Now that we have a built-in parser abstraction, which was introduced in
<https://github.com/ollama/ollama/pull/12248>, we can modify our harmony
parser to match this and then get rid of nearly all of the
harmony-specific logic in routes.go. We do have a small amount of
code that turns the parser on by default if the architecture matches and
no other built-in parser was provided.
The built-in parser interface was modified in order to handle harmony's
prefill and tool name translation requirements.
The format qwen3-coder uses is relatively unique, both in rendering and
in parsing. To implement parsing, I wrote a custom parser in similar
style to harmony. For the rendering, I found that the logic would be
much more difficult to follow in a template, so I introduced the concept
of a built-in renderer that uses go code, rather than a template to
generate prompts.
I set us up for future built-in parsers and renderers by making it so
they can be specified in a Modelfile like so:
```
RENDERER "qwen3-coder"
PARSER "qwen3-coder"
```
These need to be provided explicitly because the architecture alone is
not enough to understand what format the model expects to receive, and
what format we expect it to output (e.g., qwen3-coder is `qwen3moe`,
which includes other qwen3-family models as well)
I haven't converted harmony to be one of these "built-ins" yet, since
some of it is in flux with the changes @ParthSareen has been making to
move harmony to the runner. It is likely that many other built-ins will
need to move to the runner as well, but I'm able to slightly defer that
decision since qwen3-coder doesn't have thinking (and therefore doesn't
need to be in the runner to make structured outputs work). I expect to
unify harmony with this approach very soon.
Whether a particular model supports tools or thinking was previously
inferred from templates, but without a template we now also use the
parser itself to declare what it supports. If we have future models that
re-use the same parsing format, but have different capabilities, we'll
want to parameterize them and give them different names to be specified
as a `PARSER`.
Misc changes:
- I worked on the renderer by diffing outputs from the reference
implementation and ours. To make it easier to do this, I extended
<https://github.com/ollama/ollama/pull/11875> to also support
returning the prompt via the openai compat layer
* 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.
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
* TEMPORARY: Update the llama.cpp upstream to my fork's Granite Four branch
This will be redone once my branch is merged upstream in llama.cpp
* feat: Update all patches
There are a number that are no longer needed at all:
- 0003-embeddings: Embeddings entirely overhauled on master
- 0008-ensure-KV-cache-is-fully-defragmented: KV caching entirely
overhauled on master
- 0019-metal-add-mean-kernel-14267: Merged upstream
- 0020-CUDA-add-mean-operation-14313: Merged upstream
* feat: Sync llama.cpp and ggml
* fix: Update rsync-filter for all moved/new/removed files
* fix: Add files missing from sync
* fix: Update ggml rsync-filter for new ggml-cpu/arch subdirs
* fix: Add ggml files missing from sync
* fix: Narrow llama.cpp rsync-filter to not include mtmd main tool cpp files
* fix: Remove mtmd main cpp files
* fix: Add missing include in sampling_ext.cpp
* fix: Update llama.go to use mtmd instead of clip/llava
* fix: Add patch for mtmd_input_text
* chore: Ignore *.patched in the patch directory
* fix: Fix support for arch-specific ggml-cpu source files with new arrangement
In https://github.com/ggml-org/llama.cpp/pull/13892, all arch-specific
implementations were split out into a nested tree structure under
ggml-cpu/arch. This conflicts with standard CGO layout where all
arch-specific source files are expected to live in the same directory as
the parent go module and use suffixes based on GOOS and GOARCH. As such,
there were really two options for getting this to work:
1. Add a patch on top of the GGML sync to rearrange the files to match the
GO layout convention
2. Use CGO directives to conditionally include the nested source files in
the compilation units
This commit does (2) in order to minimize the set of changes needed on top
of the upstream file layout. To get this to work, there are two key things
needed:
1. In cpu.go, #cgo directives are added to explicitly set __${GOARCH}__ in
the preprocessor directives
2. In arch-impls.c|cpp, use an #ifdef | #elif defined | #endif chain to
explicitly include the .c|.cpp files for the given architecture from the
nested directory
* fix: Use mtmd_helper to correctly load the bitmap for the image
* fix: Apply patch for mtmd_text_input
* fix: Add missing stb to llama.cpp rsync-filter
* fix: Add sync'ed stb vendored header
* fix: Use c++17 and include vendor for go wrapper modules
* fix: Update patch 0015 for upstream implementation of uuid
* feat: Bump to the latest tip of the branch
* fix: Update patches for bump
* feat: Bump back to the cenral repo and point at the latest master
This includes granite 4 and a number of other model architectures!
