Deep Learning › Training at Scale › Day 233

Hands-on lab — Day 233: Mixed Precision and Performance

Commands

Setup

pip install -r requirements/requirements.txt

Run

python3 examples/mixed_precision_and_performance_lib.py

Test

./tests/run_tests.sh

File tree

conftest.py
examples/mixed_precision_and_performance_lib.py
examples/test_mixed_precision_and_performance_lib.py
expected-output/examples-run.txt
expected-output/FIELDS.md
expected-output/measured-values.txt
expected-output/starter-run.txt
expected-output/test-run.txt
metadata.yml
README.md
requirements/requirements.txt
security.md
starter/mixed_precision_and_performance_lib.py
starter/test_mixed_precision_and_performance_lib.py
tests/run_tests.sh
tests/test_mixed_precision_and_performance_lib.py
troubleshooting.md

Lab README

Lab: Day 233 -- Mixed Precision and Performance

Lesson

Day number: 233 of 365. Course: Course05-SS02 (Deep Learning - Sequence Models and Transformers). Topic: Mixed Precision Training, Floating-Point Formats, and Loss Scaling.

Purpose

Build and test a modular mixed-precision simulation and VRAM memory profiling engine in Python. Implement precision memory allocation calculations for AdamW training states and construct a dynamic loss scaling engine with overflow detection.

Learning objectives

  • Calculate exact static GPU VRAM memory requirements for mixed-precision training.
  • Implement the dynamic loss scaling algorithm with backoff and growth intervals.
  • Analyze floating-point precision trade-offs across FP32, FP16, BF16, and FP8.
  • Verify numerical stability safeguards during training execution.

Prerequisites

  • Day 232 (GPUs and AI Hardware).
  • Python 3.11+ with PyTorch and NumPy.

Supported operating systems

  • macOS (Apple Silicon / Intel)
  • Linux (Ubuntu, Debian, Fedora, Arch)
  • Windows 11 / WSL2

Hardware requirements

  • 1+ CPU cores.
  • 1 GB RAM.
  • 100 MB disk space.

Required software

  • Python 3.11 or newer.
  • pip package manager.
  • virtualenv or venv module.

Free and open-source options

PyTorch is open-source software maintained by the Linux Foundation under a modified BSD license.

Installation

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements/requirements.txt

File structure

  • starter/mixed_precision_and_performance_lib.py: Student scaffold file.
  • examples/mixed_precision_and_performance_lib.py: Complete reference implementation.
  • tests/test_mixed_precision_and_performance_lib.py: Pytest automated validation suite.
  • expected-output/: Verified output logs and baseline values.

How to run

Execute the reference demonstration script:

python3 examples/mixed_precision_and_performance_lib.py

What the commands do

  • Calculates VRAM memory footprints across model parameter scales.
  • Simulates dynamic loss scaling adjustments.
  • Runs unit test assertions.

Expected output

Mixed Precision Demo: 7B Model Static VRAM = 104.31 GB, Scale after overflow = 32768.0

Validation steps

  1. Verify calculate_adamw_vram_gb allocates 16 bytes per model parameter.
  2. Confirm MockDynamicLossScaler halves scale upon overflow detection.
  3. Confirm MockDynamicLossScaler doubles scale after successful steps reach growth_interval.
  4. Ensure all unit test assertions pass.

Tests

Run the test runner script:

./tests/run_tests.sh

Cleanup

find . -type d -name "__pycache__" -exec rm -rf {} +
find . -type d -name ".pytest_cache" -exec rm -rf {} +

Troubleshooting

  • Scale overflow: Ensure scale factors remain within positive finite bounds ($S > 1.0$).

Security notes

All performance benchmarks execute locally in process memory.

Extension exercises

  1. Profile torch.compile speedup on a Transformer encoder block.
  2. Implement an FP8 simulated quantizer.

Expected output

FIELDS.md

# Expected Output Fields: Day 233

- `Static VRAM`: Total static VRAM memory in GB required for 7B parameters in mixed-precision AdamW.
- `Scale after overflow`: Adjusted loss scale factor following overflow detection.

examples-run.txt

Mixed Precision Demo: 7B Model Static VRAM = 104.31 GB, Scale after overflow = 32768.0

measured-values.txt

Static VRAM: 104.31 GB
Scale after overflow: 32768.0

starter-run.txt

Starter scaffold executed. Ready for student implementation.

