The route assembly in Handler lacked clear organization making it
difficult scan for routes and their relationships to each other. This
commit aims to fix that by reordering the assembly of routes to group
them by category and purpose.
Also, be more specific about what "config" refers to (it is about CORS
if you were wondering... I was.)
There are two benefits to doing this:
- Provide a library function that models can use, reducing code for
each model implementation
- Enables a single place to drop in optimized implementations of
attention based on the backend or other factors. One is provided for
GGML.
On CUDA this improves token generation rate by about 3%. It does not
have a significant effect on Metal.
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Currently Rows is called as the last step in a model computation
to get the values for the output tokens. However, if we move it
earlier in the process then we can trim out computations that
never get used. This is similar to how models are defined in
llama.cpp.
Changing the model definition in this way improves token generation
performance by approximately 8%.
We don't need to create and destroy the GGML scheduler for every
context. This introduces extra CPU overhead for every forward
pass and extra memory for contexts that don't actually get scheduled
(for example, KV caches). We can instead just have one scheduler
for the backend and reset it each time we call Compute.
This improves token generation performance by 1-2% and removes
scheduler create/destroy from profile traces.
Currently the following parameters are in the runner but not used:
- numGPULayers
- mainGPU
- threads
- tensorSplit
This passes them through to the backend, which is where they would
actually get used. However, the GGML backend does not yet do anything
with them.
Added unit tests to verify error handling behavior in the Client.stream and Client.do methods.
Tests cover various error scenarios including:
- Error responses with status codes >= 400
- Error messages with successful status codes
- Empty error messages
- Successful responses
clang outputs are faster. we were previously building with clang via gcc
wrapper in cgo but this was missed during the build updates so there was
a drop in performance
sapphire rapids has amx support but it ends up having a negative
performance impact.
emerald rapids also has amx support with a positive performance impact
however there's no reasonable way in ggml to differentiate between the
two. the impact is small (~6%) so disable amx entirely for simplicity
set owner and group when building the linux tarball so extracted files
are consistent. this is the behaviour of release tarballs in version
0.5.7 and lower
The previous commit fixed flickering in the progress bar itself. Cursor
flickering is harder to address.
Cursor flickering could be fixed by hiding the cursor altogether while
the progress bar is displayed. The downside of this is that if the
program is killed in such a way that it can't clean up its state, it
would leave the cursor invisible.
Instead, this commit introduces an output buffer. All of the escape
codes and content for a single progress update are written to a buffer,
which is then flushed to the terminal all at once. This significantly
decreases the time during which the terminal has seen the cursor-hiding
code but has not yet seen the cursor-showing code, thus minimizing (but
not 100% eliminating) cursor flickering.
For more context, see:
https://gitlab.gnome.org/GNOME/vte/-/issues/2837#note_2269501
Previous code cleared the display before writing new content, creating a
window where the terminal could (and in some cases did) render empty lines.
Instead, we now write new content over the old content, only clearing
the trailing end of lines for cases where the new line is shorter.
Fixes#1664
We currently print system info before the GGML backends are loaded.
This results in only getting information about the default lowest
common denominator runner. If we move up the GGML init then we can
see what we are actually running.
Before:
time=2025-02-14T11:15:07.606-08:00 level=INFO source=runner.go:935 msg=system info="CPU : LLAMAFILE = 1 | CPU : LLAMAFILE = 1 | cgo(gcc)" threads=24
After:
time=2025-02-14T11:16:02.936-08:00 level=INFO source=runner.go:935 msg=system info="CPU : LLAMAFILE = 1 | CPU : LLAMAFILE = 1 | CUDA : ARCHS = 890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | cgo(gcc)" threads=24
provides a better approach to #9088 that will attempt to
evaluate symlinks (important for macOS where 'ollama' is
often a symlink), but use the result of os.Executable()
as a fallback in scenarios where filepath.EvalSymlinks
fails due to permission erorrs or other issues
In some cases, the directories in the executable path read by
filepath.EvalSymlinks are not accessible, resulting in permission
errors which results in an error when running models. It also
doesn't work well on long paths on windows, also resulting in
errors. This change removes filepath.EvalSymlinks when accessing
os.Executable() altogether
This provides integration with the new Ollama engine
(5824541 next ollama runner (#7913)) and the rest of the Ollama
infrastructure such as the runner and Ollama server.
In addition, it also builds out the KV cache infrastructure to
support requirements of how Ollama runs models such as:
- Parallel processing
- Memory management for defragmentation and shifting
- Multi-modal modals
Both old and new engines continue to be supported. By default, only
the old engine is used. To enable the new engine:
Start the server with the OLLAMA_NEW_ENGINE environment variable set:
OLLAMA_NEW_ENGINE=1 ./ollama serve
Start a model that is supported by the Ollama engine. This one is Llama 3.1 8b Q4_K_M:
./ollama run jessegross/llama3.1
Special tokens are currently read as uint32 from the model metadata.
However, all other parts of the system (including the tokenizer) use
int32 to represent tokens so it is impossible to represent the high
portion of the unsigned range. For consistency and to avoid casts,
we should just use int32 everywhere.
Currently, if a model uses an interface for its data structures (as mllama
does) then the tensor data in the structs implementing that interface will
not get loaded.
We need to sync before retrieving data after async computation.
It is also important to ensure that the Go buffer is not moved by
the GC across function calls so we do a synchronous copy.
Passing in a Go buffer is not safe because the garbage collector could
free or move the memory while the context is still open. However, if
we pass in the size and a nil pointer then GGML will allocate it from
the C side.
Most tensor backends try to optimize performance by using a lower
precision for matmuls. However, some operations (such as kq) on
some models are sensitive to this and require full precision.
There are two cases where we may not have an output after computing:
- Prompt processing where the length of the input exceeds the batch
size
- Internal memory management operations such as cache defrag and shift
Currently there is a mixture of int and int64 used when dealing with
tensor dimensions and shapes, which causes unnecessary conversions -
they all should be the same type.
In general, most interfaces (such as Pytorch) use int64 for
generality but most implementations (such as CUDA) use int32 for
performance. There isn't much benefit to us to being more flexible
than the implementations we are likely to run on.
In addition, as a practical matter, a model with a tensor with a single
dimension larger than 32 bits is unlikely to run on a 32-bit machine.
It is not common to return errors with close/free operations - most
people won't check it and even if they did there's probably not much
that can do. It's better to not give implementations false expectations.
feat: add new Ollama engine using ggml through cgo
This change introduces a new way to run pretrained models. It introduces 3 high level interfaces and a bunch of smaller helper interfaces to facilitate this.
- `model.Model` defines the interface for a model architecture. Models such as `llama` and `mllama`, which are provided as examples, can implement the model's forward propagation in the `Forward` method. This method will be called to generate completions. This interface can be found in `model/model.go`
- `ml.Backend` defines the interface for a backend tensor library, in this case `ggml`. Among other things, a Backend is responsible for loading a pretrained model into hardware (GPU, CPU, etc) and providing an interface for Models to access loaded tensors. This interface can be found in `ml/backend.go`
- `ml.Tensor` defines the interface for a tensor and tensor operations
This is the first implementation of the new engine. Follow up PRs will implement more features:
- non-greedy sampling (#8410)
- integration with Ollama and KV caching (#8301)
- more model support (#9080) with more coming soon
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
