GitLab

We do not need to bypass the prompt caching in the ollama runner yet, as
only embedding models needed to bypass the prompt caching. When embedding
models are implemented they can skip initializing this cache completely.

Currently there is a single context per sequence, shared all by
all multimodal inputs. Since we build a vision encoder graph per
image, with a large number of inputs we can eventually hit the
maximum number of graph nodes per context.

This changes to use a separate context for each image, ensuring
that available resource limits are consistent.

Models may require that a set of inputs all be processed as part
of the same batch. For example, if an image has multiple patches
with fully connected attention between them, we should not split
the batch in the middle of an image.

Fixes #9697

This commit refactors the LLM subsystem by removing internal subprocess
request and response types. It consolidates duplicate type definitions
across the codebase, moving them to centralized locations. The change also
standardizes interfaces between components, simplifies the ServerStatusResp
struct, and moves the ParseDurationMs function to a common package. This
cleanup reduces code duplication between different runner implementations
(llamarunner and ollamarunner).

This reverts commit c7eae586b899083acebcd9b3847b89ea78c2850c.

This is useful for a few things:
 - Work around bugs, such as having 2 images in one batch
 - Keep the image in a single batch for fully connected attention
 - Improve performance by not evaluating embeddings multiple times

The encoder cache needs to know the position of images in the input
stream so that it knows when to delete them. Previously images didn't
have a position, so we implied one by breaking batches before an
image and then assuming the image was in the first position. However,
multimodal objects are now given explicit positions in the input
stream, so we can use that instead.

Breaking batches was also a way to simulate a cross attention mask
for mllama. However, given that it only supports a single sequence
and a single image, this mask doesn't serve any real purpose.
Removing the batch break does not appear to affect the quality of
the output.

Most of this is simply moving the input data structures to a new
package to avoid import cycles.

It's ok to fail on startup but we shouldn't panic during runtime
based on user input. Downgrade the panic to a warning.

Debug logging of every token has previously caused test timeouts
on slower machines.

This change bring in various interface cleanups along with greatly improving the performance of the sampler.

Tested with llama3.2 on local machine.
Improves performance from ~ 70 tokens/s -> 135 tokens/s with topK(40) enabled.
Without topK performance is ~ 110 tokens/s

Various vision models have different requirements for how they
receive their inputs. For example:
 - Mllama wants images together with text and the image embeddings
   don't themselves have positions or get stored in the main KV cache
 - Llava-style models feed in embeddings similar to tokens and
   images correspond to a varying number of tokens in the cache.

In addition, the strategy for providing inputs must support batching
and multiple sequences, which are managed by the runner. At the same
time, we want to keep data handling fully in the model so that new
architectures are not bottlenecked by runner code which does not
understand their particular requirements.

This provides a method for models to edit the input stream so that
it meets their needs while still being in a format that the runner
understands. This allows the runner to avoid special processing
for different models.

In addition, this fixes a regression where non-vision models may
try to incorrectly interpret images.

We sometimes tokenize partial strings. For example, with
multimodal inputs, we split the input string around the images
and then tokenize each piece. In these cases, we should only add
the special tokens on the first piece.

- output backend system info when initializing the backend. this ensures
  this information is always present without needing to be called
  explicitly
- convert to structured logging
- enumerate devices rather than backends since devices are ordered
- track device indices grouped by device name

The GGML flash attention kernel has specific requirements for
padding and permutation. This adds support to the KV cache
for conforming to these requirements so that flash attention
can be enabled.

Flash attention can be used in the same situations as the llama
engine and is enabled by the user in the same way.

defer the cancel to guarantee it runs

As are adding support for weighted sampling we have seen some performance
regressions, bypassing the sampler logic for now and defaulting to greedy
until we can benchmark the new sampler logic.

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.

This provides integration with the new Ollama engine
(58245413 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

Sometimes the KV cache requires defragmentation even without
triggering the threshold heuristic. In this case, decoding
will not being able to find a KV cache slot. This is particularly
difficult for the caller to handle if it happens in between
ubatches. To avoid this, we should immediately trigger a defrag.

In addition, a heavily fragmented cache can require more than
max_moves to defragment. Currently, we stop when we hit the limit
but this can leave a cache that still does not have adequate space
even after defragmentation is triggered. Instead, we should do
multiple batches of processing until everything is complete.

Fixes #7949

* llama: wire up builtin runner

This adds a new entrypoint into the ollama CLI to run the cgo built runner.
On Mac arm64, this will have GPU support, but on all other platforms it will
be the lowest common denominator CPU build.  After we fully transition
to the new Go runners more tech-debt can be removed and we can stop building
the "default" runner via make and rely on the builtin always.

* build: Make target improvements

Add a few new targets and help for building locally.
This also adjusts the runner lookup to favor local builds, then
runners relative to the executable, and finally payloads.

