[Perf] Use bisect for mm feature lookup in model runner v2 (#45566)

Signed-off-by: Roger Wang <hey@rogerw.io>

vllm/v1/worker/gpu/mm/encoder_runner.py

@@ -5,7 +5,7 @@ import torch
 
 from vllm.model_executor.models.interfaces import SupportsMultiModal
 from vllm.multimodal.inputs import MultiModalKwargsItem
-from vllm.multimodal.utils import group_and_batch_mm_kwargs
+from vllm.multimodal.utils import get_mm_features_in_window, group_and_batch_mm_kwargs
 from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache
 from vllm.v1.worker.utils import sanity_check_mm_encoder_outputs
 
@@ -91,19 +91,17 @@ class EncoderRunner:
                 continue
 
             mm_features = self.encoder_cache.mm_features[req_id]
-            for mm_feature in mm_features:
+            lo, hi = get_mm_features_in_window(
+                mm_features,
+                start=query_start[i],
+                end=query_end[i],
+            )
+            for idx in range(lo, hi):
+                mm_feature = mm_features[idx]
                 pos_info = mm_feature.mm_position
                 start_pos = pos_info.offset
                 num_encoder_tokens = pos_info.length
 
-                if start_pos >= query_end[i]:
-                    # The encoder output is not needed in this step.
-                    break
-                if start_pos + num_encoder_tokens <= query_start[i]:
-                    # The encoder output is already processed and stored
-                    # in the decoder's KV cache.
-                    continue
-
                 start_idx = max(query_start[i] - start_pos, 0)
                 end_idx = min(query_end[i] - start_pos, num_encoder_tokens)
                 assert start_idx < end_idx