[TRTLLM-9493][feat] Custom AllToAll for helix parallelism

Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>

Author: brb-nv

cpp/tensorrt_llm/common/envUtils.cpp

@@ -554,6 +554,12 @@ bool getEnvEplbForceGdrcopy()
     return getBoolEnv("TRTLLM_EPLB_FORCE_GDRCOPY");
 }
 
+bool getEnvUseNcclForHelix()
+{
+    static bool const useNcclForHelix = getBoolEnv("TRTLLM_USE_NCCL_FOR_HELIX");
+    return useNcclForHelix;
+}
+
 } // namespace common
 
 TRTLLM_NAMESPACE_END

cpp/tensorrt_llm/common/envUtils.h

@@ -161,6 +161,9 @@ bool getEnvKVCacheTransferAllBlocksForWindow();
 
 bool getEnvEplbForceGdrcopy();
 
+// Whether to use NCCL-based all-to-all for Helix parallelism.
+bool getEnvUseNcclForHelix();
+
 } // namespace common
 
 TRTLLM_NAMESPACE_END