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
* broad lint fixes to sidestep CI scope glitch
* runner: Remove CGO engines, use llama-server exclusively for GGML models
Remove the vendored GGML and llama.cpp backend, CGO runner, Go model
implementations, and sample. llama-server (built from upstream llama.cpp via
FetchContent) is now the sole inference engine for GGUF-based models.
(Safetensor based models continue to run on the new MLX engine.) This allows
us to more rapidly pick up new capabilities and fixes from llama.cpp as they
come out.
On windows this now requires recent AMD driver versions to support ROCm v7 as
llama.cpp currently does not support building against v6.
* llama/compat: load Ollama-format GGUFs in llama-server
Squashed from upstream/jmorganca/llama-compat on 2026-04-29.
Source tip: 0c33775d37.
Original source commits:
- 25223160d llama/compat: add in-memory shim so llama-server can load Ollama-format GGUFs
- 7449b539a llm,server: route Ollama-format gemma3 blobs through llama/compat
- 436f2e2b1 llama/compat: make patch-apply idempotent
- 8c2c9d4c8 llama/compat: extend gemma3 handler to cover 1B and 270M blobs
- 021389f7b llama/compat: shrink clip.cpp injection from 18 lines to 1
- 61b367ec2 llama/compat: shrink patch to pure call-site hooks (34 -> 20 lines)
- 36049361c llama/compat: simplify shim (gemma3-tested)
- 8fa664865 llama/compat: add qwen35moe text handler
- db0c74530 llama/compat: add qwen35moe vision (clip) support
- 2a388da77 llama/compat: split shared infra into a util TU
- 9a69a17dc llama/compat: document non-public API dependencies
- d0f38a915 llama/compat: add gpt-oss and lfm2 handlers
- 086071822 llama/compat: add mistral3 text handler (vision TODO)
- 63bde9ff7 llama/compat: add mistral3 vision (clip) support
- 3a57b89d5 llama/compat: apply LLaMA RoPE permute to mistral3 vision Q/K
- 99cb87439 llama/compat: add qwen35, gemma4, deepseek-ocr handlers
- 2c7850dba llama/compat: add nemotron_h_moe handler (latent FFN + MTP skip)
- 9e3b54225 llama/compat: add llama4 text + clip handlers
- 034fee349 llama/compat: add gemma4 clip handler (gemma4v projector)
- 9945c5a93 server: remove dhiltgen/* compat redirect table
- 5d4539101 llama/compat: rewrite gemma4 tokenizer model to BPE
- 7e0765327 llama/compat: add glm-ocr text handler + text-loader load-op hook
- f1bd1a25a llama/compat: add glm-ocr clip handler (glm4v projector)
- 4b5cf3420 llama/compat: collapse text-loader hook back to one new patch line
- eb4ecf4fc llama/compat: extend gemma4 clip handler to gemma4a (audio)
- a23a5e76f llama/compat: fix gemma4a per-block norm tensor mapping
- cd2dcaff4 llama/compat: add embeddinggemma handler
- 1ce8a6b26 llama/compat: add qwen3-vl + qwen2.5-vl handlers
- fd98ffa1e llama/compat: add gemma3n + glm4moelite handlers
- cc7bdf0bc llama/compat: handle null buft in maybe_load_tensor
- 0c33775d3 llama/compat: disable mmap when load_op transforms text-side tensors
* refine implementation
* ci: fix windows MLX build
* ci: fix windows llama-server build
* ci: fix windows rocm build
* ci: windows mlx tuning
Shorten long-tail on build, and get OllamaSetup.exe back under 2g limit
* ci: fix windows dependencies
* win: fix dependency gathering
* disable openmp
* win: arm64 cross-compile build
also DRY out CI steps
* scheduler improvements
* ci: improvements from #15982
* win: favor ninja for faster developer builds
* win: fix build
* win: fix arm64 cross-compile
* win: avoid spaces in compiler path
* misc discovery fixes, and bos handling
* lint fixes
* win: fix arm cross-compile build/CI bugs
* llama.cpp update
* win: handle multiple CRT dirs
* vulkan: add windows iGPU detection
* fix creation bugs for patched models, other refactoring work
* tune batch size for better performance
* ci and lint fixes
* fix repeat_last_n bug
* build: revamp build for better developer UX
* amd, sampler, qwen3next fixes
* version bump
* fix mlx build
* revamp GPU discovery
Scanning the output of llama-server is turning out to be too error prone across
llama.cpp updates, so this switches to a thin dynamic library load against the
bundled GGML libraries so more details can be gathered from the API.
