Add support for Deepseek

[franklucky001]

deepseek-ai/DeepSeek-V2-Chat

[bnusunny]

Two new models:

https://huggingface.co/deepseek-ai/DeepSeek-V3

https://huggingface.co/deepseek-ai/DeepSeek-R1

[phdbrianlee]

code models using mix of expert archtecture:
https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct
https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-lite-Instruct

[super-fun-surf]

is there work on this for candle?

[EricLBuehler]

@franklucky001 @bnusunny @phdbrianlee @super-fun-surf

I'm working on implementing these. The PRs are not yet uploaded, but the following has been implemented

• Working DeepSeek V2/V2 Lite model
• DeepSeek V3/R1 (loading but not yet working)

[super-fun-surf]

rad! great to hear.

[franklucky001]

great!!

[EricLBuehler]

@franklucky001 @bnusunny @phdbrianlee @super-fun-surf

Both DeepSeek V2 and V3/R1 PRs have been opened:

• DeepSeek V2: Implement DeepSeek V2 #2744
• DeepSeek V3/R1: Implement DeepSeek V3/R1 #2745

@LaurentMazare the DeepSeek V3/R1 model is very large - do you think it would be more appropriate to implement it similar to the llama multiprocess example?

[super-fun-surf]

rad!. the quantized / distilled versions of R1 should work on smaller gpus.