[KV-Offloading][TP] : Expand replicated_layout detection to multi-group MLA (#57652)
This commit is contained in:
@@ -658,6 +658,7 @@ class DeepseekV32IndexerCache(torch.nn.Module, AttentionLayerBase):
|
||||
return MLAAttentionSpec(
|
||||
block_size=self.cache_config.block_size,
|
||||
num_kv_heads=1,
|
||||
max_tp_shards=1,
|
||||
head_size=self.head_dim,
|
||||
dtype=self.dtype,
|
||||
cache_role=SparseCacheRole.INDEXER,
|
||||
|
||||
Reference in New Issue
Block a user