[MagpieTTS][bugfix] defaults to force_map_dataset=True for validation datast to avoid duplicates by a factor of num_workers.

[XuesongYang]

Summary

Fix validation dataloader data duplication for Lhotse Shar datasets by adding force_map_dataset: true to the MagpieTTS validation config.

Problem

When using lhotse_shar data with force_map_dataset=False (the default), the validation dataloader uses an iterable dataset path that causes two compounding issues:

1. No DDP data partitioning -- The Lhotse sampler is created with rank=0, world_size=1 (hardcoded for iterable datasets), so every GPU independently iterates through the entire validation dataset instead of its 1/world_size share.

2. Worker-level data duplication -- Each DataLoader worker gets a full copy of the IterableDatasetWrapper and independently iterates all shards. With num_workers=N, data is duplicated N× per GPU.

Combined, each GPU processes num_workers × total_dataset_batches instead of the correct total_dataset_batches / world_size.

This design is intentional for training (infinite datasets with force_finite=False), where unique per-worker seeds and infinite repetition avoid explicit data splitting. But for finite validation (force_finite=True), it results in massive redundant computation and metrics computed on duplicated data.

Empirical Validation

Tested on LibriTTS dev-clean (5,620 records = 176 batches at batch_size=32, num_workers=2, quadratic_duration=null):

force_map_dataset	8 GPUs	16 GPUs	Expected
False (before)	352	352	num_workers × total = 2 × 176 = 352 (GPU-count independent)
True (after)	22	11	total / world_size = 176/8=22, 176/16=11 (proper DDP scaling)

With force_map_dataset=True:

• Validation iterations scale inversely with GPU count (correct DDP behavior)
• No worker duplication (map dataset dispatches work to workers without duplication)
• Reduction factor = num_workers × world_size (e.g., 2×8 = 16× fewer iterations on 8 GPUs)

Fix

Setting force_map_dataset: true in the validation config switches from iterable dataset to map dataset, where:

• The sampler uses the actual global_rank and world_size to partition data across GPUs
• The DataLoader manages worker dispatch without duplication

This follows the same pattern used in speechlm2/data/datamodule.py for validation/test dataloaders. Unit test at tests/collections/common/test_lhotse_dataloading.py::test_force_map_dataset validate the effectiveness.