Merge branch 'v0.9.2-dev' into v0.9.2-dev-ds

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py

@@ -55,11 +55,6 @@ class ReqMeta:
             slot_mapping=slot_mapping,
         )
 
-        self.parallel_config = vllm_config.parallel_config
-        self.model_config = vllm_config.model_config
-        self.total_num_hidden_layers = getattr(self.model_config.hf_text_config,
-                                              "num_hidden_layers", 0)
-        self.pp_size = self.parallel_config.pipeline_parallel_size
         
 @dataclass
 class P2pNcclConnectorMetadata(KVConnectorMetadata):
@@ -100,6 +95,12 @@ class P2pNcclConnector(KVConnectorBase_V1):
             hostname="",
             port_offset=self._rank,
         ) if role == KVConnectorRole.WORKER else None
+        
+        self.parallel_config = vllm_config.parallel_config
+        self.model_config = vllm_config.model_config
+        self.total_num_hidden_layers = getattr(self.model_config.hf_text_config,
+                                              "num_hidden_layers", 0)
+        self.pp_size = self.parallel_config.pipeline_parallel_size
 
     # ==============================
     # Worker-side methods

vllm/distributed/kv_transfer/kv_connector/v1/p2p/tensor_memory_pool.py

@@ -63,7 +63,7 @@ class TensorMemoryPool:
             than min_block_size
     """
 
-    def __init__(self, max_block_size: int, min_block_size: int = 512):
+    def __init__(self, max_block_size: int, min_block_size: int = 128):
         if max_block_size <= 0 or min_block_size <= 0:
             raise ValueError("Block sizes must be positive")
         if max_block_size < min_block_size:

vllm/platforms/rocm.py

@@ -279,6 +279,7 @@ class RocmPlatform(Platform):
                 logger.info_once("Using Flash Attention backend on V1 engine. (only supports block size 64)")
                 return FLASH_ATTN_V1
             else:
+                os.environ['VLLM_USE_FLASH_ATTN_PA'] = '0'
                 logger.info_once("Using Triton backend on V1 engine.")
                 return TRITON_ATTN_VLLM_V1