Commit Graph

281 Commits

Author SHA1 Message Date
Tim Dettmers
097b1cc5da Fixed bug caused by undefined default type of absmax. #553 2023-07-13 21:23:33 -07:00
Tim Dettmers
90b0ac57b0 Fixed missing bias in bnb.matmul_4bit for inference; more tests. 2023-07-11 17:13:33 -07:00
Tim Dettmers
ba51d95d43 Added more extensive gemv tests; blocksize guard for gemv. 2023-07-11 05:55:49 -07:00
Tim Dettmers
5f492d437e Merge remote-tracking branch 'origin/inference' 2023-07-10 06:24:24 -07:00
Tim Dettmers
196d6f5dc1
Merge pull request #469 from shadeMe/linear-layer-device
Add `device` parameter to `Linear` subclasses and `Embedding`
2023-07-10 06:17:13 -07:00
Tim Dettmers
5fab673442 Added fp32 compute type for gemv_4bit. 2023-07-09 21:06:01 -07:00
Tim Dettmers
cef519c89e Added test for Param4bit.to() and fixed double quant behavior. 2023-07-09 17:16:50 -07:00
Tim Dettmers
6a905be5ce Fixed a bug where gemv_4bit would return a wrongly sized tensor. 2023-07-09 15:34:02 -07:00
Tim Dettmers
0f0390acb2 Added double quantization support and tests. 2023-07-09 15:32:03 -07:00
Tim Dettmers
94168d79d7 Added FP4 fast inference support. 2023-07-09 14:46:19 -07:00
Tim Dettmers
4b88d69de7 Added abitrary data types; fixed a bug for small matrices. 2023-07-09 12:04:09 -07:00
Tim Dettmers
02fd80cb81 Added bfloat16 quantizations and tests. 2023-07-04 19:58:31 -07:00
Tim Dettmers
f89ff93e26 Initial 4-bit naive batch size 1, 81 vs 185. 2023-07-03 18:45:38 -07:00
Max Ryabinin
b599fdb197 Only rearrange weight if it exists 2023-06-14 19:27:13 +02:00
Max Ryabinin
c1f3f56d2c Rearrange the weights directly in state dict before loading 2023-06-09 21:58:39 +02:00
Max Ryabinin
f734076e94 Improve memory efficiency of 8-bit serialization 2023-06-09 21:39:57 +02:00
Max Ryabinin
4fb37d45c1 Extract get_tile_inds to a separate function 2023-06-09 21:39:37 +02:00
shadeMe
db49ad43ab
Add device parameter to Embedding 2023-06-01 17:43:49 +02:00
shadeMe
9cac5dd1b6
Add device parameter to Linear subclasses 2023-06-01 17:43:30 +02:00
Tim Dettmers
1b8772a8f3 Added PagedLion and bf16 Lion. 2023-05-23 19:37:38 -07:00
Tim Dettmers
2bce175d15 Fixed Makefile. 2023-05-23 18:42:19 -07:00
Tim Dettmers
4bd1151829 Fixed gradient accumulation test. 2023-05-07 15:06:17 -07:00
Tim Dettmers
675baa79d2 Merge remote-tracking branch 'origin/main' into merge 2023-05-07 13:34:03 -07:00
Tim Dettmers
f64cfe65aa Fixed prefetch bug for non-paged tensors; added benchmark. 2023-05-06 21:49:16 -07:00
Tim Dettmers
41a9c70814 Changed prefetching. 2023-05-06 18:59:59 -07:00
Tim Dettmers
44d68ff29c Added paged optimizers. 2023-05-06 14:59:29 -07:00
Tim Dettmers
ec38ba95b0 Added paging. 2023-05-06 11:14:06 -07:00
Tim Dettmers
264a948539 4-bit draft; 128 vector load 240. 2023-05-02 16:15:38 -07:00
Tim Dettmers
f9bfea8f23 Baseline for debugging. 2023-05-02 07:24:12 -07:00
Tim Dettmers
21723f796a 4-bit draft. 2023-04-29 21:52:47 -07:00
Tim Dettmers
f6df4aef6a Added fp16 and thread/item template. 2023-04-28 18:26:52 -07:00
Tim Dettmers
3aef78342a Added template refactor. 2023-04-28 17:34:08 -07:00
Tim Dettmers
c1bfb210c5 First baseline kernel. 2023-04-28 17:19:02 -07:00
Tim Dettmers
9cab14a3ff Adedd pipeline draft. 2023-04-27 15:12:49 -07:00
Tim Dettmers
d1c4c20568 Added non-cutlass template. 2023-04-27 15:11:26 -07:00
Tim Dettmers
0afc8e9e2f Best attempt at cutlass3. 2023-04-26 17:12:34 -07:00
Tim Dettmers
84964db937 CUTLASS compiles. 2023-04-25 17:15:51 -07:00
Tim Dettmers
0f9d30207f Added nested quantization for blockwise quantization. 2023-04-19 11:48:47 -07:00
Tim Dettmers
7dc198feb7 Added 32-bit optimizer for bfloat16 gradients. 2023-04-17 18:01:49 -07:00
Tim Dettmers
9e7cdc9ea9 Added last SwitchBack refactors. All tests green. 2023-04-12 13:41:30 -07:00
Tim Dettmers
008dfff9b4 Added triton utils. 2023-04-12 12:57:46 -07:00
Tim Dettmers
b8ea2b416d Fixed bias conversion in Linear4bit 2023-04-12 12:28:35 -07:00
Tim Dettmers
5b612bc6df Added is_available_triton guard to Triton SwitchBackLinear. 2023-04-12 12:16:55 -07:00
Tim Dettmers
c3d87e4435 Added is_available_triton guard. 2023-04-12 12:10:34 -07:00
Tim Dettmers
7140c01405 Merge branch 'main' into fp8_merge 2023-04-12 11:44:39 -07:00
Tim Dettmers
dd562c24f1 Refactored simulated fp8 modules into research.nn. 2023-04-12 11:24:44 -07:00
Tim Dettmers
e67bfccbcd Added missing triton and fp8 files. 2023-04-12 10:06:18 -07:00
Tim Dettmers
ec1ea63711 Refactored triton into its own folder. Refactored fp8 matmuls. 2023-04-12 09:39:39 -07:00
Tim Dettmers
7c651012fc Added better error message for debugging on CUDA not detected failures. 2023-04-12 07:56:52 -07:00
Tim Dettmers
659a7dfc71 Fixing #300. 2023-04-11 16:14:29 -07:00