pieces[] is used to cache pending responses and is currently being
passed around to different functions. Move it into the sequences
where it logically belongs.
If a connection is closed while a sequence is being decoded, tokens
will continue to be added to the channel without anyone to read them.
This will result in the sender blocking, which will in turn block
all other decoding and sending for other sequences.
This is not limited to just the connection between Ollama and the
runner process. If the connection to the Ollama API is closed by
the user then Ollama will close the connection to the runner,
triggering this issue.
There are multiple causes and paths that result in a sequence
ending. Not all of these free the sampling context or reset the
pieces slice. This factors out the removal code so that all
paths release resources.
Currently the entire KV cache is shared by all parallel requestors.
This gives maximum resource utilization but there is a potential for
overflow and unfairness if multiple requests are trying to use
significant context. Instead, it is better to have a hard partition
of KV cache space.
It is not safe to hold a mutex only while we are waiting for the
condition variable to signal that a new sequence has been added. It's
possible that a sequence could be added in the middle of batch
processing. For example, if a new sequence is added while Decode()
is running, it will get picked up for sampling, despite not having
been added to the original batch.
This change holds a mutex for the majority of the time when active
processing is happening, releasing it only for a brief period each
time around the loop. Depending on the workload and the scheduler
is may result in unfairness between different requests. However,
this was not actually observed in testing.
This addresses the correctness issue - better performance and fairness
can be achieved with additional improvements in the future.
We should process a batch of tokens for each parallel request, rather
than having a shared pool. Otherwise, a single request can fill the
batch and then subsequent ones will fail or get starved.
Server.cpp used the KV cache size allocated for each parallel request
as the allocated size for the batch. This is the upper bound for the
batch but since we know how many tokens we will actually put in a batch
there is no need to over allocate.
The fields for inference parameters are very similar between the
Ollama API and Ollama/runners. However, some of the names are
slightly different. For these fields (such as NumKeep and
NumPredict), the values from Ollama were never read properly and
defaults were always used.
In the future, we can share a single interface rather than duplicating
structs. However, this keeps the interface consistent with minimal
changes in Ollama as long as we continue to use server.cpp
Currently, once the KV cache is full, text generation stops. Instead,
we should shift out the oldest context so that new generation can
continue based on more recent context.
This uses the algorithm from llama.cpp that is currently used by Ollama
with the server.cpp code. There are others but they are never turned
on through Ollama, so this restores parity.
The algorithm is:
- Retain a configurable number of tokens at the beginning (for things
like beginning of sequence tokens
- Drop the oldest half of the remaining tokens
- Shift the remaining new tokens to the back of the cache
If nothing has been added to a batch then decoding will fail if
attempted. This can happen, for example, if the run loop is woken
up but we realize that we have the generation limit.
If the number of input tokens exceeds the size of the batch, multiple
batches will be submitted but they will all contain the first tokens.
This processes the input tokens as expected so that each batch has
the next set of tokens.
The cgo binding for llama_token_to_piece uses a fixed 12 byte buffer,
which is usually but not always enough to hold a token. This increase
the buffer size if needed, similar to what llama.cpp does internally.
Batch size defaults to 512 but is configurable. However, llama.go uses
a fixed size buffer, causing crashes is the batch size is increase.
This changes the array size to follow the configuration.
When adding tokens to a batch, the index is zero based but is
checked against being greater than the max batch size. This results
in an out-of-bounds access when the final token is added.
