[ci, test] chore: Split L0 converter launch script into conversion and generation

[cuichenx]

Summary

• Split L0_Launch_converter.sh into two scripts so the heavier generation tests run as a separate CI matrix entry from the lighter import/export and multi-GPU conversion tests, halving per-job wall clock for either side.
• Drop an existing duplication: test_hf_fsdp_conversion.py was running both in the catch-all converter script and in the dedicated L0_Launch_converter_fsdp.sh. The split removes the duplicate run.
• No workflow changes — cicd-main.yml's generate-test-matrix scans h100/active/L0_*.sh dynamically, so the new L0_Launch_converter_generate.sh is picked up automatically.

Script	Tests covered
L0_Launch_converter.sh (modified)	test_checkpoint_conversion.py + test_multi_gpu_conversion.py
L0_Launch_converter_generate.sh (new)	test_generate_from_hf.py + test_generate_vlm_from_hf.py
L0_Launch_converter_fsdp.sh (unchanged)	test_hf_fsdp_conversion.py

Background

This was triggered by a CI timeout investigation on the converter L0 job. Locally on H100 with a populated HF cache, the full 10-test set passes cleanly (PP test = 47s call + 11s setup), so the CI timeout looks environmental (cold HF cache, hf_xet network behavior, runner contention) rather than a logic bug. Splitting the job gives each half its own timeout budget and isolates the heavier distributed-launch tests.

Test plan

• CI: L0_Launch_converter passes (covers test_checkpoint_conversion + test_multi_gpu_conversion)
• CI: L0_Launch_converter_generate passes (covers test_generate_from_hf + test_generate_vlm_from_hf)
• CI: L0_Launch_converter_fsdp still passes (unchanged)
• Verify the matrix generator picks up the new script (it scans h100/active/L0_*.sh automatically)

Follow-up (not in this PR)

• Same split could be applied symmetrically to tests/functional_tests/launch_scripts/gb200/active/L0_Launch_converter.sh for consistency.

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[cuichenx]

/ok to test 0f129b2

[claude]

LGTM — clean split. All five converter test files are covered across the three launch scripts (L0_Launch_converter.sh, L0_Launch_converter_generate.sh, L0_Launch_converter_fsdp.sh), the new file is executable, and the duplicate test_hf_fsdp_conversion.py run is removed.Suggested test casesNo perf tests impacted.CI validation:- L0_Launch_converter — should pass covering test_checkpoint_conversion.py + test_multi_gpu_conversion.py- L0_Launch_converter_generate — should pass covering test_generate_from_hf.py + test_generate_vlm_from_hf.py- L0_Launch_converter_fsdp — should pass unchanged covering test_hf_fsdp_conversion.py- Verify CI matrix generator picks up L0_Launch_converter_generate.sh automatically (no workflow changes needed)