[Core] Async Scheduling X Spec Decoding Compatibility (#24799)

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: Benjamin Chislett <chislett.ben@gmail.com>
Co-authored-by: Nick Hill <nhill@redhat.com>
Co-authored-by: Benjamin Chislett <chislett.ben@gmail.com>

vllm/v1/core/sched/scheduler.py

@@ -348,7 +348,10 @@ class Scheduler(SchedulerInterface):
             # Speculative decode related.
             if request.spec_token_ids:
                 num_scheduled_spec_tokens = (
-                    num_new_tokens + request.num_computed_tokens - request.num_tokens
+                    num_new_tokens
+                    + request.num_computed_tokens
+                    - request.num_tokens
+                    - request.num_output_placeholders
                 )
                 if num_scheduled_spec_tokens > 0:
                     # Trim spec_token_ids list to num_scheduled_spec_tokens.
@@ -1024,7 +1027,12 @@ class Scheduler(SchedulerInterface):
                 # tokens and rejections. If some tokens are rejected,
                 # num_computed_tokens is decreased by the number of rejected
                 # tokens.
-                request.num_computed_tokens -= num_rejected
+                if request.num_computed_tokens > 0:
+                    request.num_computed_tokens -= num_rejected
+                # If async scheduling, num_output_placeholders also includes
+                # the scheduled spec tokens count and so is similarly adjusted.
+                if request.num_output_placeholders > 0:
+                    request.num_output_placeholders -= num_rejected
                 spec_decoding_stats = self.make_spec_decoding_stats(
                     spec_decoding_stats,
                     num_draft_tokens=num_draft_tokens,