core.tensor_parallel.gtp_api

Generalized Tensor Parallelism (GTP) public API.

Thin facade over the core implementation in generalized_tensor_parallelism and CUDA-graph lifecycle support in gtp_cuda_graphs. GTP depends on TransformerEngine: if TE is missing or too old the core module imports cleanly but reports HAVE_TE = False, mirrored here as HAVE_GTP = False. Consumers gate every GTP code path behind if HAVE_GTP:, so no core module uses GTP symbols without TE.

Module Contents

Data

__all__

API

core.tensor_parallel.gtp_api.__all__

[‘HAVE_GTP’, ‘GTP_CONFIG’, ‘GTPChain’, ‘GTPEmbeddingWeight’, ‘attach_gtp_to_presharded_module’, ‘cla…