Add logging for cudagraph related info (#29825)

Signed-off-by: Yong Hoon Shin <yhshin@meta.com>

vllm/compilation/cuda_graph.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import dataclasses
+from collections import Counter
 from collections.abc import Callable
 from contextlib import ExitStack
 from typing import Any
@@ -22,6 +23,99 @@ from vllm.utils.torch_utils import weak_ref_tensors
 logger = init_logger(__name__)
 
 
+@dataclasses.dataclass(frozen=True)
+class CUDAGraphStat:
+    num_unpadded_tokens: int
+    num_padded_tokens: int
+    num_paddings: int
+    runtime_mode: str
+
+
+class CUDAGraphLogging:
+    """Aggregate and log cudagraph metrics"""
+
+    COLUMN_HEADERS = [
+        "Unpadded Tokens",
+        "Padded Tokens",
+        "Num Paddings",
+        "Runtime Mode",
+        "Count",
+    ]
+
+    def __init__(self, cg_mode: CUDAGraphMode, cg_capture_sizes: list[int] | None):
+        self.reset()
+        self.cg_mode = str(cg_mode)
+        self.cg_capture_sizes = str(cg_capture_sizes or [])
+
+        self.settings_header = (
+            "**CUDAGraph Config Settings:**\n\n"
+            f"- Mode: {self.cg_mode}\n"
+            f"- Capture sizes: {self.cg_capture_sizes}\n\n"
+            "**CUDAGraph Stats:**\n\n"
+        )
+
+    def reset(self):
+        self.stats = []
+
+    def observe(self, cudagraph_stat: CUDAGraphStat):
+        self.stats.append(cudagraph_stat)
+
+    def generate_metric_table(self) -> str:
+        stats_counts = Counter(self.stats)
+
+        # Convert stats to rows of strings, in descending order of observed frequencies
+        rows = []
+        for stat, count in sorted(
+            stats_counts.items(), key=lambda item: item[1], reverse=True
+        ):
+            rows.append(
+                [
+                    str(stat.num_unpadded_tokens),
+                    str(stat.num_padded_tokens),
+                    str(stat.num_paddings),
+                    stat.runtime_mode,
+                    str(count),
+                ]
+            )
+
+        # Calculate column widths (max of header and data)
+        col_widths = []
+        for i, header_text in enumerate(self.COLUMN_HEADERS):
+            max_width = len(header_text)
+            for row in rows:
+                max_width = max(max_width, len(row[i]))
+            col_widths.append(max_width)
+
+        table_header_list = [
+            h.ljust(w) for h, w in zip(self.COLUMN_HEADERS, col_widths)
+        ]
+        table_header = "| " + " | ".join(table_header_list) + " |\n"
+
+        table_separator = "|" + "|".join("-" * (w + 2) for w in col_widths) + "|\n"
+
+        # Create data rows with proper alignment
+        data_rows = []
+        for row in rows:
+            formatted_row = [
+                str(val).ljust(width) for val, width in zip(row, col_widths)
+            ]
+            data_rows.append("| " + " | ".join(formatted_row) + " |")
+
+        return (
+            self.settings_header
+            + table_header
+            + table_separator
+            + "\n".join(data_rows)
+            + "\n"
+        )
+
+    def log(self, log_fn=logger.info):
+        if not self.stats:
+            return
+        log_fn(self.generate_metric_table())
+        self.reset()
+
+
 @dataclasses.dataclass
 class CUDAGraphEntry:
     batch_descriptor: BatchDescriptor