* fix: Revert changes to ggml export GPU UUID patch
* fix: Add patch for GGML_VERSION and GGML_COMMIT constants
* feat: Sync all patched code
* build: Include cmake/common.cmake in ggml sync
* build: Add top-level include for GNUINstallDirs in CMakeLists.txt
This is used to populate CMAKE_INSTALL_BINDIR
* fix: Add a patch to avoid power throttling API on non-msvc windows builds
* fix: Sync patch changes for ggml-cpu.c
* feat: Bump llama.cpp to 4a4f42
This picks up support for Kimi K2 and PLaMO-2
* feat: Sync llama.cpp
* fix: Handle multi-chunk image encodings from mtmd
* fix: Re-number patches after merge with `main`
* feat: Bump to 41e78c in the makefile
* fix: Fix Solar and argsort/copy patches after bump
* fix: Remove Gemma3n CUDA Graphs patch
It was implemented upstream:
https://github.com/ggml-org/llama.cpp/pull/14741
* feat: Sync llama.cpp / ggml after latest bump
* build: Remove unnecessary CFLAGS definitions in cpu.go
* fix: Remove unnecessary additions in the rsync-filter
* fix: Remove unused vendored code for chat template parsing
* Revert "fix: Remove Gemma3n CUDA Graphs patch"
This reverts commit d724caced3.
* fix: Update 0020 CUDA Graphs for gemma3n to keep both llama.cpp and ollama fixes
https://github.com/ollama/ollama/pull/11195#issuecomment-3137312394
* fix: Sync ggml-cuda.cu after keeping both style cuda graph fixes for gemma3n
* unwind mxfp4 patch
Prepare to bump ggml with their impl for mxfp4
* bump
* fix windows build error
* Convert tensors at load time
Repack the mxfp4 tensors as ggmls kernels expect them to be.
* convert mlp bf16 to f32
* buffer the conversion better
* reshape earlier
* openai swiglu
* add ids
* split qkv, gate_up
* fix nested alt tags
* fast attention
* remove debug messages
* fix lint
* remove redundant test
* remap values only if source/target are different
* add back i32->i32 copy
* refactor cpu quants
* clean up vendor
* update patch instructions
* clean up patches
* remove webgpu
* update mem
* also handle gpt-oss
* revert convert changes
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* bf16
* tests
* gpt-oss
* enable gptoss for engine
* rough estimate
* convert to mxfp4
* handle safetensors U8
* clamp glu/linear
* update tokenizer
* MXFP4 support
This implements the Open Compute Microscaling (MX) FP4 format
as a tensor type with backend implementations focusing
on mulmat and mulmatid on CPU, CUDA, and Metal.
* Unit tests for MXFP4 support
This exercises various operations and shapes on both CPU and GPU (if detected
on the system)
* cuda graph
* unit test adjustments
* cuda: optimize memory access
Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4
* mac: fix crash on old macos versions
cblas_sgemm is only supported on v13.3 and up, however bf16 is
only supported on v14+ so we were falling back to ggml-blas and
crashing on bf16 tensors. Checking for the function being null
seems to be the simplest way to condittionally avoid registering the
backend.
* server: Minimum context length for gptoss
This model requires a minimum context length of 8192 to function
effectively. Users can set higher values through all normal mechanisms
but lower values will be silently reset.
* ggml: Multiply by numParallel for gptoss sliding window
When computing the graph size estimate, the context size is already
multiplied by numParallel so estimates reflect that. However, since
sliding window models use a smaller, fixed context size, they need
to manually take numParallel into account.
* gpt-oss integration
includes harmony parser and thinking levels, etc.
* fix sync
* fix tests
* fix lint
---------
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
* Enable CUDA Graphs for gemma3n.
Similar to
https://github.com/ggml-org/llama.cpp/pull/14741,
though ollama has a slightly different model graph
than llama.cpp which requires different workaround
checks.
* Remove residual check by reshaping differently in gemma3n model
This should make the heuristics more robust
- Both `/api/generate` and `/api/chat` now accept a `"think"`
option that allows specifying whether thinking mode should be on or
not
- Templates get passed this new option so, e.g., qwen3's template can
put `/think` or `/no_think` in the system prompt depending on the
value of the setting
- Models' thinking support is inferred by inspecting model templates.
The prefix and suffix the parser uses to identify thinking support is
also automatically inferred from templates
- Thinking control & parsing is opt-in via the API to prevent breaking
existing API consumers. If the `"think"` option is not specified, the
behavior is unchanged from previous versions of ollama
- Add parsing for thinking blocks in both streaming/non-streaming mode
in both `/generate` and `/chat`
- Update the CLI to make use of these changes. Users can pass `--think`
or `--think=false` to control thinking, or during an interactive
session they can use the commands `/set think` or `/set nothink`
- A `--hidethinking` option has also been added to the CLI. This makes
it easy to use thinking in scripting scenarios like
`ollama run qwen3 --think --hidethinking "my question here"` where you
just want to see the answer but still want the benefits of thinking
models
FromFloatSlice and FromIntSlice return an error if the shape doesn't
match the passed data or if memory can't be allocated. Since these
are inputs, the memory being allocated is system memory rather than VRAM.
In many cases, the caller can't really handle the error and panics.
Empty and Zeros directly panic if they can't allocate memory.
This makes things consistent by panicing for the first two cases,
removing a fair amount of error handling code. This is also consistent
with how Go typically handles these situations.
* fix mllama convert
- transform attn_gate and ffn_gate
- swap attention heads for vision models
* fix mllama
the mlp gate which was applied in the wrong place
setting samebatch on the vision start token is problematic because it
will be shared with other inputs that also use images. this will cause
the input to be cached and the runner will not see SameBatch. SameBatch
will also be incorrect since it may be for a different image.
assigning samebatch to the input tokens resolves this by ensure it's
assigned correctly to inputs corresponding to the image.
not setting same batch correctly may cause panics during inference since
images are no longer guaranteed to be in the same batch.
Currently, when the backend is created, the tensors are loaded at the
same time, which is a slow operation. This separates them to be two
steps:
- Create backend, including enumerating tensors and memory allocation
- Loading tensor data
This allows more flexibility in managing model loading.
* get eos_token_id from generation_config.json
* refactor
* include both ids and strings in trace
* comments
* remove special case for gemma3 special vocab (#10743)
For some multimodal models (such as gemma3), we create a single
graph that generates the image embedding and then use this in the
text model. The embedding tensor is completely opaque to the runner.
However, this doesn't work if we need to use the embedding in multiple
batches. This can arise if the embedding is larger than the batch size.
In these cases (as with llama4), we would like to create views that
are more appropriately sized. However, if we do this then the original
source tensor is used in multiple graphs, which isn't allowed. To
avoid that problem, models with this pattern compute the embedding
tensor on first use and recreate the individual views. There is no
longer a single vision and text graph.
