GitLab

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.

Avoid excessive log spew and make consistent with chat logging

Github actions matrix strategy can't access env settings

update llama3.2 vision memory estimation

The runtime and management libraries may not always have
identical ordering, so use the device UUID to correlate instead of ID.

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

Until we have full NUMA support, this adjusts the default thread selection
algorithm to count up the number of performance cores across all sockets.

-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>

This will no longer error if built with regular gcc on windows.  To help
triage issues that may come in related to different compilers, the runner now
reports the compier used by cgo.

* Remove llama.cpp submodule and shift new build to top

* CI: install msys and clang gcc on win

Needed for deepseek to work properly on windows

* windows: Support alt install paths

Advanced users are leveraging innosetup's /DIR switch to target
an alternate location, but we get confused by things not existing in the LocalAppData dir.
This also hardens the server path lookup code for a future attempt to unify with a ./bin prefix

* Fit and finish improvements for windows app

Document alternate install location instructions for binaries and model.
Pop up progress UI for upgrades (automatic, with cancel button).
Expose non-default port in menu to disambiguate mutiple instances.
Set minimum Windows version to 10 22H2

* Switch over to clang for deepseek on windows

The patch for deepseek requires clang on windows. gcc on windows
has a buggy c++ library and can't handle the unicode characters

* Fail fast with wrong compiler on windows

Avoid users mistakenly building with GCC when we need clang

Llama.cpp sometimes returns NULL as a return value to report an
error. We should explicitly check for this and convert it to a Go
error rather than putting NULL in our data structures and waiting
for it to blow up later.

On windows compiled with gcc the c++ regex library failed to handle
the characters

* Better support for AMD multi-GPU

This resolves a number of problems related to AMD multi-GPU setups on linux.

The numeric IDs used by rocm are not the same as the numeric IDs exposed in
sysfs although the ordering is consistent.  We have to count up from the first
valid gfx (major/minor/patch with non-zero values) we find starting at zero.

There are 3 different env vars for selecting GPUs, and only ROCR_VISIBLE_DEVICES
supports UUID based identification, so we should favor that one, and try
to use UUIDs if detected to avoid potential ordering bugs with numeric IDs

* ROCR_VISIBLE_DEVICES only works on linux

Use the numeric ID only HIP_VISIBLE_DEVICES on windows

If we're not writing out to a terminal, avoid setting the console mode
on windows, which corrupts the output file.

The common src/hdr defs should be in the common definitions, not gpu specific.

This unfies the rocm/cuda dependency logic into the makefile
and fixes a missing define which broke windows rocm

---------
Co-authored-by: Bill Wang <bill.wang@bill.wang>

Use cosine similarity to make the embeddings tests more robust

On windows ensure windows version define is properly set for rocm.
Remove duplicate rocm arch flags.
Resolve wildcards in the targets so parallel builds don't race.
Use readlink to resolve rocm dependencies since wildcards omit libelf
Keep windows rocm deps aligned with unified packaging model

We check for partial unicode characters and accumulate them before
sending. However, when we did send, we still sent each individual piece
separately, leading to broken output. This combines everything into
a single group, which is also more efficient.

This also switches to the built-in check for valid unicode characters,
which is stricter. After this, we should never send back an invalid
sequence.

Fixes #7290