The caches, the speculation binding, and the drafter were each built
lazily inside the first request: begin constructed the caches, and open
bound the cache partition and made a fresh drafter every time. All of
it is a property of the loaded model, so build it once at load.
newPrefixCache replaces the lazy construction in begin, speculation
binds when it is created, and the drafter splits the way speculation
does: a persistent mtpDrafter constructed at load opens each request's
mtpDraftSession, whose constructor syncs the pairing cursor to the
draft caches' restored offset.
Keeping the recurrent conv state small was handled unevenly: the committed
live state was recopied on every commit — wasted work on single-token
decode, where the window is already tiny — while boundary states captured as
snapshots could still be plain slices of the forward-sized convolution
buffer. A cached slice pins that whole buffer even though the trie's eviction
accounting only counts the slice's bytes, so recurrent cache memory piled up
across requests and eviction could never reclaim it.
Compact each boundary state to its real size once, where it is produced in
the conv wrapper, so live state and snapshots own only their own bytes and
eviction sees the true cost. Single-token decode leaves the already-tiny
window as a slice.
Fixes#16698
The MTP validation forward schedules a snapshot at every drafted token, which
made CausalConv1D re-run the depthwise conv once per segment to recover each
boundary's conv tail. A conv boundary state is just the trailing convTail
input positions, so run the conv once over the whole window and slice each
boundary tail from the shared buffer, removing the per-token conv launches.
MLX now requires a macOS 26.2 deployment target for NAX kernels. Ollama's Metal v4 build still targeted 26.0, so recent MLX bumps silently built mlx_metal_v4 without NAX kernels.
Qwen3.5/Qwen3-Next architecture strings contain the substring "qwen3", so the
broad qwen3 match claimed them for the generic parser and qwen3-coder
renderer, whose template doesn't frame the thinking block — an empty
<think></think> leaked into content and think=false was ignored. Match the
family first via isQwen35Family so the parser, renderer, and
thinking-capability checks share one variant list.
* llm: allow iGPU mmproj offload with fit padding
llama.cpp's fit pass sizes text-model placement before the multimodal projector is loaded. Ollama had been avoiding that risk on non-Metal iGPUs by disabling projector offload entirely, which forces CLIP onto CPU on GB10 and Strix Halo even when the projector has ample memory available.
Let integrated GPUs use the same projector-memory check as other GPUs. When projector offload is enabled, add the estimated projector memory plus the existing 1 GiB headroom to Ollama-owned LLAMA_ARG_FIT_TARGET so fit leaves space for the later projector allocation. If Ollama/device setup already supplied a fit target, add the projector pad to it. If the user set LLAMA_ARG_FIT_TARGET explicitly, leave it exactly as provided.
Fixes#16419
* review comments
This is a rewrite of the create functionality for the MLX engine.
The core idea behind the create functionality is to break the import/convert into a pipeline of distinct phases:
* Read (scan the safetensors directory for the various bits of metadata)
* Classify (determine what the import type)
* Plan (determine any transforms that need to be done)
* Write (transform any data as necessary and write out the blobs)
* Create the manifest
Each architecture has a "policy" which determines how to convert the model correctly. A number of different formats for safetensors are supported including:
* nvfp4 (two formats: model optimized, torch)
* fp8 datatypes (convert to mxfp8)
* standard bf16 based weights
A number of cleanups/simplifications have been done including:
* using the baked in names for the tensors instead of munging them into something else
* unified 3d expert tensors (instead of separate per expert tensors)
* fewer unnecessary transforms to the various tensors in a model (keep a model as close to the source as possible)
* unified capability checking
* draft model handling (for MTP) is done on the same path
Image generation has been intentionally removed.
Recent upstream Pascal kernel fixes let us compile native SM60/SM61 kernels again instead of relying on PTX JIT, so allow Flash Attention auto at runtime for CC 6.x devices.
Fixes#16591Fixes#16754
Use ggml_fopen for compat tensor reads so Windows paths with Unicode characters are converted through the same UTF-8-to-wide path as llama.cpp model loading.
Fixes#16493
The presets and docs had fallen out of sync with what our current ROCm versions on Linux and Windows actually support. We rely on Vulkan now to cover these older unsupported devices.
* discover: use the SBSA CUDA build on JetPack 7 (L4T r38+)
JetPack 7 supports SBSA-based CUDA, so the standard cuda_v13 build — shipped
in the base linux-arm64 package, and given the Orin arch (CC 8.7) in #16628
— runs on these devices.
