Some tensors were being taken out of inference mode during each
iteration of exllama's load_autosplit_gen. This causes errors since
autograd is off.
Therefore, make the shared load_gen_sync function have an overarching
inference_mode context to prevent forward issues. This should allow for
the generator to iterate across each thread call.
Signed-off-by: kingbri <bdashore3@proton.me>
Finish_reason was giving an empty offset. Fix this by grabbing the
finish reason first and then handling the static generation as normal.
Signed-off-by: kingbri <bdashore3@proton.me>
Some tokenizer variables don't get cleaned up on init, so these can
persist. Clean these up manually before creating a new tokenizer for
now.
Signed-off-by: kingbri <bdashore3@proton.me>
When the model is processing a prompt, add the ability to abort
on request cancellation. This is also a catch for a SIGINT.
Signed-off-by: kingbri <bdashore3@proton.me>
Yielding the finish reason before the logging causes the function to
terminate early. Instead, log before yielding and breaking out of the
generation loop.
Signed-off-by: kingbri <bdashore3@proton.me>
If max_tokens is None, it automatically scales to fill up the context.
This does not mean the generation will fill up that context since
EOS stops also exist.
Originally suggested by #86
Signed-off-by: kingbri <bdashore3@proton.me>
Max output len should be hardcoded to 16 since it's the amount of
tokens to predict per forward pass. 16 is a good value for both
normal inference and speculative decoding which also helps save
vram compared to 2048 which was the previous default.
Signed-off-by: kingbri <bdashore3@proton.me>
OAI expects finish_reason to be "stop" or "length" (there are others,
but they're not in the current scope of this project).
Make all completions and chat completions responses return this
from the model generation itself rather than putting a placeholder.
Signed-off-by: kingbri <bdashore3@proton.me>
Run these iterators on the background thread. On startup, the API
spawns a background thread as needed to run sync code on without blocking
the event loop.
Use asyncio's run_thread function since it allows for errors to be
propegated.
Signed-off-by: kingbri <bdashore3@proton.me>
Async generation helps remove many roadblocks to managing tasks
using threads. It should allow for abortables and modern-day paradigms.
NOTE: Exllamav2 itself is not an asynchronous library. It's just
been added into tabby's async nature to allow for a fast and concurrent
API server. It's still being debated to run stream_ex in a separate
thread or manually manage it using asyncio.sleep(0)
Signed-off-by: kingbri <bdashore3@proton.me>
Speculative ngram decoding is like speculative decoding without the
draft model. It's not as useful because it only decodes on predictable
sequences, but it depends on the usecase.
Signed-off-by: kingbri <bdashore3@proton.me>
This is a shared module which manages the model container and provides
extra utility functions around it to help slim down the API.
Signed-off-by: kingbri <bdashore3@proton.me>
Loguru is a flexible logger that allows for easier hooking and imports
into Rich with no problems. Also makes progress bars stick to the
bottom of the terminal window.
Signed-off-by: kingbri <bdashore3@proton.me>
Add this in addition to 8bit cache and 16bit cache. Passing "Q4" with
the cache_mode request parameter will set this on model load.
Signed-off-by: kingbri <bdashore3@proton.me>
The semaphore/queue model for Tabby is as follows:
- Any load requests go through the semaphore by default
- Any load request can include the skip_queue parameter to bypass
the semaphore
- Any unload requests are immediately executed
- All completion requests are placed inside the semaphore by default
This model preserves the parallelism of single-user mode with extra
convenience methods for queues in multi-user. It also helps mitigate
problems that were previously present in the concurrency stack.
Also change how the program's loop runs so it exits when the API thread
dies.
Signed-off-by: kingbri <bdashore3@proton.me>
This is the first in many future commits that will overhaul the API
to be more robust and concurrent. The model is admin-first where the
admin can do anything in-case something goes awry.
Previously, calls to long running synchronous background tasks would
block the entire API, making it ignore any terminal signals until
generation is completed.
To fix this, levrage FastAPI's run_in_threadpool to offload the long
running tasks to another thread. However, signals to abort the process
still kept the background thread running and made the terminal hang.
This was due to an issue with Uvicorn not propegating the SIGINT signal
across threads in its event loop. To fix this in a catch-all way, run
the API processes in a separate thread so the main thread can still
kill the process if needed.
In addition, make request error logging more robust and refer to the
console for full error logs rather than creating a long message on the
client-side.
Finally, add state checks to see if a model is fully loaded before
generating a completion.
Signed-off-by: kingbri <bdashore3@proton.me>
Using the Outlines library, add support to supply EBNF strings and
pass them to the library for parsing.
From there, a wrapper is created and a filter is passed to generation.
Replace with an in-house solution at some point that's more flexible.
Signed-off-by: kingbri <bdashore3@proton.me>
Add the ability to constrain the return value of a model to be JSON.
Built using the JSON schema standard to define the properties of what
the model should return.
This feature should be more accurate than using GBNF/EBNF to yield
the same results due to the use of lmformatenforcer.
GBNF/EBNF will be added in a different commit/branch.
Signed-off-by: kingbri <bdashore3@proton.me>
Accidentally checked on the token bias tensor which didn't contain
the token IDs. Check if the index exists on the id_to_piece list
instead.
Signed-off-by: kingbri <bdashore3@proton.me>
Automatically unload the existing model when calling /load. This was
requested many times, and does make more sense in the long run.
Signed-off-by: kingbri <bdashore3@proton.me>
For safety reasons, always use auto unless a manual split is provided
and auto is forced off.
If auto is forced off and a manual split isn't provided, a manual
split will be attempted.
Signed-off-by: kingbri <bdashore3@proton.me>
Previously, pre-sampling logprobs were used from the raw logits,
but newer versions of exl2 allow for returning token probs post-sampling.
Convert these to logprobs and send to the user.
Signed-off-by: kingbri <bdashore3@proton.me>
The GPU reserve is used as a VRAM buffer to prevent GPU overflow
when automatically deciding how to load a model on multiple GPUs.
Make this configurable.
Signed-off-by: kingbri <bdashore3@proton.me>
Take a log of the token probs since they're already normalized which
reflects the proper value. Also, don't error out if a token prob
doesn't exist in the dict and return None instead from zip.
Signed-off-by: kingbri <bdashore3@proton.me>
Adds chat completion logprob support using OAI's spec. Tokens are
not converted to tiktoken here since that will add an extra dependency
for no real reason.
Signed-off-by: kingbri <bdashore3@proton.me>
Returns token offsets, selected tokens, probabilities of tokens
post-sampling, and normalized probability of selecting a token
pre-sampling (for efficiency purposes).
Only for text completions. Chat completions in a later commit.
Signed-off-by: kingbri <bdashore3@proton.me>
Split the get tokens function into separate wrapper encode and decode
functions for overall code cleanliness.
Signed-off-by: kingbri <bdashore3@proton.me>
It makes more sense to use gpu split parameters when the user has
>1 GPUs. Otherwise, set split and split_auto to False and save
the user some VRAM.
Signed-off-by: kingbri <bdashore3@proton.me>
The model loader was using more VRAM on a single GPU compared to
base exllamav2's loader. This was because single GPUs were running
using the autosplit config which allocates an extra vram buffer
for safe loading. Turn this off for single-GPU setups (and turn
it off by default).
This change should allow users to run models which require the
entire card with hopefully faster T/s. For example, Mixtral with
3.75bpw increased from ~30T/s to 50T/s due to the extra vram headroom
on Windows.
Signed-off-by: kingbri <bdashore3@proton.me>
