Qwen-next script

script_qwen/qwen.sh

@@ -0,0 +1,41 @@
+export HF_HOME=/mnt/raid0/models/
+export VLLM_USE_V1=1
+export SAFETENSORS_FAST_GPU=1
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MOE=1
+export VLLM_USE_TRITON_FLASH_ATTN=0
+export VLLM_ROCM_USE_AITER_MHA=1
+
+TP=8
+EP=8
+PREFIX=/home/zhiwei/
+MODEL_NAME="qwen3-480b-ptpc"
+LOG_PATH="${HOME}/logs/${MODEL_NAME}"
+# mkdir -p ${PROFILE_RESULT_PATH}
+# export VLLM_TORCH_PROFILER_DIR=${PROFILE_RESULT_PATH}
+# export VLLM_TORCH_PROFILER_WITH_STACK=1
+# export VLLM_TORCH_PROFILER_RECORD_SHAPES=1
+
+mkdir -p ${PREFIX}/temp/vllm
+mkdir -p ${PREFIX}/temp/inductor
+rm -rf ${PREFIX}/temp/vllm/*
+rm -rf ${PREFIX}/temp/inductor/*
+export VLLM_CACHE_ROOT=${PREFIX}/temp/vllm
+export TORCHINDUCTOR_CACHE_DIR=${PREFIX}/temp/inductor
+
+SERVER_LOG_FILE="${LOG_PATH}/tp${TP}_ep${EP}_server.log"
+
+#vllm serve /mnt/raid0/models/Qwen3-Coder-480B-A35B-Instruct-FP8-ptpc \
+#--trust-remote-code \
+#--disable-log-requests \
+#--max-model-len 32768 \
+#--tensor-parallel-size ${TP} \
+#--max_seq_len_to_capture 32768 \
+#--no-enable-prefix-caching \
+#--enable-expert-parallel \
+#--compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
+#--max_num_batched_tokens 32768 2>&1 | tee ${SERVER_LOG_FILE}
+
+
+VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve Qwen/Qwen3-Next-80B-A3B-Instruct --port 8000 --tensor-parallel-size 4 --max-model-len 262144 --block-size=1 --enforce-eager
+