Cross GPU device mapping feature

[joshpopelka20]

I'm working with a long context model (gradientai/Llama-3-8B-Instruct-262k) that exceeds the memory of a single A100 GPU. While the model weights are loaded, when I try to run inference, I get CUDA Out of Memory exception.

Requesting a new feature to allow users to use cross GPU device mapping.

[b0xtch]

Related issue: https://github.com/huggingface/candle/issues/2007

There was an attempt to do tensor parallelism:

https://github.com/EricLBuehler/mistral.rs/pull/72

[EricLBuehler]

Hi @joshpopelka20 and @b0xtch! I just merged #462 which adds cross-GPU device mapping support (including for Python). I plan on implementing tensor parallelism, too, in the future.

[joshpopelka20]

@EricLBuehler thanks for adding this feature.

When using the pypi package, I'm getting this error:

        = note: /usr/bin/ld: /tmp/pip-install-8y1wtzkm/mistralrs-cuda_36180abfef7b4f0687d7842e9b298d9e/target/release/build/mistralrs-core-f5a6ed7a31cbdb62/out/libmistralcuda.a(nonzero_bitwise-b50867152df76f01.o): relocation R_X86_64_32 against symbol `_Z17transform_indicesPKjjS0_jPj' can not be used when making a shared object; recompile with -fPIC
                /usr/bin/ld: final link failed: Nonrepresentable section on output
                collect2: error: ld returned 1 exit status

I'm passing the CUDA_NVCC_FLAGS flag so not sure why it's saying to "recompile with -fPIC". These are the commands I'm using:

env["CUDA_NVCC_FLAGS"] = "-fPIE" result = subprocess.run(['pip', 'install', 'mistralrs-cuda'], env=env)

[EricLBuehler]

I'm passing the CUDA_NVCC_FLAGS flag so not sure why it's saying to "recompile with -fPIC". These are the commands I'm using:

Can you try:

env["CUDA_NVCC_FLAGS"] = "-fPIC"
result = subprocess.run(['pip', 'install', 'mistralrs-cuda'], env=env)

The -fPIC requirement may stem from your Linux distribution (some require -fPIE, I'll add this to the README)?

[joshpopelka20gmail]

Sorry, may not have been clear. I got that error when running with the flag set to '-fPIC'. I haven't tried running the code using the git repo and cargo. Do you want me to verify if that works?

I'd guess they'd be the same though.

[EricLBuehler]

Sorry, may not have been clear. I got that error when running with the flag set to '-fPIC'. I haven't tried running the code using the git repo and cargo. Do you want me to verify if that works?

Ah, ok. Not sure if we discussed this before, but what Linux distribution are you using?

[joshpopelka20]

I'm using a Sagemaker Jupyter notebook so Amazon Linux. This is the distro info:

NAME="Amazon Linux"
VERSION="2"
ID="amzn"
ID_LIKE="centos rhel fedora"
VERSION_ID="2"
PRETTY_NAME="Amazon Linux 2"
ANSI_COLOR="0;33"
CPE_NAME="cpe:2.3:o:amazon:amazon_linux:2"
HOME_URL="https://amazonlinux.com/"
SUPPORT_END="2025-06-30"
Amazon Linux release 2 (Karoo)

[joshpopelka20]

Adding some observations:

When I run CUDA_NVCC_FLAGS=-fPIE cargo build --release --features "cuda cudnn" and CUDA_NVCC_FLAGS=-fPIE cargo build --release --features "cuda flash-attn", I don't get the error.

When I run CUDA_NVCC_FLAGS=-fPIE cargo build --release --features "flash-attn cudnn" or CUDA_NVCC_FLAGS=-fPIE cargo build --release --features "cuda flash-attn cudnn", I get the same error:

= note: /usr/bin/ld: /home/ec2-user/mistral.rs/target/release/build/mistralrs-core-9c498c55121e0e87/out/libmistralcuda.a(nonzero_bitwise-b50867152df76f01.o): relocation R_X86_64_32 against symbol `_Z17transform_indicesPKjjS0_jPj' can not be used when making a shared object; recompile with -fPIC

Running CUDA_NVCC_FLAGS=-fPIE cargo build --release --features "cudnn" by itself is also successful.

So it doesn't seem like compiling with both cudnn and flash-attn.

[joshpopelka20]

I'm thinking the issue is related to mistralrs-core/build.rs file. I tried to add

.arg("--compiler-options")
.arg("-fPIC")

but it didn't help.

