* SimpleChat:DU:BringIn local helper js modules using importmap
Use it to bring in a simple trim garbage at end logic, which is
used to trim received response.
Also given that importmap assumes esm / standard js modules, so
also global variables arent implicitly available outside the
modules. So add it has a member of document for now
* SimpleChat:DU: Add trim garbage at end in loop helper
* SimpleChat:DU:TrimGarbage if unable try skip char and retry
* SimpleChat:DU: Try trim using histogram based info
TODO: May have to add max number of uniq chars in histogram at
end of learning phase.
* SimpleChat:DU: Switch trim garbage hist based to maxUniq simple
Instead of blindly building histogram for specified substring
length, and then checking if any new char within specified min
garbage length limit, NOW exit learn state when specified maxUniq
chars are found. Inturn there should be no new chars with in
the specified min garbage length required limit.
TODO: Need to track char classes like alphabets, numerals and
special/other chars.
* SimpleChat:DU: Bring in maxType to the mix along with maxUniq
Allow for more uniq chars, but then ensure that a given type of
char ie numerals or alphabets or other types dont cross the
specified maxType limit. This allows intermixed text garbage
to be identified and trimmed.
* SimpleChat:DU: Cleanup debug log messages
* SimpleChat:UI: Move html ui base helpers into its own module
* SimpleChat:DU:Avoid setting frequence/Presence penalty
Some models like llama3 found to try to be over intelligent by
repeating garbage still, but by tweaking the garbage a bit so that
it is not exactly same. So avoid setting these penalties and let
the model's default behaviour work out, as is.
Also the simple minded histogram based garbage trimming from end,
works to an extent, when the garbage is more predictable and
repeatative.
* SimpleChat:UI: Add and use a para-create-append helper
Also update the config params dump to indicate that now one needs
to use document to get hold of gMe global object, this is bcas of
moving to module type js.
Also add ui.mjs to importmap
* SimpleChat:UI: Helper to create bool button and use it wrt settings
* SimpleChat:UI: Add Select helper and use it wrt ChatHistoryInCtxt
* SimpleChat:UI:Select: dict-name-value, value wrt default, change
Take a dict/object of name-value pairs instead of just names.
Inturn specify the actual value wrt default, rather than the
string representing that value.
Trap the needed change event rather than click wrt select.
* SimpleChat:UI: Add Div wrapped label+element helpers
Move settings related elements to use the new div wrapped ones.
* SimpleChat:UI:Add settings button and bring in settings ui
* SimpleChat:UI:Settings make boolean button text show meaning
* SimpleChat: Update a bit wrt readme and notes in du
* SimpleChat: GarbageTrim enable/disable, show trimmed part ifany
* SimpleChat: highlight trim, garbage trimming bitmore aggressive
Make it easy for end user to identified the trimmed text.
Make garbage trimming logic, consider a longer repeat garbage
substring.
* SimpleChat: Cleanup a bit wrt Api end point related flow
Consolidate many of the Api end point related basic meta data into
ApiEP class.
Remove the hardcoded ApiEP/Mode settings from html+js, instead use
the generic select helper logic, inturn in the settings block.
Move helper to generate the appropriate request json string based
on ApiEP into SimpleChat class itself.
* SimpleChat:Move extracting assistant response to SimpleChat class
so also the trimming of garbage.
* SimpleChat:DU: Bring in both trim garbage logics to try trim
* SimpleChat: Cleanup readme a bit, add one more chathistory length
* SimpleChat:Stream:Initial handshake skeleton
Parse the got stream responses and try extract the data from it.
It allows for a part read to get a single data line or multiple
data line. Inturn extract the json body and inturn the delta
content/message in it.
* SimpleChat: Move handling oneshot mode server response
Move handling of the oneshot mode server response into SimpleChat.
Also add plumbing for moving multipart server response into same.
* SimpleChat: Move multi part server response handling in
* SimpleChat: Add MultiPart Response handling, common trimming
Add logic to call into multipart/stream server response handling.
Move trimming of garbage at the end into the common handle_response
helper.
