A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 21:15 UTC
[UX] Bump version in CG memory profiling log message (#40465)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Apr 21, 2026 at 11:26 UTC
ab5666eb7cb5a98f2ff260523c9aa469da23a004
1 file changed
+2
-2
vllm/v1/worker/gpu_worker.py
+2
-2
@@ -456,7 +456,7 @@ class Worker(WorkerBase):
456
logger.info(
457
"CUDA graph memory profiling is enabled "
458
"(VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1). "
459
- "This will become the default in v0.19. "
459
+ "This will become the default in v0.21. "
460
"The current --gpu-memory-utilization=%.4f is equivalent "
461
"to --gpu-memory-utilization=%.4f without CUDA graph "
462
"memory profiling. To maintain the same effective KV "
@@ -472,7 +472,7 @@ class Worker(WorkerBase):
472
1.0,
473
)
474
logger.info(
475
- "In v0.19, CUDA graph memory profiling will be enabled "
475
+ "In v0.21, CUDA graph memory profiling will be enabled "
476
"by default (VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1), "
477
"which more accurately accounts for CUDA graph memory "
478
"during KV cache allocation. To try it now, set "