1. Merge benchmark of Llama and Phi4

README.MD

@@ -0,0 +1,12 @@
+# Description
+`./test/0-1k`, `./merge_bench/` and `./merge_bench1/` have same eval data.  
+The data split includes math_tasks and mcq_tasks.  
+```
+math_tasks = ["mm|aime24|0", "mm|math_500|0", "mm|gsm8k|0"]
+mcq_tasks = ["mm|mmlu_pro|0", "mm|truthfulqa|0", "mm|commonsenseqa|0", "mm|arc_easy|0", "mm|arc_challenge|0", "mm|gpqa_diamond|0"]
+```
+
+And those only contain data samples whose generation length < 1k from respective reasoning model, e.g. DS-R1-Llama3 and Phi4-mini-reasoning. But currently all sample is from phi4-mini-reasoning
+
+The difference between `./merge_bench/` and `./merge_bench1/` is `./merge_bench1/` merged all layers of Phi4, while `./merge_bench/` missed `lm_head`.
+Note that the series of Llama in `./merge_bench/` is reaasonable, since those are merged by `mergekit`.

merge_bench/logs/llama_darelinear_1.log

@@ -0,0 +1,96 @@
+INFO 06-28 18:47:54 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 18:47:56 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 18:48:03 [config.py:717] This model supports multiple tasks: {'classify', 'score', 'reward', 'embed', 'generate'}. Defaulting to 'generate'.
+INFO 06-28 18:48:03 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 18:48:03 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 18:48:05 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 18:48:05 [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 06-28 18:48:05 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_06919893'), local_subscribe_addr='ipc:///tmp/d4f9c938-0474-4c85-8776-76fae2cfb900', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 18:48:05 [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 0x1515569ebc70>
+WARNING 06-28 18:48:05 [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 0x151554dbca90>
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:05 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_607805b6'), local_subscribe_addr='ipc:///tmp/446ff4e9-7682-40ee-a3fc-0784e08ffb01', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 18:48:05 [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 0x1515569ebd30>
+WARNING 06-28 18:48:05 [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 0x1515569eb9a0>
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:05 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_0e41e491'), local_subscribe_addr='ipc:///tmp/d8e738b3-c034-45e6-b1c5-2dfc295238ed', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:05 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_f2d47f78'), local_subscribe_addr='ipc:///tmp/fa2c9b8a-3b1c-4803-b18c-24205bbd5985', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:05 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_5a2a3f4f'), local_subscribe_addr='ipc:///tmp/730a59de-f5a9-4c2f-a5ea-44ed30623ac6', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:07 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:07 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:07 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:07 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:07 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:07 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:07 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:07 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3598716) WARNING 06-28 18:48:08 [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=3598717) WARNING 06-28 18:48:08 [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=3598715) WARNING 06-28 18:48:08 [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=3598714) WARNING 06-28 18:48:08 [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=3598714) INFO 06-28 18:48:08 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_1ad62c91'), local_subscribe_addr='ipc:///tmp/01e2f2dc-b1dd-4a71-b920-27675c6a453e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:08 [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=3598715) INFO 06-28 18:48:08 [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=3598717) INFO 06-28 18:48:08 [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=3598714) INFO 06-28 18:48:08 [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=3598716) INFO 06-28 18:48:08 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:08 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:08 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:08 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3598716) WARNING 06-28 18:48:08 [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=3598717) WARNING 06-28 18:48:08 [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=3598715) WARNING 06-28 18:48:08 [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=3598714) WARNING 06-28 18:48:08 [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=3598716) INFO 06-28 18:48:08 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:08 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:08 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:08 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:13 [loader.py:458] Loading weights took 4.51 seconds
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:13 [loader.py:458] Loading weights took 4.51 seconds
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:13 [loader.py:458] Loading weights took 4.52 seconds
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:13 [loader.py:458] Loading weights took 4.51 seconds
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:13 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 4.901312 seconds
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:13 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 4.908657 seconds
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:13 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 4.896791 seconds
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:13 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 4.908974 seconds
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:20 [backends.py:430] Dynamo bytecode transform time: 7.13 s
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:20 [backends.py:430] Dynamo bytecode transform time: 7.13 s
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:20 [backends.py:430] Dynamo bytecode transform time: 7.13 s
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:20 [backends.py:430] Dynamo bytecode transform time: 7.13 s
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:25 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.435 s
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:25 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.433 s
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:25 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.445 s
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:26 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.826 s
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:31 [monitor.py:33] torch.compile takes 7.13 s in total
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:31 [monitor.py:33] torch.compile takes 7.13 s in total
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:31 [monitor.py:33] torch.compile takes 7.13 s in total
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:31 [monitor.py:33] torch.compile takes 7.13 s in total
+INFO 06-28 18:48:33 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 18:48:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 18:48:33 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 18:48:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 18:48:33 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 18:48:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 18:48:33 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 18:48:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=0 pid=3598714) INFO 06-28 18:48:56 [gpu_model_runner.py:1686] Graph capturing finished in 23 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3598717) INFO 06-28 18:48:56 [gpu_model_runner.py:1686] Graph capturing finished in 23 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3598716) INFO 06-28 18:48:56 [gpu_model_runner.py:1686] Graph capturing finished in 23 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3598715) INFO 06-28 18:48:56 [gpu_model_runner.py:1686] Graph capturing finished in 23 secs, took 2.96 GiB
+INFO 06-28 18:48:56 [core.py:159] init engine (profile, create kv cache, warmup model) took 42.92 seconds
+INFO 06-28 18:48:56 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 19:01:31 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 19:01:31 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5201|±  |0.0281|
+|                  |       |math_pass@1:1_samples|0.7488|±  |0.0440|
+|mm\|arc_challenge\|0|      0|sem                  |0.6010|±  |0.0251|
+|mm\|arc_easy\|0     |      0|sem                  |0.6304|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4938|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7226|±  |0.0212|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7750|±  |0.0669|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3554|±  |0.0437|
+

merge_bench/logs/llama_darelinear_3.log

@@ -0,0 +1,96 @@
+INFO 06-28 19:01:30 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 19:01:32 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 19:01:39 [config.py:717] This model supports multiple tasks: {'score', 'generate', 'reward', 'embed', 'classify'}. Defaulting to 'generate'.
+INFO 06-28 19:01:39 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 19:01:39 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 19:01:40 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 19:01:40 [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 06-28 19:01:40 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_64032cd1'), local_subscribe_addr='ipc:///tmp/45893f5d-8e26-4aa9-9824-5b019d5989cf', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:01:40 [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 0x14abee988b20>
+WARNING 06-28 19:01:40 [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 0x14ac0444bdc0>
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:01:40 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_4ee5119d'), local_subscribe_addr='ipc:///tmp/6e2ed57a-78ff-4635-b216-0cc45fbb3fd6', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:01:40 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_8718327d'), local_subscribe_addr='ipc:///tmp/5be42818-9ba4-4977-b053-4709c5ac33b7', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:01:40 [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 0x14ac0444bac0>
+WARNING 06-28 19:01:41 [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 0x14ac0444bd00>
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:01:41 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_175f72fa'), local_subscribe_addr='ipc:///tmp/987bf39f-efca-4e1f-a76e-9fd79a0830a7', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:01:41 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_37f6e3ef'), local_subscribe_addr='ipc:///tmp/bc5d6131-6a85-46be-9c6b-7a2502f865ec', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:01:52 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:01:52 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:01:52 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:01:52 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:01:52 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:01:52 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:01:52 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:01:52 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3603787) WARNING 06-28 19:01:53 [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=3603788) WARNING 06-28 19:01:53 [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=3603786) WARNING 06-28 19:01:53 [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=3603785) WARNING 06-28 19:01:53 [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=3603785) INFO 06-28 19:01:53 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_01e6e72a'), local_subscribe_addr='ipc:///tmp/867fb721-a28c-49e2-a558-3ce978a6e3f7', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:01:53 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:01:53 [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=3603786) INFO 06-28 19:01:53 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:01:53 [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=3603787) INFO 06-28 19:01:53 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:01:53 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3603787) WARNING 06-28 19:01:53 [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=3603788) WARNING 06-28 19:01:53 [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=3603786) INFO 06-28 19:01:53 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:01:53 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3603786) WARNING 06-28 19:01:53 [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=3603785) WARNING 06-28 19:01:53 [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=3603787) INFO 06-28 19:01:53 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:01:53 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:01:53 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:01:53 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:01:54 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:01:54 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:01:54 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:01:54 [loader.py:458] Loading weights took 0.77 seconds
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:01:54 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.871916 seconds
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:01:54 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.871097 seconds
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:01:54 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.927502 seconds
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:01:54 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.986968 seconds
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:02:00 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:02:00 [backends.py:430] Dynamo bytecode transform time: 5.67 s
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:02:00 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:02:00 [backends.py:430] Dynamo bytecode transform time: 5.73 s
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:02:00 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:02:00 [backends.py:430] Dynamo bytecode transform time: 5.78 s
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:02:00 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:02:00 [backends.py:430] Dynamo bytecode transform time: 5.83 s
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:02:05 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.400 s
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:02:05 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.352 s
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:02:05 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.368 s
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:02:05 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.440 s
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:02:11 [monitor.py:33] torch.compile takes 5.78 s in total
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:02:11 [monitor.py:33] torch.compile takes 5.67 s in total
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:02:11 [monitor.py:33] torch.compile takes 5.73 s in total
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:02:11 [monitor.py:33] torch.compile takes 5.83 s in total
+INFO 06-28 19:02:12 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 19:02:12 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 19:02:12 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:02:12 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:02:12 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:02:12 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:02:12 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 19:02:12 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=1 pid=3603786) INFO 06-28 19:02:36 [gpu_model_runner.py:1686] Graph capturing finished in 24 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3603785) INFO 06-28 19:02:36 [gpu_model_runner.py:1686] Graph capturing finished in 24 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3603788) INFO 06-28 19:02:36 [gpu_model_runner.py:1686] Graph capturing finished in 24 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3603787) INFO 06-28 19:02:36 [gpu_model_runner.py:1686] Graph capturing finished in 24 secs, took 2.96 GiB
+INFO 06-28 19:02:36 [core.py:159] init engine (profile, create kv cache, warmup model) took 41.72 seconds
+INFO 06-28 19:02:37 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 19:15:24 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 19:15:24 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5004|±  |0.0274|
+|                  |       |math_pass@1:1_samples|0.8055|±  |0.0369|
+|mm\|arc_challenge\|0|      0|sem                  |0.6037|±  |0.0251|
+|mm\|arc_easy\|0     |      0|sem                  |0.6315|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4938|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7360|±  |0.0209|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8750|±  |0.0530|
+|mm\|truthfulqa\|0   |      0|sem                  |0.2727|±  |0.0407|
+

merge_bench/logs/llama_darelinear_5.log

@@ -0,0 +1,96 @@
+INFO 06-28 19:15:23 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 19:15:24 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 19:15:31 [config.py:717] This model supports multiple tasks: {'embed', 'reward', 'classify', 'score', 'generate'}. Defaulting to 'generate'.
+INFO 06-28 19:15:31 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 19:15:31 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 19:15:33 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 19:15:33 [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 06-28 19:15:33 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_10577a21'), local_subscribe_addr='ipc:///tmp/004d8b89-cc85-469e-a0a1-eba5bc07a552', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:15:33 [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 0x145d5d377dc0>
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:33 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_558e8955'), local_subscribe_addr='ipc:///tmp/f7c82864-a4a8-4c22-8842-dc5a67f67a87', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:15:33 [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 0x145d4f8dcb20>
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:33 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_27b3cabf'), local_subscribe_addr='ipc:///tmp/b6a9ea73-5188-4e6a-950a-8c81376233d5', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:15:33 [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 0x145d5d377ac0>
+WARNING 06-28 19:15:33 [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 0x145d5d377d00>
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:33 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_c356ca19'), local_subscribe_addr='ipc:///tmp/639cd23e-c9cf-4e49-b6d2-dac211cfea8e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:33 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_db80b59c'), local_subscribe_addr='ipc:///tmp/90eb8b9b-bfd8-42a1-ba6c-6cb01a8bd850', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:35 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:35 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:35 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:35 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:35 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:35 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:35 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:35 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3609849) WARNING 06-28 19:15:36 [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=2 pid=3609848) WARNING 06-28 19:15:36 [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=3609846) WARNING 06-28 19:15:36 [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=3609847) WARNING 06-28 19:15:36 [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=3609846) INFO 06-28 19:15:36 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_4a6645ac'), local_subscribe_addr='ipc:///tmp/96d20b0c-d027-4c6e-a850-071c00e81e80', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:36 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:36 [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=3609846) INFO 06-28 19:15:36 [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=3609848) INFO 06-28 19:15:36 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:36 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:36 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3609849) WARNING 06-28 19:15:36 [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=3609848) WARNING 06-28 19:15:36 [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=3609846) WARNING 06-28 19:15:36 [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=3609848) INFO 06-28 19:15:36 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:36 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:36 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:36 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3609847) WARNING 06-28 19:15:36 [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=3609847) INFO 06-28 19:15:36 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:36 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:37 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:37 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:37 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:37 [loader.py:458] Loading weights took 0.76 seconds
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:37 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.878368 seconds
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:37 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.874506 seconds
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:37 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.975043 seconds
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:37 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.928218 seconds
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:43 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:43 [backends.py:430] Dynamo bytecode transform time: 5.53 s
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:43 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:43 [backends.py:430] Dynamo bytecode transform time: 5.55 s
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:43 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:43 [backends.py:430] Dynamo bytecode transform time: 5.57 s
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:43 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:43 [backends.py:430] Dynamo bytecode transform time: 5.58 s
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:48 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.393 s
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:48 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.397 s
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:48 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.404 s
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:48 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.454 s
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:15:54 [monitor.py:33] torch.compile takes 5.57 s in total
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:15:54 [monitor.py:33] torch.compile takes 5.55 s in total
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:15:54 [monitor.py:33] torch.compile takes 5.53 s in total
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:15:54 [monitor.py:33] torch.compile takes 5.58 s in total
+INFO 06-28 19:15:55 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 19:15:55 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 19:15:55 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:15:55 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:15:55 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:15:55 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:15:55 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 19:15:55 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3609849) INFO 06-28 19:16:19 [gpu_model_runner.py:1686] Graph capturing finished in 23 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3609848) INFO 06-28 19:16:19 [gpu_model_runner.py:1686] Graph capturing finished in 23 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3609847) INFO 06-28 19:16:19 [gpu_model_runner.py:1686] Graph capturing finished in 23 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3609846) INFO 06-28 19:16:19 [gpu_model_runner.py:1686] Graph capturing finished in 23 secs, took 2.96 GiB
+INFO 06-28 19:16:19 [core.py:159] init engine (profile, create kv cache, warmup model) took 41.19 seconds
+INFO 06-28 19:16:19 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 19:28:57 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 19:28:57 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5105|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7999|±  |0.0371|
+|mm\|arc_challenge\|0|      0|sem                  |0.5853|±  |0.0253|
+|mm\|arc_easy\|0     |      0|sem                  |0.6336|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4844|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7248|±  |0.0211|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8750|±  |0.0530|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3388|±  |0.0432|
+

merge_bench/logs/llama_darelinear_7.log

@@ -0,0 +1,96 @@
+INFO 06-28 19:28:56 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 19:28:57 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 19:29:04 [config.py:717] This model supports multiple tasks: {'reward', 'score', 'classify', 'generate', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 19:29:04 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 19:29:04 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 19:29:06 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 19:29:06 [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 06-28 19:29:06 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_26ef43fb'), local_subscribe_addr='ipc:///tmp/1503da60-c19b-48f3-9809-e34d8853a309', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:29:06 [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 0x151ce2144b50>
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:06 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_11981bd0'), local_subscribe_addr='ipc:///tmp/9b1f1f4f-9671-425c-ae4d-18e28195a4bc', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:29:06 [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 0x151ce3b7fd30>
+WARNING 06-28 19:29:06 [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 0x151ce3b7fdf0>
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:06 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_2ab12ec2'), local_subscribe_addr='ipc:///tmp/9c3831dc-0da1-47c5-92b2-caa01026898b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:29:06 [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 0x151ce3b7faf0>
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:06 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_26bbf412'), local_subscribe_addr='ipc:///tmp/a3d59c06-7d41-4866-ad94-b254fd1dee6e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:06 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_2e0f0ae9'), local_subscribe_addr='ipc:///tmp/1cbf6108-c340-4068-a696-3ce96130e9fb', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:13 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:13 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:13 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:13 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:13 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:13 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:13 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:13 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3613747) WARNING 06-28 19:29:13 [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=2 pid=3613746) WARNING 06-28 19:29:13 [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=3613742) WARNING 06-28 19:29:13 [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=3613743) WARNING 06-28 19:29:13 [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=3613742) INFO 06-28 19:29:13 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_1c6a8c3a'), local_subscribe_addr='ipc:///tmp/4e1b6783-859b-428c-a617-d9ff90c87a4f', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:13 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:13 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3613742) WARNING 06-28 19:29:13 [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=3613747) INFO 06-28 19:29:13 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:13 [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=3613743) INFO 06-28 19:29:13 [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=3613747) INFO 06-28 19:29:13 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:13 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3613747) WARNING 06-28 19:29:13 [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=3613746) WARNING 06-28 19:29:13 [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=3613743) INFO 06-28 19:29:13 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3613743) WARNING 06-28 19:29:13 [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=3613742) INFO 06-28 19:29:13 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:13 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:13 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:13 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:14 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:14 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:14 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:14 [loader.py:458] Loading weights took 0.77 seconds
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:14 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.905328 seconds
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:14 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.900747 seconds
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:14 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.926662 seconds
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:15 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 1.000593 seconds
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:20 [backends.py:430] Dynamo bytecode transform time: 5.62 s
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:20 [backends.py:430] Dynamo bytecode transform time: 5.68 s
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:21 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:21 [backends.py:430] Dynamo bytecode transform time: 5.85 s
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:21 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:21 [backends.py:430] Dynamo bytecode transform time: 5.92 s
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:25 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.372 s
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:25 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.360 s
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:26 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.434 s
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:26 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.386 s
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:31 [monitor.py:33] torch.compile takes 5.68 s in total
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:31 [monitor.py:33] torch.compile takes 5.92 s in total
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:31 [monitor.py:33] torch.compile takes 5.85 s in total
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:31 [monitor.py:33] torch.compile takes 5.62 s in total
+INFO 06-28 19:29:33 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 19:29:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 19:29:33 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:29:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:29:33 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:29:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:29:33 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 19:29:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3613747) INFO 06-28 19:29:58 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3613743) INFO 06-28 19:29:58 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3613746) INFO 06-28 19:29:58 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3613742) INFO 06-28 19:29:58 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 19:29:58 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.73 seconds
+INFO 06-28 19:29:59 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 19:42:40 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 19:42:40 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5357|±  |0.0281|
+|                  |       |math_pass@1:1_samples|0.7499|±  |0.0440|
+|mm\|arc_challenge\|0|      0|sem                  |0.5984|±  |0.0251|
+|mm\|arc_easy\|0     |      0|sem                  |0.6452|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5437|±  |0.0279|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7248|±  |0.0211|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7750|±  |0.0669|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3554|±  |0.0437|
+

