add model_glm code

[zhipuch]

The full glm model code eliminates formatting and logic issues during inference using the inherited llama method.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[Isotr0py]

@zhipuch Could you elaborate about the logic issues?

Since the hf-format glm model is a llama-style one, I don't think we should copy most of code from llama.py, this will increase the difficulty of maintaining.

[zhipuch]

@zhipuch Could you elaborate about the logic issues?

Since the hf-format glm model is a llama-style one, I don't think we should copy most of code from llama.py, this will increase the difficulty of maintaining.

OK, I revisited the issue and found the main bug now：

1. layer.self_attn.rotary_emb.rotary_dim //= 2，lead to rotary_dim is 64, 32, 16, 8, 4, 2, 1, 0, 0, 0, .....
2. if layer.self_attn.rotary_emb.rotary_dim = self.config.head_dim / 2, since the get_rope function has already been initialized, this statement will simply change its parameters.
   So，I tried the following code and it worked. If you have better ideas, please share them with me：
   layer.self_attn.rotary_emb = get_rope(
   self.config.head_dim,
   rotary_dim=self.config.head_dim / 2,
   max_position=self.config.max_position_embeddings,
   base=self.config.rope_theta,
   rope_scaling=None,
   is_neox_style=False,
   )
   layer.self_attn.o_proj.bias = None
   layer.self_attn.o_proj.skip_bias_add = True

[Isotr0py]

Oh, my bad, rotary_dim is used to calculate inverse frequency and cos sin cache in initialization, I didn't realize this before. 😂

I‘m fine to use get_rope to re-initialize rotary_embedding if it's compatible with torch.compile.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @zhipuch.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork