Deepseek V3 support added

[saood06]

Very direct port of ggml-org/llama.cpp#11049.

Tested working with IQ4_K_R4 and IQ4_K. No tests so far on any quant that is supported by llama.cpp so that performance can be compared.

Tested on dual socket Xeon E5-2690 v3
Prompt processing:11.5 t/s for IQ4_K, 9.8 t/s IQ4_K_R4
Token generation: 2.75 t/s for IQ4_K, 3.10 t/s for IQ4_K_R4

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[ikawrakow]

@saood06

Quick question: current llama.cpp has this check for Deepseek-V3:

    } else if (tmpl_contains(LU8("<｜Assistant｜>")) && tmpl_contains(LU8("<｜User｜>")) && tmpl_contains(LU8("<｜end▁of▁sentence｜>"))) {
        return LLM_CHAT_TEMPLATE_DEEPSEEK_3;

while the check you added with this PR is

    else if (tmpl == "deepseek3" || tmpl_contains(LU8("'<｜Assistant｜>' + message['content'] + '<｜end▁of▁sentence｜>'"))) {

The check for tmpl == "deepseek3" is done before in llama.cpp, so this is not an issue, but the remainder is not the same. Is this a problem? Or would it be a problem if I just made it the same as llama.cpp ?

[saood06]

The change you are referencing happened in ggml-org/llama.cpp@ec7f3ac I was not aware of that till now.

Is this a problem? Or would it be a problem if I just made it the same as llama.cpp ?

You can change it if you want but both work, based on the chat_templates for the models that have been released.