JETSON_JETPACK=7 previously selected a nonexistent jetpack7 runner, so
runner.go skipped every CUDA library and discovery fell back to CPU. The L4T
releases JetPack 7 uses (r38 on Thor, r39 on Orin) also hit the unrecognized
branch, and install.sh warned the version was unsupported. Map JetPack 7+
(L4T r38 and newer) to cuda_v13 (returned as "" from cudaJetpack); no
Jetson-specific download is needed, so install.sh no longer warns.
Fixes#16602
* discover: fall back to standard CUDA when the JetPack runner is absent
Per review, drop the L4T-version mapping (in cudaJetpack and install.sh) and
instead clear the jetpack override in runner.go when the detected cuda_jetpack
runner isn't installed. Normal discovery then selects the standard cuda_v13
build, which supports Orin (CC 8.7) on JetPack 7.
This change allows .experts.gate_proj / .up_proj / .down_proj tensor names to each
be used for both quantized (i.e. nvfp4 and mxfp8) and non-quantized (bf16) models.
Previous to this only non-quantized models used that tensor naming scheme.
Parser.done() counted the tag's open/close characters ({}, []) without
tracking JSON string context, so a streamed tool call whose string
argument value contained a closing brace or bracket (e.g.
{"code": "if (x) { y }"}) was treated as complete too early and flushed to
the user as plain text instead of being parsed as a tool call.
findArguments() in the same file already tracks string context; apply the
same handling in done() so open/close characters inside string values are
ignored.
In the PS output, expose the scheduler selected size (clamped by model context size) instead of always reporting the model max context. This will help provide a hint to clients to keep the context size below this value to avoid paging and poor performance on smaller VRAM systems.
Our cuda_v12 build requires nvcc fatbin compression, which in turn requires driver 550 or newer. This change filters incompatible CUDA devices based on the runtime and driver version. This allows users to build from source with older toolkits to support older drivers.
Fixes#16449
* server: align generate with native chat templates
/api/generate rebuilt chat-like prompts through the Go template path even when the model selected its native GGUF Jinja chat template, so the same model rendered differently between generate and chat.
Route chat-like generate requests through the shared native chat preparation path, keep deprecated context and image handling working there, and keep explicit OLLAMA_GO_TEMPLATE overrides intact.
Fixes#16792
* review comments
Fall back to "{{ .Prompt }}" when lacking templates
* llm: fix ollama ps double-counting mmap'd weights on partial offload
With mmap enabled, llama-server reports each CPU_Mapped model buffer as the
file-offset span of its CPU-resident tensors. During partial offload that span
covers nearly the whole file because the first and last tensors stay on CPU, so
the parsed buffer sizes count the offloaded weights twice and ollama ps shows
roughly 2x the real size with a false CPU/GPU split. Model weights can never
exceed the model file on disk, so trim the excess over the file size from the
mmap-backed portion when computing MemorySize. This makes the reported size
independent of use_mmap; VRAM accounting and scheduler placement are unchanged.
* llm: exclude repacked model buffers from the mmap overlap trim
The trim that corrects mmap double-counting computed the overlap from all
model buffers, including real copies such as CPU_REPACK. On a CPU-only
repacked model that inflated the excess and trimmed the repack out,
undercounting by the repack size (llama3.2 reported ~1918 MiB instead of
~3218 MiB).
Compute the overlap from file-backed buffers only: mmap views and direct
device copies, whose spans can overlap the file on partial offload.
Repacked or host-pinned CPU copies are separate allocations that never
overlap the on-disk weights, so leave them intact. Adds a CPU_Mapped +
CPU_REPACK regression test and corrects the Metal case to the real total.
Bump MLX to the latest selected upstream ref and update the MLX/imagegen
wrappers and tests for the new API behavior.
Fix the CUDA MLX archive so runtime NVRTC kernels work after deployment:
package CUTE/CUTLASS headers, include the CUDA runtime header closure, and
stage a coherent CUDA-toolkit-matched CCCL tree instead of MLX's fetched CCCL
for CUDA payloads. The previous archive could build successfully but crash at
runtime due to missing or incompatible JIT headers.
Existing qwen2.5vl GGUFs can contain an empty qwen25vl.vision.fullatt_block_indexes array. The compat layer translated the projector metadata but left clip.vision.n_wa_pattern unset, causing llama-server to fail loading the CLIP model.
Default the runtime compat value to the standard Qwen2.5-VL pattern when the key cannot be derived, and make the converter emit the same default for nil or empty fullatt block metadata.
Fixes#16540
Jesse Gross