It's possible to get prompts that consist entirely of whitespace -
this is most likely to happen when generating embeddings. Currently,
we will trim this away, leaving an empty prompt, which will then
generate an error.
Generating embeddings from whitespace should not trigger an error,
as this may break pipelines. It's better to just leave the whitespace
in place and process what we are given. This is consistent with
past versions of Ollama.
Bug #7578
NUM_PARALEL is currently enforced by the Ollama server process - it
will only issue requests to the runner if the maximum number of
concurrent requests has not been exceeded. Although this should
be sufficient, it is good for the runner to protect its own data
structures. Currently, if too many requests get through to the
runner, they will just get stuck and never return.
This may help with reports of Ollama hanging, though it is unclear
how it would actually occur.
Bug #7573
Docker uses the container filesystem for name resolution, so we can't guide users
to use the name of the host group. Instead they must specify the numeric ID.
The structure of the accounting for KV cache shifting was carried
over from the old runner but it now doesn't feel natural with the new
runner. There are a number of invariants that should hold true but
are difficult to reason about. There is at least one bug report
that would imply that the invariants are not holding.
This reduces the number of implicit assumptions and is more forgiving
of unexpected situations. It also improves behavior around which input
tokens are kept when truncation occurs.
Bug #7545
The Go runner does not have a problem with supporting parallel
requests for most multimodal models. Now that we won't be potentially
falling back to server.cpp, this restriction can be lifted.
However, the new mllama model can't support parallel requests, so we
will need to keep a restriction for that.
One potential failure mode is an empty file which bubbles up as an EOF error,
leading to all pulls and listing operations failing. Instead, continue and
warn about the corrupt manifest. This also allows re-pulling the corrupt
manifest to repair the system.
Currently we assume that images take 768 tokens of context size for
the purposes of clipping old messages that exceed the context window.
However, our mllama implementation stores the full image embedding
in a single token. As a result, there is significant waste of context
space.
Ideally, we would handle this more generically and have the
implementation report the number of tokens. However, at the moment
this would just result in a similar set of 'if' conditions in the
runner plus APIs to report it back. So for now, we just keep this
simple.
This leverages caching, and some reduced installer scope to try
to speed up builds. It also tidies up some windows build logic
that was only relevant for the older generate/cmake builds.
Check for NULL return values from llama.cpp in more places and
convert them into Go errors, which should make debugging easier
in the future rather than having hidden surprises in our data
structures.
Mllama has large embeddings (100 MB per image) and each embedding is
represented as 1 token when passed to llama.cpp. Batches are pre-
allocated for the size of the tokens times the batch size, so this
results in allocations of over 50 GB at the default batch size.
On some systems, these mallocs will fail.
Since an image is represented as a single token and mllama doesn't
support more than 1 image per request, we only need to allocate a
batch size of 1, which is much more reasonable. In addition, for
non-multimodal models, we don't need to allocate the embedding
batches at all.
Fixes#7464
Currently if an input has embeddings at any point then we will set
cross attention to true from the beginning. This means that any
tokens before the embeddings are sent will incorrectly have cross
attention layers applied.
This only sets cross attention when we have an embedding, either
previously in this sequence or in the cache. It also makes cross
attention capable of supporting parallelism at the runner level,
though the mllama implementation doesn't support that yet.
* Give unicode test more time to run
Some slower GPUs (or partial CPU/GPU loads) can take more than the default 30s to complete this test
* Give more time for concurrency test
CPU inference can be very slow under stress
-Update mllama to take the cross attention state as embeddings in
a batch, more similar to how Llava handles it. This improves
integration with the input cache.
-Pass locations in a prompt for embeddings using tags similar to Llava.
-Abstract interface to vision models so the main runner accesses Clip
and Mllama similarly
Co-authored-by: Michael Yang <mxyng@pm.me>
Jesse Gross
Michael Yang
Blake Mizerany
Daniel Hiltgen
Joey Zheng
Evan
Prasad Bhalerao
Ivo Stoykov
frances720
Arhan Busam
Edward J. Schwartz
RAPID ARCHITECT
Med Marrouchi
suncloudsmoon