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.
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.
This patch modifies Ollama to allow grouping GPUs to memory-fit to the requested model, instead of the former algorithm of using one GPU distributing over all available GPUs.
Benefits:
- Lower amount of (PCIe-)bus communication between GPUs - especially when they are not very high speed
- Allowing unallocated GPUs to get into power-saving mode.
- Significantly reduce VRAM allocation when using more than 2 GPUs in a system
- Due to the reduced memory allocation, you can run more models simultaneously.
The current scheduler algorithm of picking the paralellism based on available
VRAM complicates the upcoming dynamic layer memory allocation algorithm. This
changes the default to 1, with the intent going forward that parallelism is
explicit and will no longer be dynamically determined. Removal of the dynamic
logic will come in a follow up.
When the same model is being reloaded rapidly with client connections
being canceled before the model finishes loading, the queued unload
event could cause a leak of runners by deleting a different runner from
the loaded list.
If a model is loading, and the request context is canceled during the load
by a client closing the connection, and another request is inbound for the
same model with a different configuration (context size, etc.) thus requiring
a reload, two unload events can be in flight. The first shuts down the
original model load, but the second one caused the loss of the new
reloading runner reference, thus triggering the leak.
The primary fix is detecting the duplicate unload and ignoring the second
instance. The load routine is also hardened to ensure we detect
clobbering an already present runner and unload it with a warning.
This enhances our logging in the scheduler. The initial "waiting for server" log
no longer claims an initial error state (now "not responding" which better reflects
the actual state). Runners now have slog wiring to report more details about the
runner, including PID.
* Adjust initial scheduler refCount
Ensure we only set the refCount on success
* sched: fix lock order inversion deadlock
Under certain race conditions, there was a scenario where the scheduler would
get into a deadlock while trying to update free space information while a model
was trying to unload.
this is in part to "pay" for #10452, which doubled the default context length. The combination isn't fully neutral though, because even though the old 4x2k limit and the new 2x4k limit are memory equivalent, the 1x fallback is larger with 4k
* increase default context length to 4096
We lower the default numParallel from 4 to 2 and use these "savings" to
double the default context length from 2048 to 4096.
We're memory neutral in cases when we previously would've used
numParallel == 4, but we add the following mitigation to handle some
cases where we would have previously fallen back to 1x2048 due to low
VRAM: we decide between 2048 and 4096 using a runtime check, choosing
2048 if we're on a one GPU system with total VRAM of <= 4 GB. We
purposefully don't check the available VRAM because we don't want the
context window size to change unexpectedly based on the available VRAM.
We plan on making the default even larger, but this is a relatively
low-risk change we can make to quickly double it.
* fix tests
add an explicit context length so they don't get truncated. The code
that converts -1 from being a signal for doing a runtime check isn't
running as part of these tests.
* tweak small gpu message
* clarify context length default
also make it actually show up in `ollama serve --help`
With support for multimodal models becoming more varied and common it is important for clients to be able to easily see what capabilities a model has. Retuning these from the show endpoint will allow clients to easily see what a model can do.
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
feat: add new Ollama engine using ggml through cgo
This change introduces a new way to run pretrained models. It introduces 3 high level interfaces and a bunch of smaller helper interfaces to facilitate this.
- `model.Model` defines the interface for a model architecture. Models such as `llama` and `mllama`, which are provided as examples, can implement the model's forward propagation in the `Forward` method. This method will be called to generate completions. This interface can be found in `model/model.go`
- `ml.Backend` defines the interface for a backend tensor library, in this case `ggml`. Among other things, a Backend is responsible for loading a pretrained model into hardware (GPU, CPU, etc) and providing an interface for Models to access loaded tensors. This interface can be found in `ml/backend.go`
- `ml.Tensor` defines the interface for a tensor and tensor operations
This is the first implementation of the new engine. Follow up PRs will implement more features:
- non-greedy sampling (#8410)
- integration with Ollama and KV caching (#8301)
- more model support (#9080) with more coming soon
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
The Go runner does not have a problem with supporting parallel
requests for most multimodal models. Now that we won't be potentially
falling back to server.cpp, this restriction can be lifted.
However, the new mllama model can't support parallel requests, so we
will need to keep a restriction for that.
* Fix embeddings memory corruption
The patch was leading to a buffer overrun corruption. Once removed though, parallism
in server.cpp lead to hitting an assert due to slot/seq IDs being >= token count. To
work around this, only use slot 0 for embeddings.
* Fix embed integration test assumption
The token eval count has changed with recent llama.cpp bumps (0.3.5+)
In mult-brand GPU setups, if we couldn't fully load the model we
would fall through the scheduler and mistakenly try to load across
a mix of brands. This makes sure we find the set of GPU(s) that
best fit for the partial load.
This change fixes the handling of keep_alive so that if client
request omits the setting, we only set this on initial load. Once
the model is loaded, if new requests leave this unset, we'll keep
whatever keep_alive was there.
Users may not realize the siny new model they're trying to load
fits on their disk, but can't load into system+GPU memory. Today
we crash, but with this fix, we'll give them a better error message
before even trying to load it.
Previously, some costly things were causing the loading of GGUF files
and their metadata and tensor information to be VERY slow:
* Too many allocations when decoding strings
* Hitting disk for each read of each key and value, resulting in a
not-okay amount of syscalls/disk I/O.
The show API is now down to 33ms from 800ms+ for llama3 on a macbook pro
m3.
This commit also prevents collecting large arrays of values when
decoding GGUFs (if desired). When such keys are encountered, their
values are null, and are encoded as such in JSON.
Also, this fixes a broken test that was not encoding valid GGUF.
Until ROCm v6.2 ships, we wont be able to get accurate free memory
reporting on windows, which makes automatic concurrency too risky.
Users can still opt-in but will need to pay attention to model sizes otherwise they may thrash/page VRAM or cause OOM crashes.
All other platforms and GPUs have accurate VRAM reporting wired
up now, so we can turn on concurrency by default.
This adjusts our default settings to enable multiple models and parallel
requests to a single model. Users can still override these by the same
env var settings as before. Parallel has a direct impact on
num_ctx, which in turn can have a significant impact on small VRAM GPUs
so this change also refines the algorithm so that when parallel is not
explicitly set by the user, we try to find a reasonable default that fits
the model on their GPU(s). As before, multiple models will only load
concurrently if they fully fit in VRAM.
Daniel Hiltgen
russcoss
Jesse Gross
Daniel Andersen
Devon Rifkin
Michael Yang
Jeffrey Morgan
Ire Gaddr
Bruce MacDonald
frob
Stefan Weil
Patrick Devine
Richard Lyons
Daniel Hiltgen
Blake Mizerany