SemiAnalysisAI · cquil11 · Mar 9, 2026 · Mar 5, 2026 · Mar 5, 2026 · Mar 7, 2026
@@ -413,7 +413,7 @@ minimaxm2.5-fp8-mi325x-vllm:
     - { tp: 4, conc-start: 4, conc-end: 64 }
 
 gptoss-fp4-mi300x-vllm:
-  image: vllm/vllm-openai-rocm:v0.16.0
+  image: vllm/vllm-openai-rocm:v0.17.0
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: mi300x
@@ -444,7 +444,7 @@ gptoss-fp4-mi300x-vllm:
     - { tp: 8, conc-start: 4, conc-end: 16 }
 
 gptoss-fp4-mi325x-vllm:
-  image: vllm/vllm-openai-rocm:v0.16.0
+  image: vllm/vllm-openai-rocm:v0.17.0
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: mi325x
@@ -475,8 +475,8 @@ gptoss-fp4-mi325x-vllm:
     - { tp: 8, conc-start: 4, conc-end: 16 }
 
 gptoss-fp4-mi355x-vllm:
-  image: vllm/vllm-openai-rocm:v0.16.0
-  model: openai/gpt-oss-120b
+  image: vllm/vllm-openai-rocm:v0.17.0
+  model: amd/gpt-oss-120b-w-mxfp4-a-fp8
   model-prefix: gptoss
   runner: mi355x
   precision: fp4

diff --git a/benchmarks/single_node/gptoss_fp4_mi300x.sh b/benchmarks/single_node/gptoss_fp4_mi300x.sh
@@ -33,23 +33,23 @@ if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
     export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
 fi
 
+export AMDGCN_USE_BUFFER_OPS=0
 export VLLM_ROCM_USE_AITER=1
-export VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1
-export VLLM_ROCM_USE_AITER_MHA=0
 export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+ATTN_BACKEND="--attention-backend ROCM_AITER_UNIFIED_ATTN"
+FUSE_ROPE_KVCACHE="-cc.pass_config.fuse_rope_kvcache=True -cc.use_inductor_graph_partition=True"
 
 SERVER_LOG=/workspace/server.log
 PORT=${PORT:-8888}
 
 set -x
 vllm serve $MODEL --port $PORT \
---tensor-parallel-size=$TP \
---gpu-memory-utilization 0.95 \
---max-model-len $MAX_MODEL_LEN \
---compilation-config  '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
---block-size=64 \
---no-enable-prefix-caching \
---disable-log-requests > $SERVER_LOG 2>&1 &
+  $ATTN_BACKEND $FUSE_ROPE_KVCACHE \
+  --tensor-parallel-size=$TP \
+  --gpu-memory-utilization 0.95 \
+  --max-model-len $MAX_MODEL_LEN \
+  --block-size=64 \
+  --no-enable-prefix-caching > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 

diff --git a/benchmarks/single_node/gptoss_fp4_mi325x.sh b/benchmarks/single_node/gptoss_fp4_mi325x.sh
@@ -33,22 +33,23 @@ if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
     export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
 fi
 
+export AMDGCN_USE_BUFFER_OPS=0
 export VLLM_ROCM_USE_AITER=1
-export VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1
-export VLLM_ROCM_USE_AITER_MHA=0
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+ATTN_BACKEND="--attention-backend ROCM_AITER_UNIFIED_ATTN"
+FUSE_ROPE_KVCACHE="-cc.pass_config.fuse_rope_kvcache=True -cc.use_inductor_graph_partition=True"
 
 SERVER_LOG=/workspace/server.log
 PORT=${PORT:-8888}
 
 set -x
 vllm serve $MODEL --port $PORT \
---tensor-parallel-size=$TP \
---gpu-memory-utilization 0.95 \
---max-model-len $MAX_MODEL_LEN \
---compilation-config  '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
---block-size=64 \
---no-enable-prefix-caching \
---disable-log-requests > $SERVER_LOG 2>&1 &
+  $ATTN_BACKEND $FUSE_ROPE_KVCACHE \
+  --tensor-parallel-size=$TP \
+  --gpu-memory-utilization 0.95 \
+  --max-model-len $MAX_MODEL_LEN \
+  --block-size=64 \
+  --no-enable-prefix-caching > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 

diff --git a/benchmarks/single_node/gptoss_fp4_mi355x.sh b/benchmarks/single_node/gptoss_fp4_mi355x.sh
@@ -33,22 +33,24 @@ if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
     export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
 fi
 
