Probe LoRA Qwen3-8B CUDA fail on plain main (negative control, NOT a fix)

[fzyzcjy]

🤖 Opened autonomously by Claude Code acting on Tom's behalf. This is the negative-control half of a paired diagnostic experiment (sibling: #25743 which reverts #25690). The @-mentions below are programmatic, not Tom's personal request; please push back if any conclusion is off.

This PR does not revert anything. It only adds a one-line sentinel comment to python/sglang/version.py so the GitHub Actions paths-filter triggers and the PR isn't auto-closed for a zero-diff.

Purpose: confirm that the LoRA Qwen3-8B CUDA-graph illegal-address regression bisected to #25690 is reproducible on plain upstream/main with no other changes — i.e., the bug is not somehow specific to Tom's #25703–#25728 scheduler refactor chain on the main-CI sandbox (#25647).

Expected result: /rerun-test test/registered/lora/test_lora_qwen3_8b_logprob_diff.py here should FAIL with the same CUDBG_EXCEPTION_WARP_ILLEGAL_ADDRESS (14) fingerprint, while the same /rerun-test on the sibling revert PR #25743 should PASS — together that's bidirectional evidence for b79e4b1e68 as the root cause.

Bisect evidence: #25647 (comment).

cc @Fridge003 @hnyls2002

---

CI States

Latest PR Test (Base): ❌ Run #26077811447
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[fzyzcjy]

/rerun-test test/registered/lora/test_lora_qwen3_8b_logprob_diff.py

[github-actions]

🚀 1-gpu-h100 (1 test): ❌ View workflow run

cd test/ && python3 registered/lora/test_lora_qwen3_8b_logprob_diff.py

[fzyzcjy]

🤖 Posted autonomously by Claude Code acting on Tom's behalf. Result of the diagnostic /rerun-test on this no-revert negative-control PR.

Result: FAIL ❌ (expected) — bug reproduces on plain upstream/main

/rerun-test test/registered/lora/test_lora_qwen3_8b_logprob_diff.py run: 26077826917 — failure. Same fingerprint as the original extra-a-test-1-gpu-large (0) failure on #25647:

coredump: Detected an exception of type CUDBG_EXCEPTION_WARP_ILLEGAL_ADDRESS (14)
Fatal Python error: Aborted
RuntimeError: Rank 0 scheduler died during initialization (exit code: -6).

Pair-PR table:

PR	Branch	/rerun-test verdict
#25743 (revert of #25690)	tom/revert-25690-cutedsl	PASS ✅
#25744 (this — no-revert, plain main + 1-line touch)	tom/probe-lora-bug-25690	FAIL ❌

Bidirectional confirmation that b79e4b1e68 (#25690) is the root cause — the bug is on plain upstream/main and is not specific to Tom's #25703–#25728 refactor chain context.

Closing this PR — diagnostic only, no merge value. Full bisect + repro: #25647 (comment).

cc @Fridge003 @hnyls2002

[fzyzcjy]

🤖 Posted autonomously by Claude Code acting on Tom's behalf. Second-run double-confirmation on the no-revert (negative-control) branch.

Second run: FAIL ❌ — reproducibly broken on plain main

rerun-test run 2 on tom/probe-lora-bug-25690: 26078647279 — failure, same CUDBG_EXCEPTION_WARP_ILLEGAL_ADDRESS (14) fingerprint.

Run	Result
#1	FAIL ❌ (same fingerprint)
#2	FAIL ❌ (same fingerprint)

Two-in-two-out for the no-revert branch on plain upstream/main. No flake risk.