vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

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

Browse source 
* 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>
This commit is contained in:
Georgi Gerganov 2023-09-28 19:04:36 +03:00 • committed by GitHub
parent 45855b3f1c
commit ec893798b7
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
35 changed files with 2700 additions and 673 deletions
Download patch file Download diff file Expand all files Collapse all files

8
examples/llama-bench/llama-bench.cpp
View file

@ -891,7 +891,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
int n_processed = 0;
while (n_processed < n_prompt) {
int n_tokens = std::min(n_prompt - n_processed, n_batch);
llama_eval(ctx, tokens.data(), n_tokens, n_past + n_processed, n_threads);
llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0), n_threads);
n_processed += n_tokens;
}
}
@ -899,7 +899,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
llama_token token = llama_token_bos(ctx);
for (int i = 0; i < n_gen; i++) {
llama_eval(ctx, &token, 1, n_past + i, n_threads);
llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0), n_threads);
}
}
@ -977,6 +977,8 @@ int main(int argc, char ** argv) {
test t(inst, lmodel, ctx);
llama_kv_cache_tokens_rm(ctx, -1, -1);
// warmup run
if (t.n_prompt > 0) {
test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
@ -986,6 +988,8 @@ int main(int argc, char ** argv) {
}
for (int i = 0; i < params.reps; i++) {
llama_kv_cache_tokens_rm(ctx, -1, -1);
uint64_t t_start = get_time_ns();
if (t.n_prompt > 0) {
test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);