* version bump
* missing file
* ci: fix cache miss on rocm build
* refine vulkan dep handling
* fix ps reporting bug on full GPU load
* improve cmake wiring for customized local builds
* version bump
* docker build arg cleanup
* improve windows exit error logs
* fix community gemma4 support and ci flakes
* fix mlx unit test
* tighten up ps logic to avoid double counting fit log lines
* version bump
* fix ps view for full gpu layer offload
* add MTP wiring for llama-server and create with GGUFs
* pick best template by capabilities
* version bump
* ci: harden apt repos
* remove unused cpu core discovery
* adjust batch default logic to reduce OOMs
* support larger tool calls
* fix audio support, template show
* qwen35 mtp patch support
* flesh out dtypes
* rocm deps
* version bump
* lint fix
* block broken gfx1150 on windows
* fix qwen3.5 moe mtp tensors in patch
* mmproj oom fallback and vulkan on by default
* qwen MTP compat fix
* version bump
* ci: fix WoA cross-compile
* ci: workaround ui tool in cross-compile
* version bump
* win: enable OpenMP for CPU builds
* build: improve developer UX
* ci: windows path workaround for CPU build
* win: fix WoA dependencies
* win: fix large offset reads for mmproj patched loads
* version bump
* fix vulkan dup detection
* add OLLAMA_IGPU_ENABLE and largely disable iGPUs by default
* opt-in MTP, win large offset, integraton fixes
* fix unit test scheduler interaction hang
* fix multi-gpu filtering
* version bump
* review comments
* fix thinking level
* fix linux rocm ordering and granite 3.3 template
* version bump
* ci fix - non-shallow MLX checkout
* bypass linux sysfs unit test on windows
---------
Co-authored-by: jmorganca <jmorganca@gmail.com>
* metal: harden for ggml initialization failures
ggml_metal_device_init performs a probe to verify the tensor API compiles. On
some systems this passes, even though kernel coverage isn't complete, which
results in a later crash when compiling the real kernels. This change adds a
single retry if any of the error strings match this failure mode to disable the
tensor API. It also hardens an error case in the Go initDevices to detect
device initialization failures and panic instead of crashing later on a nil
array entry.
Fixes#15734
* review comments
* review comments
cublasGemmBatchedEx fails during graph capture when pool allocations
return fake pointers. This is triggered when NUM_PARALLEL is greater
than 1 for models like gemma4 that use batched matmuls. Skip it
during reservation since the memory tracking is already handled by
the pool allocations.
Fixes#15249
* bench: add prompt calibration, context size flag, and NumCtx reporting
Add --num-ctx flag to set context size, and report NumCtx in model info
header. Calibrate tokens-per-word ratio during warmup using actual
tokenization metrics from the model, replacing the fixed 1.3 heuristic.
This produces more accurate prompt token counts for --prompt-tokens.
Also add fetchContextLength() to query running model context via /api/ps.
* integration: improve vision test robustness and add thinking tests
Add skipIfNoVisionOverride() to skip vision tests when OLLAMA_TEST_MODEL
is set to a non-vision model. Add Think:false to context exhaustion test
to prevent thinking models from using all context before the test can
measure it. Add third test image (ollama homepage) and replace OCR test
with ImageDescription test using it. Relax match strings for broader
model compatibility. Add TestThinkingEnabled and TestThinkingSuppressed
to verify thinking output and channel tag handling.
* gemma4: add Gemma 4 GGML model support
Add full Gemma 4 model family support (E2B, E4B, 26B MoE, 31B Dense)
for the GGML backend including text, vision, converter, parser, and
renderer.
Text model features:
- Sliding window + full attention with per-layer patterns
- KV sharing across layers with donor map
- Per-layer embeddings (PLE) with learned projections
- MoE routing with RMSNorm + learned scale
- Proportional RoPE with freq_factors for global attention
- Final logit softcapping
Vision model features:
- SigLIP vision encoder with 2D RoPE
- ClippableLinear with input/output clamping via packed v.clamp_data
- Adaptive average pooling with nMerge kernel
- Multi-modal projection with unweighted RMSNorm
Converter:
- Safetensors to GGUF with vision tensor renaming
- Fused MoE gate_up_proj splitting
- Vision patch embedding reshape (HF to Conv2D layout)
- Packed clamp data tensor for ClippableLinear bounds
- Proportional RoPE freq_factors generation
Also includes:
- BackendGet() on ml.Tensor for reading weight tensor data
- Q6_K CUDA get_rows kernel support
- MoE-aware ffn_down quantization layer counting
- Gemma4 parser with tool calling and thinking support
- Gemma4 renderer with structured tool format
- Architecture-based auto-detection of renderer/parser/stop tokens
- Integration test gemma4 model list additions
* gemma4: add audio support with USM conformer encoder
Add audio encoding for Gemma 4 using the USM conformer architecture:
- Converter: audio tensor mapping, SSCP/conformer/embedder name replacements,
softplus repacker for per_dim_scale, F32 enforcement for conv weights
- GGML backend: Conv1DDW and PadExt tensor ops
- Audio encoder: SSCP Conv2D, 12 conformer blocks (FFW + block-local
attention with relative position embeddings + LightConv1d + FFW),
output projection, audio-to-text embedding projector
- Audio preprocessing: WAV decode, mel spectrogram, FFT (pure Go)
- Model wiring: WAV detection, audio token handling, unified PostTokenize
Correctly transcribes "why is the sky blue" from test audio.