merge_bench/logs/llama_darelinear_9.log

@@ -0,0 +1,96 @@
+INFO 06-28 19:42:39 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 19:42:41 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 19:42:48 [config.py:717] This model supports multiple tasks: {'score', 'embed', 'classify', 'generate', 'reward'}. Defaulting to 'generate'.
+INFO 06-28 19:42:48 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 19:42:48 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 19:42:50 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 19:42:50 [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 06-28 19:42:50 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_c90e6d0c'), local_subscribe_addr='ipc:///tmp/966a24d0-22af-4b35-b61c-287d01dabdde', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:42:50 [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 0x14946571fd90>
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:42:50 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_6dc61b5f'), local_subscribe_addr='ipc:///tmp/e48c70ef-ba23-4cb5-91df-362ff41efa0d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:42:50 [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 0x14942fd6caf0>
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:42:50 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_1646518f'), local_subscribe_addr='ipc:///tmp/b6ee7bc0-17bb-4f38-b44b-ec7473a9d4bb', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:42:50 [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 0x14946571fcd0>
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:42:50 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_5fc7d511'), local_subscribe_addr='ipc:///tmp/dc1587a5-cd34-4d51-9ef8-72e1e473fa0d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:42:50 [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 0x14946571fa90>
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:42:50 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_cceaf416'), local_subscribe_addr='ipc:///tmp/c17485cf-1cc7-433c-9b80-d2e33392d8cd', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:42:56 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:42:56 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:42:56 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:42:56 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:42:56 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:42:56 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:42:56 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:42:56 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3616477) WARNING 06-28 19:42:57 [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=3616478) WARNING 06-28 19:42:57 [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=3616476) WARNING 06-28 19:42:57 [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=3616475) WARNING 06-28 19:42:57 [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=3616475) INFO 06-28 19:42:57 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_7d14ff7a'), local_subscribe_addr='ipc:///tmp/54dc6b12-0377-4d94-b6c9-a54dfc6fe0b4', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:42:57 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:42:57 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:42:57 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:42:57 [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=3616477) INFO 06-28 19:42:57 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3616477) WARNING 06-28 19:42:57 [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=3616478) INFO 06-28 19:42:57 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:42:57 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3616478) WARNING 06-28 19:42:57 [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=3616476) WARNING 06-28 19:42:57 [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=3616475) INFO 06-28 19:42:57 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3616475) WARNING 06-28 19:42:57 [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=3616477) INFO 06-28 19:42:57 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:42:57 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:42:57 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:42:57 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:42:58 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:42:58 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:42:58 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:42:58 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:42:58 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.868548 seconds
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:42:58 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.867938 seconds
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:42:58 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.942615 seconds
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:42:58 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.920874 seconds
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:43:04 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:43:04 [backends.py:430] Dynamo bytecode transform time: 5.57 s
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:43:04 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:43:04 [backends.py:430] Dynamo bytecode transform time: 5.75 s
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:43:04 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:43:04 [backends.py:430] Dynamo bytecode transform time: 5.90 s
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:43:04 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:43:04 [backends.py:430] Dynamo bytecode transform time: 6.00 s
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:43:09 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.353 s
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:43:09 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.393 s
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:43:09 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.390 s
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:43:09 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.436 s
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:43:15 [monitor.py:33] torch.compile takes 5.75 s in total
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:43:15 [monitor.py:33] torch.compile takes 5.90 s in total
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:43:15 [monitor.py:33] torch.compile takes 6.00 s in total
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:43:15 [monitor.py:33] torch.compile takes 5.57 s in total
+INFO 06-28 19:43:16 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 19:43:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 19:43:16 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:43:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:43:16 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:43:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:43:16 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 19:43:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3616478) INFO 06-28 19:43:42 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3616477) INFO 06-28 19:43:42 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3616475) INFO 06-28 19:43:42 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3616476) INFO 06-28 19:43:42 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 19:43:42 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.87 seconds
+INFO 06-28 19:43:43 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 19:56:27 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 19:56:27 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5197|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7193|±  |0.0465|
+|mm\|arc_challenge\|0|      0|sem                  |0.5906|±  |0.0252|
+|mm\|arc_easy\|0     |      0|sem                  |0.6367|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5125|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7136|±  |0.0214|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7250|±  |0.0715|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3388|±  |0.0432|
+

merge_bench/logs/llama_linear_1.log

@@ -0,0 +1,96 @@
+INFO 06-28 19:56:26 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 19:56:27 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 19:56:34 [config.py:717] This model supports multiple tasks: {'score', 'reward', 'classify', 'generate', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 19:56:34 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 19:56:34 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 19:56:36 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 19:56:36 [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 06-28 19:56:36 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_e47c5064'), local_subscribe_addr='ipc:///tmp/e6ad432d-f508-4f32-bd1f-0d7c0725974d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:56:36 [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 0x14c86e0c8a90>
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:36 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e6b0e3d6'), local_subscribe_addr='ipc:///tmp/f98aba7d-bfc4-4b64-b79d-83126ab2f88c', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:56:36 [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 0x14c86fa33c70>
+WARNING 06-28 19:56:36 [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 0x14c86fa33d30>
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:36 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_c6a69884'), local_subscribe_addr='ipc:///tmp/307da573-bf26-4a86-b0d1-2e4f53d94f88', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 19:56:36 [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 0x14c86fa339a0>
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:36 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_28462e05'), local_subscribe_addr='ipc:///tmp/8a976985-b5bc-4a59-a4a6-9019466bb558', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:36 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_2d722b0e'), local_subscribe_addr='ipc:///tmp/3a2db13c-7d51-4363-8f9e-c42a98ab0208', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:39 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:39 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:39 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:39 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:39 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:39 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:39 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:39 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3618691) WARNING 06-28 19:56:40 [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=3618692) WARNING 06-28 19:56:40 [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=3618690) WARNING 06-28 19:56:40 [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=3618689) WARNING 06-28 19:56:40 [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=3618689) INFO 06-28 19:56:40 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_6c78750e'), local_subscribe_addr='ipc:///tmp/55c8e627-7a54-4c21-9f63-0cfcbd1725bc', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:40 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:40 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3618690) WARNING 06-28 19:56:40 [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=3618692) INFO 06-28 19:56:40 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:40 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:40 [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=3618691) INFO 06-28 19:56:40 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:40 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3618691) WARNING 06-28 19:56:40 [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=3618692) WARNING 06-28 19:56:40 [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=3618689) INFO 06-28 19:56:40 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3618689) WARNING 06-28 19:56:40 [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=3618690) INFO 06-28 19:56:40 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:40 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:40 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:40 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:40 [loader.py:458] Loading weights took 0.66 seconds
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:40 [loader.py:458] Loading weights took 0.65 seconds
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:40 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:41 [loader.py:458] Loading weights took 0.75 seconds
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:41 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.840570 seconds
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:41 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.842686 seconds
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:41 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.982921 seconds
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:41 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.901269 seconds
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:47 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:47 [backends.py:430] Dynamo bytecode transform time: 5.76 s
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:47 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:47 [backends.py:430] Dynamo bytecode transform time: 5.86 s
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:47 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:47 [backends.py:430] Dynamo bytecode transform time: 5.94 s
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:47 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:47 [backends.py:430] Dynamo bytecode transform time: 5.96 s
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:52 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.336 s
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:52 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.390 s
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:52 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.406 s
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:52 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.512 s
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:56:58 [monitor.py:33] torch.compile takes 5.76 s in total
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:56:58 [monitor.py:33] torch.compile takes 5.96 s in total
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:56:58 [monitor.py:33] torch.compile takes 5.86 s in total
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:56:58 [monitor.py:33] torch.compile takes 5.94 s in total
+INFO 06-28 19:56:59 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 19:56:59 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 19:56:59 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:56:59 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:56:59 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 19:56:59 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 19:56:59 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 19:56:59 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=1 pid=3618690) INFO 06-28 19:57:25 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3618691) INFO 06-28 19:57:25 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3618689) INFO 06-28 19:57:25 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3618692) INFO 06-28 19:57:25 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 19:57:25 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.39 seconds
+INFO 06-28 19:57:26 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 20:10:01 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 20:10:01 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5198|±  |0.0282|
+|                  |       |math_pass@1:1_samples|0.7070|±  |0.0467|
+|mm\|arc_challenge\|0|      0|sem                  |0.6037|±  |0.0251|
+|mm\|arc_easy\|0     |      0|sem                  |0.6336|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4781|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.6890|±  |0.0219|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7250|±  |0.0715|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3636|±  |0.0439|
+

merge_bench/logs/llama_linear_3.log

@@ -0,0 +1,96 @@
+INFO 06-28 20:10:00 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 20:10:02 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 20:10:09 [config.py:717] This model supports multiple tasks: {'generate', 'reward', 'score', 'embed', 'classify'}. Defaulting to 'generate'.
+INFO 06-28 20:10:09 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 20:10:09 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 20:10:11 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 20:10:11 [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 06-28 20:10:11 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_6257e474'), local_subscribe_addr='ipc:///tmp/d597435c-2e4d-456f-89db-fcbedcefafc3', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:10:11 [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 0x14db4ad97dc0>
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:11 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_6a207ca3'), local_subscribe_addr='ipc:///tmp/46e20e0c-64bd-473c-b401-bac5e8cadcaa', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:10:11 [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 0x14db49360b20>
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:11 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_0799eacd'), local_subscribe_addr='ipc:///tmp/4c8f80b6-f69b-4626-91a1-b0f3e0c81543', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:10:11 [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 0x14db4ad97d00>
+WARNING 06-28 20:10:11 [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 0x14db4ad97ac0>
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:11 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_5c611e31'), local_subscribe_addr='ipc:///tmp/0996616a-5d2f-4b7b-aaae-5d95c304730e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:11 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_3d01097d'), local_subscribe_addr='ipc:///tmp/77a0146d-a0a9-4b05-86ed-60c0860a40fe', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:13 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:13 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:13 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:13 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:13 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:13 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:13 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:13 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3620671) WARNING 06-28 20:10:14 [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=3620672) WARNING 06-28 20:10:14 [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=3620670) WARNING 06-28 20:10:14 [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=3620669) WARNING 06-28 20:10:14 [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=3620669) INFO 06-28 20:10:14 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_5c739929'), local_subscribe_addr='ipc:///tmp/27185b44-3b72-4a03-bb20-e0fcf7dc7d56', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:14 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:14 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:14 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:14 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:14 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:14 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3620672) WARNING 06-28 20:10:14 [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=3620671) WARNING 06-28 20:10:14 [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=3620670) INFO 06-28 20:10:14 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:14 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3620670) WARNING 06-28 20:10:14 [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=3620669) WARNING 06-28 20:10:14 [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=3620671) INFO 06-28 20:10:14 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:14 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:14 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:14 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:15 [loader.py:458] Loading weights took 0.66 seconds
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:15 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:15 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:15 [loader.py:458] Loading weights took 0.74 seconds
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:15 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.846524 seconds
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:15 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.885501 seconds
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:15 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.946978 seconds
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:15 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.930972 seconds
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:21 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:21 [backends.py:430] Dynamo bytecode transform time: 5.60 s
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:21 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:21 [backends.py:430] Dynamo bytecode transform time: 5.71 s
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:21 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:21 [backends.py:430] Dynamo bytecode transform time: 5.73 s
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:21 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:21 [backends.py:430] Dynamo bytecode transform time: 5.75 s
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:26 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.402 s
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:26 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.438 s
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:26 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.448 s
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:26 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.464 s
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:32 [monitor.py:33] torch.compile takes 5.71 s in total
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:32 [monitor.py:33] torch.compile takes 5.73 s in total
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:32 [monitor.py:33] torch.compile takes 5.60 s in total
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:32 [monitor.py:33] torch.compile takes 5.75 s in total
+INFO 06-28 20:10:33 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 20:10:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 20:10:33 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 20:10:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 20:10:33 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 20:10:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 20:10:33 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 20:10:33 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3620672) INFO 06-28 20:10:56 [gpu_model_runner.py:1686] Graph capturing finished in 24 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3620671) INFO 06-28 20:10:56 [gpu_model_runner.py:1686] Graph capturing finished in 24 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3620669) INFO 06-28 20:10:56 [gpu_model_runner.py:1686] Graph capturing finished in 24 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3620670) INFO 06-28 20:10:56 [gpu_model_runner.py:1686] Graph capturing finished in 24 secs, took 2.96 GiB
+INFO 06-28 20:10:56 [core.py:159] init engine (profile, create kv cache, warmup model) took 41.41 seconds
+INFO 06-28 20:10:57 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 20:23:36 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 20:23:36 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.4992|±  |0.0276|
+|                  |       |math_pass@1:1_samples|0.8236|±  |0.0343|
+|mm\|arc_challenge\|0|      0|sem                  |0.5879|±  |0.0252|
+|mm\|arc_easy\|0     |      0|sem                  |0.6135|±  |0.0158|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5062|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7472|±  |0.0206|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.9000|±  |0.0480|
+|mm\|truthfulqa\|0   |      0|sem                  |0.2893|±  |0.0414|
+

merge_bench/logs/llama_linear_5.log

@@ -0,0 +1,96 @@
+INFO 06-28 20:23:35 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 20:23:36 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 20:23:43 [config.py:717] This model supports multiple tasks: {'score', 'embed', 'reward', 'generate', 'classify'}. Defaulting to 'generate'.
+INFO 06-28 20:23:43 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 20:23:43 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 20:23:45 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 20:23:45 [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 06-28 20:23:45 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_b6685efa'), local_subscribe_addr='ipc:///tmp/299a1d34-2ed2-4341-988b-660b2e51724a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:23:45 [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 0x148dc2563d60>
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:45 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_919d78a9'), local_subscribe_addr='ipc:///tmp/e31e068f-fcfb-41a8-94de-c14d5b28536d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:23:45 [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 0x148dc0c08ac0>
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:45 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_6406296c'), local_subscribe_addr='ipc:///tmp/036b4ed3-da30-4873-bf68-9a5a4cdd976b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:23:45 [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 0x148dc2563ca0>
+WARNING 06-28 20:23:45 [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 0x148dc25639d0>
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:45 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_c676c60a'), local_subscribe_addr='ipc:///tmp/1c4487f2-e80f-4d6a-a055-378524306660', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:23:45 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_6cc8c80e'), local_subscribe_addr='ipc:///tmp/27e1dd2c-d6e2-4155-87cb-32e683bab21b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:47 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:47 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:47 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:47 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:47 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:23:47 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:47 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:23:47 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3622892) WARNING 06-28 20:23:48 [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=2 pid=3622891) WARNING 06-28 20:23:48 [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=3622889) WARNING 06-28 20:23:48 [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=3622890) WARNING 06-28 20:23:48 [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=3622889) INFO 06-28 20:23:48 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_e5cd3014'), local_subscribe_addr='ipc:///tmp/28a2752c-3acb-4fc3-9bbf-2d7f0e4908ab', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:23:48 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:48 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:48 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:48 [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=3622892) INFO 06-28 20:23:48 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:48 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3622892) WARNING 06-28 20:23:48 [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=3622891) WARNING 06-28 20:23:48 [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=3622889) INFO 06-28 20:23:48 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:48 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3622889) WARNING 06-28 20:23:48 [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=3622890) WARNING 06-28 20:23:48 [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=3622892) INFO 06-28 20:23:48 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:48 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:48 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:48 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:49 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:23:49 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:49 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:49 [loader.py:458] Loading weights took 0.74 seconds
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:23:49 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.870155 seconds
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:49 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.867179 seconds
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:49 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.924282 seconds
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:49 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.974858 seconds
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:23:55 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:23:55 [backends.py:430] Dynamo bytecode transform time: 5.56 s
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:55 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:23:55 [backends.py:430] Dynamo bytecode transform time: 5.67 s
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:55 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:23:55 [backends.py:430] Dynamo bytecode transform time: 5.76 s
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:55 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:23:55 [backends.py:430] Dynamo bytecode transform time: 5.82 s
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:24:00 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.652 s
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:24:00 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.663 s
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:24:00 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.576 s
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:24:00 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.604 s
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:24:06 [monitor.py:33] torch.compile takes 5.82 s in total
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:24:06 [monitor.py:33] torch.compile takes 5.76 s in total
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:24:06 [monitor.py:33] torch.compile takes 5.67 s in total
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:24:06 [monitor.py:33] torch.compile takes 5.56 s in total
+INFO 06-28 20:24:07 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 20:24:07 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 20:24:07 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 20:24:07 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 20:24:07 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 20:24:07 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 20:24:07 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 20:24:07 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3622892) INFO 06-28 20:24:34 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3622891) INFO 06-28 20:24:34 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3622889) INFO 06-28 20:24:34 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3622890) INFO 06-28 20:24:34 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+INFO 06-28 20:24:34 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.75 seconds
+INFO 06-28 20:24:34 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 20:37:14 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 20:37:14 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5256|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7443|±  |0.0441|
+|mm\|arc_challenge\|0|      0|sem                  |0.6115|±  |0.0250|
+|mm\|arc_easy\|0     |      0|sem                  |0.6251|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5188|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7136|±  |0.0214|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7750|±  |0.0669|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3471|±  |0.0435|
+

