IQ3_S improvements #5829
Merged
IQ3_S improvements #5829
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
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.
Still dog slow - 10.7 t/s up from 9.9 t/s.
10.7 -> 11.0 t/s. Using vmulq_s8 is faster than the xor - sub trick that works best on AVX2.
49.4 t/s -> 50.3 t/s
E.g., for a context of 4096 LLaMA-v2-7B goes to 5.1340 from 5.1653.
How did the actually slow archs do with the other approach? I (EDIT: may) be misunderstanding the description, but it sounds to me as AVX2 PP and ARM PP/TG performance is bad with the lookup approach, but you decided to not look into the multipliers any further because the already good CUDA PP/TG and AVX2 TG performance with lookup didn't get any better with multipliers...? |
because the quantization error about 1-1.5% higher than lookup |
|
I notice you change the grid in |
Sorry, forgot to mention. Yes, previous quantization is still usable. You can re0quantize to take advantage of the slight PPL improvement. |
As in one percentage point higher, not 1.01x higher? I see, that makes sense, thanks for pointing that out. |
|
If the quantization error is 3% with lookup, it will be in the 4 to 4.5% range with the best multiplier I was able to find. Example: PPL = 5 for |
ggerganov
approved these changes
Mar 2, 2024
NeoZhangJianyu
pushed a commit
to NeoZhangJianyu/llama.cpp
that referenced
this pull request
Mar 5, 2024
* 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>
abhilash1910
pushed a commit
that referenced
this pull request
Mar 5, 2024
* 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 (#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 #5820 (comment) * 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 (#5830) * scripts : add pod-llama.sh * ggml : IQ3_S improvements (#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 (#5825) * convert : automatically fall back to HfVocab if tokenizer.model doesn't exist (#5821) * ggml : fix IQ3_S AVX implementation (#5834) ggml-ci * llama : add abort_callback to interrupt computation (#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 (#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 (#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 (#5836) * ci : schedule slow server tests only on Release or on demand (#5839) * llama : fix llama_copy_state_data with fragmented KV cache (#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 (#5841) Co-authored-by: Black_Fox <radekliska@gmail.com> * llama : allow for user specified embedding pooling type (#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 (#5853) * main : support special tokens as reverse/anti prompt (#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 (#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 (#5858) * speculative : implement stochastic speculative sampling (#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 (#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 (#0) * flake : fix * llama : fix embeddings (#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 (#5814) * fix speculative decoding build on windows (#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>
hazelnutcloud
pushed a commit
to hazelnutcloud/llama.cpp
that referenced
this pull request
Mar 10, 2024
* 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>
hazelnutcloud
pushed a commit
to hazelnutcloud/llama.cpp
that referenced
this pull request
Mar 10, 2024
* 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 (ggerganov#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 ggerganov#5820 (comment) * 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 (ggerganov#5830) * scripts : add pod-llama.sh * ggml : IQ3_S improvements (ggerganov#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 (ggerganov#5825) * convert : automatically fall back to HfVocab if tokenizer.model doesn't exist (ggerganov#5821) * ggml : fix IQ3_S AVX implementation (ggerganov#5834) ggml-ci * llama : add abort_callback to interrupt computation (ggerganov#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 (ggerganov#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 (ggerganov#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 (ggerganov#5836) * ci : schedule slow server tests only on Release or on demand (ggerganov#5839) * llama : fix llama_copy_state_data with fragmented KV cache (ggerganov#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 (ggerganov#5841) Co-authored-by: Black_Fox <radekliska@gmail.com> * llama : allow for user specified embedding pooling type (ggerganov#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 (ggerganov#5853) * main : support special tokens as reverse/anti prompt (ggerganov#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 (ggerganov#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 (ggerganov#5858) * speculative : implement stochastic speculative sampling (ggerganov#5625) * (WIP) Implement stochastic speculative decoding * sample from residual distribution on draft accept failure * fix ggerganov#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 (ggerganov#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 (#0) * flake : fix * llama : fix embeddings (ggerganov#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 (ggerganov#5814) * fix speculative decoding build on windows (ggerganov#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>
|
@ikawrakow I was testing some iq3_m quant today I found the PPL was not as good as before. After digging into your recent PRs, I think this grid change to iq3_s causes a PPL increase for my model as iq3_m uses a log of iq3_s. (i use qwen1.5-7b-chat for testing) So seems this grid value is pretty important to the final perforhamce for IQ quantization method, and I am seeing opposite results compared to yours. Would you mind elobrating how you choose gird values? |
|
@JianbangZ Sorry to hear that this PR is increasing PPL for your model. It did improve PPL on 8 of the 9 models tested, and it was about the same for one. Concerning grid points: during quantization the quants can take one in 8 values in |
Thank you. I tried. I revert the serach range to 15 and PPL improved slightly. Here is my data. I think grid points are the main contributor to the accuracy change. |
Since grid changed, and dequant seems use the new grid value, how was the the old grid quant model compatiable? |
jordankanter
pushed a commit
to jordankanter/llama.cpp
that referenced
this pull request
Mar 13, 2024
* 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>
jordankanter
pushed a commit
to jordankanter/llama.cpp
that referenced
this pull request
Mar 13, 2024
* 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 (ggerganov#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 ggerganov#5820 (comment) * 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 (ggerganov#5830) * scripts : add pod-llama.sh * ggml : IQ3_S improvements (ggerganov#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 (ggerganov#5825) * convert : automatically fall back to HfVocab if tokenizer.model doesn't exist (ggerganov#5821) * ggml : fix IQ3_S AVX implementation (ggerganov#5834) ggml-ci * llama : add abort_callback to interrupt computation (ggerganov#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 (ggerganov#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 (ggerganov#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 (ggerganov#5836) * ci : schedule slow server tests only on Release or on demand (ggerganov#5839) * llama : fix llama_copy_state_data with fragmented KV cache (ggerganov#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 (ggerganov#5841) Co-authored-by: Black_Fox <radekliska@gmail.com> * llama : allow for user specified embedding pooling type (ggerganov#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 (ggerganov#5853) * main : support special tokens as reverse/anti prompt (ggerganov#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 (ggerganov#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 (ggerganov#5858) * speculative : implement stochastic speculative sampling (ggerganov#5625) * (WIP) Implement stochastic speculative decoding * sample from residual distribution on draft accept failure * fix ggerganov#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 (ggerganov#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 (#0) * flake : fix * llama : fix embeddings (ggerganov#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 (ggerganov#5814) * fix speculative decoding build on windows (ggerganov#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>
hodlen
pushed a commit
to hodlen/llama.cpp
that referenced
this pull request
Apr 1, 2024
* 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>
hodlen
pushed a commit
to hodlen/llama.cpp
that referenced
this pull request
Apr 1, 2024
* 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 (ggerganov#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 ggerganov#5820 (comment) * 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 (ggerganov#5830) * scripts : add pod-llama.sh * ggml : IQ3_S improvements (ggerganov#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 (ggerganov#5825) * convert : automatically fall back to HfVocab if tokenizer.model doesn't exist (ggerganov#5821) * ggml : fix IQ3_S AVX implementation (ggerganov#5834) ggml-ci * llama : add abort_callback to interrupt computation (ggerganov#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 (ggerganov#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 (ggerganov#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 (ggerganov#5836) * ci : schedule slow server tests only on Release or on demand (ggerganov#5839) * llama : fix llama_copy_state_data with fragmented KV cache (ggerganov#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 (ggerganov#5841) Co-authored-by: Black_Fox <radekliska@gmail.com> * llama : allow for user specified embedding pooling type (ggerganov#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 (ggerganov#5853) * main : support special tokens as reverse/anti prompt (ggerganov#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 (ggerganov#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 (ggerganov#5858) * speculative : implement stochastic speculative sampling (ggerganov#5625) * (WIP) Implement stochastic speculative decoding * sample from residual distribution on draft accept failure * fix ggerganov#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 (ggerganov#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 (#0) * flake : fix * llama : fix embeddings (ggerganov#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 (ggerganov#5814) * fix speculative decoding build on windows (ggerganov#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>
|
@ikawrakow Could you make a different GGML_type of the old IQ3_S, so we can quantize with it? At PPL 512 : Miqu IQ2_M : approx 4.62 (stopped around 500 chunks) Or maybe this PR #5834 is causing problems (I don't know if I tested this new IQ3_S with it without the PR back in the days) ? |
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.
This PR makes some improvements to the
IQ3_Squants. Nothing earth shuttering, but I think still worth having. We getThis table shows a comparison between the quantization error on master and this PR. The quantization error is defined as
PPL(IQ3_S)/PPL(fp16) - 1. The context is 4096 for LLaMA-v2 and 2048 for LLaMA-v1This table shows a comparison in inference speed on the platforms where I can test:
ARM performance is still pathetic. Metal performance is acceptable (0.95X for TG and 1.02X for PP compared to
Q3_K_S). AVX2 Performance is acceptable for TG (0.88X compared toQ3_K_S) but bad for PP (0.38X ofQ3_K_S. I guess, for PP the loading of the lattice points from the lookup table becomes I/O bound and severely limits performance). CUDA performance is excellent (1.04X/1.015X for TG/PP compared toQ3_K_S, 1.18X compared toQ4_0TG).The main reason this PR exists is @PeterReid's suggestion to replace the lookup table with a multiplier that generates the codebook. While working on that I noticed that there is still room for improvement for the lookup based
IQ3_S, and this PR is the result. I did find a few multipliers that work quite OK (quantization error about 1-1.5% higher than lookup), but there was no performance improvement on CUDA and TG on AVX2, so I decided to not further pursue this idea.