[KV-Offloading][TP] : Expand replicated_layout detection to multi-group MLA (#57652)

This commit is contained in:
Varun Sundar Rabindranath
2026-09-30 09:11:03 -04:00
committed by GitHub
parent f28a508162
commit 5463fe4962
10 changed files with 553 additions and 91 deletions
@@ -1350,6 +1350,7 @@ class MLAAttention(nn.Module, AttentionLayerBase):
common_kwargs = dict(
block_size=vllm_config.cache_config.block_size,
num_kv_heads=1,
max_tp_shards=1,
head_size=self.head_size,
dtype=kv_cache_dtype,
cache_dtype_str=self.kv_cache_dtype,
@@ -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,