Conversation
|
Do we also want to maybe edit the |
This will allow users to do the following: 1. Cli where if they can pass in anything that's "cepo_<name-of-attribute>" 2. Yaml file where if they pass it in as "<name-of-attribute>" 3. If none of them have a specific attribute, we use the default setting 4. If both of them have the specific attribute, we error out
erich-cerebras
left a comment
erich-cerebras
left a comment
There was a problem hiding this comment.
Just a couple of small comments, besides that it LGTM!
Add modification of CePO configs through yaml and cli arguments
…ass for single source of truth
vithursant
left a comment
vithursant
left a comment
There was a problem hiding this comment.
Left some comments, overall looking good.
| @@ -0,0 +1,314 @@ | |||
| import re | |||
There was a problem hiding this comment.
@pawelf-cerebras I think we should add LICENSE for the cepo code. In the past we've done something like this.
https://github.com/CerebrasResearch/Sparse-IFT/blob/master/cbsparse/LICENSE
There was a problem hiding this comment.
@pawelf-cerebras did you have any thoughts on this?
There was a problem hiding this comment.
Yes, thanks for pointing this out. We're going to add the license. I mentioned this last week to Emma and she's making sure with the legal team that everything is good. I want to leave this comment unresolved until I hear from her and add the license.
- yield response to allow clients that expect streaming
- add dockerignore - parameterize port in dockerfile
prepare for new release
prepare for new release
- handle multiple responses as multiple attempts
- prepare for new release
- update max tokens
fix bug
add protobuf
fix inference on amd gpu
This reverts commit c0f0893.
This will allow users to do the following: 1. Cli where if they can pass in anything that's "cepo_<name-of-attribute>" 2. Yaml file where if they pass it in as "<name-of-attribute>" 3. If none of them have a specific attribute, we use the default setting 4. If both of them have the specific attribute, we error out
7 participants
Adding Cerebras Planning and Optimization (CePO). On high-level, in CePO, we make m attempts to generate n step-by-step plans, refine the plans, check inconsistencies between them, use the above feedback to generate the final plan and produce the answer. This process is them repeated N times in a classical best of n manner.