GitLab

* perf: build graph for next batch in parallel to keep GPU busy

This refactors the main run loop of the ollama runner to perform the main GPU
intensive tasks (Compute+Floats) in a go routine so we can prepare the next
batch in parallel to reduce the amount of time the GPU stalls waiting for the
next batch of work.

* tests: tune integration tests for ollama engine

This tunes the integration tests to focus more on models supported
by the new engine.

Signed-off-by: zoupingshi <hangfachang@outlook.com>

In order to iteratively find the best memory allocation, we need to
be able to free backend memory so we can try again.

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which are no longer valid.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.

The sliding window cache trims entries that are outside the window for
the latest token. This works when we are extending the cache, such as
when the conversation continues. However, if we have a partial overlap
in conversation (including the BOS tokens), then we resume from a past
point in the conversation and the needed tokens are no longer stored
in memory. This verifies that the new window overlaps with the old one
before reusing the cache.
Co-authored-by: Jesse Gross <jesse@ollama.com>

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.

Clear KV cache when shift operation is not supported by model.
Added KvCacheCanShift() check to handle models that can't perform cache shifts,
falling back to full cache clear while preserving logical token history to
maintain expected behavior when context window fills up.

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.

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.

This commit refactors the LLM subsystem by removing internal subprocess
request and response types. It consolidates duplicate type definitions
across the codebase, moving them to centralized locations. The change also
standardizes interfaces between components, simplifies the ServerStatusResp
struct, and moves the ParseDurationMs function to a common package. This
cleanup reduces code duplication between different runner implementations
(llamarunner and ollamarunner).

The encoder cache needs to know the position of images in the input
stream so that it knows when to delete them. Previously images didn't
have a position, so we implied one by breaking batches before an
image and then assuming the image was in the first position. However,
multimodal objects are now given explicit positions in the input
stream, so we can use that instead.

Breaking batches was also a way to simulate a cross attention mask
for mllama. However, given that it only supports a single sequence
and a single image, this mask doesn't serve any real purpose.
Removing the batch break does not appear to affect the quality of
the output.

Most of this is simply moving the input data structures to a new
package to avoid import cycles.

Similar to the llama engine, quantizing the KV cache requires
flash attention to be enabled through the Ollama server.

Various vision models have different requirements for how they
receive their inputs. For example:
 - Mllama wants images together with text and the image embeddings
   don't themselves have positions or get stored in the main KV cache
 - Llava-style models feed in embeddings similar to tokens and
   images correspond to a varying number of tokens in the cache.

In addition, the strategy for providing inputs must support batching
and multiple sequences, which are managed by the runner. At the same
time, we want to keep data handling fully in the model so that new
architectures are not bottlenecked by runner code which does not
understand their particular requirements.

This provides a method for models to edit the input stream so that
it meets their needs while still being in a format that the runner
understands. This allows the runner to avoid special processing
for different models.

In addition, this fixes a regression where non-vision models may
try to incorrectly interpret images.

This provides integration with the new Ollama engine
(58245413 next ollama runner (#7913)) and the rest of the Ollama
infrastructure such as the runner and Ollama server.

In addition, it also builds out the KV cache infrastructure to
support requirements of how Ollama runs models such as:
 - Parallel processing
 - Memory management for defragmentation and shifting
 - Multi-modal modals

Both old and new engines continue to be supported. By default, only
the old engine is used. To enable the new engine:

Start the server with the OLLAMA_NEW_ENGINE environment variable set:
OLLAMA_NEW_ENGINE=1 ./ollama serve

Start a model that is supported by the Ollama engine. This one is Llama 3.1 8b Q4_K_M:
./ollama run jessegross/llama3.1

* llama: wire up builtin runner

This adds a new entrypoint into the ollama CLI to run the cgo built runner.
On Mac arm64, this will have GPU support, but on all other platforms it will
be the lowest common denominator CPU build.  After we fully transition
to the new Go runners more tech-debt can be removed and we can stop building
the "default" runner via make and rely on the builtin always.

* build: Make target improvements

Add a few new targets and help for building locally.
This also adjusts the runner lookup to favor local builds, then
runners relative to the executable, and finally payloads.

* Support customized CPU flags for runners

This implements a simplified custom CPU flags pattern for the runners.
When built without overrides, the runner name contains the vector flag
we check for (AVX) to ensure we don't try to run on unsupported systems
and crash.  If the user builds a customized set, we omit the naming
scheme and don't check for compatibility.  This avoids checking
requirements at runtime, so that logic has been removed as well.  This
can be used to build GPU runners with no vector flags, or CPU/GPU
runners with additional flags (e.g. AVX512) enabled.

* Use relative paths

If the user checks out the repo in a path that contains spaces, make gets
really confused so use relative paths for everything in-repo to avoid breakage.

* Remove payloads from main binary

* install: clean up prior libraries

This removes support for v0.3.6 and older versions (before the tar bundle)
and ensures we clean up prior libraries before extracting the bundle(s).
Without this change, runners and dependent libraries could leak when we
update and lead to subtle runtime errors.

This also makes it easier to truncate long inputs the same as
shifting but does not actually implement it. This type of
truncation has a trade off between quality and time to first
token.

We need to track which tokens are in the cache ourselves. We currently
add tokens to the cache tracker when we add them to batch but they are
not actually in the cache until we call Decode. This can cause
confusion when we are shifting the cache.

Avoids "could not find a KV slot for the batch" issues.

Bug #7545

The structure of the accounting for KV cache shifting was carried
over from the old runner but it now doesn't feel natural with the new
runner. There are a number of invariants that should hold true but
are difficult to reason about. There is at least one bug report
that would imply that the invariants are not holding.

This reduces the number of implicit assumptions and is more forgiving
of unexpected situations. It also improves behavior around which input
tokens are kept when truncation occurs.

Bug #7545

-Update mllama to take the cross attention state as embeddings in
a batch, more similar to how Llava handles it. This improves
integration with the input cache.
-Pass locations in a prompt for embeddings using tags similar to Llava.
-Abstract interface to vision models so the main runner accesses Clip
and Mllama similarly
Co-authored-by: Michael Yang <mxyng@pm.me>

* Re-introduce the llama package

This PR brings back the llama package, making it possible to call llama.cpp and
ggml APIs from Go directly via CGo. This has a few advantages:

- C APIs can be called directly from Go without needing to use the previous
  "server" REST API
- On macOS and for CPU builds on Linux and Windows, Ollama can be built without
  a go generate ./... step, making it easy to get up and running to hack on
  parts of Ollama that don't require fast inference
- Faster build times for AVX,AVX2,CUDA and ROCM (a full build of all runners
  takes <5 min on a fast CPU)
- No git submodule making it easier to clone and build from source

This is a big PR, but much of it is vendor code except for:

- llama.go CGo bindings
- example/: a simple example of running inference
- runner/: a subprocess server designed to replace the llm/ext_server package
- Makefile an as minimal as possible Makefile to build the runner package for
  different targets (cpu, avx, avx2, cuda, rocm)
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>

* cache: Clear old KV cache entries when evicting a slot

When forking a cache entry, if no empty slots are available we
evict the least recently used one and copy over the KV entries
from the closest match. However, this copy does not overwrite
existing values but only adds new ones. Therefore, we need to
clear the old slot first.

This change fixes two issues:
 - The KV cache fills up and runs out of space even though we think
   we are managing it correctly
 - Performance gets worse over time as we use new cache entries that
   are not hot in the processor caches

* doc: explain golang objc linker warning (#6830)

* llama: gather transitive dependencies for rocm for dist packaging (#6848)

* Refine go server makefiles to be more DRY (#6924)

This breaks up the monolithic Makefile for the Go based runners into a
set of utility files as well as recursive Makefiles for the runners.
Files starting with the name "Makefile" are buildable, while files that
end with ".make" are utilities to include in other Makefiles.  This
reduces the amount of nearly identical targets and helps set a pattern
for future community contributions for new GPU runner architectures.

When we are ready to switch over to the Go runners, these files should
move to the top of the repo, and we should add targets for the main CLI,
as well as a helper "install" (put all the built binaries on the local
system in a runnable state) and "dist" target (generate the various
tar/zip files for distribution) for local developer use.

* llama: don't create extraneous directories (#6988)

* llama: Exercise the new build in CI (#6989)

Wire up some basic sanity testing in CI for the Go runner.  GPU runners are not covered yet.

* llama: Refine developer docs for Go server (#6842)

This enhances the documentation for development focusing on the new Go
server.  After we complete the transition further doc refinements
can remove the "transition" discussion.

* runner.go: Allocate batches for all sequences during init

We should tell the model that we could have full batches for all
sequences. We already do this when we allocate the batches but it was
missed during initialization.

* llama.go: Don't return nil from Tokenize on zero length input

Potentially receiving nil in a non-error condition is surprising to
most callers - it's better to return an empty slice.

* runner.go: Remove stop tokens from cache

If the last token is EOG then we don't return this and it isn't
present in the cache (because it was never submitted to Decode).
This works well for extending the cache entry with a new sequence.

However, for multi-token stop sequences, we won't return any of the
tokens but all but the last one will be in the cache. This means
when the conversation continues the cache will contain tokens that
don't overlap with the new prompt.

This works (we will pick up the portion where there is overlap) but
it causes unnecessary cache thrashing because we will fork the original
cache entry as it is not a perfect match.

By trimming the cache to the tokens that we actually return this
issue can be avoided.

* runner.go: Simplify flushing of pending tokens

* runner.go: Update TODOs

* runner.go: Don't panic when processing sequences

If there is an error processing a sequence, we should return a
clean HTTP error back to Ollama rather than panicing. This will
make us more resilient to transient failures.

Panics can still occur during startup as there is no way to serve
requests if that fails.
Co-authored-by: jmorganca <jmorganca@gmail.com>

* runner.go: More accurately capture timings

Currently prompt processing time doesn't capture the that it takes
to tokenize the input, only decoding time. We should capture the
full process to more accurately reflect reality. This is especially
true once we start processing images where the initial processing
can take significant time. This is also more consistent with the
existing C++ runner.

* runner.go: Support for vision models

In addition to bringing feature parity with the C++ runner, this also
incorporates several improvements:
 - Cache prompting works with images, avoiding the need to re-decode
   embeddings for every message in a conversation
 - Parallelism is supported, avoiding the need to restrict to one
   sequence at a time. (Though for now Ollama will not schedule
   them while we might need to fall back to the old runner.)
Co-authored-by: jmorganca <jmorganca@gmail.com>

* runner.go: Move Unicode checking code and add tests

* runner.go: Export external cache members

Runner and cache are in the same package so the change doesn't
affect anything but it is more internally consistent.

* runner.go: Image embedding cache

Generating embeddings from images can take significant time (on
my machine between 100ms and 8s depending on the model). Although
we already cache the result of decoding these images, the embeddings
need to be regenerated every time. This is not necessary if we get
the same image over and over again, for example, during a conversation.

This currently uses a very small cache with a very simple algorithm
but it is easy to improve as is warranted.

* llama: catch up on patches

Carry forward solar-pro and cli-unicode patches

* runner.go: Don't re-allocate memory for every batch

We can reuse memory allocated from batch to batch since batch
size is fixed. This both saves the cost of reallocation as well
keeps the cache lines hot.

This results in a roughly 1% performance improvement for token
generation with Nvidia GPUs on Linux.

* runner.go: Default to classic input cache policy

The input cache as part of the go runner implemented a cache
policy that aims to maximize hit rate in both single and multi-
user scenarios. When there is a cache hit, the response is
very fast.

However, performance is actually slower when there is an input
cache miss due to worse GPU VRAM locality. This means that
performance is generally better overall for multi-user scenarios
(better input cache hit rate, locality was relatively poor already).
But worse for single users (input cache hit rate is about the same,
locality is now worse).

This defaults the policy back to the old one to avoid a regression
but keeps the new one available through an environment variable
OLLAMA_MULTIUSER_CACHE. This is left undocumented as the goal is
to improve this in the future to get the best of both worlds
without user configuration.

For inputs that result in cache misses, on Nvidia/Linux this
change improves performance by 31% for prompt processing and
13% for token generation.

* runner.go: Increase size of response channel

Generally the CPU can easily keep up with handling reponses that
are generated but there's no reason not to let generation continue
and handle things in larger batches if needed.

* llama: Add CI to verify all vendored changes have patches (#7066)

Make sure we don't accidentally merge changes in the vendored code
that aren't also reflected in the patches.

* llama: adjust clip patch for mingw utf-16 (#7065)

* llama: adjust clip patch for mingw utf-16

* llama: ensure static linking of runtime libs

Avoid runtime dependencies on non-standard libraries

* runner.go: Enable llamafile (all platforms) and BLAS (Mac OS)

These are two features that are shown on llama.cpp's system info
that are currently different between the two runners. On my test
systems the performance difference is very small to negligible
but it is probably still good to equalize the features.

* llm: Don't add BOS/EOS for tokenize requests

This is consistent with what server.cpp currently does. It affects
things like token processing counts for embedding requests.

* runner.go: Don't cache prompts for embeddings

Our integration with server.cpp implicitly disables prompt caching
because it is not part of the JSON object being parsed, this makes
the Go runner behavior similarly.

Prompt caching has been seen to affect the results of text completions
on certain hardware. The results are not wrong either way but they
are non-deterministic. However, embeddings seem to be affected even
on hardware that does not show this behavior for completions. For
now, it is best to maintain consistency with the existing behavior.

* runner.go: Adjust debug log levels

Add system info printed at startup and quiet down noisier logging.

* llama: fix compiler flag differences (#7082)

Adjust the flags for the new Go server to more closely match the
generate flow

* llama: refine developer docs (#7121)

* llama: doc and example clean up (#7122)

* llama: doc and example clean up

* llama: Move new dockerfile into llama dir

Temporary home until we fully transition to the Go server

* llama: runner doc cleanup

* llama.go: Add description for Tokenize error case

---------
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com>