github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-06-21 21:22:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

handle empty index

Browse Source
This commit is contained in:
Tong Li 2025-05-15 18:30:27 +08:00 committed by YeAnbang
parent 957e3a521a
commit 1644adf684
2 changed files with 37 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
applications/ColossalChat/coati/distributed/consumer.py
View File

@ -113,7 +113,6 @@ class BaseConsumer:
) as pbar: ) as pbar:
for step in pbar: for step in pbar:
i = 0 i = 0
allow_sync_model = False
for _ in range(self.num_recv_per_update): for _ in range(self.num_recv_per_update):
# receive data from producers # receive data from producers
for r in range(self.num_producers): for r in range(self.num_producers):
@ -139,7 +138,6 @@ class BaseConsumer:
else: else:
self.buffer = self.buffer[self.dp_size * self.minibatch_size :] self.buffer = self.buffer[self.dp_size * self.minibatch_size :]
if loss is not None: if loss is not None:
allow_sync_model = True
pbar.set_postfix({"loss": loss}) pbar.set_postfix({"loss": loss})
i += 1 i += 1
if self.lr_scheduler is not None: if self.lr_scheduler is not None:
@ -153,7 +151,6 @@ class BaseConsumer:
print(f"Saved model checkpoint at step {step + 1} in folder {save_path}") print(f"Saved model checkpoint at step {step + 1} in folder {save_path}")
if episode != self.num_episodes - 1 or step != self.num_update_per_episode - 1: if episode != self.num_episodes - 1 or step != self.num_update_per_episode - 1:
if allow_sync_model:
if self.pp_size > 1: if self.pp_size > 1:
print( print(
f"[T{dist.get_rank()}] Sync model PP stage {self.pp_rank} episode {episode} step {step}" f"[T{dist.get_rank()}] Sync model PP stage {self.pp_rank} episode {episode} step {step}"
@ -177,7 +174,6 @@ class BaseConsumer:
) )
del state_dict del state_dict
torch.cuda.empty_cache() torch.cuda.empty_cache()
allow_sync_model = False
@ray.remote @ray.remote

7
applications/ColossalChat/coati/distributed/grpo_consumer.py
View File

@ -245,7 +245,10 @@ class GRPOConsumer(BaseConsumer):
# TODO: customize excessive prompts calculation. # TODO: customize excessive prompts calculation.
if excessive_prompts_per_rank != 0: if excessive_prompts_per_rank != 0:
# Mask excessive prompts to False # Mask excessive prompts to False
true_indices = torch.nonzero(effective_prompts_mask).squeeze() true_indices = torch.nonzero(effective_prompts_mask)
# Make sure the indices are not empty.
if true_indices.numel() > 0:
true_indices = true_indices.squeeze()
if excessive_prompts_per_rank <= len(true_indices): if excessive_prompts_per_rank <= len(true_indices):
excessive_prompts_idx = true_indices[-excessive_prompts_per_rank:] excessive_prompts_idx = true_indices[-excessive_prompts_per_rank:]
else: else:
@ -256,6 +259,8 @@ class GRPOConsumer(BaseConsumer):
if effective_prompts_mask[mask_idx] == False: if effective_prompts_mask[mask_idx] == False:
# Update loss mask. # Update loss mask.
loss_mask[mask_idx] = False loss_mask[mask_idx] = False
else:
excessive_prompts_idx = torch.empty([0])
else: else:
# If dynamic batching is disabled, we need to use all samples for training. # If dynamic batching is disabled, we need to use all samples for training.
need_update = (step_idx + 1) % self.num_microbatches == 0 need_update = (step_idx + 1) % self.num_microbatches == 0