[misc]: surface GenerationResult.peak_memory_mb in the MiniMax H3 examples

[KyleNeverGivesUp]

Problem

Every generation already computes its own peak. The worker calls torch.cuda.max_memory_allocated() at fastvideo/worker/multiproc_executor.py:699 and the value reaches the caller as GenerationResult.peak_memory_mb. Both MiniMax H3 examples receive it and print only the generation time.

On a memory-constrained device that number is the one that decides whether a configuration runs at all. It is also what a reader has to reproduce in order to check any memory claim about this model, and right now they cannot: the field is populated, handed over, and dropped on the floor.

examples/inference/optimizations/spark_benchmark.py already prints it, so this only brings the H3 examples in line with an existing convention.

What this changes

Two lines in each example, guarded on the value being present so nothing changes for a backend that does not report one.

Generation time: 902.31s
Peak memory: 70880.0 MB

basic_minimax_h3_t2v.py prints it for the first request and for each --repeats iteration, next to the generation-time line those iterations already emit. basic_fasth3.py prints it inside the same per-run block that already reports wall time, generation time and denoise time.

Why it matters right now

Two open PRs report peak memory on one GB10 and neither is reproducible without a local patch to these files:

• [perf]: MiniMax H3 on GB10 - load pipeline components on demand instead of all at once #1761 reports component residency and the ceiling between 222,208 and 347,200 packed tokens.
• [bugfix]: apply FP8 to MiniMax H3's feed-forward #1780 reports 44.23 GiB of resident weights falling to 32.99 GiB once FP8 reaches the feed-forward stack.

Every one of those numbers was read from a three-line local edit that has never been committed. With this change a reader can run the same command and see the same line.

Testing

No test parses the output of either example, so nothing is invalidated. grep -rn "Generation time:" fastvideo/tests/ returns nothing.

Checked on one GB10, 768x1344, 124 frames, MiniMax H3 r16:

Generation time: 902.31s
Peak memory: 70880.0 MB

which matches torch.cuda.max_memory_allocated() as reported by the worker for the same run.

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews and 🤖 CI

	Protection	Waiting on
🔴	PR merge requirements	👀 reviews and 🤖 CI

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=full-suite-passed

This rule is failing.

• #approved-reviews-by>=1
• check-success=full-suite-passed
• check-success=fastcheck-passed
• check-success~=pre-commit
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]