On this tutorial, we implement a complicated, sensible implementation of the NVIDIA Transformer Engine in Python, specializing in how mixed-precision acceleration might be explored in a sensible deep studying workflow. We arrange the setting, confirm GPU and CUDA readiness, try to put in the required Transformer Engine parts, and deal with compatibility points gracefully in order that the pocket book stays runnable even when the total extension can’t be constructed. As we transfer by way of every step, we construct instructor and pupil networks, evaluate a baseline PyTorch path with a Transformer Engine-enabled path, practice each fashions, benchmark their velocity and reminiscence utilization, and visualize the outcomes, giving us a transparent hands-on understanding of how performance-oriented coaching workflows are structured in apply.
import os
import sys
import json
import time
import math
import random
import shutil
import platform
import subprocess
import statistics
def run(cmd, test=True):
print(“n[RUN]”, ” “.be a part of(cmd))
consequence = subprocess.run(cmd, textual content=True, capture_output=True)
if consequence.stdout.strip():
print(consequence.stdout[-4000:])
if consequence.returncode != 0 and consequence.stderr.strip():
print(consequence.stderr[-4000:])
if test and consequence.returncode != 0:
elevate subprocess.CalledProcessError(consequence.returncode, cmd)
return consequence
def has_cmd(title):
return shutil.which(title) will not be None
run([sys.executable, “-m”, “pip”, “install”, “-q”, “–upgrade”, “pip”])
run([sys.executable, “-m”, “pip”, “install”, “-q”, “ninja”, “packaging”, “matplotlib”])
import torch
import torch.nn as nn
import torch.nn.useful as F
import matplotlib.pyplot as plt
assert torch.cuda.is_available(), “This pocket book wants a GPU runtime in Colab.”
gpu_name = torch.cuda.get_device_name(0)
cc_major, cc_minor = torch.cuda.get_device_capability(0)
cuda_runtime = torch.model.cuda
python_version = sys.model.break up()[0]
torch_version = torch.__version__
cuda_home = os.environ.get(“CUDA_HOME”, “/usr/native/cuda”)
nvcc_path = shutil.which(“nvcc”) or os.path.be a part of(cuda_home, “bin”, “nvcc”)
cudnn_header_candidates = [
os.path.join(cuda_home, “include”, “cudnn.h”),
“/usr/include/cudnn.h”,
“/usr/local/include/cudnn.h”,
]
nvcc_exists = os.path.exists(nvcc_path)
cudnn_header_exists = any(os.path.exists(p) for p in cudnn_header_candidates)
print(“=” * 120)
print(“ENVIRONMENT CHECK”)
print(“=” * 120)
print(json.dumps({
“python”: python_version,
“platform”: platform.platform(),
“torch”: torch_version,
“torch_cuda”: cuda_runtime,
“gpu_name”: gpu_name,
“compute_capability”: f”{cc_major}.{cc_minor}”,
“cuda_home”: cuda_home,
“nvcc_exists”: nvcc_exists,
“nvcc_path”: nvcc_path if nvcc_exists else None,
“cudnn_header_exists”: cudnn_header_exists,
}, indent=2))
print(“=” * 120)
We put together the Colab setting by importing the required Python libraries, defining a helper operate for executing shell instructions, and putting in the core dependencies for the tutorial. We then import PyTorch and Matplotlib, confirm {that a} GPU is out there, and gather key setting particulars, together with the GPU title, CUDA model, Python model, and toolkit paths. This offers us a transparent view of the system state earlier than we try any Transformer Engine set up or mannequin execution.
