drbh d888bc2828 feat: support force downcast after FastRMSNorm multiply for Gemma (#1658) ... This PR adds `force_downcast_after` to `FastRMSNorm.forward` which is used in the Gemma model. References https://github.com/huggingface/transformers/pull/29402 and https://github.com/huggingface/transformers/pull/29729 Setting `force_downcast_after=True` will perform the `hidden_states * weight` multiplication in f32 and then downcast to half. This differs slightly from the current implementation which first casts the `hidden_states` to a half and then multiples.		2024-04-25 12:32:42 +03:00
..
custom_modeling	feat: support force downcast after FastRMSNorm multiply for Gemma (#1658)	2024-04-25 12:32:42 +03:00
__init__.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
bloom.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
cache_manager.py	feat: add mistral model (#1071)	2023-09-28 09:55:47 +02:00
causal_lm.py	fix: Handle concurrent grammar requests (#1610)	2024-04-25 10:11:40 +03:00
flash_causal_lm.py	fix: Handle concurrent grammar requests (#1610)	2024-04-25 10:11:40 +03:00
flash_gemma.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
flash_llama.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
flash_mistral.py	feat: Qwen2 (#1608)	2024-04-25 09:21:22 +03:00
flash_mixtral.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
flash_neox.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
flash_phi.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
flash_qwen2.py	feat: Qwen2 (#1608)	2024-04-25 09:21:22 +03:00
flash_rw.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
flash_santacoder.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
flash_starcoder2.py	feat: Qwen2 (#1608)	2024-04-25 09:21:22 +03:00
galactica.py	feat: Qwen2 (#1608)	2024-04-25 09:21:22 +03:00
globals.py	Small cleanup. (#1560)	2024-04-24 14:42:35 +03:00
gpt_neox.py	feat: Qwen2 (#1608)	2024-04-25 09:21:22 +03:00
idefics_causal_lm.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
idefics.py	Fix idefics default. (#1614)	2024-04-25 10:12:13 +03:00
mamba.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
model.py	feat: add more latency metrics in forward (#1346)	2024-04-19 13:41:34 +03:00
mpt.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
opt.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
phi.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
rw.py	feat: Qwen2 (#1608)	2024-04-25 09:21:22 +03:00
santacoder.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
seq2seq_lm.py	feat: Qwen2 (#1608)	2024-04-25 09:21:22 +03:00
t5.py	Revamp medusa implementation so that every model can benefit. (#1588)	2024-04-25 09:13:03 +03:00
types.py	chore: add pre-commit (#1569)	2024-04-24 15:32:02 +03:00