[UX] Add Feedback During CUDAGraph Capture (vllm-project#19501)

Signed-off-by: [email protected] <[email protected]>

Author: robertgshaw2-redhat

vllm/v1/worker/gpu_model_runner.py

@@ -12,6 +12,7 @@
 import torch
 import torch.distributed
 import torch.nn as nn
+from tqdm import tqdm
 
 import vllm.envs as envs
 from vllm.attention import AttentionType, get_attn_backend
@@ -2034,7 +2035,9 @@ def capture_model(self) -> None:
         # can reuse the memory pool allocated for the large shapes.
         with graph_capture(device=self.device):
             skip_attn = not self.vllm_config.compilation_config.full_cuda_graph
-            for num_tokens in reversed(self.cudagraph_batch_sizes):
+            for num_tokens in tqdm(reversed(self.cudagraph_batch_sizes),
+                                   desc="Capturing CUDA graphs",
+                                   total=len(self.cudagraph_batch_sizes)):
                 for _ in range(self.vllm_config.compilation_config.
                                cudagraph_num_of_warmups):
                     self._dummy_run(num_tokens, skip_attn=skip_attn)