ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-20 21:24:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Enable CUDA graphs for MoE models + GPT-OSS support (#689)

Browse Source 
* gmp-oss: common

* gpt-oss: attnetion sinks, swiglu_oai

* gpt-oss: WIP llama

Model loads and runs (CPU only), but PPL is much to high
(~1500 for 1st batch vs ~200 in mainline).
Is it because of SWA, because of vocab, or did I introduce a bug somewhere?

* gpt-oss: CPU seems to be working

It was the SWA thta was missing in the previous commit.

There are issues with EOG tokens, so this still needs to be added.

* CUDA: ADD_ID

Just a copy from mainline

* gpt-oss: Seems to be working on CUDA

* gpt-oss: add sinks to the attn-vec kernels

* CUDA: add head size of 64 to new mma

Haven't turned it on yet, but observe slightly better PP and slightly
worse TG performance with that.

* gpt-oss: add ability to use -fmoe (only CUDA for now)

* Move row sums to the write place

* Add sinks to iqk flash attention

* gpt_oss: Implement -fmoe on the CPU

* Simdify swiglu_oai

Turning it off for now as performance becomes more variable,
so perhaps I'm running into thermal trottling imore often
because of making the CPU work too hard.

* llama: factor out model loader

* Builds successfully

* It runs, but mmap does not work

* Fix llama_mmap so mmap works

* Minor

* Fix CUDA after latest changes

* Attempt to use CUDA graphs with MoE models - not working

* CUDA graphs WIP - still not working

* CUDA graphs - seems to be working

Likely not all MLA variants are working.
I no longer remember why I added the q8_0 cpy that
transposes the tensor, but if really needed, this is now
missing. Also missing is q6_0.

* Make q8_0 cache work for DeepSeek models with CUDA graphs

* cuda: cpy for q6_0

* Fix llama_mmap on non-Linux platforms

* Adding forgotten file

* Iterating on Windows build failures

* cuda: re-add q8_0 -> q8_0 transpose

so mla = 2 can be used with CUDA graphs and q8_0 cache.

* Disable graphs without -fmoe

* Minor

* Turn graphs on by default

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-08-15 09:18:07 +03:00
committed by GitHub
parent c00335684c
commit 633e0617b0
56 changed files with 8720 additions and 5115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
src/llama-sampling.cpp
View File

@@ -734,7 +734,7 @@ llama_token llama_sample_token_impl(struct llama_sampling * smpl, llama_token_da
// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
static void get_overlapping_token_sequences(const llama_vocab& vocab, const std::string& str, std::unordered_multimap<llama_token, std::vector<llama_token>>& token_sequences, int max_tail_len = -1) {
for (llama_token token_id = 0; token_id < (llama_token)vocab.n_tokens(); token_id++) {
std::string word = llama_detokenize(vocab, { token_id }, true);
auto word = vocab.detokenize( { token_id }, true);
if (word.find(str) != std::string::npos) {
token_sequences.emplace(token_id, std::vector<llama_token>());
}
@@ -751,7 +751,8 @@ static void get_overlapping_token_sequences(const llama_vocab& vocab, const std:
}
}
if (match) {
std::vector<llama_token> tokenization = llama_tokenize_internal(vocab, str.substr(i), false, false);
auto tokenization = vocab.tokenize(str.substr(i), false, false);
//std::vector<llama_token> tokenization = llama_tokenize_internal(vocab, str.substr(i), false, false);
if (max_tail_len >= 0 && tokenization.size() > (size_t)max_tail_len) {
tokenization.resize(max_tail_len);
}