Remove the x/imagegen tree (MLX image generation engine, Flux2/zimage
models, cache, C bindings) and all imagegen integration points:
- server: drop imagegen routes, scheduling, and generate handling
- api/cmd/docs: remove image generation API surface and docs
- middleware/openai: remove image endpoint support
- integration: remove imagegen test suites
- x/create: adopt the rewritten create pipeline from main; drop
imagegen create path (CreateImageGenModel, IsTensorModelDir,
model_index.json detection, Flux2KleinPipeline vision hack)
- retain x/imagegen/manifest (Ollama-store safetensors manifest
loader), still used by x/mlxrunner and x/create/client
- fix Windows MLX dl.dll install, MLX CMake version path, and the
show command after removing safetensors models
Align Qwen parser behavior with Transformers serve by allowing <tool_call> parsing while still in thinking collection.
Changes:
- qwen3vl: detect <tool_call> before </think> in thinking state and transition to tool parsing
- qwen3: same thinking-state tool detection and partial-tag overlap handling
- tests: update qwen3vl thinking/tool interleaving expectations
- tests: add qwen3 cases for tool call before </think> and split <tool_call> streaming
The recent change in #14322 added tryLoadByName() which attempts to
load libmlxc.dylib via rpath before searching directories. This is an
optimization for Homebrew installations where rpath is correctly set.
However, when rpath isn't set (which is the common case for app bundle
installations), dlopen fails and the CHECK macro prints an error to
stderr:
ERROR - dynamic.c:21 - CHECK failed: handle->ctx != NULL
This error is misleading because it's an expected failure path - the
code correctly falls back to searching the executable directory and
loads the library successfully. The error message causes user confusion
and makes it appear that something is broken.
Replace the CHECK macro with a simple return code so the C code fails
silently. The Go code already handles error logging appropriately:
tryLoadByName() fails silently (intentional fallback), while
tryLoadFromDir() logs via slog.Error() when explicit path loading fails.
Parse the default_num_ctx from the server's "vram-based default context"
log line and expose it through the inference compute API. This eliminates
duplicate VRAM tier calculation logic in the frontend.
- Add InferenceInfo struct with Computes and DefaultContextLength
- Rename GetInferenceComputer to GetInferenceInfo
- Handle missing default context line gracefully (older servers)
- Add DefaultContextLength to InferenceComputeResponse
- Update Settings UI to use server's default, disable slider while loading
- Add disabled prop to Slider component (grays out + hides handle)
- Migrate existing users with context_length=4096 to 0 (auto mode)
If a sequence is replaced in s.seqs while a batch is computing, the old logits can be decoded into the new sequence. This change rechecks the sequence pointer after compute and skips decoding for replaced entries, preventing stale results from being applied.
Change the truncation algorithm to start with all messages and remove
from the front until it fits, rather than adding messages one at a time
from the back. This reduces tokenization calls from O(n) to O(1) in the
common case where all messages fit in context.
When launching OpenClaw without prior onboarding, run the onboarding
wizard instead of going straight to gateway. This ensures proper
gateway configuration (mode, token, etc.) before first use.
- Add onboarded() to check for wizard.lastRunAt marker in config
- Run onboard with --auth-choice skip --gateway-token ollama for fresh installs
- Existing installs (onboarding completed) run gateway directly
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
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
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.
Add --norsrc flag to ditto commands when creating Ollama-darwin.zip
to exclude AppleDouble resource fork files (._* files) from the archive.
The mlx.metallib file has extended attributes, which causes ditto to
include a ._mlx.metallib AppleDouble file in the zip. Since this file
is not part of the code signature seal, macOS rejects the bundle during
auto-update verification with:
"a sealed resource is missing or invalid"
"file added: .../._mlx.metallib"
The --norsrc flag prevents ditto from preserving resource forks and
extended attributes, ensuring only signed files are included in the
release archive.
The CMake condition for installing mlx.metallib checks
CMAKE_OSX_ARCHITECTURES, but this variable is only set when explicitly
passed - not auto-detected. The arm64 build was missing this flag,
causing the metallib to not be installed, which then caused codesign
to fail on the unexpanded glob pattern.
- Install mlx.metallib for arm64 builds (required for Metal GPU acceleration)
- Apply rpath settings to all macOS builds, not just x86_64
- Add CMAKE_BUILD_WITH_INSTALL_RPATH to avoid install_name_tool errors
- Update build_darwin.sh to copy, sign, and package the metallib
TeaCache:
- Timestep embedding similarity caching for diffusion models
- Polynomial rescaling with configurable thresholds
- Reduces transformer forward passes by ~30-50%
FP8 quantization:
- Support for FP8 quantized models (8-bit weights with scales)
- QuantizedMatmul on Metal, Dequantize on CUDA
- Client-side quantization via ollama create --quantize fp8
Other bug fixes:
- Fix `/api/show` API for image generation models
- Server properly returns model info (architecture, parameters, quantization)
- Memory allocation optimizations
- CLI improvements for image generation
RemoveLayers was calling Manifests() for each layer to check if it was
shared with other models. For models with many blobs (e.g., tensor
models), this caused O(N*M) manifest reads.
Now loads manifests once and builds a set of in-use digests.
Removes 5-minute HTTP client timeout that caused "context deadline
exceeded" errors on large file downloads. Stall detection (10s)
already handles unresponsive connections.
Fixes progress bar total going down on resume by calculating total
from all blobs upfront and reporting already-downloaded bytes
as completed immediately.
Jeffrey Morgan