diff --git a/benchmarks/70b_fp4_b200_slurm.sh b/benchmarks/70b_fp4_b200_slurm.sh
index 110facaf5..14896c7bb 100644
--- a/benchmarks/70b_fp4_b200_slurm.sh
+++ b/benchmarks/70b_fp4_b200_slurm.sh
@@ -21,6 +21,18 @@ hf download $MODEL
 
 pip install datasets pandas
 
+nvidia-smi
+
+sed -i '102,108d' /usr/local/lib/python3.12/dist-packages/flashinfer/jit/cubin_loader.py
+
+# Calculate max-model-len based on ISL and OSL
+if [ "$ISL" = "1024" ] && [ "$OSL" = "1024" ]; then
+    CALCULATED_MAX_MODEL_LEN=$((ISL + OSL + 20))
+elif [ "$ISL" = "8192" ] || [ "$OSL" = "8192" ]; then
+    CALCULATED_MAX_MODEL_LEN=$((ISL + OSL + 200))
+else
+    CALCULATED_MAX_MODEL_LEN=${MAX_MODEL_LEN:-10240}  
+fi
 
 cat > config.yaml << EOF
 kv-cache-dtype: fp8
@@ -28,22 +40,17 @@ compilation-config: '{"pass_config":{"enable_fi_allreduce_fusion":true,"enable_a
 async-scheduling: true
 no-enable-prefix-caching: true
 max-num-batched-tokens: 8192
-max-model-len: 10240
+max-model-len: $CALCULATED_MAX_MODEL_LEN
 EOF
 
-SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
-PORT=$(( 8888 + $PORT_OFFSET ))
-
-
 export TORCH_CUDA_ARCH_LIST="10.0"
 export VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB='{"2":32,"4":32,"8":8}'
+export PYTHONNOUSERSITE=1
 
 set -x
-
-
-PYTHONNOUSERSITE=1 vllm serve $MODEL --host 0.0.0.0 --port $PORT --config config.yaml \
- --gpu-memory-utilization 0.9 --tensor-parallel-size $TP --max-num-seqs 512  \
- --disable-log-requests > $SERVER_LOG 2>&1 &
+vllm serve $MODEL --host 0.0.0.0 --port $PORT --config config.yaml \
+--gpu-memory-utilization 0.9 --tensor-parallel-size $TP --max-num-seqs 512 \
+--disable-log-requests > $SERVER_LOG 2>&1 &
 
 set +x
 while IFS= read -r line; do
diff --git a/benchmarks/70b_fp8_b200_slurm.sh b/benchmarks/70b_fp8_b200_slurm.sh
index 110facaf5..176a3a0a2 100644
--- a/benchmarks/70b_fp8_b200_slurm.sh
+++ b/benchmarks/70b_fp8_b200_slurm.sh
@@ -21,16 +21,28 @@ hf download $MODEL
 
 pip install datasets pandas
 
+nvidia-smi
 
-cat > config.yaml << EOF
+sed -i '102,108d' /usr/local/lib/python3.12/dist-packages/flashinfer/jit/cubin_loader.py
+
+
+FUSION_FLAG='{'\
+'"pass_config": {"enable_fi_allreduce_fusion": true, "enable_attn_fusion": true, "enable_noop": true},'\
+'"custom_ops": ["+quant_fp8", "+rms_norm"],'\
+'"cudagraph_mode": "FULL_DECODE_ONLY",'\
+'"splitting_ops": []'\
+'}'
+cat > config.yaml <<-EOF
 kv-cache-dtype: fp8
-compilation-config: '{"pass_config":{"enable_fi_allreduce_fusion":true,"enable_attn_fusion":true,"enable_noop":true},"custom_ops":["+quant_fp8","+rms_norm"],"cudagraph_mode":"FULL_DECODE_ONLY","splitting_ops":[]}'
+compilation-config: '$FUSION_FLAG'
 async-scheduling: true
 no-enable-prefix-caching: true
 max-num-batched-tokens: 8192
-max-model-len: 10240
+max-model-len: $MAX_MODEL_LEN
 EOF
 
+cat config.yaml  # Debugging
+
 SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
 PORT=$(( 8888 + $PORT_OFFSET ))
 
@@ -38,12 +50,15 @@ PORT=$(( 8888 + $PORT_OFFSET ))
 export TORCH_CUDA_ARCH_LIST="10.0"
 export VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB='{"2":32,"4":32,"8":8}'
 
-set -x
-
+export PYTHONNOUSERSITE=1
 
-PYTHONNOUSERSITE=1 vllm serve $MODEL --host 0.0.0.0 --port $PORT --config config.yaml \
- --gpu-memory-utilization 0.9 --tensor-parallel-size $TP --max-num-seqs 512  \
- --disable-log-requests > $SERVER_LOG 2>&1 &
+set -x
+vllm serve $MODEL --host=0.0.0.0 --port=$PORT \
+--gpu-memory-utilization=0.9 \
+--tensor-parallel-size=$TP \
+--max-num-seqs=512 \
+--config config.yaml \
+--disable-log-requests > $SERVER_LOG 2>&1 &
 
 set +x
 while IFS= read -r line; do
diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index d3a6b2222..f4a5175b0 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -9,16 +9,16 @@ set -x
 PORT=$(( 8888 + $PORT_OFFSET ))
 
 
-# Default: recv every ~10 requests; if CONC ≥ 16, relax to ~30 requests between scheduler recv polls.
-
-SCHEDULER_RECV_INTERVAL=10
+sed -i '102,108d' /usr/local/lib/python3.12/dist-packages/flashinfer/jit/cubin_loader.py
 
+# Default: recv every ~10 requests; if CONC ≥ 16, relax to ~30 requests between scheduler recv polls.
 if [[ $CONC -ge 16 ]]; then
-    SCHEDULER_RECV_INTERVAL=30
+  SCHEDULER_RECV_INTERVAL=30
+else
+  SCHEDULER_RECV_INTERVAL=10
 fi
 
 
-
 echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"
 
 
@@ -26,8 +26,8 @@ PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path $MODEL --host 0.
 --tensor-parallel-size=$TP --data-parallel-size=1 \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \
 --chunked-prefill-size 16384 \
---enable-ep-moe --quantization modelopt_fp4  --enable-flashinfer-allreduce-fusion --scheduler-recv-interval $SCHEDULER_RECV_INTERVAL \
---enable-symm-mem  --disable-radix-cache --attention-backend trtllm_mla --enable-flashinfer-trtllm-moe --stream-interval 10 \
+--enable-ep-moe --quantization modelopt_fp4 --enable-flashinfer-allreduce-fusion --scheduler-recv-interval $SCHEDULER_RECV_INTERVAL \
+--enable-symm-mem --disable-radix-cache --attention-backend trtllm_mla --enable-flashinfer-trtllm-moe --stream-interval 10 \
 > $SERVER_LOG 2>&1 &
 
 set +x
diff --git a/benchmarks/dsr1_fp8_b200_slurm.sh b/benchmarks/dsr1_fp8_b200_slurm.sh
index 28e9e2a32..e4247593d 100644
--- a/benchmarks/dsr1_fp8_b200_slurm.sh
+++ b/benchmarks/dsr1_fp8_b200_slurm.sh
@@ -7,13 +7,26 @@ SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
 
 set -x
 PORT=$(( 8888 + $PORT_OFFSET ))
-export SGL_ENABLE_JIT_DEEPGEMM=false 
+
+sed -i '102,108d' /usr/local/lib/python3.12/dist-packages/flashinfer/jit/cubin_loader.py
+
+export SGL_ENABLE_JIT_DEEPGEMM=false
 export SGLANG_ENABLE_FLASHINFER_GEMM=true
-python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
+
+# Default: recv every ~10 requests; if CONC ≥ 16, relax to ~30 requests between scheduler recv polls.
+if [[ $CONC -ge 16 ]]; then
+  SCHEDULER_RECV_INTERVAL=30
+else
+  SCHEDULER_RECV_INTERVAL=10
+fi
+
+set -x
+PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path=$MODEL --host=0.0.0.0 --port=$PORT \
 --tensor-parallel-size=$TP --data-parallel-size=1 \
---cuda-graph-max-bs 128 --max-running-requests 128 --mem-fraction-static 0.82 --kv-cache-dtype fp8_e4m3 \
---chunked-prefill-size 32768 --max-prefill-tokens 32768 \
---disable-radix-cache --attention-backend trtllm_mla --enable-flashinfer-trtllm-moe --stream-interval 1 \
+--cuda-graph-max-bs 128 --max-running-requests 128 \
+--mem-fraction-static 0.82 --kv-cache-dtype fp8_e4m3 --chunked-prefill-size 32768 --max-prefill-tokens 32768 \
+--enable-flashinfer-allreduce-fusion --scheduler-recv-interval $SCHEDULER_RECV_INTERVAL --disable-radix-cache \
+--attention-backend trtllm_mla --stream-interval 30 --enable-flashinfer-trtllm-moe --quantization fp8 \
 > $SERVER_LOG 2>&1 &
 
 set +x
diff --git a/benchmarks/gptoss_fp4_b200_slurm.sh b/benchmarks/gptoss_fp4_b200_slurm.sh
index 8e7161611..096602856 100644
--- a/benchmarks/gptoss_fp4_b200_slurm.sh
+++ b/benchmarks/gptoss_fp4_b200_slurm.sh
@@ -21,6 +21,19 @@ hf download $MODEL
 
 pip install datasets pandas
 
+nvidia-smi
+
+sed -i '102,108d' /usr/local/lib/python3.12/dist-packages/flashinfer/jit/cubin_loader.py
+
+
+# Calculate max-model-len based on ISL and OSL
+if [ "$ISL" = "1024" ] && [ "$OSL" = "1024" ]; then
+    CALCULATED_MAX_MODEL_LEN=$((ISL + OSL + 20))
+elif [ "$ISL" = "8192" ] || [ "$OSL" = "8192" ]; then
+    CALCULATED_MAX_MODEL_LEN=$((ISL + OSL + 200))
+else
+    CALCULATED_MAX_MODEL_LEN=${MAX_MODEL_LEN:-10240}  
+fi
 
 cat > config.yaml << EOF
 compilation-config: '{"pass_config":{"enable_fi_allreduce_fusion":true,"enable_attn_fusion":true,"enable_noop":true},"custom_ops":["+rms_norm"],"cudagraph_mode":"FULL_AND_PIECEWISE"}'
@@ -28,7 +41,7 @@ async-scheduling: true
 no-enable-prefix-caching: true
 cuda-graph-sizes: 2048
 max-num-batched-tokens: 8192
-max-model-len: 10240
+max-model-len: $CALCULATED_MAX_MODEL_LEN
 EOF
 
 SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)