Skip to content

Fix logprob does not handle temperature #68

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
fzyzcjy merged 2 commits into from
Oct 23, 2025
Merged

Fix logprob does not handle temperature #68

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
dc2e969
more
fzyzcjy Oct 23, 2025
6f490cf
add temperature for cur_log_probs
fzyzcjy Oct 23, 2025
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
14 changes: 11 additions & 3 deletions miles/backends/fsdp_utils/actor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -215,7 +215,7 @@ def compute_log_prob(
model_args["pixel_values"] = batch["pixel_values"]
logits = self.model(**model_args).logits
batch[f"{store_prefix}log_probs"] = gather_log_probs_packed(
logits, batch["tokens"], self.args.rollout_temperature
logits, batch["tokens"], temperature=self.args.rollout_temperature
)
return rollout_data

Expand Down Expand Up @@ -371,7 +371,9 @@ def train(self, rollout_id: int, rollout_data_ref: Box) -> None:
).logits

# Handle packed sequences
log_probs = gather_log_probs_packed(logits, packed_batch["tokens"], packed_batch["cu_seqlens"])
log_probs = gather_log_probs_packed(
logits, packed_batch["tokens"], packed_batch["cu_seqlens"], temperature=self.args.rollout_temperature
)
packed_batch["cur_log_probs"] = log_probs
unpacked_batches = unpack_sequences(packed_batch)

Expand Down Expand Up @@ -654,7 +656,10 @@ def gather_log_probs(logits: torch.Tensor, input_ids: torch.Tensor, rollout_temp


def gather_log_probs_packed(
logits: torch.Tensor, input_ids: torch.Tensor, cu_seqlens: torch.Tensor | float | None = None
logits: torch.Tensor,
input_ids: torch.Tensor,
cu_seqlens: torch.Tensor | float | None = None,
temperature: torch.Tensor | None = None,
) -> torch.Tensor:
"""Gather next-token log probabilities for packed sequences.

Expand All @@ -673,6 +678,9 @@ def gather_log_probs_packed(
logits = logits.squeeze(0)
input_ids = input_ids.squeeze(0)

if temperature is not None:
logits = logits.div(temperature)

# Shift for next-token prediction: logits[:-1] predicts input_ids[1:]
log_probs = torch.log_softmax(logits[:-1], dim=-1)
targets = input_ids[1:].to(device=log_probs.device)
Expand Down