GitLab

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.

Model eviction happens when we have at least one other model
loaded and are unable to load all layers into VRAM. However, on
CPU-only systems we can never load layers into VRAM, so this
constantly triggered eviction.

Fixes #13227

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.

Originally, llamaServer represented old memory estimates, which
could be used with either the old or new engine. ollamaServer was
used only for the new estimates and new engine. Since these
implementations did not map directly to engine, there was engine-
specific code in common code paths.

Now that new estimates are always used for the new engine, there is
a direct mapping between server type and engine. This separates out
most of the engine-specific code into the correct implementation
to make things easier to understand.

Currently for both the old and new engines, there is code to
calculate how much memory is required for a model and lay out
the layers onto GPUs. This reuses the new engine's lay out code
for the old engine as well, bringing them closer together. The
old engine continues to use its current method of estimating
required memory.

This reduces maintainence effort and improves consistency, as new
features only need to be implemented in one place. The newer code
is also more accurate, especially with multiple GPUs.

Adds logprobs support to Ollama's API including support for Ollama's
OpenAI-compatible API. By specifying the new 'logprobs' boolean parameter
in the API, Ollama will return the log probabilities for each token generated.
'top_logprobs', an integer value can also be specified up to the value 20.
When specified, the API will also provide the number of most likely tokens to
return at each token position
Co-authored-by: Baptiste Jamin <baptiste@crisp.chat>

windows gets confused when we try to hand the stderr file descriptor to the subprocess children.  This ensures the log output
always shows up.

This reverts commit 5d347f6d.

Currently, checking the length of prompts for embeddings to ensure
they fit in the context window (and possible truncation) occurs in
two places - the Ollama server and runner. This can lead to
inconsistencies in both the checks and reported number of tokens
processed. Since we have to do this processing in the runner, this
consolidates all of the logic there.

If we create a memory layout that should fit based on report free VRAM
but allocation still fails, we start applying a backoff. This reduces
free VRAM by an exponential percentage (1%, 2%, 4%...). However, the
points chosen tend to be too dense at the beginning and too sparse at
the end. Therefore, this switches to an incremental backoff (10%, 20%,
30%...).

* DRY out the runner lifecycle code

Now that discovery uses the runners as well, this unifies the runner spawning code
into a single place.  This also unifies GPU discovery types with the newer ml.DeviceInfo

* win: make incremental builds better

Place build artifacts in discrete directories so incremental builds don't have to start fresh

* Adjust sort order to consider iGPUs

* handle cpu inference oom scenarios

* review comments

Currently, if you set num_gpu then this forces the model to
load with that number of layers in the current configuration.
This is done regardless of any other information, which means
that no eviction is performed even if another model is loaded.

This behavior is different from the old estimates (and still
happens for models that runs on the llama engine). In those
cases, models would be evicted if needed to load at the requested
number of layers. That behavior is more useful and less surprising,
so this changes the new estimates to match.

Fixes #12580

* implement the vulkan C backend

* add support in gpu.go

* add support in gen_linux.sh

* it builds

* fix segfault

* fix compilation

* fix free memory monitor

* fix total memory monitor

* update gpu.go

* fix build

* fix check_perfmon len

* remove cap_get_bound check

* fix vulkan handle releasing

* fix build on federa 40

* fix vulkan on windows

* making amdgpu work on arm achitecutre with vulkan

* add x86_64 lines in VulkanGlobs and capLinuxGlobs

* add aarch64 lines in vulkanGlobs and capLinuxGlobs

* Fix variable name

* Add vulkan build patch from @jmorganca

* Sync vendored ggml to add Vulkan support

* Updated dockerfile

https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2660836871

Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Installing rocm library
Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* This version works well

built based on this: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2660836871

Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Applied 00-fix-vulkan-building.patch

Work done by McBane87 here: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2660836871

Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Fixed the "detached head" issues
Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Merged in the right direction
Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Merging the latest stable (#2)

* Applied 00-fix-vulkan-building.patch

* Implemented vulkan backend based on the work done by whyvl, Dts0, McBane87 and others

