github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

[Gemini] hotfix the unittest bugs (#2125)

pull/2126/head
Jiarui Fang 2022-12-13 14:14:55 +08:00 committed by GitHub
parent 05bb28aacf
commit 5efda69735
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 11 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
colossalai/gemini/memory_tracer/param_runtime_order.py
View File

@ -36,7 +36,7 @@ class OrderedParamGenerator(ParamGenerator):
del visited_set del visited_set
def is_empty(self): def is_empty(self):
return len(self.param_visited_order) > 0 return len(self.param_visited_order) == 0
def clear(self): def clear(self):
self.param_visited_order = [] self.param_visited_order = []

18
tests/test_gemini/update/test_gemini_use_rmt.py
View File

@ -45,11 +45,15 @@ def run_gemini_use_rmt(placement_policy, keep_gather, model_name: str, use_grad_
run_fwd_bwd(runtime_mem_tracer, input_ids, label, criterion, runtime_mem_tracer) run_fwd_bwd(runtime_mem_tracer, input_ids, label, criterion, runtime_mem_tracer)
memstats = runtime_mem_tracer.memstats() memstats = runtime_mem_tracer.memstats()
runtime_tracer_non_model_data = runtime_mem_tracer._memstats._non_model_data_cuda_list runtime_tracer_non_model_data = runtime_mem_tracer._memstats._non_model_data_cuda_list
print('runtime tracer: ', runtime_tracer_non_model_data) print('runtime tracer non model data points: ', len(runtime_tracer_non_model_data))
print([memstats.param_used_timestep(p) for p in model.parameters()])
model = GeminiDDP(model, device='cuda', placement_policy=placement_policy, search_range_mb=1, memstats=memstats) world_size = torch.distributed.get_world_size()
zero_optim = GeminiAdamOptimizer(model, lr=1e-3, initial_scale=1) config_dict, _ = search_chunk_configuration(model, search_range_mb=1, search_interval_byte=100)
config_dict[world_size]['chunk_size'] = 5000
config_dict[world_size]['keep_gathered'] = keep_gather
chunk_manager = ChunkManager(config_dict)
gemini_manager = GeminiManager(placement_policy, chunk_manager, memstats)
model = ZeroDDP(model, gemini_manager, pin_memory=True)
pg = ProcessGroup() pg = ProcessGroup()
set_seed(pg.dp_local_rank()) set_seed(pg.dp_local_rank())
@ -61,12 +65,10 @@ def run_gemini_use_rmt(placement_policy, keep_gather, model_name: str, use_grad_
break break
input_ids, label = input_ids.cuda(), label.cuda() input_ids, label = input_ids.cuda(), label.cuda()
zero_optim.zero_grad()
set_seed(42) set_seed(42)
loss = run_fwd_bwd(model, input_ids, label, criterion, zero_optim) loss = run_fwd_bwd(model, input_ids, label, criterion, model)
zero_optim.step()
gemini_non_model_data = model.gemini_manager._mem_stats_collector._memstats.non_model_data_list('cuda') gemini_non_model_data = gemini_manager._mem_stats_collector._memstats.non_model_data_list('cuda')
# print('gemini non model data:', gemini_non_model_data) # print('gemini non model data:', gemini_non_model_data)