From e85c8826ce2a810367e8a70eedb987e31b140800 Mon Sep 17 00:00:00 2001
From: "Zhiyao(Ian) Yu" <105995086+Ianniu123@users.noreply.github.com>
Date: Mon, 14 Sep 2026 14:00:20 -0400
Subject: [PATCH] [Bugfix] Make KV cache and MFU log lines backend-neutral
 (#55650)

Signed-off-by: Ianniu <985683179yzy@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 vllm/v1/core/kv_cache_utils.py | 4 +++-
 vllm/v1/metrics/loggers.py     | 8 ++++++--
 vllm/v1/metrics/perf.py        | 2 +-
 3 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/vllm/v1/core/kv_cache_utils.py b/vllm/v1/core/kv_cache_utils.py
index b47df61add..28478b69f5 100644
--- a/vllm/v1/core/kv_cache_utils.py
+++ b/vllm/v1/core/kv_cache_utils.py
@@ -2391,9 +2391,11 @@ def update_kv_cache_capacity(
     vllm_config.cache_config.kv_cache_size_tokens = num_tokens
     vllm_config.cache_config.kv_cache_max_concurrency = max_concurrency
     max_model_len = vllm_config.model_config.max_model_len
+    device_type = vllm_config.device_config.device_type
     logger.info_once(
-        "GPU KV cache size: %s tokens, "
+        "%s KV cache size: %s tokens, "
         "Maximum concurrency for %s tokens per request: %.2fx",
+        "GPU" if device_type == "cuda" else device_type.upper(),
         f"{num_tokens:,}",
         f"{max_model_len:,}",
         max_concurrency,
diff --git a/vllm/v1/metrics/loggers.py b/vllm/v1/metrics/loggers.py
index 9e3f9c63a6..3a8adb6e32 100644
--- a/vllm/v1/metrics/loggers.py
+++ b/vllm/v1/metrics/loggers.py
@@ -101,6 +101,8 @@ class LoggingStatLogger(StatLoggerBase):
     def __init__(self, vllm_config: VllmConfig, engine_index: int = 0):
         self.engine_index = engine_index
         self.vllm_config = vllm_config
+        device_type = vllm_config.device_config.device_type
+        self.kv_cache_device = "GPU" if device_type == "cuda" else device_type.upper()
         self._reset(time.monotonic())
 
         self.last_scheduler_stats = SchedulerStats()
@@ -184,7 +186,7 @@ class LoggingStatLogger(StatLoggerBase):
             "%sIteration(%d): %d context requests, %d context tokens, "
             "%d generation requests, %d generation tokens, "
             "iteration elapsed time: %.2f ms%s, "
-            "GPU KV cache usage: %.1f%%%s",
+            "%s KV cache usage: %.1f%%%s",
             self._log_prefix_for_engine(engine_idx),
             details.iteration_index,
             details.num_ctx_requests,
@@ -193,6 +195,7 @@ class LoggingStatLogger(StatLoggerBase):
             details.num_generation_tokens,
             details.elapsed_ms,
             " (dummy)" if details.is_dummy else "",
+            self.kv_cache_device,
             scheduler_stats.kv_cache_usage * 100,
             encoder_msg,
         )
@@ -287,12 +290,13 @@ class LoggingStatLogger(StatLoggerBase):
 
         log_parts.extend(
             [
-                "GPU KV cache usage: %.1f%%",
+                "%s KV cache usage: %.1f%%",
                 "Prefix cache hit rate: %.1f%%",
             ]
         )
         log_args.extend(
             [
+                self.kv_cache_device,
                 self.last_scheduler_stats.kv_cache_usage * 100,
                 self.prefix_caching_metrics.hit_rate * 100,
             ]
diff --git a/vllm/v1/metrics/perf.py b/vllm/v1/metrics/perf.py
index ab90ffcbe1..ebb676e66d 100644
--- a/vllm/v1/metrics/perf.py
+++ b/vllm/v1/metrics/perf.py
@@ -1749,7 +1749,7 @@ class PerfMetricsLogging:
             )
 
         log_fn(
-            "%sMFU: %.1f TF/s/GPU %.1f GB/s/GPU",
+            "%sMFU: %.1f TF/s/device %.1f GB/s/device",
             log_prefix,
             avg_tflops_per_gpu,
             avg_gbps_per_gpu,