Skip to content

make training more stable #620

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Jul 5, 2023
Merged

make training more stable #620

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
138e3d2
make training more stable
yaozhewei Jun 30, 2023
8ea8c15
Merge branch 'master' into zheweiyao/stablize_training
awan-10 Jul 5, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 16 additions & 8 deletions applications/DeepSpeed-Chat/training/step3_rlhf_finetuning/ppo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,7 +58,7 @@ def __init__(self, rlhf_engine, args):
args.end_of_conversation_token)['input_ids'][-1]

# Those value can be changed
self.kl_ctl = 0.02
self.kl_ctl = 0.1
self.clip_reward_value = 5
self.cliprange = 0.2
self.cliprange_value = 0.2
Expand All @@ -70,17 +70,20 @@ def _generate_sequence(self, prompts, mask):
max_min_length = self.max_answer_seq_len + prompts.shape[1]

with torch.no_grad():
seq = self.actor_model.module.generate(prompts,
attention_mask=mask,
max_length=max_min_length,
min_length=max_min_length)
seq = self.actor_model.module.generate(
prompts,
attention_mask=mask,
max_length=max_min_length,
pad_token_id=self.tokenizer.pad_token_id,
# min_length=max_min_length
)

# Filter out seq with no answers (or very short). This happens when users directly use the pre-training ckpt without supervised finetuning
# NOTE: this will causes each GPU has different number of examples
batch_size = seq.shape[0]
prompt_length = prompts.shape[1]
ans = seq[:, prompt_length:]
self.prompt_length = prompt_length
ans = seq[:, prompt_length:]
valid_ans_len = (ans != self.tokenizer.pad_token_id).sum(dim=-1)
out_seq = []
for i in range(batch_size):
Expand All @@ -100,7 +103,6 @@ def generate_experience(self, prompts, mask):

pad_token_id = self.tokenizer.pad_token_id
attention_mask = seq.not_equal(pad_token_id).long()

with torch.no_grad():
output = self.actor_model(seq, attention_mask=attention_mask)
output_ref = self.ref_model(seq, attention_mask=attention_mask)
Expand Down Expand Up @@ -131,7 +133,7 @@ def compute_rewards(self, prompts, log_probs, ref_log_probs, reward_score,
kl_divergence_estimate = -self.kl_ctl * (log_probs - ref_log_probs)
rewards = kl_divergence_estimate
start = prompts.shape[1] - 1
ends = start + action_mask[:, start:].sum(1)
ends = start + action_mask[:, start:].sum(1) + 1
reward_clip = torch.clamp(reward_score, -self.clip_reward_value,
self.clip_reward_value)
batch_size = log_probs.shape[0]
Expand Down Expand Up @@ -159,6 +161,12 @@ def train_rlhf(self, inputs):
old_rewards = self.compute_rewards(prompts, log_probs,
ref_log_probs, reward_score,
action_mask)
ends = start + action_mask[:, start:].sum(1) + 1
# we need to zero out the reward and value after the end of the conversation
# otherwise the advantage/return will be wrong
for i in range(old_rewards.shape[0]):
old_rewards[i, ends[i]:] = 0
old_values[i, ends[i]:] = 0
advantages, returns = self.get_advantages_and_returns(
old_values, old_rewards, start)

Expand Down