tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

Commit Graph

  • 6d3f4b9ab5 keep some softmax layers rala Wing Lian 2025-01-15 15:08:32 -05:00
  • 12aade921a various fixes Wing Lian 2024-12-28 15:03:22 -05:00
  • 198f01f902 remove bias term in phi and add custom modeling code Wing Lian 2024-12-26 20:27:18 -05:00
  • 2e6265090f remove lr_groups from other branch Wing Lian 2024-12-25 21:45:27 -05:00
  • 1c5b78621c fix forward sig Wing Lian 2024-12-21 00:27:59 -05:00
  • 32258c247e no batching for kd chat templates Wing Lian 2025-01-15 08:22:29 -05:00
  • 04efcb102f don't shift student logits for kd Wing Lian 2025-01-15 01:07:48 -05:00
  • 483defb9ae try tests for kd on l40s Wing Lian 2025-01-14 23:56:00 -05:00
  • 35a84f2cb8 more fixes kd-trainer-v2 Wing Lian 2025-01-14 21:37:10 -05:00
  • 510cf45317 improve logprob masking and shift in trainer Wing Lian 2025-01-14 20:10:05 -05:00
  • 7232cbdeab chore: lint kd-trainer-pre Wing Lian 2025-01-13 14:05:56 -05:00
  • e8fceb7091 chore: lint Wing Lian 2025-01-10 02:18:55 -05:00
  • a5e0671738 make sure to use tensorboard to capture loss for checks Wing Lian 2025-01-08 22:30:13 -05:00
  • b9847553af fix adapter model check Wing Lian 2025-01-08 20:15:42 -05:00
  • 513ec9e03b make sure to use the correct tokenizer Wing Lian 2025-01-08 17:54:48 -05:00
  • 530347856d make sure to set tokenizer from l3 70b and save safetensors Wing Lian 2025-01-08 14:05:49 -05:00
  • 261e4fb619 lower lr Wing Lian 2025-01-08 13:45:12 -05:00
  • 158071e95f set lora_dropout explicitly Wing Lian 2025-01-08 12:14:16 -05:00
  • 432f65f5e6 make the kd e2e fit in vram for ci and add lora version Wing Lian 2025-01-08 11:07:29 -05:00
  • 1d039f5486 rename test files so it gets picked up Wing Lian 2025-01-08 09:21:17 -05:00
  • b9a42b396f linting Wing Lian 2025-01-08 08:31:28 -05:00
  • ff2fb0fc1b add kd trainer e2e test Wing Lian 2025-01-08 08:19:10 -05:00
  • 317f290186 reward model doesn't work well with batched Wing Lian 2025-01-07 18:19:42 -05:00
  • ab690f3f01 improve check for batched Wing Lian 2025-01-07 16:57:47 -05:00
  • 47932f21c4 fix reward trainer calls for tokenization Wing Lian 2025-01-07 15:41:40 -05:00
  • 808328e041 reward can use same batch check Wing Lian 2025-01-07 15:11:07 -05:00
  • 6784822cfb tweak check for batched prompt data Wing Lian 2025-01-07 14:54:32 -05:00
  • 684b38291f ensure that batch vs single is done properly Wing Lian 2025-01-07 14:03:49 -05:00
  • 01896b1bde improve iterable support Wing Lian 2025-01-02 13:50:35 -05:00
  • e659c01646 support streaming for processing sft datasts? Wing Lian 2025-01-01 09:11:14 -05:00
  • 204d6c43b4 make loss torch script compat Wing Lian 2024-12-30 21:34:46 -05:00
  • d3c2b7ce9d kd sample packing Wing Lian 2024-12-30 20:10:47 -05:00
  • 93dfff92f1 be a bit pickier about loading dynamic prompt strategies Wing Lian 2024-12-30 16:52:41 -05:00
  • 6e409d2d88 more info on preprocess for kd and fix import Wing Lian 2024-12-30 15:58:02 -05:00
  • d5bc214300 remove duplicate code Wing Lian 2024-12-30 14:16:33 -05:00
  • 92c6c1087e add copyrights Wing Lian 2024-12-30 14:12:02 -05:00
  • feed96f95e increase logging around loading plugins Wing Lian 2024-12-30 13:33:56 -05:00
  • cba6165ae1 make plugin setup concise Wing Lian 2024-12-30 13:25:25 -05:00
  • cdfcd69afa remove moved class from import Wing Lian 2024-12-30 13:17:11 -05:00
  • 885653d52e move more things to kd plugin Wing Lian 2024-12-30 13:15:28 -05:00
  • 27faacbf5a refactor kd chat template loader Wing Lian 2024-12-30 12:57:11 -05:00
  • c51b0337c1 support for custom trainer classes from plugins Wing Lian 2024-12-30 12:20:45 -05:00
  • fa055f9f69 handle token/logprob shifting Wing Lian 2024-12-30 11:21:19 -05:00
  • f60c623af0 remove references to triton kd for now Wing Lian 2024-12-30 10:40:05 -05:00
  • 746891eb5c add license block Wing Lian 2024-12-29 16:18:05 -05:00
  • f09b5da60b refactor so we can easily add new loss functions Wing Lian 2024-12-29 16:15:47 -05:00
  • 689e1c10ba chore: lint Wing Lian 2024-12-28 16:02:06 -05:00
  • a5c085e003 var naming and add todo Wing Lian 2024-12-25 21:41:06 -05:00
  • 63146300b7 fix kd loss so it's causal (fixes repeating tokens) Wing Lian 2024-12-25 18:59:30 -05:00
