INFO 07-06 02:07:56 [__init__.py:239] Automatically detected platform cuda.
INFO 07-06 02:07:58 [config.py:209] Replacing legacy 'type' key with 'rope_type'
INFO 07-06 02:07:58 [config.py:2968] Downcasting torch.float32 to torch.float16.
INFO 07-06 02:08:05 [config.py:717] This model supports multiple tasks: {'classify', 'generate', 'embed', 'reward', 'score'}. Defaulting to 'generate'.
INFO 07-06 02:08:05 [config.py:1770] Defaulting to use mp for distributed inference
INFO 07-06 02:08:05 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
INFO 07-06 02:08:07 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./merged1/phi_darelinear_7', speculative_config=None, tokenizer='./merged1/phi_darelinear_7', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='auto', reasoning_backend=None), observability_config=ObservabilityConfig(show_hidden_metrics=False, otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=None, served_model_name=./merged1/phi_darelinear_7, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, disable_mm_preprocessor_cache=False, mm_processor_kwargs=None, pooler_config=None, compilation_config={"level":3,"custom_ops":["none"],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output"],"use_inductor":true,"compile_sizes":[],"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"max_capture_size":512}
WARNING 07-06 02:08:07 [multiproc_worker_utils.py:306] Reducing Torch parallelism from 128 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
INFO 07-06 02:08:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_1bcb022b'), local_subscribe_addr='ipc:///tmp/ee948401-c9f8-41a2-a34e-44dd790f1865', remote_subscribe_addr=None, remote_addr_ipv6=False)
WARNING 07-06 02:08:07 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x14903a1a7fd0>
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_bfa2d2a8'), local_subscribe_addr='ipc:///tmp/86d51acd-0efd-4dea-8ca3-c4f87555bace', remote_subscribe_addr=None, remote_addr_ipv6=False)
WARNING 07-06 02:08:07 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x149038780f10>
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_79dfef4a'), local_subscribe_addr='ipc:///tmp/1c167278-2df4-417a-b744-d664c84373d5', remote_subscribe_addr=None, remote_addr_ipv6=False)
WARNING 07-06 02:08:07 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x14903a1a7d30>
WARNING 07-06 02:08:07 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x14903a1a7310>
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_79b3846e'), local_subscribe_addr='ipc:///tmp/ae256d6f-7965-455a-8523-b801fb96df48', remote_subscribe_addr=None, remote_addr_ipv6=False)
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_2861b1af'), local_subscribe_addr='ipc:///tmp/8e727f98-dd48-43de-8f1d-86cbb27336b8', remote_subscribe_addr=None, remote_addr_ipv6=False)
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:18 [utils.py:1055] Found nccl from library libnccl.so.2
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:18 [utils.py:1055] Found nccl from library libnccl.so.2
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:18 [pynccl.py:69] vLLM is using nccl==2.21.5
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:18 [pynccl.py:69] vLLM is using nccl==2.21.5
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:19 [utils.py:1055] Found nccl from library libnccl.so.2
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:19 [pynccl.py:69] vLLM is using nccl==2.21.5
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:19 [utils.py:1055] Found nccl from library libnccl.so.2
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:19 [pynccl.py:69] vLLM is using nccl==2.21.5
(VllmWorker rank=2 pid=3923543) WARNING 07-06 02:08:19 [custom_all_reduce.py:136] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
(VllmWorker rank=3 pid=3923544) WARNING 07-06 02:08:19 [custom_all_reduce.py:136] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
(VllmWorker rank=1 pid=3923542) WARNING 07-06 02:08:19 [custom_all_reduce.py:136] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
(VllmWorker rank=0 pid=3923541) WARNING 07-06 02:08:19 [custom_all_reduce.py:136] Custom allreduce is disabled because it's not supported on more than two PCIe-only GPUs. To silence this warning, specify disable_custom_all_reduce=True explicitly.
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:19 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_f31c9e4e'), local_subscribe_addr='ipc:///tmp/051d2fb3-e6ad-4c1d-b095-62c5dea7f720', remote_subscribe_addr=None, remote_addr_ipv6=False)
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:19 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:19 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:19 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:19 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:19 [cuda.py:221] Using Flash Attention backend on V1 engine.
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:19 [cuda.py:221] Using Flash Attention backend on V1 engine.
(VllmWorker rank=2 pid=3923543) WARNING 07-06 02:08:19 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
(VllmWorker rank=3 pid=3923544) WARNING 07-06 02:08:19 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:19 [cuda.py:221] Using Flash Attention backend on V1 engine.
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:19 [cuda.py:221] Using Flash Attention backend on V1 engine.
(VllmWorker rank=1 pid=3923542) WARNING 07-06 02:08:19 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
(VllmWorker rank=0 pid=3923541) WARNING 07-06 02:08:19 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:19 [gpu_model_runner.py:1329] Starting to load model ./merged1/phi_darelinear_7...