* integration: add gemma4 audio tests including OpenAI API coverage
Test audio transcription and response via the Ollama native API, plus
two new tests exercising the OpenAI-compatible endpoints:
- /v1/audio/transcriptions (multipart form upload)
- /v1/chat/completions with input_audio content type
All tests use capability checks and skip models without audio support.
* gemma4: add OpenAI audio API support and capability detection
- Add CapabilityAudio and detect from audio.block_count in GGUF
- Add /v1/audio/transcriptions endpoint with TranscriptionMiddleware
- Add input_audio content type support in /v1/chat/completions
- Add TranscriptionRequest/Response types in openai package
* gemma4: add audio input support for run command
- /audio toggle in interactive mode for voice chat
- Platform-specific microphone recording (AVFoundation on macOS,
PulseAudio/ALSA on Linux, WASAPI on Windows)
- Space to start/stop recording, automatic chunking for long audio
* gemma4: add transcribe command (ollama transcribe MODEL)
- Interactive mode with readline prompt and slash commands
- Non-interactive mode for piped audio or record-until-Ctrl+C
- Chunked streaming transcription for long recordings
- Word-wrapped output matching run command style
* gemma4: add parser, renderer, and integration test plumbing
* gemma4: fix renderer to emit BOS token
* gemma4: add OpenAI audio transcription API and input_audio support
* gemma4: update converter for new weight drop naming
* gemma4: add per_expert_scale to MoE router and fix moe_intermediate_size config
* gemma4: rewrite renderer to match HF Jinja2 template exactly
Fix 8 bugs found by building 55 reference tests verified against the
HF Jinja2 chat template (VERIFY_JINJA2=1 shells out to Python):
- Tool responses use separate <|turn>tool turns (not inline tags)
- Tool calls emitted before content in assistant messages
- Thinking content stripped from assistant history (strip_thinking)
- User, tool, and system content trimmed (template does | trim)
- Empty system message still emits system turn (check role, not content)
- Nested object properties rendered recursively with required field
- Array items specification rendered for array-type properties
- OBJECT/ARRAY type-specific rendering comma logic matches template
Also adds Required field to api.ToolProperty for nested object schemas,
replaces old gemma4_test.go with comprehensive gemma4_reference_test.go,
and commits the Jinja2 template as testdata for verification.
* gemma4: fix MoE fused gate_up split and multiline tool-call arg parsing
- Text MoE: split `ffn_gate_up_exps` into contiguous `[gate|up]` halves instead of stride-2 slices.
- Parser: escape control characters in `<|"|>...<|"|>` string literals when converting tool-call args to JSON.
- Fixes warnings like `invalid character '\n' in string literal` for multiline tool arguments.
- Add Gemma4 parser regressions for multiline tool-call args and `gemma4ArgsToJSON`.
* cmd: simplify audio input to dropped file attachments
* gemma4: use full SWA memory for better cache reuse
* gemma4: initialize clamps after backend load
* convert: align gemma4 audio tensor renames with llama.cpp
* Remove redundant comments in gemma4 vision model
* Format Gemma4 MoE block field alignment
* use 4096 kvcache.NewSWAMemCache
* convert: support new Gemma4 audio_tower tensor naming (#15221)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* fix integration test defaults for audio
* review comments and lint fixes
* remove unused audio/video files
---------
Co-authored-by: jmorganca <jmorganca@gmail.com>
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.
* 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
* Revert "add support for NVIDIA Nemotron 3 Nano"
This reverts commit e7d2ae9d69421012e9a8765c06a3fdf0e45b12f3.
* GGML update to 380b4c984
Remove MaskBatchPadding as GGML_KQ_MASK_PAD is no longer present (no
padding required)
* update to c45f89d55
* ec98e2002
solar pro needed more adjusting - needs verification
* review comments
* flash attn: add auto mode for llama engine
If the user does not specify fa in the environment, use auto-mode.
