On the llama engine, when we compute the memory layout, we reserve
a buffer to allow for some flexibility for incorrect estimates.
This is subtracted from GPU free memory and on GPUs with limited
memory, it may underflow.
Fixes#13494
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.
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 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.
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.
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.
The initial implementation of qwen3-vl:235b exceeded the maximum graph
size based on the number of tensors. Although this was later fixed
through the use of the mrope operation, we are close to the limit in
some cases. This updates to track the current llama.cpp usage of GGML.
We pass invalid pointers when we check the size of the required
compute graph before fitting. Some CUDA APIs validate these pointers
but we can just skip them during this phase. cudaMemsetAsync is one
of these that we weren't skipping but never took the code path that
used it. Now that we have enabled op_offload, we can hit it in
memory pressured situations.
When a model is partially offloaded to system RAM, we can either
do the calculations on the CPU or we can temporarily transfer the
data to the GPU to do the calculations there. Small batches tend
to be better on the CPU, large batches on the GPU.
The llamarunner used the GPU in most cases and the ollamarunner
used the CPU. Although the ollamarunner saw an improvement in
token generation performance, there was a large performance hit
in prompt processing (3-10x).
There is an existing heuristic to dynamically switch between these
two modes but in practice it doesn't have enough information to
accurately make that decision. This adds authoritative data to make
the check work to get the best of both worlds.
Fixes#12037
We currently allocate the worst case batch for max sized
batches, which corresponds to prompt processing. However,
there are some cases where the generated graph is different
for small and large batches. To ensure that we don't need
to allocate memory later after layout has taken place, we
should run the worst case batch both ways and take the larger
amount of memory.
This does not noticeably affect loading speed as the most expensive
part of this logic is from image processing and that does not
occur during token generation.
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%...).
We currently short circuit generation of the cache mask and just
generate an empty tensor of the correct size. However, in some
cases, this can also skip a cast operation. This can result in the
worst case graph being not fully worst case.
We don't actually need the fast path for mask generation, so it's
better to just use the normal code path.
Currently, we only record the time for the last batch when processing
the prompt. This results in unrealistically high numbers for the
old llama runner.
Before:
total duration: 31.273112939s
load duration: 4.97054657s
prompt eval count: 32768 token(s)
prompt eval duration: 235.137439ms
prompt eval rate: 139356.80 tokens/s
eval count: 1873 token(s)
eval duration: 18.173182374s
eval rate: 103.06 tokens/s
After:
total duration: 30.024798033s
load duration: 4.758588663s
prompt eval count: 32768 token(s)
prompt eval duration: 7.779621548s
prompt eval rate: 4212.03 tokens/s
eval count: 1769 token(s)
eval duration: 17.148014223s
eval rate: 103.16 tokens/s
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
Sliding windows models (e.g. gpt-oss, gemma3) remove tokens that
are out of the cache's window each time we start a new forward pass.
The cache storage needs to handle the window size for each sequence
plus the batch size, since the batch needs to attend to the full
window size. This means that we have greater than a window size
stored while processing the batch.
When the next batch comes, we are currently only looking at the
sequences in the incoming batch to slide the window forward.
However, we also need to clean up the other sequences that might
be occupying space in the batch processing buffer to ensure each
sequence is only using its window size of storage. Failure to do
this can result in "no kv cache slot found" errors.
Fixes: #10127
GGML picks the wrong kernel and these systems fail with:
Sep 28 22:25:39 xavier ollama[48999]: //ml/backend/ggml/ggml/src/ggml-cuda/fattn-wmma-f16.cu:437:
ERROR: CUDA kernel flash_attn_ext_f16 has no device code compatible with CUDA arch 720. ggml-cuda.cu
was compiled for: __CUDA_ARCH_LIST__
Fixes#12442
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.
The GGML CUDA backend allocates additional memory for intermediate
results during calculation. This memory isn't currently allocated
during worst case graph reservation and therefore not included in
scheduling. This means that as these buffers potentially grow
with context length, we could crash.
This extends the memory allocation system down layer from the GGML
graph to the CUDA layer, preallocating the worst case memory there
as well.
Fixes#11753
The GGML scale kernel uses signed 32-bit ints to represent
the number of elements in the tensor. For large images,
mistral-small3.2 overflows this, triggering CUDA errors due
to negative arguments.
Currently, this can happen when the user passes a large image
to mistral-small3.2. However, with upcoming changes to reserve
CUDA memory, it happens every time mistral-small is loaded as
we reserve using a worst case batch.
This patch is part of an upstream GGML commit and should be removed
after GGML is updated past 0a1b398 "ggml: add ops for WAN video model
(cuda && cpu) (#15669)".
Fixes#10388
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.
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.
Our new engine implementation of gemma2 doesn't support flash
attention, which means that it also doesn't support KV cache
quantization. Currently, it is possible to turn these two on,
which will result in a crash.
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
The recent memory management changes caused all GPUs to be visible
to the runner, regardless of whether they are ultimately used. This
caused CUDA devices to allocate a primary context (~300 MB VRAM) on
each GPU, for each model. This is unnecessary, so we can both avoid
touching GPUs that we exclude in the early stage of allocation and
freeing the memory for any that we touch but don't use.
The issue will continue to exist for the old engine, since it touches
all devices during initialization.
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
Flash attention kernels require the mask of the KV cache be a F16
rather than an F32. We can use the GGML operation ggml_cast to do
this rather than doing it ourselves, which allows reuse of a
preallocated buffer in the graph rather than allocating a new one
for each batch. This improves token generation performance with
flash attention by 10-30% (with gpt-oss). This also makes performance
with flash attention better than without it, as expected.
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.
Some AMD GPUs do not provide UUIDs and report only "XX". In these
cases, we should use the ordinal ID as an alternate identifier.
This is the same as we always need to do on Windows for AMD.
In addition, this prints out the ID for each GPU when enumerating
them for easier debugging in the future.
Callers can set a backend buffer type to be no-alloc, meaning that
it does not allocate memory for tensors or operations. This can
be used for calculating memory requirements. Tensors and graphs
must be recreated with no-alloc set to false before loading data.
Defaults to false for newly created backend buffer types.
For many backend data structures, GGML defines a typedef of a pointer
type and returns these from functions. In most cases, CGo understands
that these are interchangable but some parts of Go (such as generics)
think they are two different types. We should prefer the form that
GGML uses.
gpt-oss works best with a context length of at least 8k. However,
for GPUs with limited amount of VRAM, there is a significant
performance hit to this increased context. In these cases, we
switch to the Ollama default of 4k
KV cache quantization has a dependency on the flash attention kernel.
We currently cannot use flash attention with gpt-oss as it requires
additional operations.
The model definition does not call flash attention, so it works
regardless of the setting but the cache will pick up the
quantization type. This updates the flash attention setting earlier
in the loading flow so that all downstream settings are also set correctly.
Fixes: #11671
There is a bug when using sliding window attention where we run
out of KV cache slots. This is likely due to not correctly removing
all of the entries as they slide out of range. This adds additional
logging when this occurs to track down the source.
Bug #10127
Models that use sliding window attention can only resume a sequence
from the cache if it falls within the saved windows. This works well
if the next message picks up where the old one left off. However, it
generally prevents a partial prefix match unless the entire conversation
falls within the sliding window.
This can be a problem with reasoning models where the traces are
supposed to be removed from future messages, forcing the entire
history to be re-evaluated.
This change allows models to specify that a larger amount of the
history be retained in memory, to allow more partial resumption.
It still respects the window that the model was trained on for
token generation.
When we context shift, we delete half the context and apply RoPE
with an offset to the other half. We used to RoPE across the entire
context in a single pass with a zero offset for the deleted
section. With the change to shifting in batches, we can skip any
batches where all of the offsets would be zero. This typically
reduces the number of operations by half.
Currently, when we need to do a shift on the cache, it is one
RoPE operation on the entire size of the cache (per layer). In
some cases, this can create a compute graph that is larger than
the forward pass since the forward pass is working in batches.
Since we don't consider shifting in our memory estimates, it's
possible for this to cause a crash if we run out of memory.
By limiting the size of the RoPE calls to batch size chunks, we
ensure that the shift will never exceed the size of the forward
pass, since the forward pass will also contain a RoPE of the same
size. This does not have a sigificant impact on performance since
RoPE is a math operation that is mostly proportional to the size
of its inputs.
In theory defrag could have the same issue since it also creates a
compute graph outside of the forward pass, however, since it is
only copies, it does not require any working space.
Reporting params.NumGPULayers can be misleading because it is the
requested number of layers, not the actual number that is loaded.
While they are often the same, there are cases where they might mismatch,
such as if the GPU backend is missing.
We don't get valid UUIDs for AMD GPUs on Windows, so the best option
is to use the ordinal IDs. This brings us in line with what we currently
do on the Ollama server - the only exception is AMD GPUs on Linux, which
falls back to using ordinal IDs. The GGML implementation has no fallback
but it doesn't appear to occur for any of the GPUs that we support.
It's also possible that there are collisions between ordinal IDs for
different libraries - however the only places where we use them are
AMD on Windows and Metal on Mac, which can never occur on the same
system.
This is causing segfaults, so disable it. Currently UUIDs are only
used for debugging purposes, although they planned to be used in
additional ways in the future.
Bug #11211
We don't check the return status after computing the graph, which
can silently lead to bad outputs if we try to keep going and future
computation succeeds. This appears to happens in certain cases on
Apple M2 devices.
Fixes#11070
This enables matching up devices and information reported by the backend
with system management libraries such as nvml to get accurate free
memory reporting.
"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.
Computing an attention mask for a large context and max batch is
expensive - over 100ms. Models like Gemma3 that have multiple types
of caches and custom attention masks need to do this 4 times, so this
adds approximately 500ms to startup time when using 128k context
When we are reserving the worst case graph, we don't need the mask,
only its shape, so we can skip this.
FromFloatSlice and FromIntSlice return an error if the shape doesn't
match the passed data or if memory can't be allocated. Since these
are inputs, the memory being allocated is system memory rather than VRAM.
In many cases, the caller can't really handle the error and panics.
Empty and Zeros directly panic if they can't allocate memory.
This makes things consistent by panicing for the first two cases,
removing a fair amount of error handling code. This is also consistent
with how Go typically handles these situations.
This provides granular information about the backend memory allocations
required by the runner:
- Per backend
- Per layer
- Weights, cache and graph
- Allocation status
This can be used for debugging and validating memory estimates.
GGML has a function to report the allocated size of a backend buffer.
However, this returns 0 if we tried to allocate a buffer and it failed.
For memory management purposes, it's important to know how much we were
trying to allocate. This extends the API to report attempted sizes for
all buffers and whether it succeeeded.
This is a partial revert of 0478d44 "Fixed over vram allcation dure to
small initial layer sizes."
Previously we used the size of the first layer as an extra reserved
amount of space to buffer our memory estimates. The above commit
changed this to use the largest layer. However, this had performance
impacts on more models than the original commit was trying to fix.
There is just a heuristic without an ideal solution so this goes back
to the historic behavior.
Fixes: #10765, #10756, #10752, #10726
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.
The Llama engine always places vision projectors on the first GPU
if one exists. However, the Ollama engine groups it with the output
layer, which means the projector is only offloaded if all other layers
are offloaded. The memory estimation code always assumes the former
layout - this changes it to use the correct layout based on the engine.
This addresses two impacts of the current behavior:
- In multi-GPU setups, we can crash with OOM errors when we try to
allocate memory on a full GPU while another still has space.
- If the vision projector is large, it may prevent us from offloading
anything when we could have fit some of the text layers.
In some cases, if we fail to assign a piece of the model to a GPU then
we lose track of this data. Although it doesn't change the memory
allocation, it does affect the total size of the model reported by
tools such as ollama ps (and also the percent offloaded).
This makes it look like setting num_gpu isn't reflected in ollama ps,
which isn't true but the offloading percent may appear to not change.
Spreading the model across more GPUs will continue to impact the
reported total size of the model.
We currently preallocate compute graph memory for the worst case
batch of text tokens. This adds support for doing the same for
images.
Note that image models are more complicated than text models in
how they process their inputs so there may be cases where this
approach isn't completely generic for all models. It covers all
currently supported models though.
For some multimodal models (such as gemma3), we create a single
graph that generates the image embedding and then use this in the
text model. The embedding tensor is completely opaque to the runner.
However, this doesn't work if we need to use the embedding in multiple
batches. This can arise if the embedding is larger than the batch size.
In these cases (as with llama4), we would like to create views that
are more appropriately sized. However, if we do this then the original
source tensor is used in multiple graphs, which isn't allowed. To
avoid that problem, models with this pattern compute the embedding
tensor on first use and recreate the individual views. There is no
longer a single vision and text graph.
This codifies the pattern of separating vision and text graphs. The
logic of computing tensors on demand is moved to the runner, so models
no longer have to worry about this. It also gives the runner visibility
into the multimodal tensors, which is important for memory management.
When we restore a sequence from the cache, we split the prompt into
the already used tokens (stored in the cache) and new tokens that
need to be processed. Currently, the references to the used tokens
are coming from the stored previous sequence.
However, even though we know that the used tokens are semantically
equivalent to the prefix of the prompt, tokens can contain pointers
which are no longer valid. As a result, it is better to get the
used tokens from the prompt, which has currently valid pointers.
This doesn't currently have any impact because it isn't possible
to reuse the pointers (which are tensors) anyways. However, it
becomes an issue once we can.
The correct constant to remove all entries to the end of the sequence
for the Ollama engine is math.MaxInt32. -1 is used by the old engine.
The impact of this is currently minimal because it would only occur
in situations that are not supported by the implemented models or
rarely used options.
Successfully completing processing with an errgroup cancels the
associated context. However, we also have a goroutine that is checking
for cancelation of the context. As a result, there is a race where
the goroutine can pick up the cancelation and report an error,
replacing the sucessful error message.
To avoid that, this replaces the goroutine with a cancelation check
when we are reading files. This also has the advantage of stopping
all reads relatively quickly on error and also ensuring that there are
no outstanding I/O operations when we return in this case.
The downside is that if a file read blocks forever (for example, over
the network) then cancelation of the context effectively won't be
honored. However, this is also true for other smaller files we read
and the tensors are read in small chunks (128K), so it's consistent
and better on balance overall.
Worst case graph preallocation was disabled by a27462b
"ollamarunner: Temporarily disable worst case graph preallocation"
since it caused crashes with large batches when not using the GPU.
This backports upstream llama.cpp commit f057808
"ggml: Don't assert fail when tensor data changes (#13222)", which
fixes the underlying bug and allows reverting the previous workaround.
In some cases, we can't find a cache slot when using sliding window
attention. It would be helpful in this (and other cases) to know what
the batch size is.
Bug #10127
The context (and therefore associated input tensors) was not being
properly closed when images were being processed. We were trying to
close them but in reality we were closing over an empty list, preventing
anything from actually being freed.
Fixes#10434
When we later have a large batch running purely on a CPU, this
results the error:
GGML_ASSERT(talloc->buffer_id >= 0)
Disabling this means that we will incrementally reallocate memory
as the graph grows.
Fixes#10410
When ggml_backend_buffer_free() is called, the device memory
is released but not all backends consistently release the actual
ggml_backend_buffer_t in system RAM, causing a memory leak.
Bug #10040
For every forward pass through the model, we need to allocate input
tensors: tokens, images, positions, outputs and masks. These get
allocated in system memory.
However, when we close the context that the tensors were allocated
through, the metadata gets freed but the actual backend memory does
not. This results in a significant memory leak.
This makes it so that all the memory allocated through a context
gets freed when it is closed.
Fixes#10040
Allocating (and in particular, freeing) memory from CUDA host buffers
is expensive and can cause a significant performance hit if we do
it for every token. Using normal system memory avoids this issue
and also gives the OS more flexibility to manage it.
There is no performance impact from this patch directly (either
positive or negative) but it makes a difference once we start
freeing memory correctly.
Context is currently mixed between pointer and value receivers. Change
this to be all pointer receivers so don't have to reason about whether
the things we are updating in the struct will be retained.
Sometimes loading the GGUF file fails with:
panic: context canceled
This is probably a filesystem error but it doesn't provide any
information about what happened.
Currently, the KV cache and graph are lazily allocated as needed.
The cache is fully allocated on first use of the corresponding
layer whereas the graph grows with the size of the context.
This can be an issue if another application allocates more VRAM
after we do our calculations - Ollama will crash in the middle of
inference. If we instead allocate the maximum needed memory at
startup of the runner, we will either succeed or fail at that point
rather than at some surprising time in the future.
Currently, this only generates a worst case batch for text, which
means that vision models may get a partial allocation and continue
to lazily allocate the rest.
If there is a CUDA OOM, we currently don't check the return value
and will evetually segfault. This checks for the problem and generates
a Go error. At the moment, this will still result in a panic but having
the error is the first step to being able to handle it more gracefully.
When truncating inputs to the the context window at the beginning of
a sequence, we remove the minimum amount possible. However, this
may cause us to truncate to the middle of a set of inputs that
the model specified should not be split up. To avoid this, we
need to remove the rest of the partial batch.
If we have an error after creating a new sequence but before
finding a slot for it, we return without releasing the semaphore.
This reduces our parallel sequences and eventually leads to deadlock.
In practice this should never happen because once we have acquired
the semaphore, we should always be able to find a slot. However, the
code is clearly not correct.
With the llama runner, we can generate up to NUM_PARALLEL batches
at once, which will then get broken up to into individual batches
to get executed by llama.cpp (i.e. we add up to 2048 tokens and
this gets split into 4 batches of 512 tokens at default settings).
This splitting can improve parallelism on multi-GPU systems because
the individual batches can move though the pipeline without blocking
on the first one to fully complete. However, we don't yet support
this in the Ollama runner, partially because it makes it hard to
enforce model-specified batch constraints, which didn't exist
previously.
The result is that we will try to execute the full, unsplit batch.
This could result in out of memory or insufficient KV cache space
errors.
This triggers batch breaking when the total inputs from all sequences
exceeds the batch size, rather than per-sequence. In order to ensure
fairness, it also reintroduces round-robinning around sequences so
that we don't let one busy sequence starve the others.
Model implementations should use Input for all of their tensors
supplied to the model. This includes tensors that relate to the
outputs, which is confusing since there is also an Output funciton.
Since Output is only used internally in GGML and not used by any
model implementations, we can remove it from the interface to
reduce confusion.
Gemma3 uses sliding windows for its context on 5/6 layers, significantly
reducing memory usage but leading to uneven usage across layers,
which makes allocation to the correct GPU difficult. We currently
estimate very conservatively by assuming all layers are consistent
at the max size.
Llama3.2-vision is also inconsistent between self attention and cross
attention layers - at moment, we calculate the correct total size
and then average this across layers. In some cases, this may lead
to crashes if a large layer is placed on a GPU sized by the average.
This allows memory estimation to calculate per-layer KV cache size
and take this account when placing layers onto GPUs. We already do
this for weights that vary per-tensor, so this is a logical extension.
Fixes#9730Fixes#9890
When computing the size of the cache for sliding window attention,
we don't need to multiple the batch size by the number of parallel
sequences - the batch size is constant.
This also simplifies the check for whether to allocate the cache
size based on capacity or window size as the batch size is already
incorporated into the capacity when handled by the runner.
Currently sliding window attention allocates and uses the full
context size and just masks out any tokens that are outside of the
window. However, we really only need (roughly) the sliding window
size.
At large context sizes this improves two things:
- Memory allocated - since the fully context size is allocated up front,
memory requirements drop substantially. On Gemma3:4b with a 32k
context window, total memory usage (including weights and non-sliding
layers) drops from ~20GB to ~8GB.
- Computation - ranges that are completely outside of the sliding
window are now removed from the tensors that are returned from the
cache rather than simply being masked out. This results in more
efficient processing, scaling with the size of the context that
has actually been used.
Notable, this does not update the scheduler for any model to be aware of
the smaller memory requirements. This is difficult for Gemma3 because
the layers are heterogeneous between sliding and non-sliding attention.
As a result, while actual memory consumption will be reduced, the
scheduler will over-estimate the requirements of the model. This means
that splitting between GPUs or GPUs and CPUs will still be suboptimal.
Bug #9730
Currently the runner computes the kv size needed and creates a
cache of that size. This is the context size times number of
parallel sequences.
Cache implementations can make better decisions about their memory
usage, so instead pass in the required capacity, number of sequences
and maximum batch size. For now, the causal cache just uses this to
compute the size in the same way as before.
This enables the runner to report progress back to the Ollama server,
both for showing status to the user and also to prevent the server
from killing the runner if it thinks things have stalled.
Most of the infrastructure was already there, this extends it to
be available to the backends.
Defragging the KV cache can generate a lot of operations, so we
need to be careful that we don't overflow the number that the graph
can support. We currently account for all of the nodes that we add
to the graph for each move but we also need to include the original
cache tensors as well.
Fixes#9904
Rather than directly giving the input data to models, we can
pass a tensor instead. In the short term, this saves some duplicated
code.
Longer term, we will want to overlap setting up the next batch with
processing of the current one. In this case, we will only have the
shape of tensor but it will not be loaded with data at the time of
graph generation. By passing only a tensor to models now, we set up
this possibility and prevent them from relying on data that they won't
have in the future.
Although the same could be done for Positions and Outputs, in some
cases we either need the raw input data or don't use them at all.
Therefore, for now we leave them as they are and allow models to
convert them to tensors as needed.
Jesse Gross