Expose PyTorch profiler configuration to environment variables (#21803)

Signed-off-by: Csrayz <33659823+Csrayz@users.noreply.github.com>

docs/contributing/profiling.md

@@ -5,7 +5,12 @@
 
 ## Profile with PyTorch Profiler
 
-We support tracing vLLM workers using the `torch.profiler` module. You can enable tracing by setting the `VLLM_TORCH_PROFILER_DIR` environment variable to the directory where you want to save the traces: `VLLM_TORCH_PROFILER_DIR=/mnt/traces/`
+We support tracing vLLM workers using the `torch.profiler` module. You can enable tracing by setting the `VLLM_TORCH_PROFILER_DIR` environment variable to the directory where you want to save the traces: `VLLM_TORCH_PROFILER_DIR=/mnt/traces/`. Additionally, you can control the profiling content by specifying the following environment variables:
+
+- `VLLM_TORCH_PROFILER_RECORD_SHAPES=1` to enable recording Tensor Shapes, off by default
+- `VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY=1` to record memory, off by default
+- `VLLM_TORCH_PROFILER_WITH_STACK=1` to enable recording stack information, on by default
+- `VLLM_TORCH_PROFILER_WITH_FLOPS=1` to enable recording FLOPs, off by default
 
 The OpenAI server also needs to be started with the `VLLM_TORCH_PROFILER_DIR` environment variable set.
 

vllm/envs.py

@@ -80,6 +80,10 @@ if TYPE_CHECKING:
     VLLM_PLUGINS: Optional[list[str]] = None
     VLLM_LORA_RESOLVER_CACHE_DIR: Optional[str] = None
     VLLM_TORCH_PROFILER_DIR: Optional[str] = None
+    VLLM_TORCH_PROFILER_RECORD_SHAPES: bool = False
+    VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: bool = False
+    VLLM_TORCH_PROFILER_WITH_STACK: bool = True
+    VLLM_TORCH_PROFILER_WITH_FLOPS: bool = False
     VLLM_USE_TRITON_AWQ: bool = False
     VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
     VLLM_SKIP_P2P_CHECK: bool = False
@@ -629,6 +633,31 @@ environment_variables: dict[str, Callable[[], Any]] = {
     lambda: (None if os.getenv("VLLM_TORCH_PROFILER_DIR", None) is None else os
              .path.expanduser(os.getenv("VLLM_TORCH_PROFILER_DIR", "."))),
 
+    # Enable torch profiler to record shapes if set
+    # VLLM_TORCH_PROFILER_RECORD_SHAPES=1. If not set, torch profiler will
+    # not record shapes.
+    "VLLM_TORCH_PROFILER_RECORD_SHAPES":
+    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES", "0") != "0"),
+
+    # Enable torch profiler to profile memory if set
+    # VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY=1. If not set, torch profiler
+    # will not profile memory.
+    "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY":
+    lambda: bool(
+        os.getenv("VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY", "0") != "0"),
+
+    # Enable torch profiler to profile stack if set
+    # VLLM_TORCH_PROFILER_WITH_STACK=1. If not set, torch profiler WILL
+    # profile stack by default.
+    "VLLM_TORCH_PROFILER_WITH_STACK":
+    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_WITH_STACK", "1") != "0"),
+
+    # Enable torch profiler to profile flops if set
+    # VLLM_TORCH_PROFILER_WITH_FLOPS=1. If not set, torch profiler will
+    # not profile flops.
+    "VLLM_TORCH_PROFILER_WITH_FLOPS":
+    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_WITH_FLOPS", "0") != "0"),
+
     # If set, vLLM will use Triton implementations of AWQ.
     "VLLM_USE_TRITON_AWQ":
     lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),

vllm/v1/worker/gpu_worker.py

@@ -71,12 +71,23 @@ class Worker(WorkerBase):
             torch_profiler_trace_dir = envs.VLLM_TORCH_PROFILER_DIR
             logger.info("Profiling enabled. Traces will be saved to: %s",
                         torch_profiler_trace_dir)
+            logger.debug(
+                "Profiler config: record_shapes=%s,"
+                "profile_memory=%s,with_stack=%s,with_flops=%s",
+                envs.VLLM_TORCH_PROFILER_RECORD_SHAPES,
+                envs.VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY,
+                envs.VLLM_TORCH_PROFILER_WITH_STACK,
+                envs.VLLM_TORCH_PROFILER_WITH_FLOPS,
+            )
             self.profiler = torch.profiler.profile(
                 activities=[
                     torch.profiler.ProfilerActivity.CPU,
                     torch.profiler.ProfilerActivity.CUDA,
                 ],
-                with_stack=True,
+                record_shapes=envs.VLLM_TORCH_PROFILER_RECORD_SHAPES,
+                profile_memory=envs.VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY,
+                with_stack=envs.VLLM_TORCH_PROFILER_WITH_STACK,
+                with_flops=envs.VLLM_TORCH_PROFILER_WITH_FLOPS,
                 on_trace_ready=torch.profiler.tensorboard_trace_handler(
                     torch_profiler_trace_dir, use_gzip=True))
         else:
@@ -209,7 +220,7 @@ class Worker(WorkerBase):
 
     @torch.inference_mode()
     def determine_available_memory(self) -> int:
-        """Profiles the peak memory usage of the model to determine how much 
+        """Profiles the peak memory usage of the model to determine how much
         memory can be used for KV cache without OOMs.
 
         The engine will first conduct a profiling of the existing memory usage.

vllm/v1/worker/xpu_worker.py

@@ -41,12 +41,23 @@ class XPUWorker(Worker):
             torch_profiler_trace_dir = envs.VLLM_TORCH_PROFILER_DIR
             logger.info("Profiling enabled. Traces will be saved to: %s",
                         torch_profiler_trace_dir)
+            logger.debug(
+                "Profiler config: record_shapes=%s,"
+                "profile_memory=%s,with_stack=%s,with_flops=%s",
+                envs.VLLM_TORCH_PROFILER_RECORD_SHAPES,
+                envs.VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY,
+                envs.VLLM_TORCH_PROFILER_WITH_STACK,
+                envs.VLLM_TORCH_PROFILER_WITH_FLOPS,
+            )
             self.profiler = torch.profiler.profile(
                 activities=[
                     torch.profiler.ProfilerActivity.CPU,
                     torch.profiler.ProfilerActivity.XPU,
                 ],
-                with_stack=True,
+                record_shapes=envs.VLLM_TORCH_PROFILER_RECORD_SHAPES,
+                profile_memory=envs.VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY,
+                with_stack=envs.VLLM_TORCH_PROFILER_WITH_STACK,
+                with_flops=envs.VLLM_TORCH_PROFILER_WITH_FLOPS,
                 on_trace_ready=torch.profiler.tensorboard_trace_handler(
                     torch_profiler_trace_dir, use_gzip=True))
         else: