[None][feat] Add NvTelemetry/GXT-compliant usage telemetry

.github/tava_architecture_diagram.md

@@ -91,6 +91,16 @@ graph TB
         BatchManager --> KVCache
     end
 
+    subgraph "Usage_Telemetry"
+        ReportUsage[report_usage]
+        BgReporter[Background Reporter]
+        GxtPayload[GXT Payload Builder]
+        GxtEndpoint[NvTelemetry Endpoint]
+        ReportUsage --> BgReporter
+        BgReporter --> GxtPayload
+        GxtPayload --> GxtEndpoint
+    end
+
     subgraph "Output_Results"
         Tokens[Generated Tokens]
         Stats[Performance Stats]
@@ -99,13 +109,17 @@ graph TB
         GenVideos[Generated Videos]
     end
 
+    LLMAPI --> ReportUsage
+
     PyTorch_Flow ~~~ TensorRT_Flow
 
     TensorRT_Flow --> Output_Results
     PyTorch_Flow --> Output_Results
     AutoDeploy_Flow --> Output_Results
     Visual_Gen_Flow --> Output_Results
 
+    AutoDeploy_Flow ~~~ Usage_Telemetry
+
     %% Force Output_Results to be between PyTorch_flow and TensorRT_flow
     PyTorch_Flow ~~~ Output_Results
 
@@ -141,6 +155,10 @@ graph TB
     classDef api fill:#bfb,stroke:#333,stroke-width:2px;
     class PythonAPI,CppAPI,LLMAPI api;
 
+    %% Telemetry format
+    classDef telemetry fill:#cef,stroke:#333,stroke-width:2px;
+    class ReportUsage,BgReporter,GxtPayload,GxtEndpoint telemetry;
+
     %% Results format
     classDef result fill:#fbb,stroke:#333,stroke-width:2px;
     class Tokens,Stats,Metrics,GenImages,GenVideos result;

README.md

@@ -284,6 +284,41 @@ Deprecation is used to inform developers that some APIs and tools are no longer
 4. Removal After Migration Period
   - After the 3-month migration period ends, deprecated APIs, tools, or parameters are removed in a manner consistent with semantic versioning (major version changes may include breaking removals).
 
+## Telemetry Data Collection
+
+TensorRT-LLM collects anonymous telemetry data by default. This data is used
+in aggregate to understand usage patterns and prioritize engineering efforts.
+**This data cannot be traced back to any individual user.** No prompts,
+user-identifying information, or persistent identifiers are collected. Any
+deployment identifiers are ephemeral, randomly generated per deployment, and
+not linked to users. The data we collect includes:
+
+- Ingress point (e.g., LLM API, CLI, serve command)
+- Deployment duration (via periodic heartbeats)
+- GPU SKUs, count, memory, and CUDA version
+- Model architecture class name (e.g., `LlamaForCausalLM`)
+- Parallelism configuration (TP/PP/CP/MoE-EP/MoE-TP sizes), quantization algorithm, dtype, KV cache dtype
+- System information (OS platform, Python version, CPU architecture, CPU count)
+- TRT-LLM version and backend
+- Feature flags (LoRA, speculative decoding, prefix caching, CUDA graphs, chunked context, data parallelism)
+- Disaggregated serving metadata (role and deployment ID)
+
+Telemetry is automatically disabled in CI and test environments.
+
+### Opting Out of Telemetry Data Collection
+
+To disable telemetry data collection, use any of the following methods:
+
+- **Environment variable**: Set `TRTLLM_NO_USAGE_STATS=1`, `DO_NOT_TRACK=1`, or `TELEMETRY_DISABLED=true`
+- **File-based**: Create the file `~/.config/trtllm/do_not_track`
+- **Python API**: Pass `TelemetryConfig(disabled=True)` to `LLM()`
+- **CLI flag**: Use `--no-telemetry` on `trtllm-serve`, `trtllm-bench`, or `trtllm-eval`
+
+The telemetry collection code is fully open source and auditable at
+[`tensorrt_llm/usage/`](./tensorrt_llm/usage/). For a detailed field-by-field
+reference of exactly what is collected, see the
+[schema documentation](./tensorrt_llm/usage/schemas/README.md).
+
 ## Useful Links
 - [Quantized models on Hugging Face](https://huggingface.co/collections/nvidia/model-optimizer-66aa84f7966b3150262481a4): A growing collection of quantized (e.g., FP8, FP4) and optimized LLMs, including [DeepSeek FP4](https://huggingface.co/nvidia/DeepSeek-R1-FP4), ready for fast inference with TensorRT LLM.
 - [NVIDIA Dynamo](https://github.com/ai-dynamo/dynamo): A datacenter scale distributed inference serving framework that works seamlessly with TensorRT LLM.

setup.py

@@ -170,7 +170,8 @@ def has_ext_modules(self):
     "_torch/auto_deploy/config/*.yaml",
     # Include CUDA source for fused MoE align extension so runtime JIT can find it in wheels
     '_torch/auto_deploy/custom_ops/fused_moe/moe_align_kernel.cu',
-    '_torch/auto_deploy/custom_ops/fused_moe/triton_fused_moe_configs/*'
+    '_torch/auto_deploy/custom_ops/fused_moe/triton_fused_moe_configs/*',
+    'usage/schemas/*.json',
 ]
 
 

tensorrt_llm/bench/benchmark/low_latency.py

@@ -292,6 +292,8 @@ def latency_command(
     llm = None
     kwargs = kwargs | runtime_config.get_llm_args()
     kwargs['backend'] = options.backend
+    if bench_env.telemetry_config is not None:
+        kwargs["telemetry_config"] = bench_env.telemetry_config
 
     # Set environment variables for setting runtime options.
     default_env_overrides = {

tensorrt_llm/bench/benchmark/throughput.py

@@ -428,6 +428,8 @@ def throughput_command(
         kwargs = kwargs | runtime_config.get_llm_args()
         kwargs['skip_tokenizer_init'] = not no_skip_tokenizer_init
         kwargs['backend'] = options.backend
+        if bench_env.telemetry_config is not None:
+            kwargs["telemetry_config"] = bench_env.telemetry_config
 
         llm = get_llm(runtime_config, kwargs)
 

tensorrt_llm/bench/build/build.py

@@ -332,7 +332,8 @@ def build_command(
               quant_config=quant_config,
               workspace=str(bench_env.workspace),
               load_format=load_format,
-              trust_remote_code=trust_remote_code)
+              trust_remote_code=trust_remote_code,
+              telemetry_config=bench_env.telemetry_config)
     # Save the engine.
     llm.save(engine_dir)
     llm.shutdown()

tensorrt_llm/bench/dataclasses/general.py

@@ -15,6 +15,7 @@ class BenchmarkEnvironment(BaseModel):
     checkpoint_path: Optional[Path]
     workspace: Path
     revision: Optional[str] = None
+    telemetry_config: Optional[Any] = None
 
 
 class InferenceRequest(BaseModel):

tensorrt_llm/commands/bench.py

@@ -10,6 +10,7 @@
 from tensorrt_llm.bench.dataclasses.general import BenchmarkEnvironment
 from tensorrt_llm.bench.dataset.prepare_dataset import prepare_dataset
 from tensorrt_llm.logger import logger, severity_map
+from tensorrt_llm.usage import config as _telemetry_config
 
 
 class NotRequiredForHelp(click.Option):
@@ -56,6 +57,9 @@ def handle_parse_result(self, ctx, opts, args):
               default=None,
               help="The revision to use for the HuggingFace model "
               "(branch name, tag name, or commit id).")
+@click.option("--telemetry/--no-telemetry",
+              default=True,
+              help="Enable or disable anonymous usage telemetry collection.")
 @click.pass_context
 def main(
     ctx,
@@ -64,15 +68,21 @@ def main(
     workspace: Path,
     log_level: str,
     revision: Optional[str],
+    telemetry: bool,
 ) -> None:
     logger.set_level(log_level)
     if model is None:
         return
 