Add new global flag to control between oneshot and multipart/stream
mode of fetching response. Allow same to be controlled by user.
If in multipart/stream mode, send the stream flag to the server.
* SimpleChat: show streamed generative text as it becomes available
Now that the extracting of streamed generated text is implemented,
add logic to show the same on the screen.
* SimpleChat:DU: Add NewLines helper class
To work with an array of new lines. Allow adding, appending,
shifting, ...
* SimpleChat:DU: Make NewLines shift more robust and flexible
* SimpleChat:HandleResponseMultiPart using NewLines helper
Make handle_response_multipart logic better and cleaner. Now it
allows for working with the situation, where the delta data line
got from server in stream mode, could be split up when recving,
but still the logic will handle it appropriately.
ALERT: Rather except (for now) for last data line wrt a request's
response.
* SimpleChat: Disable console debug by default by making it dummy
Parallely save a reference to the original func.
* SimpleChat:MultiPart/Stream flow cleanup
Dont try utf8-decode and newlines-add_append if no data to work on.
If there is no more data to get (ie done is set), then let NewLines
instance return line without newline at end, So that we dont miss
out on any last-data-line without newline kind of scenario.
Pass stream flag wrt utf-8 decode, so that if any multi-byte char
is only partly present in the passed buffer, it can be accounted
for along with subsequent buffer. At sametime, bcas of utf-8's
characteristics there shouldnt be any unaccounted bytes at end,
for valid block of utf8 data split across chunks, so not bothering
calling with stream set to false at end. LATER: Look at TextDecoder's
implementation, for any over intelligence, it may be doing..
If needed, one can use done flag to account wrt both cases.
* SimpleChat: Move baseUrl to Me and inturn gMe
This should allow easy updating of the base url at runtime by the
end user.
* SimpleChat:UI: Add input element helper
* SimpleChat: Add support for changing the base url
This ensures that if the user is running the server with a
different port or wants to try connect to server on a different
machine, then this can be used.
* SimpleChat: Move request headers into Me and gMe
Inturn allow Authorization to be sent, if not empty.
* SimpleChat: Rather need to use append to insert headers
* SimpleChat: Allow Authorization header to be set by end user
* SimpleChat:UI+: Return div and element wrt creatediv helpers
use it to set placeholder wrt Authorization header.
Also fix copy-paste oversight.
* SimpleChat: readme wrt authorization, maybe minimal openai testing
* SimpleChat: model request field for openai/equivalent compat
May help testing with openai/equivalent web services, if they
require this field.
* SimpleChat: readme stream-utf-8 trim-english deps, exception2error
* Readme: Add a entry for simplechat in the http server section
* SimpleChat:WIP:Collate internally, Stream mode Trap exceptions
This can help ensure that data fetched till that point, can be
made use of, rather than losing it.
On some platforms, the time taken wrt generating a long response,
may lead to the network connection being broken when it enters
some user-no-interaction related power saving mode.
* SimpleChat:theResp-origMsg: Undo a prev change to fix non trim
When the response handling was moved into SimpleChat, I had changed
a flow bit unnecessarily and carelessly, which resulted in the non
trim flow, missing out on retaining the ai assistant response.
This has been fixed now.
* SimpleChat: Save message internally in handle_response itself
This ensures that throwing the caught exception again for higher
up logic, doesnt lose the response collated till that time.
Go through theResp.assistant in catch block, just to keep simple
consistency wrt backtracing just in case.
Update the readme file.
* SimpleChat:Cleanup: Add spacing wrt shown req-options
* SimpleChat:UI: CreateDiv Divs map to GridX2 class
This allows the settings ui to be cleaner structured.
* SimpleChat: Show Non SettingsUI config field by default
* SimpleChat: Allow for multiline system prompt
Convert SystemPrompt into a textarea with 2 rows. Reduce
user-input-textarea to 2 rows from 3, so that overall
vertical space usage remains same.
Shorten usage messages a bit, cleanup to sync with settings ui.
* SimpleChat: Add basic skeleton for saving and loading chat
Inturn when ever a chat message (system/user/model) is added,
the chat will be saved into browser's localStorage.
* SimpleChat:ODS: Add a prefix to chatid wrt ondiskstorage key
* SimpleChat:ODS:WIP:TMP: Add UI to load previously saved chat
This is a temporary flow
* SimpleChat:ODS:Move restore/load saved chat btn setup to Me
This also allows being able to set the common system prompt
ui element to loaded chat's system prompt.
* SimpleChat:Readme updated wrt save and restore chat session info
* SimpleChat:Show chat session restore button, only if saved session
* SimpleChat: AutoCreate ChatRequestOptions settings to an extent
* SimpleChat: Update main README wrt usage with server
Supercedes #4024 and #4813.
CMake's native HIP support has become the
recommended way to add HIP code into a project (see
[here](https://rocm.docs.amd.com/en/docs-6.0.0/conceptual/cmake-packages.html#using-hip-in-cmake)).
This PR makes the following changes:
1. The environment variable `HIPCXX` or CMake option
`CMAKE_HIP_COMPILER` should be used to specify the HIP
compiler. Notably this shouldn't be `hipcc`, but ROCm's clang,
which usually resides in `$ROCM_PATH/llvm/bin/clang`. Previously
this was control by `CMAKE_C_COMPILER` and `CMAKE_CXX_COMPILER`.
Note that since native CMake HIP support is not yet available on
Windows, on Windows we fall back to the old behavior.
2. CMake option `CMAKE_HIP_ARCHITECTURES` is used to control the
GPU architectures to build for. Previously this was controled by
`GPU_TARGETS`.
3. Updated the Nix recipe to account for these new changes.
4. The GPU targets to build against in the Nix recipe is now
consistent with the supported GPU targets in nixpkgs.
5. Added CI checks for HIP on both Linux and Windows. On Linux, we test
both the new and old behavior.
The most important part about this PR is the separation of the
HIP compiler and the C/C++ compiler. This allows users to choose
a different C/C++ compiler if desired, compared to the current
situation where when building for ROCm support, everything must be
compiled with ROCm's clang.
~~Makefile is unchanged. Please let me know if we want to be
consistent on variables' naming because Makefile still uses
`GPU_TARGETS` to control architectures to build for, but I feel
like setting `CMAKE_HIP_ARCHITECTURES` is a bit awkward when you're
calling `make`.~~ Makefile used `GPU_TARGETS` but the README says
to use `AMDGPU_TARGETS`. For consistency with CMake, all usage of
`GPU_TARGETS` in Makefile has been updated to `AMDGPU_TARGETS`.
Thanks to the suggestion of @jin-eld, to maintain backwards
compatibility (and not break too many downstream users' builds), if
`CMAKE_CXX_COMPILER` ends with `hipcc`, then we still compile using
the original behavior and emit a warning that recommends switching
to the new HIP support. Similarly, if `AMDGPU_TARGETS` is set but
`CMAKE_HIP_ARCHITECTURES` is not, then we forward `AMDGPU_TARGETS`
to `CMAKE_HIP_ARCHITECTURES` to ease the transition to the new
HIP support.
Signed-off-by: Gavin Zhao <git@gzgz.dev>
* Revert "Revert "llava : add support for moondream vision language model (#6899)""
This reverts commit 9da243b36a.
* Fix num_positions and embeddings initialization
* Further tidy on Android instructions README.md
Fixed some logic when following readme direction
* Clean up redundent information
A new user arriving will see simple directions on llama.cpp homepage
* corrected puncuation
Period after cmake, colon after termux
* re-word for clarity
method seems to be more correct, instead of alternative in this context
* Organized required packages per build type
building llama.cpp with NDK on a pc doesn't require installing clang, cmake, git, or wget in termux.
* README.md
corrected title
* fix trailing whitespace
* add support for moondream vision language model
This required making the following changes to the CLIP model:
1. Support for patch embedding bias.
2. Make class embedding and pre-layernorm optional.
3. Add support for post-layernorm.
* Update examples/llava/clip.cpp
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* added fedora to list of distros that may need the package (the packages have the same name on Fedora)
* how to add clblast that is avalible in the fedora repos
* llama : save and restore kv cache for single seq id
* remove trailing whitespace
* respond error in case there's no space in the kv cache
* add kv seq save restore to test case
* add --slot-save-path arg to enable save restore and restrict save location
* Returning 0 for some cases, instead of asserting.
* cleanup error cases
* rename sequence state functions
* rename state get set functions
* add previous function names back in with DEPRECATED notice
* update doc
* adjust endpoints to preferred style
* fix restoring zero cell count
* handle seq rm return value
* unused param
* keep in the size check
* fix return types
* add server test case for slot save restore
* cleanup
* add cake
* cleanup style
* add special
* removing a whole sequence never fails
* move sequence state file functionality from server to llama to match session api and add version tags
* catch exceptions on save as well
* error log messages
* check types for stricter restore
* update server doc
* readme : update API changes date
* strict filename validation
* move include, reject bom as well
* also reject empty filename
* reject whitespace and trailing dot
---------
Co-authored-by: Martin Evans <martindevans@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* initial commit for sealion support
* add sealion support
* minor fix
* q/k ln and pos_embd only if required
* Apply suggestions from code review
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* minor : clear whitespaces
---------
Co-authored-by: bryan <bryansiow@aisingapore.org>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix Vulkan no kv offload incoherence
* Add k-quant mul mat mat shaders
* Rework working buffer allocation, reduces vram use noticeably
Clean up cpu assist code, replaced with ggml-backend offload function
* Default to all dedicated GPUs
* Add fallback for integrated GPUs if no dedicated GPUs are found
* Add debug info which device is allocating memory
* Fix Intel dequant issue
Fix validation issue
* Fix Vulkan GGML_OP_GET_ROWS implementation
* Clean up merge artifacts
* Remove Vulkan warning
* Support xverse model convert to gguf format.
* 1. Convert xverse models to gguf;
2. Add LLM_ARCH_XVERSE inference in llama.cpp;
3. Add xverse item in Supported models in README.md;
* * gguf-py: remove redundant logs
* llama: remove the init_mapping_prefetch custom parameter
* llama.cpp: Include the changes from #6122 to exclude the unused outputs of the last layers.
* - Fix format issues
- Remove duplicate set kqv_out to llm_build_kv
* Update llama.cpp
---------
Co-authored-by: willhe <willhe@xverse.cn>
Co-authored-by: willhe <hexin@xverse.cn>
* llama : greatly reduce logits memory usage
* llama : more compact state saving and reloading
* llama : fix lctx.n_outputs not being set before building graph
* perplexity : adapt to the logits API changes
* perplexity : fix Winogrande, use correct logits for second choice start
The first logits used to evaluate the second choice were not from
the end of the common prefix; instead, they were the logits from the end
of the first choice. This has been corrected.
The previous implementation sometimes had outliers in the scores of
choices for some tasks, and the logic to skip choices words
in the log-likelihood evaluation probably was an attempt to reduce those,
but it was complex and didn't quite seem to be the right thing.
This is simpler now, and the outlier scores aren't there anymore.
* perplexity : normalize spaces and punctuation in Winogrande sentences
* llama : fix embedding conditions
* llama : fix llama_get_embeddings_ith when the resulting id is 0
* llama : fix wrong n_outputs in llama_set_inputs
A mismatch happened when using a smaller n_ubatch than n_batch and then using
llama_batch_get_one(). The decision of what n_outputs should be now almost
fully depends on how lctx.n_outputs is set in llama_decode_internal.
The conditions are simpler this way.
* llama : when saving the state, recalculate n_outputs
This ensures the correct number of outputs for the entire previous batch
is stored in the session file, even when n_ubatch is smaller than n_batch.
* llama : fix not-skipping outputs of non-causal models
* llama : fix running a batch with n_outputs == 0
It previously worked because lctx.inp_out_ids was not initialized,
so it pointed to some garbage address which was somehow still valid when I
ran my tests.
* llama : keep same graph topology even when n_outputs == 0
* ggml : saner ggml_can_repeat with empty tensors
* ggml : future-proof ggml_is_empty by using GGML_MAX_DIMS - 1
* ggml : do not multi-thread ops returning empty tensors
* ggml : make ggml_is_empty public and work with views
* llama : use a vector for ctx->output_ids
* llama : rework reallocation logic for llama_output_reserve
Now comparing the actual size with the new total size of the output buffer
to allow more efficient enabling and disabling of the embeddings
and/or logits output in the future.
* ggml : skip empty tensors in all backends
* llama : fix llama_output_reserve nullptr deref when new_size is 0
* perplexity : make Winogrande work as it does on master
The problems with the Winogrande implementation will
need to be fixed in a separate PR to ease review.
* llama : clearer error messages for invalid logits or embeddings ids
* llama : assert all models that can have inp_out_ids
Since the graph topology is now constant, this presence check
can be done even when there are no outputs.
* llama : assert logits and embd buffers exist before writing to them
* llama : handle errors from llama_output_reserve at call sites
* perplexity : make hellaswag and multiple-choice outputs identical to master
Due to how the KV cache is updated, the logprobs for tokens in a batch
are very slightly affected by the other tokens present in the batch,
so to make hellaswag and multiple-choice return exactly the same results
as on master, the last token of each sequence needs to be evaluated
even though its output is not used at all.
This will probably be changed back in the future to make these benchmarks
a tiny bit faster.
* perplexity : fix division by zero when using less than 100 multiple-choice tasks
* llama : allow loading state saved with a different ctx size
When loading a session file, the context size is now only required to be
at least enough to load the KV cells contained in that session file,
instead of requiring to use exactly the same context size as when saving.
Doing this enables the use-case of extending or shrinking the context size
of a saved session.
This breaks existing session files because the meaning of kv_buf_size
is slightly changed (previously it was the size of the whole KV cache,
now it's only the size of the saved part of it). This allows for
finer-grained sanity checks when loading in an effort to keep kv_buf_size
useful even when the kv_size is changed.
* llama : minor
ggml-ci
* readme : update recent API changes, and warn about Vulkan
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* split: support in llama_model_loader
* avoid copying the entire vector
Co-authored-by: slaren <slarengh@gmail.com>
* split: move llama_tensor_offset to llama_model_loader
* llama_model_loader: PR feedbacks:
- use only one gguf_context for metadata only
- store all ggml_context in a vector as the files and mappings
- store all weights in a vector along with the source tensor
- rename ctx_gguf to meta
- rename ctx_meta to contexts
* avoid copying the entire vector
* Simplify this by making these optional, switch some layer creation tensor optional
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Handle optional tensors
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama_model_loader: fail if backend cannot allocate buffer
* fix mmap buffer management
* llama_model_loader: map file to backend buffer if the allocation succeeds only
* llama_model_loader: only map tensors included in the context
* llama_model_loader: minor, use same variable name for consistency, fix spacing in types cast
* llama_model_loader: fail if any of backend buffer cannot be allocated
* spacing
Co-authored-by: slaren <slarengh@gmail.com>
* fix loop over pointer
Co-authored-by: slaren <slarengh@gmail.com>
* llama_model_loader: if n_tensors declared not equals to loaded tensors in split, throw an exception instead of asserting
* llama_model_loader: ensure mappings vector has the expected size
* llama_model_loader: use at instead of operator[] if this should never add to the map.
* llama_model_loader: immediately add the backend buffer to the model buffers in order to free them if an error occurs in the next allocation. Reserve the expected size.
* llama_model_loader: be sure the model mappings has enough capacity before allocating backend buffer
* llama_model_loader: fix map -> unordered map
* llama_split_prefix: use a clearer version, not pass split path len but dest max len.
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
* llama : minor
ggml-ci
* llama : introduce some typedef helpers
* docs: add model shard in hot topic
* llama_model_loader: put mapping in a unique_ptr from the moment it is allocated
Co-authored-by: slaren <slarengh@gmail.com>
* fix llama_split_prefix
---------
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
Information about the Command-R 35B model (128k context) can be found at:
https://huggingface.co/CohereForAI/c4ai-command-r-v01
Based on the llama2 model with a few changes:
1) New hyper parameter to scale output logits (logit_scale)
2) Uses LayerNorm instead of RMSNorm
3) Transfomer layers have a single shared LayerNorm that feeds into both the
self-attention and FFN layers in parallel. There is no post-attention LayerNorm.
4) No support for Rotary Position Embeddings (RoPE) scaling
5) No biases used
Find GGUF files here:
https://huggingface.co/andrewcanis/c4ai-command-r-v01-GGUF
To convert model to GGUF format yourself:
1) Download Command-R Hugging Face safetensors:
git lfs install
git clone https://huggingface.co/CohereForAI/c4ai-command-r-v01
2) Run:
python3 convert-hf-to-gguf.py --outtype f16 ./c4ai-command-r-v01
* mamba : begin working on support for Mamba SSM
* mamba : begin figuring out how to (ab)use the kv cache for Mamba
* mamba : recurrent inference almost works, but incoherent
* mamba : recurrent inference WORKS!!!
* convert : optionally use d_conv and d_state from config.json for Mamba
* mamba : refactor recurrent conv, resulting in 20% perf increase
It's still slower than I'd like, but I did not really optimize `ggml_exp` yet.
I also refactored `ggml_exp` to work with tensors with more than 2 dimensions.
* ggml : parallelize ggml_exp
This results in 8% faster token generation for Mamba-130M.
* mamba : simplify the conv step with a self-overlapping view
Turns out the conv_state can be made smaller by one column.
Note that this breaks existing GGUFs of Mamba,
because the key_value_length field is tied to the conv_state size.
Convolution with a self-overlapping view is cool!
And it's much simpler than what I initially thought would be necessary
to make the convolution step work with more than 1 token at a time.
Next step is to make the SSM step work on batches of tokens too,
and thus I need to figure out a way to make a parallel selective scan
which will keep the ssm_state small and won't make it bigger
by a factor of (n_layer * batch_size).
* llama : fix Mamba KV self size wrongly displaying as f16 instead of f32
Relatedly, I also tried to see if other types than f32 worked for the states,
but they don't, because of the operators used.
It's probably better anyway to keep lots of precision there,
since the states are small anyway.
* mamba : fix self-overlapping view depth stride
* mamba : handle batches of more than 1 token
This means running Mamba no longer crashes when using the default settings!
And probably also slightly faster prompt processing.
Both batched and non-batched processing yield the same output.
Previously, the state was not cleared when starting a sequence.
Next step is to make the KV cache API work as expected for Mamba models.
* ggml: add ggml_ssm_scan to help with parallel selective scan
If the selective scan was implemented without a custom operator,
there would be waaay too many nodes in the graph. For example,
for Mamba-130M, with a batch size of 512 (the default),
a naive selective scan could add at least 24*512=12288 nodes,
which is more than LLAMA_MAX_NODES (8192),
and that's only for the smallest Mamba model.
So it's much cleaner with a custom operator.
Not sure about the name, though.
* ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation
This will help with performance on CPU if ggml_vec_mul_f32
and ggml_vec_add_f32 are ever optimized with SIMD.
* mamba : very basic quantization support
Mostly works, but there is currently no difference
between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same).
Most of the SSM-specific weights can be kept in f32 without affecting
the size that much, since they are relatively small.
(the linear projection weights are responsible for most of Mamba's size)
Too much quantization seems to make the state degrade quite fast, and
the model begins to output gibberish.
It seems to affect bigger models to a lesser extent than small models,
but I'm not sure by how much.
Experimentation will be needed to figure out which weights are more important
for the _M (and _L?) variants of k-quants for Mamba.
* convert : fix wrong name for layer norm weight of offical Mamba models
I was using Q-bert/Mamba-* models before, which have a slighlty different
naming scheme for the weights.
