github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-07-13 07:04:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

[fix] rm comments;

Browse Source
This commit is contained in:
duanjunwen 2024-09-20 07:29:41 +00:00
parent c6d6ee39bd
commit b6616f544e
1 changed files with 0 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
colossalai/pipeline/schedule/zero_bubble_pp.py
View File

@ -489,12 +489,6 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# For chunk 0 stage 0, use micro_batch as input_obj_ # For chunk 0 stage 0, use micro_batch as input_obj_
if model_chunk_id == 0 and self.stage_manager.is_first_stage(ignore_chunk=True): if model_chunk_id == 0 and self.stage_manager.is_first_stage(ignore_chunk=True):
# for k, v in micro_batch.items():
# if v.requires_grad:
# input_obj_.append(micro_batch[k])
# output_obj_.append(output_obj[k]) # y
# output_obj_grad_.append(output_obj_grad[k]) # dy
input_obj_, _ = tree_flatten(micro_batch) input_obj_, _ = tree_flatten(micro_batch)
output_obj_, _ = tree_flatten(output_obj) # y output_obj_, _ = tree_flatten(output_obj) # y
output_obj_grad_, _ = tree_flatten(output_obj_grad) # dy output_obj_grad_, _ = tree_flatten(output_obj_grad) # dy
@ -502,22 +496,12 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# For loss backward; output_obj is loss; output_obj_grad should be None # For loss backward; output_obj is loss; output_obj_grad should be None
elif model_chunk_id == 1 and self.stage_manager.is_first_stage(ignore_chunk=True): elif model_chunk_id == 1 and self.stage_manager.is_first_stage(ignore_chunk=True):
assert output_obj_grad is None assert output_obj_grad is None
# for k, v in input_obj.items():
# if v.requires_grad:
# input_obj_.append(input_obj[k])
input_obj_, _ = tree_flatten(input_obj) input_obj_, _ = tree_flatten(input_obj)
# output_obj_.append(output_obj) # LOSS
# output_obj_grad_.append(output_obj_grad) # None
output_obj_, _ = tree_flatten(output_obj) # LOSS output_obj_, _ = tree_flatten(output_obj) # LOSS
output_obj_grad_, _ = tree_flatten(output_obj_grad) # None output_obj_grad_, _ = tree_flatten(output_obj_grad) # None
# For other chunk stage, use input_obj as input_obj_; # For other chunk stage, use input_obj as input_obj_;
else: else:
# for k, v in input_obj.items():
# if v.requires_grad:
# input_obj_.append(input_obj[k])
# output_obj_.append(output_obj[k]) # y
# output_obj_grad_.append(output_obj_grad[k]) # dy
input_obj_, _ = tree_flatten(input_obj) input_obj_, _ = tree_flatten(input_obj)
output_obj_, _ = tree_flatten(output_obj) # y output_obj_, _ = tree_flatten(output_obj) # y
output_obj_grad_, _ = tree_flatten(output_obj_grad) # dy output_obj_grad_, _ = tree_flatten(output_obj_grad) # dy
@ -572,10 +556,6 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
output_obj_.append(output_obj) # LOSS output_obj_.append(output_obj) # LOSS
output_obj_grad_.append(None) # None output_obj_grad_.append(None) # None
else: else:
# for k, v in output_obj.items():
# if v.requires_grad:
# output_obj_.append(output_obj[k])
# output_obj_grad_.append(output_obj_grad[k])
output_obj_, _ = tree_flatten(output_obj) # y output_obj_, _ = tree_flatten(output_obj) # y
output_obj_grad_, _ = tree_flatten(output_obj_grad) # dy output_obj_grad_, _ = tree_flatten(output_obj_grad) # dy
@ -653,7 +633,6 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
tree_map(deallocate, deallocate_output_obj) tree_map(deallocate, deallocate_output_obj)
# add input and output object for backward b # add input and output object for backward b
self.input_tensors[model_chunk_id].append((micro_batch, input_obj)) self.input_tensors[model_chunk_id].append((micro_batch, input_obj))
# for bwd b&w, we only need the graph(grad_fn) of output_obj # for bwd b&w, we only need the graph(grad_fn) of output_obj