quill-sort 6.0.4

Adaptive ultra-sort: auto-dispatching, correctness-first sorting that beats np.sort by a mile on the array fast path

quill-sort

Adaptive Ultra-Sort — Quill auto-dispatches to the fastest correct sort for your data and your machine, and never loses to the baseline.

import numpy as np
import quill

# The list API: a drop-in for sorted() / list.sort(), ~3-4x faster on numeric.
quill.quill_sort([3, 1, 4, 1, 5, 9])              # → [1, 1, 3, 4, 5, 9]
quill.quill_sort(records, key=lambda r: r['age']) # sort objects by a key

# The array API: the "by a mile" path — dispatches to Rust radix / GPU / polars.
a = np.random.randint(0, 2**40, 20_000_000)
quill.sort_array(a)                                # ~5x faster than np.sort

---

The one thing to know: list API vs. array API

Quill has two doors, and which one you use determines how big the win is.

API	Input / output	Speed (numeric, this machine)	Why
quill_sort(list)	list in, list out	matches np.sort, ~3.8x vs sorted() / list.sort()	It does np.asarray(...) then .tolist() around a fast kernel. Those two conversions dominate the runtime (Amdahl's law), so the kernel speedup is mostly hidden — you get np.sort-class throughput, not the "by a mile" number.
quill.sort_array(ndarray)	ndarray in, ndarray out	~5.3x int64, ~3.3x float64, ~4x GPU vs np.sort	Zero conversions. It dispatches the raw buffer straight to the fastest installed backend. This is where Quill actually pulls away from numpy.

Plain version: if your data starts and ends as a Python list, the conversion wall caps the win — quill_sort is np.sort-fast and ~4x faster than the built-in sorted(), which is already a great deal, but it is not 5x. If you keep your data in numpy arrays and call sort_array, you get the full multi-threaded / GPU speedup. Don't believe any library that claims a 5x list-in / list-out sort — the conversion alone forbids it.

All numbers below are measured on this development box (Windows 11, 28-core CPU, RTX 4060 Ti, numpy 2.4.6) with a freshly-shuffled array on every timed run, so they reflect real cold-sort throughput, not warm/already-sorted artifacts. Your mileage will vary with CPU, GPU, array size, and dtype.

---

Installation

pip install quill-sort              # core — ZERO required deps, always correct
pip install quill-sort[fast]        # + numpy + psutil (the baseline fast path)
pip install quill-sort[polars]      # + polars  (no-compile multi-threaded sort)
pip install quill-sort[gpu]         # + cupy    (NVIDIA GPU sort)
pip install quill-sort[all]         # numpy + pandas + psutil + polars

The compiled Rust backend (the ~5x path) ships as a pre-built binary wheel for common platforms — pip install quill-sort just gets it, no Rust toolchain needed. If no wheel matches your platform, pip falls back to the pure-Python sdist and Quill transparently uses the next-fastest backend (GPU → polars → numpy → np.sort). The install never fails for lack of a compiler.

---

How it picks (the dispatch chain)

sort_array() profiles the array and walks a priority chain, taking the first backend that is installed and supports the array. Every step has a measured crossover (min_n) below which it isn't worth it, and any backend error falls straight back to np.sort — correctness is never at risk.

sort_array(ndarray)
        │
        ▼
  eligible?  (1-D, dtype kind i/u/f, itemsize ≤ 8, value-only, C-contiguous)
        │ no ──────────────────────────────────────────────► np.sort
        │ yes
        ▼
  dense bounded int64/uint64?  ──► counting sort (np.bincount)   ~1.7-2.8x, single-thread
        │ no
        ▼
  ┌──────────────────────────────────────────────────────────┐
  │  try, in priority order, the first available + supporting │
  │    rust_voracious   (compiled radix, n ≥ 1M)              │
  │    cupy_gpu         (GPU, n ≥ 2M, fits in free VRAM)      │
  │    polars           (no-compile MT sort, n ≥ 200k)        │
  │    numpy_parallel   (int-only partition sort, n ≥ 5M)     │
  └──────────────────────────────────────────────────────────┘
        │ any error / nothing eligible
        ▼
  np.sort   ← the floor; always correct, never slower than numpy

NaN is stripped before any backend runs and re-appended at the end (numpy convention), so a backend that can't order NaN never sees one. Descending is a post-sort reverse. quill.available_backends() shows what your machine will actually use, in this order.