This codifies the pattern of separating vision and text graphs. The
logic of computing tensors on demand is moved to the runner, so models
no longer have to worry about this. It also gives the runner visibility
into the multimodal tensors, which is important for memory management.
Currently, the KV cache and graph are lazily allocated as needed.
The cache is fully allocated on first use of the corresponding
layer whereas the graph grows with the size of the context.
This can be an issue if another application allocates more VRAM
after we do our calculations - Ollama will crash in the middle of
inference. If we instead allocate the maximum needed memory at
startup of the runner, we will either succeed or fail at that point
rather than at some surprising time in the future.
Currently, this only generates a worst case batch for text, which
means that vision models may get a partial allocation and continue
to lazily allocate the rest.
Mistral is a popular research lab making open source models. This updates
the forward pass of llama architecture models to support both llama models
and mistral models by accounting for additional metadata present in mistral
models, and finding the correct dimensions for the output projection.
Rather than directly giving the input data to models, we can
pass a tensor instead. In the short term, this saves some duplicated
code.
Longer term, we will want to overlap setting up the next batch with
processing of the current one. In this case, we will only have the
shape of tensor but it will not be loaded with data at the time of
graph generation. By passing only a tensor to models now, we set up
this possibility and prevent them from relying on data that they won't
have in the future.
Although the same could be done for Positions and Outputs, in some
cases we either need the raw input data or don't use them at all.
Therefore, for now we leave them as they are and allow models to
convert them to tensors as needed.
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
Softcap isn't in the whitepaper/implementation for the language model so we should remove it. There is no discernible difference in output with it removed.
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.
use a similar strategy as llama.cpp for deciding where tensors should be
allocated. this will be improved later to be aware of usable memory
before assigning the tensor
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.
* Include unified vision layers in memory prediction
For newer vision models with a single gguf, include
the projection estimates.
* Adjust CLI to handle both styles of vision model metadata
* Wire up new tokenizers for new engine
If we're loading the new engine, utilize the new model
text processor instead of calling into cgo wrappers for
llama.cpp. This also cleans up some tech debt from the
older tokenization flow for the C++ server which was
no longer used.
This also adjusts the grammar handling logic to pass
through to the new engine instead of utilizing the cgo
schema to grammar call.
* Lay foundation for auto selection of new engine
Prior to performing attention, we need to permute query, key
and value. Currently we call Contiguous after each of these
permutations, which is correct but expensive. Avoiding the
3 calls to Contiguous increases performance by over 20%.
The permutations of query and key do not violate the continuity
rules for mulmat and the Contiguous call can be simply removed.
Value requires a different permutation and does require Contiguous.
However, we can use the copy into the cache as a way to perform this
without further overhead.
To support this and avoid unexpected tensor shapes that are seen by
models, we need tighter integration between attention, cache
and backend. Future optimization will also likely need this structure
- for example, flash attention has special padding requirements in
the cache and other backends may have their own needs.
This further contains the operations that go into attention so that
these and other optimizations can be handled transparently. Models
that have special requirements for attention can still implement
their own version of it.
update Context.Forward to accept multiple tensors to match
Context.Compute signature
update Context.Forward to return Context such that it can be chained
with Context.Compute
There are two benefits to doing this:
- Provide a library function that models can use, reducing code for
each model implementation
- Enables a single place to drop in optimized implementations of
attention based on the backend or other factors. One is provided for
GGML.
On CUDA this improves token generation rate by about 3%. It does not
have a significant effect on Metal.
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Currently Rows is called as the last step in a model computation
to get the values for the output tokens. However, if we move it
earlier in the process then we can trim out computations that
never get used. This is similar to how models are defined in
llama.cpp.
Changing the model definition in this way improves token generation
performance by approximately 8%.
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.
Devon Rifkin
Michael Yang
Daniel Hiltgen
Oliver Simons
Jeffrey Morgan
Jesse Gross
Bruce MacDonald
Michael Yang
Parth Sareen
Patrick Devine