[Feature] Refactor examples/configs filename

[jiapingW]

Motivation

Standardize example configuration filenames by removing redundant online, offline, and disaggregated labels already represented by the directory structure.

Modifications

• Rename YAML recipes using -[-server-dp][-].yaml.
• Preserve topology variants and place hardware suffixes such as npu, amd, and h200 last.
• Update references in documentation, launch scripts, and tests.
• Update recipe discovery tests to identify configurations by relative path, allowing identical filenames in different directories.

Related Issues

N/A.

Accuracy Test

No model-side changes or accuracy impact expected.

Validation:

• 158 tests and 563 subtests passed.
• All renamed YAML configurations retain their original parsed values.
• Verified 41 referenced YAML paths and documentation links; no stale filename references remain.
• Updated shell scripts pass bash -n; pre-commit checks pass.

Benchmark & Profiling

Checklist

• [✅] Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.

[maocheng23]

Nit: now that the name only says -1server-dp4, could we add a short header comment noting how this differs from its sibling qwen3-8b-dflash.yaml? Besides 4 vs 8 trainer ranks, it also uses report_to: none, the default attention backend (no flex_attention), and drops embedding_key/torch_dtype/log_interval. For example:

# 4-rank trainer variant of qwen3-8b-dflash.yaml (1 external capture server).
# Also differs from the 8-rank recipe: report_to=none, default attention backend.

[maocheng23]

Same nit as the dflash dp4 file: a header comment would help. This one differs from qwen3-8b-domino.yaml by more than topology: it trains on perfectblend_qwen3-8b_regen.jsonl (vs ShareGPT), and also sets mask_token_id, trust_remote_code, report_to: none, and different save/log intervals. Someone choosing between the two by filename alone wouldn't expect a different dataset.

[maocheng23]

Nit: same here, consider a header comment. Compared with qwen3.6-27b-dflash.yaml, this uses max_length: 2048 (vs 4096) and save_interval: 1000, not just 2 vs 8 trainer ranks.

[maocheng23]

LGTM. I checked that no references to the 78 old filenames remain (also after merging current main), that no new path reuses an old one (so stale --config paths fail loudly instead of loading a different recipe), and that the <N>server-dp<M> suffixes match the configs (dp = trainer ranks). Keeping run_id/output_dir unchanged is a good call: it keeps W&B and output-dir continuity, and the run IDs are still unique.

Two small non-blocking nits are inline: one README sentence about when the topology suffix is used, and header comments on the three -1server-dpN variants that differ from their siblings by more than topology.

[maocheng23]

Nit: in practice the topology suffix only appears where two topologies of the same recipe exist side by side. Most recipes have none whatever their layout, e.g. external/qwen3-8b-dflash.yaml (1 server / 8 ranks), external/qwen3.8-27b-dflash2.yaml (8 server URLs / 8 ranks), managed-local/qwen3.5-4b-mtp-npu.yaml (6 servers / 10 ranks). Maybe add a sentence like: "The -<servers>server-dp<ranks> suffix is added only to tell topology variants of the same recipe apart; other recipes record their topology under deployment only." That way readers don't assume an unsuffixed name implies some particular topology.

[FrankLeeeee]

Can we merge this now?

[jiapingW]

Have updated description according to the comment.