* review comments
* ensure kv cache quantized types have FA explicitly enabled
additional review comments
* feat: Bump llama.cpp to the latest master (17f7f4b)
This brings in significant improvements to prefill performance for all
models using the SSM_CONV and SSM_SCAN ops (granite4, jamba, falcon-h,
nemotron-h, Qwen3 Next) on Apple Metal.
See https://github.com/ggml-org/llama.cpp/pull/17876
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Update patches 1-4
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Update patches 5-12
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Update patches 13-18
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Update patch 20
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Update patches 21-31
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Sync vendored code
The two files I'm not sure about here are the swap from gemma3-iswa.cpp to
gemma3.cpp (I chose to include this because I think it's required), and the
inclusion of `ggml-zendnn.h` which I chose to omit.
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Although the vision component of multimodal models typically already
call the optimized nn.Attention, it is converted into non-fused
operations. That is because the backend-specific fused kernels may
have requirements, such as padding, and they is performed by the
cache, which vision encoders don't use.
This implements a fallback path in the backend, softening the
requirements into optimizations. In turn, this allows flash attention
to be used for vision encoders, saving a significant amount of VRAM
and improving performance.
We currently use cache padding of 32 when not using flash attention
and 256 with flash attention, which is based on the historic alignment
requirements of these kernels. The restrictions have since been
loosened but there are still performance benefits, such as better
CUDA graph reuse.
Since the requirement is no longer kernel-specific, set the padding
uniformly to 256, as llama.cpp has.
* Revert "vulkan: temporary cary of vulkan fixes (#12971)"
This reverts commit 3a9e8e9fd4.
* ggml update to b7087
* fix argsort on metal
* update to b7108
* fix bakllava regression
This model lacks the metadata for the projector type.
* update to b7209
* fix TopK perf
* only build arm code on arm
We now do a deeper probe of CUDA devices to verify the library version has
the correct compute capability coverage for the device. Due to ROCm also
interpreting the CUDA env var to filter AMD devices, we try to avoid setting
it which leads to problems in mixed vendor systems. However without setting
it for this deeper probe, each CUDA library subprocess discovers all CUDA GPUs
and on systems with lots of GPUs, this can lead to hitting timeouts. The fix is
to turn on the CUDA visibility env var just for this deeper probe use-case.
We currently copy data into the KV cache in contiguous buffers using
ggml_cpy(). ggml_set_rows() was introduced to allow scatter operation
so that contiguous buffers are no longer required. The direct primary
benefit of this is that we no longer need to perform defragmentation.
However, GGML recently removed an optimization for ggml_cpy() and
we picked it up in 544b673 "ggml update to b6840 (#12791)". This
caused a roughly 40% drop in token generation performance on CUDA
due to CUDA graphs no longer being used. By switching to
ggml_set_rows(), the original optimization is no longer necessary
and CUDA performance is restored.
Fixes#13112
GGML requires tensors to be contiguous for reshape and if
this is not the case, it will assert fail. Contiguous is an
expensive operation, so it's best to do it lazily when it is
actually required rather than ahead of time when it may not
be needed.
Calling abort on windows triggers the C++ runtime to attempt a debugger
attach, which causes the crashed runners to hang instead of exit, leading
to a timeout instead of a fast failure during discovery.
* build: optimize dockerfile context for iterating
This moves the copy of the source into the layer AFTER
doing software installs so we don't have to go through
the RPM install for cuda, etc. every time you touch a
source file.
* amd: implement linux sysfs based VRAM lookup
This adds a C++ implementation of sysfs DRM VRAM discovery
for more accurate free VRAM data on linux for AMD GPUs.
We currently assign model layers to GPUs according to free VRAM,
which assumes that GPU performance is roughly equal. This does not
work well for mixed dGPU and iGPU systems because iGPUs typically
use system memory which is large but their performance is slow.
This instead assigns layers to dGPUs first and then iGPUs.
In the future, this could be generalized to have a more fine grained
notion of GPU performance but dGPU vs. iGPU performance is the most
extreme.
We used to control the way that llama.cpp saw devices using
CUDA_VISIBLE_DEVICES or similar. This would ensure that the layers
offloaded to a device were actually the ones intended. This is
particularly important because we might reorder devices based on
free memory or performance.
When we started explicitly scheduling layers, this logic went
away but the llamarunner didn't have any way to set the correct
order of devices. This meant that the correct number of layers
would be assigned to a device but not necessarily the layers
that were expected. This change sets up the devices correctly
based on the offload information.
Daniel Hiltgen