[None][fix] Guard changeBeamWidth on mInflightReqIds.empty() in error handler

In pipeline-parallel multi-micro-batch mode, requests from other micro-batches
may still be tracked in mInflightReqIds after the error handler erases only the
current activeRequests.  changeBeamWidth asserts mInflightReqIds.empty(), so
calling it unconditionally would throw (caught by the inner try-catch) and skip
the intended buffer reset.

Add an explicit guard so the reset is skipped when other micro-batches are still
in-flight.  The next successful forwardAsync iteration will perform the reset via
the normal changeBeamWidth call in verifyRequests once the set is clear.  This
makes the skip explicit rather than relying on TLLM_CHECK as control flow.

Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com>

Author: achartier

cpp/tensorrt_llm/batch_manager/trtGptModelInflightBatching.cpp

@@ -1223,8 +1223,11 @@ void TrtGptModelInflightBatching::forwardAsync(RequestList const& activeRequests
             }
             // Force buffer/decoder reset to clean up any partial state from the aborted batch
             // (e.g. partially-filled cross-KV block offsets from mid-context-chunk processing).
-            // This prevents subsequent requests from reusing stale RuntimeBuffers.
-            if (mWorldConfig.isLastPipelineParallelRank())
+            // Guard on mInflightReqIds.empty(): in pipeline-parallel multi-micro-batch mode,
+            // other micro-batches may still have requests tracked here; changeBeamWidth asserts
+            // emptiness so we skip the reset and let the next successful forwardAsync iteration
+            // perform it when the set is clear.
+            if (mWorldConfig.isLastPipelineParallelRank() && mInflightReqIds.empty())
             {
                 changeBeamWidth(mOperatingBeamWidth);
             }