Add GLM5 SFT support

[samaritan1998]

Summary

• Add a GLM5-specific SFT loss mask type that follows GLM-style stop markers.
• Append the correct GLM5 final stop marker for assistant turns: <|user|> for normal replies and <|observation|> for final tool calls.
• Add a GLM5 SFT launch script using sft_loss and the existing GLM5 Megatron model config.
• Add unit coverage for multi-turn GLM5 masking, tool calls, final tool-call stop markers, and step_loss_mask handling.
• Apply the repository pre-commit formatting/lint fixes required by the all-files pre-commit CI job.

Details

The GLM5 mask generation renders the full chat template, supervises assistant spans, and treats <|user|>, <|observation|>, and <|endoftext|> as stop boundaries. This mirrors GLM5 chat-template behavior: regular assistant replies learn the <|user|> boundary, while assistant tool calls learn the <|observation|> boundary so the model stops before tool results are appended externally.

Validation

• /tmp/slime-precommit-venv/bin/pre-commit run --all-files --show-diff-on-failure --color=always
• Ran GLM5 mask tests via direct Python invocation with a lightweight transformers stub, because this local environment does not have pytest or transformers installed.
• bash -n scripts/run-glm5-744B-A40B-sft.sh
• git diff --check

Could not run python -m pytest directly because pytest is not installed in this local environment.

[samaritan1998]

Local validation passed for GLM5 loss mask behavior and script syntax.

[stevewx]

Hi samaritan1998, this might be a duplicate of #1824

[samaritan1998]

Hi samaritan1998, this might be a duplicate of #1824
我看到了 #1824 但是#1824的实现是不是有bug 没有学eos token(<|user|>) 会导致模型不停止

[stevewx]

ahh yes you're right, closed #1824 in favor of this PR