Slot: parallel_mode

How trial-axis parallelism is realised at codegen time. vmap batches all trials in parallel (fast, peak memory ~n_trials × per-trial working set). lax_map runs them sequentially via jax.lax.map (slower, peak memory bounded by one trial). pmap shards across devices for multi-GPU. auto (default) picks vmap when the estimated batched memory fits, lax_map otherwise.

URI: tvbo:slot/parallel_mode

Applicable Classes

Name Description Modifies Slot
Exploration Parameter space exploration (grid search, sweep) no

Properties

Type and Range

Property Value
Range ParallelMode
Domain Of Exploration

Cardinality and Requirements

Property Value

Slot Characteristics

Property Value
Owner Exploration

Identifier and Mapping Information

Schema Source

  • from schema: https://w3id.org/tvbo

Mappings

Mapping Type Mapped Value
self tvbo:parallel_mode
native tvbo:parallel_mode

LinkML Source

name: parallel_mode
description: How trial-axis parallelism is realised at codegen time. ``vmap`` batches
  all trials in parallel (fast, peak memory ~n_trials × per-trial working set). ``lax_map``
  runs them sequentially via ``jax.lax.map`` (slower, peak memory bounded by one trial).
  ``pmap`` shards across devices for multi-GPU. ``auto`` (default) picks ``vmap``
  when the estimated batched memory fits, ``lax_map`` otherwise.
from_schema: https://w3id.org/tvbo
rank: 1000
owner: Exploration
domain_of:
- Exploration
range: ParallelMode