Still not complete, needs some refinement to our prediction to understand the
discrete GPUs available space so we can see how many layers fit in each one
since we can't split one layer across multiple GPUs we can't treat free space
as one logical block
The GPU drivers take a while to update their free memory reporting, so we need
to wait until the values converge with what we're expecting before proceeding
to start another runner in order to get an accurate picture.
Trying to live off the land for cuda libraries was not the right strategy. We need to use the version we compiled against to ensure things work properly
This moves all the env var reading into one central module
and logs the loaded config once at startup which should
help in troubleshooting user server logs
Now that the llm runner is an executable and not just a dll, more users are facing
problems with security policy configurations on windows that prevent users
writing to directories and then executing binaries from the same location.
This change removes payloads from the main executable on windows and shifts them
over to be packaged in the installer and discovered based on the executables location.
This also adds a new zip file for people who want to "roll their own" installation model.
This change adds support for multiple concurrent requests, as well as
loading multiple models by spawning multiple runners. The default
settings are currently set at 1 concurrent request per model and only 1
loaded model at a time, but these can be adjusted by setting
OLLAMA_NUM_PARALLEL and OLLAMA_MAX_LOADED_MODELS.
Leaving the cudart library loaded kept ~30m of memory
pinned in the GPU in the main process. This change ensures
we don't hold GPU resources when idle.
This should resolve a number of memory leak and stability defects by allowing
us to isolate llama.cpp in a separate process and shutdown when idle, and
gracefully restart if it has problems. This also serves as a first step to be
able to run multiple copies to support multiple models concurrently.
If expanding the runners fails, don't leave a corrupt/incomplete payloads dir
We now write a pid file out to the tmpdir, which allows us to scan for stale tmpdirs
and remove this as long as there isn't still a process running.
This fixes a few bugs in the new sysfs discovery logic. iGPUs are now
correctly identified by their <1G VRAM reported. the sysfs IDs are off
by one compared to what HIP wants due to the CPU being reported
in amdgpu, but HIP only cares about GPUs.
This allows people who package up ollama on their own to place
the rocm dependencies in a peer directory to the ollama executable
much like our windows install flow.
The recent ROCm change partially removed idempotent
payloads, but the ggml-metal.metal file for mac was still
idempotent. This finishes switching to always extract
the payloads, and now that idempotentcy is gone, the
version directory is no longer useful.
This refines where we extract the LLM libraries to by adding a new
OLLAMA_HOME env var, that defaults to `~/.ollama` The logic was already
idempotenent, so this should speed up startups after the first time a
new release is deployed. It also cleans up after itself.
We now build only a single ROCm version (latest major) on both windows
and linux. Given the large size of ROCms tensor files, we split the
dependency out. It's bundled into the installer on windows, and a
separate download on windows. The linux install script is now smart and
detects the presence of AMD GPUs and looks to see if rocm v6 is already
present, and if not, then downloads our dependency tar file.
For Linux discovery, we now use sysfs and check each GPU against what
ROCm supports so we can degrade to CPU gracefully instead of having
llama.cpp+rocm assert/crash on us. For Windows, we now use go's windows
dynamic library loading logic to access the amdhip64.dll APIs to query
the GPU information.
* read iogpu.wired_limit_mb on macOS
Fix for https://github.com/ollama/ollama/issues/1826
* improved determination of available vram on macOS
read the recommended maximal vram on macOS via Metal API
* Removed macOS-specific logging
* Remove logging from gpu_darwin.go
* release Core Foundation object
fixes a possible memory leak
This wires up some new logic to start using sysfs to discover AMD GPU
information and detects old cards we can't yet support so we can fallback to CPU mode.
At least with the ROCm libraries, its possible to have the library
present with zero GPUs. This fix avoids a divide by zero bug in llm.go
when we try to calculate GPU memory with zero GPUs.
We build the GPU libraries with AVX enabled to ensure that if not all
layers fit on the GPU we get better performance in a mixed mode.
If the user is using a virtualization/emulation system that lacks AVX
this used to result in an illegal instruction error and crash before this
fix. Now we will report a warning in the server log, and just use
CPU mode to ensure we don't crash.
Fix an ordering glitch of dlerr/dlclose and add more logging to help
root cause some crashes users are hitting. This also refines the
function pointer names to use the underlying function names instead
of simplified names for readability.
This adds additional calls to both CUDA and ROCm management libraries to
discover additional attributes about the GPU(s) detected in the system, and
wires up runtime verbosity selection. When users hit problems with GPUs we can
ask them to run with `OLLAMA_DEBUG=1 ollama serve` and share the results.
Update gpu.go initGPUHandles() to declare gpuHandles variable before
reading it. This resolves an "invalid memory address or nil pointer
dereference" error.
Update dyn_ext_server.c to avoid setting the RTLD_DEEPBIND flag under
__TERMUX__ (Android).
The memory changes and multi-variant change had some merge
glitches I missed. This fixes them so we actually get the cpu llm lib
and best variant for the given system.
This reduces the built-in linux version to not use any vector extensions
which enables the resulting builds to run under Rosetta on MacOS in
Docker. Then at runtime it checks for the actual CPU vector
extensions and loads the best CPU library available
In some cases we may want multiple variants for a given GPU type or CPU.
This adds logic to have an optional Variant which we can use to select
an optimal library, but also allows us to try multiple variants in case
some fail to load.
This can be useful for scenarios such as ROCm v5 vs v6 incompatibility
or potentially CPU features.
* increase minimum cuda overhead and fix minimum overhead for multi-gpu
* fix multi gpu overhead
* limit overhead to 10% of all gpus
* better wording
* allocate fixed amount before layers
* fixed only includes graph alloc
When there are multiple management libraries installed on a system
not every one will be compatible with the current driver. This change
improves our management library algorithm to build up a set of discovered
libraries based on glob patterns, and then try all of them until we're able to
load one without error.
If you attempt to run the current CUDA build on compute capability 5.2
cards, you'll hit the following failure:
cuBLAS error 15 at ggml-cuda.cu:7956: the requested functionality is not supported
* select layers based on estimated model memory usage
* always account for scratch vram
* dont load +1 layers
* better estmation for graph alloc
* Update gpu/gpu_darwin.go
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
* Update llm/llm.go
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
* Update llm/llm.go
* add overhead for cuda memory
* Update llm/llm.go
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
* fix build error on linux
* address comments
---------
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
If we try to load the CUDA library on an old GPU, it panics and crashes
the server. This checks the compute capability before we load the
library so we can gracefully fall back to CPU mode.