Add cpath for mpi env var fix

Signed-off-by: Faraz Khoubsirat <58580514+farazkh80@users.noreply.github.com>

Author: farazkh80

docker/Dockerfile.multi

@@ -31,6 +31,11 @@ FROM base AS devel
 # NB: PyTorch requires this to be < 1.0
 ENV PYTORCH_CUDA_ALLOC_CONF="garbage_collection_threshold:0.99999"
 
+# Set CUDA include path and Triton PTXAS path for Triton runtime compilation
+# Required for Triton to find CUDA headers when compiling cuda_utils.c on remote MPI nodes
+ENV CPATH=/usr/local/cuda/include
+ENV TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas
+
 # Copy all installation scripts at once to reduce layers
 COPY docker/common/install.sh \
      docker/common/install_base.sh \

security_scanning/metadata.json

@@ -1,4 +1,4 @@
 {
   "commit_hash": "05b5336ab6135e368157600da1d16b090ee9a00a",
   "timestamp": "2025-11-14T18:16:21Z"
-}
+}

tensorrt_llm/_torch/modules/linear.py

@@ -74,7 +74,7 @@ def load_weight_shard(
         # For integrated GPU systems (e.g., DGX Spark), CPU and GPU share limited physical memory.
         # Avoiding device transfers reduces memory consumption and unnecessary data copies,
         # enabling support for larger models on memory-constrained systems.
-        logger.warning(
+        logger.debug(
             f"[load_weight_shard] Skipping device transfer from {weight.device} to {device} on integrated GPU to conserve shared memory."
         )
         device = weight.device