* wip: llama : separate recurrent states from the KV cache
This will be necessary to support Jamba
(and other recurrent models mixed with Attention).
Doesn't compile yet, and finding a slot isn't yet done correctly for recurrent states.
* llama : use std::find for seq_nodes in llama_rs_cache
* llama : state checkpoints for recurrent models
* llama : correctly handle more edge cases for the rs cache
* llama : rename many llama_kv_cache_* functions
* llama : remove useless return value for some llama_cache_* functions
* llama : rethink recurrent state cell counts
* llama : begin work on support for variable GQA
This will also be useful for Jamba if we consider the Mamba layers
to have 0 KV heads.
* llama : gracefully fail when not finding hybrid slot
* llama : support Jamba
* llama : fix BERT inference without KV cache
* convert-hf : check for unprocessed Jamba experts
* convert-hf : support Mini-Jamba conversion
* llama : fix Jamba quantization sanity checks
* llama : sequence-length-aware batch splitting
* llama : use equal-sequence-length sub-batches for recurrent models
* ggml : simplify SSM-related operators
* llama : make recurrent state slot allocation contiguous
* llama : adapt internal uses of batches to llama_ubatch
* llama : fix batch split output count for embeddings
* llama : minimize swaps when reordering logits
This reduces overhead when running hellaswag
on thousands of sequences with very small 100k params Mamba models.
* llama : fix edge case finding batch seq_id of split recurrent cell
This otherwise was a problem when running the HellaSwag benchmark
with small batch sizes, making it crash.
* llama : avoid copies for simple batch splits
* ggml : make ggml_ssm_scan not modify its source tensors
* llama : fix shared recurrent tail cell count for small ubatch sizes
Otherwise it was impossible to run the 'parallel' example with '-ub 1'
with a Mamba or Jamba model.
* llama : fix .base() compilation error on Windows
* llama : allow doing the equivalent of SSM_CONV with SUM_ROWS and MUL
* ggml : allow GGML_OP_CONCAT to work on non-contiguous tensors
The implementation already supported it,
and this makes Mamba's conv step slightly faster.
* mamba : fix non-contiguous usage of ggml_silu
* llama : session saving and reloading for hybrid models
* convert_hf : fix Jamba conversion
* llama : fix mixed signedness comparison
* llama : use unused n_embd_k_gqa in k_shift
This also slightly reduces the diff from the master branch
* llama : begin renaming llama_past back to llama_kv_cache
* llama : remove implicit recurrent state rollbacks
* llama : partially apply clang-format style
* convert : fix jamba conv1d shape squeezing
* graph : add back hybrid memory graph input
But this time it contains the sub-cache graph inputs.
This *should* make it easier to handle updating the inputs
when caching the graph (eventually).
* model : add Jamba to Mamba-specific hparams printing
* jamba : remove redundant nullptr initializations
* model : remove unnecessary prefix for tensor loading constants
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* model : use ggml_swiglu_split for Mamba
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* model : make falcon-h1 use shared mamba2 layer builder
* memory : avoid referring to KV in recurrent cache logs
* gguf-py : avoid adding duplicate tensor mappings for Jamba
Some of the tensor names are common with Llama4
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Splits producing more than one ubatch per batch for recurrent models
were broken with #14512.
This fixes it by moving the completeness check after the ubatch split loop.
* llama : initial Mamba-2 support
* ggml : SIMD ggml_ssm_scan for Mamba-2
* ggml : improve ggml_mul speed when masking recurrent states
* llama : support running Mamba-Codestral-7B-v0.1
* llama : fix Mamba-2 conv state saving
* ggml : make the ggml_mul fast broadcast path more consistently formatted
* llama : remove unused variable
* llama : add missing break
* convert_hf : prefer SentencePiece tokenizer for Mamba-2 when present
The tokenzier.json of Mamba-Codestral-7B-v0.1 otherwise requires
workarounds to work correctly.
* llama : avoid redundant state copy for Mamba 1 and 2
* metal : attempt to adapt SSM_SCAN for Mamba-2
* metal : fix SSM_SCAN pipeline scope
* metal : use log and exp instead of log1pf and expf in SSM_SCAN
* metal : remove unused arguments for SSM_SCAN
The max index is 31, so trimming the arguments is necessary.
* metal : add back n_seqs to SSM_SCAN args
Whoops, this is needed for the offset in the concatenated output.
* metal : fix SSM_SCAN state head offset
* metal : fix wrong number of tokens per sequence in SSM_SCAN
* ggml : remove unused fast broadcast path in GGML_MUL
This was initially added because states were masked with ggml_mul,
but this is no longer done and so this "optimisation" is no longer
necessary, or at least not worth the additional code complexity.
* ggml : avoid multiply by D in GGML_OP_SSM_SCAN
This makes the weight buft detection in src/llama.cpp simpler.
* convert : transpose Mamba-2 A, D and reshape SSM_NORM
This breaks existing conversions of Mamba-2 models
to avoid some reshapes.
Not sure if it's a good idea,
but it makes the graph slightly cleaner.
* llama : more appropriate SSM_SCAN and SSM_CONV buft support checks
* convert : fix flake8 lint
* metal : fix confusion between ; and ,
* metal : add missing args for nb references in ssm_scan_f32_group
* metal : single-user mamba2 inference works
* kv-cache : remove const_cast when setting inputs for s_copy
And also fix multi-user inference for recurrent models
by using cell_id instead of i as the kv cell index
when populating s_copy.
* convert : avoid AutoConfig for Mamba and Mamba2 hparams
* kv-cache : allow context shift for recurrent models
* graph : fix recurrent state copies when avoiding copies
Works, but using lambda functions might not be that clean.
* ggml : fix mamba2 ssm scan when compiled with SVE
* ggml-cpu : reorder SVE FMA for consistency with other SIMD arches
* cuda : implement ssm scan for Mamba2
There is still room for improvement, but it works!
* cuda : adapt Mamba1 ssm scan to shape changes from Mamba2
* mamba : fix mismatched new and delete size for llm_build_mamba
Subclasses of llm_graph_context cannot have extra fields,
because the called destructor is not the one from the subclass.
This otherwise would cause problems when runnning Mamba-(1|2) inference
when compiled -DGGML_SANITIZE_ADDRESS=ON
* cuda : graceful fallback for Mamba-1 models with weird embd size
* kv-cache : avoid modifying recurrent cells when setting inputs
* kv-cache : remove inp_s_mask
It was replaced with equivalent and simpler functionality
with rs_z (the first zeroed state) and the already-existing inp_s_copy.
* kv-cache : fix non-consecutive token pos warning for recurrent models
The problem was apparently caused by how the tail cells were swapped.
* graph : simplify logic for recurrent state copies
* kv-cache : use cell without src refs for rs_z in recurrent cache
* llama-graph : fix recurrent state copy
The `state_copy` shuffle assumes everything is moved at once,
which is not true when `states_extra` is copied back to the cache
before copying the range of states between `head` and `head + n_seqs`.
This is only a problem if any of the cells in [`head`, `head + n_seqs`)
have an `src` in [`head + n_seqs`, `head + n_kv`),
which does happen when `n_ubatch > 1` in the `llama-parallel` example.
Changing the order of the operations avoids the potential overwrite
before use, although when copies are avoided (like with Mamba2),
this will require further changes.
* llama-graph : rename n_state to state_size in build_recurrent_state
This naming should reduce confusion between the state size
and the number of states.
- gguf-py : remove gguf-py/gguf/scripts/__init__.py because it's not needed
Implicit namespaces are supported since Python 3.3 (https://peps.python.org/pep-0420/),
and the entrypoints in pyproject.toml can directly refer to the main functions.
* gguf-py : support lazy tensor splitting
Splitting usually involves returning tuples of tensors,
which need to be handled properly to avoid early eager evaluation.
* gguf-py : fix flake8 lint
* convert : refactor rope_freqs generation
This should also fix vocab-only conversion for Phi-3.
* convert : adapt MiniCPM3 to separate rope_freqs insertion
MiniCPM3's tokenizer is treated as a SentencePiece tokenizer to avoid
having to run its custom Python code which mixes tokenization
in the same file as tool calls.
gguf-py : add long and short RoPE factors to tensor mappings
Empty, but the key names are used to populate the mappings.
* ggml-quants : 1.625 bpw ternary packing for BitNet 1.58b
* ggml-quants : faster 1.625 bpw AVX2 vec_dot
Not using a lookup table anymore makes it match q4_0 speed.
* gguf-py : fix formatting
* llama : remove spaces on empty line
* ggml-quants : subtract 1 when back in epi8
This makes the 1.625 bpw type go faster than q4_0. Still not the fastest.
* ggml-quants : Q2_2 now faster than Q4_K on with AVX2
* ggml-quants : cleanup Q1_3 code formatting
* ggml-quants : ARM NEON vec_dot for q2_2 and q1_3
* ggml-quants : use ceiling division when quantizing q1_3
* convert-hf : simplify BitNet pre-quantization
This still results in the exact same tensor weights and scales,
but it reveals some weirdness in the current algorithm.
* convert-hf : allow converting the weird BitNet 1.3B
Its FFN size is 5460 which is not convenient.
The offending tensors are kept in F16,
which makes the final model 5.01 bpw.
* bitnet : replace 1.58b with b1.58, as in the paper
* ggml-quants : fix build failure on Windows
* ggml-quants : attempt to fix Arm 32-bit support
* ggml : add some informative comments in q1_3 vec_dot
* ggml : add TQ1_0 and TQ2_0 ternary quantization types
* ggml : even faster TQ2_0
* ggml : also faster TQ1_0
Same optimization as for TQ2_0 by offsetting the sum instead of the weights.
This makes TQ1_0 almost as fast as Q8_0 on AVX2.
* ggml : fix build issues in certain environments
* ggml : add NEON vec_dot implementation for TQ1_0 and TQ2_0
* ggml : avoid directly using vmlal_high_s8, for 32-bit ARM compat
The compiler seems smart enough to use the same instruction
even when using vget_high_s8 instead.
* ggml : remove q1_3 and q2_2
No more 1.625 bpw and 2.000 bpw,
now instead using 1.6875 bpw and 2.0625 bpw
with TQ1_0 and TQ2_0, respectively.
* llama : remove the separate scale tensors of BitNet b1.58
They won't be needed, since the remaining ternary quant types have
built-in scales.
* ggml-quants : rename fields of TQ1_0 and TQ2_0 structs for consistency
* ggml-quants : allow using vdotq_s32 in TQ2_0 vec_dot
Not yet tested on hardware which supports it,
might not work or might not even compile. But also it might.
It should make the performance better on recent ARM CPUs.
* ggml-quants : remove comment about possible format change of TQ2_0
Making it slightly more convenient for AVX512
but less convenient for everything else is not worth the trouble.
* gguf-py : Numpy (de)quantization for TQ1_0 and TQ2_0
* ggml-quants : use roundf instead of nearest_int for TQ1_0 and TQ2_0
This does not change anything for ternary models,
since their values should never end up being in halfway cases anyway.
* convert : allow direct conversion to TQ1_0 and TQ2_0
The token embeddings and output tensors are kept in F16
to allow quantizing them to Q4_K and Q6_K with llama-quantize.
* llama : handle fallback for TQ1_0 and TQ2_0 with Q4_0
Q4_0 is not completely symmetric (so not lossless for ternary models),
but it should be good enough.
* ggml-quants : allow using ARM dot product instructions for TQ1_0
* ggml-quants : deduplicate TQ1_0 and TQ2_0 __ARM_FEATURE_DOTPROD support
* ggml : remove unused ggml_mul special case
It would otherwise conflict with the more general
optimization coming with Mamba-2.
* ggml : handle TQ1_0 and TQ2_0 in dequantization-based operators
* test-backend-ops : add TQ1_0 and TQ2_0 comments for later
Not yet adding uncommented, because some backends like SYCL and Metal
do not properly handle unknown types in supports_op for GGML_OP_MUL_MAT.
(and Metal also doesn't handle it with GGML_OP_GET_ROWS)
Support for TQ1_0 and TQ2_0 for other backends than CPU
will be added in follow-up pull requests.
* llama : advanced batch splits
This includes equal-sequence-length batch splits which are useful
to simplify recurrent model operators.
* llama : always make recurrent state slots contiguous
* ggml : simplify mamba operators
* llama : fix integer signedness mixing
* llama : logits_all has priority over batch->logits
Otherwise, the server embeddings tests failed.
This was likely an existing problem but was only detected here
because of an additional assertion.
* llama : apply suggestions
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix t5 segfault
* llama : fix Mamba session save and restore
* llama : minor cosmetic changes
* llama : rename llama_reorder_outputs to llama_output_reorder
Also move it closer to llama_output_reserve.
* llama : fix pooled embeddings when using batches with equal_seqs
* minor : add struct members for clarity
ggml-ci
* llama : fix T5 segfault again
* llama : fix Mamba pooled embeddings with multiple sequences
Until the pooled embeddings are refactored to allow splitting
across ubatches for causal embeddings,
recurrent models can only process a single sequence per ubatch
when calculating pooled embeddings.
* llama : add llama_model_is_recurrent to simplify figuring that out
This will make it easier to more cleanly support RWKV-v6 and Mamba-2.
* llama : fix simple splits when the batch contains embeddings
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* gguf-py : use classes for quants
* convert_hf : simplify internal quantization type selection
* gguf-py : fix flake8 lint
* gguf-py : fix BF16 numpy view type
* gguf-py : remove LlamaFileTypeMap
Too specific to 'llama.cpp', and would be a maintenance burden
to keep up to date.
* gguf-py : add generic quantize and dequantize functions
The quant classes no longer need to be known,
only the target or the source type,
for 'quantize' and 'dequantize', respectively.
* llama : refactor session file management
* llama : saving and restoring state checks for overflow
The size of the buffers should now be given to the functions working
with them, otherwise a truncated file could cause out of bound reads.
* llama : stream from session file instead of copying into a big buffer
Loading session files should no longer cause a memory usage spike.
* llama : llama_state_get_size returns the actual size instead of max
This is a breaking change, but makes that function *much* easier
to keep up to date, and it also makes it reflect the behavior
of llama_state_seq_get_size.
* llama : share code between whole and seq_id-specific state saving
Both session file types now use a more similar format.
* llama : no longer store all hparams in session files
Instead, the model arch name is stored.
The layer count and the embedding dimensions of the KV cache
are still verified when loading.
Storing all the hparams is not necessary.
* llama : fix uint64_t format type
* llama : various integer type cast and format string fixes
Some platforms use "%lu" and others "%llu" for uint64_t.
Not sure how to handle that, so casting to size_t when displaying errors.
* llama : remove _context suffix for llama_data_context
* llama : fix session file loading
llama_state_get_size cannot be used to get the max size anymore.
* llama : more graceful error handling of invalid session files
* llama : remove LLAMA_MAX_RNG_STATE
It's no longer necessary to limit the size of the RNG state,
because the max size of session files is not estimated anymore.
* llama : cast seq_id in comparison with unsigned n_seq_max
* gguf-py : fix some metadata name extraction edge cases
* convert_lora : use the lora dir for the model card path
* gguf-py : more metadata edge cases fixes
Multiple finetune versions are now joined together,
and the removal of the basename annotation on trailing versions
is more robust.
* gguf-py : add more name metadata extraction tests
* convert_lora : fix default filename
The default filename was previously hardcoded.
