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 14, 05:30 UTC
[WideEP][P/D] Add usage stats for DP+EP and KV Connector (#26836)
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Tyler Michael Smith committed
Oct 14, 2025 at 19:51 UTC
579d2e5458b19c442f48e0cba0ba71c5d4abf6ea
1 file changed
+14
-1
vllm/v1/utils.py
+14
-1
@@ -345,13 +345,17 @@ def report_usage_stats(
345
346
parallel_config = vllm_config.parallel_config
347
348
+ # Prepare KV connector string if applicable
349
+ kv_connector = None
350
+ if vllm_config.kv_transfer_config is not None:
351
+ kv_connector = vllm_config.kv_transfer_config.kv_connector
352
+
353
usage_message.report_usage(
354
get_architecture_class_name(vllm_config.model_config),
355
usage_context,
356
extra_kvs={
357
# Common configuration
358
"dtype": str(vllm_config.model_config.dtype),
354
- "tensor_parallel_size": parallel_config.tensor_parallel_size,
359
"block_size": vllm_config.cache_config.block_size,
360
"gpu_memory_utilization": vllm_config.cache_config.gpu_memory_utilization,
361
"kv_cache_memory_bytes": vllm_config.cache_config.kv_cache_memory_bytes,
@@ -363,6 +367,15 @@ def report_usage_stats(
367
"enable_prefix_caching": vllm_config.cache_config.enable_prefix_caching,
368
"enforce_eager": vllm_config.model_config.enforce_eager,
369
"disable_custom_all_reduce": parallel_config.disable_custom_all_reduce,
370
+ # Distributed parallelism settings
371
+ "tensor_parallel_size": parallel_config.tensor_parallel_size,
372
+ "data_parallel_size": parallel_config.data_parallel_size,
373
+ "pipeline_parallel_size": parallel_config.pipeline_parallel_size,
374
+ "enable_expert_parallel": parallel_config.enable_expert_parallel,
375
+ # All2All backend for MoE expert parallel
376
+ "all2all_backend": parallel_config.all2all_backend,
377
+ # KV connector used
378
+ "kv_connector": kv_connector,
379
},
380
)
381