seqpro 0.13.0__tar.gz → 0.14.0__tar.gz

{seqpro-0.13.0 → seqpro-0.14.0}/.github/workflows/release.yaml

@@ -1,9 +1,6 @@
 name: Create release
 
 on:
-  push:
-    tags:
-      - '[0-9]+.[0-9]+.[0-9]+'
   workflow_dispatch:
     inputs:
       tag:
@@ -30,15 +27,11 @@ jobs:
       - uses: actions/checkout@v4
         with:
           fetch-depth: 0
+          ref: ${{ inputs.tag }}
 
       - name: Resolve tag
         id: tag
-        run: |
-          case "${{ github.event_name }}" in
-            push)          echo "tag=${GITHUB_REF#refs/tags/}" >> $GITHUB_OUTPUT ;;
-            workflow_dispatch) echo "tag=${{ inputs.tag }}" >> $GITHUB_OUTPUT ;;
-            workflow_call) echo "tag=${{ inputs.tag }}" >> $GITHUB_OUTPUT ;;
-          esac
+        run: echo "tag=${{ inputs.tag }}" >> $GITHUB_OUTPUT
 
       - name: Extract changelog section
         run: |

{seqpro-0.13.0 → seqpro-0.14.0}/CHANGELOG.md

@@ -1,3 +1,25 @@
+## 0.14.0 (2026-06-01)
+
+### Feat
+
+- **rag**: add Ragged.to_packed method and export to_packed
+- **rag**: add Numba-parallelized to_packed for flat layouts
+
+### Fix
+
+- **ci**: check out release tag and drop duplicate release trigger
+- **rag**: produce canonical list-of-records layout in to_packed
+
+### Refactor
+
+- **rag**: make Ragged.to_packed copy keyword-only; add multibyte-trailing test
+- **rag**: drop unused as_contiguous packing path in unbox
+
+### Perf
+
+- **rag**: microbenchmark to_packed throughput vs ak.to_packed
+- **rag**: use to_packed at internal ak.to_packed call sites
+
 ## 0.13.0 (2026-06-01)
 
 ### Feat

{seqpro-0.13.0 → seqpro-0.14.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: seqpro
-Version: 0.13.0
+Version: 0.14.0
 Classifier: Programming Language :: Rust
 Classifier: Programming Language :: Python :: Implementation :: CPython
 Classifier: Programming Language :: Python :: Implementation :: PyPy