* Support customized CPU flags for runners

This implements a simplified custom CPU flags pattern for the runners.
When built without overrides, the runner name contains the vector flag
we check for (AVX) to ensure we don't try to run on unsupported systems
and crash.  If the user builds a customized set, we omit the naming
scheme and don't check for compatibility.  This avoids checking
requirements at runtime, so that logic has been removed as well.  This
can be used to build GPU runners with no vector flags, or CPU/GPU
runners with additional flags (e.g. AVX512) enabled.

* Use relative paths

If the user checks out the repo in a path that contains spaces, make gets
really confused so use relative paths for everything in-repo to avoid breakage.

* Remove payloads from main binary

* install: clean up prior libraries

This removes support for v0.3.6 and older versions (before the tar bundle)
and ensures we clean up prior libraries before extracting the bundle(s).
Without this change, runners and dependent libraries could leak when we
update and lead to subtle runtime errors.

Closes #7627

When processing a prompt, we look for image tags of the form
[img-0], which are inserted by the Ollama server process.
However, this can cause errors if the original prompt has these
tags - typically an image not found error is returned.

This changes tag searching behavior to be similar to the 0.3.x
series, which will largely avoid these problems. However,they can
still happen when input text with these tags is used with image
models. The correct solution is to escape the tags but this is a
larger issue with special sequences in general so this is an
incremental fix that should avoid the problem for the majority
of cases.

This also makes it easier to truncate long inputs the same as
shifting but does not actually implement it. This type of
truncation has a trade off between quality and time to first
token.

If there are no avilable slots for new sequences then a request
will not be added to the processing queue but will continue on
to wait for a response that never comes. Besides never giving a
response to the request, this prevents the model from being
unloaded due to the outstanding request.

To prevent this, there are semaphores that prevent more requests
from being processed than there are slots - one in the Ollama
server and one in the runner.
 - The Ollama server one works but it is not designed to protect
the runner's data internal structures and the runner can return a
final response before clearing its data structures.
 - The internal runner semaphore has similar behavior where it
 can release the semaphore when it issues a response. This is
 wrong - it should only release the semaphore after it has
 cleared the data structure.

In addition, we should return an error if a slot is not found
rather than deadlocking in the event we ever get to this spot.

Fixes #7779

Users get confused by "Failed to acquire semaphore" error="context canceled"
messages in the logs, which are actually clients giving up.  While there could be
a legitimate hang bug in the system, sometimes this is just short client timeouts
with an overloaded system, so this should help users understand what's going on
better.

Previous versions of the runner would truncate inputs to the context
window before beginning processing. The main processing loop relied
on this behavior if the context needed to be shifted later (due to
token generation). If truncation did not occur then invariants
would be broken, causing crashes or infinite loops.

Later versions attempted to fix these bugs and make the logic less
subtle so that all inputs could be handled. Truncation was removed
to make things consistent.

However, truncation is much faster than processing and shifting, so
removing it caused performance problems when the input vastly exceeded
the context size. This restores the input truncation as a performance
optimization while keeping the more robust processing logic.

Fixes #7762

We need to track which tokens are in the cache ourselves. We currently
add tokens to the cache tracker when we add them to batch but they are
not actually in the cache until we call Decode. This can cause
confusion when we are shifting the cache.

Avoids "could not find a KV slot for the batch" issues.

Bug #7545

We try to recover from errors by dropping the tokens that caused the
problem and re-trying. However, dropping the tokens is not correct
and continuing often leads to infinite loops. To avoid, this we
end the sequence if such a condition is detected, which is also
surprising.

At this point, it is better to just report the error. This will make
it easier to find problems and the alternatives are perhaps even more
surprising to users.

This is not a very satisfactory solution either - we should isolate
the error and return it to the user without killing the whole process.
However, this is an incremental step and consistent with most other
failures (which either manifest as abort() or panic).

Fragmentation of the KV cache can occur due to cache shifting or
different sequences getting processed. Decode uses a heuristic to
decide if it should defrag. However, this heuristic isn't 100%
accurate, so decoding can sometimes fail by surprise.

For these cases, if decode indicates that there is no KV cache space,
we should defrag and then try again.

This doesn't have any impact currently because NUM_PARALLEL is forced
to 1 for embeddings, so both indicies will always be 0.

This is a partial revert of 8a35bb92
"runner.go: Increase survivability of main processing loop", removing
the panic handler.

Although we want to avoid errors taking down the runner, we also
should make the user aware of problems when they happen. In the
future, we can restructure things so both parts are true.

Currently, if an error occurs during the prep stages (such as
tokenizing) of a single request, it will only affect that request.
However, if an error happens during decoding, it can take down the
entire runner.

Instead, it's better to drop the tokens that triggered the error and try to
keep going. However, we also need to stop when we run out of tokens,
otherwise, this just causes an infinite loop. This is likely the cause
of at least some of the hanging issues that have been reported.

Bug #7573