[MLA] Simplification to batch P/D reordering (vllm-project#16673)

Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: Yang Wang <elainewy@meta.com>

Author: njhill

vllm/v1/attention/backends/mla/common.py

@@ -415,20 +415,18 @@ def reorder_batch(self, input_batch: "InputBatch",
         # the above loop
         num_decodes = len(decodes)
         num_prefills = len(prefills)
-        first_prefill = 0
         modified_batch = False
 
         for i in range(1, min(num_decodes, num_prefills) + 1):
             # If the decode is at the "back" of the batch, i, we can swap it
             # with the prefill closest to the front of the batch
-            if decodes[num_decodes - i] >= num_decodes:
-                input_batch.swap_states(prefills[first_prefill],
-                                        decodes[num_decodes - i])
-                first_prefill += 1
-                modified_batch = True
-            else:
+            decode_idx = decodes[num_decodes - i]
+            if decode_idx < num_decodes:
                 break
 
+            input_batch.swap_states(prefills[i - 1], decode_idx)
+            modified_batch = True
+
         # Save for next `build` call
         # TODO(lucas): this is a bit of a hack, we should probably have a
         # better way of doing this

vllm/v1/worker/gpu_model_runner.py

@@ -458,7 +458,13 @@ def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
         if removed_req_indices:
             self.input_batch.condense(removed_req_indices)
 
-        if batch_changed:
+        # Some attention backends (namely MLA) may want to separate requests
+        # based on if the attention computation will be compute-bound or
+        # memory-bound. This gives them a hook to do that.
+        batch_reordered = self.attn_metadata_builder.reorder_batch(
+            self.input_batch, scheduler_output)
+
+        if batch_changed or batch_reordered:
             self.input_batch.refresh_sampling_metadata()
 
     def _prepare_inputs(
@@ -471,14 +477,6 @@ def _prepare_inputs(
         num_reqs = self.input_batch.num_reqs
         assert num_reqs > 0
 
-        # Some attention backends (namely MLA) may want to separate requests
-        # based on if the attention computation will be compute-bound or
-        # memory-bound. This gives them a hook to do that.
-        modified_batch = self.attn_metadata_builder.reorder_batch(
-            self.input_batch, scheduler_output)
-        if modified_batch:
-            self.input_batch.refresh_sampling_metadata()
-
         # OPTIMIZATION: Start copying the block table first.
         # This way, we can overlap the copy with the following CPU operations.
         self.input_batch.block_table.commit(num_reqs)