te_available = False
te_mode = “fallback”
te_import_error = None
attempt:
run([sys.executable, “-m”, “pip”, “install”, “-q”, “transformer_engine[core_cu12]”])
besides Exception as e:
print(“Core wheel set up failed:”, repr(e))
can_try_te_torch = nvcc_exists and cudnn_header_exists
if can_try_te_torch:
env = os.environ.copy()
env[“NVTE_FRAMEWORK”] = “pytorch”
env[“MAX_JOBS”] = “1”
env[“NVTE_BUILD_THREADS_PER_JOB”] = “1”
env[“CUDA_PATH”] = cuda_home
env[“CUDA_HOME”] = cuda_home
attempt:
print(“nAttempting to construct the PyTorch extension for Transformer Engine…”)
consequence = subprocess.run(
[sys.executable, “-m”, “pip”, “install”, “-q”, “–no-build-isolation”, “transformer_engine[pytorch]”],
textual content=True,
capture_output=True,
env=env,
)
if consequence.stdout.strip():
print(consequence.stdout[-4000:])
if consequence.returncode != 0 and consequence.stderr.strip():
print(consequence.stderr[-4000:])
if consequence.returncode == 0:
import transformer_engine.pytorch as te
from transformer_engine.frequent import recipe
te_available = True
te_mode = “transformer_engine”
else:
te_import_error = consequence.stderr[-4000:] if consequence.stderr else “Unknown pip construct error”
besides Exception as e:
te_import_error = repr(e)
else:
te_import_error = “Lacking nvcc or cuDNN headers on this Colab runtime, so TE PyTorch extension can’t be constructed right here.”
if te_available:
attempt:
fp8_available, fp8_reason = te.is_fp8_available(return_reason=True)
besides Exception as e:
fp8_available, fp8_reason = False, f”Couldn’t question FP8 availability: {e}”
attempt:
bf16_available = te.is_bf16_available()
besides Exception:
bf16_available = torch.cuda.is_bf16_supported()
else:
fp8_available = False
fp8_reason = “Transformer Engine not put in; utilizing fallback PyTorch path.”
bf16_available = torch.cuda.is_bf16_supported()
amp_dtype = torch.bfloat16 if bf16_available else torch.float16
print(“n” + “=” * 120)
print(“INSTALL STATUS”)
print(“=” * 120)
print(json.dumps({
“te_available”: te_available,
“te_mode”: te_mode,
“fp8_available”: fp8_available,
“fp8_reason”: fp8_reason,
“te_import_error”: te_import_error,
“amp_dtype”: str(amp_dtype),
}, indent=2))
print(“=” * 120)
system = “cuda”
random.seed(42)
torch.manual_seed(42)
torch.cuda.manual_seed_all(42)
if te_available:
fp8_recipe = recipe.DelayedScaling(margin=0, fp8_format=recipe.Format.E4M3)
def baseline_autocast():
return torch.autocast(device_type=”cuda”, dtype=amp_dtype)
def te_forward_context(use_fp8):
if te_available and use_fp8:
return te.autocast(enabled=True, recipe=fp8_recipe)
return baseline_autocast()
We try to put in the Transformer Engine core bundle after which test whether or not the Colab runtime can construct the PyTorch extension by verifying the presence of nvcc and cuDNN headers. If the setting helps it, we attempt to set up the Transformer Engine PyTorch backend after which examine whether or not FP8 and BF16 can be found on the present {hardware}. We additionally configure the precision mode and outline the autocast contexts that later enable us to modify between normal blended precision and Transformer Engine execution.
class TeacherNet(nn.Module):
def __init__(self, hidden_size=512, intermediate_size=2048, num_layers=3, vocab_size=4096):
tremendous().__init__()
self.embed = nn.Embedding(vocab_size, hidden_size)
self.layers = nn.ModuleList([
nn.Sequential(
nn.LayerNorm(hidden_size),
nn.Linear(hidden_size, intermediate_size),
nn.GELU(),
nn.Linear(intermediate_size, hidden_size),
) for _ in range(num_layers)
])
self.head = nn.Linear(hidden_size, hidden_size)
def ahead(self, token_ids):
x = self.embed(token_ids)
for layer in self.layers:
x = x + layer(x)
return self.head(x)
class BaselineStudent(nn.Module):
def __init__(self, hidden_size=512, intermediate_size=2048, num_layers=3, vocab_size=4096):
tremendous().__init__()
self.embed = nn.Embedding(vocab_size, hidden_size)
self.norms = nn.ModuleList([nn.LayerNorm(hidden_size) for _ in range(num_layers)])
self.fc1 = nn.ModuleList([nn.Linear(hidden_size, intermediate_size) for _ in range(num_layers)])
self.fc2 = nn.ModuleList([nn.Linear(intermediate_size, hidden_size) for _ in range(num_layers)])
self.head = nn.Linear(hidden_size, hidden_size)
def ahead(self, token_ids):
x = self.embed(token_ids)
for ln, fc1, fc2 in zip(self.norms, self.fc1, self.fc2):
residual = x
x = ln(x)
x = fc1(x)
x = F.gelu(x, approximate=”tanh”)
x = fc2(x)
x = x + residual
return self.head(x)
if te_available:
class TEStudent(nn.Module):
def __init__(self, hidden_size=512, intermediate_size=2048, num_layers=3, vocab_size=4096):
tremendous().__init__()
self.embed = nn.Embedding(vocab_size, hidden_size)
self.norms = nn.ModuleList([te.LayerNorm(hidden_size) for _ in range(num_layers)])
self.fc1 = nn.ModuleList([te.Linear(hidden_size, intermediate_size, bias=True) for _ in range(num_layers)])
self.fc2 = nn.ModuleList([te.Linear(intermediate_size, hidden_size, bias=True) for _ in range(num_layers)])
self.head = te.Linear(hidden_size, hidden_size, bias=True)
def ahead(self, token_ids, use_fp8=False):
x = self.embed(token_ids)
with te_forward_context(use_fp8):
for ln, fc1, fc2 in zip(self.norms, self.fc1, self.fc2):
residual = x
x = ln(x)
x = fc1(x)
x = F.gelu(x, approximate=”tanh”)
x = fc2(x)
x = x + residual
x = self.head(x)
return x
else:
class TEStudent(nn.Module):
def __init__(self, hidden_size=512, intermediate_size=2048, num_layers=3, vocab_size=4096):
tremendous().__init__()
self.embed = nn.Embedding(vocab_size, hidden_size)
self.norms = nn.ModuleList([nn.LayerNorm(hidden_size) for _ in range(num_layers)])
self.fc1 = nn.ModuleList([nn.Linear(hidden_size, intermediate_size) for _ in range(num_layers)])
self.fc2 = nn.ModuleList([nn.Linear(intermediate_size, hidden_size) for _ in range(num_layers)])
self.head = nn.Linear(hidden_size, hidden_size)
def ahead(self, token_ids, use_fp8=False):
x = self.embed(token_ids)
with baseline_autocast():
for ln, fc1, fc2 in zip(self.norms, self.fc1, self.fc2):
residual = x
x = ln(x)
x = fc1(x)
x = F.gelu(x, approximate=”tanh”)
x = fc2(x)
x = x + residual
x = self.head(x)
return x
def count_params(mannequin):
return sum(p.numel() for p in mannequin.parameters() if p.requires_grad)
def format_millions(n):
return f”{n / 1e6:.2f}M”
We outline the neural community architectures used all through the tutorial, together with the instructor mannequin, the baseline pupil mannequin, and the Transformer Engine pupil path. We hold the mannequin constructions aligned in order that the comparability stays significant whereas permitting the TE path to swap in Transformer Engine layers when the extension is out there. We additionally outline small utility features for counting parameters and formatting mannequin dimension, which assist us examine the size of the fashions earlier than coaching begins.
hidden_size = 512
intermediate_size = 2048
num_layers = 3
vocab_size = 4096
seq_len = 128
batch_size = 8
steps = 25
benchmark_iters = 20
lr = 2e-4
weight_decay = 1e-2
instructor = TeacherNet(hidden_size, intermediate_size, num_layers, vocab_size).to(system).eval()
baseline_model = BaselineStudent(hidden_size, intermediate_size, num_layers, vocab_size).to(system)
te_model = TEStudent(hidden_size, intermediate_size, num_layers, vocab_size).to(system)
optimizer_baseline = torch.optim.AdamW(baseline_model.parameters(), lr=lr, weight_decay=weight_decay)
optimizer_te = torch.optim.AdamW(te_model.parameters(), lr=lr, weight_decay=weight_decay)
print(“Instructor params :”, format_millions(count_params(instructor)))
print(“Baseline params:”, format_millions(count_params(baseline_model)))
print(“TE-path params :”, format_millions(count_params(te_model)))
def make_batch(batch_size, seq_len, vocab_size, system):
tokens = torch.randint(0, vocab_size, (batch_size, seq_len), system=system)
with torch.no_grad():
goal = instructor(tokens)
return tokens, goal
def peak_mem_mb():
return torch.cuda.max_memory_allocated() / (1024 ** 2)
def train_baseline_step():
baseline_model.practice()
optimizer_baseline.zero_grad(set_to_none=True)
tokens, goal = make_batch(batch_size, seq_len, vocab_size, system)
with baseline_autocast():
pred = baseline_model(tokens)
loss = F.mse_loss(pred, goal)
loss.backward()
optimizer_baseline.step()
return float(loss.detach().merchandise())
def train_te_step(use_fp8):
te_model.practice()
optimizer_te.zero_grad(set_to_none=True)
tokens, goal = make_batch(batch_size, seq_len, vocab_size, system)
pred = te_model(tokens, use_fp8=use_fp8)
loss = F.mse_loss(pred, goal)
loss.backward()
optimizer_te.step()
return float(loss.detach().merchandise())
We set the principle experiment hyperparameters, instantiate all fashions on the GPU, and create the optimizers that will likely be used throughout coaching. We additionally print the parameter counts to substantiate that the baseline and TE paths are comparable when it comes to mannequin dimension. As well as, we outline the batch-generation logic, reminiscence monitoring operate, and the person training-step features that execute one optimization step for every mannequin path.
baseline_losses = []
te_losses = []
mode_name = “TE-FP8” if (te_available and fp8_available) else (“TE-BF16/FP16” if te_available else “Fallback-PyTorch”)
print(“n” + “=” * 120)
print(“TRAINING”)
print(“=” * 120)
for step in vary(1, steps + 1):
b_loss = train_baseline_step()
t_loss = train_te_step(use_fp8=fp8_available)
baseline_losses.append(b_loss)
te_losses.append(t_loss)
if step == 1 or step % 5 == 0 or step == steps:
print(f”step={step:02d} | baseline_loss={b_loss:.6f} | te_path_loss={t_loss:.6f} | mode={mode_name}”)
@torch.no_grad()
def evaluate_model(mannequin, is_te=False, use_fp8=False, eval_batches=8):
mannequin.eval()
vals = []
for _ in vary(eval_batches):
tokens, goal = make_batch(batch_size, seq_len, vocab_size, system)
if is_te:
pred = mannequin(tokens, use_fp8=use_fp8)
else:
with baseline_autocast():
pred = mannequin(tokens)
vals.append(float(F.mse_loss(pred, goal).merchandise()))
return sum(vals) / len(vals)
baseline_eval = evaluate_model(baseline_model, is_te=False)
te_eval = evaluate_model(te_model, is_te=True, use_fp8=fp8_available)
def benchmark_train_step(mannequin, optimizer, is_te=False, use_fp8=False, warmup=5, iters=20):
times_ms = []
mems_mb = []
for _ in vary(warmup):
optimizer.zero_grad(set_to_none=True)
tokens, goal = make_batch(batch_size, seq_len, vocab_size, system)
if is_te:
pred = mannequin(tokens, use_fp8=use_fp8)
else:
with baseline_autocast():
pred = mannequin(tokens)
loss = F.mse_loss(pred, goal)
loss.backward()
optimizer.step()
torch.cuda.synchronize()
for _ in vary(iters):
torch.cuda.reset_peak_memory_stats()
optimizer.zero_grad(set_to_none=True)
tokens, goal = make_batch(batch_size, seq_len, vocab_size, system)
begin = time.perf_counter()
if is_te:
pred = mannequin(tokens, use_fp8=use_fp8)
else:
with baseline_autocast():
pred = mannequin(tokens)
loss = F.mse_loss(pred, goal)
loss.backward()
optimizer.step()
torch.cuda.synchronize()
finish = time.perf_counter()
times_ms.append((finish – begin) * 1000.0)
mems_mb.append(peak_mem_mb())
return {
“mean_ms”: statistics.imply(times_ms),
“median_ms”: statistics.median(times_ms),
“max_memory_mb”: max(mems_mb),
}
baseline_bench = benchmark_train_step(baseline_model, optimizer_baseline, is_te=False, use_fp8=False, iters=benchmark_iters)
te_bench = benchmark_train_step(te_model, optimizer_te, is_te=True, use_fp8=fp8_available, iters=benchmark_iters)
We run the principle coaching loop for each the baseline mannequin and the TE path, monitoring their losses over a number of steps. We then outline and execute the analysis operate to measure how nicely every mannequin matches the instructor’s outputs after coaching. Lastly, we implement the benchmarking routine to measure per-step runtime and peak CUDA reminiscence utilization, enabling quantitative comparability of efficiency traits.
abstract = {
“gpu_name”: gpu_name,
“compute_capability”: f”{cc_major}.{cc_minor}”,
“te_available”: te_available,
“fp8_available”: fp8_available,
“fp8_reason”: fp8_reason,
“mode”: mode_name,
“baseline_eval_mse”: baseline_eval,
“te_path_eval_mse”: te_eval,
“baseline_mean_step_ms”: baseline_bench[“mean_ms”],
“te_path_mean_step_ms”: te_bench[“mean_ms”],
“baseline_peak_mem_mb”: baseline_bench[“max_memory_mb”],
“te_path_peak_mem_mb”: te_bench[“max_memory_mb”],
}
print(“n” + “=” * 120)
print(“SUMMARY”)
print(“=” * 120)
print(json.dumps(abstract, indent=2))
plt.determine(figsize=(10, 5))
plt.plot(baseline_losses, label=”Baseline loss”)
plt.plot(te_losses, label=f”{mode_name} loss”)
plt.xlabel(“Coaching step”)
plt.ylabel(“MSE loss”)
plt.title(“Coaching Loss Comparability”)
plt.legend()
plt.grid(True)
plt.present()
plt.determine(figsize=(8, 5))
plt.bar([“Baseline”, mode_name], [baseline_bench[“mean_ms”], te_bench[“mean_ms”]])
plt.ylabel(“Imply practice step time (ms)”)
plt.title(“Velocity Comparability”)
plt.grid(True, axis=”y”)
plt.present()
plt.determine(figsize=(8, 5))
plt.bar([“Baseline”, mode_name], [baseline_bench[“max_memory_mb”], te_bench[“max_memory_mb”]])
plt.ylabel(“Peak reminiscence (MB)”)
plt.title(“Peak CUDA Reminiscence Comparability”)
plt.grid(True, axis=”y”)
plt.present()
We collect all ultimate metrics right into a abstract dictionary and print the experiment’s consolidated leads to a structured format. We then generate visualizations of coaching loss, imply training-step time, and peak reminiscence utilization to extra intuitively interpret the variations between the baseline and TE paths. This ultimate part helps us transfer from uncooked numbers to sensible insights by exhibiting how the 2 implementations behave throughout accuracy, velocity, and reminiscence.
In conclusion, we constructed way over a easy set up walkthrough; we created a whole experimental pipeline that helps us perceive how the NVIDIA Transformer Engine matches into trendy GPU-accelerated mannequin coaching. We examined the runtime setting, tailored to Colab limitations, preserved a working fallback path, after which educated, evaluated, and benchmarked two implementations facet by facet to watch sensible variations in effectivity, precision conduct, and useful resource utilization. On the finish, we understood the right way to use the Transformer Engine in a Colab-friendly setting and gained a reusable basis that we are able to prolong to bigger transformer architectures, richer benchmarking eventualities, and extra production-oriented optimization workflows.
Try the Full Codes/Pocket book right here. Additionally, be at liberty to comply with us on Twitter and don’t overlook to affix our 120k+ ML SubReddit and Subscribe to our E-newsletter. Wait! are you on telegram? now you possibly can be a part of us on telegram as nicely.
Have to associate with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and many others.? Join with us
