Allow exporting a view of the KV cache #4180
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
This may be useful for stuff other than just debugging as well. For example, you could use the view to calculate exactly what the largest possible batch size is without having to call |
|
This is amazing, seems to fix #4035 then! |
ggerganov
approved these changes
Nov 23, 2023
common/common.cpp
Outdated
| // | ||
|
|
||
| void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) { | ||
| printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d\n", |
There was a problem hiding this comment.
Isn't it better to use LOG_TEE instead of printf?
There was a problem hiding this comment.
Sure, I can make that change. I find the logging behavior sort of confusing. I.E. if I compile without logging I'll see the LOG_TEE output. However, if I compile with logging but --disable-logs it will just never show up on the console.
There was a problem hiding this comment.
Hmm, actually... Using the LOG stuff is possibly sort of weird if I'm printing stuff out in pieces rather than line by line like those functions do. What do you think, still change it?
Sure, just keep in mind the example dump functions I made are in common and not in the However, just exporting the KV cache view allows writing functions like that so you could dump however you want in your application that uses Unrelated thought: It would be nice we could also export the token for a slot as well. However, this would require changing the actual KV cache structure. I was thinking it was per sequence, but that's actually wrong isn't it. The whole cell has one token id, or am I just crazy? If I'm not crazy, then that actually would be really easy to add. If I am crazy, then it would require changing the |
That is fine for me. And it may still be a good alternative example dump for others.
That would be amazing. And as far as I understand one token embedding will lead to one KV cache entry. |
Make dump functions deal with lengths or sequences counts > 10 better
Eliminate cell sequence struct; use llama_seq_id directly Minor cleanups
|
Okay, I think I'm about done messing with this. I can change the Actually, I guess there's possibly one other change that could potentially be useful: adding a boolean flag to the view update function that controls where it populates the individual cell data. Since this can now be used to calculate the max permissible batch size, it occurs to me we might want to call it and just generate the stats like that, not necessarily generate the whole detailed view of the cache. |
|
We just need to put the KV dumping behind a command line flag, because it would likely affect performance so it should be enabled only on demand.
Hm, I'm not following. Can you clarify? |
The example of just dumping every decode in
Currently when you call In addition to copying the individual cell information, it populates some informational fields in the view struct: // An updateable view of the KV cache.
struct llama_kv_cache_view {
// Number of KV cache cells. This will be the same as the context size.
int32_t n_cells;
// Maximum number of sequences that can exist in a cell. It's not an error
// if there are more sequences in a cell than this value, however they will
// not be visible in the view cells_sequences.
int32_t n_max_seq;
// Number of tokens in the cache. For example, if there are two populated
// cells, the first with 1 sequence id in it and the second with 2 sequence
// ids then you'll have 3 tokens.
int32_t token_count;
// Number of populated cache cells.
int32_t used_cells;
// Maximum contiguous empty slots in the cache.
int32_t max_contiguous;
// Index to the start of the max_contiguous slot range. Can be negative
// when cache is full.
