qpkit

qpkit is a Python library for downloading NOAA precipitation GRIB2 products — MRMS QPE, WPC QPF, and HRRR — and, for MRMS QPE and HRRR, writing the downloaded grids directly into HEC-DSS files for use in hydrologic models.

Installation

Clone the repository and install qpkit in editable mode:

git clone https://github.com/gyanz/qpkit.git
cd qpkit
python -m pip install -e .

Editable mode requires a local clone, but if you just want to install the latest code without one, pip can install directly from a Git URL:

python -m pip install git+https://github.com/gyanz/qpkit.git

An optional dev extra adds developer tools (pytest, ruff, mypy, sphinx):

python -m pip install -e ".[dev]"

Products

qpkit.QPKit is the high-level entry point for every supported product. Use it as a context manager so the underlying HTTP client closes cleanly:

from qpkit import QPKit

with QPKit() as kit:
    ...

Request	Product family	Source	Writes to DSS
QPERequest	MRMS QPE — MultiSensor_Pass1, MultiSensor_Pass2, RadarOnly, MultiSensor_Pass1_Pass2	MRMS (live) / AWS S3 (archive)	Yes — QPEGridOptions
QPFRequest	WPC gridded QPF	WPC 2.5 km GRIB FTP	Yes — QPFGridOptions
HRRRRequest	HRRR — APCP (accumulated precip) or PRATE (precip rate)	NOMADS GRIB filter	Yes — HRRRGridOptions

download_qpe / download_qpf / download_hrrr (and the generic download) return a DownloadResult:

Field	Type	Meaning
succeeded	tuple[Path, ...]	Paths that were written.
failed	tuple[str, ...]	URLs that did not download.
skipped	tuple[Path, ...]	Paths to files already present locally.
plan	DownloadPlan	The plan that was executed.

DownloadResult.all_files returns a sorted tuple[Path, ...] combining succeeded and skipped — convenient when feeding the full set of available files into a downstream processing step.

Build a plan without downloading — build_plan dispatches on the request type (QPERequest, QPFRequest, or HRRRRequest); product-specific build_plan_qpe / build_plan_qpf / build_plan_hrrr are also available:

plan = kit.build_plan(request, "download")
for item in plan.items:
    print(item.filename, item.url, item.valid_time)

Each DownloadItem exposes filename, url, and an optional valid_time (populated from the embedded timestamp where the product encodes one). A DownloadPlan carries the output_dir it was built for, so two-stage execution is just kit.downloader.download(plan) — no second argument.

QPE (MRMS)

Important QPERequest parameters:

• product — MultiSensor_Pass1 (lower latency, less QA/QC), MultiSensor_Pass2 (higher quality, more latency), RadarOnly (radar-derived, no gauge correction), or MultiSensor_Pass1_Pass2 (hybrid: Pass1 for the most recent ~25 hours, Pass2 for everything older — requires start/end).
• interval — accumulation window in hours: 1, 3, 6, 12, 24, 48, 72.
• start / end — optional timezone-aware datetimes for an archive time range (AWS S3, available from 2020-10-14 onward).
• source — "auto" (NCEP for live requests, AWS when a time range is given), "aws", or "ncep".

Download QPE GRIB2 files:

from pathlib import Path
from qpkit import QPERequest, QPKit

with QPKit() as kit:
    request = QPERequest(product="MultiSensor_Pass1", interval=1)
    result = kit.download_qpe(request, Path("download"))

    print(f"downloaded={len(result.succeeded)} "
          f"skipped={len(result.skipped)} "
          f"failed={len(result.failed)}")