merge_bench/logs/llama_linear_7.log

@@ -0,0 +1,96 @@
+INFO 06-28 20:37:13 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 20:37:14 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 20:37:21 [config.py:717] This model supports multiple tasks: {'generate', 'reward', 'embed', 'score', 'classify'}. Defaulting to 'generate'.
+INFO 06-28 20:37:21 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 20:37:21 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 20:37:23 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 20:37:23 [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 06-28 20:37:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_93b18dc7'), local_subscribe_addr='ipc:///tmp/7537e22e-27ae-4eed-8ba9-8f8926cf4814', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:37:23 [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 0x152d55a33d60>
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:37:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_d66e4aef'), local_subscribe_addr='ipc:///tmp/5956dee2-1734-4465-acfb-51251cb047b9', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:37:23 [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 0x152d4ffc0ac0>
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_56d69c19'), local_subscribe_addr='ipc:///tmp/eee01736-5f87-422d-81df-6f613a8a4f39', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:37:23 [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 0x152d55a33ca0>
+WARNING 06-28 20:37:23 [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 0x152d55a339d0>
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_b35e06ee'), local_subscribe_addr='ipc:///tmp/70baefd3-b12c-4835-9947-e5bce282520d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:37:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_9b3261e8'), local_subscribe_addr='ipc:///tmp/5affabae-a840-41f8-95e7-09f8fb3b8037', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:50 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:37:50 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:50 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:37:50 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:50 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:50 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:37:50 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:37:50 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3625205) WARNING 06-28 20:37:51 [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=3625203) WARNING 06-28 20:37:51 [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=2 pid=3625204) WARNING 06-28 20:37:51 [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=3625202) WARNING 06-28 20:37:51 [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=3625202) INFO 06-28 20:37:51 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_304efd39'), local_subscribe_addr='ipc:///tmp/4e473a91-f106-49ce-a859-61afa613a874', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:51 [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=3625203) INFO 06-28 20:37:51 [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=3625205) INFO 06-28 20:37:51 [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=3625202) INFO 06-28 20:37:51 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:37:51 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:51 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3625205) WARNING 06-28 20:37:51 [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=3625203) INFO 06-28 20:37:51 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:51 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3625204) WARNING 06-28 20:37:51 [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=3625203) WARNING 06-28 20:37:51 [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=3625202) WARNING 06-28 20:37:51 [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=3625205) INFO 06-28 20:37:51 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:51 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:37:51 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:51 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:37:52 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:52 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:37:52 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:52 [loader.py:458] Loading weights took 0.74 seconds
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:37:52 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.864541 seconds
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:52 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.913058 seconds
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:52 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.977701 seconds
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:37:52 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.913645 seconds
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:37:58 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:37:58 [backends.py:430] Dynamo bytecode transform time: 5.66 s
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:37:58 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:37:58 [backends.py:430] Dynamo bytecode transform time: 5.72 s
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:58 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:37:58 [backends.py:430] Dynamo bytecode transform time: 5.76 s
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:58 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:37:58 [backends.py:430] Dynamo bytecode transform time: 5.78 s
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:38:03 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.423 s
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:38:03 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.432 s
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:38:03 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.545 s
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:38:03 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.650 s
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:38:09 [monitor.py:33] torch.compile takes 5.76 s in total
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:38:09 [monitor.py:33] torch.compile takes 5.66 s in total
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:38:09 [monitor.py:33] torch.compile takes 5.72 s in total
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:38:09 [monitor.py:33] torch.compile takes 5.78 s in total
+INFO 06-28 20:38:10 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 20:38:10 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 20:38:10 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 20:38:10 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 20:38:10 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 20:38:10 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 20:38:10 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 20:38:10 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=0 pid=3625202) INFO 06-28 20:38:37 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3625205) INFO 06-28 20:38:37 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3625203) INFO 06-28 20:38:37 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3625204) INFO 06-28 20:38:37 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+INFO 06-28 20:38:37 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.85 seconds
+INFO 06-28 20:38:37 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 20:51:18 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 20:51:18 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5158|±  |0.0282|
+|                  |       |math_pass@1:1_samples|0.6988|±  |0.0481|
+|mm\|arc_challenge\|0|      0|sem                  |0.5669|±  |0.0254|
+|mm\|arc_easy\|0     |      0|sem                  |0.6209|±  |0.0158|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5281|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7226|±  |0.0212|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.6750|±  |0.0750|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3471|±  |0.0435|
+

merge_bench/logs/llama_linear_9.log

@@ -0,0 +1,96 @@
+INFO 06-28 20:51:17 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 20:51:18 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 20:51:25 [config.py:717] This model supports multiple tasks: {'reward', 'score', 'embed', 'classify', 'generate'}. Defaulting to 'generate'.
+INFO 06-28 20:51:25 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 20:51:25 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 20:51:27 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 20:51:27 [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 06-28 20:51:27 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_bb84ef60'), local_subscribe_addr='ipc:///tmp/9f532b5b-5fab-4bd7-a386-12ac4cb074df', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:51:27 [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 0x1529f08dbd60>
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:27 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e583dcba'), local_subscribe_addr='ipc:///tmp/bb8329a0-d171-447f-bbb2-4bfaa38bf85a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:51:27 [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 0x1529e2f2cac0>
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:27 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_2e196d70'), local_subscribe_addr='ipc:///tmp/6c61b285-be86-41cb-9cb1-705e39daaba9', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 20:51:27 [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 0x1529f08dbca0>
+WARNING 06-28 20:51:27 [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 0x1529f08db9d0>
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:27 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_4864d8cf'), local_subscribe_addr='ipc:///tmp/79d64628-22b1-4ea9-a2ec-ee94d71df77d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:27 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a9de9a12'), local_subscribe_addr='ipc:///tmp/943c88f2-b4dd-40ee-a3e3-13e59c2e57fc', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:30 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:30 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:30 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:30 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:30 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:30 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:30 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:30 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3627171) WARNING 06-28 20:51:30 [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=3627172) WARNING 06-28 20:51:30 [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=3627170) WARNING 06-28 20:51:30 [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=3627169) WARNING 06-28 20:51:30 [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=3627169) INFO 06-28 20:51:30 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_af3ce749'), local_subscribe_addr='ipc:///tmp/f1cecbba-d83a-49da-abec-058737ee8492', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:30 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:30 [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=3627170) INFO 06-28 20:51:30 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:30 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:30 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:30 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3627172) WARNING 06-28 20:51:30 [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=3627171) WARNING 06-28 20:51:30 [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=3627169) INFO 06-28 20:51:30 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:30 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3627169) WARNING 06-28 20:51:30 [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=3627170) WARNING 06-28 20:51:30 [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=3627172) INFO 06-28 20:51:30 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:30 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:30 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:30 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:31 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:31 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:31 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:31 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:31 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.869730 seconds
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:31 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.849842 seconds
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:32 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.906712 seconds
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:32 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.951992 seconds
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:37 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:37 [backends.py:430] Dynamo bytecode transform time: 5.56 s
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:37 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:37 [backends.py:430] Dynamo bytecode transform time: 5.70 s
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:37 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:37 [backends.py:430] Dynamo bytecode transform time: 5.72 s
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:38 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:38 [backends.py:430] Dynamo bytecode transform time: 5.98 s
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:42 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.382 s
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:43 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.404 s
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:43 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.443 s
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:43 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.471 s
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:51:49 [monitor.py:33] torch.compile takes 5.56 s in total
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:51:49 [monitor.py:33] torch.compile takes 5.70 s in total
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:51:49 [monitor.py:33] torch.compile takes 5.72 s in total
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:51:49 [monitor.py:33] torch.compile takes 5.98 s in total
+INFO 06-28 20:51:50 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 20:51:50 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 20:51:50 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 20:51:50 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 20:51:50 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 20:51:50 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 20:51:50 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 20:51:50 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=1 pid=3627170) INFO 06-28 20:52:16 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3627171) INFO 06-28 20:52:16 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3627169) INFO 06-28 20:52:16 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3627172) INFO 06-28 20:52:16 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 20:52:16 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.15 seconds
+INFO 06-28 20:52:16 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 21:04:54 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 21:04:54 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5069|±  |0.0276|
+|                  |       |math_pass@1:1_samples|0.7874|±  |0.0392|
+|mm\|arc_challenge\|0|      0|sem                  |0.6142|±  |0.0250|
+|mm\|arc_easy\|0     |      0|sem                  |0.6283|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4875|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7248|±  |0.0211|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8500|±  |0.0572|
+|mm\|truthfulqa\|0   |      0|sem                  |0.2975|±  |0.0417|
+

merge_bench/logs/llama_ties_1.log

@@ -0,0 +1,96 @@
+INFO 06-28 21:04:53 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 21:04:55 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 21:05:02 [config.py:717] This model supports multiple tasks: {'generate', 'score', 'reward', 'classify', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 21:05:02 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 21:05:02 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 21:05:03 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 21:05:03 [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 06-28 21:05:03 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_49535664'), local_subscribe_addr='ipc:///tmp/396eb2be-5260-482c-8468-18f9912dab8c', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:05:04 [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 0x14c8da423c70>
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:04 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_32f083a9'), local_subscribe_addr='ipc:///tmp/4580c9ae-5fa5-4852-a517-8d5b635bf792', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:05:04 [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 0x14c8d8ab49d0>
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:04 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_df472ef4'), local_subscribe_addr='ipc:///tmp/3bfded55-edfb-4719-9266-d163d3b02918', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:05:04 [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 0x14c8da423bb0>
+WARNING 06-28 21:05:04 [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 0x14c8da4238e0>
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:04 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_478853c5'), local_subscribe_addr='ipc:///tmp/d8b38252-237a-42ce-937a-196da3b5790e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:04 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_ae4c4856'), local_subscribe_addr='ipc:///tmp/6ba7770b-4581-4bbb-806f-3a89604afdf9', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:06 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:06 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:06 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:06 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:06 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:06 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:06 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:06 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3629140) WARNING 06-28 21:05:07 [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=2 pid=3629139) WARNING 06-28 21:05:07 [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=3629138) WARNING 06-28 21:05:07 [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=3629137) WARNING 06-28 21:05:07 [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=3629137) INFO 06-28 21:05:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_af396e84'), local_subscribe_addr='ipc:///tmp/5bd0bf44-5b41-424c-9852-44131328ab72', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:07 [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=3629140) INFO 06-28 21:05:07 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:07 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:07 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:07 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:07 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:07 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3629140) WARNING 06-28 21:05:07 [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=3629139) WARNING 06-28 21:05:07 [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=3629138) WARNING 06-28 21:05:07 [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=3629137) INFO 06-28 21:05:07 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3629137) WARNING 06-28 21:05:07 [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=3629138) INFO 06-28 21:05:07 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:07 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:07 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:07 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:07 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:07 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:08 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:08 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:08 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.863475 seconds
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:08 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.861492 seconds
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:08 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.951411 seconds
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:08 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.920031 seconds
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:13 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:13 [backends.py:430] Dynamo bytecode transform time: 5.54 s
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:14 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:14 [backends.py:430] Dynamo bytecode transform time: 5.68 s
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:14 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:14 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:14 [backends.py:430] Dynamo bytecode transform time: 5.79 s
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:14 [backends.py:430] Dynamo bytecode transform time: 5.79 s
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:18 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.368 s
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:19 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.363 s
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:19 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.341 s
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:19 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.401 s
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:25 [monitor.py:33] torch.compile takes 5.54 s in total
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:25 [monitor.py:33] torch.compile takes 5.68 s in total
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:25 [monitor.py:33] torch.compile takes 5.79 s in total
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:25 [monitor.py:33] torch.compile takes 5.79 s in total
+INFO 06-28 21:05:26 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 21:05:26 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 21:05:26 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 21:05:26 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 21:05:26 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 21:05:26 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 21:05:26 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 21:05:26 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=1 pid=3629138) INFO 06-28 21:05:56 [gpu_model_runner.py:1686] Graph capturing finished in 30 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3629140) INFO 06-28 21:05:56 [gpu_model_runner.py:1686] Graph capturing finished in 30 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3629137) INFO 06-28 21:05:56 [gpu_model_runner.py:1686] Graph capturing finished in 30 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3629139) INFO 06-28 21:05:56 [gpu_model_runner.py:1686] Graph capturing finished in 30 secs, took 2.96 GiB
+INFO 06-28 21:05:56 [core.py:159] init engine (profile, create kv cache, warmup model) took 48.16 seconds
+INFO 06-28 21:05:56 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 21:18:38 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 21:18:38 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5101|±  |0.0276|
+|                  |       |math_pass@1:1_samples|0.8100|±  |0.0368|
+|mm\|arc_challenge\|0|      0|sem                  |0.6063|±  |0.0251|
+|mm\|arc_easy\|0     |      0|sem                  |0.6336|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5031|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7450|±  |0.0206|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8750|±  |0.0530|
+|mm\|truthfulqa\|0   |      0|sem                  |0.2975|±  |0.0417|
+

merge_bench/logs/llama_ties_3.log

@@ -0,0 +1,96 @@
+INFO 06-28 21:18:37 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 21:18:38 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 21:18:45 [config.py:717] This model supports multiple tasks: {'reward', 'embed', 'generate', 'classify', 'score'}. Defaulting to 'generate'.
+INFO 06-28 21:18:45 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 21:18:45 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 21:18:47 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 21:18:47 [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 06-28 21:18:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_397c2c8c'), local_subscribe_addr='ipc:///tmp/9b35e708-cffe-4aa9-be48-15fbf9844d84', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:18:47 [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 0x154f3f43ca90>
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:18:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_3d5102a4'), local_subscribe_addr='ipc:///tmp/3c5e6239-1d7c-4267-9676-9ac237de72d8', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:18:47 [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 0x154f44dffc70>
+WARNING 06-28 21:18:47 [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 0x154f44dffd30>
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:18:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e7b03a50'), local_subscribe_addr='ipc:///tmp/f48b8442-bcbe-4798-8547-c0a4ea2f1765', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:18:47 [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 0x154f44dff9a0>
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:18:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_29183d5c'), local_subscribe_addr='ipc:///tmp/a52d762d-6263-45f1-9161-419a79d4cf72', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:18:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_570fd23d'), local_subscribe_addr='ipc:///tmp/c0c9db08-7e7c-48aa-86fa-0a111f16fdfa', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:18:59 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:18:59 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:18:59 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:18:59 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:18:59 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:18:59 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:18:59 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:18:59 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3631106) WARNING 06-28 21:18:59 [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=2 pid=3631105) WARNING 06-28 21:18:59 [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=3631103) WARNING 06-28 21:18:59 [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=3631104) WARNING 06-28 21:18:59 [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=3631103) INFO 06-28 21:18:59 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_0515b019'), local_subscribe_addr='ipc:///tmp/b7ff17de-8d57-419c-97f6-014d6d347e56', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:18:59 [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=3631105) INFO 06-28 21:18:59 [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=3631104) INFO 06-28 21:18:59 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:18:59 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:18:59 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3631105) WARNING 06-28 21:18:59 [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=3631103) INFO 06-28 21:18:59 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:18:59 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:18:59 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3631106) WARNING 06-28 21:18:59 [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=3631103) WARNING 06-28 21:18:59 [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=3631104) WARNING 06-28 21:18:59 [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=3631106) INFO 06-28 21:18:59 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:18:59 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:18:59 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:18:59 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:19:00 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:19:00 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:19:00 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:19:00 [loader.py:458] Loading weights took 0.75 seconds
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:19:00 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.847711 seconds
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:19:00 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.853132 seconds
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:19:00 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.918891 seconds
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:19:01 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.961888 seconds
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:19:06 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:19:06 [backends.py:430] Dynamo bytecode transform time: 5.62 s
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:19:06 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:19:06 [backends.py:430] Dynamo bytecode transform time: 5.66 s
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:19:06 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:19:06 [backends.py:430] Dynamo bytecode transform time: 5.66 s
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:19:06 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:19:06 [backends.py:430] Dynamo bytecode transform time: 5.66 s
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:19:11 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.368 s
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:19:11 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.404 s
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:19:11 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.387 s
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:19:12 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.400 s
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:19:17 [monitor.py:33] torch.compile takes 5.62 s in total
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:19:17 [monitor.py:33] torch.compile takes 5.66 s in total
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:19:17 [monitor.py:33] torch.compile takes 5.66 s in total
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:19:17 [monitor.py:33] torch.compile takes 5.66 s in total
+INFO 06-28 21:19:18 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 21:19:18 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 21:19:18 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 21:19:18 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 21:19:18 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 21:19:18 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 21:19:18 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 21:19:18 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3631105) INFO 06-28 21:19:45 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3631106) INFO 06-28 21:19:45 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3631103) INFO 06-28 21:19:45 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3631104) INFO 06-28 21:19:45 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 21:19:45 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.15 seconds
+INFO 06-28 21:19:45 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 21:32:27 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 21:32:27 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.4954|±  |0.0275|
+|                  |       |math_pass@1:1_samples|0.8201|±  |0.0365|
+|mm\|arc_challenge\|0|      0|sem                  |0.5774|±  |0.0253|
+|mm\|arc_easy\|0     |      0|sem                  |0.6452|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4781|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7651|±  |0.0201|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8750|±  |0.0530|
+|mm\|truthfulqa\|0   |      0|sem                  |0.2810|±  |0.0410|
+

merge_bench/logs/llama_ties_5.log

@@ -0,0 +1,96 @@
+INFO 06-28 21:32:26 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 21:32:28 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 21:32:35 [config.py:717] This model supports multiple tasks: {'embed', 'classify', 'reward', 'generate', 'score'}. Defaulting to 'generate'.
+INFO 06-28 21:32:35 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 21:32:35 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 21:32:37 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 21:32:37 [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 06-28 21:32:37 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_f77824a9'), local_subscribe_addr='ipc:///tmp/0c065325-9bd3-44ae-b570-badee1d8a29a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:32:37 [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 0x14a180e77dc0>
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:37 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e1ffa07c'), local_subscribe_addr='ipc:///tmp/188479cc-b446-481e-a6cf-70a39b355001', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:32:37 [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 0x14a17710cb20>
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:37 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_f0f19da6'), local_subscribe_addr='ipc:///tmp/ba9731f7-cc39-4a0e-b49e-d77734e64886', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:32:37 [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 0x14a180e77d00>
+WARNING 06-28 21:32:37 [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 0x14a180e77ac0>
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:37 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_6669c6e5'), local_subscribe_addr='ipc:///tmp/b31583d8-d023-4014-bf0c-ae58a5bca37a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:37 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_9793bc93'), local_subscribe_addr='ipc:///tmp/11e9b52f-d9ea-4a23-b8e3-05261cad20b0', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:40 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:40 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:40 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:40 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:40 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:40 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:40 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:40 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3633070) WARNING 06-28 21:32:40 [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=2 pid=3633071) WARNING 06-28 21:32:40 [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=3633069) WARNING 06-28 21:32:40 [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=3633072) WARNING 06-28 21:32:40 [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=3633069) INFO 06-28 21:32:40 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_d275a2db'), local_subscribe_addr='ipc:///tmp/b1872141-efd7-4f10-9c2b-1c8b9c1e6159', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:40 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:40 [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=3633070) INFO 06-28 21:32:40 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:40 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:40 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:40 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3633072) WARNING 06-28 21:32:40 [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=3633071) WARNING 06-28 21:32:40 [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=3633070) INFO 06-28 21:32:40 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:40 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3633070) WARNING 06-28 21:32:40 [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=3633069) WARNING 06-28 21:32:40 [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=3633072) INFO 06-28 21:32:40 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:40 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:40 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:40 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:41 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:41 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:41 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:41 [loader.py:458] Loading weights took 0.75 seconds
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:41 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.852514 seconds
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:41 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.853413 seconds
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:41 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.973925 seconds
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:42 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.914584 seconds
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:47 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:47 [backends.py:430] Dynamo bytecode transform time: 5.87 s
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:47 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:47 [backends.py:430] Dynamo bytecode transform time: 5.91 s
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:47 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:47 [backends.py:430] Dynamo bytecode transform time: 5.96 s
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:48 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:48 [backends.py:430] Dynamo bytecode transform time: 5.97 s
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:53 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.429 s
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:53 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.414 s
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:53 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.475 s
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:53 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.497 s
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:32:58 [monitor.py:33] torch.compile takes 5.97 s in total
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:32:58 [monitor.py:33] torch.compile takes 5.91 s in total
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:32:58 [monitor.py:33] torch.compile takes 5.87 s in total
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:32:58 [monitor.py:33] torch.compile takes 5.96 s in total
+INFO 06-28 21:33:00 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 21:33:00 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 21:33:00 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 21:33:00 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 21:33:00 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 21:33:00 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 21:33:00 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 21:33:00 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3633071) INFO 06-28 21:33:27 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3633072) INFO 06-28 21:33:27 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3633069) INFO 06-28 21:33:27 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3633070) INFO 06-28 21:33:27 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+INFO 06-28 21:33:27 [core.py:159] init engine (profile, create kv cache, warmup model) took 45.17 seconds
+INFO 06-28 21:33:27 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 21:46:09 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 21:46:09 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5174|±  |0.0277|
+|                  |       |math_pass@1:1_samples|0.7736|±  |0.0423|
+|mm\|arc_challenge\|0|      0|sem                  |0.6220|±  |0.0249|
+|mm\|arc_easy\|0     |      0|sem                  |0.6304|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5031|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7472|±  |0.0206|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8000|±  |0.0641|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3140|±  |0.0424|
+

merge_bench/logs/llama_ties_7.log

@@ -0,0 +1,96 @@
+INFO 06-28 21:46:08 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 21:46:10 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 21:46:17 [config.py:717] This model supports multiple tasks: {'classify', 'reward', 'generate', 'score', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 21:46:17 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 21:46:17 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 21:46:18 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 21:46:18 [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 06-28 21:46:18 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_ba897e64'), local_subscribe_addr='ipc:///tmp/e3fa3541-40e9-45f6-9069-61120d744d93', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:46:19 [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 0x146fbc723df0>
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:19 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e708e433'), local_subscribe_addr='ipc:///tmp/9b2b5f3c-b689-414b-9c6f-aa51bbbdd8b6', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:46:19 [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 0x146fa6d2cb50>
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:19 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a3d2107f'), local_subscribe_addr='ipc:///tmp/49648420-1505-4079-a94a-512c322bc00f', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 21:46:19 [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 0x146fbc723d30>
+WARNING 06-28 21:46:19 [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 0x146fbc723af0>
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:19 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_028e2b27'), local_subscribe_addr='ipc:///tmp/60eb7f5a-ec4e-4d9f-8aa9-28dff9f36b3c', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:19 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_5d554696'), local_subscribe_addr='ipc:///tmp/e1da4acd-28b5-4113-8560-1bba01f5f16a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:26 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:26 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:26 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:26 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:26 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:26 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:26 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:26 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3635043) WARNING 06-28 21:46:26 [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=2 pid=3635042) WARNING 06-28 21:46:26 [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=3635040) WARNING 06-28 21:46:26 [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=3635041) WARNING 06-28 21:46:26 [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=3635040) INFO 06-28 21:46:26 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_e2916777'), local_subscribe_addr='ipc:///tmp/3e96f702-a53e-4867-97a9-d4ef2f1ac5d1', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:26 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:26 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:26 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:26 [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=3635043) INFO 06-28 21:46:26 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3635043) WARNING 06-28 21:46:26 [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=3635042) INFO 06-28 21:46:26 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:26 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:26 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3635042) WARNING 06-28 21:46:26 [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=3635040) WARNING 06-28 21:46:26 [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=3635041) WARNING 06-28 21:46:26 [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=3635043) INFO 06-28 21:46:26 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:26 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:26 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:26 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:27 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:27 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:27 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:27 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:27 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.871172 seconds
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:27 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.899707 seconds
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:28 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.912291 seconds
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:28 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.940351 seconds
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:33 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:33 [backends.py:430] Dynamo bytecode transform time: 5.50 s
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:33 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:33 [backends.py:430] Dynamo bytecode transform time: 5.60 s
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:33 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:33 [backends.py:430] Dynamo bytecode transform time: 5.61 s
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:33 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:33 [backends.py:430] Dynamo bytecode transform time: 5.65 s
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:38 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.333 s
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:38 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.358 s
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:38 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.400 s
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:38 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.373 s
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:46:44 [monitor.py:33] torch.compile takes 5.50 s in total
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:46:44 [monitor.py:33] torch.compile takes 5.61 s in total
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:46:44 [monitor.py:33] torch.compile takes 5.65 s in total
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:46:44 [monitor.py:33] torch.compile takes 5.60 s in total
+INFO 06-28 21:46:45 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 21:46:45 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 21:46:45 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 21:46:45 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 21:46:45 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 21:46:45 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 21:46:45 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 21:46:45 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3635043) INFO 06-28 21:47:10 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3635042) INFO 06-28 21:47:10 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3635040) INFO 06-28 21:47:10 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3635041) INFO 06-28 21:47:10 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+INFO 06-28 21:47:10 [core.py:159] init engine (profile, create kv cache, warmup model) took 42.66 seconds
+INFO 06-28 21:47:11 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 21:59:57 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 21:59:57 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5003|±  |0.0276|
+|                  |       |math_pass@1:1_samples|0.7906|±  |0.0406|
+|mm\|arc_challenge\|0|      0|sem                  |0.5774|±  |0.0253|
+|mm\|arc_easy\|0     |      0|sem                  |0.6283|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5062|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7562|±  |0.0203|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8250|±  |0.0608|
+|mm\|truthfulqa\|0   |      0|sem                  |0.2893|±  |0.0414|
+

merge_bench/logs/llama_ties_9.log

@@ -0,0 +1,96 @@
+INFO 06-28 21:59:56 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 21:59:58 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 22:00:05 [config.py:717] This model supports multiple tasks: {'embed', 'score', 'generate', 'reward', 'classify'}. Defaulting to 'generate'.
+INFO 06-28 22:00:05 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 22:00:05 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 22:00:07 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 22:00:07 [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 06-28 22:00:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_072005e2'), local_subscribe_addr='ipc:///tmp/3b8324da-8e55-4477-8b70-faf81399ad67', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 22:00:07 [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 0x1530cb067d30>
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_333f2675'), local_subscribe_addr='ipc:///tmp/6023b97e-9a60-41ff-8484-ab9fbab5e5b6', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 22:00:07 [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 0x1530c9634a90>
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_133d5830'), local_subscribe_addr='ipc:///tmp/49b7a9c2-4dc6-4fca-ae72-30cc06e6a06a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 22:00:07 [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 0x1530cb067c70>
+WARNING 06-28 22:00:07 [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 0x1530cb0679a0>
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_84d0fca4'), local_subscribe_addr='ipc:///tmp/9511444a-afdf-4220-aac2-d0231e605465', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_f4fb0ebf'), local_subscribe_addr='ipc:///tmp/7da8aa6f-3ad8-49da-95f2-d8239fe6d553', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:15 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:15 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:15 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:15 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:15 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:15 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:15 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:15 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3637009) WARNING 06-28 22:00:15 [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=2 pid=3637008) WARNING 06-28 22:00:15 [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=3637006) WARNING 06-28 22:00:15 [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=3637007) WARNING 06-28 22:00:15 [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=3637006) INFO 06-28 22:00:15 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_911522ef'), local_subscribe_addr='ipc:///tmp/c75ce496-9112-4e97-84ea-5fb9f862786a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:15 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:15 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:15 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:15 [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=3637009) INFO 06-28 22:00:15 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:15 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:15 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3637008) WARNING 06-28 22:00:15 [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=3637009) WARNING 06-28 22:00:15 [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=3637006) WARNING 06-28 22:00:15 [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=3637007) INFO 06-28 22:00:15 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3637007) WARNING 06-28 22:00:15 [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=3637008) INFO 06-28 22:00:15 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:15 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:15 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:15 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:16 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:16 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:16 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:16 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:16 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.855716 seconds
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:17 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.858431 seconds
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:17 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.920326 seconds
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:17 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.960038 seconds
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:22 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:22 [backends.py:430] Dynamo bytecode transform time: 5.70 s
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:22 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:22 [backends.py:430] Dynamo bytecode transform time: 5.72 s
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:23 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:23 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:23 [backends.py:430] Dynamo bytecode transform time: 5.77 s
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:23 [backends.py:430] Dynamo bytecode transform time: 5.77 s
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:28 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.422 s
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:28 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.445 s
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:28 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.425 s
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:28 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.461 s
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:00:33 [monitor.py:33] torch.compile takes 5.77 s in total
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:00:33 [monitor.py:33] torch.compile takes 5.77 s in total
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:00:33 [monitor.py:33] torch.compile takes 5.70 s in total
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:00:33 [monitor.py:33] torch.compile takes 5.72 s in total
+INFO 06-28 22:00:34 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 22:00:34 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 22:00:34 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 22:00:34 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 22:00:34 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 22:00:34 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 22:00:34 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 22:00:34 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3637008) INFO 06-28 22:01:00 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3637009) INFO 06-28 22:01:00 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3637006) INFO 06-28 22:01:00 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3637007) INFO 06-28 22:01:00 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 22:01:00 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.61 seconds
+INFO 06-28 22:01:01 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 22:13:43 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 22:13:43 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5212|±  |0.0279|
+|                  |       |math_pass@1:1_samples|0.7986|±  |0.0389|
+|mm\|arc_challenge\|0|      0|sem                  |0.6142|±  |0.0250|
+|mm\|arc_easy\|0     |      0|sem                  |0.6399|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5000|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7472|±  |0.0206|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8500|±  |0.0572|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3306|±  |0.0429|
+

merge_bench/logs/phi_darelinear_1.log

@@ -0,0 +1,96 @@
+INFO 06-28 01:21:52 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 01:21:54 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 01:22:01 [config.py:717] This model supports multiple tasks: {'reward', 'generate', 'score', 'classify', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 01:22:01 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 01:22:01 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 01:22:03 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 01:22:03 [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 06-28 01:22:03 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_b2217354'), local_subscribe_addr='ipc:///tmp/a3e8bc96-bab3-4345-8a48-730fe105e3e1', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 01:22:03 [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 0x1459a4d4fcd0>
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:03 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_3acfa44b'), local_subscribe_addr='ipc:///tmp/4ef91927-c90f-43eb-a030-37c127c3362d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 01:22:03 [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 0x1459a4d4fc10>
+WARNING 06-28 01:22:03 [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 0x1459a4d4f940>
+WARNING 06-28 01:22:03 [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 0x14599f62ca30>
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:03 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_b59b84f7'), local_subscribe_addr='ipc:///tmp/a968292a-9ad5-4bce-89ea-6eaff6531d1c', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:03 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_3d01424e'), local_subscribe_addr='ipc:///tmp/1d75b13e-e9fb-472c-a09d-757ce058c078', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:03 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_1b952b25'), local_subscribe_addr='ipc:///tmp/00d52892-dc3a-4cf8-babd-7ba75c78873b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:05 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:05 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:05 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:05 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:05 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:05 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:05 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:05 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3509493) WARNING 06-28 01:22:06 [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=3509490) WARNING 06-28 01:22:06 [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=2 pid=3509492) WARNING 06-28 01:22:06 [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=3509491) WARNING 06-28 01:22:06 [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=3509490) INFO 06-28 01:22:06 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_55349baa'), local_subscribe_addr='ipc:///tmp/c13ba9bd-36a6-4b0c-a6a9-cccb260e6d14', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:06 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:06 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:06 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:06 [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=3509493) INFO 06-28 01:22:06 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:06 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3509493) WARNING 06-28 01:22:06 [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=3509490) INFO 06-28 01:22:06 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3509492) WARNING 06-28 01:22:06 [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=3509491) INFO 06-28 01:22:06 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3509490) WARNING 06-28 01:22:06 [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=3509491) WARNING 06-28 01:22:06 [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=3509493) INFO 06-28 01:22:06 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:06 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:06 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:06 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:07 [loader.py:458] Loading weights took 0.75 seconds
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:07 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:07 [loader.py:458] Loading weights took 0.77 seconds
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:07 [loader.py:458] Loading weights took 0.79 seconds
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:07 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.940849 seconds
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:07 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.983826 seconds
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:07 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.939358 seconds
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:07 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 1.013340 seconds
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:13 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:13 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:13 [backends.py:430] Dynamo bytecode transform time: 5.83 s
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:13 [backends.py:430] Dynamo bytecode transform time: 5.83 s
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:13 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:13 [backends.py:430] Dynamo bytecode transform time: 5.83 s
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:13 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:13 [backends.py:430] Dynamo bytecode transform time: 5.87 s
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:18 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.394 s
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:18 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.398 s
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:18 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.453 s
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:18 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.430 s
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:24 [monitor.py:33] torch.compile takes 5.83 s in total
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:24 [monitor.py:33] torch.compile takes 5.83 s in total
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:24 [monitor.py:33] torch.compile takes 5.83 s in total
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:24 [monitor.py:33] torch.compile takes 5.87 s in total
+INFO 06-28 01:22:25 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 01:22:25 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 01:22:25 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 01:22:25 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 01:22:25 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 01:22:25 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 01:22:25 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 01:22:25 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=0 pid=3509490) INFO 06-28 01:22:51 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3509491) INFO 06-28 01:22:51 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3509493) INFO 06-28 01:22:51 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3509492) INFO 06-28 01:22:51 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 01:22:51 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.42 seconds
+INFO 06-28 01:22:52 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 01:35:24 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 01:35:24 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5259|±  |0.0278|
+|                  |       |math_pass@1:1_samples|0.7702|±  |0.0424|
+|mm\|arc_challenge\|0|      0|sem                  |0.6194|±  |0.0249|
+|mm\|arc_easy\|0     |      0|sem                  |0.6367|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5250|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7405|±  |0.0208|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8000|±  |0.0641|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3223|±  |0.0427|
+

merge_bench/logs/phi_darelinear_3.log

@@ -0,0 +1,96 @@
+INFO 06-28 01:35:23 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 01:35:25 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 01:35:32 [config.py:717] This model supports multiple tasks: {'score', 'generate', 'classify', 'embed', 'reward'}. Defaulting to 'generate'.
+INFO 06-28 01:35:32 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 01:35:32 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 01:35:34 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 01:35:34 [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 06-28 01:35:34 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_f1a88531'), local_subscribe_addr='ipc:///tmp/4b8fdbbe-bfe9-49ea-81e6-583208874c6d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 01:35:34 [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 0x149cd159bd90>
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:34 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_191ebca1'), local_subscribe_addr='ipc:///tmp/b850358a-2e43-4778-a548-506d0ca4be92', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 01:35:34 [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 0x149c93bd0af0>
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:34 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_dc536c27'), local_subscribe_addr='ipc:///tmp/17308bf4-154b-4d0f-9777-fd0d7e8f6a83', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 01:35:34 [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 0x149cd159bcd0>
+WARNING 06-28 01:35:34 [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 0x149cd159ba90>
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:34 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_3aa9fd19'), local_subscribe_addr='ipc:///tmp/d6edfdbd-538b-4086-8421-706a2dbd4119', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:34 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_99ba7d8f'), local_subscribe_addr='ipc:///tmp/714e0361-fdc5-4bb8-b328-ecc573c57fd8', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:36 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:36 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:36 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:36 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:36 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:36 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:36 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:36 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3512149) WARNING 06-28 01:35:37 [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=2 pid=3512148) WARNING 06-28 01:35:37 [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=3512147) WARNING 06-28 01:35:37 [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=3512146) WARNING 06-28 01:35:37 [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=3512146) INFO 06-28 01:35:37 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_431a8311'), local_subscribe_addr='ipc:///tmp/8b3b34f6-f867-4925-b5a0-7cd3f6afe61c', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:37 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:37 [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=3512147) INFO 06-28 01:35:37 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:37 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:37 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:37 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3512149) WARNING 06-28 01:35:37 [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=3512147) INFO 06-28 01:35:37 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3512148) WARNING 06-28 01:35:37 [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=3512147) WARNING 06-28 01:35:37 [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=3512146) INFO 06-28 01:35:37 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3512146) WARNING 06-28 01:35:37 [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=3512149) INFO 06-28 01:35:37 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:37 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:37 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:37 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:38 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:38 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:38 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:38 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:38 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.892795 seconds
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:38 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.888533 seconds
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:38 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.959265 seconds
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:38 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.910843 seconds
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:44 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:44 [backends.py:430] Dynamo bytecode transform time: 5.58 s
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:44 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:44 [backends.py:430] Dynamo bytecode transform time: 5.62 s
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:44 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:44 [backends.py:430] Dynamo bytecode transform time: 5.74 s
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:44 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:44 [backends.py:430] Dynamo bytecode transform time: 5.80 s
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:49 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.361 s
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:49 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.393 s
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:49 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.462 s
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:49 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.444 s
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:35:55 [monitor.py:33] torch.compile takes 5.62 s in total
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:35:55 [monitor.py:33] torch.compile takes 5.74 s in total
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:35:55 [monitor.py:33] torch.compile takes 5.58 s in total
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:35:55 [monitor.py:33] torch.compile takes 5.80 s in total
+INFO 06-28 01:35:56 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 01:35:56 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 01:35:56 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 01:35:56 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 01:35:56 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 01:35:56 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 01:35:56 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 01:35:56 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3512148) INFO 06-28 01:36:22 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3512149) INFO 06-28 01:36:22 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3512146) INFO 06-28 01:36:22 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3512147) INFO 06-28 01:36:22 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 01:36:22 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.87 seconds
+INFO 06-28 01:36:22 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 01:48:58 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 01:48:58 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5141|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7988|±  |0.0371|
+|mm\|arc_challenge\|0|      0|sem                  |0.5801|±  |0.0253|
+|mm\|arc_easy\|0     |      0|sem                  |0.6209|±  |0.0158|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5250|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7226|±  |0.0212|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8750|±  |0.0530|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3306|±  |0.0429|
+

merge_bench/logs/phi_darelinear_5.log

@@ -0,0 +1,96 @@
+INFO 06-28 01:48:57 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 01:48:59 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 01:49:06 [config.py:717] This model supports multiple tasks: {'classify', 'generate', 'score', 'reward', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 01:49:06 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 01:49:06 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 01:49:07 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 01:49:07 [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 06-28 01:49:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_e6ed1dc2'), local_subscribe_addr='ipc:///tmp/7eae7c1e-515b-41c5-b887-41e9ee2cf4ea', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 01:49:07 [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 0x14b8ef05bd90>
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_1080e32a'), local_subscribe_addr='ipc:///tmp/7db8a905-31e6-402f-8555-ad1d4729fe0e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 01:49:07 [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 0x14b8ed628af0>
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a26fcf28'), local_subscribe_addr='ipc:///tmp/bd04efc5-4fbe-476e-9a5e-f560a6840f61', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 01:49:07 [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 0x14b8ef05ba90>
+WARNING 06-28 01:49:07 [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 0x14b8ef05bcd0>
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_189894f8'), local_subscribe_addr='ipc:///tmp/d29cdad2-23d9-499b-9441-e80446dc5912', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:07 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_508e7988'), local_subscribe_addr='ipc:///tmp/96aceb6e-f7bf-41d1-a76c-0c8d028246dc', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:10 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:10 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:10 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:10 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:10 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:10 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:10 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:10 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3515448) WARNING 06-28 01:49:10 [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=3515449) WARNING 06-28 01:49:10 [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=3515446) WARNING 06-28 01:49:10 [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=3515447) WARNING 06-28 01:49:10 [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=3515446) INFO 06-28 01:49:10 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_110906cd'), local_subscribe_addr='ipc:///tmp/bae02c6c-f9ee-49c6-81f8-e4d4afacfdf2', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:11 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:11 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:11 [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=3515448) INFO 06-28 01:49:11 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:11 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:11 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3515448) WARNING 06-28 01:49:11 [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=3515449) WARNING 06-28 01:49:11 [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=3515446) WARNING 06-28 01:49:11 [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=3515447) INFO 06-28 01:49:11 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:11 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:11 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:11 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3515447) WARNING 06-28 01:49:11 [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=3515446) INFO 06-28 01:49:11 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:11 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:11 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:11 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:11 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:11 [loader.py:458] Loading weights took 0.75 seconds
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:12 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.911455 seconds
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:12 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.883583 seconds
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:12 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.973162 seconds
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:12 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.938663 seconds
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:17 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:17 [backends.py:430] Dynamo bytecode transform time: 5.61 s
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:18 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:18 [backends.py:430] Dynamo bytecode transform time: 5.72 s
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:18 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:18 [backends.py:430] Dynamo bytecode transform time: 5.78 s
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:18 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:18 [backends.py:430] Dynamo bytecode transform time: 5.81 s
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:22 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.387 s
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:23 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.424 s
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:23 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.454 s
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:23 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.450 s
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:28 [monitor.py:33] torch.compile takes 5.78 s in total
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:28 [monitor.py:33] torch.compile takes 5.72 s in total
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:28 [monitor.py:33] torch.compile takes 5.81 s in total
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:28 [monitor.py:33] torch.compile takes 5.61 s in total
+INFO 06-28 01:49:30 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 01:49:30 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 01:49:30 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 01:49:30 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 01:49:30 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 01:49:30 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 01:49:30 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 01:49:30 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=1 pid=3515447) INFO 06-28 01:49:55 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3515449) INFO 06-28 01:49:55 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3515448) INFO 06-28 01:49:55 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3515446) INFO 06-28 01:49:55 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+INFO 06-28 01:49:55 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.09 seconds
+INFO 06-28 01:49:55 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 02:02:37 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 02:02:37 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5234|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7316|±  |0.0462|
+|mm\|arc_challenge\|0|      0|sem                  |0.6089|±  |0.0250|
+|mm\|arc_easy\|0     |      0|sem                  |0.6315|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5062|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7383|±  |0.0208|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7250|±  |0.0715|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3471|±  |0.0435|
+

merge_bench/logs/phi_darelinear_7.log

@@ -0,0 +1,96 @@
+INFO 06-28 02:02:36 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 02:02:38 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 02:02:45 [config.py:717] This model supports multiple tasks: {'reward', 'classify', 'score', 'generate', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 02:02:45 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 02:02:45 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 02:02:46 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 02:02:46 [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 06-28 02:02:46 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_e30e996e'), local_subscribe_addr='ipc:///tmp/82a2a047-ddbc-4d57-8204-54b364f14611', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:02:47 [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 0x1465321a7df0>
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e04fbba5'), local_subscribe_addr='ipc:///tmp/d2b21b36-91d6-4916-b86d-bd92d2be12f8', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:02:47 [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 0x14653077cb50>
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_cf8334ce'), local_subscribe_addr='ipc:///tmp/565158bd-bf30-47f2-b1bf-3d8242588a09', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:02:47 [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 0x1465321a7d30>
+WARNING 06-28 02:02:47 [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 0x1465321a7af0>
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:02:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_c3766619'), local_subscribe_addr='ipc:///tmp/25817364-7ca4-4212-b549-8c2349e2cdf9', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:47 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_c40b7e23'), local_subscribe_addr='ipc:///tmp/d70aee33-5ed7-4dfd-ad56-427d114db39a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:49 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:49 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:49 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:49 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:02:49 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:49 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:02:49 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:49 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3520096) WARNING 06-28 02:02:49 [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=3520097) WARNING 06-28 02:02:49 [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=3520095) WARNING 06-28 02:02:49 [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=3520094) WARNING 06-28 02:02:49 [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=3520094) INFO 06-28 02:02:49 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_43887bb8'), local_subscribe_addr='ipc:///tmp/9f047318-5779-440e-81ee-e007b90cf083', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:49 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:49 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:49 [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=3520096) INFO 06-28 02:02:49 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:49 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:02:49 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3520097) WARNING 06-28 02:02:49 [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=3520096) WARNING 06-28 02:02:49 [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=3520094) INFO 06-28 02:02:49 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:49 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3520095) WARNING 06-28 02:02:49 [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=3520094) WARNING 06-28 02:02:49 [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=3520096) INFO 06-28 02:02:49 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:49 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:49 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:49 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:50 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:02:50 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:50 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:50 [loader.py:458] Loading weights took 0.74 seconds
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:02:51 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.870194 seconds
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:51 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.857993 seconds
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:51 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.926725 seconds
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:51 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.956879 seconds
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:02:56 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:02:56 [backends.py:430] Dynamo bytecode transform time: 5.65 s
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:56 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:02:56 [backends.py:430] Dynamo bytecode transform time: 5.67 s
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:57 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:02:57 [backends.py:430] Dynamo bytecode transform time: 5.73 s
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:57 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:02:57 [backends.py:430] Dynamo bytecode transform time: 5.77 s
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:03:02 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.403 s
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:03:02 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.440 s
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:03:02 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.459 s
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:03:02 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.458 s
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:03:07 [monitor.py:33] torch.compile takes 5.77 s in total
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:03:07 [monitor.py:33] torch.compile takes 5.73 s in total
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:03:07 [monitor.py:33] torch.compile takes 5.67 s in total
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:03:07 [monitor.py:33] torch.compile takes 5.65 s in total
+INFO 06-28 02:03:09 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 02:03:09 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 02:03:09 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:03:09 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:03:09 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:03:09 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:03:09 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 02:03:09 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3520096) INFO 06-28 02:03:34 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3520097) INFO 06-28 02:03:34 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3520095) INFO 06-28 02:03:34 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3520094) INFO 06-28 02:03:34 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+INFO 06-28 02:03:34 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.12 seconds
+INFO 06-28 02:03:34 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 02:16:05 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 02:16:05 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5132|±  |0.0281|
+|                  |       |math_pass@1:1_samples|0.7533|±  |0.0439|
+|mm\|arc_challenge\|0|      0|sem                  |0.5932|±  |0.0252|
+|mm\|arc_easy\|0     |      0|sem                  |0.6220|±  |0.0158|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4906|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7315|±  |0.0210|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7750|±  |0.0669|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3471|±  |0.0435|
+

merge_bench/logs/phi_darelinear_9.log

@@ -0,0 +1,96 @@
+INFO 06-28 02:16:04 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 02:16:06 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 02:16:13 [config.py:717] This model supports multiple tasks: {'embed', 'generate', 'reward', 'score', 'classify'}. Defaulting to 'generate'.
+INFO 06-28 02:16:13 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 02:16:13 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 02:16:15 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 02:16:15 [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 06-28 02:16:15 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_3d0d1676'), local_subscribe_addr='ipc:///tmp/53082177-1424-40f4-ba33-0aea5b7c1554', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:16:15 [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 0x151621b93df0>
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:15 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a80e1a15'), local_subscribe_addr='ipc:///tmp/99211d1d-a8ab-458b-ad09-678db2b1d0cc', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:16:15 [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 0x1516204d8b50>
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:15 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_8a177959'), local_subscribe_addr='ipc:///tmp/8a42daca-2ec9-48f4-9764-9e1d8675c552', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:16:15 [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 0x151621b93d30>
+WARNING 06-28 02:16:15 [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 0x151621b93af0>
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:15 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_414db17a'), local_subscribe_addr='ipc:///tmp/1608ff44-ab8e-4556-af45-11be3dbe1b61', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:15 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_feaba932'), local_subscribe_addr='ipc:///tmp/c9cc0f59-c525-49dc-8ca9-f743a3e96e03', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:17 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:17 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:17 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:17 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:17 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:17 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:17 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:17 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3523120) WARNING 06-28 02:16:17 [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=2 pid=3523119) WARNING 06-28 02:16:17 [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=3523118) WARNING 06-28 02:16:17 [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=3523117) WARNING 06-28 02:16:17 [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=3523117) INFO 06-28 02:16:17 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_f636ee5c'), local_subscribe_addr='ipc:///tmp/1bb8e3a6-99a5-43af-97d6-b5e194c11329', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:17 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:17 [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=3523117) INFO 06-28 02:16:17 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:17 [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=3523120) INFO 06-28 02:16:17 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:17 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3523120) WARNING 06-28 02:16:17 [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=3523119) WARNING 06-28 02:16:17 [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=3523118) INFO 06-28 02:16:17 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:17 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3523118) WARNING 06-28 02:16:17 [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=3523117) WARNING 06-28 02:16:17 [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=3523120) INFO 06-28 02:16:17 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:17 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:17 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:17 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:18 [loader.py:458] Loading weights took 0.65 seconds
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:18 [loader.py:458] Loading weights took 0.74 seconds
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:18 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:18 [loader.py:458] Loading weights took 0.75 seconds
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:19 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.934303 seconds
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:19 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.928185 seconds
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:19 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.946575 seconds
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:19 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.890070 seconds
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:24 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:24 [backends.py:430] Dynamo bytecode transform time: 5.66 s
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:24 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:24 [backends.py:430] Dynamo bytecode transform time: 5.69 s
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:25 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:25 [backends.py:430] Dynamo bytecode transform time: 5.85 s
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:25 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:25 [backends.py:430] Dynamo bytecode transform time: 5.88 s
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:29 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.407 s
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:30 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.415 s
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:30 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.419 s
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:30 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.421 s
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:16:36 [monitor.py:33] torch.compile takes 5.66 s in total
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:16:36 [monitor.py:33] torch.compile takes 5.88 s in total
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:16:36 [monitor.py:33] torch.compile takes 5.85 s in total
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:16:36 [monitor.py:33] torch.compile takes 5.69 s in total
+INFO 06-28 02:16:37 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 02:16:37 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 02:16:37 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:16:37 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:16:37 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:16:37 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:16:37 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 02:16:37 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3523119) INFO 06-28 02:17:03 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3523118) INFO 06-28 02:17:03 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3523120) INFO 06-28 02:17:03 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3523117) INFO 06-28 02:17:03 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 02:17:03 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.03 seconds
+INFO 06-28 02:17:03 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 02:29:52 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 02:29:52 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5253|±  |0.0279|
+|                  |       |math_pass@1:1_samples|0.7848|±  |0.0420|
+|mm\|arc_challenge\|0|      0|sem                  |0.6142|±  |0.0250|
+|mm\|arc_easy\|0     |      0|sem                  |0.6378|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5188|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7696|±  |0.0199|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8000|±  |0.0641|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3306|±  |0.0429|
+

merge_bench/logs/phi_linear_1.log

@@ -0,0 +1,100 @@
+INFO 06-27 02:28:22 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 02:28:24 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 02:28:30 [config.py:717] This model supports multiple tasks: {'score', 'reward', 'generate', 'classify', 'embed'}. Defaulting to 'generate'.
+INFO 06-27 02:28:31 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 02:28:31 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 02:28:32 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 02:28:32 [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 06-27 02:28:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_64c2b29a'), local_subscribe_addr='ipc:///tmp/4c604289-52fe-4fbf-9aff-29f47c927adc', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:28:32 [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 0x14ba9dc07d30>
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_ef238695'), local_subscribe_addr='ipc:///tmp/f8eead83-296e-4a81-beb2-2f42053e6457', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:28:32 [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 0x14ba9c2aca90>
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_41841aee'), local_subscribe_addr='ipc:///tmp/9a6df31c-9333-4ead-9b5d-3f7288cd7b00', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:28:32 [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 0x14ba9dc07c70>
+WARNING 06-27 02:28:32 [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 0x14ba9dc079a0>
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_7c043b9e'), local_subscribe_addr='ipc:///tmp/360601b1-8515-49b4-955a-b18734598b9d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_d3023a17'), local_subscribe_addr='ipc:///tmp/82620656-c113-4dc5-bd83-39d18cd8bf2d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3429733) WARNING 06-27 02:28:35 [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=3429734) WARNING 06-27 02:28:35 [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=3429732) WARNING 06-27 02:28:35 [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=3429731) WARNING 06-27 02:28:35 [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=3429731) INFO 06-27 02:28:35 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_1e3a61da'), local_subscribe_addr='ipc:///tmp/b4f1a266-7798-40fa-82df-2c25b5062d45', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:35 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:35 [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=3429732) INFO 06-27 02:28:35 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:35 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:35 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:35 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3429734) WARNING 06-27 02:28:35 [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=3429733) WARNING 06-27 02:28:35 [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=3429731) INFO 06-27 02:28:35 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:35 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3429731) WARNING 06-27 02:28:35 [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=3429732) WARNING 06-27 02:28:35 [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=3429734) INFO 06-27 02:28:35 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:35 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:35 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:35 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:36 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:36 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:36 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:36 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.866745 seconds
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.870209 seconds
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.912511 seconds
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.957005 seconds
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:42 [backends.py:430] Dynamo bytecode transform time: 5.68 s
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:42 [backends.py:430] Dynamo bytecode transform time: 5.73 s
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:42 [backends.py:430] Dynamo bytecode transform time: 5.80 s
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:42 [backends.py:430] Dynamo bytecode transform time: 5.88 s
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:28:46 [backends.py:136] Cache the graph of shape None for later use
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:28:46 [backends.py:136] Cache the graph of shape None for later use
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:28:46 [backends.py:136] Cache the graph of shape None for later use
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:28:47 [backends.py:136] Cache the graph of shape None for later use
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:29:07 [backends.py:148] Compiling a graph for general shape takes 24.70 s
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:29:07 [backends.py:148] Compiling a graph for general shape takes 24.72 s
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:29:07 [backends.py:148] Compiling a graph for general shape takes 24.83 s
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:29:08 [backends.py:148] Compiling a graph for general shape takes 24.88 s
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:29:29 [monitor.py:33] torch.compile takes 30.76 s in total
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:29:29 [monitor.py:33] torch.compile takes 30.63 s in total
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:29:29 [monitor.py:33] torch.compile takes 30.37 s in total
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:29:29 [monitor.py:33] torch.compile takes 30.45 s in total
+INFO 06-27 02:29:31 [kv_cache_utils.py:634] GPU KV cache size: 1,999,536 tokens
+INFO 06-27 02:29:31 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 976.34x
+INFO 06-27 02:29:31 [kv_cache_utils.py:634] GPU KV cache size: 1,999,280 tokens
+INFO 06-27 02:29:31 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 976.21x
+INFO 06-27 02:29:31 [kv_cache_utils.py:634] GPU KV cache size: 1,999,280 tokens
+INFO 06-27 02:29:31 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 976.21x
+INFO 06-27 02:29:31 [kv_cache_utils.py:634] GPU KV cache size: 2,000,560 tokens
+INFO 06-27 02:29:31 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 976.84x
+(VllmWorker rank=3 pid=3429734) INFO 06-27 02:30:01 [gpu_model_runner.py:1686] Graph capturing finished in 30 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3429733) INFO 06-27 02:30:01 [gpu_model_runner.py:1686] Graph capturing finished in 30 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3429732) INFO 06-27 02:30:01 [gpu_model_runner.py:1686] Graph capturing finished in 30 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3429731) INFO 06-27 02:30:01 [gpu_model_runner.py:1686] Graph capturing finished in 30 secs, took 2.96 GiB
+INFO 06-27 02:30:01 [core.py:159] init engine (profile, create kv cache, warmup model) took 84.99 seconds
+INFO 06-27 02:30:01 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 02:42:44 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 02:42:44 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5197|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7193|±  |0.0465|
+|mm\|arc_challenge\|0|      0|sem                  |0.5906|±  |0.0252|
+|mm\|arc_easy\|0     |      0|sem                  |0.6367|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5125|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7136|±  |0.0214|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7250|±  |0.0715|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3388|±  |0.0432|
+

merge_bench/logs/phi_linear_2.log

@@ -0,0 +1,96 @@
+INFO 06-27 02:42:43 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 02:42:45 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 02:42:52 [config.py:717] This model supports multiple tasks: {'classify', 'reward', 'generate', 'score', 'embed'}. Defaulting to 'generate'.
+INFO 06-27 02:42:52 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 02:42:52 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 02:42:54 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 02:42:54 [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 06-27 02:42:54 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_f75c7f2e'), local_subscribe_addr='ipc:///tmp/128f64f5-39ca-4318-b6d6-702afc25e764', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:42:54 [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 0x14ea69747d60>
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:42:54 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_8c6a899a'), local_subscribe_addr='ipc:///tmp/bfe1aa7b-3caf-4823-aee3-3b040821364d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:42:54 [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 0x14ea5bcd0ac0>
+WARNING 06-27 02:42:54 [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 0x14ea69747ca0>
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:42:54 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_91075db3'), local_subscribe_addr='ipc:///tmp/fb40d5e0-49c2-414a-b62e-75f1e168124f', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:42:54 [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 0x14ea697479d0>
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:42:54 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_56dad22d'), local_subscribe_addr='ipc:///tmp/bdc6a620-d195-4169-96e6-e1afc95097d1', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:42:54 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_bbd791fe'), local_subscribe_addr='ipc:///tmp/8bc2120a-e48f-4551-be13-deeee6692171', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:42:56 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:42:56 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:42:56 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:42:56 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:42:56 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:42:56 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:42:56 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:42:56 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3438994) WARNING 06-27 02:42:56 [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=2 pid=3438993) WARNING 06-27 02:42:56 [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=3438992) WARNING 06-27 02:42:56 [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=3438991) WARNING 06-27 02:42:56 [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=3438991) INFO 06-27 02:42:56 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_835e5bd0'), local_subscribe_addr='ipc:///tmp/53797ae2-24e6-41c5-a70a-50496a13bf1a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:42:56 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:42:56 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:42:56 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3438992) WARNING 06-27 02:42:56 [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=3438991) INFO 06-27 02:42:56 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:42:56 [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=3438991) WARNING 06-27 02:42:56 [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=3438993) INFO 06-27 02:42:56 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:42:56 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3438994) WARNING 06-27 02:42:56 [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=3438993) INFO 06-27 02:42:56 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3438993) WARNING 06-27 02:42:56 [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=3438992) INFO 06-27 02:42:56 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:42:56 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:42:56 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:42:56 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:42:57 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:42:57 [loader.py:458] Loading weights took 0.66 seconds
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:42:57 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:42:57 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:42:57 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.881809 seconds
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:42:57 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.906570 seconds
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:42:58 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.904886 seconds
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:42:58 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.923940 seconds
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:43:03 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:43:03 [backends.py:430] Dynamo bytecode transform time: 5.62 s
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:43:04 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:43:04 [backends.py:430] Dynamo bytecode transform time: 5.69 s
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:43:04 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:43:04 [backends.py:430] Dynamo bytecode transform time: 5.76 s
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:43:04 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:43:04 [backends.py:430] Dynamo bytecode transform time: 5.94 s
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:43:09 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.426 s
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:43:09 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.417 s
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:43:09 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.500 s
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:43:09 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.542 s
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:43:15 [monitor.py:33] torch.compile takes 5.69 s in total
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:43:15 [monitor.py:33] torch.compile takes 5.94 s in total
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:43:15 [monitor.py:33] torch.compile takes 5.62 s in total
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:43:15 [monitor.py:33] torch.compile takes 5.76 s in total
+INFO 06-27 02:43:16 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-27 02:43:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-27 02:43:16 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 02:43:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 02:43:16 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 02:43:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 02:43:16 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-27 02:43:16 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3438994) INFO 06-27 02:43:42 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3438993) INFO 06-27 02:43:42 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3438992) INFO 06-27 02:43:42 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3438991) INFO 06-27 02:43:42 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-27 02:43:42 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.34 seconds
+INFO 06-27 02:43:43 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 02:56:22 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 02:56:22 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5228|±  |0.0278|
+|                  |       |math_pass@1:1_samples|0.7669|±  |0.0425|
+|mm\|arc_challenge\|0|      0|sem                  |0.6168|±  |0.0249|
+|mm\|arc_easy\|0     |      0|sem                  |0.6241|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5281|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7338|±  |0.0209|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8000|±  |0.0641|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3223|±  |0.0427|
+

merge_bench/logs/phi_linear_3.log

@@ -0,0 +1,96 @@
+INFO 06-27 02:56:21 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 02:56:23 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 02:56:30 [config.py:717] This model supports multiple tasks: {'classify', 'generate', 'reward', 'embed', 'score'}. Defaulting to 'generate'.
+INFO 06-27 02:56:30 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 02:56:30 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 02:56:32 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 02:56:32 [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 06-27 02:56:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_033fa3da'), local_subscribe_addr='ipc:///tmp/a1051549-2dec-4688-9494-0718440d0c2a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:56:32 [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 0x14d51a9dbd60>
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_6383bb93'), local_subscribe_addr='ipc:///tmp/94832504-1a54-45c4-a0b8-da03a934d306', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:56:32 [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 0x14d518db4ac0>
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_4b8e0cc5'), local_subscribe_addr='ipc:///tmp/9db3c5c5-8593-4669-b9ec-1b2dfcd93261', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 02:56:32 [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 0x14d51a9db9d0>
+WARNING 06-27 02:56:32 [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 0x14d51a9dbca0>
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_b6e0d56f'), local_subscribe_addr='ipc:///tmp/3ebc8f6d-94f2-4041-b265-1596b1a7e5c3', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:32 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_43b80aeb'), local_subscribe_addr='ipc:///tmp/33b51447-7fe1-46a4-a147-ba820a37718b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3447778) WARNING 06-27 02:56:34 [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=2 pid=3447776) WARNING 06-27 02:56:34 [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=3447772) WARNING 06-27 02:56:34 [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=3447773) WARNING 06-27 02:56:34 [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=3447772) INFO 06-27 02:56:34 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_4f17e707'), local_subscribe_addr='ipc:///tmp/cc1369ad-3d5f-4a77-a9bf-bfb609385bef', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:34 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:34 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:34 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:34 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:34 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:34 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3447776) WARNING 06-27 02:56:34 [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=3447778) WARNING 06-27 02:56:34 [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=3447772) INFO 06-27 02:56:34 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:34 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3447772) WARNING 06-27 02:56:34 [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=3447773) WARNING 06-27 02:56:34 [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=3447778) INFO 06-27 02:56:34 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:34 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:34 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:34 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:35 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:35 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:35 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:35 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.878122 seconds
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.878944 seconds
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.937358 seconds
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.907204 seconds
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:41 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:41 [backends.py:430] Dynamo bytecode transform time: 5.68 s
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:42 [backends.py:430] Dynamo bytecode transform time: 5.80 s
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:42 [backends.py:430] Dynamo bytecode transform time: 5.82 s
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:42 [backends.py:430] Dynamo bytecode transform time: 5.88 s
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:47 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.405 s
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:47 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.381 s
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:47 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.368 s
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:47 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.486 s
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:56:52 [monitor.py:33] torch.compile takes 5.82 s in total
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:56:52 [monitor.py:33] torch.compile takes 5.68 s in total
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:56:52 [monitor.py:33] torch.compile takes 5.80 s in total
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:56:52 [monitor.py:33] torch.compile takes 5.88 s in total
+INFO 06-27 02:56:54 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-27 02:56:54 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-27 02:56:54 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 02:56:54 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 02:56:54 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 02:56:54 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 02:56:54 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-27 02:56:54 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3447778) INFO 06-27 02:57:20 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3447776) INFO 06-27 02:57:20 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3447772) INFO 06-27 02:57:20 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3447773) INFO 06-27 02:57:20 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-27 02:57:20 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.10 seconds
+INFO 06-27 02:57:20 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 03:10:01 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 03:10:01 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5275|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7814|±  |0.0421|
+|mm\|arc_challenge\|0|      0|sem                  |0.6115|±  |0.0250|
+|mm\|arc_easy\|0     |      0|sem                  |0.6315|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5281|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7629|±  |0.0201|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8000|±  |0.0641|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3388|±  |0.0432|
+

merge_bench/logs/phi_linear_4.log

@@ -0,0 +1,96 @@
+INFO 06-27 03:10:00 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 03:10:02 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 03:10:08 [config.py:717] This model supports multiple tasks: {'embed', 'classify', 'score', 'generate', 'reward'}. Defaulting to 'generate'.
+INFO 06-27 03:10:09 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 03:10:09 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 03:10:10 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 03:10:10 [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 06-27 03:10:10 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_8f74b1a6'), local_subscribe_addr='ipc:///tmp/d6a3a804-af20-451f-bbde-88cafad6d09e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:10:10 [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 0x14d5646ffcd0>
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:10 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_b833bb60'), local_subscribe_addr='ipc:///tmp/9b12d81b-f5a0-4a07-abff-2a89c6e7efc3', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:10:10 [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 0x14d552c40a30>
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:10 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_df1b06e4'), local_subscribe_addr='ipc:///tmp/5fbbbe13-7f22-462d-806c-b109618e0050', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:10:10 [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 0x14d5646ffc10>
+WARNING 06-27 03:10:10 [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 0x14d5646ff940>
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:10 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_fd88f513'), local_subscribe_addr='ipc:///tmp/5519fa92-0b12-4661-87d3-60cc8720f215', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:10 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_b47742a0'), local_subscribe_addr='ipc:///tmp/df1ac424-4f25-4f30-8aa8-4fa48b068a28', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:12 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:12 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:12 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:12 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:12 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:12 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:12 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:12 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3455003) WARNING 06-27 03:10:13 [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=3455001) WARNING 06-27 03:10:13 [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=3455004) WARNING 06-27 03:10:13 [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=3455002) WARNING 06-27 03:10:13 [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=3455001) INFO 06-27 03:10:13 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_c686f274'), local_subscribe_addr='ipc:///tmp/6acb8797-fe95-4e20-b79c-c3db8cabe649', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:13 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:13 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:13 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:13 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:13 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:13 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:13 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:13 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3455003) WARNING 06-27 03:10:13 [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=3455004) WARNING 06-27 03:10:13 [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=3455001) WARNING 06-27 03:10:13 [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=3455002) WARNING 06-27 03:10:13 [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=3455003) INFO 06-27 03:10:13 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:13 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:13 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:13 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:14 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:14 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:14 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:14 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:14 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.892547 seconds
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:14 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.893978 seconds
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:14 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.888170 seconds
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:15 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.959422 seconds
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:20 [backends.py:430] Dynamo bytecode transform time: 5.52 s
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:20 [backends.py:430] Dynamo bytecode transform time: 5.59 s
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:20 [backends.py:430] Dynamo bytecode transform time: 5.64 s
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:20 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:20 [backends.py:430] Dynamo bytecode transform time: 5.79 s
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:25 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.333 s
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:25 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.369 s
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:25 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.354 s
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:26 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.428 s
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:31 [monitor.py:33] torch.compile takes 5.59 s in total
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:31 [monitor.py:33] torch.compile takes 5.52 s in total
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:31 [monitor.py:33] torch.compile takes 5.79 s in total
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:31 [monitor.py:33] torch.compile takes 5.64 s in total
+INFO 06-27 03:10:32 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-27 03:10:32 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-27 03:10:32 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 03:10:32 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 03:10:32 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 03:10:32 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 03:10:32 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-27 03:10:32 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3455003) INFO 06-27 03:10:58 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3455001) INFO 06-27 03:10:58 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3455004) INFO 06-27 03:10:58 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3455002) INFO 06-27 03:10:58 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-27 03:10:59 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.01 seconds
+INFO 06-27 03:10:59 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 03:23:34 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 03:23:34 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5022|±  |0.0273|
+|                  |       |math_pass@1:1_samples|0.7850|±  |0.0407|
+|mm\|arc_challenge\|0|      0|sem                  |0.6142|±  |0.0250|
+|mm\|arc_easy\|0     |      0|sem                  |0.6283|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4938|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7450|±  |0.0206|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8250|±  |0.0608|
+|mm\|truthfulqa\|0   |      0|sem                  |0.2727|±  |0.0407|
+

merge_bench/logs/phi_linear_5.log

@@ -0,0 +1,96 @@
+INFO 06-27 03:23:33 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 03:23:35 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 03:23:42 [config.py:717] This model supports multiple tasks: {'score', 'embed', 'reward', 'generate', 'classify'}. Defaulting to 'generate'.
+INFO 06-27 03:23:42 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 03:23:42 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 03:23:43 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 03:23:43 [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 06-27 03:23:43 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_f203064f'), local_subscribe_addr='ipc:///tmp/580126a8-c433-4576-8303-f43bfd9ef804', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:23:44 [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 0x15411adc7d60>
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:44 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_22ad3f81'), local_subscribe_addr='ipc:///tmp/17f69624-7605-41fa-9a3c-7ef518f34993', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:23:44 [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 0x154119194ac0>
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:44 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_48bc76d6'), local_subscribe_addr='ipc:///tmp/d0301c4f-b794-4755-a83d-b48f0a57b58d', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:23:44 [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 0x15411adc7ca0>
+WARNING 06-27 03:23:44 [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 0x15411adc79d0>
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:44 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_32db5fc4'), local_subscribe_addr='ipc:///tmp/8781eb5d-4016-45ed-85cf-f2e24cd0b7f0', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:44 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_34e3bba9'), local_subscribe_addr='ipc:///tmp/f7be10cc-06a3-46b4-ba2f-e37bbc8351e1', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:45 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:45 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:45 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:45 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:45 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:45 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:45 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:45 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3457338) WARNING 06-27 03:23:46 [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=3457339) WARNING 06-27 03:23:46 [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=3457335) WARNING 06-27 03:23:46 [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=3457333) WARNING 06-27 03:23:46 [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=3457333) INFO 06-27 03:23:46 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_77f60c8a'), local_subscribe_addr='ipc:///tmp/7bc40d89-5a96-45a9-b7ee-9c97d05ab0cb', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:46 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:46 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:46 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:46 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:46 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:46 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3457335) WARNING 06-27 03:23:46 [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=3457338) WARNING 06-27 03:23:46 [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=3457333) WARNING 06-27 03:23:46 [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=3457339) INFO 06-27 03:23:46 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:46 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:46 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3457339) WARNING 06-27 03:23:46 [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=3457338) INFO 06-27 03:23:46 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:46 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:46 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:47 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:47 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:47 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:47 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:47 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.863905 seconds
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:47 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.873327 seconds
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:47 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.905206 seconds
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:47 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.967568 seconds
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:53 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:53 [backends.py:430] Dynamo bytecode transform time: 5.66 s
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:53 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:53 [backends.py:430] Dynamo bytecode transform time: 5.68 s
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:53 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:53 [backends.py:430] Dynamo bytecode transform time: 5.69 s
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:53 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:53 [backends.py:430] Dynamo bytecode transform time: 5.75 s
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:23:58 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.403 s
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:23:58 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.397 s
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:23:58 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.397 s
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:23:58 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.404 s
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:24:04 [monitor.py:33] torch.compile takes 5.66 s in total
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:24:04 [monitor.py:33] torch.compile takes 5.68 s in total
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:24:04 [monitor.py:33] torch.compile takes 5.75 s in total
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:24:04 [monitor.py:33] torch.compile takes 5.69 s in total
+INFO 06-27 03:24:05 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-27 03:24:05 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-27 03:24:05 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 03:24:05 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 03:24:05 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 03:24:05 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 03:24:05 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-27 03:24:05 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3457338) INFO 06-27 03:24:31 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3457339) INFO 06-27 03:24:31 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3457335) INFO 06-27 03:24:31 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3457333) INFO 06-27 03:24:31 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-27 03:24:31 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.73 seconds
+INFO 06-27 03:24:31 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 03:37:15 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 03:37:15 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5198|±  |0.0283|
+|                  |       |math_pass@1:1_samples|0.7397|±  |0.0452|
+|mm\|arc_challenge\|0|      0|sem                  |0.5748|±  |0.0254|
+|mm\|arc_easy\|0     |      0|sem                  |0.6220|±  |0.0158|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5188|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7293|±  |0.0210|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7500|±  |0.0693|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3636|±  |0.0439|
+

merge_bench/logs/phi_linear_6.log

@@ -0,0 +1,96 @@
+INFO 06-27 03:37:14 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 03:37:16 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 03:37:23 [config.py:717] This model supports multiple tasks: {'score', 'classify', 'generate', 'reward', 'embed'}. Defaulting to 'generate'.
+INFO 06-27 03:37:23 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 03:37:23 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 03:37:25 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 03:37:25 [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 06-27 03:37:25 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_580fe053'), local_subscribe_addr='ipc:///tmp/75627ddd-1278-43f9-9357-73bdfd261eb0', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:37:25 [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 0x14bbc75b3ca0>
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:25 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_38aa069a'), local_subscribe_addr='ipc:///tmp/27c05aef-fc8a-4305-9b84-771d29610914', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:37:25 [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 0x14bbc5994a00>
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:25 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a6e9872f'), local_subscribe_addr='ipc:///tmp/40a3df37-ef3f-4f32-8162-8a898c97b508', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:37:25 [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 0x14bbc75b3be0>
+WARNING 06-27 03:37:25 [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 0x14bbc75b3910>
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:25 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_738e08a9'), local_subscribe_addr='ipc:///tmp/af4d4034-226e-4702-9021-b27ebe87aab7', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:25 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_21fb3f2e'), local_subscribe_addr='ipc:///tmp/fbbf66d1-93ad-4810-be1a-c28e80ccb952', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:27 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:27 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:27 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:27 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:27 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:27 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:27 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:27 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3459154) WARNING 06-27 03:37:27 [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=2 pid=3459152) WARNING 06-27 03:37:27 [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=3459150) WARNING 06-27 03:37:27 [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=3459149) WARNING 06-27 03:37:27 [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=3459149) INFO 06-27 03:37:27 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_ff219e32'), local_subscribe_addr='ipc:///tmp/f13b016b-7c21-4dec-9987-b15d524f465b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:27 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:27 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:27 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:27 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:27 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:27 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3459152) WARNING 06-27 03:37:27 [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=3459154) WARNING 06-27 03:37:27 [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=3459150) INFO 06-27 03:37:27 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:27 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3459150) WARNING 06-27 03:37:27 [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=3459149) WARNING 06-27 03:37:27 [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=3459154) INFO 06-27 03:37:27 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:27 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:27 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:27 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:28 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:28 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:28 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:28 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:28 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.868632 seconds
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:29 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.862858 seconds
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:29 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.945122 seconds
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:29 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.898845 seconds
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:34 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:34 [backends.py:430] Dynamo bytecode transform time: 5.62 s
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:34 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:34 [backends.py:430] Dynamo bytecode transform time: 5.63 s
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:34 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:34 [backends.py:430] Dynamo bytecode transform time: 5.74 s
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:35 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:35 [backends.py:430] Dynamo bytecode transform time: 5.90 s
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:39 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.378 s
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:39 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.414 s
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:40 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.504 s
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:40 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.487 s
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:37:46 [monitor.py:33] torch.compile takes 5.62 s in total
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:37:46 [monitor.py:33] torch.compile takes 5.74 s in total
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:37:46 [monitor.py:33] torch.compile takes 5.63 s in total
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:37:46 [monitor.py:33] torch.compile takes 5.90 s in total
+INFO 06-27 03:37:47 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-27 03:37:47 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-27 03:37:47 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 03:37:47 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 03:37:47 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 03:37:47 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 03:37:47 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-27 03:37:47 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3459152) INFO 06-27 03:38:13 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3459149) INFO 06-27 03:38:13 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3459154) INFO 06-27 03:38:13 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3459150) INFO 06-27 03:38:13 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-27 03:38:13 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.52 seconds
+INFO 06-27 03:38:14 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 03:50:53 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 03:50:53 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5278|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7443|±  |0.0441|
+|mm\|arc_challenge\|0|      0|sem                  |0.6220|±  |0.0249|
+|mm\|arc_easy\|0     |      0|sem                  |0.6325|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5094|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7136|±  |0.0214|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7750|±  |0.0669|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3471|±  |0.0435|
+

merge_bench/logs/phi_linear_7.log

@@ -0,0 +1,96 @@
+INFO 06-27 03:50:52 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 03:50:54 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 03:51:01 [config.py:717] This model supports multiple tasks: {'embed', 'score', 'generate', 'reward', 'classify'}. Defaulting to 'generate'.
+INFO 06-27 03:51:01 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 03:51:01 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 03:51:02 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 03:51:02 [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 06-27 03:51:02 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_714b3dea'), local_subscribe_addr='ipc:///tmp/2f2fd7a1-d76f-40e9-9afb-0889889d0481', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:51:02 [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 0x1460b75f8b20>
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:02 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_862c6351'), local_subscribe_addr='ipc:///tmp/aea04bd6-bf04-43a9-b63b-7e1d4d38328a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:51:02 [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 0x1460bcf9bdc0>
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:02 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a03f8b8d'), local_subscribe_addr='ipc:///tmp/4b33314d-7b7a-4a0f-a474-f73a7288f7e5', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 03:51:02 [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 0x1460bcf9bac0>
+WARNING 06-27 03:51:03 [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 0x1460bcf9bd00>
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:03 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_334cdaef'), local_subscribe_addr='ipc:///tmp/09f5481c-e720-4801-a905-6174d455376a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:03 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_f940b541'), local_subscribe_addr='ipc:///tmp/8c21ebc4-9e4a-4abe-9b85-0b5185ba2ca4', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:04 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:04 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:04 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:04 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:04 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:04 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:04 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:04 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3461659) WARNING 06-27 03:51:05 [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=3461660) WARNING 06-27 03:51:05 [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=3461658) WARNING 06-27 03:51:05 [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=3461657) WARNING 06-27 03:51:05 [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=3461657) INFO 06-27 03:51:05 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_61595aee'), local_subscribe_addr='ipc:///tmp/cf3c5a85-6794-4c17-b25b-e36130b1eef1', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:05 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:05 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:05 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:05 [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=3461660) INFO 06-27 03:51:05 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:05 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3461660) WARNING 06-27 03:51:05 [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=3461659) WARNING 06-27 03:51:05 [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=3461657) INFO 06-27 03:51:05 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:05 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3461657) WARNING 06-27 03:51:05 [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=3461658) WARNING 06-27 03:51:05 [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=3461660) INFO 06-27 03:51:05 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:05 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:05 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:05 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:06 [loader.py:458] Loading weights took 0.67 seconds
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:06 [loader.py:458] Loading weights took 0.66 seconds
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:06 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:06 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:06 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.882061 seconds
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:06 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.863736 seconds
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:06 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.935206 seconds
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:06 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.874226 seconds
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:12 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:12 [backends.py:430] Dynamo bytecode transform time: 5.56 s
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:12 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:12 [backends.py:430] Dynamo bytecode transform time: 5.64 s
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:12 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:12 [backends.py:430] Dynamo bytecode transform time: 5.68 s
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:12 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:12 [backends.py:430] Dynamo bytecode transform time: 5.71 s
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:17 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.375 s
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:17 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.451 s
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:17 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.414 s
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:17 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.425 s
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:23 [monitor.py:33] torch.compile takes 5.56 s in total
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:23 [monitor.py:33] torch.compile takes 5.64 s in total
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:23 [monitor.py:33] torch.compile takes 5.68 s in total
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:23 [monitor.py:33] torch.compile takes 5.71 s in total
+INFO 06-27 03:51:24 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-27 03:51:24 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-27 03:51:24 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 03:51:24 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 03:51:24 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 03:51:24 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 03:51:24 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-27 03:51:24 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3461660) INFO 06-27 03:51:50 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3461659) INFO 06-27 03:51:50 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3461658) INFO 06-27 03:51:50 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3461657) INFO 06-27 03:51:50 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-27 03:51:50 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.27 seconds
+INFO 06-27 03:51:51 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 04:04:33 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 04:04:33 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5260|±  |0.0279|
+|                  |       |math_pass@1:1_samples|0.7533|±  |0.0439|
+|mm\|arc_challenge\|0|      0|sem                  |0.6220|±  |0.0249|
+|mm\|arc_easy\|0     |      0|sem                  |0.6357|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5156|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7315|±  |0.0210|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7750|±  |0.0669|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3306|±  |0.0429|
+

