原文 · 保持原样，未翻译原文 · 未翻译
In this tutorial, we implement NVIDIA cuML as a GPU-accelerated machine learning framework and build a practical workflow that demonstrates how RAPIDS can accelerate familiar data science and machine learning tasks. We begin by configuring the GPU environment and examining cuml.accel, which lets us accelerate existing scikit-learn workloads with minimal code changes, before moving to the native cuML API for direct CuPy and cuDF interoperability. We then benchmark CPU and GPU implementations of PCA, K-Means, nearest-neighbor search, logistic regression, random forests, and DBSCAN, while using synchronized timing to obtain meaningful performance measurements. We also build GPU-based manifold-learning and clustering pipelines with UMAP, t-SNE, HDBSCAN, and trustworthiness metrics; explore high-throughput forest inference with FIL; validate GPU-generated SHAP explanations; perform hyperparameter optimization with scikit-learn meta-estimators; and finally serialize trained models while examining portability between GPU and CPU environments.
import os
import sys
import time
import json
import shutil
import warnings
import subprocess
import importlib
import traceback
warnings.filterwarnings("ignore")
QUICK = False
SEED = 42
SCALE = 0.25 if QUICK else 1.0
N_MAIN = int(200_000 * SCALE)
D_MAIN = 64
N_RF = int(50_000 * SCALE)
D_RF = 32
N_NN_INDEX = int(50_000 * SCALE)
N_NN_QUERY = int(5_000 * SCALE)
N_DBSCAN = int(20_000 * SCALE)
N_MANIFOLD = int(60_000 * SCALE)
N_ACCEL = int(80_000 * SCALE)
RESULTS = []
NOTES = []
def banner(title):
line = "=" * 78
print(f"\n{line}\n {title}\n{line}", flush=True)
def section(title, fn, *args, **kwargs):
banner(title)
t0 = time.perf_counter()
try:
fn(*args, **kwargs)
except Exception:
print(f"[!] Section skipped due to an error:\n{traceback.format_exc()}")
print(f"[section wall time: {time.perf_counter() - t0:.1f}s]", flush=True)
def bootstrap():
if shutil.which("nvidia-smi") is None:
raise SystemExit(
"No NVIDIA GPU found. In Colab: Runtime > Change runtime type > GPU."
)
print(subprocess.run(
["nvidia-smi",
"--query-gpu=name,memory.total,compute_cap,driver_version",
"--format=csv"],
capture_output=True, text=True).stdout)
try:
import cuml
print("cuML already available — skipping install.")
except ImportError:
print("Installing RAPIDS cuML (this takes ~1-3 minutes)...")
pin = ""
try:
import cudf
major_minor = ".".join(cudf.__version__.split("+")[0].split(".")[:2])
pin = f"=={major_minor}.*"
print(f" Pinning to the preinstalled cuDF line: cuml-cu12{pin}")
except Exception:
print(" cuDF not found; installing the latest stable cuml-cu12.")
cmd = [sys.executable, "-m", "pip", "install", "-q",
"--extra-index-url=https://pypi.nvidia.com", f"cuml-cu12{pin}"]
print("$ " + " ".join(cmd))
rc = subprocess.run(cmd).returncode
if rc != 0:
raise SystemExit(
"pip install failed. Alternative that always works on Colab:\n"
" !git clone https://github.com/rapidsai/rapidsai-csp-utils.git\n"
" !python rapidsai-csp-utils/colab/pip-install.py"
)
importlib.invalidate_caches()
import cuml
import cupy
print(f"cuml {cuml.__version__}")
print(f"cupy {cupy.__version__}")
try:
import cudf
print(f"cudf {cudf.__version__}")
except Exception:
pass
import sklearn
print(f"sklearn {sklearn.__version__} (cuML requires scikit-learn >= 1.6)")
bootstrap()
import numpy as np
import cupy as cp
import cuml
import matplotlib.pyplot as plt
from cuml.datasets import make_classification as gpu_make_classification
from cuml.datasets import make_blobs as gpu_make_blobs
rng = np.random.RandomState(SEED)
cp.random.seed(SEED)
class Timer:
def __init__(self, label, sync=True):
self.label = label
self.sync = sync
def __enter__(self):
if self.sync:
cp.cuda.runtime.deviceSynchronize()
self.t0 = time.perf_counter()
return self
def __exit__(self, *exc):
if self.sync:
cp.cuda.runtime.deviceSynchronize()
self.dt = time.perf_counter() - self.t0
print(f" {self.label:<44s} {self.dt:8.3f}s")
return False
def to_numpy(a):
if isinstance(a, cp.ndarray):
return cp.asnumpy(a)
if hasattr(a, "to_numpy"):
return a.to_numpy()
return np.asarray(a)
def record(task, cpu_s, gpu_s):
RESULTS.append((task, cpu_s, gpu_s))
if cpu_s and gpu_s:
print(f" -> {task}: {cpu_s / gpu_s:.1f}x speedup\n")
ACCEL_SCRIPT = f'''
import time
import numpy as np
from sklearn.datasets import make_blobs
from sklearn.decomposition import PCA
from sklearn.cluster import KMeans
from sklearn.neighbors import NearestNeighbors
from sklearn.linear_model import Ridge
X, y = make_blobs(n_samples={N_ACCEL}, n_features=32, centers=12, random_state=0)
X = X.astype("float32"); y = y.astype("float32")
t0 = time.perf_counter()
PCA(n_components=8).fit_transform(X)
KMeans(n_clusters=12, n_init=1, random_state=0).fit(X)
NearestNeighbors(n_neighbors=8).fit(X[:{N_ACCEL // 2}]).kneighbors(X[:5000])
Ridge(alpha=1.0).fit(X, y)
Ridge(alpha=1.0, positive=True).fit(X[:5000], y[:5000])
print("MODELTIME %.3f" % (time.perf_counter() - t0))
'''
def demo_accel():
path = "/content/_accel_demo.py" if os.path.isdir("/content") else "_accel_demo.py"
with open(path, "w") as f:
f.write(ACCEL_SCRIPT)
def run(cmd, label):
print(f"\n$ {' '.join(cmd[1:])}")
t0 = time.perf_counter()
p = subprocess.run(cmd, capture_output=True, text=True)
wall = time.perf_counter() - t0
out = p.stdout + p.stderr
model_s = None
for line in out.splitlines():
if line.startswith("MODELTIME"):
model_s = float(line.split()[1])
print(out.strip()[:4000])
print(f"[{label}] model time = {model_s}s | process wall = {wall:.1f}s")
return model_s
cpu_s = run([sys.executable, path], "stock sklearn")
cmd = [sys.executable, "-m", "cuml.accel", "--profile", path]
gpu_s = run(cmd, "cuml.accel")
if gpu_s is None:
gpu_s = run([sys.executable, "-m", "cuml.accel", path], "cuml.accel")
record("cuml.accel (sklearn script, unmodified)", cpu_s, gpu_s)
NOTES.append(
"cuml.accel needed ZERO source changes; the profile table above shows "
"which calls ran on GPU and why Ridge(positive=True) fell back to CPU."
)
We configure the tutorial environment, define dataset sizes and benchmarking utilities, and verify that an NVIDIA GPU is available. We install and initialize RAPIDS cuML when necessary, set up CuPy and reproducibility controls, and create synchronized timing and result-tracking helpers. We also demonstrate cuml.accel by running an unmodified scikit-learn workload and comparing its CPU execution with GPU-accelerated execution.
def demo_native_api():
from cuml.preprocessing import StandardScaler
from cuml.model_selection import train_test_split
X, y = gpu_make_blobs(n_samples=50_000, n_features=8, centers=5,
random_state=SEED, dtype=np.float32)
print(f"cuml.datasets output lives on device: {type(X).__module__}, "
f"shape={X.shape}, dtype={X.dtype}")
try:
import cudf
df = cudf.DataFrame(X, columns=[f"f{i}" for i in range(X.shape[1])])
back = df.values
ptr_a = X.__cuda_array_interface__["data"][0]
ptr_b = back.__cuda_array_interface__["data"][0]
print(f"CuPy ptr = {hex(ptr_a)}")
print(f"cuDF->CuPy= {hex(ptr_b)}")
print("Same device pointer (true zero-copy)? ", ptr_a == ptr_b)
print("Note: a column-major DataFrame round trip may re-pack; what "
"matters is that no host (CPU) round trip ever happens.")
scaled = StandardScaler().fit_transform(df)
print(f"StandardScaler(cuDF) -> {type(scaled).__name__}")
except Exception as e:
print(f"cuDF interop skipped: {e}")
from cuml.decomposition import PCA
pca = PCA(n_components=3).fit(X)
print(f"default (mirrors input) -> {type(pca.transform(X)).__name__}")
with cuml.using_output_type("numpy"):
print(f"inside using_output_type() -> {type(pca.transform(X)).__name__}")
print(f"after the context manager -> {type(pca.transform(X)).__name__}")
NOTES.append(
"Keep output_type as CuPy/cuDF inside a pipeline; converting to NumPy "
"on every step forces a device->host copy and eats the speedup."
