Triton inference server support #26
Conversation
| from googleapiclient.discovery import build | ||
|
|
||
|
|
||
| class Tool: |
There was a problem hiding this comment.
needs an update function that can be called, retrieval and calendar will need this
There was a problem hiding this comment.
what should the update do?
There was a problem hiding this comment.
pass in the data, the update will use either the text, url, or etc to setup for the generation
| selected_start_tokens = probs[:, insert_api_at, start_tokens].argmax().item() | ||
|
|
||
| for i in range(m): | ||
| _, api_calls = await model( |
There was a problem hiding this comment.
need to add in the tool here before sending, e.g. [Calendar(
There was a problem hiding this comment.
hmm, shouldn't that already be in the annotation prompt?
| iter_data = islice(iter_data, max_datapoints) | ||
|
|
||
| async def sample_and_filter_api_calls(tool, text, top_k, n_gen): | ||
| async for tool_use in sample_api_calls( |
There was a problem hiding this comment.
Should chunk this, otherwise we're cutting off everything past 512? tokens
Needs a tool update function for current data
|
Cool with putting this into a dev branch and working on the stuff I found independently if that sounds good, or we can update this |
2 participants
This PR adds support for triton inference server. Might be worth @dmahan93 or @conceptofmind trying it out to verify. Also maybe worth adding a non-triton
infer_modelfunction that just loads the.ptfile and runs it in process.