---

Backends

Backend	How to get it	Measured (int64, vs np.sort)	Notes
rust_voracious	pip install quill-sort (bundled wheel)	~5.3x (float64 ~3.3x)	Compiled PyO3 + voracious_radix_sort, multi-threaded radix. Fastest measured path; leads the chain.
cupy_gpu	pip install quill-sort[gpu]	~4x (float64 ~2.5x)	GPU radix via CuPy. Counts the host→device→host round trip and still wins for large arrays that fit in VRAM.
polars	pip install quill-sort[polars]	~2.3x (float64 ~1.7x)	Delegates to polars' multi-threaded Rust sort. No compiler needed — the "fast for everyone" fallback.
numpy_parallel	pip install quill-sort[fast]	~1.1x (int only)	Thread-parallel np.partition sample-sort. Small but reliable int win; deliberately uses few workers (the gain saturates at memory bandwidth).
counting sort	pip install quill-sort[fast]	~1.7-2.8x	np.bincount for dense bounded int64/uint64. O(n+k), single-threaded, beats every comparison backend in its niche.
np.sort	always (with numpy)	1.0x (the floor)	The never-lose baseline. Reached on any error or ineligible dtype.

Without numpy at all, Quill still sorts correctly via the stdlib Timsort — it is simply slower. Correctness has zero required dependencies.

---

The array API — sort_array()

import numpy as np
import quill

a = np.random.randint(0, 2**40, 20_000_000)

s = quill.sort_array(a)                      # sorted COPY (a untouched), fastest backend
quill.sort_array(a, inplace=True)            # sort a IN PLACE, returns a
d = quill.sort_array(a, descending=True)     # reverse order

quill.available_backends()
# → ['rust_voracious', 'cupy_gpu', 'polars', 'numpy_parallel']  (priority order)

sort_array matches np.sort exactly — including negatives, mixed int/float promotion, and NaN-to-end — and is never slower than np.sort on any input.

Top-k — quill_topk()

When you only need the k smallest (or largest), don't sort the whole array. quill_topk uses numpy's argpartition (introselect, O(n)) instead of a full O(n log n) sort:

quill.quill_topk(scores, 10)                  # 10 smallest, sorted ascending
quill.quill_topk(scores, 10, largest=True)    # 10 largest,  sorted descending
quill.quill_topk(rows, 5, key=lambda r: r.size)

Measured ~8x faster than sorted(data, reverse=True)[:10] for k=10 over 10M elements on this machine.

analyze() — inspect the profile

quill.analyze([3, 1, 4, 1, 5, 9])
# {'n': 6, 'dtype': 'int_pos', 'presorted': False, 'dense': True, ...}

---

The list API — quill_sort() / quill_sorted()

A drop-in replacement for sorted() / list.sort() that routes numeric lists through the fast numeric kernel and everything else through Timsort. It returns a list, so it pays the conversion tax described above — expect np.sort-class speed and ~3-4x over the built-ins on numeric data, not the array-API number.