I think it's more of an issue with this line of code println!("cargo:rustc-link-lib=mistralcuda"). How do I find that C library and determine if it was compiled with the -fPIC flag?

[EricLBuehler]

I'm thinking the issue is related to mistralrs-core/build.rs file. I tried to add

I added support for the NVCC flag envvar to that, so it should be seamless to use the envvar instead of changing the code, now.

I think it's more of an issue with this line of code println!("cargo:rustc-link-lib=mistralcuda"). How do I find that C library and determine if it was compiled with the -fPIC flag?

It's in whatever the OUT_DIR environment variable is. Perhaps you can panic! on it: panic!("{build_dir}");?

[joshpopelka20]

I added a pull request with a fix for the issue https://github.com/EricLBuehler/mistral.rs/pull/471. Looks like it was a divide by zero issue.

I didn't add any error message; I just let it continue to run. I ran llama and there was no issues.

Also, I'm wondering if there should be other code added to the build.rs file. Like in the candle project:

    let target = std::env::var("TARGET").unwrap();
    if target.contains("msvc") {
        // nothing to link to
    } else if target.contains("apple") || target.contains("freebsd") || target.contains("openbsd") {
        println!("cargo:rustc-link-lib=dylib=c++");
    } else if target.contains("android") {
        println!("cargo:rustc-link-lib=dylib=c++_shared");
    } else {
        println!("cargo:rustc-link-lib=dylib=stdc++");
    }

I didn't have any issues with these, but someone else might be using those OSs in the future.

[EricLBuehler]

Also, I'm wondering if there should be other code added to the build.rs file. Like in the candle project:

I just merged #472 which adds this, thanks for pointing that out.

[joshpopelka20]

The layers are now distributed across my 4 A10G GPUs

One request I have is: can there be a progress bar added when the model is being loaded? For larger models (40+ gb), it takes about 20 mins and is hard to know what is going on.

No rush on this, but it would be a nice enhancement.

[EricLBuehler]

Hi @joshpopelka20! I just merged #479 which adds a loading bar while loading the repeating layers. It would be great if you could install from source with maturin ahead of the PyPI rolling release (in ~2 days) to try it out!

[joshpopelka20]

There were no issues building from source. Also, the 2 day delay is not an issue for me.

Finally, I think the ask for this issue is complete. Would you like me to leave it open for adding tensor parallelism in the future? I'm not sure how you are tracking that.

[EricLBuehler]

There were no issues building from source. Also, the 2 day delay is not an issue for me.

Great, just one thing to confirm: does the progress bar function to show the loading?

Would you like me to leave it open for adding tensor parallelism in the future? I'm not sure how you are tracking that.

I'll create a separate issue, as device mapping is a bit different from tensor parallelism.

[b0xtch]

Amazing stuff! The tensor parallelism I am guessing will be on the core candle repo? or do you plan to abstract that in some way under this repo?

I have linked the device mapping issue with one in candle. https://github.com/huggingface/candle/issues/2007

[EricLBuehler]

Amazing stuff! The tensor parallelism I am guessing will be on the core candle repo? or do you plan to abstract that in some way under this repo?

I plan on implementing the higher-level aspects here: the synchronization, the reduce ops, etc., which can all be done with the public Candle APIs. I actually have a fork of Candle which I maintain (https://github.com/EricLBuehler/candle). I have this because some of the features which make mistral.rs faster than Candle would not get merged quickly enough/fit that project's goal well. However, I do want to contribute any progress I make with tensor parallelism and so I'll try to contribute what makes sense!

[joshpopelka20]

When I run CUDA_NVCC_FLAGS="-fPIE" maturin develop -r --features "cuda flash-attn cudnn" and try to load the model with the Runner class, I get this error:

panicked at /home/ec2-user/.cargo/registry/src/index.crates.io-6f17d22bba15001f/cudarc-0.11.6/src/driver/result.rs:63:43:
thread panicked while processing panic. aborting.

I don't normally run from the command line so I tried with the current pip package (mistralrs-cuda 0.1.22). That was able to load the model.

[EricLBuehler]

When I run CUDA_NVCC_FLAGS="-fPIE" maturin develop -r --features "cuda flash-attn cudnn" and try to load the model with the Runner class, I get this error:

Can you try to run that with RUST_BACKTRACE=1?

[joshpopelka20]

I tried with RUST_BACKTRACE=1 CUDA_NVCC_FLAGS="-fPIE" maturin develop -r --features "cuda flash-attn cudnn" and export RUST_BACKTRACE=1, it isn't giving me any additional output. Not sure if I'm doing it wrong or that's it for the stacktrace.

