[v1] Refactor KVCacheConfig (#14079)

Signed-off-by: Chen Zhang <zhangch99@outlook.com>

vllm/v1/executor/abstract.py

@@ -62,14 +62,11 @@ class Executor(ExecutorBase):
                             args=(kv_cache_configs, ))
         self.collective_rpc("compile_or_warm_up_model")
 
-    def determine_available_memory(self) -> int:  # in bytes
+    def determine_available_memory(self) -> list[int]:  # in bytes
         output = self.collective_rpc("determine_available_memory")
-        # Since we use a shared centralized controller, we take the minimum
-        # memory size across all workers to make sure all the memory
-        # operators can be applied to all workers.
-        return min(output)
+        return output
 
-    def get_kv_cache_specs(self) -> list[KVCacheSpec]:
+    def get_kv_cache_specs(self) -> list[dict[str, KVCacheSpec]]:
         output = self.collective_rpc("get_kv_cache_spec")
         return output
 
@@ -95,7 +92,7 @@ class UniProcExecutor(UniProcExecutorV0, Executor):
 
 class ExecutorWithExternalLauncher(ExecutorWithExternalLauncherV0, Executor):
 
-    def determine_available_memory(self) -> int:  # in bytes
+    def determine_available_memory(self) -> list[int]:  # in bytes
         # same as determine_num_available_blocks in v0,
         # we need to get the min across all ranks.
         memory = super().determine_available_memory()
@@ -103,4 +100,4 @@ class ExecutorWithExternalLauncher(ExecutorWithExternalLauncherV0, Executor):
         cpu_group = get_world_group().cpu_group
         memory_tensor = torch.tensor([memory], device="cpu", dtype=torch.int64)
         dist.all_reduce(memory_tensor, group=cpu_group, op=dist.ReduceOp.MIN)
-        return memory_tensor.item()
+        return [memory_tensor.item()]