github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-02 09:38:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

[zero] add strict ddp mode (#2508)

Browse Source 
* [zero] add strict ddp mode

* [polish] add comments for strict ddp mode

* [zero] fix test error
This commit is contained in:
HELSON
2023-01-20 14:04:38 +08:00
committed by GitHub
parent c04f183237
commit 2d1a7dfe5f
5 changed files with 31 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
examples/language/gpt/gemini/commons/model_zoo.py
View File

@@ -53,6 +53,14 @@ def gpt2_24b(checkpoint=True):
return GPTLMModel(hidden_size=8192, num_layers=30, num_attention_heads=16, checkpoint=checkpoint)
def gpt2_30b(checkpoint=True):
return GPTLMModel(hidden_size=8192, num_layers=37, num_attention_heads=16, checkpoint=checkpoint)
def gpt2_40b(checkpoint=True):
return GPTLMModel(hidden_size=8192, num_layers=50, num_attention_heads=16, checkpoint=checkpoint)
def model_builder(model_size: str) -> callable:
if model_size == "gpt2_medium":
return gpt2_medium
@@ -66,6 +74,10 @@ def model_builder(model_size: str) -> callable:
return gpt2_20b
elif model_size == "gpt2_24b":
return gpt2_24b
elif model_size == "gpt2_30b":
return gpt2_30b
elif model_size == "gpt2_40b":
return gpt2_40b
else:
raise TypeError(f"model_builder {model_size}")

10
examples/language/gpt/gemini/train_gpt_demo.py
View File

@@ -187,17 +187,18 @@ def tensor_parallelize(model: torch.nn.Module, pg: ProcessGroup):
# Gemini + ZeRO DDP
def build_gemini(model: torch.nn.Module, pg: ProcessGroup, placement_policy: str = "auto"):
def build_gemini(model: torch.nn.Module, pg: ProcessGroup, placement_policy: str = "auto", ddp_flag: bool = True):
fp16_init_scale = 2**5
gpu_margin_mem_ratio_for_auto = 0
if version.parse(CAI_VERSION) > version.parse("0.1.10"):
model = GeminiDDP(model,
strict_ddp_mode=ddp_flag,
device=get_current_device(),
placement_policy=placement_policy,
pin_memory=True,
hidden_dim=model.config.n_embd,
search_range_mb=64)
search_range_mb=128)
# configure the const policy
if placement_policy == 'const':
model.gemini_manager._placement_policy.set_const_memory_boundary(2 * 1024)
@@ -279,11 +280,12 @@ def main():
tp_pg = ProcessGroup(tp_degree=args.tp_degree)
# Tensor Parallelism (TP)
# You should notice that v0.1.10 is not compatible with TP degree > 1
tensor_parallelize(model, tp_pg)
if args.tp_degree > 1:
tensor_parallelize(model, tp_pg)
# build a Gemini model and a highly optimized cpu optimizer
# Gemini + ZeRO DP, Note it must be used after TP
model, optimizer = build_gemini(model, tp_pg, args.placement)
model, optimizer = build_gemini(model, tp_pg, args.placement, args.tp_degree == 1)
logger.info(get_mem_info(prefix='After init optim, '), ranks=[0])
else: