agray3 [Wed, 15 May 2024 13:44:49 +0000 (14:44 +0100)]
Avoid unnecessarily disabling CUDA graphs (#7302)
As discussed in PR #6766, CUDA graphs were being disabled in the presence of long prompts.
This fixes the issue by avoiding the consective update counter from incrementing unnecessarily
for tokens in which cuda graphs are disabled due to batch size > 1.
Haggai Nuchi [Tue, 14 May 2024 05:25:56 +0000 (22:25 -0700)]
Add left recursion check: quit early instead of going into an infinite loop (#7083)
* Add left recursion check: quit early instead of going into an infinite loop
* Remove custom enum, rename left recursion check and move to "grammar internal" section, add handling for edge case where a leftmost nonterminal may be empty
Ryuei [Tue, 14 May 2024 05:20:47 +0000 (14:20 +0900)]
docs: Fix typo and update description for --embeddings flag (#7026)
- Change '--embedding' to '--embeddings' in the README
- Update the description to match the latest --help output
- Added a caution about defining physical batch size
compilade [Sat, 11 May 2024 15:06:26 +0000 (11:06 -0400)]
convert-hf : support bfloat16 conversion (#7158)
* convert-hf : support bfloat16 conversion
* gguf-py : flake8 fixes
* convert-hf : add missing space after comma
* convert-hf : get bit-exact same output as ./quantize
The quantization version was missing.
* convert-hf : don't round bf16 NANs
* convert-hf : save some memory with np.int16 intermediate bf16 weights
* convert-hf : more closely match llama.cpp with which weights to keep in f32
* convert-hf : add --outtype auto-f16
A reason for this to exist is for model quantizers who want an initial
GGUF with the most fidelity to the original model while still using
a 16-bit float type instead of 32-bit floats.
* convert-hf : remove a semicolon because flake8 doesn't like it
It's a reflex from when programming in C/C++, I guess.
* convert-hf : support outtype templating in outfile name
* convert-hf : rename --outtype auto-f16 to --outtype auto
Steve Grubb [Sat, 11 May 2024 08:13:02 +0000 (04:13 -0400)]
server : free llama_batch on exit (#7212)
* [server] Cleanup a memory leak on exit
There are a couple memory leaks on exit of the server. This hides others.
After cleaning this up, you can see leaks on slots. But that is another
patch to be sent after this.
Justine Tunney [Fri, 10 May 2024 11:01:08 +0000 (07:01 -0400)]
Fix memory bug in grammar parser (#7194)
The llama.cpp grammar parser had a bug where forgetting to add a closing
quotation mark to strings would cause parsing to crash. Anyone running a
server on a public endpoint is advised to upgrade. To reproduce this bug
./llamafile -m foo.gguf -p bar --grammar 'root::="'
Credit for discovering and reporting this issue goes to Eclypsium
Security Researcher Richard Johnson <redacted>.
This commit changes the value assigned to llama_timings.n_p_eval when
ctx->n_p_eval is 0 to be 1 instead of 1 which is the current value.
The motivation for this change is that if session caching is enabled,
for example using the `--prompt-cache main-session.txt` command line
argument for the main example, and if the same prompt is used then on
subsequent runs, the prompt tokens will not actually be passed to
llama_decode, and n_p_eval will not be updated by llama_synchoronize.
But the value of n_p_eval will be set 1 by llama_get_timings because
ctx->n_p_eval will be 0. This could be interpreted as 1 token was
evaluated for the prompt which could be misleading for applications
using this value.
Sigbjørn Skjæret [Thu, 9 May 2024 10:56:00 +0000 (12:56 +0200)]
gguf-py : add special token modification capability (#7166)
* Add special token modification capability
To be able to fix/amend special tokens in a GGUF let's add two new arguments:
* `--special-token <name> <value>` where `<name>` can be bos, eos, prefix, middle, etc. while `<value>` is the token value, f.ex. `"<｜fim▁begin｜>"`
* `--special-token-by-id <name> <id>` where `<id>` is the ID of the token, f.ex. 32006
So, in order to f.ex. add fill-in-middle tokens to a GGUF you would do the following:
```bash
python3 gguf-new-metadata.py input.gguf output.gguf --special-token prefix "<｜fim▁begin｜>" --special-token middle "<｜fim▁hole｜>" --special-token suffix "<｜fim▁end｜>"
```
* improve help text
* flake--
* fix multiple tokens warning
* make script executable
* switch to namedtuple, no need to dataclass
* typing++
* add progress bar
* Add special token modification capability
To be able to fix/amend special tokens in a GGUF let's add two new arguments:
* `--special-token <name> <value>` where `<name>` can be bos, eos, prefix, middle, etc. while `<value>` is the token value, f.ex. `"<｜fim▁begin｜>"`
* `--special-token-by-id <name> <id>` where `<id>` is the ID of the token, f.ex. 32006
So, in order to f.ex. add fill-in-middle tokens to a GGUF you would do the following:
```bash
gguf-new-metadata.py input.gguf output.gguf --special-token prefix "<｜fim▁begin｜>" --special-token middle "<｜fim▁end｜>" --special-token suffix "<｜fim▁hole｜>"
```
(yes, fim_end is the `middle` token, because completion is a `prefix`/`suffix`/`middle` sequence (where `middle` is unfilled))
or
```bash
gguf-new-metadata.py input.gguf output.gguf --special-token prefix "<fim_prefix>" --special-token middle "<fim_middle>" --special-token suffix "<fim_suffix>"
```
etc...
NB: The tokens have to exist already, trying to add non-existent token name/IDs will be ignored (with a warning), while non-existent values will fail (with an error).
compilade [Wed, 8 May 2024 22:16:38 +0000 (18:16 -0400)]
convert-hf : save memory with lazy evaluation (#7075)
* convert-hf : begin refactoring write_tensor
* convert : upgrade to sentencepiece v0.2.0
* convert-hf : remove unused n_dims in extra_*_tensors
* convert-hf : simplify MoE weights stacking
* convert-hf : flake8 linter doesn't like semicolons
* convert-hf : allow unusual model part names
For example, loading `model-00001-of-00001.safetensors` now works.
* convert-hf : fix stacking MoE expert tensors
`torch.stack` and `torch.cat` don't do the same thing.
* convert-hf : fix Mamba conversion
Tested to work even with a SentencePiece-based tokenizer.
* convert : use a string for the SentencePiece tokenizer path
* convert-hf : display tensor shape
* convert-hf : convert norms to f32 by default
* convert-hf : sort model part names
`os.listdir` is said to list files in arbitrary order.
Sorting the file names should let "model-00009-of-00042.safetensors"
be loaded before "model-00010-of-00042.safetensors".
* convert-hf : use an ABC for Model again
It seems Protocol can't be used as a statically type-checked ABC,
because its subclasses also can't be instantiated. (why did it seem to work?)
At least there's still a way to throw an error when forgetting to define
the `model_arch` property of any registered Model subclasses.
* convert-hf : use a plain class for Model, and forbid direct instantiation
There are no abstract methods used anyway,
so using ABC isn't really necessary.
* convert-hf : more consistent formatting of cmdline args
* convert-hf : align the message logged for converted tensors
* convert-hf : fix Refact conversion
* convert-hf : save memory with lazy evaluation
* convert-hf : flake8 doesn't like lowercase L as a variable name
* convert-hf : remove einops requirement for InternLM2
* convert-hf : faster model parts loading
Instead of pre-loading them all into a dict, iterate on the tensors
in the model parts progressively as needed in Model.write_tensors
Conversion for some architectures relies on checking for the presence
of specific tensor names, so for multi-part models, the weight map is read
from the relevant json file to quickly get these names up-front.
* convert-hf : minor changes for consistency
* gguf-py : add tqdm as a dependency
It's small, and used for a progress bar
in GGUFWriter.write_tensors_to_file
* Fall back if graph capture fails and address other comments
* - renamed GGML_ALLOW_CUDA_GRAPHS to GGML_CUDA_USE_GRAPHS
- rename env variable to disable CUDA graphs to GGML_CUDA_DISABLE_GRAPHS
- updated Makefile build to enable CUDA graphs
- removed graph capture failure checking in ggml_cuda_error
using a global variable to track this is not thread safe, but I am also not safistied with checking an error by string
if this is necessary to workaround some issues with graph capture with eg. cuBLAS, we can pass the ggml_backend_cuda_context to the error checking macro and store the result in the context
- fixed several resource leaks
- fixed issue with zero node graphs
- changed fixed size arrays to vectors
- removed the count of number of evaluations before start capturing, and instead changed the capture mode to relaxed
- removed the check for multiple devices so that it is still possible to use a single device, instead checks for split buffers to disable cuda graphs with -sm row
- changed the op for checking batch size to GGML_OP_ADD, should be more reliable than GGML_OP_SOFT_MAX
- code style fixes
- things to look into
- VRAM usage of the cudaGraphExec_t, if it is significant we may need to make it optional
- possibility of using cudaStreamBeginCaptureToGraph to keep track of which ggml graph nodes correspond to which cuda graph nodes
This encoding has the same number of exponent bits as float32. That
makes conversion relatively straightforward, even in the absence of
hardware support. For example, converting brain16 to binary32 means
simply shifting 16 bits to the left.
The issue is that converting bf16 to fp16 can result in information
loss. Only 13% of bf16 numbers can be precisely represented in fp16
which in practice ends up being 99.71% of Mistral 7b v0.2's weights
however there is currently no way other than fp32 to get the others
This change fixes that, by adding a bf16 data type to GGML. Support
for CPU inference has been implemented along with optimizations for
the AVX2, AVX512, and AVX512BF16 ISAs. Perplexity on Mistral 7b 0.2
improves somewhere around -0.0024 to -0.0046 compared to using fp16
* Remove GGML code that's not needed
* Minimize the GGML API surface area for BF16
* Remove bf16 luts
* Make the GGML header look nicer
* Fix documentation
* Apply ggerganov's fixes for test-backend-ops
* Add BF16 code for new ggml_validate_row_data() function
jukofyork [Wed, 8 May 2024 00:24:16 +0000 (01:24 +0100)]
Fixed save_imatrix to match old behaviour for MoE (#7099)
* Fixed save_imatrix to match old behaviour for MoE
This fix is simple and clear, but unnecessarily doubles the memory overhead..
* Fixed missing idx variable
* Unconditionally increment ncall
Co-authored-by: slaren <redacted>
* Fixed 2 bugs in save_imatrix()
- Fixed segfault bug because the counts vector needed to be created.
- Fixed pre-existing bug didn't actually add to the counts for "--combine" option.
William Tambellini [Mon, 6 May 2024 18:12:14 +0000 (11:12 -0700)]
Add an option to build without CUDA VMM (#7067)
Add an option to build ggml cuda without CUDA VMM
resolves
https://github.com/ggerganov/llama.cpp/issues/6889
https://forums.developer.nvidia.com/t/potential-nvshmem-allocated-memory-performance-issue/275416/4
overview / pkg / ggml / sources / llama.cpp / log