-
Mitch Naylor authored
* add mega file structure and plain pytorch version of mega source code * added config class with old naming conventions * filled in mega documentation * added config class and embeddings with optional token types * updated notes * starting the conversion process, deleted intermediate and added use_cache back to config * renamed config attributes in modeling_mega.py * checkpointing before refactoring incremental decoding functions * removed stateful incremental key/values for EMA and self-attention * refactored MovingAverageGatedAttention to remove stateful k/v history and use unified attention mask * MovingAverageGatedAttention works with incremental decoding + past values, added sequence length enforcement * more comments in MovingAverageGatedAttention + checkpointing before GatedCrossAttention * bug fix in attention mask handling in MovingAverageGatedAttention * removed incremental state from GatedCrossAttention and removed IncrementalState class * finished gated cross attention and got MegaLayer working * fixed causal masking in mega decoder * fixed how padding and causal masks are passed through MegaLayer with and without k/v caching * finished MegaModel; tested with encoder, decoder-only, and cross-attention type inputs; started work on downstream classes; removed mentions of position_ids * added optional dense hidden layer for masked and causal LM classes * docstring updates in MultiHeadEMA and GatedCrossAttention, removed unnecessary inputs in cross-attention * removed before_attn_fn in Mega class and updated docstrings and comments up to there * bug fix in MovingAverageGatedAttention masking * working conversion of MLM checkpoint in scratchpad script -- perfect matches * moved arg for hidden dense layer in LM head to config; discovered issue where from_pretrained is renaming gamma and beta parameters * renamed gamma and beta parameters to avoid HF renaming when loading from checkpoint * finished checkpoint conversion script * cleanup old class in mega config script * removed 'copied from' statements and passing integration tests * added num_attention_heads=1 to config for integration compatibility, decoder tests working, generation tests failing * fixed tuple output of megamodel * all common tests passing after fixing issues in decoder, gradient retention, and initialization * added mega-specific tests, ready for more documentation and style checks * updated docstrings; checkpoint before style fixes * style and quality checks, fixed initialization problem in float_tensor, ready for PR * added mega to toctree * removed unnecessary arg in megaconfig * removed unused arg and fixed code samples with leftover roberta models * Apply suggestions from code review Applied all suggestions except the one renaming a class, as I'll need to update that througout Co-authored-by:
Arthur <48595927+ArthurZucker@users.noreply.github.com> * fixed issue where .view breaks batch dimension, conversion script fixed with absolute imports, updated readme with Mega->MEGA * removed asserts in Mega code, renamed sequencenorm, gatedcrossattention, and NFFN, replaced get_activation_fn with ACTFN, and added sequencenorm to layer norms * reformatted .forward() docstrings to match style and removed unused mask input in cross-attention * removed all reset_parameters() methods and rolled into MegaPreTrainedModel._init_weights() * renamed all single-letter variables and improved readability in tensor size comments, Mega->MEGA in 2 documentation files * variable names in NFFN * manual Mega->MEGA changes in docs * Mega->MEGA in config auto * style and quality fixes * Apply suggestions from code review Co-authored-by:
Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * added Mega prefixes where missing, reverted MegaSequenceNorm to if-else, other module renaming requested in PR * style and quality fixes + readme updates pointing to main --------- Co-authored-by:
57f25f4b