)
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.2, random_state=SEED)
print(f"train_test_split -> {Xtr.shape} / {Xte.shape}, still on device: "
f"{isinstance(Xtr, cp.ndarray)}")
We work directly with the native cuML API and explore how GPU-resident data moves between CuPy, cuDF, and cuML components. We inspect device pointers to understand zero-copy interoperability and use cuML output-type controls to manage whether results remain on the GPU or return as NumPy arrays. We also perform a GPU-native train-test split so that our data remains on the device throughout the workflow.
def demo_benchmarks():
from sklearn.decomposition import PCA as skPCA
from sklearn.cluster import KMeans as skKMeans, DBSCAN as skDBSCAN
from sklearn.neighbors import NearestNeighbors as skNN
from sklearn.linear_model import LogisticRegression as skLR
from sklearn.ensemble import RandomForestClassifier as skRF
from cuml.decomposition import PCA as cuPCA
from cuml.cluster import KMeans as cuKMeans, DBSCAN as cuDBSCAN
from cuml.neighbors import NearestNeighbors as cuNN
from cuml.linear_model import LogisticRegression as cuLR
from cuml.ensemble import RandomForestClassifier as cuRF
print(f"Generating {N_MAIN:,} x {D_MAIN} on the GPU...")
Xg, yg = gpu_make_classification(n_samples=N_MAIN, n_features=D_MAIN,
n_informative=32, n_classes=4,
random_state=SEED)
Xg = Xg.astype(cp.float32)
yg = yg.astype(cp.int32)
Xc, yc = cp.asnumpy(Xg), cp.asnumpy(yg)
print(f" device array: {Xg.nbytes / 1e6:.0f} MB\n")
print("PCA (n_components=16)")
with Timer("sklearn", sync=False) as t:
skPCA(n_components=16, random_state=SEED).fit_transform(Xc)
cpu = t.dt
with Timer("cuML") as t:
cuPCA(n_components=16, random_state=SEED).fit_transform(Xg)
record("PCA", cpu, t.dt)
print("KMeans (k=16)")
with Timer("sklearn", sync=False) as t:
skKMeans(n_clusters=16, n_init=1, max_iter=100,
random_state=SEED).fit(Xc)
cpu = t.dt
with Timer("cuML") as t:
cuKMeans(n_clusters=16, n_init=1, max_iter=100,
random_state=SEED).fit(Xg)
record("KMeans", cpu, t.dt)
print(f"NearestNeighbors k=16 ({N_NN_INDEX:,} index / {N_NN_QUERY:,} query)")
idx_g, q_g = Xg[:N_NN_INDEX], Xg[N_NN_INDEX:N_NN_INDEX + N_NN_QUERY]
idx_c, q_c = cp.asnumpy(idx_g), cp.asnumpy(q_g)
with Timer("sklearn (brute)", sync=False) as t:
skNN(n_neighbors=16, algorithm="brute", n_jobs=-1).fit(idx_c).kneighbors(q_c)
cpu = t.dt
with Timer("cuML") as t:
d_gpu, i_gpu = cuNN(n_neighbors=16).fit(idx_g).kneighbors(q_g)
record("NearestNeighbors", cpu, t.dt)
print("LogisticRegression (multinomial, lbfgs/QN)")
with Timer("sklearn", sync=False) as t:
sk_lr = skLR(max_iter=200, n_jobs=-1).fit(Xc, yc)
cpu = t.dt
with Timer("cuML") as t:
cu_lr = cuLR(max_iter=200).fit(Xg, yg)
record("LogisticRegression", cpu, t.dt)
print(f" accuracy sklearn={sk_lr.score(Xc, yc):.4f} "
f"cuML={float((cu_lr.predict(Xg) == yg).mean()):.4f} "
"(different solvers, so small differences are expected)\n")
print(f"RandomForestClassifier (100 trees, depth 12, {N_RF:,} x {D_RF})")
Xr_g, yr_g = gpu_make_classification(n_samples=N_RF, n_features=D_RF,
n_informative=16, n_classes=2,
random_state=SEED)
Xr_g = Xr_g.astype(cp.float32)
yr_g = yr_g.astype(cp.int32)
Xr_c, yr_c = cp.asnumpy(Xr_g), cp.asnumpy(yr_g)
with Timer("sklearn", sync=False) as t:
skRF(n_estimators=100, max_depth=12, n_jobs=-1,
random_state=SEED).fit(Xr_c, yr_c)
cpu = t.dt
with Timer("cuML") as t:
cu_rf = cuRF(n_estimators=100, max_depth=12, n_bins=128,
n_streams=4, random_state=SEED).fit(Xr_g, yr_g)
record("RandomForest (fit)", cpu, t.dt)
globals()["_RF_ARTIFACTS"] = (cu_rf, Xr_g, yr_g, Xr_c, yr_c)
print(f"DBSCAN ({N_DBSCAN:,} x 8)")
Xd_g, _ = gpu_make_blobs(n_samples=N_DBSCAN, n_features=8, centers=6,
cluster_std=0.6, random_state=SEED,
dtype=np.float32)
Xd_c = cp.asnumpy(Xd_g)
with Timer("sklearn", sync=False) as t:
lab_c = skDBSCAN(eps=0.9, min_samples=8, n_jobs=-1).fit_predict(Xd_c)
cpu = t.dt
with Timer("cuML") as t:
lab_g = cuDBSCAN(eps=0.9, min_samples=8).fit_predict(Xd_g)
record("DBSCAN", cpu, t.dt)
print(f" clusters found: sklearn={len(set(lab_c.tolist())) - 1}, "
f"cuML={len(set(cp.asnumpy(lab_g).tolist())) - 1}\n")
We benchmark scikit-learn and cuML implementations of PCA, K-Means, nearest neighbors, logistic regression, random forests, and DBSCAN. We generate datasets on the GPU, synchronize CUDA operations for fair timing, and record the speedup each accelerated algorithm achieves. We also compare model behavior and retain the trained cuML random forest so that we can reuse it later in the tutorial.
def demo_manifold():
from cuml.manifold import UMAP, TSNE
from cuml.metrics import trustworthiness
X, y = gpu_make_blobs(n_samples=N_MANIFOLD, n_features=48, centers=8,
cluster_std=1.6, random_state=SEED, dtype=np.float32)
print(f"data: {X.shape}")
embeddings = {}
for n_neighbors, min_dist in [(15, 0.1), (50, 0.0)]:
key = f"UMAP(n_neighbors={n_neighbors}, min_dist={min_dist})"
with Timer(key) as t:
emb = UMAP(n_neighbors=n_neighbors, min_dist=min_dist,
n_components=2, random_state=SEED).fit_transform(X)
sub = slice(0, min(5000, X.shape[0]))
tw = trustworthiness(X[sub], emb[sub], n_neighbors=10)
print(f" trustworthiness = {tw:.4f}")
embeddings[key] = (emb, t.dt, tw)
with Timer("TSNE(method='fft')") as t:
tsne_emb = TSNE(n_components=2, perplexity=30,
random_state=SEED).fit_transform(X)
embeddings["TSNE"] = (tsne_emb, t.dt, float("nan"))
best_key = max([k for k in embeddings if k.startswith("UMAP")],
key=lambda k: embeddings[k][2])
emb = embeddings[best_key][0]
print(f"\nClustering the '{best_key}' embedding with GPU HDBSCAN")
try:
from cuml.cluster import HDBSCAN
with Timer("HDBSCAN") as t:
hdb = HDBSCAN(min_cluster_size=max(int(50 * SCALE), 5),
min_samples=10, prediction_data=True).fit(emb)
labels = cp.asarray(hdb.labels_)
n_clusters = int(labels.max()) + 1
noise = float((labels == -1).mean())
print(f" clusters={n_clusters} noise fraction={noise:.3f}")
try:
from cuml.metrics.cluster import adjusted_rand_score
print(f" adjusted Rand index vs ground truth: "
f"{adjusted_rand_score(y, labels):.4f}")
except Exception as e:
print(f" ARI skipped: {e}")
try:
from cuml.cluster.hdbscan import all_points_membership_vectors
mv = all_points_membership_vectors(hdb)
print(f" soft-cluster membership matrix: {tuple(mv.shape)}")
except Exception as e:
print(f" soft clustering skipped: {e}")
except Exception as e:
print(f" HDBSCAN step skipped: {e}")
fig, axes = plt.subplots(1, 3, figsize=(16, 5))
keys = list(embeddings)[:3]
for ax, k in zip(axes, keys):
e = to_numpy(embeddings[k][0])
c = to_numpy(y)
ax.scatter(e[:, 0], e[:, 1], c=c, s=1.5, cmap="tab10", alpha=0.6)
ax.set_title(f"{k}\n{embeddings[k][1]:.2f}s", fontsize=9)
ax.set_xticks([]); ax.set_yticks([])
plt.suptitle("GPU manifold learning (colored by ground-truth cluster)")
plt.tight_layout(); plt.show()
We build an unsupervised GPU pipeline using UMAP and t-SNE to reduce high-dimensional data into two-dimensional embeddings. We evaluate UMAP configurations with the trustworthiness metric, select the strongest embedding, and apply HDBSCAN to identify clusters and noise points. We then visualize the resulting embeddings and compare their structures using the known ground-truth cluster labels.
def demo_fil():
from sklearn.ensemble import RandomForestClassifier as skRF
n = int(30_000 * SCALE)
Xg, yg = gpu_make_classification(n_samples=n, n_features=24,
n_informative=12, n_classes=2,
random_state=SEED)
Xg = Xg.astype(cp.float32)
Xc, yc = cp.asnumpy(Xg), cp.asnumpy(yg.astype(cp.int32))
print("Training a 200-tree sklearn forest on CPU (the model to be served)...")
sk_model = skRF(n_estimators=200, max_depth=10, n_jobs=-1,
random_state=SEED).fit(Xc, yc)
with Timer("sklearn.predict_proba (CPU)", sync=False) as t:
p_cpu = sk_model.predict_proba(Xc)[:, 1]
cpu = t.dt
fil = None
try:
from cuml.fil import ForestInference
except Exception as e:
print(f"cuml.fil unavailable on this build ({e}); skipping. "
"On newer stacks use the standalone nvForest library instead.")
return
for kwargs in ({"is_classifier": True, "output_type": "numpy"},
{"output_class": True, "output_type": "numpy"},
{}):
try:
fil = ForestInference.load_from_sklearn(sk_model, **kwargs)
print(f"Loaded into FIL with kwargs={kwargs or '{}'}")
break
except Exception as e:
print(f" load_from_sklearn(**{kwargs}) -> {type(e).__name__}: {e}")
if fil is None:
print("Could not load the forest into FIL on this build; skipping.")
return
try:
fil.optimize(batch_size=Xg.shape[0])
print("Ran fil.optimize() to auto-tune layout/chunk size for this batch.")
except Exception:
pass
fil.predict_proba(Xg[:1024])
with Timer("FIL predict_proba (GPU)") as t:
p_gpu = fil.predict_proba(Xg)
record("Forest inference (200 trees)", cpu, t.dt)
p_gpu = to_numpy(p_gpu)
p_gpu = p_gpu[:, 1] if p_gpu.ndim == 2 and p_gpu.shape[1] == 2 else p_gpu.ravel()
print(f" max |prob difference| vs sklearn: {np.abs(p_gpu - p_cpu).max():.2e} "
"(FIL defaults to float32, so ~1e-6 is normal)")
NOTES.append(
"FIL/nvForest is the piece that matters in production: the same trained "
"artifact, served with GPU-class throughput and no retraining."
)
We focus on accelerating inference for tree-based models after training. We train a scikit-learn random forest on the CPU, load it into the cuML Forest Inference Library when supported, and optimize the inference configuration for the current GPU batch size. We compare CPU and GPU prediction times and validate that the predicted probabilities remain numerically consistent.
def demo_explainer():
from cuml.linear_model import Ridge
from cuml.explainer import PermutationExplainer
n, d = int(20_000 * SCALE), 12
X = cp.asarray(rng.randn(n, d), dtype=cp.float32)
true_coef = cp.asarray(rng.uniform(-3, 3, size=d), dtype=cp.float32)
y = (X @ true_coef + 0.1 * cp.asarray(rng.randn(n), dtype=cp.float32))
model = Ridge(alpha=1e-3).fit(X, y)
coef = cp.asarray(model.coef_).ravel()
background = X[:200]
to_explain = X[200:220]
with Timer("PermutationExplainer (GPU)") as t:
expl = PermutationExplainer(model=model.predict, data=background,
random_state=SEED)
shap_values = expl.shap_values(to_explain)
shap_values = cp.asarray(shap_values)
analytic = (to_explain - background.mean(axis=0)) * coef
err = float(cp.abs(shap_values - analytic).max())
print(f" max |SHAP - analytical linear SHAP| = {err:.4f}")
print(" (permutation SHAP is sampling-based, so a small residual is "
"expected; the pattern must match)")
base = float(model.predict(background).mean())
recon = cp.asarray(shap_values).sum(axis=1) + base
actual = cp.asarray(model.predict(to_explain)).ravel()
print(f" additivity residual (mean |sum(phi)+base - f(x)|) = "
f"{float(cp.abs(recon - actual).mean()):.4f}")
imp = to_numpy(cp.abs(shap_values).mean(axis=0))
order = np.argsort(imp)[::-1]
plt.figure(figsize=(8, 3.5))
plt.bar(range(d), imp[order], color="#76b900")
plt.xticks(range(d), [f"f{i}" for i in order])
plt.ylabel("mean |SHAP|")
plt.title("GPU SHAP feature importance (cuml.explainer.PermutationExplainer)")
plt.tight_layout(); plt.show()
def demo_hpo():
from sklearn.model_selection import RandomizedSearchCV
from cuml.ensemble import RandomForestClassifier as cuRF
n = int(60_000 * SCALE)
X, y = gpu_make_classification(n_samples=n, n_features=24, n_informative=14,
n_classes=3, random_state=SEED)
X = cp.asnumpy(X.astype(cp.float32))
y = cp.asnumpy(y.astype(cp.int32))
param_dist = {
"n_estimators": [50, 100, 200],
"max_depth": [8, 12, 16],
"max_features": [0.3, 0.5, 0.8],
"n_bins": [64, 128, 256],
}
search = RandomizedSearchCV(
cuRF(random_state=SEED, n_streams=1),
param_distributions=param_dist,
n_iter=8, cv=3, n_jobs=1, random_state=SEED, verbose=0,
)
with Timer("RandomizedSearchCV over cuML RF (8 x 3 fits)", sync=True) as t:
search.fit(X, y)
print(f" best CV accuracy: {search.best_score_:.4f}")
print(f" best params : {json.dumps(search.best_params_)}")
NOTES.append(
"Because each fit is seconds instead of minutes, you can afford a real "
"search space instead of one hand-tuned guess."
)
We use cuML’s GPU-based permutation explainer to calculate SHAP values for a Ridge regression model and validate those explanations against the analytical linear solution. We test SHAP additivity and visualize feature importance to confirm that the computed attributions behave as expected. We also combine cuML estimators with scikit-learn’s RandomizedSearchCV to perform cross-validated hyperparameter optimization while fitting the model on the GPU.
def demo_persistence():
import pickle
art = globals().get("_RF_ARTIFACTS")
if art is None:
from cuml.ensemble import RandomForestClassifier as cuRF
Xg, yg = gpu_make_classification(n_samples=int(20_000 * SCALE),
n_features=16, n_classes=2,
random_state=SEED)
Xg = Xg.astype(cp.float32); yg = yg.astype(cp.int32)
model = cuRF(n_estimators=50, max_depth=10, random_state=SEED).fit(Xg, yg)
else:
model, Xg, yg, _, _ = art
before = to_numpy(model.predict(Xg[:1000]))
path = "/content/cuml_rf.pkl" if os.path.isdir("/content") else "cuml_rf.pkl"
with open(path, "wb") as f:
pickle.dump(model, f)
size_mb = os.path.getsize(path) / 1e6
with open(path, "rb") as f:
restored = pickle.load(f)
after = to_numpy(restored.predict(Xg[:1000]))
print(f" pickled model: {size_mb:.2f} MB at {path}")
print(f" predictions identical after round trip: {np.array_equal(before, after)}")
print(" cuML uses cloudpickle internally, so models trained under "
"cuml.accel can be loaded and used by plain scikit-learn on a "
"CPU-only machine.")
