[Misc] Remove use of CUDA_VISIBLE_DEVICES for device selection (fix DP slow startup time &c) (vllm-project#26709)

Signed-off-by: ilmarkov <markovilya197@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: Alberto Perdomo <aperdomo@redhat.com>

Author: 2 people

vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py

@@ -991,11 +991,14 @@ def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
         # Enable different block lengths for different layers when MLA is used.
         self.block_len_per_layer = list[int]()
         self.slot_size_per_layer = list[int]()  # HD bytes in kv terms
+        self.device_id = self.tp_rank
         for layer_name, cache_or_caches in xfer_buffers.items():
             cache_list = cache_or_caches if split_k_and_v else [cache_or_caches]
 
             for cache in cache_list:
                 base_addr = cache.data_ptr()
+                if not self.use_host_buffer and current_platform.is_cuda_alike():
+                    self.device_id = cache.device.index
                 if base_addr in seen_base_addresses:
                     continue
 
@@ -1023,7 +1026,7 @@ def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
                         "All kv cache tensors must have the same size"
                     )
                 caches_data.append(
-                    (base_addr, curr_tensor_size_bytes, self.tp_rank, "")
+                    (base_addr, curr_tensor_size_bytes, self.device_id, "")
                 )
 
         logger.debug(
@@ -1070,7 +1073,7 @@ def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
                 block_offset = block_id * self.block_len_per_layer[i]
                 addr = base_addr + block_offset
                 # (addr, len, device id)
-                blocks_data.append((addr, kv_block_len, self.tp_rank))
+                blocks_data.append((addr, kv_block_len, self.device_id))
 
             if self._use_flashinfer:
                 # Separate and interleave K/V regions to maintain the same
@@ -1081,12 +1084,13 @@ def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
                     addr = base_addr + block_offset
                     # Register addresses for V cache (K registered first).
                     v_addr = addr + kv_block_len
-                    blocks_data.append((v_addr, kv_block_len, self.tp_rank))
+                    blocks_data.append((v_addr, kv_block_len, self.device_id))
         logger.debug(
-            "Created %s blocks for src engine %s and rank %s",
+            "Created %s blocks for src engine %s and rank %s on device id %s",
             len(blocks_data),
             self.engine_id,
             self.tp_rank,
+            self.device_id,
         )
 
         descs = self.nixl_wrapper.get_xfer_descs(blocks_data, self.nixl_memory_type)

vllm/v1/engine/utils.py

@@ -134,9 +134,12 @@ def __init__(
         data_parallel = vllm_config.parallel_config.data_parallel_size > 1
         try:
             for proc, local_dp_rank in zip(self.processes, local_dp_ranks):
+                # Adjust device control in DP for non-CUDA platforms
+                # For CUDA platforms, setting same device id for different DP
+                # processes affects NCCL init performance.
                 with (
                     set_device_control_env_var(vllm_config, local_dp_rank)
-                    if (data_parallel)
+                    if (data_parallel and not current_platform.is_cuda_alike())
                     else contextlib.nullcontext()
                 ):
                     proc.start()

vllm/v1/worker/dp_utils.py

@@ -8,7 +8,6 @@
 from vllm.config import ParallelConfig
 from vllm.distributed.parallel_state import get_dp_group, is_global_first_rank
 from vllm.logger import init_logger
-from vllm.platforms import current_platform
 from vllm.v1.worker.ubatch_utils import (
     UBatchSlices,
     check_ubatch_thresholds,
@@ -20,7 +19,8 @@
 
 
 def _get_device_and_group(parallel_config: ParallelConfig):
-    device = current_platform.device_type
+    # Use the actual device assigned to the DP group, not just the device type
+    device = get_dp_group().device
     group = get_dp_group().device_group
 
     # Transfering this tensor from GPU to CPU will introduce a GPU sync

vllm/v1/worker/gpu_worker.py

@@ -169,6 +169,27 @@ def init_device(self):
         if self.device_config.device.type == "cuda":
             # This env var set by Ray causes exceptions with graph building.
             os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None)
+            if (
+                self.parallel_config.data_parallel_size > 1
+                and self.parallel_config.data_parallel_size_local > 0
+                and self.parallel_config.data_parallel_backend != "ray"
+            ):
+                # Use local DP rank if available, otherwise use global DP rank.
+                dp_local_rank = self.parallel_config.data_parallel_rank_local
+                if dp_local_rank is None:
+                    dp_local_rank = self.parallel_config.data_parallel_rank
+
+                tp_pp_world_size = (
+                    self.parallel_config.pipeline_parallel_size
+                    * self.parallel_config.tensor_parallel_size
+                )
+
+                # DP_LOCAL_RANK * TP_PP_WORLD_SIZE + TP_LOCAL_RANK
+                self.local_rank += dp_local_rank * tp_pp_world_size
+                assert self.local_rank <= torch.cuda.device_count(), (
+                    f"DP adjusted local rank {self.local_rank} is out of bounds. "
+                )
+
             self.device = torch.device(f"cuda:{self.local_rank}")
             current_platform.set_device(self.device)