[crashfix] Eagle + multimodal can crash on mm cache miss (#29750)

Signed-off-by: Mickael Seznec <[email protected]>
Co-authored-by: Roger Wang <[email protected]>

Author: mickaelseznec

vllm/v1/core/sched/scheduler.py

@@ -263,6 +263,7 @@ def schedule(self) -> SchedulerOutput:
                     request.num_computed_tokens,
                     num_new_tokens,
                     encoder_compute_budget,
+                    shift_computed_tokens=1 if self.use_eagle else 0,
                 )
 
             if num_new_tokens == 0:
@@ -532,6 +533,7 @@ def schedule(self) -> SchedulerOutput:
                             num_computed_tokens,
                             num_new_tokens,
                             encoder_compute_budget,
+                            shift_computed_tokens=1 if self.use_eagle else 0,
                         )
                         if num_new_tokens == 0:
                             # The request cannot be scheduled.
@@ -829,6 +831,7 @@ def _try_schedule_encoder_inputs(
         num_computed_tokens: int,
         num_new_tokens: int,
         encoder_compute_budget: int,
+        shift_computed_tokens: int = 0,
     ) -> tuple[list[int], int, int, list[int]]:
         """
         Determine which encoder inputs need to be scheduled in the current step,
@@ -873,7 +876,10 @@ def _try_schedule_encoder_inputs(
             # The encoder output is needed if the two ranges overlap:
             # [num_computed_tokens, num_computed_tokens + num_new_tokens) and
             # [start_pos, start_pos + num_encoder_tokens)
-            if start_pos >= num_computed_tokens + num_new_tokens:
+            if (
+                start_pos
+                >= num_computed_tokens + num_new_tokens + shift_computed_tokens
+            ):
                 # The encoder input is not needed in this step.
                 break
 
@@ -929,10 +935,12 @@ def _try_schedule_encoder_inputs(
                 # NOTE(woosuk): We assume that the encoder input tokens should
                 # be processed altogether, as the encoder usually uses
                 # bidirectional attention.
-                if num_computed_tokens < start_pos:
+                if num_computed_tokens + shift_computed_tokens < start_pos:
                     # We only schedule the decoder tokens just before the
                     # encoder input.
-                    num_new_tokens = start_pos - num_computed_tokens
+                    num_new_tokens = start_pos - (
+                        num_computed_tokens + shift_computed_tokens
+                    )
                 else:
                     # Because of prefix caching, num_computed_tokens is greater
                     # than start_pos even though its encoder input is not