Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

build(llama): manipulate cmake flags via cargo features #1716

Open
wants to merge 6 commits into
base: main
Choose a base branch
from
Open

build(llama): manipulate cmake flags via cargo features #1716

wants to merge 6 commits into from

Conversation

pinbraerts
Copy link

fixes #1142

@pinbraerts pinbraerts changed the title llama: manipulate cmake flags via cargo features fix-avx2: llama: manipulate cmake flags via cargo features Mar 25, 2024
@codecov Codecov
Copy link

codecov bot commented Mar 25, 2024 •
edited

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 53.94%. Comparing base (d24424d) to head (aa36e59).
Report is 2 commits behind head on main.

❗ Current head aa36e59 differs from pull request most recent head a54af19. Consider uploading reports for the commit a54af19 to get more accurate results

Additional details and impacted files 
@@            Coverage Diff             @@
##             main    #1716      +/-   ##
==========================================
- Coverage   55.46%   53.94%   -1.53%     
==========================================
  Files         125      115      -10     
  Lines       10936     9625    -1311     
==========================================
- Hits         6066     5192     -874     
+ Misses       4870     4433     -437

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

@pinbraerts
Copy link
Author

I have i7-3820 which does not support avx2. This patch works for me. I complemented default cmake parameters in llama.cpp/CMakeLists.txt

wsxiaoys
wsxiaoys reviewed Mar 26, 2024
View reviewed changes
crates/llama-cpp-bindings/Cargo.toml Outdated
@@ -4,9 +4,18 @@ version = "0.10.0-dev.0"
edition = "2021"

[features]
default = ["native", "avx", "fma"]
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
default = ["native", "avx", "fma"]
# non-windows
default = ["avx", "avx2", "fma", "f16c"]
# windows
default = ["avx", "avx2", "fma"]

This should be the actual default tabby is using, could you update?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok. What should I do with avx2? I disabled it by default in order to fix #1142. Was inspired by ollama/ollama#644 (comment)

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done with avx2 on by default

@pinbraerts pinbraerts changed the title fix-avx2: llama: manipulate cmake flags via cargo features fix(avx2, llama) manipulate cmake flags via cargo features Mar 28, 2024
@pinbraerts pinbraerts changed the title fix(avx2, llama) manipulate cmake flags via cargo features fix(avx2, llama): manipulate cmake flags via cargo features Mar 28, 2024
@pinbraerts pinbraerts changed the title fix(avx2, llama): manipulate cmake flags via cargo features build(llama): manipulate cmake flags via cargo features Mar 29, 2024
vonydev
vonydev suggested changes Apr 12, 2024
View reviewed changes
Copy link

@vonydev vonydev left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

deployed on a Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz with the changes I've mentioned.
Works nice. Thank you for the PR

crates/llama-cpp-bindings/build.rs
@@ -21,8 +21,60 @@ fn main() {

fn build_llama_cpp() {
let mut config = Config::new("llama.cpp");
config.define("LLAMA_NATIVE", "OFF");
config.define("INS_ENB", "ON");

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

from llama.cpp/CMakeLists.txt:

# instruction set specific
if (LLAMA_NATIVE)
    set(INS_ENB OFF)
else()
    set(INS_ENB ON)
endif()

option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
option(LLAMA_AVX512_VBMI                     "llama: enable AVX512-VBMI"                        OFF)
option(LLAMA_AVX512_VNNI                     "llama: enable AVX512-VNNI"                        OFF)
option(LLAMA_FMA                             "llama: enable FMA"                                ${INS_ENB})
# in MSVC F16C is implied with AVX2/AVX512
if (NOT MSVC)
    option(LLAMA_F16C                        "llama: enable F16C"                               ${INS_ENB})
endif()

so if LLAMA_NATIVE is ON then AVX, AVX2, FMA and F16C are OFF
and if LLAMA_NATIVE is OFF then AVX, AVX2, FMA and F16C are ON

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the remark, I misunderstood the interaction between this flags

crates/llama-cpp-bindings/build.rs

if cfg!(feature = "native") {
config.define("LLAMA_NATIVE", "ON");

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

definig LLAMA_NATIVE = ON turns all other options automatically OFF so the following defines have no purpose. They should be moved in the else brach

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Swapped the branches

crates/llama-cpp-bindings/build.rs Outdated
config.define("LLAMA_NATIVE", "ON");

if cfg!(not(feature = "avx")) {
config.define("LLAMA_AVX2", "OFF");
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

typo: should be LLAMA_AVX instead of LLAMA_AVX2

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fixed

crates/llama-cpp-bindings/build.rs
}
else {
config.define("LLAMA_NATIVE", "OFF");

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

definig LLAMA_NATIVE = OFF turns all other options automatically ON so the following defines have no purpose. They should switch places with the defines above

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fixed

crates/llama-cpp-bindings/build.rs Outdated
config.define("LLAMA_NATIVE", "OFF");

if cfg!(feature = "avx") {
config.define("LLAMA_AVX2", "ON");
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

typo

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fixed

@pinbraerts
llama: manipulate cmake flags via cargo features
3d2d2fa
@pinbraerts
better precise control over features
5b4f917 
LLAMA_NATIVE enables a bunch of options automatically, so if cargo
features are disabled, corresponding options will be disabled too.
Now the `native` feature only enables `-march=native`
@pinbraerts
restore default flags
b60daaf
@pinbraerts
tunnel cargo flags
468cded
@pinbraerts
fix native flag interaction
abfb376
@pinbraerts
Copy link
Author

Should I also disable default-features in crates/tabby/Cargo.toml for llama-cpp-bindings and reenable them under default feature?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@vonydev vonydev Awaiting requested review from vonydev

@wsxiaoys wsxiaoys Awaiting requested review from wsxiaoys

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

Supports cpu without avx2 instruction set
3 participants
@pinbraerts @wsxiaoys @vonydev