[joshpopelka20]

Tried 'full' as well:

[joshpopelka20]

https://github.com/EricLBuehler/mistral.rs/issues/478 also seems to have an issue with that library as well. Though in a different file: cudarc-0.11.6/src/lib.rs

Not sure if they're related.

[joshpopelka20]

It seems to be throwing the error at this line:

let err_str = self.error_string().unwrap();

From this https://users.rust-lang.org/t/how-to-prevent-thread-panicked-while-processing-panic-aborting/56508/2, it looks like it might be an issue with trying to unwrap that error string. I'll try to debug tonight.

[joshpopelka20]

I cloned https://github.com/coreylowman/cudarc and added this code to the mistral.rs root Cargo.toml (on my box):

[patch.crates-io]
cudarc = { path = "/home/ec2-user/cudarc" }

I removed the unwrap function and now I'm getting this error: Segmentation fault

Any suggestions on further debugging?

I'll see if I can get more output tomorrow; right now, that error seems like another Cuda bug.

[NiuBlibing]

I tried to run with ./target/release/mistralrs-server --port 1234 -n "0:20;1:20;2:20;3:20" plain -m ./Qwen/Qwen2-72B-Instruct/ -a qwen2 but still oom where only one gpu's memory was growing.

And ./target/release/mistralrs-server --port 1234 -n "0:6;1:6;2:6;3:6" plain -m /jr-sec-ai-train/open-models/Qwen/Qwen1.5-0.5B-Chat/ -a qwen2 is success where four gpus' memory was growing.

Is there something wrong for me? I use git version e04f8400.

[joshpopelka20]

I was able to get more logging when I used panic in line 53 of cudarc/src/driver/result.rs

   pub fn error_string(&self) -> Result<&CStr, DriverError> {
        let mut err_str = MaybeUninit::uninit();
        panic!("{:?}", err_str);