-    ctx.obj = BenchmarkEnvironment(model=model,
-                                   checkpoint_path=model_path,
-                                   workspace=workspace,
-                                   revision=revision)
+    ctx.obj = BenchmarkEnvironment(
+        model=model,
+        checkpoint_path=model_path,
+        workspace=workspace,
+        revision=revision,
+        telemetry_config=_telemetry_config.TelemetryConfig(
+            disabled=not telemetry,
+            usage_context=_telemetry_config.UsageContext.CLI_BENCH),
+    )
 
     # Create the workspace where we plan to store intermediate files.
     ctx.obj.workspace.mkdir(parents=True, exist_ok=True)

tensorrt_llm/commands/eval.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -26,6 +26,7 @@
 from ..llmapi import BuildConfig, KvCacheConfig
 from ..llmapi.llm_utils import update_llm_args_with_extra_options
 from ..logger import logger, severity_map
+from ..usage import config as _telemetry_config
 
 
 @click.group()
@@ -117,31 +118,49 @@
               is_flag=True,
               default=False,
               help="Flag for disabling KV cache reuse.")
+@click.option("--telemetry/--no-telemetry",
+              default=True,
+              help="Enable or disable anonymous usage telemetry collection.")
 @click.pass_context
 def main(ctx, model: str, tokenizer: Optional[str],
          custom_tokenizer: Optional[str], log_level: str, backend: str,
          max_beam_width: int, max_batch_size: int, max_num_tokens: int,
          max_seq_len: int, tp_size: int, pp_size: int, ep_size: Optional[int],
          gpus_per_node: Optional[int], kv_cache_free_gpu_memory_fraction: float,
          trust_remote_code: bool, revision: Optional[str],
-         extra_llm_api_options: Optional[str], disable_kv_cache_reuse: bool):
+         extra_llm_api_options: Optional[str], disable_kv_cache_reuse: bool,
+         telemetry: bool):
     logger.set_level(log_level)
 
     kv_cache_config = KvCacheConfig(
         free_gpu_memory_fraction=kv_cache_free_gpu_memory_fraction,
         enable_block_reuse=not disable_kv_cache_reuse)
 
     llm_args = {
-        "model": model,
-        "tokenizer": tokenizer,
-        "custom_tokenizer": custom_tokenizer,
-        "tensor_parallel_size": tp_size,
-        "pipeline_parallel_size": pp_size,
-        "moe_expert_parallel_size": ep_size,
-        "gpus_per_node": gpus_per_node,
-        "trust_remote_code": trust_remote_code,
-        "revision": revision,
-        "kv_cache_config": kv_cache_config,
+        "model":
+        model,
+        "tokenizer":
+        tokenizer,
+        "custom_tokenizer":
+        custom_tokenizer,
+        "tensor_parallel_size":
+        tp_size,
+        "pipeline_parallel_size":
+        pp_size,
+        "moe_expert_parallel_size":
+        ep_size,
+        "gpus_per_node":
+        gpus_per_node,
+        "trust_remote_code":
+        trust_remote_code,
+        "revision":
+        revision,
+        "kv_cache_config":
+        kv_cache_config,
+        "telemetry_config":
+        _telemetry_config.TelemetryConfig(
+            disabled=not telemetry,
+            usage_context=_telemetry_config.UsageContext.CLI_EVAL),
     }
 
     if backend == 'pytorch':
@@ -166,6 +185,11 @@ def main(ctx, model: str, tokenizer: Optional[str],
         llm_args = update_llm_args_with_extra_options(llm_args,
                                                       extra_llm_api_options)
 
+    # CLI --no-telemetry always wins over YAML config
+    if not telemetry:
+        llm_args["telemetry_config"] = llm_args["telemetry_config"].model_copy(
+            update={"disabled": True})
+
     profiler.start("trtllm init")
     llm = llm_cls(**llm_args)
     profiler.stop("trtllm init")