Using "auto" for rope alpha removes ambiguity on how to explicitly
enable automatic rope calculation. The same behavior of None -> auto
calculate still exists, but can be overwritten if a model's tabby_config.yml
includes `rope_alpha`.
Signed-off-by: kingbri <bdashore3@proton.me>
Use the tensor parallel loader when the flag is enabled. The new loader
has its own autosplit implementation, so gpu_split_auto isn't valid
here.
Also make it easier to determine which cache type to use rather than
multiple if/else statements.
Signed-off-by: kingbri <bdashore3@proton.me>
The override wasn't being passed in before. Also, the default is now
none since Exl2 can automatically calculate the max batch size.
Signed-off-by: kingbri <bdashore3@proton.me>
Embedding models are managed on a separate backend, but are run
in parallel with the model itself. Therefore, manage this in a separate
container with separate routes.
Signed-off-by: kingbri <bdashore3@proton.me>
Infinity-emb is an async batching engine for embeddings. This is
preferable to sentence-transformers since it handles scalable usecases
without the need for external thread intervention.
Signed-off-by: kingbri <bdashore3@proton.me>
Realtime process priority assigns resources to point to tabby's
processes. Running as administrator will give realtime priority
while running as a normal user will set as high priority.
Signed-off-by: kingbri <bdashore3@proton.me>
These are faster event loops for asyncio which should improve overall
performance. Gate these under an experimental flag for now to stress
test these loops.
Signed-off-by: kingbri <bdashore3@proton.me>
Always enable the core endpoints and allow servers to be selected
as needed. Use the OAI server by default.
Signed-off-by: kingbri <bdashore3@proton.me>
Log all the parts of a request if the config flag is set. The logged
fields are all server side anyways, so nothing is being exposed to
clients.
Signed-off-by: kingbri <bdashore3@proton.me>
It's possible that tracebacks can give too much info about a system
when sent over the API. Gate this under a flag to send them only
when debugging since this feature is still useful.
Signed-off-by: kingbri <bdashore3@proton.me>
Previously, the parameters under the "model" block in config.yml only
handled the loading of a model on startup. This meant that any subsequent
API request required each parameter to be filled out or use a sane default
(usually defaults to the model's config.json).
However, there are cases where admins may want an argument from the
config to apply if the parameter isn't provided in the request body.
To help alleviate this, add a mechanism that works like sampler overrides
where users can specify a flag that acts as a fallback.
Therefore, this change both preserves the source of truth of what
parameters the admin is loading and adds some convenience for users
that want customizable defaults for their requests.
This behavior may change in the future, but I think it solves the
issue for now.
Signed-off-by: kingbri <bdashore3@proton.me>
The dynamic generator requires Flash attention 2.5.7 or higher to
be installed. This is only supported on Nvidia's 30 series and higher.
If a card is AMD or lower than the 30 series, switch to compatability
mode which functions the same way as the older generator, except
without parallel batching and any features that depend on it, such as
CFG.
Signed-off-by: kingbri <bdashore3@proton.me>
Wrong class attribute name used for max_attention_size and fixes
declaration of the draft model's chunk_size.
Also expose the parameter to the end user in both config and model
load.
Signed-off-by: kingbri <bdashore3@proton.me>
HuggingFace updated transformers to provide templates in a list for
tokenizers. Update to support this new format. Providing the name
of a template for the "prompt_template" value in config.yml will also
look inside the template list.
In addition, log if there's a template exception, but continue model
loading since it shouldn't shut down the application.
Signed-off-by: kingbri <bdashore3@proton.me>
This option saves some VRAM, but does have the chance to error out.
Add this in the experimental config section.
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>
Many APIs automatically ask for request streaming without giving
the user the option to turn it off. Therefore, give the user more
freedom by giving a server-side kill switch.
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>
Add the ability to use an unsafe config flag if needed and migrate
the exl2 check to a different file within the exl2 backend code.
Signed-off-by: kingbri <bdashore3@proton.me>
The example JSON fields were changed because of the new sampler
default strategy. Fix these by manually changing the values.
Also add support for fasttensors and expose generate_window to
the API. It's recommended to not adjust generate_window as it's
dynamically scaled based on max_seq_len by default.
Signed-off-by: kingbri <bdashore3@proton.me>
Unify API sampler params into a superclass which should make them
easier to manage and inherit generic functions from.
Not all frontends expose all sampling parameters due to connections
with OAI (that handles sampling themselves with the exception of
a few sliders).
Add the ability for the user to customize fallback parameters from
server-side.
In addition, parameters can be forced to a certain value server-side
in case the repo automatically sets other sampler values in the
background that the user doesn't want.
Signed-off-by: kingbri <bdashore3@proton.me>
CFG, or classifier-free guidance helps push a model in different
directions based on what the user provides.
Currently, CFG is ignored if the negative prompt is blank (it shouldn't
be used in that way anyways).
Signed-off-by: kingbri <bdashore3@proton.me>
With the new wiki, all parameters are fully documented along with
comments in the YAML file itself. This should help new users who
pull, copy the config, and can't start the API due to subsections
being uncommented and read.
Signed-off-by: kingbri <bdashore3@proton.me>
In newer versions of exllamav2, this value is read from the model's
config.json. This value will still default to 1.0 anyways.
Signed-off-by: kingbri <bdashore3@proton.me>
This creates a massive security hole, but it's gated behind a flag
for users who only use localhost.
A warning will pop up when users disable authentication.
Signed-off-by: kingbri <bdashore3@proton.me>
Some models (such as mistral and mixtral) set their base sequence
length to 32k due to assumptions of support for sliding window
attention.
Therefore, add this parameter to override the base sequence length
of a model which helps with auto-calculation of rope alpha.
If auto-calculation of rope alpha isn't being used, the max_seq_len
parameter works fine as is.
Signed-off-by: kingbri <bdashore3@proton.me>
Previously, the max sequence length was overriden by the user's
config and never took the model's config.json into account.
Now, set the default to 4096, but include config.prepare when
selecting the max sequence length. The yaml and API request
now serve as overrides rather than parameters.
Signed-off-by: kingbri <bdashore3@proton.me>
Append generation prompts if given the flag on an OAI chat completion
request.
This appends the "assistant" message to the instruct prompt. Defaults
to true since this is intended behavior.
Signed-off-by: kingbri <bdashore3@proton.me>
Jinja2 is a lightweight template parser that's used in Transformers
for parsing chat completions. It's much more efficient than Fastchat
and can be imported as part of requirements.
Also allows for unblocking Pydantic's version.
Users now have to provide their own template if needed. A separate
repo may be usable for common prompt template storage.
Signed-off-by: kingbri <bdashore3@proton.me>
New parameter that's safe to edit in exllamav2 v0.0.11. Only recommended
for people who know what they're doing.
Signed-off-by: kingbri <bdashore3@proton.me>
Generations can be logged in the console along with sampling parameters
if the user enables it in config.
Metrics are always logged at the end of each prompt. In addition,
the model endpoint tells the user if they're being logged or not
for transparancy purposes.
Signed-off-by: kingbri <bdashore3@proton.me>
Sometimes fastchat may not be able to detect the prompt template from
the model path. Therefore, add the ability to set it in config.yml or
via the request object itself.
Also send the provided prompt template on model info request.
Signed-off-by: kingbri <bdashore3@proton.me>
* Model: Implement basic lora support
* Add ability to load loras from config on launch
* Supports loading multiple loras and lora scaling
* Add function to unload loras
* Colab: Update for basic lora support
* Model: Test vram alloc after lora load, add docs
* Git: Add loras folder to .gitignore
* API: Add basic lora-related endpoints
* Add /loras/ endpoint for querying available loras
* Add /model/lora endpoint for querying currently loaded loras
* Add /model/lora/load endpoint for loading loras
* Add /model/lora/unload endpoint for unloading loras
* Move lora config-checking logic to main.py for better compat with API endpoints
* Revert bad CRLF line ending changes
* API: Add basic lora-related endpoints (fixed)
* Add /loras/ endpoint for querying available loras
* Add /model/lora endpoint for querying currently loaded loras
* Add /model/lora/load endpoint for loading loras
* Add /model/lora/unload endpoint for unloading loras
* Move lora config-checking logic to main.py for better compat with API endpoints
* Model: Unload loras first when unloading model
* API + Models: Cleanup lora endpoints and functions
Condenses down endpoint and model load code. Also makes the routes
behave the same way as model routes to help not confuse the end user.
Signed-off-by: kingbri <bdashore3@proton.me>
* Loras: Optimize load endpoint
Return successes and failures along with consolidating the request
to the rewritten load_loras function.
Signed-off-by: kingbri <bdashore3@proton.me>
---------
Co-authored-by: kingbri <bdashore3@proton.me>
Co-authored-by: DocShotgun <126566557+DocShotgun@users.noreply.github.com>
