[Hoxfix] Fix CUDA_DEVICE_MAX_CONNECTIONS for comm overlap

Co-authored-by: Edenzzzz <[email protected]>

colossalai/initialize.py

@@ -3,6 +3,12 @@
 
 import os
 
+# set CUDA_DEVICE_MAX_CONNECTIONS=1 to ensure that when overlapping communication and computation,
+# the order of of kernel launches on GPUs are the same as on the CPU so that comm is launched first.
+# see https://github.com/NVIDIA/Megatron-LM/issues/533
+# https://forums.developer.nvidia.com/t/how-many-streams-maximum-number-of-streams/6571/16
+os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "1"
+
 import torch.distributed as dist
 
 from colossalai.accelerator import get_accelerator

colossalai/legacy/nn/layer/parallel_1d/_operation.py

@@ -81,7 +81,6 @@ def backward(ctx, grad_output):
             handle = dist.all_reduce(grad_input, group=gpc.get_group(ctx.parallel_mode), async_op=True)
             # Delay the start of weight gradient computation shortly (3us) to have
             # all-reduce scheduled first and have GPU resources allocated
-            _ = torch.empty(1, device=grad_output.device) + 1
 
         grad_weight = grad_output.t().matmul(total_input)
         grad_bias = grad_output.sum(dim=0) if use_bias else None

colossalai/shardformer/shard/shardformer.py

@@ -1,4 +1,3 @@
-import os
 from typing import Dict, List, Tuple
 
 import torch.distributed as dist
@@ -11,9 +10,6 @@
 from .shard_config import ShardConfig
 from .sharder import ModelSharder
 
-# set CUDA_DEVICE_MAX_CONNECTIONS=1 to ensure that when communication and computation overlap, the order of core scheduling is correct
-os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "1"
-
 
 class ShardFormer:
     """

examples/language/llama/benchmark.py

@@ -292,7 +292,7 @@ def empty_init():
     with get_profile_context(
         args.profile,
         args.ignore_steps,
-        len(dataloader) - 1,
+        1,  # avoid creating massive log files
         save_dir=f"profile/{time.strftime('%H:%M', time.localtime())}-{args.plugin}-llama-{args.config}",
     ) as prof:
         if isinstance(plugin, HybridParallelPlugin) and args.pp > 1: