Mixed-Precision Training Explained.

Why BF16, FP16, and FP32 coexist in one training step, how autocast and loss scaling work, and which operations must stay precise.

This chapter follows the series' four-layer pyramid: intuition first, then consequences, system design, and finally implementation-level checks. It is written to be useful both as a first explanation and as a review sheet before reading the primary papers.

What you should be able to do after reading

• Explain the mechanism without relying on the feature name.
• Trace the relevant tensors, losses, or messages through one concrete example.
• Distinguish a paper claim from an inference, implementation choice, or marketing shorthand.
• Design a minimal experiment that could prove the idea wrong.

Where this chapter fits in the ten-phase map

Low precision touches every earlier mechanism: attention projections, cached representations, expert FFNs, routers, and MTP branches. A format that works for one matrix may fail in a reduction or collective. Phase 8 therefore treats precision as part of the distributed system, not a checkpoint conversion.

The dependency is useful when debugging. If the model-level equation is correct but the measured result is poor, walk backward through representation, numerical format, memory layout, routing or communication, and finally the evaluation harness. The first broken contract is usually more actionable than the final benchmark delta.

1. Range and precision are separate

FP16 has more fraction bits than BF16 but far less exponent range. BF16 shares FP32’s eight exponent bits, making overflow and underflow less likely in training even though each value has fewer significant digits. That is why BF16 often works without loss scaling.

2. Autocast chooses per operation

Matrix multiplications and convolutions benefit from low-precision tensor cores. Reductions, normalization statistics, softmax, and loss calculations often need FP32. A good implementation does not cast the entire model blindly; it uses an allowlist informed by numerical behavior.

3. Master state preserves small updates

If weights and optimizer moments lived only in FP16, tiny updates could round to zero. Many recipes compute low-precision forward and backward operations while maintaining FP32 optimizer state or master weights. BF16 changes the tradeoff but optimizer precision remains important.

4. Loss scaling rescues FP16 gradients

Multiplying the loss scales gradients upward before backpropagation, moving small values into FP16’s representable range. After unscaling, the optimizer checks for non-finite values and adjusts a dynamic scale. Scaling cannot repair overflow inside the forward pass.

5. Numerical debugging is a first-class task

Track non-finite counts, gradient norms, loss scale, activation ranges, and the first operation producing NaN. Re-running everything in FP32 can identify a precision cause; selectively promoting the failing reduction is better than abandoning mixed precision.

Engineering lens. For every concept above, identify the tensor, state, metric, or system boundary that makes it observable. Then ask which assumption would make the claim fail. This keeps the chapter testable instead of leaving it as architecture vocabulary.

Worked example

Compare 1e-8, 1.0, and 1e5 in FP16 and BF16. Explain which underflows or overflows, then trace a gradient of 1e-9 through a loss scale of 2^16 and back to its unscaled value.

Do the arithmetic with small dimensions first. Small examples expose index shifts, hidden assumptions, and missing denominators that disappear inside a billion-parameter headline. Once the hand-worked result is correct, automate it and compare the program output against the same values.

Implementation and measurement plan

Start from framework AMP defaults, keep reductions and optimizer updates safe, and add automated finite checks. Validate convergence against FP32 or BF16 baselines at equal tokens, not merely equal steps if throughput differs.

1. State the exact model, checkpoint, hardware, and date behind every numerical claim.
2. Separate algorithmic complexity, theoretical FLOPs, measured latency, memory, and end-to-end cost.
3. Build a small reference implementation before optimizing kernels or distributing it.
4. Compare against an equal-compute or equal-parameter baseline and report the denominator.
5. Record failure cases and scope limits beside the successful result.

From a paper claim to an engineering contract

The primary anchor for this chapter is DeepSeek-V3 Technical Report from DeepSeek-V3. Reading a number from that source is only the first step. A reproducible contract has four layers:

Layer	Question to write down	Evidence
Mechanism	What operation, loss, state, or routing decision changes?	Equation, pseudocode, tensor shapes
Implementation	How is it realized on the named hardware and software stack?	Kernel, precision, layout, process groups
Measurement	Which denominator and baseline make the comparison fair?	Raw metrics, config, repeated runs
Scope	Where should the claim stop being trusted?	Failure cases, ablations, dated limitations

This separation prevents a frequent error in frontier-model writing: converting a theoretical reduction into a latency promise, or converting one internal benchmark into a universal quality ranking. The implementation can fail to realize the algorithm, and the workload can fail to expose the intended benefit.

