Fix T5 export: encoder attention mask and static decoder shapes

[duyhv-qualgo]

Problem

Three bugs in optimum/exporters/executorch/integrations.py that cause wrong outputs or export failures when exporting T5 with ExecuTorch.

Bug 1 — Encoder ignores padding tokens

Seq2SeqLMEncoderExportableModule.forward calls self.encoder(input_ids) without an attention_mask. When encoder inputs are padded to a fixed length, PAD tokens (id=0) attend to real positions and corrupt the hidden states at every token position — the encoder output is semantically wrong regardless of the input text.

Fix: compute attention_mask = (input_ids != 0).long() inside forward and pass it to the encoder. Also zero out hidden states at PAD positions so the decoder cross-attention ignores them.

Bug 2 — Decoder ignores encoder attention mask → wrong position bias

Seq2SeqLMDecoderExportableModuleWithStaticCache.forward calls self.decoder(...) without encoder_attention_mask. T5 computes a relative position bias in cross-attention scaled by key_length. Without the mask, key_length equals the full padded length (e.g. 512) instead of the real token count, producing a ~20× logit scale error and completely wrong greedy-decoding output.

Fix: add encoder_attention_mask: Tensor | None = None to forward and pass it through to self.decoder(...).

Bug 3 — Dynamic encoder dim conflicts with static KV cache (T5 export failure)

_export_decoder marks encoder_hidden_states dim-1 as a dynamic symbol (encoder_hidden_seq_length). With transformers 5.0, T5's cross-attention slices a causal mask against the static KV-cache size:

causal_mask = mask[:, :, :, : key_states.shape[-2]]  # static int at export time
position_bias = position_bias + causal_mask            # symbolic dim → shape conflict

This raises RuntimeError: tensor a (1024) must match tensor b (s96) during torch.export.

Fix: use dynamic_shapes=None for T5 decoder export (fully static). Callers pad encoder inputs to max_seq_len before encoding — export() is updated to do this automatically.

Changes

• Seq2SeqLMEncoderExportableModule.forward: compute and apply attention_mask; zero PAD positions in output.
• Seq2SeqLMDecoderExportableModuleWithStaticCache.forward: add encoder_attention_mask parameter and pass to decoder.
• Seq2SeqLMExportableModule._export_decoder (T5 path): use dynamic_shapes=None; pass encoder_attention_mask.
• Seq2SeqLMExportableModule.export: use max_seq_len-padded encoder input for T5 to match static decoder shape; build and pass encoder_attention_mask.

Verification

Tested on a Helsinki-NLP–style T5 seq2seq (en↔vi) checkpoint, max_seq_len=512, XNNPACK recipe:

	Before	After
Encoder output at PAD positions	non-zero (corrupt)	zeroed
ExecuTorch vs HF exact match	0/5	5/5
torch.export succeeds	no (shape conflict)	yes