As c doesnt have the concept of pass by reference, and inturn the
existing c api uses pointers wrt llama chat message structure, so
switching to same wrt chat_tmpl_apply logics.
Also fix a oversight in previous commit and add the remaining logic.
Initial skeletons
Update existing logics to help with same. Also the inbetween helper
was having a bad signature wrt returning status and data, thats also
fixed.
While sending the current chat session along with new user query
to the model, many models expect that a tag be added at the end
to indicate that user is expecting the model to respond, this
flags allows for the same.
Add a c api wrapper for a single message tagging scenario.
Inturn to match convention followed by existing chat_apply_template
code, make it return the size expected of the tagged message string
buffer. Update internal single logic to help with same.
Explicitly check if tmpl specified is available in the loaded json
or not and then return a error if not found.
Fix a oversight wrt key name.
Add a alert in case if passed meta json file contains begin(BoS)
wrt assistant role, similar to check for end (EoS) wrt user role.
Bcas normally both (ie EoS wrt User and BoS wrt Assistant) shouldnt
be needed.
Update main wrt begin & prefix and suffix & end addition.
Move helpers to the begining, so can avoid adding prototype
declerations/function signatures to the begining
Get the char * wrt string data in the c++ string.
Also fix a oversight wrt begin, when flag based begin adding control
was introduced.
NOTE: Currently system role suffix/end conditional adding always
triggered, if 1st system prompt seen or additional system prompt
is seen.
Now there is a simple and extended version of returning tagged
messages.
The extended version returns the tagged string, as well as the
details of the parts that make up that tagged message interms of
the type of parts and the lengths of the parts.
Now there is a simple and extended version of returning tagged
message wrt a single role and its content.
The extended version returns the tagged string, as well as the
details of the parts that make up that tagged message interms of
the type of parts and the lengths of the parts.
Dump shows user->begin.
chat-template-apply[-single] updated to work with begin and prefix
TODO: need to wrap begin in a try-catch, so that irrespective of
role, begin+prefix will work, irrespoective of whether that role
has a begin entry or not.
Was looking at the tokenized vector, and noticed that the EOS
mentioned by existing chat_apply_template of llama.cpp, is different
from what I noticed in tokenizer_config.json of deepseek llm, so
I have added two entries
* "deepseek-alt" which matches llama.cpp's chat_apply_template and
* "deepseek" which matches that in tokenizer_config.json.
This impacts the assistant suffix and reverse prompt entries.
CasOfThis: Need to look into other entries which I added previously
at a later time. However as the default logic should be picking the
EOS from model file, so I assume reverse-prompt being outofsync,
may not matter beyond a limit, potentially.
rename because they return value of specified key.
[main] update metaok to take template-id, so that one can cross
check that all needed entries are there wrt that template-id in
the chaton-meta-json file
Update the note
Rename global-prefix|suffix to global-begin|end.
Rename chat-apply-template to chat-apply-template-single, cas it
handles only a single message.
Add some debug log messages to the helper functions
* ggml : add ggml_flash_attn_ext API
* ggml : fix GQA support in ggml_flash_attn_ext
* ggml : online attention (CPU)
* metal : initial implementation
* metal : f16 precision
* metal : reduce branches
* metal : specialize for head size
* wip : 8 rows per simd group
* wip : 4 rows per simd group
* wip : template for rows per warp
* metal : parallelize across KV size
* metal : parallel reduce across heads
* metal : efficient flash_attn_f16 implementation
* metal : avoid redundant loads of the attention
* metal : scale and mask in matrix form
* metal : fix comment
* llama : avoid ggml_cast, use F32 query
* metal : add parallel reduce version (disabled)
* metal : move output into local memory + optimize
- the result from each simdgroup now stays in the registers
- significantly reduced SRAM usage
