139.6503, "continent": "Asia", "climate_tag": "humid_subtropical", "elevation": 40},
    "Seoul": {"location_id": "58", "latitude": 37.5665, "longitude": 126.9780, "continent": "Asia", "climate_tag": "humid_continental", "elevation": 38},
    "Ulaanbaatar": {"location_id": "59", "latitude": 47.8864, "longitude": 106.9057, "continent": "Asia", "climate_tag": "cold_steppe", "elevation": 1350},
    "Kathmandu": {"location_id": "60", "latitude": 27.7172, "longitude": 85.3240, "continent": "Asia", "climate_tag": "highland_subtropical", "elevation": 1400},
    "Chiang Mai": {"location_id": "61", "latitude": 18.7883, "longitude": 98.9853, "continent": "Asia", "climate_tag": "tropical_seasonal", "elevation": 300},
    "Lhasa": {"location_id": "62", "latitude": 29.6520, "longitude": 91.1721, "continent": "Asia", "climate_tag": "high_altitude_cold", "elevation": 3656},
    "Jakarta": {"location_id": "63", "latitude": -6.2088, "longitude": 106.8456, "continent": "Asia", "climate_tag": "tropical_humid", "elevation": 8},
    "Manila": {"location_id": "64", "latitude": 14.5995, "longitude": 120.9842, "continent": "Asia", "climate_tag": "tropical_humid", "elevation": 16},
    "Karachi": {"location_id": "65", "latitude": 24.8607, "longitude": 67.0011, "continent": "Asia", "climate_tag": "hot_arid", "elevation": 10},

    "Cairo": {"location_id": "66", "latitude": 30.0444, "longitude": 31.2357, "continent": "Africa", "climate_tag": "hot_arid", "elevation": 23},
    "Alexandria": {"location_id": "67", "latitude": 31.2001, "longitude": 29.9187, "continent": "Africa", "climate_tag": "coastal_mediterranean", "elevation": 5},
    "Casablanca": {"location_id": "68", "latitude": 33.5731, "longitude": -7.5898, "continent": "Africa", "climate_tag": "coastal_mediterranean", "elevation": 56},
    "Marrakech": {"location_id": "69", "latitude": 31.6295, "longitude": -7.9811, "continent": "Africa", "climate_tag": "hot_semi_arid", "elevation": 466},
    "Lagos": {"location_id": "70", "latitude": 6.5244, "longitude": 3.3792, "continent": "Africa", "climate_tag": "tropical_humid", "elevation": 41},
    "Nairobi": {"location_id": "71", "latitude": -1.2921, "longitude": 36.8219, "continent": "Africa", "climate_tag": "temperate_highland", "elevation": 1795},
    "Addis Ababa": {"location_id": "72", "latitude": 8.9806, "longitude": 38.7578, "continent": "Africa", "climate_tag": "temperate_highland", "elevation": 2355},
    "Cape Town": {"location_id": "73", "latitude": -33.9249, "longitude": 18.4241, "continent": "Africa", "climate_tag": "mediterranean", "elevation": 25},
    "Johannesburg": {"location_id": "74", "latitude": -26.2041, "longitude": 28.0473, "continent": "Africa", "climate_tag": "subtropical_highland", "elevation": 1753},
    "Windhoek": {"location_id": "75", "latitude": -22.5609, "longitude": 17.0658, "continent": "Africa", "climate_tag": "semi_arid", "elevation": 1650},
    "Accra": {"location_id": "76", "latitude": 5.6037, "longitude": -0.1870, "continent": "Africa", "climate_tag": "tropical_humid", "elevation": 61},
    "Kigali": {"location_id": "77", "latitude": -1.9441, "longitude": 30.0619, "continent": "Africa", "climate_tag": "highland_tropical", "elevation": 1567},
    "Tunis": {"location_id": "78", "latitude": 36.8065, "longitude": 10.1815, "continent": "Africa", "climate_tag": "mediterranean", "elevation": 4},
    "Dakar": {"location_id": "79", "latitude": -14.7167, "longitude": -17.4677, "continent": "Africa", "climate_tag": "hot_coastal", "elevation": 25},
    "Mombasa": {"location_id": "80", "latitude": -4.0435, "longitude": 39.6682, "continent": "Africa", "climate_tag": "tropical_coastal", "elevation": 17},

