image processing
Update model.go
Update model.go
Update model.go
no projector
no projector
vision model scaffold
...
...
wip
...
rebase
fix patch merger
tidy
...
Update model_vision.go
server: do not attempt to parse offset file as gguf
This logic was causing issues for me when importing a gguf that had some padding at the end of the file. The valid gguf would be read, but then it would try to read the offset as a different gguf file. This does not seem right.
Update process_image_test.go
apply norm
prompt processing
prompt processing
fix post tokenize
fix gguf padding + populate the split patch embeddings
...
...
another shot at patch embeddings
...
patch embedding
Update model_vision.go
split pixels
This commit refactors the Rotary Position Embedding (RoPE) implementation across the codebase to use a structured configuration approach instead of individual parameters.
Key changes:
- Add new RoPEConfig struct with fields for dimension, type, base frequency, and scaling
- Add RopeType enum to formalize different RoPE implementation variants
- Add YarnConfig struct and related configuration for YaRN (Yet Another RoPE extensioN) context extension
- Update RoPE method signature across all tensor interfaces and implementations
- Refactor all model implementations (llama, gemma2, gemma3, mllama) to use the new configuration structure
This change improves code organization, makes the RoPE configuration more explicit, and provides better support for different RoPE variants and context extension methods.
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
Bruce MacDonald
Michael Yang
Michael Yang
Parth Sareen
Jesse Gross
Jeffrey Morgan