llama : custom attention mask + parallel decoding + no context swaps (#3228)

* tests : verify that RoPE is "additive"

* llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask)

* ggml : ggml_rope now takes a vector with positions instead of n_past

* metal : add rope_f16 kernel + optimize cpy kernels

* llama : unified KV cache + batch inference API

* llama : add new llama_decode() API that works with llama_batch

* llama : add cell_max heuristic for more efficient kv_cache

* llama : extend llama_kv_cache API

* llama : more robust cell_max heuristic + wip shift

* metal : disable concurrency optimization

* llama : add llama_kv_cache_shift_seq + no more context swaps

* llama : apply K-cache roping for Falcon and Baichuan

* speculative : fix KV cache management

* parallel : example for serving multiple users in parallel

* parallel : disable hot-plug to avoid cache fragmentation

* fixes : speculative KV cache + llama worst-case graph

* llama : extend batch API to select which logits to output

* llama : fix worst case graph build

* ggml-cuda : update rope implementation for parallel decoding (#3254)

* ggml-cuda : update rope implementation for parallel decoding

* better solution for p0 computation

* fix rope

* simpler rope implementation

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* make : add parallel to build + fix static functions in llama.cpp

* simple : fix token counting

* parallel : various improvements

* llama : fix cell_max logic + rename functions

* parallel : try smaller batches when the KV cache is fragmented

* parallel : fix sequence termination criteria

* llama : silence errors KV cache errors

* parallel : remove new line from prompt

* parallel : process system prompt once + configurable paramters + llama API

* parallel : remove question with short answers

* parallel : count cache misses

* parallel : print misses on each request

* parallel : minor

* llama : fix n_kv to never become 0

* parallel : rename hot-plug to continuous-batching

* llama : improve llama_batch API + simplify parallel example

* simple : add parallel decoding support

* simple : improve comments + free batch

* ggml-cuda : add rope f16, restore performance with parallel decoding (#3272)

* ggml-cuda : add rope f16, restore performance

* offload KQ_mask with all models

* fix rope shift

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* llama : disable MPI for now

ggml-ci

* train : make KQ_pos memory buffer permanent via dummy scale op

* ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275)

ggml-ci

* parallel : fix bug (extra BOS) + smaller token_prev array

* parallel : fix cases where the input prompts can overflow the batch

* parallel : add disabled experimental batch chunking in powers of two

* llama : llama.h formatting + comments

* simple : add README.md

* llama : fix kv cache heuristic when context is less than 32

* parallel : fix crash when `-n -1`

* llama : simplify returns if/else branches

* metal : use mm kernels for batch size > 2

* examples : utilize new llama_get_logits_ith()

* examples : add example for batched decoding

* examples : do not eval prompt 2 times (close #3348)

* server : clear the KV cache beyond n_past before llama_decode

* server : avoid context swaps by shifting the KV cache

---------

Co-authored-by: slaren <slarengh@gmail.com>

examples/speculative/speculative.cpp

@@ -37,7 +37,7 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    params.perplexity = true; // HACK: enable logits_all = true
+    params.logits_all = true;
     std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
 
     // load the draft model
@@ -70,9 +70,9 @@ int main(int argc, char ** argv) {
     const auto t_enc_start = ggml_time_us();
 
     // eval the prompt with both models
-    llama_eval(ctx_tgt,  inp.data(), int(inp.size() - 1), 0, params.n_threads);
-    llama_eval(ctx_tgt, &inp.back(),      1, inp.size() - 1, params.n_threads);
-    llama_eval(ctx_dft,  inp.data(),     int(inp.size()), 0, params.n_threads);
+    llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0,           0), params.n_threads);
+    llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0), params.n_threads);
+    llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input,     0,           0), params.n_threads);
 
     const auto t_enc_end = ggml_time_us();
 
@@ -134,7 +134,7 @@ int main(int argc, char ** argv) {
 
         while (true) {
             // sample from the target model
-            const llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);
+            llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);
 
             // remember which tokens were sampled - used for repetition penalties during sampling
             last_tokens.erase(last_tokens.begin());
@@ -172,7 +172,8 @@ int main(int argc, char ** argv) {
                 LOG("out of drafted tokens\n");
             }
 
-            llama_eval(ctx_dft, &id, 1, n_past_dft, params.n_threads);
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, n_ctx);
+            llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0), params.n_threads);
             ++n_past_dft;
 
             // heuristic for n_draft
@@ -256,7 +257,8 @@ int main(int argc, char ** argv) {
             }
 
             // evaluate the drafted token on the draft model
-            llama_eval(ctx_dft, &drafted.back(), 1, n_past_cur, params.n_threads);
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, n_ctx);
+            llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0), params.n_threads);
             ++n_past_cur;
 
             if (grammar_dft != NULL) {
@@ -265,7 +267,8 @@ int main(int argc, char ** argv) {
         }
 
         // evaluate the target model on the drafted tokens
-        llama_eval(ctx_tgt, drafted.data(), drafted.size(), n_past_tgt, params.n_threads);
+        llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, n_ctx);
+        llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0), params.n_threads);
         ++n_past_tgt;
 
         // the first token is always proposed by the traget model before the speculation loop