  • ca5e397fc5 use kd_alpha in the correct loss method Wing Lian 2024-12-24 19:54:32 -05:00
  • 3416302b0d hash for temperature too Wing Lian 2024-12-24 15:48:35 -05:00
  • 7366efc4ca better rescaling for temperatures Wing Lian 2024-12-24 09:26:27 -05:00
  • d8d817eaed don't use triton for now Wing Lian 2024-12-21 16:47:11 -05:00
  • c0757e8a20 fix kwarg Wing Lian 2024-12-21 14:32:11 -05:00
  • e565694914 v3 Wing Lian 2024-12-21 14:17:30 -05:00
  • 081928e55b no torch.tensor Wing Lian 2024-12-21 14:00:01 -05:00
  • dc90c93894 no log etc Wing Lian 2024-12-21 13:54:21 -05:00
  • 18a46c338a no torch.exp inside triton kernel Wing Lian 2024-12-21 13:52:31 -05:00
  • 119d586cf4 v2 trial Wing Lian 2024-12-21 13:43:48 -05:00
  • c73acd7de0 no where support Wing Lian 2024-12-21 13:21:54 -05:00
  • 0b59a242d4 triton wip Wing Lian 2024-12-21 13:18:23 -05:00
  • ed490517da chore: lint Wing Lian 2024-12-19 02:04:25 -05:00
  • 00ce77e7ef make sure to multiply against the correct loss Wing Lian 2024-12-19 01:42:57 -05:00
  • ae545e0165 cross entropy loss coefficient during KD Wing Lian 2024-12-19 01:42:21 -05:00
  • b592c05b93 flipped the slice Wing Lian 2024-12-19 01:21:48 -05:00
  • 7fe0ad088b make it work Wing Lian 2024-12-19 00:28:02 -05:00
  • ddcf5c68b3 handle padding/collation for KD datasets Wing Lian 2024-12-18 18:07:27 -05:00
  • e633a12dbe make batch smaller Wing Lian 2024-12-18 16:23:50 -05:00
  • d584354ee4 filter bad rows Wing Lian 2024-12-18 15:47:18 -05:00
  • 303cfa71aa KD dataset loading and KD with logprobs Wing Lian 2024-12-18 15:16:45 -05:00
  • 88b3198894 refactor trainer to prevent circular dependencies later Wing Lian 2024-12-16 14:16:36 -05:00
  • 8bef307af4 Built site for gh-pages Quarto GHA Workflow Runner 2025-01-15 03:10:22 +00:00
  • 5c570ed0be Built site for gh-pages Quarto GHA Workflow Runner 2025-01-15 03:09:51 +00:00
  • 8606093921 fix for indexing error from token/embeddings mismatch (#2257) jwongTensora 2025-01-15 03:09:29 +00:00
  • cba5a457d9 fix: use text_column even when not packing for pretraining (#2254) NanoCode012 2025-01-15 10:08:56 +07:00
  • d6bb2bb498 Built site for gh-pages Quarto GHA Workflow Runner 2025-01-15 03:08:54 +00:00
  • 19cd83d408 rename references to dpo dataset prep to pref data (#2258) Wing Lian 2025-01-14 22:07:55 -05:00
  • 28694219a5 inline comment change Dan Saunders 2025-01-14 16:59:43 +00:00
  • dbcd11e533 revert seq len in multipack sampler Sunny 2025-01-14 11:45:35 -05:00
  • c06a6be915 flex_attn sample packing WIP Sunny 2025-01-14 00:22:05 -05:00
  • fd8ad6fcbf fixing negative component mixing Dan Saunders 2025-01-13 19:21:55 +00:00
  • 4a0ab11fcf chore: lint kd-trainer-2 Wing Lian 2025-01-13 14:05:56 -05:00
  • 73b6b0a580 chore: lint Wing Lian 2025-01-10 02:18:55 -05:00
  • 9db5072407 make sure to use tensorboard to capture loss for checks Wing Lian 2025-01-08 22:30:13 -05:00
  • 42d3e36a6f fix adapter model check Wing Lian 2025-01-08 20:15:42 -05:00
  • b12d93bedf make sure to use the correct tokenizer Wing Lian 2025-01-08 17:54:48 -05:00
  • 08ec9c0e5b make sure to set tokenizer from l3 70b and save safetensors Wing Lian 2025-01-08 14:05:49 -05:00
  • 9abac55f92 lower lr Wing Lian 2025-01-08 13:45:12 -05:00
  • 800e7fa41e set lora_dropout explicitly Wing Lian 2025-01-08 12:14:16 -05:00
  • 5a1c1b82d4 make the kd e2e fit in vram for ci and add lora version Wing Lian 2025-01-08 11:07:29 -05:00
  • efb3f70d38 rename test files so it gets picked up Wing Lian 2025-01-08 09:21:17 -05:00
  • 58d9896777 linting Wing Lian 2025-01-08 08:31:28 -05:00
  • f7963083b8 add kd trainer e2e test Wing Lian 2025-01-08 08:19:10 -05:00
  • f0b6581f8c reward model doesn't work well with batched Wing Lian 2025-01-07 18:19:42 -05:00
  • 27bb21c459 improve check for batched Wing Lian 2025-01-07 16:57:47 -05:00
  • 74d98ca6d8 fix reward trainer calls for tokenization Wing Lian 2025-01-07 15:41:40 -05:00
  • ec4dfb02c8 reward can use same batch check Wing Lian 2025-01-07 15:11:07 -05:00
  • 28ef5e8d5a tweak check for batched prompt data Wing Lian 2025-01-07 14:54:32 -05:00
  • 5ed2823855 ensure that batch vs single is done properly Wing Lian 2025-01-07 14:03:49 -05:00
  • fb0775d264 improve iterable support Wing Lian 2025-01-02 13:50:35 -05:00