merge_bench/logs/phi_linear_8.log

@@ -0,0 +1,96 @@
+INFO 06-27 04:04:32 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 04:04:34 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 04:04:41 [config.py:717] This model supports multiple tasks: {'generate', 'score', 'embed', 'reward', 'classify'}. Defaulting to 'generate'.
+INFO 06-27 04:04:41 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 04:04:41 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 04:04:43 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 04:04:43 [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 06-27 04:04:43 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_99c4f216'), local_subscribe_addr='ipc:///tmp/f947237b-7f6b-4094-8434-f89ab25aa6b2', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 04:04:43 [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 0x1534b0fb0af0>
+WARNING 06-27 04:04:43 [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 0x1534b2bd3d90>
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:43 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_b587ff9a'), local_subscribe_addr='ipc:///tmp/6fbaae7a-578f-44bf-b908-cc00996cd192', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:43 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_c0f62c48'), local_subscribe_addr='ipc:///tmp/7eb0d6c0-c7f6-4df3-95f8-a9553f4454b3', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 04:04:43 [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 0x1534b2bd3cd0>
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:43 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_7482f04a'), local_subscribe_addr='ipc:///tmp/7a18c3c1-b1e2-445b-88a8-cd69ddcee828', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 04:04:43 [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 0x1534b2bd3a90>
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:43 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_90cc6f29'), local_subscribe_addr='ipc:///tmp/24b5aa5a-c1df-4f7e-bf59-5856daa4fa6b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:45 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:45 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:45 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:45 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:45 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:45 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:45 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:45 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3463534) WARNING 06-27 04:04:45 [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=3463535) WARNING 06-27 04:04:45 [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=2 pid=3463536) WARNING 06-27 04:04:45 [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=3463537) WARNING 06-27 04:04:45 [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=3463534) INFO 06-27 04:04:45 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_6ce5769e'), local_subscribe_addr='ipc:///tmp/54247821-49ed-48c0-a116-4078da2aaa21', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:45 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:45 [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=3463535) INFO 06-27 04:04:45 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:45 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:45 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:45 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3463534) WARNING 06-27 04:04:45 [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=3463536) INFO 06-27 04:04:45 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3463535) WARNING 06-27 04:04:45 [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=3463537) INFO 06-27 04:04:45 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3463536) WARNING 06-27 04:04:45 [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=3463537) WARNING 06-27 04:04:45 [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=3463535) INFO 06-27 04:04:45 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:45 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:45 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:45 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:46 [loader.py:458] Loading weights took 0.66 seconds
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:46 [loader.py:458] Loading weights took 0.65 seconds
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:46 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:46 [loader.py:458] Loading weights took 0.76 seconds
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:47 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.856983 seconds
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:47 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.856359 seconds
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:47 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.928136 seconds
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:47 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.963085 seconds
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:52 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:52 [backends.py:430] Dynamo bytecode transform time: 5.60 s
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:53 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:53 [backends.py:430] Dynamo bytecode transform time: 5.66 s
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:53 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:53 [backends.py:430] Dynamo bytecode transform time: 5.72 s
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:53 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:53 [backends.py:430] Dynamo bytecode transform time: 5.74 s
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:04:57 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.346 s
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:04:58 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.373 s
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:04:58 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.446 s
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:04:58 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.457 s
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:05:03 [monitor.py:33] torch.compile takes 5.60 s in total
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:05:03 [monitor.py:33] torch.compile takes 5.72 s in total
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:05:03 [monitor.py:33] torch.compile takes 5.66 s in total
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:05:03 [monitor.py:33] torch.compile takes 5.74 s in total
+INFO 06-27 04:05:05 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-27 04:05:05 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-27 04:05:05 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 04:05:05 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 04:05:05 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 04:05:05 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 04:05:05 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-27 04:05:05 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3463537) INFO 06-27 04:05:29 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3463535) INFO 06-27 04:05:29 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3463534) INFO 06-27 04:05:29 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3463536) INFO 06-27 04:05:29 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+INFO 06-27 04:05:29 [core.py:159] init engine (profile, create kv cache, warmup model) took 42.49 seconds
+INFO 06-27 04:05:30 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 04:18:11 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 04:18:11 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5092|±  |0.0280|
+|                  |       |math_pass@1:1_samples|0.7633|±  |0.0437|
+|mm\|arc_challenge\|0|      0|sem                  |0.5748|±  |0.0254|
+|mm\|arc_easy\|0     |      0|sem                  |0.6283|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5031|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7517|±  |0.0205|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7750|±  |0.0669|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3306|±  |0.0429|
+

merge_bench/logs/phi_linear_9.log

@@ -0,0 +1,96 @@
+INFO 06-27 04:18:10 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-27 04:18:12 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-27 04:18:19 [config.py:717] This model supports multiple tasks: {'embed', 'score', 'classify', 'generate', 'reward'}. Defaulting to 'generate'.
+INFO 06-27 04:18:19 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-27 04:18:19 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-27 04:18:20 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-27 04:18:20 [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 06-27 04:18:20 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_38b9ed00'), local_subscribe_addr='ipc:///tmp/927bf6ed-7718-4102-810c-743be7f346f7', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 04:18:21 [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 0x14f57f85fd30>
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:21 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_88552a69'), local_subscribe_addr='ipc:///tmp/d6f5d2c8-7c9b-46ca-a769-431b1606bc1e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 04:18:21 [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 0x14f57def8a90>
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:21 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a1cdac22'), local_subscribe_addr='ipc:///tmp/1004e10b-1501-42e5-8d81-9e7ba71d12c2', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-27 04:18:21 [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 0x14f57f85f9a0>
+WARNING 06-27 04:18:21 [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 0x14f57f85fc70>
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:21 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_2de4b4e3'), local_subscribe_addr='ipc:///tmp/7ceea1d7-a038-4bab-815e-73854524454e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:21 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_c8ff2cd5'), local_subscribe_addr='ipc:///tmp/b866d7f8-e437-4e63-b251-1bdee58d9c83', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:22 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:22 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:22 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:22 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:22 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:22 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:22 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:22 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3465349) WARNING 06-27 04:18:23 [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=3465350) WARNING 06-27 04:18:23 [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=3465347) WARNING 06-27 04:18:23 [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=3465348) WARNING 06-27 04:18:23 [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=3465347) INFO 06-27 04:18:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_a1278e9a'), local_subscribe_addr='ipc:///tmp/d57b6989-89f2-4332-a268-20cdd1112732', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:23 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:23 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:23 [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=3465349) INFO 06-27 04:18:23 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:23 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:23 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3465350) WARNING 06-27 04:18:23 [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=3465347) INFO 06-27 04:18:23 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:23 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3465349) WARNING 06-27 04:18:23 [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=3465347) WARNING 06-27 04:18:23 [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=3465348) WARNING 06-27 04:18:23 [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=3465350) INFO 06-27 04:18:23 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:23 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:23 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:23 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:24 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:24 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:24 [loader.py:458] Loading weights took 0.66 seconds
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:24 [loader.py:458] Loading weights took 0.71 seconds
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:24 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.875188 seconds
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:24 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.871170 seconds
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:24 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.937340 seconds
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:24 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.886571 seconds
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:30 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:30 [backends.py:430] Dynamo bytecode transform time: 5.63 s
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:30 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:30 [backends.py:430] Dynamo bytecode transform time: 5.64 s
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:30 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:30 [backends.py:430] Dynamo bytecode transform time: 5.71 s
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:30 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:30 [backends.py:430] Dynamo bytecode transform time: 5.84 s
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:35 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.371 s
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:35 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.381 s
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:35 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.384 s
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:35 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.415 s
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:18:41 [monitor.py:33] torch.compile takes 5.71 s in total
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:18:41 [monitor.py:33] torch.compile takes 5.63 s in total
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:18:41 [monitor.py:33] torch.compile takes 5.84 s in total
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:18:41 [monitor.py:33] torch.compile takes 5.64 s in total
+INFO 06-27 04:18:42 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-27 04:18:42 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-27 04:18:42 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 04:18:42 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 04:18:42 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-27 04:18:42 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-27 04:18:42 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-27 04:18:42 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3465350) INFO 06-27 04:19:08 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3465349) INFO 06-27 04:19:08 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3465348) INFO 06-27 04:19:08 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3465347) INFO 06-27 04:19:08 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-27 04:19:08 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.42 seconds
+INFO 06-27 04:19:08 [core_client.py:439] Core engine process 0 ready.
+INFO 06-27 04:31:48 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-27 04:31:48 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5194|±  |0.0279|
+|                  |       |math_pass@1:1_samples|0.8031|±  |0.0388|
+|mm\|arc_challenge\|0|      0|sem                  |0.5906|±  |0.0252|
+|mm\|arc_easy\|0     |      0|sem                  |0.6272|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5375|±  |0.0279|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7562|±  |0.0203|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8500|±  |0.0572|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3223|±  |0.0427|
+

merge_bench/logs/phi_ties_1.log

@@ -0,0 +1,96 @@
+INFO 06-28 00:04:19 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 00:04:20 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 00:04:28 [config.py:717] This model supports multiple tasks: {'reward', 'generate', 'score', 'classify', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 00:04:28 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 00:04:28 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 00:04:30 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 00:04:30 [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 06-28 00:04:30 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_2abe6abe'), local_subscribe_addr='ipc:///tmp/457416df-872d-4317-b106-2134e675d0da', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 00:04:30 [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 0x1505d1c4fd90>
+WARNING 06-28 00:04:30 [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 0x1505d1c4fa90>
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:30 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_685504c9'), local_subscribe_addr='ipc:///tmp/2a5ea551-f9c4-451f-adec-163349ef193a', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 00:04:30 [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 0x1505d021caf0>
+WARNING 06-28 00:04:30 [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 0x1505d1c4fcd0>
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:30 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_ee5b58f4'), local_subscribe_addr='ipc:///tmp/dcfa2e9f-cb07-43e4-89da-be93880cfb53', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:30 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_d60d37fc'), local_subscribe_addr='ipc:///tmp/27722dbf-c5aa-42c3-a7aa-0f0e9d7aa849', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:30 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_120c3c8a'), local_subscribe_addr='ipc:///tmp/e38be149-f63e-4ee1-b9e2-7857fd7453f4', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:37 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:37 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:37 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:37 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:37 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:37 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:37 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:37 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3498583) WARNING 06-28 00:04:38 [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=3498581) WARNING 06-28 00:04:38 [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=3498580) WARNING 06-28 00:04:38 [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=2 pid=3498582) WARNING 06-28 00:04:38 [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=3498580) INFO 06-28 00:04:38 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_99e5fc0d'), local_subscribe_addr='ipc:///tmp/46fa865f-a132-40a8-8b2d-8aa796c96a28', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:38 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:38 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:38 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:38 [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=3498583) INFO 06-28 00:04:38 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:38 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:38 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:38 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3498580) WARNING 06-28 00:04:38 [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=3498581) WARNING 06-28 00:04:38 [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=3498583) WARNING 06-28 00:04:38 [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=3498582) WARNING 06-28 00:04:38 [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=3498581) INFO 06-28 00:04:38 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:38 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:38 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:38 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:39 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:39 [loader.py:458] Loading weights took 0.76 seconds
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:39 [loader.py:458] Loading weights took 0.76 seconds
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:39 [loader.py:458] Loading weights took 0.77 seconds
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:39 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.962431 seconds
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:39 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.961445 seconds
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:39 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.955792 seconds
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:39 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 1.009121 seconds
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:45 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:45 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:45 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:45 [backends.py:430] Dynamo bytecode transform time: 6.12 s
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:45 [backends.py:430] Dynamo bytecode transform time: 6.12 s
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:45 [backends.py:430] Dynamo bytecode transform time: 6.12 s
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:45 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:45 [backends.py:430] Dynamo bytecode transform time: 6.12 s
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:51 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.659 s
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:51 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.731 s
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:51 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.703 s
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:51 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.765 s
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:04:56 [monitor.py:33] torch.compile takes 6.12 s in total
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:04:56 [monitor.py:33] torch.compile takes 6.12 s in total
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:04:56 [monitor.py:33] torch.compile takes 6.12 s in total
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:04:56 [monitor.py:33] torch.compile takes 6.12 s in total
+INFO 06-28 00:04:58 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 00:04:58 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 00:04:58 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 00:04:58 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 00:04:58 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 00:04:58 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 00:04:58 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 00:04:58 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3498582) INFO 06-28 00:05:23 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3498580) INFO 06-28 00:05:23 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3498581) INFO 06-28 00:05:23 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3498583) INFO 06-28 00:05:23 [gpu_model_runner.py:1686] Graph capturing finished in 25 secs, took 2.96 GiB
+INFO 06-28 00:05:23 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.71 seconds
+INFO 06-28 00:05:23 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 00:18:11 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 00:18:12 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5163|±  |0.0281|
+|                  |       |math_pass@1:1_samples|0.7770|±  |0.0422|
+|mm\|arc_challenge\|0|      0|sem                  |0.6010|±  |0.0251|
+|mm\|arc_easy\|0     |      0|sem                  |0.6325|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4844|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7539|±  |0.0204|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8000|±  |0.0641|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3471|±  |0.0435|
+

merge_bench/logs/phi_ties_3.log

@@ -0,0 +1,96 @@
+INFO 06-28 00:18:11 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 00:18:12 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 00:18:20 [config.py:717] This model supports multiple tasks: {'reward', 'score', 'generate', 'classify', 'embed'}. Defaulting to 'generate'.
+INFO 06-28 00:18:20 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 00:18:20 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 00:18:21 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 00:18:21 [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 06-28 00:18:21 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_64a3cf43'), local_subscribe_addr='ipc:///tmp/48692c60-28a5-46d0-84af-9a1b5e3fde25', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 00:18:22 [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 0x1490b810fd90>
+WARNING 06-28 00:18:22 [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 0x1490966e0af0>
+WARNING 06-28 00:18:22 [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 0x1490b810fa90>
+WARNING 06-28 00:18:22 [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 0x1490b810fcd0>
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:22 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_3b32a873'), local_subscribe_addr='ipc:///tmp/47b03dd2-7cda-4a81-b056-a71be7086532', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:22 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_9031601a'), local_subscribe_addr='ipc:///tmp/4f5a3729-d58c-4c0c-9aff-59cfa0cf970b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:22 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_0333bbf3'), local_subscribe_addr='ipc:///tmp/f8ecc6d9-14fe-4bb9-9f76-449d44779b1b', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:22 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_d7d0086a'), local_subscribe_addr='ipc:///tmp/6342ef1b-0bd2-437e-a6a8-31352458218f', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:34 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:34 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3502091) WARNING 06-28 00:18:35 [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=3502090) WARNING 06-28 00:18:35 [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=2 pid=3502092) WARNING 06-28 00:18:35 [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=3502093) WARNING 06-28 00:18:35 [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=3502090) INFO 06-28 00:18:35 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_d8c777a2'), local_subscribe_addr='ipc:///tmp/36440de9-abb4-4834-bdcd-4771f8e0a2eb', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:35 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:35 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:35 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:35 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:35 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:35 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3502091) WARNING 06-28 00:18:35 [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=3502090) WARNING 06-28 00:18:35 [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=3502093) INFO 06-28 00:18:35 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:35 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3502093) WARNING 06-28 00:18:35 [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=3502092) WARNING 06-28 00:18:35 [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=3502091) INFO 06-28 00:18:35 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:35 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:35 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:35 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:36 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:36 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:36 [loader.py:458] Loading weights took 0.75 seconds
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:36 [loader.py:458] Loading weights took 0.80 seconds
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.876938 seconds
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.883893 seconds
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.973901 seconds
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:36 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 1.031346 seconds
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:42 [backends.py:430] Dynamo bytecode transform time: 6.00 s
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:42 [backends.py:430] Dynamo bytecode transform time: 6.00 s
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:42 [backends.py:430] Dynamo bytecode transform time: 6.14 s
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:42 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:42 [backends.py:430] Dynamo bytecode transform time: 6.19 s
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:48 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.704 s
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:48 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.755 s
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:48 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 5.082 s
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:48 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 5.240 s
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:18:54 [monitor.py:33] torch.compile takes 6.00 s in total
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:18:54 [monitor.py:33] torch.compile takes 6.00 s in total
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:18:54 [monitor.py:33] torch.compile takes 6.19 s in total
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:18:54 [monitor.py:33] torch.compile takes 6.14 s in total
+INFO 06-28 00:18:55 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 00:18:55 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 00:18:55 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 00:18:55 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 00:18:55 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 00:18:55 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 00:18:55 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 00:18:55 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=0 pid=3502090) INFO 06-28 00:19:22 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3502091) INFO 06-28 00:19:22 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3502093) INFO 06-28 00:19:23 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3502092) INFO 06-28 00:19:23 [gpu_model_runner.py:1686] Graph capturing finished in 27 secs, took 2.96 GiB
+INFO 06-28 00:19:23 [core.py:159] init engine (profile, create kv cache, warmup model) took 46.42 seconds
+INFO 06-28 00:19:23 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 00:31:57 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 00:31:57 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5150|±  |0.0277|
+|                  |       |math_pass@1:1_samples|0.7385|±  |0.0452|
+|mm\|arc_challenge\|0|      0|sem                  |0.6220|±  |0.0249|
+|mm\|arc_easy\|0     |      0|sem                  |0.6241|±  |0.0157|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5000|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7271|±  |0.0211|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.7500|±  |0.0693|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3140|±  |0.0424|
+

merge_bench/logs/phi_ties_5.log

@@ -0,0 +1,96 @@
+INFO 06-28 02:29:51 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 02:29:53 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 02:30:00 [config.py:717] This model supports multiple tasks: {'generate', 'score', 'embed', 'reward', 'classify'}. Defaulting to 'generate'.
+INFO 06-28 02:30:00 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 02:30:00 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 02:30:01 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 02:30:01 [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 06-28 02:30:01 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_604e8e56'), local_subscribe_addr='ipc:///tmp/6a5d04f5-6450-485a-b7da-b2fb9b33f0d9', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:30:02 [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 0x14b29162caf0>
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:02 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_8e0bd40b'), local_subscribe_addr='ipc:///tmp/14b598b5-accc-4bf5-a43a-15a15c4b1a6f', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:30:02 [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 0x14b293063d90>
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:02 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_937ccfd1'), local_subscribe_addr='ipc:///tmp/e1d4207d-bd2c-4669-bdb5-d2808557c43c', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:30:02 [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 0x14b293063cd0>
+WARNING 06-28 02:30:02 [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 0x14b293063a90>
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:02 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_850fb964'), local_subscribe_addr='ipc:///tmp/dfdeb6ab-4cc1-4a1c-b3d0-6a304eeeaadd', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:02 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_9434e564'), local_subscribe_addr='ipc:///tmp/789ee0fa-631c-4e35-a794-701f51a258f1', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:04 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:04 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:04 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:04 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:04 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:04 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:04 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:04 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3525089) WARNING 06-28 02:30:04 [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=3525090) WARNING 06-28 02:30:04 [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=3525088) WARNING 06-28 02:30:04 [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=3525087) WARNING 06-28 02:30:04 [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=3525087) INFO 06-28 02:30:04 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_8a453fcb'), local_subscribe_addr='ipc:///tmp/3b68802e-dd95-43b0-a83e-ed9aadda8540', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:04 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:04 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:04 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:04 [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=3525089) INFO 06-28 02:30:04 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:04 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3525089) WARNING 06-28 02:30:04 [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=3525090) WARNING 06-28 02:30:04 [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=3525088) INFO 06-28 02:30:04 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:04 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3525088) WARNING 06-28 02:30:04 [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=3525087) WARNING 06-28 02:30:04 [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=3525089) INFO 06-28 02:30:04 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:04 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:04 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:04 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:05 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:05 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:05 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:05 [loader.py:458] Loading weights took 0.74 seconds
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:05 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.918278 seconds
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:06 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.922065 seconds
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:06 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.895514 seconds
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:06 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.973592 seconds
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:11 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:11 [backends.py:430] Dynamo bytecode transform time: 5.65 s
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:11 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:11 [backends.py:430] Dynamo bytecode transform time: 5.73 s
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:11 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:11 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:11 [backends.py:430] Dynamo bytecode transform time: 5.81 s
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:11 [backends.py:430] Dynamo bytecode transform time: 5.81 s
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:16 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.407 s
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:17 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.434 s
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:17 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.482 s
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:17 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.454 s
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:22 [monitor.py:33] torch.compile takes 5.81 s in total
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:22 [monitor.py:33] torch.compile takes 5.81 s in total
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:22 [monitor.py:33] torch.compile takes 5.65 s in total
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:22 [monitor.py:33] torch.compile takes 5.73 s in total
+INFO 06-28 02:30:24 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 02:30:24 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 02:30:24 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:30:24 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:30:24 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:30:24 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:30:24 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 02:30:24 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3525090) INFO 06-28 02:30:50 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3525087) INFO 06-28 02:30:50 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3525089) INFO 06-28 02:30:50 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3525088) INFO 06-28 02:30:50 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 02:30:50 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.44 seconds
+INFO 06-28 02:30:50 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 02:43:29 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 02:43:29 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5091|±  |0.0279|
+|                  |       |math_pass@1:1_samples|0.7928|±  |0.0405|
+|mm\|arc_challenge\|0|      0|sem                  |0.5827|±  |0.0253|
+|mm\|arc_easy\|0     |      0|sem                  |0.6410|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4906|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7606|±  |0.0202|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8250|±  |0.0608|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3223|±  |0.0427|
+

merge_bench/logs/phi_ties_7.log

@@ -0,0 +1,96 @@
+INFO 06-28 02:43:28 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 02:43:30 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 02:43:37 [config.py:717] This model supports multiple tasks: {'classify', 'generate', 'reward', 'embed', 'score'}. Defaulting to 'generate'.
+INFO 06-28 02:43:37 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 02:43:37 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 02:43:39 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 02:43:39 [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 06-28 02:43:39 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_901a0879'), local_subscribe_addr='ipc:///tmp/087980f4-6715-439e-ade5-e490bb2ff57e', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:43:39 [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 0x152887198a30>
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:39 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_a12311a9'), local_subscribe_addr='ipc:///tmp/25170eb0-7a2f-4e8d-ad5a-bf695aad19fd', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:43:39 [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 0x152890c43c10>
+WARNING 06-28 02:43:39 [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 0x152890c43cd0>
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:39 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e733a3eb'), local_subscribe_addr='ipc:///tmp/5e2ea1ec-c5f6-4d6f-90a6-4c58def244aa', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:43:39 [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 0x152890c43940>
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:39 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_65524c3c'), local_subscribe_addr='ipc:///tmp/66934d46-2cd6-48fd-861e-d91c8468b582', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:39 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_96394c91'), local_subscribe_addr='ipc:///tmp/ef2160b0-e13f-4fbf-bc5a-3a12e6da0e19', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:41 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:41 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:41 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:41 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:41 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:41 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:41 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:41 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3527443) WARNING 06-28 02:43:41 [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=2 pid=3527442) WARNING 06-28 02:43:41 [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=3527441) WARNING 06-28 02:43:41 [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=3527440) WARNING 06-28 02:43:41 [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=3527440) INFO 06-28 02:43:41 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_91ddd545'), local_subscribe_addr='ipc:///tmp/79bc8596-ad99-4743-9101-657250aa290c', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:41 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:41 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:41 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:41 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3527442) WARNING 06-28 02:43:41 [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=3527443) INFO 06-28 02:43:41 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:41 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:41 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:41 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3527443) WARNING 06-28 02:43:41 [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=3527440) WARNING 06-28 02:43:41 [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=3527441) WARNING 06-28 02:43:41 [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=3527442) INFO 06-28 02:43:41 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:41 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:41 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:41 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:42 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:42 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:42 [loader.py:458] Loading weights took 0.70 seconds
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:42 [loader.py:458] Loading weights took 0.73 seconds
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:42 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.878173 seconds
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:43 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.875533 seconds
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:43 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.944981 seconds
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:43 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.922516 seconds
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:48 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:48 [backends.py:430] Dynamo bytecode transform time: 5.56 s
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:49 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:49 [backends.py:430] Dynamo bytecode transform time: 5.58 s
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:49 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:49 [backends.py:430] Dynamo bytecode transform time: 5.78 s
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:49 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:49 [backends.py:430] Dynamo bytecode transform time: 5.86 s
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:43:53 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.346 s
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:43:54 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.360 s
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:43:54 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.368 s
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:43:54 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.479 s
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:44:00 [monitor.py:33] torch.compile takes 5.56 s in total
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:44:00 [monitor.py:33] torch.compile takes 5.78 s in total
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:44:00 [monitor.py:33] torch.compile takes 5.86 s in total
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:44:00 [monitor.py:33] torch.compile takes 5.58 s in total
+INFO 06-28 02:44:01 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 02:44:01 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 02:44:01 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:44:01 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:44:01 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:44:01 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:44:01 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 02:44:01 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=2 pid=3527442) INFO 06-28 02:44:27 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3527440) INFO 06-28 02:44:27 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=3 pid=3527443) INFO 06-28 02:44:27 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3527441) INFO 06-28 02:44:27 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 02:44:27 [core.py:159] init engine (profile, create kv cache, warmup model) took 43.92 seconds
+INFO 06-28 02:44:27 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 02:57:13 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 02:57:13 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5215|±  |0.0275|
+|                  |       |math_pass@1:1_samples|0.7805|±  |0.0409|
+|mm\|arc_challenge\|0|      0|sem                  |0.6352|±  |0.0247|
+|mm\|arc_easy\|0     |      0|sem                  |0.6410|±  |0.0156|
+|mm\|commonsenseqa\|0|      0|sem                  |0.5125|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7360|±  |0.0209|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.8250|±  |0.0608|
+|mm\|truthfulqa\|0   |      0|sem                  |0.2975|±  |0.0417|
+

merge_bench/logs/phi_ties_9.log

@@ -0,0 +1,96 @@
+INFO 06-28 02:57:12 [__init__.py:239] Automatically detected platform cuda.
+INFO 06-28 02:57:14 [config.py:209] Replacing legacy 'type' key with 'rope_type'
+INFO 06-28 02:57:21 [config.py:717] This model supports multiple tasks: {'reward', 'embed', 'generate', 'classify', 'score'}. Defaulting to 'generate'.
+INFO 06-28 02:57:21 [config.py:1770] Defaulting to use mp for distributed inference
+INFO 06-28 02:57:21 [config.py:2003] Chunked prefill is enabled with max_num_batched_tokens=16384.
+INFO 06-28 02:57:22 [core.py:58] Initializing a V1 LLM engine (v0.8.5.post1) with config: model='./models/R-Phi4', speculative_config=None, tokenizer='./models/R-Phi4', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, 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=./models/R-Phi4, 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 06-28 02:57:22 [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 06-28 02:57:22 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0, 1, 2, 3], buffer_handle=(4, 10485760, 10, 'psm_e5197895'), local_subscribe_addr='ipc:///tmp/b0a44ef9-bda3-4b70-a828-3d8153cb25e8', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:57:23 [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 0x14ff9a4afd00>
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_66e46478'), local_subscribe_addr='ipc:///tmp/3efa8e16-e41b-4f7c-94f9-95ea91664c1f', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:57:23 [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 0x14ff98a78a60>
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_b98a5dc9'), local_subscribe_addr='ipc:///tmp/2870b470-eb7a-4d36-bd20-881fd3cf2c8c', remote_subscribe_addr=None, remote_addr_ipv6=False)
+WARNING 06-28 02:57:23 [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 0x14ff9a4afc40>
+WARNING 06-28 02:57:23 [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 0x14ff9a4af970>
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_263260cf'), local_subscribe_addr='ipc:///tmp/833452ad-0120-435e-b553-3cb039a6a2c9', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:23 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[0], buffer_handle=(1, 10485760, 10, 'psm_e741f63c'), local_subscribe_addr='ipc:///tmp/6de270e3-27d7-4282-8278-117949320a93', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:24 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:24 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:24 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:24 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:24 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:24 [utils.py:1055] Found nccl from library libnccl.so.2
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:24 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:24 [pynccl.py:69] vLLM is using nccl==2.21.5
+(VllmWorker rank=3 pid=3529696) WARNING 06-28 02:57:25 [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=2 pid=3529695) WARNING 06-28 02:57:25 [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=3529694) WARNING 06-28 02:57:25 [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=3529693) WARNING 06-28 02:57:25 [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=3529693) INFO 06-28 02:57:25 [shm_broadcast.py:266] vLLM message queue communication handle: Handle(local_reader_ranks=[1, 2, 3], buffer_handle=(3, 4194304, 6, 'psm_aafaa12e'), local_subscribe_addr='ipc:///tmp/67dc1f4e-7410-4b14-99a6-9177125c8985', remote_subscribe_addr=None, remote_addr_ipv6=False)
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:25 [parallel_state.py:1004] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:25 [parallel_state.py:1004] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:25 [parallel_state.py:1004] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:25 [parallel_state.py:1004] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:25 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:25 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=2 pid=3529695) WARNING 06-28 02:57:25 [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=3529696) WARNING 06-28 02:57:25 [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=3529693) INFO 06-28 02:57:25 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:25 [cuda.py:221] Using Flash Attention backend on V1 engine.
+(VllmWorker rank=0 pid=3529693) WARNING 06-28 02:57:25 [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=3529694) WARNING 06-28 02:57:25 [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=3529695) INFO 06-28 02:57:25 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:25 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:25 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:25 [gpu_model_runner.py:1329] Starting to load model ./models/R-Phi4...
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:26 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:26 [loader.py:458] Loading weights took 0.69 seconds
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:26 [loader.py:458] Loading weights took 0.68 seconds
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:26 [loader.py:458] Loading weights took 0.72 seconds
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:26 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.878988 seconds
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:26 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.886175 seconds
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:26 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.919960 seconds
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:26 [gpu_model_runner.py:1347] Model loading took 1.8196 GiB and 0.959284 seconds
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:32 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_0_0 for vLLM's torch.compile
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:32 [backends.py:430] Dynamo bytecode transform time: 5.59 s
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:32 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_1_0 for vLLM's torch.compile
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:32 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_2_0 for vLLM's torch.compile
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:32 [backends.py:430] Dynamo bytecode transform time: 5.76 s
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:32 [backends.py:430] Dynamo bytecode transform time: 5.76 s
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:32 [backends.py:420] Using cache directory: /home/jiangli/.cache/vllm/torch_compile_cache/bc6735f00d/rank_3_0 for vLLM's torch.compile
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:32 [backends.py:430] Dynamo bytecode transform time: 5.80 s
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:37 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.381 s
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:37 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.387 s
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:37 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.395 s
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:37 [backends.py:118] Directly load the compiled graph(s) for shape None from the cache, took 4.436 s
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:57:43 [monitor.py:33] torch.compile takes 5.76 s in total
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:57:43 [monitor.py:33] torch.compile takes 5.59 s in total
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:57:43 [monitor.py:33] torch.compile takes 5.76 s in total
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:57:43 [monitor.py:33] torch.compile takes 5.80 s in total
+INFO 06-28 02:57:44 [kv_cache_utils.py:634] GPU KV cache size: 2,007,088 tokens
+INFO 06-28 02:57:44 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.02x
+INFO 06-28 02:57:44 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:57:44 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:57:44 [kv_cache_utils.py:634] GPU KV cache size: 2,006,832 tokens
+INFO 06-28 02:57:44 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 979.90x
+INFO 06-28 02:57:44 [kv_cache_utils.py:634] GPU KV cache size: 2,008,112 tokens
+INFO 06-28 02:57:44 [kv_cache_utils.py:637] Maximum concurrency for 2,048 tokens per request: 980.52x
+(VllmWorker rank=3 pid=3529696) INFO 06-28 02:58:10 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=2 pid=3529695) INFO 06-28 02:58:10 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=1 pid=3529694) INFO 06-28 02:58:10 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+(VllmWorker rank=0 pid=3529693) INFO 06-28 02:58:11 [gpu_model_runner.py:1686] Graph capturing finished in 26 secs, took 2.96 GiB
+INFO 06-28 02:58:11 [core.py:159] init engine (profile, create kv cache, warmup model) took 44.11 seconds
+INFO 06-28 02:58:11 [core_client.py:439] Core engine process 0 ready.
+INFO 06-28 03:10:49 [importing.py:53] Triton module has been replaced with a placeholder.
+INFO 06-28 03:10:49 [__init__.py:239] Automatically detected platform cuda.
+|       Task       |Version|       Metric        |Value |   |Stderr|
+|------------------|------:|---------------------|-----:|---|-----:|
+|all               |       |sem                  |0.5092|±  |0.0276|
+|                  |       |math_pass@1:1_samples|0.8303|±  |0.0341|
+|mm\|arc_challenge\|0|      0|sem                  |0.6352|±  |0.0247|
+|mm\|arc_easy\|0     |      0|sem                  |0.6177|±  |0.0158|
+|mm\|commonsenseqa\|0|      0|sem                  |0.4781|±  |0.0280|
+|mm\|gsm8k\|0        |      0|math_pass@1:1_samples|0.7606|±  |0.0202|
+|mm\|math_500\|0     |      3|math_pass@1:1_samples|0.9000|±  |0.0480|
+|mm\|truthfulqa\|0   |      0|sem                  |0.3058|±  |0.0421|
+

merge_bench/outputs/._merged1_llama_darelinear_1/2025-06-23T01-52-10.258150/outputs_mm|arc_challenge|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce3917ead928e2c0b135b0486913789ebce9bf887ab523b118e4dbf44b3d98fb
+size 3529231

merge_bench/outputs/._merged1_llama_darelinear_1/2025-06-23T01-52-10.258150/outputs_mm|arc_easy|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66410b03c3e808718cb27f85ed11686da011eb81f1b30ce185527d6cb428735e
+size 8156073

merge_bench/outputs/._merged1_llama_darelinear_1/2025-06-23T01-52-10.258150/outputs_mm|commonsenseqa|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7377a29f21f91b2bfaf75bbf93a1b2f128fa753f1cefc634829bcb1583732bbd
+size 2862028

merge_bench/outputs/._merged1_llama_darelinear_1/2025-06-23T01-52-10.258150/outputs_mm|gsm8k|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbf37c3148192aa67066977a11a4870049c30d3f71317e44e3e1c65c4cc03dcf
+size 3039563

merge_bench/outputs/._merged1_llama_darelinear_1/2025-06-23T01-52-10.258150/outputs_mm|math_500|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b84257797190eecdbc0ef3e6cb8d4b4e1140e78fbe200889ef405bcc876f98e1
+size 316667

merge_bench/outputs/._merged1_llama_darelinear_1/2025-06-23T01-52-10.258150/outputs_mm|truthfulqa|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:613a394111e16909114efc1a671aa7916329532d45f0eb62d35997e3219189a9
+size 1148077

merge_bench/outputs/._merged1_llama_darelinear_3/2025-06-23T01-52-10.258150/outputs_mm|arc_challenge|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66e45619e88ade20bf58ac95773dd63dff38803210a4384bab83d8b889231794
+size 3518921

merge_bench/outputs/._merged1_llama_darelinear_3/2025-06-23T01-52-10.258150/outputs_mm|arc_easy|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7d59b8f7c561bc7d2e58d99395c5b50d9d7f0b8c7c590003c1f92a8db017111
+size 8166781

merge_bench/outputs/._merged1_llama_darelinear_3/2025-06-23T01-52-10.258150/outputs_mm|commonsenseqa|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a67f6187601b7140b54906bb296894958bc31d84dbcf18c4923a62d11105c33e
+size 2859764

merge_bench/outputs/._merged1_llama_darelinear_3/2025-06-23T01-52-10.258150/outputs_mm|gsm8k|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:081b137da3e9752b00de835a085365af6036fb0b27cfafcc2f495f3fbdad462a
+size 3039375

merge_bench/outputs/._merged1_llama_darelinear_3/2025-06-23T01-52-10.258150/outputs_mm|math_500|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:589a4bb3ac1b5db19871cb67d76a9a3d636887d0ab07c1ceefb7cee55fab7ced
+size 317874

merge_bench/outputs/._merged1_llama_darelinear_3/2025-06-23T01-52-10.258150/outputs_mm|truthfulqa|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4937003c6389d8dddc60b3dcefcfeb410aae72eb792fac7116d4a94913d2230f
+size 1146204

merge_bench/outputs/._merged1_llama_darelinear_5/2025-06-23T01-52-10.258150/outputs_mm|arc_challenge|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b66af07142cf41943266bfde37ff44bc060dccd2e6b9b05f85db2e7c4884e33f
+size 3524830

merge_bench/outputs/._merged1_llama_darelinear_5/2025-06-23T01-52-10.258150/outputs_mm|arc_easy|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb24ecc922da45dc457507d7d08ac984d0b5c371bba77756a5e951bda588ee0c
+size 8167635

merge_bench/outputs/._merged1_llama_darelinear_5/2025-06-23T01-52-10.258150/outputs_mm|commonsenseqa|0_2025-06-23T01-52-10.258150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2be5f6739cba12997dc33b547ae73dae4b91e66747ba2198768b883683cea7d
+size 2860274