Download and write the grids to a DSS file in one step with download_to_dss and QPEGridOptions:

from pydsstools.core.crs import shg
from qpkit import BoundingBox, QPERequest, QPKit
from qpkit.models import QPEGridOptions

with QPKit() as kit:
    request = QPERequest(product="MultiSensor_Pass1", interval=1)
    options = QPEGridOptions(
        part_a="MRMS",
        part_b="SEATTLE",  # name the cropped region, not CONUS, when `extents` is set
        part_c="PRECIP",
        crs=shg(),  # destination CRS for reprojection; this is also the QPEGridOptions default
        cell_size=1000,  # meters; None = native GRIB resolution
        # Crop to this bounding box before writing; omit `extents` to write full CONUS.
        extents=BoundingBox(left_lon=-122.6, right_lon=-121.9, top_lat=47.8, bottom_lat=47.3),
    )

    result = kit.download_to_dss(
        request,
        "download",
        "mrms_qpe.dss",
        grid_options=options,
        dss_version=6,
        # `force` defaults to False here and that's fine to leave as-is: MRMS
        # GRIB2 filenames encode only the product and timestamp, and the
        # downloaded grid always covers the full CONUS regardless of `extents`
        # — cropping happens later, when writing to DSS. (Contrast with HRRR,
        # where `force=True` is recommended; see the HRRR section below.)
    )

    print(f"downloaded={len(result.download.succeeded)} "
          f"written={len(result.dss.written)} "
          f"skipped={len(result.dss.skipped)} "
          f"failed={len(result.dss.failed)}")
    for pathname in result.dss.all_pathnames:
        print(pathname)

QPERequest.until_now(start=...) is a convenience constructor for archive pulls that should run up through the current completed UTC hour.

QPF (WPC)

Important QPFRequest parameters:

• interval — forecast accumulation window in hours: 6, 24, 48, 120.
• cycle_hour — UTC forecast cycle hour, 0..23.

from pathlib import Path
from qpkit import QPFRequest, QPKit

with QPKit() as kit:
    request = QPFRequest(interval=6, cycle_hour=0)
    result = kit.download_qpf(request, Path("download"))

    print(f"downloaded={len(result.succeeded)} "
          f"skipped={len(result.skipped)} "
          f"failed={len(result.failed)}")

Download and write the grids to a DSS file in one step with download_to_dss and QPFGridOptions. Each WPC QPF file is a discrete accumulation bucket; the E part is the forecast cycle time plus the forecast hour and the D part is the E part minus the accumulation interval:

from pydsstools.core.crs import shg
from qpkit import BoundingBox, QPFRequest, QPKit
from qpkit.models import QPFGridOptions

with QPKit() as kit:
    request = QPFRequest(interval=6, cycle_hour=0)
    options = QPFGridOptions(
        part_a="WPC",
        part_b="CONUS",  # name the cropped region, not CONUS, when `extents` is set
        part_c="PRECIP",
        crs=shg(),  # destination CRS for reprojection; this is also the QPFGridOptions default
        cell_size=2500,  # meters; None = native GRIB resolution
        # Crop to this bounding box before writing; omit `extents` to write full CONUS.
        # extents=BoundingBox(left_lon=-98.0, right_lon=-94.0, top_lat=33.0, bottom_lat=29.0),
    )

    result = kit.download_to_dss(
        request,
        "download",
        "wpc_qpf.dss",
        grid_options=options,
        dss_version=6,
    )

    print(f"downloaded={len(result.download.succeeded)} "
          f"written={len(result.dss.written)} "
          f"skipped={len(result.dss.skipped)} "
          f"failed={len(result.dss.failed)}")
    for pathname in result.dss.all_pathnames:
        print(pathname)

HRRR

Important HRRRRequest parameters:

• precip_type — "APCP" (accumulated precipitation) or "PRATE" (precipitation rate).
• cycle_hour — UTC hour of the model run, 0..23.
• cycle_date — date of the run, or None for today (UTC).
• forecast_hours — tuple of lead-time hours from the cycle, each 0..48.
• bbox — a BoundingBox cropping the NOMADS GRIB filter request. It isn't optional (there's no None form), but its default already spans the full CONUS extent (left_lon=-125, right_lon=-65, top_lat=50, bottom_lat=24), so omitting it downloads and writes the full CONUS grid.

Download HRRR GRIB2 files:

from datetime import date
from pathlib import Path
from qpkit import BoundingBox, HRRRRequest, QPKit

download_folder = Path("download")  # GRIB2 files are saved here, named by cycle/forecast hour

with QPKit() as kit:
    request = HRRRRequest(
        precip_type="APCP",
        cycle_date=date(2026, 5, 8),
        cycle_hour=5,
        forecast_hours=(6,),
        bbox=BoundingBox(left_lon=-122.6, right_lon=-121.9, top_lat=47.8, bottom_lat=47.3),
    )
    result = kit.download_hrrr(
        request,
        download_folder,
        # The on-disk filename only encodes date/cycle/hour/precip_type, not the
        # bbox, so re-running with a different `bbox` (or while NOMADS is still
        # serving an in-progress cycle) would otherwise skip a stale local file.
        # force=True is recommended whenever the requested extent may vary
        # between downloads.
        force=True,
    )

    print(f"downloaded={len(result.succeeded)} "
          f"skipped={len(result.skipped)} "
          f"failed={len(result.failed)}")

Download and write the grids to DSS with download_to_dss and HRRRGridOptions. HRRRGridOptions.for_apcp() / .for_prate() fill in the right destination units, data type, and unit conversion for each precip type; the interval argument to download_to_dss (HRRR APCP only) computes per-interval accumulations from a single end forecast hour instead of storing the raw running total:

from datetime import date
from pathlib import Path
from pydsstools.core.crs import shg
from qpkit import BoundingBox, HRRRRequest, QPKit
from qpkit.models import HRRRGridOptions

download_folder = Path("download")  # GRIB2 files are saved here before being written to DSS
dss_file = download_folder / "precip.dss"

with QPKit() as kit:
    request = HRRRRequest(
        precip_type="APCP",
        cycle_date=date(2026, 5, 8),
        cycle_hour=5,
        forecast_hours=(6,),  # end hour; download_to_dss derives the intermediates
        bbox=BoundingBox(left_lon=-122.6, right_lon=-121.9, top_lat=47.8, bottom_lat=47.3),
    )
    options = HRRRGridOptions.for_apcp(
        part_a="HRRR",
        part_b="SEATTLE",  # name the cropped region the bbox covers, not CONUS
        part_c="PRECIP",
        part_f="HRR-APCP-1H",
        crs=shg(),  # destination CRS for reprojection; this is also the HRRRGridOptions default
        cell_size=3000,
    )

    result = kit.download_to_dss(
        request,
        download_folder,
        dss_file,
        grid_options=options,
        dss_version=6,
        # True auto-detects the latest available NOMADS cycle, overriding
        # request.cycle_date and request.cycle_hour with that cycle's values.
        cycle_hour_latest=False,
        interval=1,
        # The on-disk filename only encodes date/cycle/hour/precip_type, not the
        # bbox, so re-running with a different `bbox` (or while NOMADS is still
        # serving an in-progress cycle) would otherwise skip a stale local file.
        # force=True is recommended whenever the requested extent may vary
        # between downloads.
        force=True,
    )

    print(f"downloaded={len(result.download.succeeded)} "
          f"written={len(result.dss.written)} "
          f"skipped={len(result.dss.skipped)} "
          f"failed={len(result.dss.failed)}")
    for pathname in result.dss.all_pathnames:
        print(pathname)

download_to_dss returns a DownloadDSSResult, combining the download phase (result.download, a DownloadResult) with the DSS write phase (result.dss, a DSSWriteResult exposing written, skipped, failed, source_files, and all_pathnames).

For two-stage workflows — writing GRIBs that came from elsewhere, or re-writing an existing set of files under different DSS conventions — use the writer services directly: kit.hrrr_dss.write(...) / kit.qpe_dss.write(...).

Validation and serialization

Every request (QPERequest, QPFRequest, HRRRRequest, BoundingBox) and every result (DownloadResult, DownloadDSSResult, DSSWriteResult, ...) is an immutable pydantic v2 model. Two consequences are worth knowing:

• Requests validate at construction time. Invalid field values or combinations raise pydantic.ValidationError immediately, before any network call is made — so mistakes surface at the point you build the request, not three steps later inside a download:

  from qpkit import HRRRRequest
  
  # forecast hours must be in 0..48 — this raises pydantic.ValidationError
  # ("forecast hour 49 out of range 0..48") instead of failing during download.
  HRRRRequest(cycle_hour=0, forecast_hours=(0, 6, 49))

• Results are easy to log or serialize. model_dump() and model_dump_json() are available on every result type, including DownloadResult, DownloadDSSResult, and DSSWriteResult.