PT2 + TritonParse + AMD GPU can't genera meaningful launch events

[FindHao]

The launch event doesn't have arguments and compilation metadata

[FindHao]

I'm @FindHao's AI agent, posting this update on their behalf.

This appears largely addressed on current main (0c6075b), particularly the Inductor launch-hook / compilation-metadata path. The arguments part still needs explicit regression coverage before we consider the issue fully resolved.

Evidence:

• 7cd8040 — Propagate launch tracing to Inductor's JIT hook gate makes init(enable_trace_launch=True) update both TORCHINDUCTOR_RUN_JIT_POST_COMPILE_HOOK and the live torch._inductor.config.run_jit_post_compile_hook setting. This fixes the missing Inductor hook activation after torch has already been imported.
• The existing test_launches_carry_compilation_metadata regression test starts with the hook gate disabled, then verifies that launches are traced and at least one carries a compilation_metadata dictionary.
• That test passed in the September 7 AMD MI350 / ROCm CI job, rather than being skipped. The log reports PyTorch 2.14.0.dev20260804+rocm7.1, Triton 3.8.0, and a successful 976-test run.

The current test only checks that compilation_metadata is a dictionary; it does not validate its contents or assert anything about extracted_args. Consequently, the passing CI job alone does not establish that the original missing-arguments symptom is resolved.

I recommend extending that test, or adding a focused AMD Inductor test, to check a small torch.compile workload for:

1. A nonempty compilation_metadata.hash that matches a captured compilation event.
2. Nonempty extracted_args, including the expected shape and dtype of representative tensor arguments.
3. Execution of these assertions in the existing ROCm CI job, while retaining the cleared-hook setup to catch activation regressions.

Once that coverage passes on AMD, we should have the evidence needed to close this issue.