int32_t max_contiguous_idx;
So what I was saying is using this view might be useful just for calculating what the batch size should be (instead of trying the normal batch size value, then trying batch_size / 2, etc). However, for that use case, the application won't care about stuff in the individual cells: it'll just care about the statistics in the view struct. So I was suggesting a boolean flag for the update function that could allow it to skip populating the individual cell data (but still generate stuff like |
|
Ok got it. I don't think it's necessary. I think that KV cache views should be used just for debugging stuff - we don't want people to start building logic around it. |
Fair enough. I think it would be good if there was some way for the application to find out what the max permissible batch size was, without having to call with some value and then guess at what to try next if that fails. Maybe the KV cache view isn't the right place for that, but it currently could be used to calculate the exact value without any guesswork. edit: By the way, sorry, I didn't remove the stuff in |
|
I'm hoping we will find a better way in the future that would not involve searching for a big enough spot. That's why I don't want to build too much around this idea for now. |
ggerganov
added a commit
that referenced
this pull request
Nov 23, 2023
* llama : keep track of used KV cells + better KV cache management * llama : zero KV cache used upon clear ggml-ci * llama : allow exporting a view of the KV cache (#4180) * Allow exporting a view of the KV cache * Allow dumping the sequences per cell in common * Track max contiguous cells value and position as well * Fix max contiguous empty cells index calculation Make dump functions deal with lengths or sequences counts > 10 better * Fix off by one error in dump_kv_cache_view * Add doc comments for KV cache view functions Eliminate cell sequence struct; use llama_seq_id directly Minor cleanups * common : add -dkvc arg for enabling kv cache dumps --------- Co-authored-by: Kerfuffle <[email protected]>
hodlen
added a commit
to hodlen/llama.cpp
that referenced
this pull request
Apr 1, 2024
llama : restore prefix space in llama tokenizer (ggerganov#4081) gguf : fix potential infinite loops while parsing (ggerganov#4100) Co-authored-by: Bernhard Gstrein <[email protected]> Respect tokenizer.ggml.add_bos_token value when tokenizing (ggerganov#4040) * gguf-py: gguf-dump: Respect --no-tensor flag in JSON mode. * Respect add_bos_token GGUF metadata value * gguf-py: Try to fix SpecialVocab giving up too easily for the Nth time llama : fix data units (ggerganov#4101) * llama : fix data units ggml-ci * Revert "llama : fix data units" This reverts commit f5feac8. * llama : disambiguate data units ggml-ci cuda : get_row_rounding F32 (ggerganov#4095) * Fix ggerganov#4017 * Update ggml-cuda.cu Co-authored-by: Jared Van Bortel <[email protected]> * Update ggml-cuda.cu Co-authored-by: Jared Van Bortel <[email protected]> --------- Co-authored-by: Jared Van Bortel <[email protected]> finetune : zero the loraB initial vectors (ggerganov#4082) * finetune : zero the loraB initial vectors Without this, the first iteration is starting out far from the base model, instead of exactly on it. Zeroing loraB is what the paper recommends. loralib also zeroes at least one of the init vector pairs (though it departs from the paper in using a different distribution for the other vector, in some cases). * tabs to spaces * Use ggml_set_zero instead of adding a new function finetune : speed-up ggml_compute_forward_out_prod_f32 via BLAS (ggerganov#4079) * Remove logically superfluous assertions and order by dimension * Use cblas_sgemm() to implement ggml_compute_forward_out_prod() * Remove ggml_compute_forward_out_prod_use_blas(), fix compiling errors on cmake/zig, remove trailing whitespace * Add openBLAS support for sgemm() in compute_forward_out_prod() llama : add functions to get the model's metadata (ggerganov#4013) * llama : add functions to get the model's metadata * format -> std::to_string * better documentation train : move number of gpu layers argument parsing to common/train.cpp (ggerganov#4074) - introduces help entry for the argument - cuts '--gpu-layers' form in order to simplify usage and documentation. Signed-off-by: Jiri Podivin <[email protected]> Co-authored-by: Jiri Podivin <[email protected]> py : remove superfluous import statements (ggerganov#4076) Signed-off-by: Jiri Podivin <[email protected]> Co-authored-by: Jiri Podivin <[email protected]> llava : fix compilation warning that fread return value is not used (ggerganov#4069) common : improve yaml log escaping (ggerganov#4080) * logging: improve escaping in yaml output * logging: include review feedback py : Falcon HF compatibility (ggerganov#4104) Falcon HF compatibility convert : use 'model' value if it exists. This allows karpathy/tinyllamas to load (ggerganov#4089) Co-authored-by: Don Mahurin <@> examples : add tokenize (ggerganov#4039) tokenize : fix trailing whitespace build : support ppc64le build for make and CMake (ggerganov#3963) * build: support ppc64le build for make and CMake * build: keep __POWER9_VECTOR__ ifdef and extend with __powerpc64__ Co-authored-by: Georgi Gerganov <[email protected]> --------- Co-authored-by: Georgi Gerganov <[email protected]> llama : increase max nodes (ggerganov#4115) Clean up ggml-cuda.cu warnings when compiling with clang (for ROCM) (ggerganov#4124) * ggml-cuda.cu: Clean up warnings when compiling with clang * ggml-cuda.cu: Move static items into anonymous namespace * ggml-cuda.cu: Fix use of namespace start macro * Revert "ggml-cuda.cu: Fix use of namespace start macro" This reverts commit 26c1149. * Revert "ggml-cuda.cu: Move static items into anonymous namespace" This reverts commit e29757e. scripts : Remove missed baichuan convert script (ggerganov#4127) tokenize example: Respect normal add BOS token behavior (ggerganov#4126) Allow building with Makefile gguf-py : export chat templates (ggerganov#4125) * gguf-py : export chat templates * llama.cpp : escape new lines in gguf kv info prints * gguf-py : bump version * gguf-py : check chat_template type * gguf-py : initialize chat_template gitignore : tokenize common : comma should be semicolon (ggerganov#4137) server : relay error messages (ggerganov#4131) finetune : add --n-gpu-layers flag info to --help (ggerganov#4128) Revert "finetune : add --n-gpu-layers flag info to --help (ggerganov#4128)" This reverts commit 05e8301. speculative : fix prompt tokenization in speculative example (ggerganov#4025) * Support special tokens and not adding BOS to prompt in speculative * Adapt to new should_add_bos function * Ensure tgt and dft have same add_bos setting ci : add flake8 to github actions (python linting) (ggerganov#4129) Disabled rules: * E203 Whitespace before ':' - disabled because we often use 'C' Style where values are aligned * E211 Whitespace before '(' (E211) - disabled because we often use 'C' Style where values are aligned * E221 Multiple spaces before operator - disabled because we often use 'C' Style where values are aligned * E225 Missing whitespace around operator - disabled because it's broken so often it seems like a standard * E231 Missing whitespace after ',', ';', or ':' - disabled because we often use 'C' Style where values are aligned * E241 Multiple spaces after ',' - disabled because we often use 'C' Style where values are aligned * E251 Unexpected spaces around keyword / parameter equals - disabled because it's broken so often it seems like a standard * E261 At least two spaces before inline comment - disabled because it's broken so often it seems like a standard * E266 Too many leading '#' for block comment - sometimes used as "section" separator * E501 Line too long - disabled because it's broken so often it seems like a standard * E701 Multiple statements on one line (colon) - broken only in convert.py when defining abstract methods (we can use# noqa instead) * E704 Multiple statements on one line - broken only in convert.py when defining abstract methods (we can use# noqa instead) main : Add ChatML functionality to main example (ggerganov#4046) Co-authored-by: Sebastian Cramond <[email protected]> readme : update ROCm Windows instructions (ggerganov#4122) * Update README.md * Update README.md Co-authored-by: Jared Van Bortel <[email protected]> --------- Co-authored-by: Jared Van Bortel <[email protected]> finetune - update readme to mention llama support only (ggerganov#4148) stablelm : simplify + speedup generation (ggerganov#4153) docs : add llama-star arch idea examples : fix typo in parallel example doc comment (ggerganov#4181) Signed-off-by: Daniel Bevenius <[email protected]> readme : update hot topics llama : KV cache view API + better KV cache management (ggerganov#4170) * llama : keep track of used KV cells + better KV cache management * llama : zero KV cache used upon clear ggml-ci * llama : allow exporting a view of the KV cache (ggerganov#4180) * Allow exporting a view of the KV cache * Allow dumping the sequences per cell in common * Track max contiguous cells value and position as well * Fix max contiguous empty cells index calculation Make dump functions deal with lengths or sequences counts > 10 better * Fix off by one error in dump_kv_cache_view * Add doc comments for KV cache view functions Eliminate cell sequence struct; use llama_seq_id directly Minor cleanups * common : add -dkvc arg for enabling kv cache dumps --------- Co-authored-by: Kerfuffle <[email protected]> Fix incorrect format strings and uninitialized variables. (ggerganov#4133) * Fix incorrect format strings and uninitialized variables. * Address comments * Add the missing include statement readme : use PATH for Windows ROCm (ggerganov#4195) * Update README.md to use PATH for Windows ROCm * Update README.md * Update README.md main.swift : fix eos checking (ggerganov#4197) llama_token_eos(const struct llama_model *) is currently getting struct llama_context type variable context as a parameter. convert : fix tensors using grad in some models (ggerganov#4173) ggml-cuda : support stablelm rope (ggerganov#4156) * ggml-cuda : support stablelm rope * remove unused freq_base kernel parameter * add n_dims parameter to llm_build_k_shift, default to n_rot via overload * llama : fix llm_build_k_shift args --------- Co-authored-by: Georgi Gerganov <[email protected]> llama : set metal log callback correctly (ggerganov#4204) server : OAI API compatibility (ggerganov#4198) * Add openai-compatible POST /v1/chat/completions API endpoint to server example * fix code style * Update server README.md * Improve server README.md * Fix server.cpp code style according to review * server : some style changes * server : indentation * server : enable special tokens during tokenization by default * server : minor code style * server : change random string generator * straightforward /v1/models endpoint --------- Co-authored-by: kir-gadjello <[email protected]> Co-authored-by: Tobi Lütke <[email protected]> readme : update hot topics Update docs for yarn_ext_factor <0.0 as unspecified instead of NaN (ggerganov#4189) llama : grammar `reserve` space in `decode_utf8` (ggerganov#4210) * reserve space for codepoints * improvement for the appended 0 scripts : Use mmap in torch load (ggerganov#4202) * Use mmap in torch load, prefer .bin files when loading * Revert .bin > .safetensors preference metal : fix yarn (ggerganov#4220) get the correct n_orig_ctx in metal lookahead : add example for lookahead decoding (ggerganov#4207) * lookahead : init * lookahead : generate and store n-grams * lookahead : use loop instead recursion to generate n-grams * lookahead : initial working implementation * lookahead : filter repeating n-grams * lookahead : use deterministic init * lookahead : add to Makefile * lookahead : fix a bug in the seq_id of the lookahead tokens * lookahead : add comments --------- Co-authored-by: slaren <[email protected]> readme : update hot topics lookahead : support `-n -1` infinite generation ggml : fix -Warray-bounds warning with gcc (ggerganov#4231) examples : iOS example with swift ui (ggerganov#4159) * copy to llama.cpp as subdir * attempt enabling metal, fails * ggml metal compiles! * Update README.md * initial conversion to new format, utf8 errors? * bug fixes, but now has an invalid memory access :( * added O3, now has insufficient memory access * begin sync with master * update to match latest code, new errors * fixed it! * fix for loop conditionals, increase result size * fix current workflow errors * attempt a llama.swiftui workflow * Update .github/workflows/build.yml Co-authored-by: Georgi Gerganov <[email protected]> --------- Co-authored-by: Georgi Gerganov <[email protected]> readme : add Amica to UI list (ggerganov#4230) cmake : fix issue with version info not getting baked into LlamaConfig.cmake (ggerganov#3970) * Split CPP generation from build-info query * Remove blank lines * Add BUILD_SHARED_LIBS option ggml : re-enable BLAS for CPU when src0 != F32 + remove redundant full offload checks in llama.cpp (ggerganov#4240) * ggml : use blas even if src0 is not F32 * llama : use n_threads_batch only when n_tokens >= 32 ggml-ci * llama : revert n_threads_batch logic ggml-ci ggml : restore abort() in GGML_ASSERT (ggerganov#4242) readme : add FreeChat (ggerganov#4248) examples : add readme files py : fix oai proxy (ggerganov#3972) * fix oai proxy fix generation not stoped while bot stop talking in chat mode fix possible `slot_id` not exist response for cors (and pre flight) * oai proxy: workaround for some client (such as Chatbox) * use stop as separator to replace hardcoded `\n` llama : fix typical sampling (ggerganov#4261) Typical sampling was broken because after copying new_candidates into canditates, the "sorted" bool is left at "true", but the new data is no longer sorted according to probability. Patch to set "sorted" to false. Test: Generating with temp=0.0001 (approx. argmax) should generate the same sequence at typical>=1.0 and typical=0.9999 (approx. disabled, but enters the typical sampling codepath). convert.py : fix llama/llama2 conversion due to vocab_size=-1 (ggerganov#4258) llama : fix alignment of general.name in print meta (ggerganov#4254) * llama: fix alignment of general.name in print meta This commit fixes the alignment of the general.name field in the llm_load_print_meta function. Currently the output looks like this: ```console llm_load_print_meta: model ftype = mostly Q4_0 llm_load_print_meta: model params = 13.02 B llm_load_print_meta: model size = 6.86 GiB (4.53 BPW) llm_load_print_meta: general.name = LLaMA v2 ``` And with this commit it looks like this: ```console llm_load_print_meta: model ftype = mostly Q4_0 llm_load_print_meta: model params = 13.02 B llm_load_print_meta: model size = 6.86 GiB (4.53 BPW) llm_load_print_meta: general.name = LLaMA v2 ``` Signed-off-by: Daniel Bevenius <[email protected]> * llama: fix alignment of special tokens Signed-off-by: Daniel Bevenius <[email protected]> --------- Signed-off-by: Daniel Bevenius <[email protected]> readme : fix typo (ggerganov#4253) llama.cpp uses GitHub Actions, not Gitlab Actions. cmake : fix the metal file foder path (ggerganov#4217) batched.swift : update README.md (ggerganov#4214) docs: update how to run docker : add finetune option (ggerganov#4211) readme : fix (ggerganov#4135) * fix: readme * chore: resolve comments * chore: resolve comments main : pass LOG_TEE callback to llama.cpp log (ggerganov#4033) * main : Call llama_log_set to use LOG_TEE * tabs to spaces llava : ShareGPT4V compatibility (vision encoder only loading) (ggerganov#4172) * ShareGPT4 compatibility (vision encoder only loading) Load only a CLIP vision encoder (as supplied by ShareGPT finetunes) Corrects the argument parsing for --img_mean and --img_std (which were previously not parsed but attempted to access) Defines defaults for img_mean and img_std which are equal to the llava 1.5 CLIP encoder, so you do not have to provide them * Update convert-image-encoder-to-gguf.py build : fix build info generation and cleanup Makefile (ggerganov#3920) * cmake : fix joining of REAL_GIT_DIR * fix includes with help from include-what-you-use * make : remove unneeded deps and add test-rope target * fix C includes in C++ source files * Revert "fix includes with help from include-what-you-use" This reverts commit 635e9fa. make : fix Apple clang determination bug (ggerganov#4272) Co-authored-by: Will Findley <[email protected]> server : add single-client multi-prompt support (ggerganov#4232) * * add multiprompt support * * cleanup * * more cleanup * * remove atomicity of id_gen, and change lock_guard to unique_lock on completion requests * * remove all references to mutex_multitasks * Update examples/server/server.cpp Co-authored-by: Jared Van Bortel <[email protected]> * Update examples/server/server.cpp Co-authored-by: Jared Van Bortel <[email protected]> * Update examples/server/server.cpp Co-authored-by: Jared Van Bortel <[email protected]> * Update examples/server/server.cpp Co-authored-by: Jared Van Bortel <[email protected]> * * change to set --------- Co-authored-by: Jared Van Bortel <[email protected]> server : add --log-disable to disable logging to file (ggerganov#4260) * * add --log-disable to disable logging to file in the server example * * typo fix ggml : add ggml_soft_max_ext (ggerganov#4256) * metal : implement soft_max_ext * cuda : implement soft_max_ext * ggml : implement soft_max_ext (CPU) * batched-bench : print threads ggml-ci * metal : simplify soft_max encoding ggml-ci * cuda : use 512 threads for soft_max instead of 32 * ggml : update soft max cpu * cuda : do warp-based block reduce * cuda : increase max block size to 1024 * cuda : fix warp reduction initialization of shared mem * metal : warp-based reduction for soft max kernel * metal : warp-based reduce for rms_norm * metal : simplify soft max kernel ggml-ci * alloc : fix build with debug py : add requirements file for convert-hf-to-gguf.py (ggerganov#4277) This commit adds a requirements file for the convert-hf-to-gguf.py script, and also add the torch and transformers packages to it. The motivation for this is that currently running convert-hf-to-gguf.py