EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix incorrect spec_num_draft_tokens in draft_extend (#7757)

Browse Source
This commit is contained in:
Cheng Wan
2025-07-05 02:18:16 -07:00
committed by GitHub
parent 77cfea689d
commit 6c903611ca
2 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/layers/dp_attention.py
View File

@@ -237,6 +237,10 @@ def _dp_gather(
assert (
local_tokens.untyped_storage() is not global_tokens.untyped_storage()
), "aliasing between global_tokens and local_tokens not allowed"
# NOTE: During draft extend, the gathered_buffer is padded to num_tokens * (speculative_num_steps + 1).
# But the size of local_tokens is total accepted tokens. We need to reduce the local_num_tokens to the
# actual size of the accepted tokens.
if forward_batch.forward_mode.is_draft_extend():
shape_tensor = local_num_tokens.new_full((), local_tokens.shape[0])
local_num_tokens = torch.minimum(local_num_tokens, shape_tensor)
@@ -291,6 +295,10 @@ def dp_scatter(
assert (
local_tokens.untyped_storage() is not global_tokens.untyped_storage()
), "aliasing between local_tokens and global_tokens not allowed"
# NOTE: During draft extend, the gathered_buffer is padded to num_tokens * (speculative_num_steps + 1).
# But the size of local_tokens is total accepted tokens. We need to reduce the local_num_tokens to the
# actual size of the accepted tokens.
if forward_batch.forward_mode.is_draft_extend():
shape_tensor = local_num_tokens.new_full((), local_tokens.shape[0])
local_num_tokens = torch.minimum(local_num_tokens, shape_tensor)

2
python/sglang/srt/speculative/eagle_worker.py
View File

@@ -844,7 +844,7 @@ class EAGLEWorker(TpModelWorker):
)
batch.return_hidden_states = False
model_worker_batch = batch.get_model_worker_batch()
model_worker_batch.spec_num_draft_tokens = self.speculative_num_draft_tokens
model_worker_batch.spec_num_draft_tokens = self.speculative_num_steps + 1
assert model_worker_batch.capture_hidden_mode == CaptureHiddenMode.LAST
forward_batch = ForwardBatch.init_new(
model_worker_batch, self.draft_model_runner