FSDP and ZeRO-style sharding
Use fsdp and zero-style sharding to move the distributed training production brief toward a defensible release.
An eight-GPU job is slower than one GPU and loses progress on failure.
This lesson isolates fsdp and zero-style sharding as one decision inside that system. The people affected are product users and operators; the learning data must carry event time, availability time, ownership, and version; and the operating envelope is eight GPUs with restart under 15 minutes.
Choose whether and how to use global batch at a declared prediction cutoff.
Measure decision utility alongside calibration, slice reliability, and system latency—not model score alone.
An eight-GPU job is slower than one GPU and loses progress on failure. An unsafe release must degrade to a named baseline or the last known-good version.
Build the mental model before the machinery.
The core move is to treat fsdp and zero-style sharding as a contract between data, a computation, and an action. Log per-rank data/compute/collective time, preserve complete training state, and keep a one-GPU equivalence test. The implementation becomes easier to debug once you can state which inputs exist, which state is learned, what output means, and what must remain invariant after serialization.
global batch
Define it in a hand-checkable form and name the prediction-time inputs.
FSDP
Connect it to the production metric and identify what it cannot guarantee.
tensor parallelism
Stress it with a slice, a temporal boundary, and a failure-safe alternative.
checkpoint resharding
Stress it with a slice, a temporal boundary, and a failure-safe alternative.
Lessons 1, 2 in this course.
Name every symbol. Check every shape.
Scaling efficiency is the central invariant for this lesson. The formula is useful only when its inputs match the production cutoff and its output maps to an action.
Scaling efficiency
| Symbol | Meaning / shape / unit |
|---|---|
W | worker/GPU count |
throughput_W | distributed samples per second |
throughput_1 | single-worker baseline |
Open derivation and numerical substitution
Start from the production quantity being optimized, substitute the observed values with their declared units, then isolate the model-controlled term. Preserve shape annotations at each step so broadcasting or aggregation cannot silently change the result.
- Write the named inputs: W, throughput_W, throughput_1.
- Substitute one small, hand-checkable batch before vectorizing.
- Calculate an independent reference value and compare within a declared tolerance.
# equation → code contract
inputs = validate_shapes_and_units(batch)
value = compute_c13(inputs)
assert is_finite(value)Calculate it small. Shape it realistically. Break it on purpose.
A result you can reproduce on paper
Two ranks average different local gradients and take identical optimizer steps.
- Write every input and unit.
- Substitute values into the scaling efficiency equation above.
- Compare the result to one simple baseline and explain the direction of the difference.
The same reasoning under real constraints
Scale a decoder to eight-GPU DDP, then shard parameters and optimizer state when one GPU cannot fit them.
The production record includes the data snapshot, transformation state, artifact identity, cutoff, score, decision, and the version of the policy that consumed it.
The attractive result you should reject
One rank skips a batch and misses an all-reduce, leaving every other rank hung.
Diagnostic: replay the smallest failing slice from immutable inputs, then compare each boundary rather than retuning the model.
Change one assumption and make the tradeoff visible.
This lab runs predefined TypeScript only. It never executes learner code. Use the slider, numeric input, reset, live text, or table—the computation is the same.
All-reduce scaling lab
Add workers and estimate delivered speedup after communication and stragglers.
Assumption: 120 ms local-compute step; ring communication and logarithmic straggler penalty.
Open nonvisual data table
| Item | Computed state | Interpretation |
|---|---|---|
| 1 workers | 120.3 ms | 1.0× |
| 2 workers | 125.7 ms | 1.9× |
| 4 workers | 131.4 ms | 3.7× |
| 8 workers | 137.8 ms | 7.0× |
| 16 workers | 145.6 ms | 13.2× |
Trace the complete operating path.
- 01
Validate and version global batch.
- 02
Compute fsdp and zero-style sharding from prediction-time-safe inputs.
- 03
Persist model, feature, and configuration identities together.
- 04
Serve or materialize behind explicit eight GPUs with restart under 15 minutes.
- 05
Join telemetry to mature outcomes and retain a rollback path.
Observability
Join service health, input quality, prediction distributions, slice behavior, and mature outcomes by exact version.
Cost
Measure storage, preprocessing, compute, queueing, and human review under a representative arrival pattern.
Failure modes
One rank skips a batch and misses an all-reduce, leaving every other rank hung. Add a detector, owner, mitigation, and stop condition for this class of failure.
Alternatives
Compare a rule, a simpler statistical baseline, and a different system boundary before adding model complexity.
Explain the contract, not just the vocabulary.
Fault-tolerant distributed trainer
Specify DDP correctness, scaling experiments, worker failure, and complete checkpoint recovery.
- Gradient/data equivalence
- 1/2/4-GPU efficiency report
- Kill-and-resume proof
- Hang/straggler runbook
Read primary material with a purpose.
Return to the opening failure.
Log per-rank data/compute/collective time, preserve complete training state, and keep a one-GPU equivalence test.
For this lesson, the release evidence is a hand-checked formal result, deterministic simulation output, a ≥80% checkpoint, the production rubric, and a named fallback. The course resolves when the system can produce efficient ddp, parallelism choices, elastic sampling, and resumable checkpoints.