* convert_hf : Model.fname_out can no longer be None
* gguf-py : do not use title case for naming convention
Some models use acronyms in lowercase,
which can't be title-cased like other words,
so it's best to simply use the same case
as in the original model name.
Note that the size label still has an uppercased suffix
to make it distinguishable from the context size of a finetune.
* convert_hf : fix Gemma v1 conversion
* convert_hf : allow renaming tokens, but with a warning
* convert_hf : fix Gemma v1 not setting BOS and EOS tokens
* convert_hf : faster lazy safetensors
This makes '--dry-run' much, much faster.
* convert_hf : fix memory leak in lazy MoE conversion
The '_lazy' queue was sometimes self-referential,
which caused reference cycles of objects old enough
to avoid garbage collection until potential memory exhaustion.
* llama : fix mpt and olmo pre-tokenizer
* llama : pre-tokenize non-special user-defined tokens first
* llama : fix detection of control-like user-defined tokens
* convert_hf : identify which user-defined tokens are control tokens
Only used in _set_vocab_gpt2() for now.
* convert_hf : identify more added control tokens for SPM tokenziers
This makes Gemma and Gemma-2 tokenize pretty much EVERYTHING correctly,
including HTML tags and consecutive spaces,
but it unfortunately requires model re-conversion.
There seems to be a weird behavior of the HF tokenizer for Gemma,
which prefers to use the 16-space token over more lengthy space tokens,
while using the SentencePiece tokenizer does not do this.
(the implementation in llama.cpp has the same behavior as SentencePiece)
* llama : fix wrong pre-tokenization of byte tokens
* llama : fix Viking pre-tokenizer regex
The order was previously wrong, which caused errors in some tests.
* llama : fix command-r detokenization
* convert_hf : reduce usages of the UNKNOWN token type
* llama : add UNKNOWN tokens in the special tokens cache
* convert_hf : reduce usages of UNKNOWN for InternLM2
This makes the changes from #8321 more consistent
with the other changes made here.
* test-tokenizer-random : reduce potential confilcts with #8379
* test-tokenizer-random : add a failing edge case for falcon
* py : type-check all Python scripts with Pyright
* server-tests : use trailing slash in openai base_url
* server-tests : add more type annotations
* server-tests : strip "chat" from base_url in oai_chat_completions
* server-tests : model metadata is a dict
* ci : disable pip cache in type-check workflow
The cache is not shared between branches, and it's 250MB in size,
so it would become quite a big part of the 10GB cache limit of the repo.
* py : fix new type errors from master branch
* tests : fix test-tokenizer-random.py
Apparently, gcc applies optimisations even when pre-processing,
which confuses pycparser.
* ci : only show warnings and errors in python type-check
The "information" level otherwise has entries
from 'examples/pydantic_models_to_grammar.py',
which could be confusing for someone trying to figure out what failed,
considering that these messages can safely be ignored
even though they look like errors.
In #7075, to fix the conversion of (some) models using model-00001-of-00001.safetensors instead of model.safetensors for a single model part we simply used the same logic as the part count to get the part names.
But this doesn't always work correctly, like when unusual additional model files like consolidated.safetensors in https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3 are present.
This commit matching both the prefix and the suffix of the model part names should fix this problem without breaking any previously-supported upstream models. But according to report by @teleprint-me there is still some
persistent problem, but shall do in the meantime.
Main changes of this PR is to consolidate GGUFWriter.add_key and GGUFWriter.add_val into GGUFWriter.add_key_value.
In addition use_temp_file is now opt-in instead of opt-out defaulting to False.
Also GGUFWriter now does not require output file name until when actually writing to it.
And GGUFWriter doesn't really need to eagerly prepare the data layout of the metadata
* convert-hf : support q8_0 conversion
* convert-hf : add missing ftype
This was messing with the checksums otherwise.
* convert-hf : add missing ftype to Baichuan and Xverse
I didn't notice these on my first pass.
* 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
* 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