- more efficient skipping of -INF blocks
- avoid simdgroup barrier in hot loop
- add comments
* metal : add tests, fix scaling, support C > 32
* metal : improve precision
* ggml : fix f16 mad
* metal : minor
* metal : support Q > 8
* tests : add ATTN tests
* metal : disable buffer allocation logs
* tests : more
* metal : faster inner loop for C == 32
* metal : fix array initialization
* tests : ifdef
* ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext
* ggml : fix ggml_soft_max mask requirement
* cuda : fix soft_max to use correct mask size
* cuda : add flash_attn kernel (wip)
* metal : optimize softmax for C > 32
* metal : optimize softmax
* tests : minor fix
* cuda : avoid zeroing fragments
* tests : update dims
* cuda : fix __hisinf() result check
* cuda : avoid warp_reduce for smax
* cuda : use int instead of int64_t
Noticeably improves performance (thanks to Johannes)
* cuda : make loops use the same loop values
Thanks Johannes again for the tip
* cuda : unroll some of the loops
* cuda : avoid __hisinf branches
* cuda : use half2 in softmax
* cuda : switch to 1 warp for bs > 16
* cuda : speed-up reduce part of the kernel
* cuda : unroll Q*K^T loop
* cuda : fix -INF block check
* cuda : simplify softmax
* cuda : fix matrix names
* cuda : minor
* llama : adapt to F16 KQ_pos
* llama : adapt new models to F16 KQ_mask
* ggml : fix F16 store (ARM NEON)
* llama : fix type of KQ_mask and KQ_pos
* ggml : fix CPU soft_max
* tests : add hs=256
* cuda : fix build
* metal : improve perf via smaller int registers
* cuda : adapt soft_max to F16 mask and pos
* CUDA: faster FlashAttention, kernel for bs == 1
* 16 cols for Phi-2
* no vec for hs, no hs==256 ncols==32 for Volta
* adjust kernel selection logic
* 4 warps, 256 stride for all D
* no ncols == 64
* Multiple parallel blocks for batch size 1
* fix compile warnings
* fix excessive KQ_b loads
* fix cmake build
* fix KV cache padding, NaN from INFINITY (#6438)
* llama : flash_attn cparam + fix defrag
* server: support flash_attn param
* server: bench: enable flash_attn param
* CUDA: refactor host code, dyn. par. blocks
* fix flash_attn_vec_f16 race condition
* flush softmax exp below threshold to 0
* store temp KQ in registers
* Calculate KQ as FP32 if KQV has GGML_PREC_F32
* Add __hgt2_mask implementation for CUDA 11
* fix KQ FP32 precision fpr parallel_blocks > 1
* llama-bench : add -fa,--flash-attn arg
* metal : add BS=1 kernel for flash attention (#6508)
* metal : add BS=1 kernel for flash attention (wip)
* metal : support more than 1 warps
* metal : opts
* metal : opt
* metal : switch to parallel reduce
* metal : reduce registers
* metal : simplify
* metal : initial FA vec kernel
* metal : use F32 attention accumulators
* batched-bench : add fattn arg
* llama : simplify llama_build_kv_store
ggml-ci
* llama : adapt build_olmo to changes
* ggml : fix arm fp16 store on windows
* metal : clean-up
* metal : clean-up kernel code
* metal : minor
* tests : remove benchmarks
ggml-ci
* ggml : fix avx512 const correctness
ggml-ci
* ggml : fix soft_max with bias on CPU
ggml-ci
* common : print --flash-attn in help
* ggml : fix num dimensions in ggml_flash_attn_ext
* llama : force disable flash attention for incompatible models
* ggml : ggml_soft_max support F16/F32 mask/pos
ggml-ci
* cuda : uint -> uint32_t
* cuda : "constexpr dim3" -> "const dim3"
ggml-ci
* cuda : try to fix __hgt2_mask
ggml-ci
* ggml : add TODO's for F16/F32 mask/pos support in other backends
* llama : replace bool need_kq_pos with use_alibi
* llama : prep ALiBi support for BERT models
ggml-ci
* llama : fix n_batch requirements
ggml-ci
* cont
* server : add help for --flash-attn arg
* llama : disable FA for AMD
* tests : remove TMP_ATTN_BENCH
ggml-ci
* llama : support save/load state with FA enabled
ggml-ci
* ci : add CUDA save-load-state tests
ggml-ci
* llama : llama_kv_cache_clear zeroes data + fix save-load seq
ggml-ci
* llama : fix copy-paste errors, add TODO
* llama : disallow incompatible states
* llama : update llama_state_get_size after v_trans field
* metal : remove tmp log
* llama : add static reminder for llama_state_get_size
* metal : fix max nsg
ggml-ci
* ci : fix arg order
ggml-ci
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
HanishKVC
viric
Andrew Downing
Johannes Gäßler
Georgi Gerganov
Olivier Chafik
cpumaxx