[no merge] bump to nightly vllm and use vllm-router

.github/configs/amd-master.yaml

@@ -993,7 +993,7 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
         - "DECODE_MTP_SIZE=2"
 
 kimik2.5-fp4-mi355x-vllm-disagg:
-  image: vllm/vllm-openai-rocm:v0.18.0
+  image: vllm/vllm-openai-rocm:nightly-100c7b65e7579c8caf4ee0b04a6410b2796b905c
   model: amd/Kimi-K2.5-MXFP4
   model-prefix: kimik2.5
   runner: mi355x-disagg
@@ -1046,7 +1046,7 @@ kimik2.5-fp4-mi355x-vllm-disagg:
         - "DECODE_NODES=2"
 
 minimaxm2.5-fp8-mi355x-vllm-disagg:
-  image: vllm/vllm-openai-rocm:v0.18.0
+  image: vllm/vllm-openai-rocm:nightly-100c7b65e7579c8caf4ee0b04a6410b2796b905c
   model: MiniMaxAI/MiniMax-M2.5
   model-prefix: minimaxm2.5
   runner: mi355x-disagg

benchmarks/multi_node/amd_utils/env.sh

@@ -32,8 +32,13 @@ fi
 export IBDEVICES
 
 # Shared: Auto-detect default network interface (portable across clusters)
-export GLOO_SOCKET_IFNAME=$(ip route | grep '^default' | awk '{print $5}' | head -n 1)
-export NCCL_SOCKET_IFNAME=$(ip route | grep '^default' | awk '{print $5}' | head -n 1)
+# Only auto-detect if not already set by the runner/environment
+if [[ -z "$GLOO_SOCKET_IFNAME" ]]; then
+    export GLOO_SOCKET_IFNAME=$(ip route 2>/dev/null | grep '^default' | awk '{print $5}' | head -n 1)
+fi
+if [[ -z "$NCCL_SOCKET_IFNAME" ]]; then
+    export NCCL_SOCKET_IFNAME=$(ip route 2>/dev/null | grep '^default' | awk '{print $5}' | head -n 1)
+fi
 
 set +x
 

benchmarks/multi_node/amd_utils/job.slurm

@@ -416,6 +416,7 @@ if [[ \"$ENGINE\" == \"vllm-disagg\" && \"$ROUTER_TYPE\" == \"vllm-router\" && \
         \"$VLLM_ROUTER_IMAGE\" \
         bash -lc \"mkdir -p /run_logs/slurm_job-${SLURM_JOB_ID} && exec vllm-router \
             --vllm-pd-disaggregation \
+            --kv-connector moriio \
             --vllm-discovery-address 0.0.0.0:${PROXY_PING_PORT} \
             --port ${ROUTER_PORT} \
             --host 0.0.0.0 \

benchmarks/multi_node/amd_utils/patches/minimax_m2.py

@@ -137,7 +137,6 @@ def __init__(
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.intermediate_size,
-            reduce_results=False,
             renormalize=True,
             scoring_func=getattr(config, "scoring_func", "softmax"),
             e_score_correction_bias=self.e_score_correction_bias,
@@ -185,7 +184,8 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
             )
             final_hidden_states = final_hidden_states[:num_tokens]
         elif self.tp_size > 1:
-            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
+            from vllm.distributed.communication_op import tensor_model_parallel_all_reduce
+            final_hidden_states = tensor_model_parallel_all_reduce(
                 final_hidden_states
             )