As of #1486, whisper.cpp uses a unified KV cache with KQ masking.
As a result, depending on their location in the batch,
identical sequences in a batch can have slightly different outputs
due to floating point rounding errors during reduction.
See the discussion in #1941 for more details.
The beam search code used "has identical sum of log probabilities"
as a shorthand for "is an identical token sequence". However, per above,
identical tokens do not necessarily result in identical probabilities.
Instead, explicitly compare on sequences.
This is linear in cost when they are identical,
but the lengths are always small and the comparisons are cheap.
This increases diversity during beam search.
This improves output quality for some short samples I've been working
with, at no detectable performance cost.
I haven't checked against larger corpuses.
Fixes#1941
All else being otherwise equal, this encourages the beam candidate
selection to re-use the same decoder, which slightly
reduces the cache size.
I wouldn't expect it to make much of a performance difference,
but it helps when debug printing the cache and beam.
Added as part of understanding #1941.
* fix mul_mat fault in cpy_f32_f16
* rm unused function
* add wait() for memcpy
* restore ci/run.sh, rename struct defination, fix bug in ggml_sycl_op_mul_mat_sycl
* fix format issue
* llama : fix segfault from unknown model arch name (llama/5820)
* llama : fix segfault from unknown model arch name
* llama : make all LLM maps const
This also requires using `std::map::at` instead of its `operator[]`
which does not exist for const maps.
* llama : name LLM_ARCH_UNKNOWN to "(unknown)"
This avoids errors from `std::map::at` when
getting the general name of the model architecture.
Using "(unknown)" instead of an empty string as per suggestion
https://github.com/ggerganov/llama.cpp/pull/5820#issuecomment-1973735284
* llama : remove redundant inner const for LLM_TENSOR_NAMES
The extra const won't do anything here as const maps
return const references to values.
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
* llama : remove redundant nullptr check in llm_arch_from_string
Since LLM_ARCH_NAMES is a const map, no spurious elements
with a NULL name are inserted anymore, so this check is dead code.
---------
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
* llama : refactor internal quantization functions (llama/5830)
* scripts : add pod-llama.sh
* ggml : IQ3_S improvements (llama/5829)
* iq3_s: somewhat faster AVX2 dot product
On Ryzen a 7950X TG-128 increases to 16 t/s from 15.5 t/s using
16 threads. For 8 threads it is 13.85 t/s vs 11.75 t/s.
PP-512 increases to 28.5 t/s from 23.8 t/s.
* iq3_s: somewhat faster ARM_NEON dot product
Still dog slow - 10.7 t/s up from 9.9 t/s.
* iq3_s: another small ARM_NEON improvement
10.7 -> 11.0 t/s. Using vmulq_s8 is faster than the xor - sub trick
that works best on AVX2.
* iq3_s: minor improvement on Metal
49.4 t/s -> 50.3 t/s
* iq3_s: PPL improvement
E.g., for a context of 4096 LLaMA-v2-7B goes to 5.1340 from 5.1653.
* iq3_s: use new grid everywhere
* Fix ARM_NEON
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* convert-hf : make model class definitions self-contained (llama/5825)
* convert : automatically fall back to HfVocab if tokenizer.model doesn't exist (llama/5821)
* ggml : fix IQ3_S AVX implementation (llama/5834)
ggml-ci
* llama : add abort_callback to interrupt computation (llama/5409)
* using abort_callback from ggml to stop llama computation
* format fix
* a brief explaining comment
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* server: tests: passkey challenge / self-extend with context shift demo (llama/5832)
* server: tests: add models endpoint scenario
* server: /v1/models add some metadata
* server: tests: add debug field in context before scenario
* server: tests: download model from HF, add batch size
* server: tests: add passkey test
* server: tests: add group attention params
* server: do not truncate prompt tokens if self-extend through group attention is enabled
* server: logs: do not truncate log values
* server: tests - passkey - first good working value of nga
* server: tests: fix server timeout
* server: tests: fix passkey, add doc, fix regex content matching, fix timeout
* server: tests: fix regex content matching
* server: tests: schedule slow tests on master
* server: metrics: fix when no prompt processed
* server: tests: self-extend add llama-2-7B and Mixtral-8x7B-v0.1
* server: tests: increase timeout for completion
* server: tests: keep only the PHI-2 test
* server: tests: passkey add a negative test
* flake.lock: Update (llama/5842)
Flake lock file updates:
• Updated input 'flake-parts':
'github:hercules-ci/flake-parts/b253292d9c0a5ead9bc98c4e9a26c6312e27d69f' (2024-02-01)
→ 'github:hercules-ci/flake-parts/f7b3c975cf067e56e7cda6cb098ebe3fb4d74ca2' (2024-03-01)
• Updated input 'flake-parts/nixpkgs-lib':
'github:NixOS/nixpkgs/97b17f32362e475016f942bbdfda4a4a72a8a652?dir=lib' (2024-01-29)
→ 'github:NixOS/nixpkgs/1536926ef5621b09bba54035ae2bb6d806d72ac8?dir=lib' (2024-02-29)
• Updated input 'nixpkgs':
'github:NixOS/nixpkgs/cbc4211f0afffe6dfd2478a62615dd5175a13f9a' (2024-02-23)
→ 'github:NixOS/nixpkgs/1536926ef5621b09bba54035ae2bb6d806d72ac8' (2024-02-29)
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
* server : init http requests thread pool with --parallel if set (llama/5836)
* ci : schedule slow server tests only on Release or on demand (llama/5839)
* llama : fix llama_copy_state_data with fragmented KV cache (llama/5840)
The row size of the saved states was based on kv_self.head while
it should be based on llama_kv_cache_cell_max.
Existing session files should still work.
* llama : fix llama_kv_cache_cell_max inability to return 1
I've also changed its return type to uint32_t,
because this function is always used to set the value of uint32_t variables,
and because the index already has this type.
* llama : fix state size calculation
Some bytes in the state were unaccounted for in llama_get_state_size.
Since the logits reserve so much space, it did not cause problems.
* gguf-dump : support i-quants (llama/5841)
Co-authored-by: Black_Fox <radekliska@gmail.com>
* llama : allow for user specified embedding pooling type (llama/5849)
* allow for user specified pooling type
* llama : use enum types over int
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* readme : add API changes section
* cuda : fix data race in soft max (llama/5853)
* main : support special tokens as reverse/anti prompt (llama/5847)
* Support special tokens as reverse/anti prompt.
* Tokenize antiprompts only once.
* main : minor
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* common : use LLAMA_DEFAULT_SEED (llama/5855)
* add some new ops, fix some operators and add batch operations to certain operators. (ggml/747)
* cuda: fix group_norm
* cuda: add batch inference support for ggml_pad/ggml_upscale
* add ggml_arrange
* add ggml_timestep_embedding
* update ggml_arange/ggml_timestep_embedding tests
* cuda: fix im2col
* add ggml_arange/ggml_timestep_embbeding support for metal backend
* fix some bugs
* fix some bugs
* Update ggml.h
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update ggml-cuda.cu
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update ggml-metal.m
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update ggml-metal.m
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update ggml-metal.metal
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* modify according to the review comments
* ggml : fix compile warnings + code style
* ggml : normalize compute_forward calls + fix seg fault in debug
* minor
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: slaren <slarengh@gmail.com>
* sync : ggml
* add alias for chat template (llama/5858)
* speculative : implement stochastic speculative sampling (llama/5625)
* (WIP) Implement stochastic speculative decoding
* sample from residual distribution on draft accept failure
* fix#5657: force greedy sampling with probs when temp is 0
* remove p_accept parameter
* fix style
* remove unused variables
* add srand() in speculative.cpp
* replace use of rand() with mt19937 sampling
* fixes based on review (@JohannesGaessler)
* fix r random generation
* randomly select next sequence to verify + fix bug in memory freeing
* fix bug in active_seqs sync
* fix uniform int distribution initialization
* remove warnings from comparison between int and size_t
* check grammar in `llama_sample_probability_distribution_impl`
* remove malloc code by utilizing vectors
* add PR link to README
* cmake : handle cases where git index is not found in .git (llama/5844)
* Update CMakeLists.txt
* Update CMakeLists.txt
* ggml : introduce ggml_status (ggml/750)
* using enum as an exit code instead of macros
* update return type from enum to unsigned int
* indentation fix
* compound update
ggml_compute_exit_code -> ggml_status
changed ggml_status from a bit-field type to simple codes
ggml_status to string cast
* ggml_status to string cast
* GGML_CALL was removed
Co-authored-by: slaren <slarengh@gmail.com>
---------
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* sync : ggml
ggml-ci
* ggml : fix unknown status (llama/0)
* flake : fix
* llama : fix embeddings (llama/5796)
* llama : fix embeddings
ggml-ci
* llama : do not use KV cache for non-causal models
ggml-ci
* embeddings : fix llama_batch_init arg
* llama : add pooling switch
* llama : distinguish token vs sequence embeddings
ggml-ci
* llama : assert pooling tensor
* llama : simplify causal mask condition
ggml-ci
* llama : assert input batch with pooling enabled
* readme : update API changes list
* nix: static build (llama/5814)
* fix speculative decoding build on windows (llama/5874)
* rebase and rm tailing space
---------
Co-authored-by: LiangtaoJin <liang-tao.jin@intel.com>
Co-authored-by: compilade <113953597+compilade@users.noreply.github.com>
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Jared Van Bortel <jared@nomic.ai>
Co-authored-by: Michael Podvitskiy <podvitskiymichael@gmail.com>
Co-authored-by: Pierrick Hymbert <pierrick.hymbert@gmail.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Nindaleth <Nindaleth@users.noreply.github.com>
Co-authored-by: Black_Fox <radekliska@gmail.com>
Co-authored-by: Douglas Hanley <thesecretaryofwar@gmail.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: DAN™ <dranger003@gmail.com>
Co-authored-by: leejet <leejet714@gmail.com>
Co-authored-by: Minsoo Cheong <54794500+mscheong01@users.noreply.github.com>
Co-authored-by: Dane Madsen <dane_madsen@hotmail.com>
Co-authored-by: hutli <6594598+hutli@users.noreply.github.com>
Co-authored-by: Jeffrey Quesnelle <emozilla@nousresearch.com>
* using enum as an exit code instead of macros
* update return type from enum to unsigned int
* indentation fix
* compound update
ggml_compute_exit_code -> ggml_status
changed ggml_status from a bit-field type to simple codes
ggml_status to string cast
* ggml_status to string cast
* GGML_CALL was removed
Co-authored-by: slaren <slarengh@gmail.com>
---------
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* iq3_s: somewhat faster AVX2 dot product
On Ryzen a 7950X TG-128 increases to 16 t/s from 15.5 t/s using
16 threads. For 8 threads it is 13.85 t/s vs 11.75 t/s.
PP-512 increases to 28.5 t/s from 23.8 t/s.
* iq3_s: somewhat faster ARM_NEON dot product
Still dog slow - 10.7 t/s up from 9.9 t/s.
* iq3_s: another small ARM_NEON improvement
10.7 -> 11.0 t/s. Using vmulq_s8 is faster than the xor - sub trick
that works best on AVX2.
* iq3_s: minor improvement on Metal
49.4 t/s -> 50.3 t/s
* iq3_s: PPL improvement
E.g., for a context of 4096 LLaMA-v2-7B goes to 5.1340 from 5.1653.
* iq3_s: use new grid everywhere
* Fix ARM_NEON
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* suport multiple cards: split-mode - layer|row
* rm warning
* rebase with master, support tow new OPs, close feature for -sm=row, fix for unit test
* update news
* fix merge error
* update according to review comments
* Use batched mul_mat pathway
* rm extra line
* Explicitly state scaled data type
---------
Co-authored-by: Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
* WIP: make i-quants work for QK_K = 64
* iq2_xs: attempt to fix AVX dot product for QK_K = 64
Tests pass, but I get gibberish.
* QK_K = 64 tests pass on ARM_NEON and Metal
Sadly, that does not mean it actually works.
* Make CUDA compile with QK_K = 64
Tests don't pass, plus we get misaligned access
* Q2_K: fixed bug in imatrix quantization for QK_K = 64
* iq1_s: turn off SIMD implementation for QK_K = 64 (it does not work)
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Try IQ4_NL with blocks of 64 - does not look good
* iq4_xs: go to super-blocks of 256 and 6-bit scales for blocks of 32
* iq4_xs: CUDA works - 133.2 t/s
* iq4_xs: AVX2 dot product
* iq4_xs: ARM_NEON dot product
* iq4_nl: Metal implementation
As usual, Metal / Apple Silicon don't like my quants.
* iq3_xs: minor fix
* iq4_xs: shrink by using IQ3_S for attn_k and attn_q
* iq4_xs: revert using IQ3_S for attn_k and attn_v
PPL vs size is good, but CPU performance suffers: on M2 Max
TG-128 drops to 21.7 t/s from 28.8, and on a Ryzen-7950X
to 14.5 t/s from 15.8 t/s. On CUDA we have 135 t/s when
using IQ3_S vs 133 t/s with pure IQ4_XS.
* Fix CI
* iq4_xs: Added forgotten check for 256 divisibility
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Adding IQ2_S and IQ2_M as a single cumulative commit
* Update examples/quantize/quantize.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* [ggml-quants] Provide ggml_vqtbl1q_u8 for 64bit compatibility
vqtbl1q_u8 is not part of arm v7 neon library
* [android-example] Remove abi filter after arm v7a fix
* [github-workflows] Do not skip Android armeabi-v7a build
* add magika inference example
* ggml : fix unaligned accesses in custom ops
* ggml : fix FP32 GELU for values that exceed the FP16 range
* use ggml_pool_1d
* add README
* Update README.md
* pad inputs if the files are too small
* cleanup
ggml-ci
