[FEATURE]: [BUILD] Install single candle-core version throughout Dynamo

[rmccorm4]

Feature request

If you build dynamo with cargo build --timings, you will get a detailed HTML report of the build times of every crate/version.

This report revealed that we have ~2 different versions of candle-core (v0.8.0 && v0.8.4) being installed in Dynamo.

We should install a single version in the workspace and re-use throughout the Dynamo build to reduce build times.

Similar to #2314

Describe the problem you're encountering

N/A

Describe alternatives you've tried

No response

[rmccorm4]

CC @da-x since you fixed #2314 already 🙂 *

[qimcis]

Took a look at this, looks like mistralrs (git dependency) is hardcoded to use candle-core v0.8.0 from git source, whereas our workspace dynamo-llm is using candle-core v0.8.4 from registry. Since mistralrs relies on v0.8.0, it seems like the best solution would be to make dynamo-llm use the same git source as mistralrs, let me know if we'd like to proceed with this! @rmccorm4

Refer to lib/engines/mistralrs/Cargo.toml and lib/llm/Cargo.toml

[grahamking]

Thanks for doing the detective work @qimcis !

mistral.rs uses it's own unpublished fork of candle, here https://github.com/EricLBuehler/candle . It don't understand why, because I think Eric is on the Candle team at Hugging Face.

Unless we make Dynamo depend on his fork in github, we're always going to have two copies of candle.