diff --git a/vllm/v1/worker/gpu/cudagraph_utils.py b/vllm/v1/worker/gpu/cudagraph_utils.py index 6bd404985d..662f92e581 100644 --- a/vllm/v1/worker/gpu/cudagraph_utils.py +++ b/vllm/v1/worker/gpu/cudagraph_utils.py @@ -9,6 +9,7 @@ import torch import torch.nn as nn from tqdm import tqdm +from vllm.compilation.counter import compilation_counter from vllm.config import VllmConfig from vllm.config.compilation import CUDAGraphMode from vllm.distributed.parallel_state import ( @@ -237,6 +238,7 @@ class CudaGraphManager: # the next forward pass. get_offloader().join_after_forward() self.graphs[desc] = graph + compilation_counter.num_cudagraph_captured += 1 self._graphs_captured = True return captured_attn_states diff --git a/vllm/v1/worker/gpu/model_runner.py b/vllm/v1/worker/gpu/model_runner.py index dfe50cb135..98d889cdbb 100644 --- a/vllm/v1/worker/gpu/model_runner.py +++ b/vllm/v1/worker/gpu/model_runner.py @@ -27,6 +27,7 @@ import numpy as np import torch import torch.nn as nn +from vllm.compilation.counter import compilation_counter from vllm.config import VllmConfig from vllm.config.compilation import CUDAGraphMode from vllm.distributed.parallel_state import ( @@ -587,6 +588,8 @@ class GPUModelRunner(LoRAModelRunnerMixin): ) return 0 + compilation_counter.num_gpu_runner_capture_triggers += 1 + start_time = time.perf_counter() gc.collect() torch.accelerator.empty_cache()