Skip to content

Fix sampler nan check when calling top_k_top_p_sampling_from_probs #5546

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 11 commits into from
Apr 20, 2025
Merged

Fix sampler nan check when calling top_k_top_p_sampling_from_probs #5546

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
3a9d12e
remove success that is not returned from flash infer any more
yubofredwang Apr 19, 2025
35f603c
remove comments
yubofredwang Apr 19, 2025
bf376f2
change function signature
yubofredwang Apr 19, 2025
09eac31
change signature
yubofredwang Apr 19, 2025
c8b2a8e
Merge remote-tracking branch 'upstream/main' into fix-page-size
yubofredwang Apr 19, 2025
dd25e88
zero out
yubofredwang Apr 19, 2025
8d195d2
fix sampling
yubofredwang Apr 19, 2025
25109cb
fix shape
yubofredwang Apr 19, 2025
0242a77
fix sampler
yubofredwang Apr 19, 2025
86ef868
Merge branch 'main' into fix-page-size
yubofredwang Apr 19, 2025
910ce92
Merge branch 'main' into fix-page-size
zhyncs Apr 20, 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
7 changes: 3 additions & 4 deletions python/sglang/srt/layers/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,17 +100,16 @@ def forward(
probs, sampling_info.min_ps
)
else:
# Check Nan will throw exception, only check when crash_on_warnings is True
check_nan = self.use_nan_detection and crash_on_warnings()
Copy link
Contributor Author

@yubofredwang yubofredwang Apr 19, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Adding another comment for future reference: because we already check for Nan for logits before they get converted into probs. The chance of probs being Nan is quite low. Even when probs contain NaN, sampling was still able to generate tokens

batch_next_token_ids = top_k_top_p_sampling_from_probs(
probs,
sampling_info.top_ks,
sampling_info.top_ps,
filter_apply_order="joint",
check_nan=check_nan,
)

if self.use_nan_detection:
Copy link
Contributor Author

@yubofredwang yubofredwang Apr 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Adding a comment here as well: we can remove this check as the new flash infer sampling implementation guarantees the generation of enough tokens. See PR: flashinfer-ai/flashinfer#912

logger.warning("Detected errors during sampling!")
batch_next_token_ids = torch.zeros_like(batch_next_token_ids)

elif global_server_args_dict["sampling_backend"] == "pytorch":
# A slower fallback implementation with torch native operations.
batch_next_token_ids = top_k_top_p_min_p_sampling_from_probs_torch(
Expand Down
10 changes: 5 additions & 5 deletions sgl-kernel/python/sgl_kernel/sampling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
from typing import Optional, Tuple, Union
from typing import Optional, Union

import torch
from sgl_kernel.utils import _to_tensor_scalar_tuple, get_cuda_stream
Expand Down Expand Up @@ -109,7 +109,7 @@ def _top_p_sampling_from_probs_internal(
top_p_val: float,
deterministic: bool,
generator: Optional[torch.Generator],
) -> Tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
with probs.device as device:
probs = probs.float()
maybe_top_p_arr = (
Expand All @@ -135,7 +135,7 @@ def top_p_sampling_from_probs(
deterministic: bool = True,
generator: Optional[torch.Generator] = None,
check_nan: bool = False,
) -> Tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
r"""Adapt from https://github.com/flashinfer-ai/flashinfer/flashinfer/sampling.py
Fused GPU kernel for top-p sampling (nucleus sampling) from probabilities,
this operator implements GPU-based rejection sampling without explicit sorting.
Expand Down Expand Up @@ -194,7 +194,7 @@ def _top_k_top_p_sampling_from_probs_internal(
top_p_val: float,
deterministic: bool,
generator: Optional[torch.Generator],
) -> Tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
with probs.device as device:
probs = probs.float()
maybe_top_k_arr = maybe_top_k_arr.int() if maybe_top_k_arr is not None else None
Expand Down Expand Up @@ -225,7 +225,7 @@ def top_k_top_p_sampling_from_probs(
deterministic: bool = True,
generator: Optional[torch.Generator] = None,
check_nan: bool = False,
) -> Tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
r"""Adapt from https://github.com/flashinfer-ai/flashinfer/flashinfer/sampling.py
Fused GPU kernel for top-k and top-p sampling from probabilities,

Expand Down
Loading