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

python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py

@@ -30,12 +30,15 @@
 from sglang.srt.speculative.eagle_info import EagleDraftExtendInput
 from sglang.srt.speculative.spec_utils import fast_topk
 from sglang.srt.utils import (
+    is_hip,
     require_attn_tp_gather,
     require_gathered_buffer,
     require_mlp_sync,
     require_mlp_tp_gather,
 )
 
+_is_hip = is_hip()
+
 if TYPE_CHECKING:
     from sglang.srt.speculative.eagle_worker import EAGLEWorker
 
@@ -401,8 +404,17 @@ def run_once():
                 forward_batch.positions,
                 forward_batch,
             )
-            probs = torch.softmax(ret.next_token_logits, dim=-1)
-            ret.topk_p, ret.topk_index = fast_topk(probs, self.topk, dim=-1)
+            # ROCm's argmax tie-breaks differently from CUDA's softmax+max
+            # path on FP8 logits, which corrupts MTP draft selection on AMD.
+            # Keep the fastpath CUDA-only.
+            if self.topk == 1 and not _is_hip:
+                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)
 
             forward_batch.out_cache_loc = output_cache_loc_backup
             forward_batch.spec_info.hidden_states = hidden_states_backup

python/sglang/srt/speculative/eagle_worker_v2.py

@@ -487,8 +487,19 @@ def draft_forward(self, forward_batch: ForwardBatch):
                 ).logits_output
             maybe_detect_nan(logits_output.next_token_logits, f"draft_forward step {i}")
             maybe_detect_inf(logits_output.next_token_logits, f"draft_forward step {i}")
-            probs = torch.softmax(logits_output.next_token_logits, dim=-1)
-            topk_p, topk_index = fast_topk(probs, self.topk, dim=-1)
+            if self.topk == 1 and not _is_hip:
+                # topk=1 → degenerate single-path tree; `topk_p` is unused
+                # downstream, so skip softmax and just argmax over logits.
+                # Gated to CUDA: on ROCm the argmax tie-break diverges from
+                # the softmax+max path on FP8 logits and corrupts MTP draft
+                # selection (DSV3.2 MTP GSM8K, see #26358).
+                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)
             maybe_detect_oob(
                 topk_index,
                 0,
@@ -660,8 +671,16 @@ def _draft_extend_for_decode(
             draft_logits_output.hidden_states = draft_logits_output.hidden_states[
                 select_index
             ]
-        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 and not _is_hip:
+            # Gated to CUDA: see #26358 — ROCm's argmax tie-break corrupts
+            # MTP draft selection on FP8 logits.
+            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)
         ret_hidden_states = draft_logits_output.hidden_states
 
         # Construct the return values