test-run.txt

============================= test session starts ==============================
collected 3 items

tests/test_mixed_precision_and_performance_lib.py::test_adamw_vram_calculation PASSED [ 33%]
tests/test_mixed_precision_and_performance_lib.py::test_dynamic_loss_scaler_overflow PASSED [ 66%]
tests/test_mixed_precision_and_performance_lib.py::test_dynamic_loss_scaler_growth PASSED [100%]

============================== 3 passed in 0.42s ===============================

Source files

conftest.py (585 bytes)
"""Make this lab importable when the suite runs.

pytest puts the test file's own directory on sys.path, not the lab root, so a
test that does `from examples.<module> import ...` -- or imports a module that
lives in examples/ directly -- fails without this. Putting it here rather than
in run_tests.sh means a bare `pytest` from the lab root works too.
"""

import os
import sys

LAB = os.path.dirname(os.path.abspath(__file__))
for candidate in (LAB, os.path.join(LAB, "examples")):
    if os.path.isdir(candidate) and candidate not in sys.path:
        sys.path.insert(0, candidate)
examples/mixed_precision_and_performance_lib.py (2031 bytes)
from typing import Dict, Any

def calculate_adamw_vram_gb(num_params_billions: float) -> Dict[str, float]:
    params = num_params_billions * 1e9
    
    fp16_params_gb = (params * 2) / (1024 ** 3)
    fp16_grads_gb = (params * 2) / (1024 ** 3)
    fp32_master_gb = (params * 4) / (1024 ** 3)
    fp32_mom1_gb = (params * 4) / (1024 ** 3)
    fp32_mom2_gb = (params * 4) / (1024 ** 3)

    total_static_gb = fp16_params_gb + fp16_grads_gb + fp32_master_gb + fp32_mom1_gb + fp32_mom2_gb

    return {
        "params_billions": num_params_billions,
        "model_weights_gb": round(fp16_params_gb, 2),
        "gradients_gb": round(fp16_grads_gb, 2),
        "master_weights_gb": round(fp32_master_gb, 2),
        "optimizer_moments_gb": round(fp32_mom1_gb + fp32_mom2_gb, 2),
        "total_static_vram_gb": round(total_static_gb, 2)
    }

class MockDynamicLossScaler:
    def __init__(self, init_scale: float = 65536.0, growth_factor: float = 2.0,
                 backoff_factor: float = 0.5, growth_interval: int = 2000):
        self.scale = init_scale
        self.growth_factor = growth_factor
        self.backoff_factor = backoff_factor
        self.growth_interval = growth_interval
        self.successful_steps = 0

    def step(self, has_overflow: bool) -> float:
        if has_overflow:
            self.scale *= self.backoff_factor
            self.successful_steps = 0
        else:
            self.successful_steps += 1
            if self.successful_steps >= self.growth_interval:
                self.scale *= self.growth_factor
                self.successful_steps = 0
        return self.scale

def run_mixed_precision_demo():
    vram = calculate_adamw_vram_gb(7.0)
    scaler = MockDynamicLossScaler(init_scale=65536.0)
    s1 = scaler.step(has_overflow=False)
    s2 = scaler.step(has_overflow=True)

    print(f"Mixed Precision Demo: 7B Model Static VRAM = {vram['total_static_vram_gb']} GB, Scale after overflow = {s2}")
    return vram, s2

if __name__ == "__main__":
    run_mixed_precision_demo()
examples/test_mixed_precision_and_performance_lib.py (849 bytes)
import pytest
from examples.mixed_precision_and_performance_lib import calculate_adamw_vram_gb, MockDynamicLossScaler

def test_adamw_vram_calculation():
    # 7 billion parameters: 16 bytes per param -> 112 GB (in decimal) ~ 104.3 GB (in binary GiB)
    res = calculate_adamw_vram_gb(7.0)
    assert res["total_static_vram_gb"] > 100.0
    assert res["master_weights_gb"] == 2 * res["model_weights_gb"]

def test_dynamic_loss_scaler_overflow():
    scaler = MockDynamicLossScaler(init_scale=65536.0, backoff_factor=0.5)
    assert scaler.step(has_overflow=False) == 65536.0
    assert scaler.step(has_overflow=True) == 32768.0

