* parsers/ministral: fix nested tool call parsing by counting brace nesting
* fix lint error
* parsers: refactor ministral parser
The old one was very tied to expecting to see only one token at a time,
which I don't like to assume (who knows what the future might hold wrt
speculative decoding, etc). This new one follows a similar structure to
qwen3-coder's parser, which incidentally makes it easier to test as well
(since we can test the individual events that come out when given
particular inputs).
---------
Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
Use the original key dimension (qkNopeHeadDim + qkRopeHeadDim = 256) for
the attention scale instead of the MLA absorbed dimension (kvLoraRank +
qkRopeHeadDim = 576).
MLA absorption is a mathematically equivalent reorganization of the
attention computation - it should not change the effective attention
scale. The scale should match training, which uses 1/sqrt(256).
This improves tool calling and model looping issues.
CGO_CFLAGS and CGO_CXXFLAGS were being set without optimization flags,
which overrides Go's default -O2 and results in unoptimized C++ code.
This caused significant performance degradation in release builds
compared to local `go build` which uses the default optimization.
- build_darwin.sh: add -O3 to CGO_CFLAGS and CGO_CXXFLAGS exports
- Dockerfile: preserve CGO_CFLAGS/CGO_CXXFLAGS from build args instead
of overwriting them
- app/README.md: update documentation to include -O3
* update README ruby link
the ollama-ai ruby gem is vastly less popular and seems unmaintained
https://rubygems.org/gems/ollama-ai
the defacto standard with the most downloads in the ruby ecosystem is ruby_llm
https://rubygems.org/gems/ruby_llm
I would link to that to avoid complication and guarantee feature compatibility with ollama.
* Update gem link ruby_llm from website to GitHub
ollama links mostly to github, not project websites, hence link to ruby_llm github.
Use nthreads=128 for ncols=4 configurations in flash attention tile
kernel to reduce shared memory usage below 48KB limit on Maxwell
architectures (sm_50/52).
With nthreads=256 and ncols=4, np=2 which caused shared memory to
exceed 48KB. With nthreads=128 and ncols=4, np=1 keeps shared memory
under the limit.
The nvidia_fp32 config for (576, 512) head sizes had nbatch_fa=32,
which caused zero-sized arrays when computing array dimensions:
nbatch_fa / (np * warp_size) = 32 / (2 * 32) = 0
This resulted in CUDA compilation failures on CUDA 12 (Windows and
Linux arm64):
- "static assertion failed with nbatch_fa % (np*warp_size) != 0"
- "the size of an array must be greater than zero"
Fix by changing nbatch_fa from 32 to 64 for all (576, 512) configs
in the nvidia_fp32 function, matching the nvidia_fp16 and AMD configs.
- Fix panic in ollama show for image gen models (safe type assertion)
- Add vision capability for Flux2KleinPipeline models at create time
- Flatten transparent PNG images onto white background for better results
* model: add MLA absorption for glm4moelite
Split the combined KV_B tensor into separate K_B and V_B tensors
during conversion, enabling MLA (Multi-head Latent Attention)
absorption which compresses the KV cache for improved efficiency.
* ggml: enable MLA flash attention for GLM-4.7-flash
Add support for gqa_ratio 4 in MLA flash attention kernels. GLM-4.7-flash
uses head size 576 with gqa_ratio 4, which was previously only supported
for gqa_ratio 16 (DeepSeek).
Metal changes:
- Enable head size 576 for flash attention
- Increase simdgroups to 8 for large heads (>=512)
- Add case 8 kernel dispatch for 8 simdgroups
CUDA changes:
- Add gqa_ratio 4 support for head 576/512
- Add tile configs for (576, 512, 4) and (576, 512, 8)
- Add MMA config cases for ncols 4
- Add template instances for ncols2=4
* model: add compatibility validation for glm4moelite architecture
Remove static VRAM estimation (EstimateVRAM, CheckMemoryRequirements)
which wasn't helpful. Instead, report the actual tensor weight size
from the manifest for ollama ps.
- Remove memory estimation check from runner startup
- Remove EstimateVRAM, CheckMemoryRequirements, modelVRAMEstimates
- Add TotalTensorSize() to get actual weight size from manifest
- Use weight size for Server.vramSize instead of estimates
Note: This is better than showing 0 or inaccurate estimates, but the
weight size is a drastic underestimation of actual memory usage since
it doesn't account for activations, intermediate tensors, or MLX
overhead. Future work should query real-time memory from MLX
(e.g., MetalGetActiveMemory) for accurate reporting.
Remove the Qwen image generation and image editing model packages
to clean up the codebase. These models will be reintroduced later.
- Delete x/imagegen/models/qwen_image/ (10 files)
- Delete x/imagegen/models/qwen_image_edit/ (5 files)
- Remove related CLI flags and imports from cmd/engine/main.go
- Update comments in cache/step.go to remove Qwen-specific references
For `/api/show`, a fully missing `model_info` field trips up various
integrators (including a recent Android Studio integration).
The primary source of missing info tends to come from models with a
remote that are also missing other data. It seems better to me to return
an empty `model_info` than making up some other fields within
`model_info` (like saying the architecture is `remote` or something like
that). So this does slightly change `/api/show`'s behavior that possibly
someone is relying on, but it seems more important to ensure the field
is always there (from a quick sampling integrations seem to be robust to
missing fields _within_ it).
Fixes: https://github.com/ollama/ollama/issues/13783
Move the unload check (empty prompt + KeepAlive=0) before the image
generation model dispatch in GenerateHandler. This prevents models like
flux from being loaded into memory just to be immediately unloaded when
running `ollama rm`.
Also fix a bug in DeleteHandler where `args[0]` was used instead of
`arg` in the delete loop, causing only the first model to be unloaded
when deleting multiple models.
Add --quantize fp4 support to ollama create for image generation models
(flux2, z-image-turbo), using MLX's affine 4-bit quantization.
Changes:
- Add fp4 to validation in CreateImageGenModel
- Add FP4 case to quantizeTensor (group_size=32, bits=4, affine mode)
- Add GetQuantization() to WeightSource interface for dynamic params
- Update LoadLinearLayer to use quantization params from model metadata
The loadImageGen function was not setting Options on the runnerRef,
causing needsReload() to always return true (since it checks if
runner.Options == nil). This resulted in the image generation
subprocess being killed and restarted for every request.
Simplify Nemotron3NanoParser by delegating tool call parsing to
Qwen3CoderParser instead of duplicating the parsing logic. The
Nemotron parser now only handles the thinking state machine and
transitions to Qwen3CoderParser for content and tool call parsing.
This also fixes an issue where tool calls without </think> would
cause the parser to get stuck in thinking mode.
* MLX - dynamic loading of mlx-c
Create a wrapper layer to indirect the dependency on mlx-c so
the main ollama binary does not have a load-time dependency on mlx-c, mlx, and on linux, cuda. Lazy load the library via dlopen
so we can adjust the path to ensure the dependencies are found
and fail gracefully if not present.
* review comments
* fix broken tests
* x: make `ollama create --experimental` import from safetensors
This change allows pulling in safetensors models into the new experimental model format, and also
fixes the `ollama show` command to be able to correctly display the model information.
* gofumpt the linter
* gofumpt the linter again
* validate the model name
Added validation to ensure auth redirects stay on the same host as the original request. The fix is a single check in getAuthorizationToken comparing the realm URL's host against the request host. Added tests for the auth flow.
Co-Authored-By: Gecko Security <188164982+geckosecurity@users.noreply.github.com>
* gofmt
---------
Co-authored-by: Gecko Security <188164982+geckosecurity@users.noreply.github.com>