Failure modes and misleading shortcuts

• Casting LayerNorm statistics down can destabilize deep networks.
• Static loss scaling may fail as gradients evolve.
• BF16’s range does not eliminate rounding noise.
• FP32 master weights increase memory beyond the visible model dtype.
• Faster steps are useless if more tokens are needed to converge.

These are not footnotes. Frontier-model engineering is dominated by boundary conditions: a method can be mathematically correct and still lose to memory traffic, data skew, numerical drift, evaluation leakage, or a poorly stated comparison. A credible result makes those boundaries visible.

How to audit claims about this topic

Rewrite each claim with its missing boundary: name the exact mechanism, identify the tensor or resource it changes, and attach the workload and measurement. Then construct a counterexample at the edge of the claim. If a sentence cannot survive that rewrite, treat it as orientation—not evidence.

Next, trace provenance. Prefer the primary report for configuration and results, the released code for implementation behavior, and your own profiler for product performance. Secondary explainers are valuable for intuition but should not silently become the source of a numerical claim.

Decision guide: when should you use this idea?

Use it when the bottleneck named in the thesis appears in profiler traces or controlled quality experiments, the necessary kernels and runtime support exist, and the added system complexity can be observed in production. Start with the smallest configuration that exposes the bottleneck.

Delay it when a dense or higher-precision baseline does not yet converge, the evaluation harness is unstable, or the claimed resource is not limiting the workload. Sophisticated architecture cannot compensate for an invalid baseline.

Reject it when its benefit exists only under a denominator irrelevant to the product—for example, theoretical FLOPs while user latency worsens—or when numerical, safety, or operational regressions exceed the measured gain.

Hands-on study lab

1. 1. Inspect an autocast trace.
2. 2. Force a softmax overflow and repair it.
3. 3. Compare FP16 dynamic loss scaling with BF16.
4. 4. Calculate optimizer-state memory by dtype.

For each exercise, save the configuration, a tiny deterministic fixture, the raw measurements, and one failed case. The goal is not merely to make the code run; it is to make the conclusion independently checkable.

Quick self-check

What is the central idea?

Mixed precision is an operation-by-operation policy: use narrow formats where tensor cores and bandwidth matter, retain wider range or accumulation where small errors compound.

What is the most common reading mistake?

Casting LayerNorm statistics down can destabilize deep networks.

What evidence should I demand?

An exact configuration, a fair baseline, primary-source support, end-to-end measurements, and failure cases at the limits of the claim.

How do I explain it to a new engineer?

Begin with the bottleneck, show one tiny worked example, trace the changed state, and only then introduce the official name. Finish by naming one situation where the method will not help.

How do I review an implementation?

Check indexing and masks, parameter sharing, dtype transitions, layouts, process-group scope, raw metric denominators, and behavior under an adversarial or worst-case fixture. A passing happy-path shape test is not enough.

Teach-back synthesis

Close the page and reconstruct the argument in five sentences: the bottleneck; the mechanism; the state or tensor that changes; the fair measurement; and the main failure mode. Then reopen the page and compare. If you can repeat the feature names but cannot state those five sentences, revisit the worked example.

Finally, connect the idea to two neighboring phases. DeepSeek's advantage is not one isolated invention: compressed attention changes the cache, sparse experts change active compute, FP8 changes arithmetic and bandwidth, distributed schedules hide communication, and reasoning training spends the resulting capacity differently. The series becomes useful when those dependencies form one mental model.

Key takeaways

• Why BF16, FP16, and FP32 coexist in one training step, how autocast and loss scaling work, and which operations must stay precise.
• The mechanism, training recipe, runtime implementation, and measured product behavior are separate layers of evidence.
• Numbers remain meaningful only with their workload, precision, hardware, context length, and date attached.
• A small reproducible test is more valuable than a large uncheckable diagram.

Primary sources and further reading

• DeepSeek-V3. DeepSeek-V3 Technical Report.
• Micikevicius et al.. FP8 Formats for Deep Learning.
• PyTorch. Automatic Mixed Precision documentation.

Source note: explanations and worked examples here are original. Numerical claims are scoped to the linked reports; rapidly changing model comparisons are dated in the article itself.