print(" SECURITY: never unpickle a model file from an untrusted source.")
def demo_summary():
rows = [(t, c, g) for (t, c, g) in RESULTS if c and g]
if not rows:
print("No comparable timings were collected.")
return
w = max(len(r[0]) for r in rows)
print(f"{'task'.ljust(w)} {'CPU (s)':>9} {'GPU (s)':>9} {'speedup':>9}")
print("-" * (w + 32))
for t, c, g in rows:
print(f"{t.ljust(w)} {c:9.3f} {g:9.3f} {c / g:8.1f}x")
labels = [r[0] for r in rows][::-1]
speeds = [r[1] / r[2] for r in rows][::-1]
plt.figure(figsize=(9, 0.55 * len(labels) + 2))
bars = plt.barh(labels, speeds, color="#76b900")
for b, s in zip(bars, speeds):
plt.text(b.get_width() * 1.02, b.get_y() + b.get_height() / 2,
f"{s:.1f}x", va="center", fontsize=9)
plt.axvline(1.0, color="grey", ls="--", lw=1)
plt.xscale("log")
plt.xlabel("speedup vs CPU (log scale, higher is better)")
plt.title(f"cuML {cuml.__version__} on this Colab GPU")
plt.tight_layout(); plt.show()
print("\nTakeaways")
for i, n in enumerate(NOTES, 1):
print(f" {i}. {n}")
print("""
Caveats worth internalizing:
* Speedups are size-dependent. Under ~10k rows, PCIe transfer and kernel
launch overhead usually dominate, and the CPU wins. Benchmark YOUR shapes.
* Always deviceSynchronize() before stopping a timer, or you time nothing.
* cuML matches scikit-learn's API, not its exact numerics: different solvers,
float32 defaults, and non-deterministic reductions produce small deltas.
* Multi-GPU / multi-node: swap cuml.X for cuml.dask.X with a LocalCUDACluster.
Where to go next:
* cuml.accel compatibility matrix : https://docs.nvidia.com/cuml/stable/cuml-accel/compatibility/
* Profiling accelerated code : %%cuml.accel.profile and %%cuml.accel.line_profile
* Multi-GPU guide : https://docs.nvidia.com/cuml/stable/dask_multigpu_guide/
* Walkthrough notebooks : https://github.com/NVIDIA/cuml/tree/main/notebooks
""")
_t_all = time.perf_counter()
section("1. cuml.accel — zero code change acceleration of stock scikit-learn",
demo_accel)
section("2. Native cuML API: cuDF/CuPy interop, zero-copy, output types",
demo_native_api)
section("3. CPU vs GPU benchmark harness", demo_benchmarks)
section("4. UMAP -> trustworthiness -> HDBSCAN pipeline", demo_manifold)
section("5. High-throughput forest inference (FIL / nvForest)", demo_fil)
section("6. GPU SHAP with cuml.explainer, validated analytically", demo_explainer)
section("7. Hyperparameter search over cuML estimators", demo_hpo)
section("8. Serialization and GPU -> CPU portability", demo_persistence)
section("9. Summary", demo_summary)
print(f"\nTotal tutorial wall time: {time.perf_counter() - _t_all:.1f}s")
We serialize a trained cuML random forest with pickle, restore it, and verify that its predictions remain unchanged after the round trip. We aggregate the CPU and GPU timing results collected throughout the tutorial and visualize the resulting speedups on a logarithmic chart. Finally, we run every tutorial section in sequence, print the accumulated practical takeaways, and report the total runtime of the complete workflow.
In conclusion, we implemented a comprehensive understanding of how NVIDIA cuML integrates GPU acceleration into both existing scikit-learn workflows and fully GPU-native machine learning pipelines. We compared computational performance across several core algorithms, managed device-resident data efficiently with CuPy and cuDF, evaluated unsupervised representations and clustering quality, accelerated tree-model inference, and generated interpretable SHAP explanations directly on the GPU. We also showed that familiar scikit-learn utilities such as RandomizedSearchCV can work alongside cuML estimators, preserving established machine learning development patterns while benefiting from GPU execution.