[EPLB] Nixl communicator optimization. Zero-copy transfers (#41633)

Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Markov Ilya <markovilya19@gmail.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Markov Ilya <markovilya19@gmail.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
This commit is contained in:
Ilya Markov
2026-06-04 03:40:34 +00:00
committed by GitHub
co-authored by Markov Ilya Nicolò Lucchesi Tyler Michael Smith Tyler Michael Smith
parent f0cd590d62
commit 4f423bd5bc
12 changed files with 287 additions and 219 deletions
@@ -470,10 +470,14 @@ class ElasticEPScalingExecutor:
module._replace_quant_method(module.quant_method.old_quant_method)
prepare_communication_buffer_for_model(self.worker.model_runner.model)
eplb_model_state.expert_buffer = [
torch.empty_like(w) for w in model.expert_weights[0]
]
eplb_model_state.communicator = create_eplb_communicator(
group_coordinator=get_eplb_group(),
backend=parallel_config.eplb_config.communicator,
expert_weights=model.expert_weights[0],
expert_weights=model.expert_weights,
expert_buffer=eplb_model_state.expert_buffer,
)
if (