Although the vision component of multimodal models typically already
call the optimized nn.Attention, it is converted into non-fused
operations. That is because the backend-specific fused kernels may
have requirements, such as padding, and they is performed by the
cache, which vision encoders don't use.
This implements a fallback path in the backend, softening the
requirements into optimizations. In turn, this allows flash attention
to be used for vision encoders, saving a significant amount of VRAM
and improving performance.
We currently use cache padding of 32 when not using flash attention
and 256 with flash attention, which is based on the historic alignment
requirements of these kernels. The restrictions have since been
loosened but there are still performance benefits, such as better
CUDA graph reuse.
Since the requirement is no longer kernel-specific, set the padding
uniformly to 256, as llama.cpp has.
* cmd/bench: support writing benchmark output to file
This changes Ollama to allow the bench command to write benchmark
results to a user-specified output file instead of stdout when the
--output flag is provided.
---------
Co-authored-by: Patrick Devine <patrick@infrahq.com>
* Revert "vulkan: temporary cary of vulkan fixes (#12971)"
This reverts commit 3a9e8e9fd4.
* ggml update to b7087
* fix argsort on metal
* update to b7108
* fix bakllava regression
This model lacks the metadata for the projector type.
* update to b7209
* fix TopK perf
* only build arm code on arm
This fixes a bug where disabling thinking on deepseek-v3.1 did not stop the model from thinking.
When thinking is not defined it should not be sent to the server since this will cause error responses in some cases where the model does not support thinking. However if it is defined as false it should still be sent.
We now do a deeper probe of CUDA devices to verify the library version has
the correct compute capability coverage for the device. Due to ROCm also
interpreting the CUDA env var to filter AMD devices, we try to avoid setting
it which leads to problems in mixed vendor systems. However without setting
it for this deeper probe, each CUDA library subprocess discovers all CUDA GPUs
and on systems with lots of GPUs, this can lead to hitting timeouts. The fix is
to turn on the CUDA visibility env var just for this deeper probe use-case.
Model eviction happens when we have at least one other model
loaded and are unable to load all layers into VRAM. However, on
CPU-only systems we can never load layers into VRAM, so this
constantly triggered eviction.
Fixes#13227
This change:
* fixes rope scaling in the mistral converter
* updates ministral to include llama4 scaling
* includes a new ministral parser for parsing reasoning and tool calling
---------
Co-authored-by: jmorganca <jmorganca@gmail.com>
If the user has somehow installed another GGML based app which places a
ggml-base lib somewhere in their PATH, we can experience runtime problems
due to incompatibilities. This change adds a warning message if we detect
a ggml-base outside of our install location to aid in troubleshooting.
While processing the response stream during a chat or generation if an error is occurred it is parsed and returned to the user. The issue with the existing code is that this assumed the response would be valid JSON, which is not a safe assumption and caused cryptic error messages to be displayed due to parsing failures:
`invalid character 'i' looking for beginning of value`
This change updates the stream function to return the raw error string if it cant be parsed as JSON. This should help with debugging issues by making sure the actual error reaches the user.
The cuda_jetpack libs will enumerate discrete GPUs on SBSA systems
which leads to runtime failures of missing kernels. This fix
requires an exact match to enable jetpacks instead of relying on
enumeration to filter out supported libraries.
We currently copy data into the KV cache in contiguous buffers using
ggml_cpy(). ggml_set_rows() was introduced to allow scatter operation
so that contiguous buffers are no longer required. The direct primary
benefit of this is that we no longer need to perform defragmentation.
However, GGML recently removed an optimization for ggml_cpy() and
we picked it up in 544b673 "ggml update to b6840 (#12791)". This
caused a roughly 40% drop in token generation performance on CUDA
due to CUDA graphs no longer being used. By switching to
ggml_set_rows(), the original optimization is no longer necessary
and CUDA performance is restored.
Fixes#13112
GGML requires tensors to be contiguous for reshape and if
this is not the case, it will assert fail. Contiguous is an
expensive operation, so it's best to do it lazily when it is
actually required rather than ahead of time when it may not
be needed.
Calling abort on windows triggers the C++ runtime to attempt a debugger
attach, which causes the crashed runners to hang instead of exit, leading
to a timeout instead of a fast failure during discovery.
Void is an open source AI code editor and Cursor alternative that supports
Ollama. It's built on VS Code and allows users to connect directly to Ollama
for private LLM usage without going through a middleman backend.
Key features:
- Open source Cursor alternative
- Direct Ollama integration
- VS Code fork with full compatibility
- Agent mode and MCP support
- Works with any open source model
Fixes#12919
Signed-off-by: Samaresh Kumar Singh <ssam3003@gmail.com>
* build: optimize dockerfile context for iterating
This moves the copy of the source into the layer AFTER
doing software installs so we don't have to go through
the RPM install for cuda, etc. every time you touch a
source file.
* amd: implement linux sysfs based VRAM lookup
This adds a C++ implementation of sysfs DRM VRAM discovery
for more accurate free VRAM data on linux for AMD GPUs.
