[V1] Move usage stats to worker and start logging TPU hardware (#16211)

vllm/usage/usage_lib.py

@@ -174,6 +174,15 @@ class UsageMessage:
                 cuda_get_device_properties(0, ("name", "total_memory")))
         if current_platform.is_cuda():
             self.cuda_runtime = torch.version.cuda
+        if current_platform.is_tpu():
+            try:
+                import torch_xla
+                self.gpu_count = torch_xla.runtime.world_size()
+                self.gpu_type = torch_xla.tpu.get_tpu_type()
+                self.gpu_memory_per_device = (
+                    torch_xla.core.xla_model.get_memory_info()["bytes_limit"])
+            except Exception:
+                pass
         self.provider = _detect_cloud_provider()
         self.architecture = platform.machine()
         self.platform = platform.platform()

vllm/v1/engine/async_llm.py

@@ -36,7 +36,6 @@ from vllm.v1.executor.abstract import Executor
 from vllm.v1.metrics.loggers import (LoggingStatLogger, PrometheusStatLogger,
                                      StatLoggerBase)
 from vllm.v1.metrics.stats import IterationStats, SchedulerStats
-from vllm.v1.utils import report_usage_stats
 
 logger = init_logger(__name__)
 
@@ -113,9 +112,6 @@ class AsyncLLM(EngineClient):
         except RuntimeError:
             pass
 
-        # If usage stat is enabled, collect relevant info.
-        report_usage_stats(vllm_config, usage_context)
-
     @classmethod
     def from_vllm_config(
         cls,

vllm/v1/engine/llm_engine.py

@@ -28,7 +28,6 @@ from vllm.v1.engine.output_processor import OutputProcessor
 from vllm.v1.engine.parallel_sampling import ParentRequest
 from vllm.v1.engine.processor import Processor
 from vllm.v1.executor.abstract import Executor
-from vllm.v1.utils import report_usage_stats
 
 logger = init_logger(__name__)
 
@@ -97,9 +96,6 @@ class LLMEngine:
             # for v0 compatibility
             self.model_executor = self.engine_core.engine_core.model_executor  # type: ignore
 
-        # If usage stat is enabled, collect relevant info.
-        report_usage_stats(vllm_config, usage_context)
-
     @classmethod
     def from_vllm_config(
         cls,

vllm/v1/utils.py

@@ -205,7 +205,9 @@ def copy_slice(from_tensor: torch.Tensor, to_tensor: torch.Tensor,
     return to_tensor[:length].copy_(from_tensor[:length], non_blocking=True)
 
 
-def report_usage_stats(vllm_config, usage_context: UsageContext) -> None:
+def report_usage_stats(
+        vllm_config,
+        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT) -> None:
     """Report usage statistics if enabled."""
 
     if not is_usage_stats_enabled():

vllm/v1/worker/gpu_worker.py

@@ -23,6 +23,7 @@ from vllm.platforms import current_platform
 from vllm.utils import GiB_bytes
 from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
 from vllm.v1.outputs import ModelRunnerOutput
+from vllm.v1.utils import report_usage_stats
 from vllm.v1.worker.gpu_model_runner import GPUModelRunner
 from vllm.v1.worker.worker_base import WorkerBase
 
@@ -141,6 +142,10 @@ class Worker(WorkerBase):
         self.model_runner: GPUModelRunner = GPUModelRunner(
             self.vllm_config, self.device)
 
+        if self.rank == 0:
+            # If usage stat is enabled, collect relevant info.
+            report_usage_stats(self.vllm_config)
+
     # FIXME(youkaichao & ywang96): Use TorchDispatchMode instead of memory pool
     # to hijack tensor allocation.
     def load_model(self) -> None:

vllm/v1/worker/tpu_worker.py

@@ -21,7 +21,7 @@ from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.kv_cache_interface import (AttentionSpec, KVCacheConfig,
                                         KVCacheSpec)
 from vllm.v1.outputs import ModelRunnerOutput
-from vllm.v1.utils import bind_kv_cache
+from vllm.v1.utils import bind_kv_cache, report_usage_stats
 from vllm.v1.worker.tpu_model_runner import TPUModelRunner
 
 logger = init_logger(__name__)
@@ -133,6 +133,10 @@ class TPUWorker:
         # Init ModelRunner here, so that we have access to self.device.
         self.model_runner = TPUModelRunner(self.vllm_config, self.device)
 
+        if rank == 0:
+            # If usage stat is enabled, collect relevant info.
+            report_usage_stats(self.vllm_config)
+
     def determine_available_memory(self) -> int:
         kv_caches: dict[str, torch.Tensor] = {}
         kv_cache_spec = self.model_runner.get_kv_cache_spec()