Add Marian encoder-decoder support (static cache)

[vortex-captain]

Summary

• Marian (Helsinki-NLP/opus-mt-*) encoder-decoder support: MarianEncoderWrapper / MarianDecoderWrapper using WinMLStaticCache (multi-dim index_put_ → ScatterND). The decoder's OnnxConfig exposes cache_position as the position input; under static cache, stock HF's position_ids=cache_position already indexes the sin/cos table correctly, so no embedding patch is applied at runtime. The PATCHING_SPECS infra (_patched_marian_sinusoidal_forward / _build_marian_patching_specs) is kept in the file with its assignment commented, for reference and future sliding-window re-export. Registered composite: ("marian", "translation") → WinMLMarianModel.

• BART (facebook/bart-large-cnn, sshleifer/distilbart-cnn-12-6) encoder-decoder support: BartEncoderWrapper / BartDecoderWrapper using WinMLStaticCache. Same static-cache approach as Marian — BartLearnedPositionalEmbedding.forward correctly receives position_ids=cache_position from stock HF and applies its +self.offset shift, so no embedding patch is applied at runtime. The _patched_bart_learned_forward patch spec remains in the file (commented out on the decoder config) for future sliding-window re-export. Registered composite: ("bart", "summarization") → WinMLBartModel.

kv_cache changes

• WinMLStaticCache.update uses multi-dim index_put_ with full (batch, head, pos) coord tuples so ONNX export emits ScatterND instead of ScatterElements.
• WinMLCache.num_layers now reads len(self.layers) after super().__init__ (HF's StaticCache.__init__ builds self.layers from config.get_text_config(decoder=True).num_hidden_layers, i.e. the decoder's layer count). Previously config.num_hidden_layers on the outer BartConfig maps to encoder_layers, which fails for asymmetric distilled variants like distilbart-cnn-12-6 (encoder_layers=12, decoder_layers=6). No-op for symmetric models.
• PastKeyValueInputGenerator reads normalized_config.head_dim unconditionally (no try/except fallback). MarianConfig / BartConfig have no native head_dim attr, so _MarianDecoderNormalizedConfig / _BartDecoderNormalizedConfig — real NormalizedConfig subclasses with UPPERCASE class-level attribute mappings (same pattern as stock NormalizedTextConfig) — expose head_dim as a computed property.

Verified

Model	Test	Result
Helsinki-NLP/opus-mt-fr-en	3 fr→en translations	3/3 exact MATCH vs HF PyTorch MarianMTModel
facebook/bart-large-cnn	2 CNN-style summarizations	2/2 PREFIX_MATCH vs HF PyTorch BartForConditionalGeneration
sshleifer/distilbart-cnn-12-6 (asymmetric: 12 enc / 6 dec)	2 CNN-style summarizations	2/2 PREFIX_MATCH vs HF PyTorch

PREFIX_MATCH means WinML ONNX and HF PyTorch agree bit-identically for the full shared prefix (e.g. 53 tokens for BART article 1, 40 for article 2) and then diverge at a single EOS-vs-continuation argmax flip near the natural sentence boundary — the expected fingerprint of fp32 drift from fused ops (gelu_fusion, matmul_add_fusion).

Commits

1. Add Marian (Helsinki-NLP/opus-mt) encoder-decoder support
2. Add BART (facebook/bart-large-cnn) encoder-decoder support
3. Fix WinMLCache.num_layers for asymmetric encoder-decoder models
4. Derive head_dim in per-model NormalizedConfig instead of kv_cache fallback

[xieofxie]

do we need tests? although no one will review them (?)

[zhenchaoni]

Reviewed the bart.py

[vortex-captain]

do we need tests? although no one will review them (?)

UTs added

[zhenchaoni]

Overall is good to me. I leveraged AI to do a check as well. I think both AI comments can be solved by refactoring the code to use the common decoder_wrapper. You can consider doing it in a separate PR.

1. Heavy structural duplication between bart.py and marian.py

These two files are ~90% structurally identical:

Component	bart.py	marian.py	Difference
*EncoderWrapper	✓	✓	class name + outer model class
*DecoderWrapper.forward	line 281-352	line 319-392	bit-identical body, just renamed class
*DecoderIOConfig.inputs/outputs	line 432-456	line 472-496	bit-identical
_*DecoderNormalizedConfig	line 388-403	line 428-443	bit-identical field map
_patched_*_forward	line 142-184	line 179-214	same shape (different embedding type)
_build_*_patching_specs	line 187-200	line 217-233	bit-identical pattern

This is independent of any other PR — even between just bart and marian, a shared _BartFamilyDecoderWrapperBase (or even just _BartFamilyDecoderIOConfigBase for the inputs/outputs body) would cut ~250-300 LOC of duplicated code. Two parallel files mean every future bug fix has to be applied twice.

Suggestion: factor at least the IOConfig inputs/outputs body and the NormalizedConfig field map into a shared base in either bart.py (and marian.py imports from there) or a new _bart_family.py. The decoder wrapper's forward body is harder to share without a true base class, but the IOConfig body and the NormalizedConfig are easy wins.

---

2. Positional args[N] indexing couples the wrapper to the IOConfig's input order — fragile

In [bart.py:293-313]:

decoder_input_ids = args[0]
encoder_hidden_states = args[1]
attention_mask = args[2]
decoder_attention_mask = args[3]
cache_position = args[4]
kv_start = 5

max_cache_len = args[kv_start].size(2)
...
for i in range(self.num_layers):
    self_attn_cache.layers[i].keys = args[kv_start + i * 2]
    self_attn_cache.layers[i].values = args[kv_start + i * 2 + 1]

The contract is implicit: "the IOConfig's inputs property declares names in this exact order; forward receives them positionally in the same order; we read args[N] by hardcoded index." Nothing in the wrapper checks that the IOConfig matches this expectation — and nothing in the IOConfig knows that the wrapper depends on its order. If anyone reorders or inserts an input in the IOConfig's inputs dict (e.g., adds head_mask), forward silently misreads and produces nonsense ONNX.

Suggestion: in forward, build a name-keyed dict from the IOConfig:

inputs = dict(zip(self.onnx_config.inputs.keys(), args, strict=True))
decoder_input_ids = inputs["decoder_input_ids"]
# ...
for i in range(self.num_layers):
    self_attn_cache.layers[i].keys = inputs[f"past_{i}_key"]

This makes the wrapper explicit about which names it depends on, fails loudly with strict=True if the count mismatches, and is robust to IOConfig reorderings.

---