TreePerfAnalyzer.from_file() always crashes when capture_trace_filepath is passed

[mgehre-amd]

Summary

TreePerfAnalyzer.from_file(..., capture_trace_filepath=...) — the documented SDK entry point for graph-replay + graph-capture inference traces — raises KeyError: t_end unconditionally. There is no value of add_python_func that makes it work.

Repro

Repo at 210fb5c8, using bundled traces:

from TraceLens.TreePerf import TreePerfAnalyzer

TreePerfAnalyzer.from_file(
    "tests/traces/inference/xdit_flux.1/profile_trace_rank_0.json.gz",
    capture_trace_filepath="tests/traces/inference/xdit_flux.1/capture_traces",
)

Traceback (most recent call last):
  File "TraceLens/TreePerf/tree_perf.py", line 240, in from_file
    return TreePerfAnalyzer(
  File "TraceLens/TreePerf/tree_perf.py", line 288, in __init__
    self.tree.build_tree(add_python_func=add_python_func)
  File "TraceLens/Trace2Tree/trace_to_tree.py", line 149, in build_host_call_stack_tree
    >= stack[-1][TraceEventUtils.TraceKeys.TimeEnd]
KeyError: <TraceKeys.TimeEnd: 't_end'>

The capture merge itself succeeds — alignment reports matched 993 capture events to 993 graph kernels before the crash. Reproduces identically with add_python_func=True and add_python_func=False, and on both xdit_flux.1 and xdit_flux.1_aiter_sage.

Cause

from_file (tree_perf.py:230-247) replaces the tree with the output of merge_capture_trace_into_graph, which returns an already-built tree. It then passes that tree to TreePerfAnalyzer.__init__, which re-runs build_tree because rebuild_tree defaults to True (tree_perf.py:288).

The merged tree contains events with no t_end, which build_host_call_stack_tree dereferences unconditionally:

events missing t_end: 17133
  ac2g          17063
  None             62
  cuda_runtime      8

Workaround

TreePerfAnalyzer.from_file(..., capture_trace_filepath=..., rebuild_tree=False)

Verified: yields a usable tree with 11,076 kernel events and correct device totals.

Suggested fix

Have from_file force rebuild_tree=False on the capture-merge path — rebuilding an already-merged tree is wrong independently of the crash.

Why CI misses this

tests/test_inference_perf_report.py calls merge_capture_trace_into_graph directly and passes the resulting tree onward, so the from_file convenience path has no coverage. Worth adding a smoke test for it.