GitLab

This adjusts our default settings to enable multiple models and parallel
requests to a single model.  Users can still override these by the same
env var settings as before.  Parallel has a direct impact on
num_ctx, which in turn can have a significant impact on small VRAM GPUs
so this change also refines the algorithm so that when parallel is not
explicitly set by the user, we try to find a reasonable default that fits
the model on their GPU(s).  As before, multiple models will only load
concurrently if they fully fit in VRAM.

If we try to use mmap when the model is larger than the system free space, loading is slower than the no-mmap approach.

Prior to this change, we logged the memory prediction multiple times
as the scheduler iterates to find a suitable configuration, which can be
confusing since only the last log before the server starts is actually valid.
This now logs once just before starting the server on the final configuration.
It also reports what library instead of always saying "offloading to gpu" when
using CPU.

On Windows, recent llama.cpp changes make mmap slower in most
cases, so default to off.  This also implements a tri-state for
use_mmap so we can detect the difference between a user provided
value of true/false, or unspecified.

We update the PATH on windows to get the CLI mapped, but this has
an unintended side effect of causing other apps that may use our bundled
DLLs to get terminated when we upgrade.

Implement support for GPU env var workarounds, and leverage
this for the Vega RX 56 which needs
HSA_ENABLE_SDMA=0 set to work properly

Still not complete, needs some refinement to our prediction to understand the
discrete GPUs available space so we can see how many layers fit in each one
since we can't split one layer across multiple GPUs we can't treat free space
as one logical block

Critical fix from llama.cpp JSON grammar to forbid un-escaped escape characters inside strings, which breaks parsing. (#3782)

* Revert "use `int32_t` for call to tokenize (#4738)"

This reverts commit 763bb65d.

* Revert "vocab only"

This reverts commit bf54c845.

* Revert "use ffi for tokenizing/detokenizing"

This reverts commit 26a00a04.

* partial offloading: allow flash attention and disable mmap

* allow mmap with num_gpu=0

On some systems, 1 minute isn't sufficient to finish the load after it
hits 100% This creates 2 distinct timers, although they're both set to
the same value for now so we can refine the timeouts further.

Signed-off-by: Lei Jitang <leijitang@outlook.com>

If the client closes the connection before we finish loading the model
we abort, so lets make the log message clearer why to help users
understand this failure mode

This doesn't expose a UX yet, but wires the initial server portion
of progress reporting during load

* put flash attention behind flag for now

* add test

* remove print

* up timeout for sheduler tests

* feat: enable flash attention if supported

* feat: enable flash attention if supported

* feat: enable flash attention if supported

* feat: add flash_attn support

Only dump env vars we care about in the logs

This reverts commit ce3b212d.

* dont clamp ctx size in `PredictServerFit`

* minimum 4 context

* remove context warning

This records more GPU usage information for eventual UX inclusion.

This will bubble up a much more informative error message if noexec
is preventing us from running the subprocess

Trying to live off the land for cuda libraries was not the right strategy.  We need to use the version we compiled against to ensure things work properly

* fix llava models not working after first request

* individual requests only for llava models

This moves all the env var reading into one central module
and logs the loaded config once at startup which should
help in troubleshooting user server logs