Commit Graph

92 Commits

Author SHA1 Message Date
Ed Addario c3b6685599
Merge branch 'master' into imatrix 2025-12-16 13:57:27 +01:00
Georgi Gerganov 254098a279
common : refactor common_sampler + grammar logic changes (#17937)
* common : refactor common_sampler + grammar logic changes

* tests : increase max_tokens to get needed response

* batched : fix uninitialized samplers
2025-12-14 10:11:13 +02:00
Ed Addario 4a0511f3a0
Remove storing tensor statistics 2025-11-23 22:18:12 +00:00
Ed Addario 8d97eee557
Improve layer 0 stats 2025-11-17 17:52:15 +00:00
Ed Addario bf9823afa7
Minor refactoring 2025-11-17 14:51:12 +00:00
Ed Addario cdc7caea97
Remove unreachable logic 2025-11-17 14:46:45 +00:00
Ed Addario 658c6a8303
Enforce tensor structure when aggregating multiple imatrix files 2025-11-17 14:46:21 +00:00
Ed Addario a2b86d7fd9
Minor refactoring 2025-11-17 14:14:05 +00:00
Ed Addario 1f3db496cc
Calculate layer_sum only for legacy 2025-11-17 13:36:28 +00:00
Ed Addario 76566b83de
Enforce same-size between compared tensors 2025-11-17 13:28:35 +00:00
Ed Addario fb2b09a43c
Skip experts with zero count (unused) 2025-11-17 13:06:37 +00:00
Ed Addario 63cbcc6dfc
Refactor legacy determination 2025-11-17 13:05:34 +00:00
Ed Addario ae1cbc707b
Warn if problem with previous layer 2025-11-17 13:04:16 +00:00
Ed Addario 5384a11b94
Initialise layer and tensor variables 2025-11-17 13:00:47 +00:00
Ed Addario 559ae9ab89
Refactor legacy imatrix handling 2025-11-17 10:19:34 +00:00
Ed Addario b2b7175e19
Fix bug when vectors are zero 2025-11-06 15:12:09 +00:00
Ed Addario ce046dcee8
Save statistics to imatrix 2025-10-30 22:43:46 +00:00
Ed Addario 7d8819f57a
Improve compute_layer_statistics() processing of mismatched tensor sizes 2025-10-29 18:36:01 +00:00
Ed Addario 006e7ef991
Improve compute_vector_statistics() processing of mismatched tensor sizes 2025-10-29 18:35:39 +00:00
Ed Addario 2a6f5d7e60
Refactor variable names 2025-10-29 18:32:47 +00:00
Ed Addario 6ff0a79e54
Minor stats report cosmetic changes 2025-10-29 07:59:40 +00:00
Ed Addario 86fabce58d
Clamp values 2025-10-28 23:10:44 +00:00
Ed Addario ab015065b8
Minor refactoring 2025-10-28 23:10:12 +00:00
Ed Addario 92a42bac3d
Type refactoring 2025-10-28 23:06:29 +00:00
Ed Addario b5068df804
Minor refactoring 2025-10-28 23:03:52 +00:00
Ed Addario 0b0381c94c
Merge Cosine Similarity and L2 Norm computation into single loop 2025-10-28 21:41:31 +00:00
Ed Addario dc4a04b5c5
Adjust size calculation and change fallback value to 0.0f 2025-10-28 21:35:35 +00:00
Ed Addario 683ef8dfb7
Fill zeros for experts with zero counts to preserve shape 2025-10-28 18:35:17 +00:00
Ed Addario 637e674da6
Avoid division by zero on zero-count matrices 2025-10-28 18:33:37 +00:00
Ed Addario c9a0874f35
Clamp CosSim to [-1, 1] to avoid float drift 2025-10-28 18:29:59 +00:00
Ed Addario af3b6aca22
Fix legacy_mode getting overwritten on each tensor bug 2025-10-28 18:27:19 +00:00
Ed Addario 8f1aa7885e
Remove activation_statistics() option 2025-08-31 14:03:19 +01:00
Ed Addario 6371902f98
Add --output-format to usage 2025-08-26 21:53:54 +01:00
Ed Addario 630750fdef
Validate number of elements if in_sum is present 2025-08-17 09:42:18 +01:00
Ed Addario 1f72bc157f
Avoid using if statements with initialiser 2025-08-17 08:35:17 +01:00
Ed Addario f6934b9417
Merge branch 'imatrix' of https://github.com/EAddario/llama.cpp into imatrix 2025-08-17 08:20:18 +01:00
Ed Addario 44ea7ddeac
Change statement order 2025-08-17 08:20:03 +01:00
Ed Addario 2e803234f4
Use { and } around conditionally-executed single line statements 2025-08-17 08:19:02 +01:00
Ed Addario a96013f720
Define one variable per line and refactor names 2025-08-17 08:16:41 +01:00
Ed Addario 12607d3203
Use { and } around single line for statement 2025-08-17 08:10:54 +01:00
Ed Addario d19e6c9afa
Use { and } around the conditionally-executed statement
Co-authored-by: compilade <git@compilade.net>
2025-08-17 08:08:26 +01:00
Ed Addario 97d839c441
Using one line per variable definition
Co-authored-by: compilade <git@compilade.net>
2025-08-17 08:06:15 +01:00
Ed Addario 4a487ea7e4
Use { and } around the conditionally-executed statement
Co-authored-by: compilade <git@compilade.net>
2025-08-17 07:26:16 +01:00
Ed Addario e3149a2168
Use the corresponding size
Co-authored-by: compilade <git@compilade.net>
2025-08-17 07:24:27 +01:00
Ed Addario d4b0d89115
Fix return type bug 2025-08-16 11:00:43 +01:00
Ed Addario 030ec53d7a
Remove unnecessary include 2025-08-16 10:46:09 +01:00
Ed Addario 42bfe3b2a3
Update stats output sort based on imatrix type 2025-08-15 21:12:56 +01:00
Ed Addario dcac206f8e
Add --activation-statistics logic to avoid doubling the imatrix size by default 2025-08-09 14:49:25 +01:00
Ed Addario c5ecdaa1a1
Add Euclidean–Cosine Score (ECS) 2025-08-07 19:04:49 +01:00
Ed Addario 5bb2def02d
Add --activation-statistics parameter 2025-08-07 17:41:21 +01:00