Optimized DeepSeek V2/V3 implementation (MLA)#11446

Closed

fairydreaming wants to merge 10 commits intoggml-org:masterfrom

fairydreaming:deepseek2-mla-exp

Commits on Jan 22, 2025

llama : experimental DeepSeek2 MLA implementation that caches latent kv representations
sszymczy
committed

Commits on Jan 24, 2025

Merge remote-tracking branch 'origin/master' into deepseek2-mla-exp
sszymczy
committed

Commits on Jan 25, 2025

llama : optimize DeepSeek MLA implementation
sszymczy
committed

Commits on Jan 26, 2025

llama : Update tensor names in DeepSeek2 MLA implementation.
sszymczy
committed
llama : add a second copy of c^KV cache in DeepSeek2 MLA to avoid transposing the cache during inference
sszymczy
committed

Commits on Jan 27, 2025

Commits on Jan 28, 2025

Commits on Jan 31, 2025

llama : avoid ggml_cont() is possible in DeepSeek V2 implementation
sszymczy
committed