Paralleelsus süvaõppes 2025/26 kevad

Practical 9

Pipeline Parallelism in PyTorch

---

Objective

By the end of this practical, students will:

• Understand pipeline execution
• Implement model partitioning across GPUs
• Apply micro-batching
• Observe pipeline efficiency (bubble problem)
• Understand differences between:
  • Naive Pipeline (Option A)
  • GPipe (Option B)
  • 1F1B (Option C)
• Analyze how scheduling affects performance

---

Provided Code Files

Students will work with three versions:

1) Naive Pipeline (Option A):

• practical9_OptionA.py Download

2) GPipe (Option B):

• practical9_OptionB.py Download

3) 1F1B (Option C):

• practical9_OptionC.py Download

Pipeline Scheduling Comparison

Method	Behavior	Efficiency
Naive	Sequential forward + backward	❌ Very Low
GPipe	Forward all → backward all	⚠️ Medium
1F1W	Interleaved forward/backward	✅ High

---

Part 1— Run and Observe (Option A)

Understand naive pipeline behavior

1) Instructions

• Run Option A code
• Observe:
  • Step time
  • GPU utilization (qualitatively)

2) Questions

• Q1) Are GPUs working in parallel?
• Q2) Where do you expect idle time?

---

Part 2— Analyze GPipe (Option B)

Understand pipeline fill and drain

1) Instructions

• Run Option B code
• Compare with Option A:
  • Execution time
  • Loss behavior

2) Questions

• Q1) What changed compared to Option A?
• Q2) Why do we separate forward and backward?
• Q3) Does this remove pipeline bubbles?

---

Part 3—Analyze 1F1B (Option C)

Understand overlapping execution

1) Instructions

• Run Option C code
• Compare with:
  • Option A
  • Option B

2) Questions

• Q1) When does backward start?
• Q2) What is different from GPipe?
• Q3) Why is this more efficient?

---

Part 4—Modify and Experiment

Understand impact of micro-batching

1) Change

• NUM_MICROBATCHES = 2, 4, 8

2) Run for each option:

• A
• B
• C

3) Record Results

Method	µB=2	µB=4	'µB=4
A
B
C

4) Questions

• Q1) Does increasing micro-batches always help?
• Q2) Which method benefits most?

---

Part 5—Code Understanding

1) Analyze Option C (1F1B)
Focus on this part:

if i > 0:
    prev_out = forward_outputs[i - 1]
    prev_target = targets[i - 1]

    loss = loss_fn(prev_out, prev_target)
    loss.backward()

2) Questions

• Q1) Why do we delay backward by one step?
• Q2) What would happen if we remove i > 0?
• Q3) What happens to pipeline overlap?

---

Final Discussion

For students, discuss:

• Why is Option A not true pipeline parallelism?
• What is the main limitation of GPipe?
• How does 1F1B reduce pipeline bubbles?
• Which approach would you use in large-scale systems?

---