Skip to content

Fix incorrect LSE gradients in cached FlashAttention for Eagle training #536

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

Open
uygnef wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Fix incorrect LSE gradients in cached FlashAttention for Eagle training #536

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f9c1cda
fix fa
uygnef Apr 16, 2026
65d8943
Improve flash attention validation metrics
uygnef Apr 16, 2026
dec444b
Add a flag for grad norm logging
uygnef Apr 16, 2026
9b3de21
Update specforge/modeling/draft/llama3_eagle.py
uygnef Apr 17, 2026
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
9 changes: 9 additions & 0 deletions scripts/train_dflash.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,6 +117,11 @@ def parse_args():
training_group.add_argument("--accumulation-steps", type=int, default=1)
training_group.add_argument("--seed", type=int, default=42)
training_group.add_argument("--resume", action="store_true")
training_group.add_argument(
"--log-grad-norm",
action="store_true",
help="Log train/pre_clip_grad_norm during training.",
)

output_group = parser.add_argument_group("output")
output_group.add_argument("--output-dir", type=str, required=True)
Expand Down Expand Up @@ -328,6 +333,9 @@ def record_metrics(

if mode == "train" and optimizer is not None:
logdict["train/lr"] = optimizer.get_learning_rate()
grad_norm = optimizer.get_last_grad_norm()
if grad_norm is not None:
logdict["train/pre_clip_grad_norm"] = grad_norm

logdict[f"{mode}/loss"] = loss
logdict[f"{mode}/accuracy"] = accuracy
Expand Down Expand Up @@ -458,6 +466,7 @@ def main():
max_grad_norm=args.max_grad_norm,
warmup_ratio=args.warmup_ratio,
total_steps=total_steps,
log_grad_norm=args.log_grad_norm,
)

if resume_state is not None:
Expand Down
11 changes: 10 additions & 1 deletion scripts/train_eagle3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,6 +165,11 @@ def parse_args() -> Tuple[ArgumentParser, Namespace]:
)
training_group.add_argument("--seed", type=int, default=0)
training_group.add_argument("--draft-accumulation-steps", type=int, default=1)
training_group.add_argument(
"--log-grad-norm",
action="store_true",
help="Log train/pre_clip_grad_norm during training.",
)

# data processing type
optimization_group = parser.add_argument_group("optimization")
Expand All @@ -181,7 +186,7 @@ def parse_args() -> Tuple[ArgumentParser, Namespace]:
"--attention-backend",
type=str,
default="flex_attention",
help="The attention backend for the draft model",
help="The attention backend for the draft model (e.g. flex_attention, fa, usp)",
)

# other args
Expand Down Expand Up @@ -693,6 +698,9 @@ def record_metrcs(

if mode == "train" and optimizer is not None:
logdict["train/lr"] = optimizer.get_learning_rate()
grad_norm = optimizer.get_last_grad_norm()
if grad_norm is not None:
logdict["train/pre_clip_grad_norm"] = grad_norm

accuracies = torch.stack(accuracies)
plosses = torch.stack(plosses)
Expand Down Expand Up @@ -826,6 +834,7 @@ def main():
max_grad_norm=args.max_grad_norm,
warmup_ratio=args.warmup_ratio,
total_steps=args.total_steps,
log_grad_norm=args.log_grad_norm,
)
print_with_rank("Initialized optimizer and scheduler")

Expand Down
Loading
Loading