    "Sydney": {"location_id": "81", "latitude": -33.8688, "longitude": 151.2093, "continent": "Oceania", "climate_tag": "humid_subtropical", "elevation": 58},
    "Melbourne": {"location_id": "82", "latitude": -37.8136, "longitude": 144.9631, "continent": "Oceania", "climate_tag": "temperate_oceanic", "elevation": 31},
}

CITY_TIMEZONES: dict[str, str] = {
    "Seattle": "America/Los_Angeles",
    "Portland": "America/Los_Angeles",
    "San Francisco": "America/Los_Angeles",
    "Los Angeles": "America/Los_Angeles",
    "Denver": "America/Denver",
    "Chicago": "America/Chicago",
    "Dallas": "America/Chicago",
    "Atlanta": "America/New_York",
    "New York": "America/New_York",
    "Miami": "America/New_York",
    "Phoenix": "America/Phoenix",
    "Salt Lake City": "America/Denver",
    "Anchorage": "America/Anchorage",
    "Minneapolis": "America/Chicago",
    "Toronto": "America/Toronto",
    "Montreal": "America/Toronto",
    "Vancouver": "America/Vancouver",
    "Mexico City": "America/Mexico_City",
    "Havana": "America/Havana",
    "San Juan": "America/Puerto_Rico",
    "Lima": "America/Lima",
    "Santiago": "America/Santiago",
    "Buenos Aires": "America/Argentina/Buenos_Aires",
    "Bogotá": "America/Bogota",
    "Quito": "America/Guayaquil",
    "Caracas": "America/Caracas",
    "Rio de Janeiro": "America/Sao_Paulo",
    "São Paulo": "America/Sao_Paulo",
    "La Paz": "America/La_Paz",
    "Cusco": "America/Lima",
    "Montevideo": "America/Montevideo",
    "Asunción": "America/Asuncion",
    "Manaus": "America/Manaus",
    "Recife": "America/Recife",
    "Punta Arenas": "America/Punta_Arenas",
    "London": "Europe/London",
    "Paris": "Europe/Paris",
    "Madrid": "Europe/Madrid",
    "Rome": "Europe/Rome",
    "Berlin": "Europe/Berlin",
    "Stockholm": "Europe/Stockholm",
    "Oslo": "Europe/Oslo",
    "Helsinki": "Europe/Helsinki",
    "Reykjavik": "Atlantic/Reykjavik",
    "Kyiv": "Europe/Kyiv",
    "Lisbon": "Europe/Lisbon",
    "Athens": "Europe/Athens",
    "Zurich": "Europe/Zurich",
    "Dublin": "Europe/Dublin",
    "Vienna": "Europe/Vienna",
    "Dubai": "Asia/Dubai",
    "Riyadh": "Asia/Riyadh",
    "Delhi": "Asia/Kolkata",
    "Mumbai": "Asia/Kolkata",
    "Bangkok": "Asia/Bangkok",
    "Singapore": "Asia/Singapore",
    "Tokyo": "Asia/Tokyo",
    "Seoul": "Asia/Seoul",
    "Ulaanbaatar": "Asia/Ulaanbaatar",
    "Kathmandu": "Asia/Kathmandu",
    "Chiang Mai": "Asia/Bangkok",
    "Lhasa": "Asia/Shanghai",
    "Jakarta": "Asia/Jakarta",
    "Manila": "Asia/Manila",
    "Karachi": "Asia/Karachi",
    "Cairo": "Africa/Cairo",
    "Alexandria": "Africa/Cairo",
    "Casablanca": "Africa/Casablanca",
    "Marrakech": "Africa/Casablanca",
    "Lagos": "Africa/Lagos",
    "Nairobi": "Africa/Nairobi",
    "Addis Ababa": "Africa/Addis_Ababa",
    "Cape Town": "Africa/Johannesburg",
    "Johannesburg": "Africa/Johannesburg",
    "Windhoek": "Africa/Windhoek",
    "Accra": "Africa/Accra",
    "Kigali": "Africa/Kigali",
    "Tunis": "Africa/Tunis",
    "Dakar": "Africa/Dakar",
    "Mombasa": "Africa/Nairobi",
    "Sydney": "Australia/Sydney",
    "Melbourne": "Australia/Melbourne",
}

# ----------------------------
# Helpers
# ----------------------------
def weather_code_to_bucket(code) -> int:
    if code is None:
        return 1
    try:
        if pd.isna(code):
            return 1
    except Exception:
        pass

    code = int(code)
    if code == 0:
        return 0
    if code in (1, 2, 3):
        return 1
    if code in (45, 48):
        return 2
    if code in (51, 53, 55, 56, 57):
        return 3
    if code in (61, 63, 65, 66, 67, 80, 81, 82):
        return 4
    if code in (71, 73, 75, 77, 85, 86):
        return 5
    if code in (95, 96, 99):
        return 6
    return 1


def cyc(x: np.ndarray, period: float) -> tuple[np.ndarray, np.ndarray]:
    angle = 2.0 * np.pi * (x / period)
    return np.sin(angle), np.cos(angle)


def clamp_array(x: np.ndarray, lo: float | None = None, hi: float | None = None) -> np.ndarray:
    return np.clip(x, lo, hi)


def request_with_backoff(session: requests.Session, url: str, params: dict[str, Any]) -> dict[str, Any]:
    last_exc: Exception | None = None
    for attempt in range(MAX_RETRIES):
        try:
            resp = session.get(url, params=params, timeout=REQUEST_TIMEOUT_S)
            if resp.status_code == 429:
                retry_after = resp.headers.get("Retry-After")
                sleep_s = float(retry_after) if retry_after else min(60.0, 2**attempt)
                print(f"Rate limited. Sleeping {sleep_s:.1f}s and retrying.", flush=True)
                time.sleep(sleep_s)
                continue
            resp.raise_for_status()
            return resp.json()
        except Exception as e:
            last_exc = e
            sleep_s = min(60.0, 2**attempt)
            print(f"Request failed: {e}. Sleeping {sleep_s:.1f}s and retrying.", flush=True)
            time.sleep(sleep_s)
    raise RuntimeError(f"Failed after {MAX_RETRIES} retries: {params}") from last_exc


def load_sequence_meta(path: str) -> dict[str, Any]:
    p = Path(path)
    if not p.exists():
        return {"location_to_id": {}}
    with open(p, "r", encoding="utf-8") as f:
        meta = json.load(f)
    meta.setdefault("location_to_id", {})
    return meta


def load_model():
    config = AutoConfig.from_pretrained(MODEL_ID, trust_remote_code=True)
    model = AutoModel.from_pretrained(MODEL_ID, config=config, trust_remote_code=True)
    model.eval()
    return model, config


def fetch_recent_history(city: str, context_hours: int) -> pd.DataFrame:
    if city not in CITY_SPECS:
        raise ValueError(f"Unknown city: {city}")

    spec = CITY_SPECS[city]
    session = requests.Session()
    session.headers.update({"User-Agent": "Mozilla/5.0"})

    params = {
        "latitude": spec["latitude"],
        "longitude": spec["longitude"],
        "hourly": ",".join(HOURLY_VARS),
        "timezone": "UTC",
        "temperature_unit": "celsius",
        "wind_speed_unit": "kmh",
        "precipitation_unit": "mm",
        "past_hours": int(context_hours) + 2,
        "forecast_hours": 0,
    }

    data = request_with_backoff(session, API_BASE_URL, params=params)
    hourly = data.get("hourly", {})
    if "time" not in hourly:
        raise ValueError(f"No hourly data returned for {city}: {data}")

    df = pd.DataFrame(hourly)
    if df.empty:
        raise ValueError(f"Empty hourly response for {city}.")

    df["time"] = pd.to_datetime(df["time"], errors="coerce", utc=True)
    df = df.dropna(subset=["time"]).sort_values("time").drop_duplicates(subset=["time"]).reset_index(drop=True)

    needed = HOURLY_VARS
    missing = [c for c in needed if c not in df.columns]
    if missing:
        raise ValueError(f"Missing hourly columns in API response: {missing}")

    for c in needed:
        df[c] = pd.to_numeric(df[c], errors="coerce")

    df["weather_code"] = df["weather_code"].fillna(1)
    df["precipitation"] = df["precipitation"].fillna(0.0)

    for c in [
        "temperature_2m",
        "relative_humidity_2m",
        "apparent_temperature",
        "precipitation",
        "pressure_msl",
        "surface_pressure",
        "cloud_cover",
        "visibility",
        "wind_speed_10m",
        "wind_direction_10m",
    ]:
        df[c] = df[c].interpolate(limit_direction="both").ffill().bfill()

    now_utc = pd.Timestamp.now(tz="UTC")
    df = df[df["time"] <= now_utc].copy()

    if len(df) < context_hours:
        raise ValueError(f"Not enough observed rows: got {len(df)}, need {context_hours}")

    return df.tail(context_hours).reset_index(drop=True)


def build_single_sequence(df: pd.DataFrame) -> np.ndarray:
    hour = df["time"].dt.hour.to_numpy()
    doy = df["time"].dt.dayofyear.to_numpy()

    hour_sin, hour_cos = cyc(hour.astype(float), 24.0)
    doy_sin, doy_cos = cyc(doy.astype(float), 365.25)

    temp = np.nan_to_num(df["temperature_2m"].astype(float).to_numpy(), nan=0.0)
    humidity = np.nan_to_num(df["relative_humidity_2m"].astype(float).to_numpy(), nan=0.0)
    apparent = np.nan_to_num(df["apparent_temperature"].astype(float).to_numpy(), nan=0.0)
    precip = np.nan_to_num(df["precipitation"].astype(float).to_numpy(), nan=0.0)
    pressure = np.nan_to_num(df["pressure_msl"].astype(float).to_numpy(), nan=0.0)
    surface_pressure = np.nan_to_num(df["surface_pressure"].astype(float).to_numpy(), nan=0.0)
    cloud_cover = np.nan_to_num(df["cloud_cover"].astype(float).to_numpy(), nan=0.0)
    visibility = np.nan_to_num(df["visibility"].astype(float).to_numpy(), nan=0.0)
    wind = np.nan_to_num(df["wind_speed_10m"].astype(float).to_numpy(), nan=0.0)
    wind_dir = np.nan_to_num(df["wind_direction_10m"].astype(float).to_numpy(), nan=0.0)

    humidity = clamp_array(humidity, 0.0, 100.0)
    cloud_cover = clamp_array(cloud_cover, 0.0, 100.0)
    precip = clamp_array(precip, 0.0, None)
    wind = clamp_array(wind, 0.0, None)
    visibility = clamp_array(visibility, 0.0, None)

    wind_dir_sin, wind_dir_cos = cyc(wind_dir, 360.0)
    weather_bucket = df["weather_code"].fillna(1).apply(weather_code_to_bucket).to_numpy(dtype=np.int64)

    rows = []
    for i in range(len(df)):
        wc_oh = np.zeros(WEATHER_CODE_BUCKETS, dtype=np.float32)
        wc_oh[weather_bucket[i]] = 1.0

        row = np.concatenate(
            [
                np.array(
                    [
                        temp[i] / TEMP_SCALE,
                        humidity[i] / HUMIDITY_SCALE,
                        apparent[i] / TEMP_SCALE,
                        np.log1p(max(precip[i], 0.0)) / 3.0,
                        pressure[i] / 1100.0,
                        surface_pressure[i] / 1100.0,
                        cloud_cover[i] / 100.0,
                        visibility[i] / 50000.0,
                        wind[i] / WIND_SCALE,
                        wind_dir_sin[i],
                        wind_dir_cos[i],
                        hour_sin[i],
                        hour_cos[i],
                        doy_sin[i],
                        doy_cos[i],
                    ],
                    dtype=np.float32,
                ),
                wc_oh,
            ]
        )
        rows.append(row)