will produce the following error: ```console $ python3 -m venv venv $ source venv/bin/activate (venv) $ pip install -r requirements.txt Collecting numpy==1.24.4 Collecting sentencepiece==0.1.98 Collecting gguf>=0.1.0 Installing collected packages: sentencepiece, numpy, gguf Successfully installed gguf-0.5.1 numpy-1.24.4 sentencepiece-0.1.98 (venv) $ python convert-hf-to-gguf.py --help Traceback (most recent call last): File "llama.cpp/convert-hf-to-gguf.py", line 16, in <module> import torch ModuleNotFoundError: No module named 'torch' ``` With this commit, and using requirements-hf-to-gguf.txt instead of requirements.txt, the script can be run and shows the help output. Signed-off-by: Daniel Bevenius <[email protected]> llama : fix integer overflow during quantization (ggerganov#4284) happens with multi-threaded quantization of Qwen-72B ggml-ci llama : add Qwen support (ggerganov#4281) * enable qwen to llama.cpp * llama : do not GPU split bias tensors --------- Co-authored-by: Georgi Gerganov <[email protected]> llama : support attention bias on LLaMA architecture (ggerganov#4283) * Support attention_bias on LLaMA architecture QKVO bias, should fix InternLM (ggerganov#3133) and works for LLaMAfied Qwen models (ggerganov#3743 (comment)). * check existence of qkvo bias while loading llama models Tested on LLaMA2, CUDA and CPU. * Update llama.cpp build : enable libstdc++ assertions for debug builds (ggerganov#4275) swift : fix token_to_piece implementation (ggerganov#4278) * Fix token_to_piece implementation in Swift * Fix errors llama : support optional tensors (ggerganov#4283) llama : avoid using "optional" keyword (ggerganov#4283) llama : pad KV cache size (ggerganov#4280) * llama : pad KV cache size to 32 * metal : try to improve batched decoding py : add grammar to oai like api (ggerganov#4294) server : fix OpenAI API `stop` field to be optional (ggerganov#4299) (cherry picked from commit Mozilla-Ocho/llamafile@e8c92bc) ggml : fix soft max out-of-bounds access (ggerganov#4307) ggml-ci ggml : reuse ggml_get_n_tasks() in ggml_graph_plan() (ggerganov#4308) * ggml : fix soft max out-of-bounds access ggml-ci * ggml : reuse ggml_get_n_tasks() in ggml_graph_plan() ggml-ci grammar-parser : fix typo (ggerganov#4318) preceeding -> preceding swift : fix prompt tokenization logic (ggerganov#4321) swift : fix concatenation method to avoid invalid UTF8 stringfication (ggerganov#4325) simple : update error message for KV cache check (ggerganov#4324) This commit updates the error message that is printed when the KV cache is not big enough to hold all the prompt and generated tokens. Specifically it removes the reference to n_parallel and replaces it with n_len. Signed-off-by: Daniel Bevenius <[email protected]> swift : revert compiler checks for swift package (ggerganov#4332) sampling : custom samplers order (ggerganov#4285) * Samplers sequence order w parameter * Cleaned commented code * Fixed formatting * Rewrote with unordered_map * Revert and rewrite, too many problems and safeguards would be needed * Fixed code style * Code style fixes according to review * More readable samplers input string, fixed help * Style fix in sampler_queue * Formatting fixes * Fixing whitespaces llama : allow overriding GGUF metadata when loading model (ggerganov#4092) * feat: Allow overriding GGUF metadata when loading model * Fix the one time GCC is stricter than clang about something * Step1 * Refactor... basically everything! * Nuke obsolete GetArrayLen struct * simplify std::string specialization * Various cleanups Add informational output when overrides are applied Warn user when an override with the wrong type is specified * Fix broken logic for parsing bool KV overrides Fix issue where overrides didn't apply when key missing in GGUF metadata Resolve merge changes * llama : rearrange model params * Update new GET_KEY call Add note that metadata KV overrides aren't reflected in initial metadata KV info dump --------- Co-authored-by: cebtenzzre <[email protected]> Co-authored-by: Georgi Gerganov <[email protected]> grammar : pre-computed pieces + reserve mem + less string copies (ggerganov#4330) * reserve space for codepoints * improvement for the appended 0 * used precomputed token text for grammar sample * reserve canidates_decoded * reserve canidates_grammar * remove candidates_decoded * Revert "remove candidates_decoded" This reverts commit 3773328. * changed decode_utf8 to take src by ref speculative : support `--color` (ggerganov#4343) * speculative: add some colors * minor : add braces --------- Co-authored-by: Georgi Gerganov <[email protected]> common : fix compile warning server : recognize cache_prompt parameter in OAI API (ggerganov#4347) train : fix ggerganov#4227 (double free in examples/train-text-from-scratch/train-text-from-scratch.cpp) (ggerganov#4351) On commit b1108 (44c117f) xaedes added ggml_allocr * alloc = NULL; ... (many lines in between) if (alloc) { ggml_allocr_free(alloc); } Which is correct, but it's easy to lose context after many lines in between. On commit b1287 (0e76a899) xaedes made a big change. From here on, alloc is freed eagerly. alloc = ggml_allocr_new(...) ... (short lines of code) ggml_allocr_free(alloc) This happens a few times, but alloc is never set to NULL, and many lines below, we still have if (alloc) { ggml_allocr_free(alloc); } which causes a double-free. llama : per-layer KV cache + quantum K cache (ggerganov#4309) * per-layer KV * remove unnecessary copies * less code duplication, offload k and v separately * llama : offload KV cache per-layer * llama : offload K shift tensors * llama : offload for rest of the model arches * llama : enable offload debug temporarily * llama : keep the KV related layers on the device * llama : remove mirrors, perform Device -> Host when partial offload * common : add command-line arg to disable KV cache offloading * llama : update session save/load * llama : support quantum K cache (ggerganov#4312) * llama : support quantum K cache (wip) * metal : add F32 -> Q8_0 copy kernel * cuda : add F32 -> Q8_0 copy kernel ggml-ci * cuda : use mmv kernel for quantum cache ops * llama : pass KV cache type through API * llama : fix build ggml-ci * metal : add F32 -> Q4_0 copy kernel * metal : add F32 -> Q4_1 copy kernel * cuda : wip * cuda : add F32 -> Q4_0 and F32 -> Q4_1 copy kernels * llama-bench : support type_k/type_v * metal : use mm kernel only for quantum KV cache * cuda : add comment * llama : remove memory_f16 and kv_f16 flags --------- Co-authored-by: slaren <[email protected]> * readme : add API change notice --------- Co-authored-by: slaren <[email protected]> sync : ggml (new ops, tests, backend, etc.) (ggerganov#4359) * sync : ggml (part 1) * sync : ggml (part 2, CUDA) * sync : ggml (part 3, Metal) * ggml : build fixes ggml-ci * cuda : restore lost changes * cuda : restore lost changes (StableLM rope) * cmake : enable separable compilation for CUDA ggml-ci * ggml-cuda : remove device side dequantize * Revert "cmake : enable separable compilation for CUDA" This reverts commit 09e35d0. * cuda : remove assert for rope * tests : add test-backend-ops * ggml : fix bug in ggml_concat * ggml : restore `ggml_get_n_tasks()` logic in `ggml_graph_plan()` * ci : try to fix macOS * ggml-backend : remove backend self-registration * ci : disable Metal for macOS cmake build ggml-ci * metal : fix "supports family" call * metal : fix assert * metal : print resource path ggml-ci --------- Co-authored-by: slaren <[email protected]> grammar : revert the replacement of llama_token_to_piece with id_to_token (ggerganov#4396) Update README.md (ggerganov#4388) Fix small typo. ggml : increased GGML_MAX_PARAMS to allow finetuning of 70b models (ggerganov#4424) server : fix local model name in server (ggerganov#4420) llama : document logits_all deprecation (ggerganov#4418) llama_context_params.logits_all is a parameter for controlling llama_eval. This documents that logits_all should not be used with llama_decode and llama_batch. build : target Windows 8 for standard mingw-w64 (ggerganov#4405) * build : target Windows 8 for standard mingw-w64 * make : fix missing console.o deps This was causing a link error with `make all` on Windows. english : use `typos` to fix comments and logs (ggerganov#4354) server : tweak default sampling parameters (ggerganov#4367) * Set a more typical Top P setting as the default * Update temp max llama : add Mixtral support (ggerganov#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (ggerganov#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: ggerganov/ggml@1914017 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <[email protected]> Co-authored-by: Radek Pilar <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For example, dumping with
dump_kv_cache_viewlooks like this after runningparallel -np 3 -cb -ns 50 -c 700for a while:The numbers represent the number of sequences each cell is populated with.
edit: Also added a way to dump the sequences in each cell:
edit: I added tracking for maximum contiguous empty KV cells and the start position. "max contiguous cells=5 @ 402" - 5 empty cells, starting at index
402