Add vllm viewer

[angelayi]

This PR adds a dedicated vllm-compile tlparse viewer which looks like (internal example):

When we encouter vllm compilation artifacts (prefixed with vllm_), we use the vllm parsers located under src/vllm. We will also choose to display a different index.html.

In vllm, we will have multiple layers of compilation. First based on dynamic shapes (shape ranges, and then specialized sizes), and then if using piecewise cudagraphs then it will compile subgraphs of the toplevel model. So we want to have a hierarchical view like

Inductor Compilation                                         
  ├── compile range [1, 16384]                                         
  │   ├── submod_0                                                                   
  │   │   └── inductor artifacts ...                     
  │   └── submod_2                                                              
  │       └── inductor artifacts ...                     
  └── size 8                                                   
      ├── submod_0                                             
      │   └── inductor artifacts ...                     
      └── submod_2  

To do this, we create a VllmState to pass around which tracks the artifacts as we parse through. In the vllm logs added in vllm-project/vllm#33213, vllm will emit a vllm_piecewise_compile_start log at the start of each subgraph compilation, containing the submod name and compile range/size. As we parse subsequent inductor artifacts (post_grad_graph, output_code, etc.), we attach them to the current subgraph in VllmState. When we encounter the next vllm_piecewise_compile_start, we know the previous subgraph is complete and start collecting artifacts for the new one.

[yushangdi]

maybe I'm missing something. why this code will only run vllm artifacts? do we need to check vllm file prefix and only add vllm artifacts to vllm_state?

[angelayi]

updated! I removed most of that logic since it was unnecessary, and moved the rest into the vllm_state.add_artifact function

[yushangdi]

since this override the normal html, can we add a more detailed description of when would vllm_state.has_artifacts() = True? e.g. it's true when there exist at least a file that starts with vllm

[angelayi]

updated the comment!