    seq = np.asarray(rows, dtype=np.float32)

    if not np.isfinite(seq).all():
        bad = np.argwhere(~np.isfinite(seq))
        raise ValueError(f"Non-finite values remain in sequence at positions like: {bad[:10].tolist()}")

    return seq


def to_iso(ts: pd.Timestamp, tz_name: str | None = None) -> str:
    if tz_name:
        try:
            return ts.tz_convert(ZoneInfo(tz_name)).isoformat()
        except Exception:
            pass
    return ts.isoformat()


def get_logits(out):
    if isinstance(out, dict) and "logits" in out:
        return out["logits"]
    if hasattr(out, "logits"):
        return out.logits
    return out


def resolve_location_index(seq_meta: dict[str, Any], city_location_id: str) -> int:
    location_to_id = seq_meta.get("location_to_id", {})

    if city_location_id in location_to_id:
        return int(location_to_id[city_location_id])

    try:
        as_int = int(city_location_id)
        if as_int in location_to_id:
            return int(location_to_id[as_int])
        if str(as_int) in location_to_id:
            return int(location_to_id[str(as_int)])
    except Exception:
        pass

    for unk_key in ("UNK", "<UNK>", "unknown", "UNKNOWN"):
        if unk_key in location_to_id:
            return int(location_to_id[unk_key])

    return 0


def predict():
    seq_meta = load_sequence_meta(SEQUENCE_META_PATH)
    model, config = load_model()

    if CITY not in CITY_SPECS:
        raise ValueError(f"Unknown city: {CITY}")

    if CONTEXT_HOURS <= 0:
        raise ValueError("CONTEXT_HOURS must be > 0")

    if hasattr(config, "seq_len") and int(config.seq_len) != CONTEXT_HOURS:
        raise ValueError(f"Set CONTEXT_HOURS to {int(config.seq_len)} for this model.")

    city_spec = CITY_SPECS[CITY]
    city_tz = CITY_TIMEZONES.get(CITY, "UTC")
    model_location_id = resolve_location_index(seq_meta, str(city_spec["location_id"]))

    df = fetch_recent_history(CITY, CONTEXT_HOURS)
    seq = build_single_sequence(df)

    X = torch.from_numpy(seq).unsqueeze(0)
    loc = torch.tensor([model_location_id], dtype=torch.long)

    target_device = torch.device(
        DEVICE if DEVICE else ("cuda" if torch.cuda.is_available() else "cpu")
    )
    model = model.to(target_device)
    X = X.to(target_device)
    loc = loc.to(target_device)

    weather_class_names = getattr(config, "weather_class_names", None)
    if not weather_class_names:
        weather_class_names = [f"class_{i}" for i in range(int(getattr(config, "num_weather_classes", 7)))]

    with torch.no_grad():
        out = model(X=X, location_id=loc)
        logits = get_logits(out)

        (
            temp_pred,
            humidity_pred,
            apparent_pred,
            precip_pred,
            sea_level_pressure_pred,
            surface_pressure_pred,
            cloud_cover_pred,
            wind_pred,
            wind_dir_sin_pred,
            wind_dir_cos_pred,
            rain_logit,
            weather_logits,
        ) = logits

