This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3
NCC_EXSP001#
The combined memory needed for the model’s activation tensors exceeds the high-bandwidth memory limit.#
There are several ways to potentially fix this issue.
Simply reduce the batch/tensor size if possible
Utilize pipeline/tensor parallelism via neuronx-distributed
Short snippet of tensor parallelism:
class ParallelSelfAttention(transformers.models.bert.modeling_bert.BertSelfAttention):
def __init__(self, config, position_embedding_type=None):
super().__init__(config, position_embedding_type)
self.query = ColumnParallelLinear(config.hidden_size,
self.all_head_size,
gather_output=False)
self.key = ColumnParallelLinear(config.hidden_size,
self.all_head_size,
gather_output=False)
self.value = ColumnParallelLinear(config.hidden_size,
self.all_head_size,
gather_output=False)
# Since we shard the number of attention heads across tensor parallel
# ranks, each rank would have a subset of heads, hence, we update
# the num_attention_heads here.
tp_size = parallel_state.get_tensor_parallel_size()
self.num_attention_heads = self.num_attention_heads // tp_size
self.all_head_size = self.all_head_size // tp_size
For more information:
This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3