feat: render torch.profiler GPU profiles in the Flyte report

Add torch_profile() to flyteplugins-pytorch: a context manager (sync and
async) that runs torch.profiler over a region of a task body and renders
the result into a Flyte report tab — summary tiles, top-ops tables, and
an interactive Perfetto timeline (the gzipped chrome trace is base64-
embedded and pushed into a lazy ui.perfetto.dev iframe via its
postMessage deep-link API), plus a trace download button. Traces over an
embed cap are uploaded to blob storage and linked instead.

Rendering is best-effort and never fails the task; body exceptions are
never suppressed. Verified end-to-end on the demo cluster (L4 GPU).

Claude-Session: https://claude.ai/code/session_01FoFxbPpsYcWMny59J63BZM
Signed-off-by: Kevin Su <pingsutw@apache.org>

Author: pingsutw

plugins/pytorch/examples/profile_matmul.py

@@ -0,0 +1,55 @@
+"""GPU profiling with torch.profiler, rendered in the Flyte report.
+
+The task runs a few matmul+add steps under `torch_profile()`; the resulting report tab shows
+summary tables and an interactive Perfetto timeline (ProfilerStep#N / matmul_add / aten::mm
+spans on CPU and CUDA tracks) right in the Flyte UI.
+
+Run:
+    make dist && FLYTE_PLUGIN_DIST=plugins/pytorch make dist-plugins
+    flyte run plugins/pytorch/examples/profile_matmul.py profile_matmul
+"""
+
+import flyte
+
+from flyteplugins.pytorch import torch_profile
+
+image = (
+    flyte.Image.from_debian_base(name="torch-profile", python_version=(3, 12))
+    .with_pip_packages("torch")
+    .with_local_v2()
+    .with_local_v2_plugins(["flyteplugins-pytorch"])
+)
+
+env = flyte.TaskEnvironment(
+    name="torch-profile",
+    image=image,
+    resources=flyte.Resources(cpu="4", memory="16Gi", gpu="L4:1"),
+)
+
+
+@env.task(report=True)
+def profile_matmul(steps: int = 8, n: int = 4096) -> str:
+    import torch
+    from torch.profiler import record_function, schedule
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    a = torch.randn(n, n, device=device)
+    b = torch.randn(n, n, device=device)
+
+    # repeat=1: keep the completed profiling cycle — without it, stepping past the active
+    # window starts a new cycle and clears the collected events (empty report).
+    with torch_profile(profile_memory=True, schedule=schedule(wait=1, warmup=2, active=4, repeat=1)) as prof:
+        for _ in range(steps):
+            with record_function("matmul_add"):
+                c = a @ b + a
+            if device == "cuda":
+                torch.cuda.synchronize()
+            prof.step()
+
+    return f"done on {device}: {c.shape}"
+
+
+if __name__ == "__main__":
+    flyte.init_from_config()
+    run = flyte.run(profile_matmul)
+    print(run.url)

plugins/pytorch/src/flyteplugins/pytorch/__init__.py

@@ -1,3 +1,4 @@
-__all__ = ["Elastic"]
+__all__ = ["Elastic", "torch_profile"]
 
+from flyteplugins.pytorch._profile import torch_profile
 from flyteplugins.pytorch.task import Elastic