[NV] llm-d: move --moe-backend out of server.sh into per-recipe extra-args

Signed-off-by: Ezra Silvera <ezra@il.ibm.com>

Author: ezrasilvera

benchmarks/multi_node/llm-d-recipes/dsr1-fp8-h200-1p1d-simple.yaml

@@ -99,6 +99,7 @@ prefill:
     --block-size 256
     --no-enable-prefix-caching
     --enforce-eager
+    --moe-backend deep_gemm
   env: {}
 
 decode:
@@ -110,6 +111,7 @@ decode:
     --max-model-len 16384
     --block-size 256
     --no-enable-prefix-caching
+    --moe-backend deep_gemm
   env: {}
 
 # ---- SLURM resource directives ----

benchmarks/multi_node/llm-d-recipes/gptoss-fp4-h200-1p1d-simple.yaml

@@ -83,6 +83,7 @@ prefill:
     --max-model-len 16384
     --block-size 256
     --no-enable-prefix-caching
+    --moe-backend triton
   env: {}
 
 decode:
@@ -93,6 +94,7 @@ decode:
     --max-model-len 16384
     --block-size 256
     --no-enable-prefix-caching
+    --moe-backend triton
   env: {}
 
 # ---- SLURM resource directives ----

benchmarks/multi_node/llm-d/server.sh

@@ -155,8 +155,10 @@ COMMON_ARGS=(
     --tensor-parallel-size "$TP_SIZE"
     --data-parallel-size "$DP_SIZE"
     --kv_transfer_config "$KV_TRANSFER_CONFIG"
-    --moe-backend deep_gemm
 )
+# --moe-backend is model-specific (DSR1-FP8 wants deep_gemm, gpt-oss-MXFP4
+# rejects it - see vllm/.../oracle/mxfp4.py:163), so each recipe sets its
+# own value via prefill/decode extra-args instead of inheriting one here.
 
 if [[ "$LWS_GROUP_SIZE" -gt 1 ]]; then
     COMMON_ARGS+=(