Revert "[Misc] Remove use of CUDA_VISIBLE_DEVICES for device selectio… (#27502)

This commit is contained in:
Zhuohan Li
2025-10-25 05:31:43 +00:00
committed by GitHub
parent 29c9cb8007
commit 56ed7609a9
4 changed files with 7 additions and 35 deletions
@@ -991,14 +991,11 @@ class NixlConnectorWorker:
# 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
@@ -1026,7 +1023,7 @@ class NixlConnectorWorker:
"All kv cache tensors must have the same size"
)
caches_data.append(
(base_addr, curr_tensor_size_bytes, self.device_id, "")
(base_addr, curr_tensor_size_bytes, self.tp_rank, "")
)
logger.debug(
@@ -1073,7 +1070,7 @@ class NixlConnectorWorker:
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.device_id))
blocks_data.append((addr, kv_block_len, self.tp_rank))
if self._use_flashinfer:
# Separate and interleave K/V regions to maintain the same
@@ -1084,13 +1081,12 @@ class NixlConnectorWorker:
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.device_id))
blocks_data.append((v_addr, kv_block_len, self.tp_rank))
logger.debug(
"Created %s blocks for src engine %s and rank %s on device id %s",
"Created %s blocks for src engine %s and rank %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)