(they start with "model.layers" instead of "backbone.layers")
* mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator
This increases performance on CPU by around 30% for prompt processing,
and by around 20% for text generation.
However, it also makes the ggml_exp and ggml_soft_plus operators unused.
Whether or not they should be kept will be decided later.
* convert : for Mamba, also consider the "MambaLMHeadModel" arch name
It's the name of the class of the official implementation,
though they don't use it (yet) in the "architectures" field of config.json
* mamba : fix vocab size problems with official models
The perplexity was waaaay to high for models with a non-round vocab size.
Not sure why, but it needed to be fixed in the metadata.
Note that this breaks existing GGUF-converted Mamba models,
but **only if** the vocab size was not already rounded.
* ggml : remove ggml_exp and ggml_soft_plus
They did not exist anyway outside of this branch,
and since ggml_ssm_scan fused operations together, they are unused.
It's always possible to bring them back if needed.
* mamba : remove some useless comments
No code change.
* convert : fix flake8 linter errors
* mamba : apply suggestions from code review
* mamba : remove unecessary branch for row-wise ssm_state and C multiplication
It was previously done to avoid permuting when only one token is processed
at a time (like when generating text), but permuting is cheap,
and dynamically changing the compute graph is not future-proof.
* ggml : in ggml_ssm_scan, use more appropriate asserts
* ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32
* mamba : multiple sequences, but one at a time
This is a step towards making this Mamba implementation usable
with the server example (the way the system prompt is kept when clearing
the client slots will need to be changed before this can work, though).
The KV cache size for this kind of model is tied to the maximum number
of sequences kept at any single time.
For now, this number is obtained from n_parallel (plus one,
to have an extra sequence to dedicate to the system prompt),
but there might be a better way to do this which won't also
make the main example use 2 cells even if only 1 is really used.
(for this specific case, --parallel 0 helps)
Simultaneous sequence processing will probably require changes to
ggml_ssm_scan, and possibly a new operator for the conv step.
* mamba : support llama_kv_cache_seq_cp
This (mis)uses the logic around K shifts, because tokens in a state
can't be shifted anyway, and because inp_K_shift has the right shape and type.
Using ggml_get_rows is a nice way to do copies, but copy chains can't work.
Fortunately, copy chains don't really seem to be used in the examples.
Each KV cell is dedicated to the sequence ID corresponding to its own index.
* mamba : use a state mask
It's cleaner than the previous heuristic of
checking for the pos of the first token in the batch.
inp_KQ_mask could not be re-used for this, because it has the wrong shape
and because it seems more suited to the next step of
simultaneous sequence processing (helping with the problem of
remembering which token belongs to which sequence(s)/state(s)).
* llama : replace the usage of n_ctx with kv_self.size in many places
* mamba : use n_tokens directly instead of n_tok
* mamba : in comments, properly refer to KV cells instead of slots
* mamba : reduce memory usage of ggml_ssm_scan
From 290.37 MiB to 140.68 MiB of CPU compute buffer size
with Mamba 3B with a batch size of 512.
The result tensor of ggml_ssm_scan was previously a big part
of the CPU compute buffer size. To make it smaller,
it does not contain the intermediate ssm states anymore.
Both y and the last ssm state are combined in the result tensor,
because it seems only a single tensor can be returned by an operator
with the way the graph is built.
* mamba : simultaneous sequence processing
A batch can now contain tokens from multiple sequences.
This is necessary for at least the parallel example, the server example,
and the HellaSwag test in the perplexity example.
However, for this to be useful, uses of llama_kv_cache_seq_rm/cp
will need to be changed to work on whole sequences.
* ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba
This operator makes it possible to use and update the correct states
for each token of the batch in the same way as ggml_ssm_scan.
Other solutions which use existing operators would need loops which would
add too many nodes to the graph (at least the ones I thought of).
Using this operator further reduces the size of the CPU compute buffer
from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512.
And (at least on CPU), it's a bit faster than before.
Note that "ggml_ssm_conv" is probably not the most appropriate name,
and it could be changed if a better one is found.
