Skip to content

Revert "[perf][spec decoding] Skip full-vocab softmax in EAGLE draft when topk == 1 (#26235)" #26358

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
Qiaolin-Yu merged 2 commits into from
May 26, 2026
Merged

Revert "[perf][spec decoding] Skip full-vocab softmax in EAGLE draft when topk == 1 (#26235)" #26358

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9a72586
Revert "[perf][spec decoding] Skip full-vocab softmax in EAGLE draft …
michaelzhang-ai May 26, 2026
401ec89
Merge branch 'main' into revert-26235-eagle-topk1-softmax-skip
michaelzhang-ai May 26, 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
10 changes: 2 additions & 8 deletions python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -401,14 +401,8 @@ def run_once():
forward_batch.positions,
forward_batch,
)
if self.topk == 1:
ret.topk_index = torch.argmax(
ret.next_token_logits, dim=-1, keepdim=True
)
ret.topk_p = torch.ones_like(ret.topk_index, dtype=torch.float32)
else:
probs = torch.softmax(ret.next_token_logits, dim=-1)
ret.topk_p, ret.topk_index = fast_topk(probs, self.topk, dim=-1)
probs = torch.softmax(ret.next_token_logits, dim=-1)
ret.topk_p, ret.topk_index = fast_topk(probs, self.topk, dim=-1)
Comment on lines +404 to +405

@gemini-code-assist gemini-code-assist Bot May 26, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

medium

Instead of reverting completely to the full-vocab softmax when self.topk == 1, we can mathematically optimize the top-1 probability computation. By using the identity $\text{softmax}(x)_{\text{argmax}} = 1 / \sum e^{x_i - \max(x)}$, we can compute topk_p and topk_index without materializing the full softmax tensor in HBM. This avoids the accuracy collapse caused by hardcoding topk_p = 1.0 while retaining the performance benefits of skipping the full-vocab softmax.

Suggested change
probs = torch.softmax(ret.next_token_logits, dim=-1)
ret.topk_p, ret.topk_index = fast_topk(probs, self.topk, dim=-1)
if self.topk == 1:
max_logits, ret.topk_index = torch.max(ret.next_token_logits, dim=-1, keepdim=True)
ret.topk_p = 1.0 / torch.exp(ret.next_token_logits - max_logits).sum(dim=-1, keepdim=True)
else:
probs = torch.softmax(ret.next_token_logits, dim=-1)
ret.topk_p, ret.topk_index = fast_topk(probs, self.topk, dim=-1)


forward_batch.out_cache_loc = output_cache_loc_backup
forward_batch.spec_info.hidden_states = hidden_states_backup
Expand Down
22 changes: 4 additions & 18 deletions python/sglang/srt/speculative/eagle_worker_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -483,16 +483,8 @@ def draft_forward(self, forward_batch: ForwardBatch):
forward_batch, skip_attn_backend_init=True
).logits_output
maybe_detect_nan(logits_output.next_token_logits, f"draft_forward step {i}")
if self.topk == 1:
# topk=1 → degenerate single-path tree; `topk_p` is unused
# downstream, so skip softmax and just argmax over logits.
topk_index = torch.argmax(
logits_output.next_token_logits, dim=-1, keepdim=True
)
topk_p = torch.ones_like(topk_index, dtype=torch.float32)
else:
probs = torch.softmax(logits_output.next_token_logits, dim=-1)
topk_p, topk_index = fast_topk(probs, self.topk, dim=-1)
probs = torch.softmax(logits_output.next_token_logits, dim=-1)
topk_p, topk_index = fast_topk(probs, self.topk, dim=-1)
Comment on lines +486 to +487

@gemini-code-assist gemini-code-assist Bot May 26, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

medium

We can optimize the top-1 probability computation here as well when self.topk == 1 by using the mathematical identity $\text{softmax}(x)_{\text{argmax}} = 1 / \sum e^{x_i - \max(x)}$. This avoids the memory bandwidth overhead of full-vocab softmax and fast_topk without causing any accuracy degradation.

            if self.topk == 1:
                max_logits, topk_index = torch.max(logits_output.next_token_logits, dim=-1, keepdim=True)
                topk_p = 1.0 / torch.exp(logits_output.next_token_logits - max_logits).sum(dim=-1, keepdim=True)
            else:
                probs = torch.softmax(logits_output.next_token_logits, dim=-1)
                topk_p, topk_index = fast_topk(probs, self.topk, dim=-1)

maybe_detect_oob(
topk_index,
0,
Expand Down Expand Up @@ -659,14 +651,8 @@ def _draft_extend_for_decode(
draft_logits_output.hidden_states = draft_logits_output.hidden_states[
select_index
]
if self.topk == 1:
ret_topk_index = torch.argmax(
draft_logits_output.next_token_logits, dim=-1, keepdim=True
)
ret_topk_p = torch.ones_like(ret_topk_index, dtype=torch.float32)
else:
probs = torch.softmax(draft_logits_output.next_token_logits, dim=-1)
ret_topk_p, ret_topk_index = fast_topk(probs, self.topk, dim=-1)
probs = torch.softmax(draft_logits_output.next_token_logits, dim=-1)
ret_topk_p, ret_topk_index = fast_topk(probs, self.topk, dim=-1)
Comment on lines +654 to +655

@gemini-code-assist gemini-code-assist Bot May 26, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

medium

Apply the same top-1 softmax optimization here to avoid full-vocab softmax and fast_topk overhead when self.topk == 1.

Suggested change
probs = torch.softmax(draft_logits_output.next_token_logits, dim=-1)
ret_topk_p, ret_topk_index = fast_topk(probs, self.topk, dim=-1)
if self.topk == 1:
max_logits, ret_topk_index = torch.max(draft_logits_output.next_token_logits, dim=-1, keepdim=True)
ret_topk_p = 1.0 / torch.exp(draft_logits_output.next_token_logits - max_logits).sum(dim=-1, keepdim=True)
else:
probs = torch.softmax(draft_logits_output.next_token_logits, dim=-1)
ret_topk_p, ret_topk_index = fast_topk(probs, self.topk, dim=-1)

ret_hidden_states = draft_logits_output.hidden_states

# Construct the return values
Expand Down
Loading