Not sure it's helpful though

thread '<unnamed>' panicked at /home/ec2-user/cudarc/src/driver/result.rs:53:9:
core::mem::maybe_uninit::MaybeUninit<*const i8>
stack backtrace:
   0:     0x7ff5fd0eb3f5 - <std::sys_common::backtrace::_print::DisplayBacktrace as core::fmt::Display>::fmt::h1e1a1972118942ad
   1:     0x7ff5fd1197cb - core::fmt::write::hc090a2ffd6b28c4a
   2:     0x7ff5fd0e74df - std::io::Write::write_fmt::h8898bac6ff039a23
   3:     0x7ff5fd0eb1ce - std::sys_common::backtrace::print::ha96650907276675e
   4:     0x7ff5fd0ec639 - std::panicking::default_hook::{{closure}}::h215c2a0a8346e0e0
   5:     0x7ff5fd0ec37d - std::panicking::default_hook::h207342be97478370
   6:     0x7ff5fd0ecad3 - std::panicking::rust_panic_with_hook::hac8bdceee1e4fe2c
   7:     0x7ff5fd0ec9b4 - std::panicking::begin_panic_handler::{{closure}}::h00d785e82757ce3c
   8:     0x7ff5fd0eb8b9 - std::sys_common::backtrace::__rust_end_short_backtrace::h1628d957bcd06996
   9:     0x7ff5fd0ec6e7 - rust_begin_unwind
  10:     0x7ff5fc369e43 - core::panicking::panic_fmt::hdc63834ffaaefae5
  11:     0x7ff5fd070eba - <&T as core::fmt::Debug>::fmt::hbb771b0a79147136
  12:     0x7ff5fd1197cb - core::fmt::write::hc090a2ffd6b28c4a
  13:     0x7ff5fd071010 - <cudarc::driver::result::DriverError as core::fmt::Display>::fmt::heb0f09e810474a5e
  14:     0x7ff5fd1197cb - core::fmt::write::hc090a2ffd6b28c4a
  15:     0x7ff5fcf92ae7 - <candle_core::error::Error as core::fmt::Display>::fmt::hf6848a77fb28bd8b
  16:     0x7ff5fc37ca6e - mistralrs::Runner::new::h62e9d9fcf7c2e3fa
  17:     0x7ff5fc3843d4 - mistralrs::Runner::__pymethod___new____::h12cf9fba34601bfc
  18:     0x7ff5fc37980a - pyo3::impl_::trampoline::trampoline::hf137faff76e4bf3b
  19:     0x7ff5fc3837c1 - mistralrs::<impl pyo3::impl_::pyclass::PyMethods<mistralrs::Runner> for pyo3::impl_::pyclass::PyClassImplCollector<mistralrs::Runner>>::py_methods::ITEMS::trampoline::hcb06f753a45992c0
  20:     0x556be9392db2 - type_call
                               at /usr/local/src/conda/python-3.10.8/Objects/typeobject.c:1123:11
  21:     0x556be9392db2 - _PyObject_MakeTpCall
                               at /usr/local/src/conda/python-3.10.8/Objects/call.c:215:18
  22:     0x556be938f097 - _PyObject_VectorcallTstate
                               at /usr/local/src/conda/python-3.10.8/Include/cpython/abstract.h:112:16
  23:     0x556be938f097 - _PyObject_VectorcallTstate
                               at /usr/local/src/conda/python-3.10.8/Include/cpython/abstract.h:99:1
  24:     0x556be938f097 - PyObject_Vectorcall
                               at /usr/local/src/conda/python-3.10.8/Include/cpython/abstract.h:123:12
  25:     0x556be938f097 - call_function
                               at /usr/local/src/conda/python-3.10.8/Python/ceval.c:5891:13
  26:     0x556be938f097 - _PyEval_EvalFrameDefault
                               at /usr/local/src/conda/python-3.10.8/Python/ceval.c:4231:19
  27:     0x556be943c732 - _PyEval_EvalFrame
                               at /usr/local/src/conda/python-3.10.8/Include/internal/pycore_ceval.h:46:12
  28:     0x556be943c732 - _PyEval_Vector
                               at /usr/local/src/conda/python-3.10.8/Python/ceval.c:5065:24
  29:     0x556be943c677 - PyEval_EvalCode
                               at /usr/local/src/conda/python-3.10.8/Python/ceval.c:1134:12
  30:     0x556be9470049 - run_eval_code_obj
                               at /usr/local/src/conda/python-3.10.8/Python/pythonrun.c:1291:9
  31:     0x556be946a964 - run_mod
                               at /usr/local/src/conda/python-3.10.8/Python/pythonrun.c:1312:19
  32:     0x556be92ee123 - pyrun_file
                               at /usr/local/src/conda/python-3.10.8/Python/pythonrun.c:1208:15
  33:     0x556be9464c9f - _PyRun_SimpleFileObject
                               at /usr/local/src/conda/python-3.10.8/Python/pythonrun.c:456:13
  34:     0x556be9464863 - _PyRun_AnyFileObject
                               at /usr/local/src/conda/python-3.10.8/Python/pythonrun.c:90:15
  35:     0x556be9461a1f - pymain_run_file_obj
                               at /usr/local/src/conda/python-3.10.8/Modules/main.c:357:15
  36:     0x556be9461a1f - pymain_run_file
                               at /usr/local/src/conda/python-3.10.8/Modules/main.c:376:15
  37:     0x556be9461a1f - pymain_run_python
                               at /usr/local/src/conda/python-3.10.8/Modules/main.c:591:21
  38:     0x556be9461a1f - Py_RunMain
                               at /usr/local/src/conda/python-3.10.8/Modules/main.c:670:5
  39:     0x556be942f969 - Py_BytesMain
                               at /usr/local/src/conda/python-3.10.8/Modules/main.c:1090:12
  40:     0x7ff60dd0b13a - __libc_start_main
  41:     0x556be942f871 - <unknown>
Traceback (most recent call last):
  File "/home/ec2-user/test.py", line 27, in <module>
    llm = Runner(
pyo3_runtime.PanicException: core::mem::maybe_uninit::MaybeUninit<*const i8>

[joshpopelka20]

I've narrowed it down to this line in mistralrs-pyo3/src/lib.rs for function not metal get_device():

let res = Device::cuda_if_available(0)?;

The error handling isn't very good to know exactly what the error is. I'm trying some add'l error handling, but nothing has worked so far.

[EricLBuehler]

I've narrowed it down to this line in mistralrs-pyo3/src/lib.rs for function not metal get_device():

The error is probably happening there because that is when the CUDA stuff will get initialized for the first time.

[joshpopelka20]

CUDA operation failed with error: CUDA_ERROR_STUB_LIBRARY

I think it's an issue with LD_LIBRARY_PATH pointing to cuda12.1. I tried to manually update it, but that didn't work. I'll get a ticket open with AWS.

Also, I'm going to create a PR with cudarc to add the additional error handling. Think that'll be beneficial going forward.