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:19 [gpu_model_runner.py:1329] Starting to load model ./merged1/phi_darelinear_7...
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:19 [gpu_model_runner.py:1329] Starting to load model ./merged1/phi_darelinear_7...
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:19 [gpu_model_runner.py:1329] Starting to load model ./merged1/phi_darelinear_7...
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:21 [loader.py:458] Loading weights took 1.67 seconds
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:21 [loader.py:458] Loading weights took 2.03 seconds
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:21 [loader.py:458] Loading weights took 1.98 seconds
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:21 [loader.py:458] Loading weights took 2.04 seconds
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:21 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 1.882851 seconds
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:22 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 2.218165 seconds
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:22 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 2.216825 seconds
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:22 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 2.216819 seconds
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:27 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/1a71ee31dc/rank_3_0 for vLLM's torch.compile
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:27 [backends.py:430] Dynamo bytecode transform time: 5.57 s
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:27 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/1a71ee31dc/rank_2_0 for vLLM's torch.compile
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:27 [backends.py:430] Dynamo bytecode transform time: 5.58 s
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:27 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/1a71ee31dc/rank_1_0 for vLLM's torch.compile
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:27 [backends.py:430] Dynamo bytecode transform time: 5.63 s
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:27 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/1a71ee31dc/rank_0_0 for vLLM's torch.compile
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:27 [backends.py:430] Dynamo bytecode transform time: 5.69 s
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:32 [backends.py:136] Cache the graph of shape None for later use
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:32 [backends.py:136] Cache the graph of shape None for later use
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:32 [backends.py:136] Cache the graph of shape None for later use
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:32 [backends.py:136] Cache the graph of shape None for later use
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:08:53 [backends.py:148] Compiling a graph for general shape takes 24.72 s
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:08:53 [backends.py:148] Compiling a graph for general shape takes 24.79 s
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:08:53 [backends.py:148] Compiling a graph for general shape takes 24.92 s
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:08:53 [backends.py:148] Compiling a graph for general shape takes 24.81 s
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:09:15 [monitor.py:33] torch.compile takes 30.34 s in total
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:09:15 [monitor.py:33] torch.compile takes 30.37 s in total
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:09:15 [monitor.py:33] torch.compile takes 30.49 s in total
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:09:15 [monitor.py:33] torch.compile takes 30.50 s in total
INFO 07-06 02:09:16 [kv_cache_utils.py:634] GPU KV cache size: 1,999,536 tokens
INFO 07-06 02:09:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 976.34x
INFO 07-06 02:09:16 [kv_cache_utils.py:634] GPU KV cache size: 1,999,280 tokens
INFO 07-06 02:09:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 976.21x
INFO 07-06 02:09:16 [kv_cache_utils.py:634] GPU KV cache size: 1,999,280 tokens
INFO 07-06 02:09:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 976.21x
INFO 07-06 02:09:16 [kv_cache_utils.py:634] GPU KV cache size: 2,000,560 tokens
INFO 07-06 02:09:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 976.84x
(VllmWorker rank=2 pid=3923543) INFO 07-06 02:09:51 [gpu_model_runner.py:1686] Graph capturing finished in 35 secs, took 3.00 GiB
(VllmWorker rank=3 pid=3923544) INFO 07-06 02:09:51 [gpu_model_runner.py:1686] Graph capturing finished in 35 secs, took 3.00 GiB
(VllmWorker rank=1 pid=3923542) INFO 07-06 02:09:51 [gpu_model_runner.py:1686] Graph capturing finished in 35 secs, took 3.00 GiB
(VllmWorker rank=0 pid=3923541) INFO 07-06 02:09:51 [gpu_model_runner.py:1686] Graph capturing finished in 35 secs, took 3.00 GiB
INFO 07-06 02:09:51 [core.py:159] init engine (profile, create kv cache, warmup model) took 89.50 seconds
INFO 07-06 02:09:52 [core_client.py:439] Core engine process 0 ready.
INFO 07-06 02:21:17 [importing.py:53] Triton module has been replaced with a placeholder.
INFO 07-06 02:21:17 [__init__.py:239] Automatically detected platform cuda.
|       Task       |Version|       Metric        |Value |   |Stderr|
|------------------|------:|---------------------|-----:|---|-----:|
|all               |       |sem                  |0.2600|±  |0.0241|
|                  |       |math_pass@1:1_samples|0.7685|±  |0.0450|
|mm\|arc_challenge\|0|      0|sem                  |0.3465|±  |0.0244|
|mm\|arc_easy\|0     |      0|sem                  |0.3411|±  |0.0154|
|mm\|commonsenseqa\|0|      0|sem                  |0.1375|±  |0.0193|
|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.8121|±  |0.0185|
|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7250|±  |0.0715|
|mm\|truthfulqa\|0   |      0|sem                  |0.2149|±  |0.0375|