* llama : add inp_s_seq as a new input tensor
The most convenient implementation to select the correct state (for Mamba)
for each token is to directly get the correct index from a tensor.
This is why inp_s_seq is storing int32_t and not floats.
The other, less convenient way to select the correct state would be
to have inp_KQ_mask contain 1.0f for each state used by a token
and 0.0f otherwise. This complicates quickly fetching the first used
state of a token, and is also less efficient because a whole row
of the mask would always need to be read for each token.
Using indexes makes it easy to stop searching when there are
no more sequences for a token, and the first sequence assigned
is always very quickly available (it's the first element of each row).
* mamba : support llama_kv_cache_seq_cp copy chains
* mamba : support shifting and dividing the kv cache pos
* mamba : make the server and parallel examples work with whole sequences
A seq_id is dedicated to the system prompt in both cases.
* llama : make llama_kv_cache_seq_rm return whether it succeeded or not
* mamba : dedicate an input tensor for state copy indices
This is cleaner and makes it easier to adapt when/if token positions
(and by extension, inp_K_shift) are no longer integers.
* mamba : adapt perplexity, batched, and batched-bench examples
* perplexity : limit the max number of sequences
This adapts to what the loaded model can provide.
* llama : add llama_n_max_seq to get the upper limit for seq_ids
Used by the perplexity example.
* batched : pass n_parallel to the model's context params
This should have been there already, but it wasn't.
* batched-bench : reserve sequences to support Mamba
* batched-bench : fix tokens being put in wrong sequences
Generation quality isn't what's measured in there anyway,
but at least using the correct sequences avoids using non-consecutive
token positions.
* mamba : stop abusing attention metadata
This breaks existing converted-to-GGUF Mamba models,
but will allow supporting mixed architectures like MambaFormer
without needing to break Mamba models.
This will also allow changing the size of Mamba's states
without having to reconvert models in the future.
(e.g. using something else than d_conv - 1 columns for the conv_states
will not require breaking existing converted Mamba models again)
* gguf-py : add new KV metadata key-value pairs for Mamba
* llama : add new metadata key-value pairs for Mamba
* llama : guard against divisions by zero when n_head is 0
* mamba : rename "unlimited" KV cache property to "recurrent"
* mamba : more correctly update the "used" field of the KV cache
* ggml : in ggml_ssm_scan, use a threshold for soft_plus
This is how the official Mamba implementation does it,
and it's also what torch.nn.Softplus does.
* convert : for Mamba, fallback to internal NeoX tokenizer
The resulting models are exactly the same
as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there.
* mamba : support state saving and restoring
* ggml : implicitly pass src tensors through dst for Mamba-related ops
* mamba : clarify some comments
* server : fix cache_tokens not getting correctly resized
Otherwise, when the "we have to evaluate at least 1 token" special case
was triggered, an extra token was kept in cache_tokens even if it was
removed from the KV cache.
For Mamba, this caused useless prompt reprocessing when the previous
request triggered the above case.
* convert-hf : support new metadata keys for Mamba
For the models available at
https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406
* mamba : rename metadata to be more similar to transformers library
This breaks existing converted-to-GGUF models,
but the metadata names are more "standard".
* mamba : support mamba-*-hf models
These models share their token_embd.weight with their output.weight
* mamba : add missing spaces
This is purely a formatting change.
* convert-hf : omit output.weight when identical with token_embd.weight
Only for Mamba for now, but it might be relevant for other models eventually.
Most Mamba models actually share these two tensors, albeit implicitly.
* readme : add Mamba to supported models, and add recent API changes
* mamba : move state_seq and state_mask views outside layer loop
A few tensors were also missing `struct` in front of `ggml_tensor`.
* server: docs - refresh and tease a little bit more the http server
* Rephrase README.md server doc
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update examples/server/README.md
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update examples/server/README.md
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update README.md
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
There are couple things in this architecture:
1. Shared input and output embedding parameters.
2. Key length and value length are not derived from `n_embd`.
More information about the models can be found at
https://ai.google.dev/gemma. GGUFs can be downloaded from
https://huggingface.co/google.