def test_dynamic_loss_scaler_growth():
    scaler = MockDynamicLossScaler(init_scale=1000.0, growth_factor=2.0, growth_interval=2)
    scaler.step(has_overflow=False)
    assert scaler.step(has_overflow=False) == 2000.0
metadata.yml (444 bytes)
lesson_id: D233
day: 233
kind: lab
languages:
  - python
setup_commands:
  - 'pip install -r requirements/requirements.txt'
run_commands:
  - 'python3 examples/mixed_precision_and_performance_lib.py'
test_commands:
  - './tests/run_tests.sh'
cleanup_commands:
  - 'find . -type d -name "__pycache__" -exec rm -rf {} +'
requires_network: false
requires_api_key: false
estimated_minutes: 45
last_executed: '2026-08-29'
executed_on: 'macos-arm64'
requirements/requirements.txt (41 bytes)
torch>=2.2.0
numpy>=1.26.0
pytest>=8.0.0
starter/mixed_precision_and_performance_lib.py (701 bytes)
from typing import Dict, Any

def calculate_adamw_vram_gb(num_params_billions: float) -> Dict[str, float]:
    # TODO: Calculate static VRAM memory breakdown for mixed-precision AdamW
    pass

class MockDynamicLossScaler:
    def __init__(self, init_scale: float = 65536.0, growth_factor: float = 2.0,
                 backoff_factor: float = 0.5, growth_interval: int = 2000):
        self.scale = init_scale
        self.growth_factor = growth_factor
        self.backoff_factor = backoff_factor
        self.growth_interval = growth_interval
        self.successful_steps = 0

    def step(self, has_overflow: bool) -> float:
        # TODO: Implement dynamic scale factor adjustment
        pass
starter/test_mixed_precision_and_performance_lib.py (849 bytes)
import pytest
from examples.mixed_precision_and_performance_lib import calculate_adamw_vram_gb, MockDynamicLossScaler

def test_adamw_vram_calculation():
    # 7 billion parameters: 16 bytes per param -> 112 GB (in decimal) ~ 104.3 GB (in binary GiB)
    res = calculate_adamw_vram_gb(7.0)
    assert res["total_static_vram_gb"] > 100.0
    assert res["master_weights_gb"] == 2 * res["model_weights_gb"]

def test_dynamic_loss_scaler_overflow():
    scaler = MockDynamicLossScaler(init_scale=65536.0, backoff_factor=0.5)
    assert scaler.step(has_overflow=False) == 65536.0
    assert scaler.step(has_overflow=True) == 32768.0

def test_dynamic_loss_scaler_growth():
    scaler = MockDynamicLossScaler(init_scale=1000.0, growth_factor=2.0, growth_interval=2)
    scaler.step(has_overflow=False)
    assert scaler.step(has_overflow=False) == 2000.0
tests/run_tests.sh (227 bytes)
#!/usr/bin/env bash
set -euo pipefail
echo "========================================"
echo "Running Day 233 Lab Test Suite"
echo "========================================"
pytest tests/ -v
echo "All tests passed successfully."
tests/test_mixed_precision_and_performance_lib.py (849 bytes)
import pytest
from examples.mixed_precision_and_performance_lib import calculate_adamw_vram_gb, MockDynamicLossScaler

def test_adamw_vram_calculation():
    # 7 billion parameters: 16 bytes per param -> 112 GB (in decimal) ~ 104.3 GB (in binary GiB)
    res = calculate_adamw_vram_gb(7.0)
    assert res["total_static_vram_gb"] > 100.0
    assert res["master_weights_gb"] == 2 * res["model_weights_gb"]

def test_dynamic_loss_scaler_overflow():
    scaler = MockDynamicLossScaler(init_scale=65536.0, backoff_factor=0.5)
    assert scaler.step(has_overflow=False) == 65536.0
    assert scaler.step(has_overflow=True) == 32768.0

def test_dynamic_loss_scaler_growth():
    scaler = MockDynamicLossScaler(init_scale=1000.0, growth_factor=2.0, growth_interval=2)
    scaler.step(has_overflow=False)
    assert scaler.step(has_overflow=False) == 2000.0

Troubleshooting

Troubleshooting: Day 233 - Mixed Precision and Performance

Common Issues

  1. Scaler scale factor collapsing to zero:
    • Cause: Consecutive NaN gradients without learning rate warmup.
    • Fix: Lower base learning rate and add warmup.

Security notes

Security & Privacy: Day 233 - Mixed Precision and Performance

Security Guidance

  • All precision profiling computations execute locally in memory.