- remove ToolConfig, reduce to a single `tool_format` argument and hard-code extra args like start/end tokens
- dispatch to short, self-contained (and probably easily vibe coded) parser for each model type
- remove autodetection (seems infeasible since parsing effectively starts during streaming, and there is overlap between tool formats for different models)
- streamline xml parser and dedicate to qwen3_coder models
- add parsers for glm4.x, minimax-m2.x and mistral (seems shaky, probably because mistralai don't validate against hf)
- update docs
- move tool config from template_vars to separate yml config
- new per-gen stream collector used for both streaming and non-streaming requests to ensure logic is consistent for both
- move responsibility for switching between phases to stream collector
- collect tool calls during streaming and parse at the end of each gen
- prevent streaming empty content spans (be nice to clients)
- correctly aggregate usage stats for n>1 requests, always emit with last chunk in last gen to finish
- collect logprobs in model wrapper and correctly handle logprobs for multi-token chars etc.
- respect top_logprobs argument in request
- handle a number of edge cases like <think> tag being part of held string, etc.
- retain tool parsing and inference-abort fixes from #413, apply similar fix to non-stream request as well
Still TODO:
- testing and validation with more models and tool schemas (tested on Qwen so far)
- enable JSON constraint for JSON tool models
- possibly some pydantification
- documentation
The default is the minimum between max_position_embeddings and cache_size.
On AMD and older than Ampere NVIDIA GPUs, cache_size is ignored due
to not being supported by batching on exl2.
Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
Since cache_size is a more important parameter now for multi-user
setups, mark it as such by placing it below max_seq_len.
Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
- Cache size is now given only by the cache_size config option. Default is 4096 (user should always override to max out VRAM)
- max_seq_len, if not overridden in the config, will default to the model's config.json
- max_seq_len is reduced to be no larger than the cache
This allows for users to use nccl or native depending on the GPU setup.
NCCL is only available with Linux built wheels.
Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
A common problem in TabbyAPI is that users who want to get up and
running with a model always had issues with max_seq_len causing OOMs.
This is because model devs set max context values in the millions which
requires a lot of VRAM.
To idiot-proof first time setup, make the fallback default 4096 so
users can run their models. If a user still wants to use the model's
max_seq_len, set it to -1.
Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
This shouldn't even be an exposed option since changing it always
breaks inference with the model. Let the model's config.json handle
it.
Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
In ExllamaV2, if a model has YaRN support, linear RoPE options are
not applied. Users can set max_seq_len and exl2 will take care of
the rest.
Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
Uvloop/Winloop does provide advantages to asyncio vs the standard
Proactor loop, so remove experimental status.
Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
The previous code overrode the existing gpu split and device idx
values. This now sets an independent draft_gpu_split value and
adjusts the gpu_devices check only if the draft_gpu_split array
is larger than the gpu_split array.
Draft gpu split is not Tensor Parallel, and defaults to gpu_split_auto
if a split is not provided.
Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
If an API key sends a dummy model, it shouldn't error as the server
is catering to clients that expect specific OAI model names. This
is a problem with inline model loading since these names would error
by default. Therefore, add an exception if the provided name is in the
dummy model names (which also doubles as inline strict exceptions).
However, the dummy model names weren't configurable, so add a new
option to specify exception names, otherwise the default is gpt-3.5-turbo.
Signed-off-by: kingbri <bdashore3@proton.me>
Adds the ability to load vision parts of text + image models. Requires
an explicit flag in config because there isn't a way to automatically
determine whether the vision tower should be used.
Signed-off-by: kingbri <bdashore3@proton.me>
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>