quill.quill_sort([3, 1, 4, 1, 5, 9])              # in-place, returns the list
quill.quill_sort(data, key=lambda x: x['score'])  # objects via key
quill.quill_sort(data, reverse=True)              # descending
quill.quill_sort(data, inplace=False)             # return a new list, leave input alone
quill.quill_sort(data, parallel=True)             # force multi-core
quill.quill_sort(data, stable=False)              # unstable, max speed on numeric
result = quill.quill_sorted(iterable)             # non-mutating, mirrors sorted()

quill.quill_sort(
    data,                        # list, generator, range, ndarray, Series, DataFrame
    key=None,                    # sort key function (like sorted())
    reverse=False,               # descending order
    inplace=True,                # mutate in place (False → return a new list)
    parallel=False,              # use multiple cores (auto on big numeric data)
    high_performance_mode=False, # skip the interactive prompt on the external-sort path
    silent=False,                # suppress status output
    stable=True,                 # True = matches sorted() exactly; False = max speed
    stats=False,                 # return (sorted_list, stats_dict)
)

stable parameter

stable=True (default) guarantees equal elements keep their original relative order — identical to Python's sorted(). stable=False allows the benchmark-optimal unstable kernels on numeric data for extra speed when order among equal elements doesn't matter.

quill.quill_sort(data)                 # stable, safe for all use cases
quill.quill_sort(data, stable=False)   # unstable, faster for large int32/int64

stats — timing

result, stats = quill.quill_sort(data, stats=True)
# stats = {'time_ms': 12.3, 'n': 1000000}

---

The never-lose guarantee

Every fast path in Quill is wrapped so that any failure — a missing backend, a GPU OOM, a compiled-extension panic, an unsupported dtype — falls back to np.sort (or stdlib Timsort if numpy is absent). The Rust extension is built with panic = "unwind", so even a native panic becomes a catchable Python exception and the process never aborts. You can never get a result that is wrong or a sort that is slower than the numpy baseline.

# To debug a backend instead of silently falling back, surface the error:
QUILL_BACKEND_DEBUG=1 python your_script.py

---

Correctness contract

Quill's output matches sorted() / np.sort on every supported input:

None values sort to the end (start with reverse=True):

quill.quill_sort([3, None, 1, None, 2])
# → [1, 2, 3, None, None]

NaN values in float data go to the end (start with reverse=True), matching numpy's convention:

quill.quill_sort([3.0, float('nan'), 1.0, 2.0])
# → [1.0, 2.0, 3.0, nan]

Mixed int/float lists are promoted to float64 and take the fast float path (not the slow object fallback):

quill.quill_sort([1, 2.5, 3, 4.0])
# → [1, 2.5, 3, 4.0]

Negative integers are handled natively (two's-complement radix; no object fallback), and keyed sorts are stable by default.

---

Supported types

• int, float, str, bytes — native fast paths
• Mixed int + float — promoted to float64
• Negative integers — handled natively
• None values — sorted to the end (start with reverse=True)
• float('nan') — sorted to the end (start with reverse=True)
• numpy.ndarray — sorted via the backend chain (use sort_array for the full win)
• pandas.Series — sorted, returned as a new Series
• pandas.DataFrame — sorted by column(s) via key='column_name'
• Any generator / iterator — materialised to a list, sorted, returned

---

Plugin system

Teach Quill to sort your own types. A plugin declares which types it handles and a prepare() that converts them to something Quill sorts natively, plus an optional postprocess to rebuild your objects.

from quill import register_plugin, QuillPlugin

class MyPlugin(QuillPlugin):
    handles = (MyCustomClass,)
    name    = "my_custom_class"

    @staticmethod
    def prepare(data, key, reverse):
        items       = [x.value for x in data]
        postprocess = lambda sorted_vals: [MyCustomClass(v) for v in sorted_vals]
        return items, key, postprocess   # (list_to_sort, key_fn, postprocess_fn)

register_plugin(MyPlugin)
quill.quill_sort(list_of_my_objects)

Built-in plugins cover numpy.ndarray, pandas.Series, pandas.DataFrame, range, and any generator/iterator. You can also drop in a custom backend (not just a type plugin) with quill._backends.register_backend(...) if you have a faster kernel for some dtype.