        temp_pred = temp_pred.squeeze(0).detach().cpu().numpy()
        humidity_pred = humidity_pred.squeeze(0).detach().cpu().numpy()
        apparent_pred = apparent_pred.squeeze(0).detach().cpu().numpy()
        precip_pred = precip_pred.squeeze(0).detach().cpu().numpy()
        sea_level_pressure_pred = sea_level_pressure_pred.squeeze(0).detach().cpu().numpy()
        surface_pressure_pred = surface_pressure_pred.squeeze(0).detach().cpu().numpy()
        cloud_cover_pred = cloud_cover_pred.squeeze(0).detach().cpu().numpy()
        wind_pred = wind_pred.squeeze(0).detach().cpu().numpy()
        rain_prob = torch.sigmoid(rain_logit).squeeze(0).detach().cpu().numpy()
        weather_probs = torch.softmax(weather_logits, dim=-1).squeeze(0).detach().cpu().numpy()
        weather_idx = np.argmax(weather_probs, axis=-1).astype(np.int64)

        humidity_pred = np.clip(humidity_pred, 0.0, 100.0)
        cloud_cover_pred = np.clip(cloud_cover_pred, 0.0, 100.0)
        precip_pred = np.clip(precip_pred, 0.0, None)
        wind_pred = np.clip(wind_pred, 0.0, None)
        rain_prob = np.clip(rain_prob, 0.0, 1.0)

    context_start = df["time"].iloc[0]
    context_end = df["time"].iloc[-1]
    requested_at_utc = pd.Timestamp.now(tz="UTC")

    horizon = min(
        int(FORECAST_HOURS),
        int(temp_pred.shape[0]),
        int(humidity_pred.shape[0]),
        int(weather_idx.shape[0]),
    )

    forecast = []
    for lead in range(1, horizon + 1):
        target_time = context_end + pd.Timedelta(hours=lead)
        idx = lead - 1
        w_idx = int(weather_idx[idx])

        forecast.append(
            {
                "lead_hours": lead,
                "target_utc": target_time.isoformat(),
                "target_local": to_iso(target_time, city_tz),
                "temperature_2m_c": float(temp_pred[idx]),
                "relative_humidity_2m_pct": float(humidity_pred[idx]),
                "apparent_temperature_c": float(apparent_pred[idx]),
                "precipitation_mm": float(precip_pred[idx]),
                "pressure_msl_hpa": float(sea_level_pressure_pred[idx]),
                "surface_pressure_hpa": float(surface_pressure_pred[idx]),
                "cloud_cover_pct": float(cloud_cover_pred[idx]),
                "wind_speed_10m_kmh": float(wind_pred[idx]),
                "rain_probability": float(rain_prob[idx]),
                "weather_class": w_idx,
                "weather_class_name": weather_class_names[w_idx] if w_idx < len(weather_class_names) else f"class_{w_idx}",
                "weather_class_probabilities": {
                    name: float(prob) for name, prob in zip(weather_class_names, weather_probs[idx])
                },
            }
        )

    result = {
        "city": CITY,
        "location_id": str(city_spec["location_id"]),
        "model_location_id": int(model_location_id),
        "data_source": "open-meteo forecast api (past-hours context only)",
        "requested_at_utc": requested_at_utc.isoformat(),
        "context": {
            "hours": int(len(df)),
            "start_utc": context_start.isoformat(),
            "end_utc": context_end.isoformat(),
            "start_local": to_iso(context_start, city_tz),
            "end_local": to_iso(context_end, city_tz),
        },
        "model": {
            "model_id": MODEL_ID,
            "encoder_type": getattr(config, "encoder_type", None),
            "seq_len": int(getattr(config, "seq_len", CONTEXT_HOURS)),
            "input_dim": int(getattr(config, "input_dim", seq.shape[1])),
            "num_weather_classes": int(getattr(config, "num_weather_classes", len(weather_class_names))),
        },
        "forecast": forecast,
        "sanity": {
            "sequence_shape": list(seq.shape),
            "finite_features": bool(np.isfinite(seq).all()),
        },
    }

    print(json.dumps(result, indent=2))


if __name__ == "__main__":
    predict()

Related Models

1. DistilHweh-446k

Citation

@misc{Hweh-6m,
  title = {Hweh-6M: A 6M-Parameter LSTM for Short-Term Multivariate Weather Forecasting},
  author    = {Paul Courneya; Harley-ml},
  year      = {2026},
  url       = {https://huggingface.co/Harley-ml/Hweh-6M}
}