Skip to content

Fix Illegal Instruction/IMA errors when using DP attention with DeepSeek-V3.2 models #12052

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

Closed
YAMY1234 wants to merge 7 commits into from
Closed

Fix Illegal Instruction/IMA errors when using DP attention with DeepSeek-V3.2 models #12052

Changes from 6 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
e582cd9
successfully run version
YAMY1234 Oct 24, 2025
2b8bc4f
remove debug prints
YAMY1234 Oct 24, 2025
7113ceb
format fix
YAMY1234 Oct 24, 2025
a12df55
fix gemini suggestion
YAMY1234 Oct 24, 2025
abbd610
Merge branch 'main' into fix_gpqa
Fridge003 Oct 24, 2025
fdf8345
format fix
YAMY1234 Oct 24, 2025
b4c0f5d
LogitsMetadata judge move to _dp_gather_via_all_reduce
YAMY1234 Oct 24, 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
25 changes: 22 additions & 3 deletions python/sglang/srt/layers/dp_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -355,16 +355,27 @@ def disable_dp_size():


def get_dp_local_info(forward_batch: ForwardBatch) -> Tuple[torch.Tensor, torch.Tensor]:
# `get_dp_local_info` is only called in global DP gather and scatter. We use global DP rank here.
# Compute local token range for current DP rank.
# For LogitsMetadata, use global_num_tokens_for_logprob_gpu (pruned tokens needing logits).
# For ForwardBatch (MLP/Attention), use global_num_tokens_gpu (all tokens in batch).
dp_rank = get_attention_dp_rank()

if forward_batch.dp_local_start_pos is None:
cumtokens = torch.cumsum(forward_batch.global_num_tokens_gpu, dim=0)
# Select metadata source based on context
if (

@ch-wan ch-wan Oct 24, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

How can dp_local_start_pos be None in LogitsMetadata? It is expected to be processed here:

sglang/python/sglang/srt/layers/logits_processor.py

Line 577 in 0bfa394

logits_metadata.compute_dp_attention_metadata()
.

LogitsMetadata has a separate processing logic. We should keep it separated. Perhaps we should add assertion like assert isinstance(forward_batch, ForwardBatch) to avoid misuse.

@YAMY1234 YAMY1234 Oct 24, 2025

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

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

Thanks, understand the design! The call logic should indeed stay separated at the outer level. Will have another patch soon after verified

@YAMY1234 YAMY1234 Oct 24, 2025

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

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

@ch-wan Just pushed another patch, could you take another look? Thanks

type(forward_batch).__name__ == "LogitsMetadata"
and forward_batch.global_num_tokens_for_logprob_gpu is not None
):
global_num_tokens_source = forward_batch.global_num_tokens_for_logprob_gpu
else:
global_num_tokens_source = forward_batch.global_num_tokens_gpu

cumtokens = torch.cumsum(global_num_tokens_source, dim=0)
if dp_rank == 0:
local_start_pos = torch.zeros_like(cumtokens[0])
else:
local_start_pos = cumtokens[dp_rank - 1]
local_num_tokens = forward_batch.global_num_tokens_gpu[dp_rank]
local_num_tokens = global_num_tokens_source[dp_rank]

forward_batch.dp_local_start_pos = local_start_pos
forward_batch.dp_local_num_tokens = local_num_tokens
Expand Down Expand Up @@ -421,6 +432,14 @@ def _dp_gather_via_all_reduce(
):
local_start_pos, local_num_tokens = get_dp_local_info(forward_batch)

# Fix potential metadata mismatch: scheduler's num_tokens_for_logprob may not account
# for pruning done by logits_processor. Use actual tensor size as authoritative source.
actual_local_size = local_tokens.shape[0]
if isinstance(local_num_tokens, torch.Tensor):
local_num_tokens.fill_(actual_local_size) # CUDA graph compatibility
else:
local_num_tokens = actual_local_size

global_tokens.fill_(0)
assert local_tokens.is_contiguous()
assert global_tokens.is_contiguous()
Expand Down
Loading