seqpro-0.14.0/benchmarks/bench_to_packed.py

@@ -0,0 +1,187 @@
+"""Microbenchmark: seqpro.rag.to_packed vs ak.to_packed vs NumPy gather.
+
+Run in the bench env:
+    pixi run -e bench python benchmarks/bench_to_packed.py --out bench_to_packed
+
+Outputs <out>.csv and <out>_*.png. Sweeps n_rows, mean length, length
+distribution, source (RAM vs memmap), dtype itemsize, and thread count, and
+prints an effect-size ranking of which axes most affect the speedup ratio.
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import time
+from pathlib import Path
+
+import numpy as np
+
+
+def make_lengths(rng, n_rows, mean_len, distribution):
+    if distribution == "uniform":
+        lo, hi = max(1, mean_len // 2), mean_len * 2
+        return rng.integers(lo, hi + 1, size=n_rows)
+    # long-tailed: most rows short, a few very long (load imbalance)
+    lengths = rng.integers(1, max(2, mean_len // 4), size=n_rows)
+    n_big = max(1, n_rows // 100)
+    lengths[rng.choice(n_rows, n_big, replace=False)] = mean_len * 50
+    return lengths
+
+
+def build(rng, n_rows, mean_len, distribution, dtype, source, tmpdir):
+    import seqpro.rag as spr
+
+    lengths = make_lengths(rng, n_rows, mean_len, distribution).astype(np.int64)
+    total = int(lengths.sum())
+    if source == "memmap":
+        path = Path(tmpdir) / "data.dat"
+        data = np.memmap(path, dtype=dtype, mode="w+", shape=(total,))
+        data[:] = rng.integers(0, 255, size=total).astype(dtype)
+    else:
+        data = rng.integers(0, 255, size=total).astype(dtype)
+    rag = spr.Ragged.from_lengths(data, lengths)
+    # reorder so the layout is an (unpacked) ListArray, forcing a real gather
+    return rag[rng.permutation(n_rows)], total * np.dtype(dtype).itemsize
+
+
+def numpy_gather(rag):
+    import seqpro.rag as spr
+
+    parts = rag._parts
+    offs = parts.offsets
+    starts, stops = (offs[0], offs[1]) if offs.ndim == 2 else (offs[:-1], offs[1:])
+    lengths = stops - starts
+    out_off = np.empty(lengths.size + 1, dtype=np.int64)
+    out_off[0] = 0
+    np.cumsum(lengths, out=out_off[1:])
+    idx = np.repeat(starts - out_off[:-1], lengths) + np.arange(int(out_off[-1]))
+    return spr.Ragged.from_offsets(parts.data[idx], parts.shape, out_off)
+
+
+def timeit(fn, arg, repeats):
+    fn(arg)  # warm up (JIT, caches)
+    times = []
+    for _ in range(repeats):
+        t0 = time.perf_counter()
+        fn(arg)
+        times.append(time.perf_counter() - t0)
+    return float(np.median(times))
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument("--out", default="bench_to_packed")
+    p.add_argument("--repeats", type=int, default=5)
+    p.add_argument(
+        "--n-rows", type=int, nargs="+", default=[1000, 10000, 100000, 1000000]
+    )
+    p.add_argument("--mean-len", type=int, nargs="+", default=[16, 256, 4096])
+    p.add_argument("--distributions", nargs="+", default=["uniform", "longtail"])
+    p.add_argument("--sources", nargs="+", default=["ram", "memmap"])
+    p.add_argument("--dtypes", nargs="+", default=["uint8", "float64"])
+    p.add_argument("--threads", type=int, nargs="+", default=[int(os.cpu_count() or 1)])
+    args = p.parse_args()
+
+    import awkward as ak
+    import pandas as pd
+    import seqpro.rag as spr  # noqa: F401  (ensures kernel import)
+    import tempfile
+
+    rng = np.random.default_rng(0)
+    rows = []
+    with tempfile.TemporaryDirectory() as tmp:
+        for nthreads in args.threads:
+            import numba
+
+            numba.set_num_threads(nthreads)
+            impls = {
+                "seqpro": lambda r: r.to_packed(),
+                "ak": lambda r: spr.Ragged(ak.to_packed(r)),
+                "numpy": numpy_gather,
+            }
+            for n_rows in args.n_rows:
+                for mean_len in args.mean_len:
+                    for dist in args.distributions:
+                        for src in args.sources:
+                            for dt in args.dtypes:
+                                rag, nbytes = build(
+                                    rng, n_rows, mean_len, dist, dt, src, tmp
+                                )
+                                for name, fn in impls.items():
+                                    t = timeit(fn, rag, args.repeats)
+                                    rows.append(
+                                        dict(
+                                            impl=name,
+                                            n_rows=n_rows,
+                                            mean_len=mean_len,
+                                            distribution=dist,
+                                            source=src,
+                                            dtype=dt,
+                                            threads=nthreads,
+                                            seconds=t,
+                                            gbps=nbytes / t / 1e9,
+                                        )
+                                    )
+
+    df = pd.DataFrame(rows)
+    df.to_csv(f"{args.out}.csv", index=False)
+    _plots_and_summary(df, args.out)
+
+
+def _plots_and_summary(df, out):
+    import matplotlib
+
+    matplotlib.use("Agg")
+    import matplotlib.pyplot as plt
+    import seaborn as sns
+
+    # throughput vs n_rows, faceted by mean_len x dtype, hue impl
+    g = sns.relplot(
+        data=df,
+        x="n_rows",
+        y="gbps",
+        hue="impl",
+        style="source",
+        col="mean_len",
+        row="dtype",
+        kind="line",
+        marker="o",
+        facet_kws=dict(sharey=False),
+    )
+    g.set(xscale="log")
+    g.savefig(f"{out}_throughput.png", dpi=120)
+
+    # thread scaling for seqpro
+    if df["threads"].nunique() > 1:
+        sub = df[df.impl == "seqpro"]
+        gt = sns.relplot(
+            data=sub, x="threads", y="gbps", hue="mean_len", kind="line", marker="o"
+        )
+        gt.savefig(f"{out}_threads.png", dpi=120)
+    plt.close("all")
+
+    # effect-size ranking on speedup (seqpro / ak) — goal 2
+    wide = df.pivot_table(
+        index=["n_rows", "mean_len", "distribution", "source", "dtype", "threads"],
+        columns="impl",
+        values="gbps",
+    ).reset_index()
+    wide["speedup"] = wide["seqpro"] / wide["ak"]
+    print("\n=== Speedup (seqpro / ak) summary ===")
+    print(wide["speedup"].describe())
+    print("\n=== Mean speedup by axis (effect size) ===")
+    axes = ["n_rows", "mean_len", "distribution", "source", "dtype", "threads"]
+    spreads = {}
+    for ax in axes:
+        means = wide.groupby(ax)["speedup"].mean()
+        spreads[ax] = means.max() - means.min()
+        print(f"\n{ax}:")
+        print(means)
+    print("\n=== Axes ranked by speedup spread (most influential first) ===")
+    for ax, s in sorted(spreads.items(), key=lambda kv: -kv[1]):
+        print(f"  {ax:14s} spread={s:.2f}")
+
+
+if __name__ == "__main__":
+    main()