---

CLI / demo / setup

python -m quill       # animated benchmark demo across data types
quill                 # same, if installed via pip
quill setup           # calibration wizard: measures the parallel/GPU crossovers
                      #   on YOUR box and persists them to ~/.quill/config.json

Quill also ships a small visualizer/wizard and add-on helpers layered on the core engine (the calibration wizard above, profile inspection via analyze(), and the quill_topk add-on). They are optional conveniences — the core sort API needs none of them.

---

How it works (the list path, in detail)

For a list, Quill profiles the data at intake and routes to the best strategy:

Data shape	Strategy	Complexity
Dense bounded integer range	np.bincount counting sort	O(n + k)
int64 / float64 (value-only)	fastest backend via the array chain	~O(n) radix / O(n log n)
Mixed int/float	promote to float64, then numeric path	O(n log n)
Strings / bytes	Python Timsort	O(n log n)
Objects with a key	extract keys → numpy argsort (stable)	O(n log n)
Nearly sorted (asc_ratio > 0.90)	Timsort bypass	~O(n)
Larger than RAM	external radix-partition + parallel bucket sort	disk-bound

The conversion to/from numpy is the fixed cost that caps the list-path speedup — see the very first table. For maximum throughput, keep your data in arrays and use sort_array.

---

Performance tuning

• Install the right extra. [fast] (numpy + psutil) is the baseline; add [polars] for a no-compile multi-threaded sort, or [gpu] if you have an NVIDIA card. The bundled Rust wheel already gives you the top path.
• Use sort_array on arrays. It skips the conversion wall entirely.
• stable=False on the list path for extra speed on numeric data when you don't need stable ordering.
• quill setup measures your machine's crossovers so auto-parallel and GPU engage exactly where they win.
• Install psutil for accurate RAM sensing (avoids spurious external-sort triggers); without it, Quill assumes 2 GB available.

---

Troubleshooting

• "quill_sort(list) isn't 5x faster." It can't be — the asarray/tolist round trip dominates (Amdahl). It is np.sort-fast and ~3-4x over sorted(). For the big win, keep arrays and call sort_array.
• "sort_array isn't using the GPU / Rust path." Check quill.available_backends(). The backend may not be installed, the array may be below its crossover (min_n), or the dtype/itemsize may be ineligible. Set QUILL_BACKEND_DEBUG=1 to surface the real reason instead of falling back silently.
• "GPU path errors on a huge array." It only engages when the array fits in free VRAM with headroom; otherwise it falls back to the CPU radix or np.sort.
• "Equal elements reordered." Use stable=True (the default). stable=False is explicitly allowed to reorder equal elements.
• "External sort triggered unexpectedly." Install psutil for accurate RAM sensing; pass high_performance_mode=True to skip the interactive prompt.

---

Development

git clone https://github.com/invariant-games/quill.git
cd quill
pip install -e ".[all,dev]"
pytest                          # full suite
pytest -m slow                  # parallel / large-data tests
python -m quill                 # animated benchmark demo

# Build the compiled Rust backend locally (optional; CI ships wheels):
cd rustext && maturin develop --release

The CI workflow .github/workflows/wheels.yml builds the rustext crate into abi3 binary wheels (one wheel per platform covers Python ≥ 3.8) for Windows, manylinux x86_64/aarch64, and macOS x86_64/arm64, plus a pure-Python sdist as the never-fail-install fallback.

---

Requirements

• Python 3.8+
• numpy — optional but strongly recommended (pip install quill-sort[fast])
• psutil — optional, for accurate RAM sensing
• polars — optional backend (pip install quill-sort[polars])
• cupy — optional GPU backend (pip install quill-sort[gpu], NVIDIA only)
• pandas — optional (Series/DataFrame support)

A C/Rust toolchain is not required to install — the compiled backend is shipped as a pre-built wheel, with a pure-Python sdist fallback.

---

License

MIT — by Isaiah Tucker