fix(glA): init default conv_state at first prefill to avoid NoneType unpack 72dc299 mehdi999 commited on Oct 30
fix: let decoder manage its own cache (avoid None conv_state in FLA) 750aa27 mehdi999 commited on Oct 30
deps: drop causal-conv1d (needs torch at build-time; not required for demo) 5fd5d30 mehdi999 commited on Oct 29
fix: SwiGLU ctor signature (drop out_features); add causal-conv1d 7314545 mehdi999 commited on Oct 29
fix: expose SimpleGLADecoder.config for registry instantiate_from_config ca4ede6 mehdi999 commited on Oct 29
fix(GLA): provide safe past_key_values with default conv_state=(None,None,None) bb32e4f mehdi999 commited on Oct 29
fix(GLA): provide safe past_key_values with default conv_state=(None,None,None) 80ec7ec mehdi999 commited on Oct 29
fix(GLA): provide safe past_key_values with default conv_state=(None,None,None) a69b69a mehdi999 commited on Oct 29
fix(GLA): provide safe past_key_values with default conv_state=(None,None,None) 159df2e mehdi999 commited on Oct 29
inference: remove empty cache (GLA expects prefetched conv_state) d7cfd9b mehdi999 commited on Oct 29
fix: handle VelocityHeadSamplingParams signature; better error logging 0a0019f mehdi999 commited on Oct 29
deps: add flash-linear-attention; bump torch/torchaudio to 2.5.1 and transformers to 4.53.0 c99b6f0 mehdi999 commited on Oct 29
inference-only: drop zcodec deps (relative imports) and avoid training/tokenizer imports at package init b650fff mehdi999 commited on Oct 29
deps: drop flash-linear-attention to resolve Torch/Transformers conflicts 0d64ef9 mehdi999 commited on Oct 29
deps: align with pyproject; torch/torchaudio pinned to cu121-compatible 6a0a374 mehdi999 commited on Oct 29
