Fixing the free algorithm to handle times where the common prefix is

smaller.

backends/v3/src/radix.rs

@@ -172,7 +172,6 @@ impl Allocator for RadixAllocator {
                         )
                         // Unwrap, failing is a programming error.
                         .expect("Failed to store prefill tokens");
-
                     // We can have a prefill with the following structure:
                     //
                     // |---| From the prefix cache.
@@ -182,12 +181,14 @@ impl Allocator for RadixAllocator {
                     // This means that while processing this request there was a
                     // partially overlapping request that had A..=E in its
                     // prefill. In this case we need to free the blocks D E.
+                    if prefix_len > allocation.cached_prefix_len {
                         self.free_blocks.extend(
                             &blocks[allocation.cached_prefix_len / self.block_size as usize
                                 ..prefix_len / self.block_size as usize],
                         );
                     }
                 }
+            }
 
             // Free non-prefill blocks.
             self.free_blocks