diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 58469dcc4..9386d9c27 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2666,7 +2666,7 @@ minimaxm3-fp4-mi355x-vllm:
 # tokens. Search space mirrors the MI355X MXFP8 MTP entry, trimming the base
 # FP4 sweep at extreme concurrency where speculative decoding loses value.
 minimaxm3-fp4-mi355x-vllm-mtp:
-  image: vllm/vllm-openai-rocm:nightly-3f5a1e1733200760169ff31ebe60a271072b199e
+  image: vllm/vllm-openai-rocm:nightly-4559c43a9526597c00cbcc4f59979496500268d1
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_vllm_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_vllm_mtp.sh
index 96a560493..8a15b8c89 100755
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_vllm_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_vllm_mtp.sh
@@ -36,6 +36,18 @@ fi
 SERVER_LOG=/workspace/server.log
 export VLLM_ENGINE_READY_TIMEOUT_S=3600
 export VLLM_USE_BREAKABLE_CUDAGRAPH=0
+# Use AITER MoE for the MXFP4 experts, matching minimaxm3_fp4_mi355x_vllm.sh.
+# This is required for ALL configs including expert parallelism: with EP enabled
+# and moe_backend=auto, the AITER MXFP4 backend is skipped and selection falls
+# back to Mxfp4MoeBackend.EMULATION, which triggers a first-time build of the
+# Quark hw-emulation C++ kernel (kernel_ext, 9 ROCm arches) on every worker at
+# warmup. Concurrent EP workers deadlock on the shared torch_extensions build
+# lock, hanging engine-core for hours. Forcing --moe-backend aiter selects the
+# AITER_MXFP4_MXFP4 backend instead (verified working under TP4+EP4 with EAGLE3
+# spec decoding), avoiding the emulation build entirely.
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MOE=1
+export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1
 
 if [ "${EVAL_ONLY}" = "true" ]; then
     setup_eval_context
@@ -65,6 +77,7 @@ vllm serve "$MODEL" --port "$PORT" \
     --language-model-only \
     --max-model-len "$MAX_MODEL_LEN" \
     --attention-backend TRITON_ATTN \
+    --moe-backend aiter \
     --speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS}" \
     --tool-call-parser minimax_m3 \
     --enable-auto-tool-choice \
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 068e37a7e..450226250 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4344,6 +4344,14 @@
     - "Reuse the existing MXFP8 B300 topology and concurrency matrix across 15 srt-slurm recipes, while dropping the FP8-only Marlin override from TP4 decode"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1931
 
+- config-keys:
+    - minimaxm3-fp4-mi355x-vllm-mtp
+  description:
+    - "Enable AITER MoE on MiniMax-M3 MXFP4 MI355X single-node vLLM MTP (EAGLE3), mirroring the STP recipe: export VLLM_ROCM_USE_AITER=1, VLLM_ROCM_USE_AITER_MOE=1, VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1; pass --moe-backend aiter unconditionally (including expert parallelism)."
+    - "Fixes the ~8h engine-core startup hang on EP configs: with moe_backend=auto, EP fell back to Mxfp4MoeBackend.EMULATION, which deadlocked all expert-parallel workers building the Quark hw-emulation C++ kernel into a shared torch_extensions dir. Forcing --moe-backend aiter selects AITER_MXFP4_MXFP4 (no emulation build)."
+    - "Pin vllm/vllm-openai-rocm:nightly-4559c43a9526597c00cbcc4f59979496500268d1 (from nightly-3f5a1e1733200760169ff31ebe60a271072b199e), matching the STP recipe."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1964
+
 - config-keys:
     - minimaxm3-fp4-b300-dynamo-vllm
   description: