ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Port speculative decoding from upstream to llama-server (#645)

Browse Source 
* server : integrate speculative decoding

* server: Fix field names

* server: fix include, whitespace

* fix compile errors in speculative.cpp

* add llama_sampling_sample_and_accept_n to sampling

* finish porting speculative decoding in server

* port functions from common/speculative, common/sampling

* remove arg

* fix function names

* init params_dft to none

* correct value for n_ctx

* prefix kv cache tensors with model name to avoid conflict

* fix call arguments

* fix spec decoding args

* correct slot.id

* use n_max

* port the rest of sampling funcs

* fix func arguments

* slot.id starts at 1?

* Revert "prefix kv cache tensors with model name to avoid conflict"

This reverts commit fbd5dfd866.

* disable draft logging

* disable logging in speculative.cpp

in mainline, these would be LOG_DEBUG, but since ik_llama doesnt support
it, logging is disabled entirely

* add more draft model parameters

* fix

* pass flash_attn

* add speculative params for parity

* set speculative params in launch_slot_with_task instead
This commit is contained in:
g2mt
2025-08-15 21:26:44 -07:00
committed by GitHub
parent 2e2abddaa8
commit b6bc5eedad
8 changed files with 655 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
common/common.h
View File

@@ -83,10 +83,13 @@ struct gpt_params {
int32_t n_threads_batch_draft = -1;
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 0; // context size
int32_t n_ctx_draft = 0; // context size for draft model
int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_ubatch = 512; // physical batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_draft = 5; // number of tokens to draft during speculative decoding
int32_t n_draft = 16; // number of tokens to draft during speculative decoding
int32_t n_draft_min = 1; // minimum number of tokens to draft during speculative decoding
float p_draft_min = 0.8f; // minimum speculative decoding probability (greedy)
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_parallel = 1; // number of parallel sequences to decode
int32_t n_sequences = 1; // number of sequences to decode
@@ -207,6 +210,8 @@ struct gpt_params {
std::string cache_type_k = "f16"; // KV cache data type for the K
std::string cache_type_v = "f16"; // KV cache data type for the V
std::string cache_type_k_draft = ""; // KV cache data type for K for the draft model
std::string cache_type_v_draft = ""; // KV cache data type for V for the draft model
// multimodal models (see examples/llava)
std::string mmproj = ""; // path to multimodal projector