code/RL_model/verl/verl_train/script/vllm.sh

# Support-check BN model (port 8090). reward_new_v6_bn.py uses VLLM_SUPPORT_CHECK_BN_API_BASE (default http://localhost:8090/v1).
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 python3 -m vllm.entrypoints.openai.api_server \
    --model /home/mshahidul/readctrl_model/support_checking_bn/gemma-3-4b-it \
    --gpu-memory-utilization 0.47 \
    --served-model-name support-check \
    --port 8090 \
    --max-model-len 8192 \
    --trust-remote-code \
    --tensor-parallel-size 1 \
    --enable-prefix-caching \
    --dtype bfloat16 \
    --max-num-seqs 256
    



# Classifier BN model (port 8040). reward_new_v6_bn.py uses VLLM_CLASSIFIER_BN_API_BASE (default http://localhost:8040/v1).
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 python3 -m vllm.entrypoints.openai.api_server \
    --model /home/mshahidul/readctrl_model/text_classifier_bn/gemma-3-4b-it \
    --served-model-name classifier \
    --gpu-memory-utilization 0.47 \
    --port 8040 \
    --max-model-len 8192 \
    --trust-remote-code \
    --tensor-parallel-size 1 \
    --enable-prefix-caching \
    --dtype bfloat16 \
    --max-num-seqs 256

# Qwen/Qwen3-30B-A3B-Instruct-2507
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=3 python3 -m vllm.entrypoints.openai.api_server \
    --model Qwen/Qwen3-30B-A3B-Instruct-2507 \
    --served-model-name subclaim-extractor \
    --gpu-memory-utilization 0.9 \
    --port 8051 \
    --max-model-len 16384 \
    --trust-remote-code \
    --tensor-parallel-size 1 \
    --enable-prefix-caching

# google/gemma-3-27b-it
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=5 python3 -m vllm.entrypoints.openai.api_server \
    --model google/gemma-3-27b-it \
    --served-model-name subclaim-extractor \
    --gpu-memory-utilization 0.9 \
    --port 8052 \
    --max-model-len 16384 \
    --trust-remote-code \
    --tensor-parallel-size 1 \
    --enable-prefix-caching
   

# Qwen/Qwen3-30B-A3B-Instruct-2507
# cyankiwi/Qwen3-Coder-Next-AWQ-4bit
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=1 vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507 \
  --max-model-len 16384 \
  --served-model-name newclaw \
  --enable-expert-parallel \
  --tensor-parallel-size 1 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml
  --dtype bfloat16 \
  --gpu-memory-utilization 0.9 \
  --port 8095 \
  --enable-reasoning \
  --reasoning-parser deepseek_r1


# Single file, default port 8050
python3 /home/mshahidul/readctrl/code/finetune-inference/subclaim_support_extraction/extract_bn_subclaims_vllm.py --input_file "/home/mshahidul/readctrl/data/translated_data/translation_testing_3396/multiclinsum_test_en2bn_gemma(0_1000)_3396.json" --port 8050

python3 /home/mshahidul/readctrl/code/finetune-inference/subclaim_support_extraction/extract_bn_subclaims_vllm.py --input_file "/home/mshahidul/readctrl/data/translated_data/translation_testing_3396/multiclinsum_test_en2bn_gemma(1000_2000)_3396.json" --port 8051

python3 /home/mshahidul/readctrl/code/finetune-inference/subclaim_support_extraction/extract_bn_subclaims_vllm.py --input_file "/home/mshahidul/readctrl/data/translated_data/translation_testing_3396/multiclinsum_test_en2bn_gemma(2000_3396)_3396.json" --port 8052