[V1] Only print cudagraph tqdm on rank 0 with is_global_first_rank (#19516)
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -26,7 +26,7 @@ from vllm.distributed.kv_transfer import (get_kv_transfer_group,
|
||||
has_kv_transfer_group)
|
||||
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
|
||||
from vllm.distributed.parallel_state import (
|
||||
get_pp_group, get_tp_group, graph_capture,
|
||||
get_pp_group, get_tp_group, graph_capture, is_global_first_rank,
|
||||
prepare_communication_buffer_for_model)
|
||||
from vllm.forward_context import (DPMetadata, get_forward_context,
|
||||
set_forward_context)
|
||||
@@ -2285,9 +2285,12 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
# can reuse the memory pool allocated for the large shapes.
|
||||
with graph_capture(device=self.device):
|
||||
full_cg = self.full_cuda_graph
|
||||
for num_tokens in tqdm(reversed(self.cudagraph_batch_sizes),
|
||||
desc="Capturing CUDA graphs",
|
||||
total=len(self.cudagraph_batch_sizes)):
|
||||
# Only rank 0 should print progress bar during capture
|
||||
compilation_cases = reversed(self.cudagraph_batch_sizes)
|
||||
if is_global_first_rank():
|
||||
compilation_cases = tqdm(list(compilation_cases),
|
||||
desc="Capturing CUDA graph shapes")
|
||||
for num_tokens in compilation_cases:
|
||||
# We skip EPLB here since we don't want to record dummy metrics
|
||||
for _ in range(
|
||||
self.compilation_config.cudagraph_num_of_warmups):
|
||||
|
||||
Reference in New Issue
Block a user