+export AMDGCN_USE_BUFFER_OPS=0
 export VLLM_ROCM_USE_AITER=1
-export VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1
-export VLLM_ROCM_USE_AITER_MHA=0
+export VLLM_ROCM_USE_AITER_TRITON_ROPE=1
+export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+ATTN_BACKEND="--attention-backend ROCM_AITER_UNIFIED_ATTN"
+FUSE_ROPE_KVCACHE="-cc.pass_config.fuse_rope_kvcache=True -cc.use_inductor_graph_partition=True"
 
 SERVER_LOG=/workspace/server.log
 PORT=${PORT:-8888}
 
 set -x
 vllm serve $MODEL --port $PORT \
---tensor-parallel-size=$TP \
---gpu-memory-utilization 0.95 \
---max-model-len $MAX_MODEL_LEN \
---compilation-config  '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
---block-size=64 \
---no-enable-prefix-caching \
---disable-log-requests > $SERVER_LOG 2>&1 &
+  $ATTN_BACKEND $FUSE_ROPE_KVCACHE \
+  --tensor-parallel-size=$TP \
+  --gpu-memory-utilization 0.95 \
+  --max-model-len $MAX_MODEL_LEN \
+  --block-size=64 \
+  --no-enable-prefix-caching > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
@@ -911,3 +911,17 @@
     - "Redo qwen eval"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/892
   evals-only: true
+
+- config-keys:
+    - gptoss-fp4-mi300x-vllm
+    - gptoss-fp4-mi325x-vllm
+    - gptoss-fp4-mi355x-vllm
+  description:
+    - "Update AMD GPT-OSS vLLM image from v0.16.0 to v0.17.0 for MI300X, MI325X, and MI355X"
+    - "MI355X: Switch model to amd/gpt-oss-120b-w-mxfp4-a-fp8 (MXFP4 weights + FP8 activations)"
+    - "MI355X: Add VLLM_ROCM_USE_AITER_TRITON_ROPE=1 for AITER triton RoPE kernel"
+    - "Add AMDGCN_USE_BUFFER_OPS=0 and VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 env vars"
+    - "Switch to --attention-backend ROCM_AITER_UNIFIED_ATTN and add fuse_rope_kvcache compilation pass"
+    - "Remove deprecated VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION/VLLM_ROCM_USE_AITER_MHA env vars and compilation-config cudagraph_mode"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/867
+
diff --git a/runners/launch_mi355x-amds.sh b/runners/launch_mi355x-amds.sh
@@ -156,23 +156,30 @@ else
 
     PARTITION="compute"
     SQUASH_FILE="/var/lib/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+    LOCK_FILE="${SQUASH_FILE}.lock"
 
     set -x
     salloc --partition=$PARTITION --gres=gpu:$TP --cpus-per-task=128 --time=180 --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -h -o %A | head -n1)
 
     srun --jobid=$JOB_ID bash -c "docker stop \$(docker ps -a -q)"
 
-    if [[ "$FRAMEWORK" == "atom" ]]; then
-        srun --jobid=$JOB_ID bash -c "rm $SQUASH_FILE"
-    fi
+    # Use flock to serialize concurrent imports to the same squash file
+    srun --jobid=$JOB_ID bash -c "
+        exec 9>\"$LOCK_FILE\"
+        flock -w 600 9 || { echo 'Failed to acquire lock for $SQUASH_FILE'; exit 1; }
+        if [[ \"$FRAMEWORK\" == \"atom\" ]]; then
+            rm -f \"$SQUASH_FILE\"
+        fi
+        if unsquashfs -l \"$SQUASH_FILE\" > /dev/null 2>&1; then
+            echo 'Squash file already exists and is valid, skipping import'
+        else
+            rm -f \"$SQUASH_FILE\"
+            enroot import -o \"$SQUASH_FILE\" docker://$IMAGE
+        fi
+    "
 
-    srun --jobid=$JOB_ID bash -c "enroot import -o $SQUASH_FILE docker://$IMAGE"
-    if ! srun --jobid=$JOB_ID bash -c "unsquashfs -l $SQUASH_FILE > /dev/null"; then
-        echo "unsquashfs failed, removing $SQUASH_FILE and re-importing..."
-        srun --jobid=$JOB_ID bash -c "rm -f $SQUASH_FILE"
-        srun --jobid=$JOB_ID bash -c "enroot import -o $SQUASH_FILE docker://$IMAGE"
-    fi
+    export VLLM_CACHE_ROOT="/it-share/gharunners/.cache/vllm"
 
     srun --jobid=$JOB_ID \
         --container-image=$SQUASH_FILE \