Tested on AMD Ryzen 7 8845HS w/ Radeon 780M Graphics with ROCm disabled

```
[GIN-debug] POST   /v1/chat/completions      --> github.com/ollama/ollama/server.(*Server).ChatHandler-fm (6 handlers)
[GIN-debug] POST   /v1/completions           --> github.com/ollama/ollama/server.(*Server).GenerateHandler-fm (6 handlers)
[GIN-debug] POST   /v1/embeddings            --> github.com/ollama/ollama/server.(*Server).EmbedHandler-fm (6 handlers)
[GIN-debug] GET    /v1/models                --> github.com/ollama/ollama/server.(*Server).ListHandler-fm (6 handlers)
[GIN-debug] GET    /v1/models/:model         --> github.com/ollama/ollama/server.(*Server).ShowHandler-fm (6 handlers)
time=2025-03-11T13:00:40.793Z level=INFO source=gpu.go:199 msg="vulkan: load libvulkan and libcap ok"
time=2025-03-11T13:00:40.877Z level=INFO source=gpu.go:421 msg="error looking up vulkan GPU memory" error="device is a CPU"
time=2025-03-11T13:00:40.878Z level=WARN source=amd_linux.go:443 msg="amdgpu detected, but no compatible rocm library found.  Either install rocm v6, or follow manual install instructions at https://github.com/ollama/ollama/blob/main/docs/linux.md#manual-install"
time=2025-03-11T13:00:40.878Z level=WARN source=amd_linux.go:348 msg="unable to verify rocm library: no suitable rocm found, falling back to CPU"
time=2025-03-11T13:00:40.879Z level=INFO source=types.go:137 msg="inference compute" id=0 library=vulkan variant="" compute=1.3 driver=1.3 name="AMD Radeon Graphics (RADV GFX1103_R1)" total="15.6 GiB" available="15.6 GiB"
```

```
 # ollama run phi4:14b
>>> /set verbose
Set 'verbose' mode.
>>> how's it going?
Hello! I'm here to help you with any questions or tasks you have. How can I assist you today? 😊



total duration:       3.341959745s
load duration:        18.165612ms
prompt eval count:    15 token(s)
prompt eval duration: 475ms
prompt eval rate:     31.58 tokens/s
eval count:           26 token(s)
eval duration:        2.846s
eval rate:            9.14 tokens/s
>>>
```

* This is no longer needed
Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Fixes SIGSEGV: segmentation violation running gemma3 models on ollama 0.6.0 #21

Patch provided by McBane87 on https://github.com/whyvl/ollama-vulkan/issues/21

Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Applied 04-disable-mmap-vulkan.patch

From: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2660836871

Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Pulled new upstream code for ggml-bulkan backend
Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Merged latest ollama 0.6.2 and nasrally's Flash Attention patches (#5)

* readme: add Ellama to list of community integrations (#9800)

* readme: add screenpipe to community integrations (#9786)

* Add support for ROCm gfx1151 (#9773)

* conditionally enable parallel pipelines

* sample: make mutations in transforms explicit (#9743)

* updated minP to use early exit making use of sorted tokens

* ml/backend/ggml: allocate memory with malloc when loading model (#9822)

* runner: remove cache prompt flag from ollama runner (#9826)

We do not need to bypass the prompt caching in the ollama runner yet, as
only embedding models needed to bypass the prompt caching. When embedding
models are implemented they can skip initializing this cache completely.

* ollamarunner: Check for minBatch of context space when shifting

Models can specify that a group of inputs need to be handled a single
batch. However, context shifting didn't respect this and could trigger
a break anyways. In this case, we should instead trigger a context
shift earlier so that it occurs before the grouped batch.

Note that there still some corner cases:
 - A long prompt that exceeds the context window can get truncated
   in the middle of an image. With the current models, this will
   result in the model not recognizing the image at all, which is
   pretty much the expected result with truncation.
 - The context window is set less than the minimum batch size. The
   only solution to this is to refuse to load the model with these
   settings. However, this can never occur with current models and
   default settings.

Since users are unlikely to run into these scenarios, fixing them is
left as a follow up.

* Applied latest patches from McBane87

See this for details: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2708820861

Signed-off-by: Vadim Grinco <vadim@grinco.eu>

* Add ability to enable flash attention on vulkan (#4

)

* discover: add flash attention handling for vulkan
* envconfig: fix typo in config.go

As part of the process some code was refactored and I added a new field
FlashAttention to GpuInfo since the previous solution didn't allow for a
granular check via vulkan extensions. As a side effect, this now allows
for granular per-device FA support checking in other places

---------
Signed-off-by: Vadim Grinco <vadim@grinco.eu>
Co-authored-by: zeo <108888572+zeozeozeo@users.noreply.github.com>
Co-authored-by: Louis Beaumont <louis.beaumont@gmail.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
Co-authored-by: Michael Yang <mxyng@pm.me>
Co-authored-by: Parth Sareen <parth.sareen@ollama.com>
Co-authored-by: Jeffrey Morgan <jmorganca@gmail.com>
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Nikita <50599445+nasrally@users.noreply.github.com>

* Revert Readme changes

* Revert

* Revert changes in amd_linux.go

* Revert changes in amd_linux.go

* Remove flashattention setting gpu.go

* Revert whitespace changes in gpu.go

* Revert changes in transforms_test.go

* Revert changes in runner.go

* Revert changes in Makefile.sync

* Revert some unintented changes in Dockerfile

* Revert vulkan copy changes in Dockerfile

* Update Vulkan Code to de4c07f93783a1a96456a44dc16b9db538ee1618

* Fixed duplicate sync in ggml.go

* Revert changes in ggml.go

* Revert chnages in ggml.go

* enable falsh attention on vulkan

* revert remove parenthesis

* fixed flash attention logic enabling

* vk_check_flash_attention 0 means supported

* Update gpu.go

* Add vulkan to Windows Build script

* Remove commented out code

* Enable Vulkan Flash attention in FlashAttentionSupported

* Fix logging

* Update Vulkan backend to e54d41befcc1575f4c898c5ff4ef43970cead75f

* Removed libcap related code

libcap is not directly related to Vulkan and should be added by its own PR. It adds additional library dependencies for building and also requires users to run setcap or run ollama as root, which is not ideal for easy use

* Fix Unit Test (Add Vulkan Library)

* Add vulkan to TestHomogeneousGPUs
Test

* vulkan: get GPU ID (ollama v0.11.5)
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* disable mmap for vulkan

* Reduce Changes remove TestHomogeneousGPUs (doesn't exist on master)

* Update vulkan version to the version used in llama.cpp

* rename gpu patch to correct number

* added Vulkan API to get correct Device UUID

current UUID from pipelineCacheUUID does not match CUDA

* Fix GPU ID Patch

* Remove Code not in llama.cpp

* modified UUID code inside ggml

* Fix Patch

* Copied minimal definition from vulkan header

* Fix compile error in Mac

Metal is preferred so we're disabling Vulkan for now

* Removed unused code

Fix linter error in CI

* Fix patches apply

* fixing lint error

* Removed unneeded function call

Somehow removing this call fixed the crashing when Vulkan header was removed

* added missing NL

* Fixed missing members in Vulkan header

also added zero clear for some structs

* Fixed wrong structure ID

* Fixed Vulkan header

More aligned with official header definition now

* buildvulkanAsSeperateFunction

* Vulkan on Windows Test

* temporarly comment out gate to run windows task

* use temporarly windows-latest for build

* Commenting out other presets to build vulkan

* reenable cpu

* commenting out error action stop

* temporarly commenting out rocm

* set vulkan path

* comment out cude for faster turnaround

* correct vulkan install

* correct vulkan silent install

* fixed install command

* revert debugging changes (vulkan builds on windows)

* revert windows-latest

* trying to build vulkan for linux

* temporarly disable cuda and rocm

* try again linux build

* fix version

* trying to fix

* trying again

* trying again

* fix version

* fixed vulkan-sdk name

* try again

* trying again

* try without version number

* try again

* add some more extra

* trying to use version 1.4.313

* revert debugging changes

* Filter out already supported gpus

* revert debug code

* Use runners for GPU discovery

This revamps how we discover GPUs in the system by leveraging the Ollama
runner.  This should eliminate inconsistency between our GPU discovery and the
runners capabilities at runtime, particularly for cases where we try to filter
out unsupported GPUs.  Now the runner does that implicitly based on the actual
device list.  In some cases free VRAM reporting can be unreliable which can
leaad to scheduling mistakes, so this also includes a patch to leverage more
reliable VRAM reporting libraries if available.

Automatic workarounds have been removed as only one GPU leveraged this, which
is now documented. This GPU will soon fall off the support matrix with the next
ROCm bump.

Additional cleanup of the scheduler and discovery packages can be done in the
future once we have switched on the new memory management code, and removed
support for the llama runner.

* timing info for runner

* WIP - wire up Vulkan with the new engine based discovery

Not a complete implementation - free VRAM is better, but not accurate on
windows

* fix - trust the library paths from discovery when starting runner

* fix index bug

* fix vulkan ids to be underlying

* fix - give bootstrapping more time on slow systems

* Test if Vulkan device is supported

* vk_check_flash_attention is not needed (coompat2 coopmapt and scalar implementation exist)

* Handle GGML_VK_VISIBLE_DEVICES

* ask for supported first

* win: fix CPU query buffer handling

Try in a short loop until we get the size right.

* test: harden integration tests for slow start

If the server takes a while to start up, block
tests from starting until it's online to avoid
setting large timeouts in individual test cases.

* gofumpt fix

* fix build

* merge fixes

* merge fixes

* fixed build

* merge fixes

* fixing build

* fixed build

* fixed formatting

* fixed build

* fix vulkan gpu id patch

* sync llama.cpp vulkan code

* update build windows script

* merge fixes

* fix format

* fixed vulkan casing

* handle igpu as gpu

* improve case

* print out unknown library

* rturn Vulkan for vulkan library

* Revert "rturn Vulkan for vulkan library"

This reverts commit 690461a12fd5e93295d174c97edefb2bc33285b1.

* fixed patch number

* return Library Name

* remvoe debug code

* return integrated in vulkan backend

* Return pci Properties

* update patch

* directly get pci proeprties without parsing

* workaround for filtering devices. Correct way is to have a LibraryPosition Parameter in the deviceInfo

* Revert "directly get pci proeprties without parsing"

This reverts commit 8e0624851f5ed7d9f74518f574dfb422e4dd4dc2.

* Set FilteredID for Environment Filtering

* ROCm Library is named ROCm

* revert changes in patch

* Create 0028-vulkan-pci-and-memory.patch

* vulkan memory patch

* casing fix

* Add more pci properties

* Added better memory management

* Added better memory managament

* fixed patch

* Fixed patch

* FilterID creation group by library

* filter out vulkan supported by other gpu

* fixing deviceid compare

* Vulkan Fix FA coopmat1 invalid array indexing

* Use everywhere the same Vulkan Version 1.4.321.1

* Remove unneeded patch

* vulkan update

* sync vulkan glsl files

* only use for vulkan the filteredid (numeric device number)

* simplify code

---------
Signed-off-by: Vadim Grinco <vadim@grinco.eu>
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
Co-authored-by: pufferffish <github@bandersnatch.anonaddy.com>
Co-authored-by: KOISHI KOMEIJI FROM TOUHOU 11 <fuck>
Co-authored-by: DSLstandard <qgeneral35@gmail.com>
Co-authored-by: pufferffish <me@windtfw.com>
Co-authored-by: yeongbba <yeongmo.lee@logpresso.com>
Co-authored-by: tomaThomas <tomathomas@mailbox.org>
Co-authored-by: Antoine Viallon <antoine@lesviallon.fr>
Co-authored-by: Vadim Grinco <vadim@grinco.eu>
Co-authored-by: zeo <108888572+zeozeozeo@users.noreply.github.com>
Co-authored-by: Louis Beaumont <louis.beaumont@gmail.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
Co-authored-by: Michael Yang <mxyng@pm.me>
Co-authored-by: Parth Sareen <parth.sareen@ollama.com>
Co-authored-by: Jeffrey Morgan <jmorganca@gmail.com>
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Nikita <50599445+nasrally@users.noreply.github.com>
Co-authored-by: Masato Nakasaka <masato.nakasaka@intel.com>
Co-authored-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>

* logs: quiet down context canceled on completion

If the client closes the connection before Completion finishes, we were
logging at error level implying the runner crashed which was misleading.

time=2025-10-08T22:59:20.566-07:00 level=ERROR source=server.go:1490 msg="post predict" error="Post \"http://127.0.0.1:57736/completion\": context canceled"

* quiet down scheduler log error on expected case

Since we don't hold the lock while performing memory load calculations, other
runners can unload in parallel, so finding no runner to unload is a valid scenario
which we shouldn't log at error level.

This reverts commit 6a62b894.

This variable isn't currently documented or intended as something the user can
override, but if the user happens to set OLLAMA_LIBRARY_PATH we were doubling
this in the subprocess environment which will cause problems with the new
bootstrap discovery logic.

As we automatically enable flash attention for more models, there
are likely some cases where we get it wrong. This allows setting
OLLAMA_FLASH_ATTENTION=0 to disable it, even for models that usually
have flash attention.

This revamps how we discover GPUs in the system by leveraging the Ollama
runner.  This should eliminate inconsistency between our GPU discovery and the
runners capabilities at runtime, particularly for cases where we try to filter
out unsupported GPUs.  Now the runner does that implicitly based on the actual
device list.  In some cases free VRAM reporting can be unreliable which can
leaad to scheduling mistakes, so this also includes a patch to leverage more
reliable VRAM reporting libraries if available.

Automatic workarounds have been removed as only one GPU leveraged this, which
is now documented. This GPU will soon fall off the support matrix with the next
ROCm bump.

Additional cleanup of the scheduler and discovery packages can be done in the
future once we have switched on the new memory management code, and removed
support for the llama runner.

For each memory allocation we report the size of the (attempted)
allocation and whether it succeeded or failed. The latter status
reporting proved to be not that useful in practice as systems
such as Windows can automatically overflow from VRAM into RAM,
resultings in successful allocations even when there isn't
enough memory where we wanted.

As a result, this information is only used for debug logging,
which isn't worthwhile enough for the amount of code. It
also isn't fully accurate, as multiple allocations may result
in partial failures.

This reverts commit 1a558f98.

This reverts commit 8d6fffae.

If a model with a split vision projector is loaded in the Ollama
engine, the projector will be ignored and the model will hallucinate
a response. Instead, fallback and try to load the model in the llama
engine.

New memory estimates (see #11090 for more information) are now
enabled automatically for all models running on the Ollama engine,
improving both stability and performance through more accurate sizing
and allocation. Models running on the llama engine will continue to
use the original style of memory estimation.

If flash attention is enabled without KV cache quanitization, we will
currently always get this warning:
level=WARN source=server.go:226 msg="kv cache type not supported by model" type=""

The context must always be able to store the current batch, so
if the user requests a small context then we should also shrink
the batch to match. This also fixes the TestLongInputContext
test on the new engine. (The old engine already has this behavior.)

If a GPU's free memory is less than the reserved amount, we might get
an underflow. Since it is an unsigned uint64, we print this as a large
number rather than the more correct 0. This only affects logging, the
actual layout code already handles this correctly.

Bug #12138

* Always filter devices

Avoid crashing on unsupported AMD iGPUs

* Remove cuda device filtering

This interferes with mixed setups

With old memory estimates, it's currently impossible to load more
than one model at a time when no GPUs are available. This is because
the check for whether we need to evict a model looks to see if all
layers of the new model can be loaded onto GPUs, which is never true
if there are no GPUs. Before the memory management changes, there
was a special code path for CPU-only systems.

This problem does not exist with new memory estimates.

Fixes #11974

We dump out our best memory estimate after we complete processing
for any reason, including errors. This is helpful for finding what
what stopped us in error conditions but in some cases we might not
have gotten even the first result yet.

Fixes #11957

This changes the memory allocation strategy from upfront estimation to
tracking actual allocations done by the engine and reacting to that. The
goal is avoid issues caused by both under-estimation (crashing) and
over-estimation (low performance due to under-utilized GPUs).

It is currently opt-in and can be enabled for models running on the
Ollama engine by setting OLLAMA_NEW_ESTIMATES=1. Behavior in other
cases is unchanged and will continue to use the existing estimates.

For smaller context models, make sure we do not exceed the training size.

* Re-remove cuda v11

Revert the revert - drop v11 support requiring drivers newer than Feb 23

This reverts commit c6bcdc42.

* Simplify layout

With only one version of the GPU libraries, we can simplify things down somewhat.  (Jetsons still require special handling)

* distinct sbsa variant for linux arm64

This avoids accidentally trying to load the sbsa cuda libraries on
a jetson system which results in crashes.

* temporary prevent rocm+cuda mixed loading

"POST predict" basically means that the runner has crashed, which
can have many reasons. However, many people think this is a specific
error and either report only this message or group together unrelated
bugs. This replaces it with a more friendly and helpful message.

Currently, when the backend is created, the tensors are loaded at the
same time, which is a slow operation. This separates them to be two
steps:
 - Create backend, including enumerating tensors and memory allocation
 - Loading tensor data

This allows more flexibility in managing model loading.