This adjusts linux to follow a similar model to windows with a discrete archive
(zip/tgz) to cary the primary executable, and dependent libraries. Runners are
still carried as payloads inside the main binary
Darwin retain the payload model where the go binary is fully self contained.
The file.Truncate call on windows will write the whole file
unless you set the sparse flag, leading to heavy I/O at the
beginning of download. This should improve our
I/O behavior on windows and put less stress on the users disk.
If the system has multiple numa nodes, enable numa support in llama.cpp
If we detect numactl in the path, use that, else use the basic "distribute" mode.
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.
If we detect an NVIDIA GPU, but nvidia doesn't support the os/arch,
this will report a better error for the user and point them to docs
to self-install the drivers if possible.
Make sure if something goes wrong spawning the process, the user gets
enough info to be able to try to self correct, or at least file a bug
with details so we can fix it. Once the process starts, we immediately
change back to the recommended setting to prevent the blocking dialog.
This ensures if the model fails to load (OOM, unsupported model type,
etc.) the process will exit quickly and we can scan the stdout/stderr
of the subprocess for the reason to report via API.
The OLLAMA_MAX_VRAM env var was a temporary workaround for OOM
scenarios. With Concurrency this was no longer wired up, and the simplistic
value doesn't map to multi-GPU setups. Users can still set `num_gpu`
to limit memory usage to avoid OOM if we get our predictions wrong.
On windows, the exit status winds up being the search term many
users search for and end up piling in on issues that are unrelated.
This refines the reporting so that if we have a more detailed message
we'll suppress the exit status portion of the message.
The v5 hip library returns unsupported GPUs which wont enumerate at
inference time in the runner so this makes sure we align discovery. The
gfx906 cards are no longer supported so we shouldn't compile with that
GPU type as it wont enumerate at runtime.
This also adjusts our algorithm to favor our bundled ROCm.
I've confirmed VRAM reporting still doesn't work properly so we
can't yet enable concurrency by default.
This adds logic to detect skew between the driver and
management library which can be attributed to OS overhead
and records that so we can adjust subsequent management
library free VRAM updates and avoid OOM scenarios.
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.
When ollama is running a long time, tmp cleaners can remove the
runners. This tightens up a few corner cases on arm macs where
we failed with "server cpu not listed in available servers map[]"
On windows, if the model dir contained unicode characters
clip models would fail to load. This fixes the file name
handling in clip.cpp to support utf16 on windows.
Refine the way we log GPU discovery to improve the non-debug
output, and report more actionable log messages when possible
to help users troubleshoot on their own.
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.
The recent refactoring of the memory prediction assumed all layers
are the same size, but for some models (like deepseek-coder-v2) this
is not the case, so our predictions were significantly off.
Prior to this change, we logged the memory prediction multiple times
as the scheduler iterates to find a suitable configuration, which can be
confusing since only the last log before the server starts is actually valid.
This now logs once just before starting the server on the final configuration.
It also reports what library instead of always saying "offloading to gpu" when
using CPU.
On Windows, recent llama.cpp changes make mmap slower in most
cases, so default to off. This also implements a tri-state for
use_mmap so we can detect the difference between a user provided
value of true/false, or unspecified.
We update the PATH on windows to get the CLI mapped, but this has
an unintended side effect of causing other apps that may use our bundled
DLLs to get terminated when we upgrade.
This implements the release logic we want via gh cli
to support updating releases with rc tags in place and retain
release notes and other community reactions.
While models are loading, the VRAM metrics are dynamic, so try
to load on a GPU that doesn't have a model actively loading, or wait
to avoid races that lead to OOMs
Our default behavior today is to try to fit into a single GPU if possible.
Some users would prefer the old behavior of always spreading across
multiple GPUs even if the model can fit into one. This exposes that
tunable behavior.
Still not complete, needs some refinement to our prediction to understand the
discrete GPUs available space so we can see how many layers fit in each one
since we can't split one layer across multiple GPUs we can't treat free space
as one logical block
On some systems, 1 minute isn't sufficient to finish the load after it
hits 100% This creates 2 distinct timers, although they're both set to
the same value for now so we can refine the timeouts further.
If the client closes the connection before we finish loading the model
we abort, so lets make the log message clearer why to help users
understand this failure mode
This test needs to be able to adjust the queue size down from
our default setting for a reliable test, so it needs to skip on
remote test execution mode.
Daniel Hiltgen