The decoder CrossAttention module (mstar/model/components/attention.py) currently uses plain nn.Linear projections and is not TP/SP-compatible — no column/row-parallel splits over heads. Serving the encoder-decoder decoder under tensor parallelism needs a TP/SP-aware cross-attention variant (and a corresponding TP config, e.g. for whisper_large). Marked in code as TODO(#160).
The decoder
CrossAttentionmodule (mstar/model/components/attention.py) currently uses plain nn.Linear projections and is not TP/SP-compatible — no column/row-parallel splits over heads. Serving the encoder-decoder decoder under tensor parallelism needs a TP/SP-aware cross-attention variant (and a corresponding TP config, e.g. for whisper_large). Marked in code as TODO(#160).