If a model with a split vision projector is loaded in the Ollama
engine, the projector will be ignored and the model will hallucinate
a response. Instead, fallback and try to load the model in the llama
engine.
New memory estimates (see #11090 for more information) are now
enabled automatically for all models running on the Ollama engine,
improving both stability and performance through more accurate sizing
and allocation. Models running on the llama engine will continue to
use the original style of memory estimation.
Our new engine implementation of gemma2 doesn't support flash
attention, which means that it also doesn't support KV cache
quantization. Currently, it is possible to turn these two on,
which will result in a crash.
If flash attention is enabled without KV cache quanitization, we will
currently always get this warning:
level=WARN source=server.go:226 msg="kv cache type not supported by model" type=""
* Add support for upcoming NVIDIA Jetsons
The latest Jetsons with JetPack 7 are moving to an SBSA compatible model and
will not require building a JetPack specific variant.
* cuda: bring back dual versions
This adds back dual CUDA versions for our releases,
with v11 and v13 to cover a broad set of GPUs and
driver versions.
* win: break up native builds in build_windows.ps1
* v11 build working on windows and linux
* switch to cuda v12.8 not JIT
* Set CUDA compression to size
* enhance manual install linux docs
* tests: reduce stress on CPU to 2 models
This should avoid flakes due to systems getting overloaded with 3 (or more) models running concurrently
* tests: allow slow systems to pass on timeout
If a slow system is still streaming a response, and the response
will pass validation, don't fail just because the system is slow.
* test: unload embedding models more quickly
The context must always be able to store the current batch, so
if the user requests a small context then we should also shrink
the batch to match. This also fixes the TestLongInputContext
test on the new engine. (The old engine already has this behavior.)
This PR updates the memory size estimate logic to better handle recurrent and hybrid-recurrent models which are currently being badly overestimated because the default logic assumes full attention for all layers.
The logic for the sizing of the recurrent layers comes from the llama.cpp implementation
ggml_tensor * r = ggml_new_tensor_1d(ctx, type_r, hparams.n_embd_r()*mem_size);
ggml_tensor * s = ggml_new_tensor_1d(ctx, type_s, hparams.n_embd_s()*mem_size);
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
If a GPU's free memory is less than the reserved amount, we might get
an underflow. Since it is an unsigned uint64, we print this as a large
number rather than the more correct 0. This only affects logging, the
actual layout code already handles this correctly.
Bug #12138
* perf: build graph for next batch in parallel to keep GPU busy
This refactors the main run loop of the ollama runner to perform the main GPU
intensive tasks (Compute+Floats) in a go routine so we can prepare the next
batch in parallel to reduce the amount of time the GPU stalls waiting for the
next batch of work.
* tests: tune integration tests for ollama engine
This tunes the integration tests to focus more on models supported
by the new engine.
The recent memory management changes caused all GPUs to be visible
to the runner, regardless of whether they are ultimately used. This
caused CUDA devices to allocate a primary context (~300 MB VRAM) on
each GPU, for each model. This is unnecessary, so we can both avoid
touching GPUs that we exclude in the early stage of allocation and
freeing the memory for any that we touch but don't use.
The issue will continue to exist for the old engine, since it touches
all devices during initialization.
there's two bugs here.
1. the check for a layer id is incorrect and should be >= 0 since layer
0 is valid
2. if both tensors have an layer identifier, it will only compare the
layer id which will return 0 if the tensors are in the same layer.
instead it should fallback to comparing the full tensor name
The thinking parser will automatically transition to being a
pass-through if non-whitespace is seen before an opening tag. However,
we weren't clearing the buffer after the first non-whitespace input, so
in practice the first token would be emitted twice.
Added a test that demonstrated this, and then fixed the bug.
With old memory estimates, it's currently impossible to load more
than one model at a time when no GPUs are available. This is because
the check for whether we need to evict a model looks to see if all
layers of the new model can be loaded onto GPUs, which is never true
if there are no GPUs. Before the memory management changes, there
was a special code path for CPU-only systems.
This problem does not exist with new memory estimates.
Fixes#11974
0x007e is a tilde and was getting adjusted (+0x00a2) to 0x0120 in the
encode, but then in the decode it was getting adjusted down (-0x0100) to
0x0020. The boundary for the +0x00a2 case has been adjusted to fix this
Fixes: #11966
Flash attention kernels require the mask of the KV cache be a F16
rather than an F32. We can use the GGML operation ggml_cast to do
this rather than doing it ourselves, which allows reuse of a
preallocated buffer in the graph rather than allocating a new one
for each batch. This improves token generation performance with
flash attention by 10-30% (with gpt-oss). This also makes performance
with flash attention better than without it, as expected.
