Debug [skip ci]

Author: UNIDY2002

mooncake-ep/src/mooncake_backend.cpp

@@ -1069,30 +1069,39 @@ void MooncakeBackend::advanceRecvOp(RecvOpStateData& opData) {
                 }
             }
 
+            // If we can allocate based on sequence/capacity, check if there's actually a free slot
             if (canAllocate) {
-                opData.state = RecvOpState::READY_TO_ALLOC;
+                // Try to find an available slot that doesn't conflict with other operations
+                int startSlot = opData.baseSlot;
+                int attempts = 0;
+                bool foundSlot = false;
+                
+                while (attempts < static_cast<int>(kP2PNumSlots)) {
+                    // Check if current slot is free and doesn't cause wrap-around
+                    if (allocatedRecvSlots_.find(opData.baseSlot) == allocatedRecvSlots_.end() &&
+                        opData.baseSlot + opData.numSlotsNeeded <= static_cast<int>(kP2PNumSlots)) {
+                        // Found a valid slot
+                        foundSlot = true;
+                        break;
+                    }
+                    // Try next slot, wrapping around if needed
+                    opData.baseSlot = (opData.baseSlot + 1) % static_cast<int>(kP2PNumSlots);
+                    attempts++;
+                }
+                
+                if (foundSlot) {
+                    opData.state = RecvOpState::READY_TO_ALLOC;
+                }
+                // If no slot found, stay in WAITING_SLOT_AVAIL and wait for slots to be freed
             }
             break;
         }
 
         case RecvOpState::READY_TO_ALLOC: {
-            // Check if the slot is already allocated by another active operation
-            // If so, find the next available slot that doesn't cause wrap-around
-            int attempts = 0;
-            while (attempts < static_cast<int>(kP2PNumSlots)) {
-                // Check if current slot is free and doesn't cause wrap-around
-                if (allocatedRecvSlots_.find(opData.baseSlot) == allocatedRecvSlots_.end() &&
-                    opData.baseSlot + opData.numSlotsNeeded <= static_cast<int>(kP2PNumSlots)) {
-                    // Found a valid slot
-                    break;
-                }
-                // Try next slot, wrapping around if needed
-                opData.baseSlot = (opData.baseSlot + 1) % static_cast<int>(kP2PNumSlots);
-                attempts++;
-            }
-            
-            TORCH_CHECK(attempts < static_cast<int>(kP2PNumSlots),
-                        "P2P recv: failed to find available slot after ", attempts, " attempts");
+            // At this point, we've already found a valid slot in WAITING_SLOT_AVAIL
+            // Just verify it's still available and allocate it
+            TORCH_CHECK(allocatedRecvSlots_.find(opData.baseSlot) == allocatedRecvSlots_.end(),
+                        "P2P recv: slot ", opData.baseSlot, " was allocated by another operation");
             TORCH_CHECK(opData.baseSlot + opData.numSlotsNeeded <= static_cast<int>(kP2PNumSlots),
                         "P2P recv: slot range would exceed buffer: baseSlot=", opData.baseSlot,
                         ", numSlots=", opData.numSlotsNeeded);