Address PR #236 review: drop Gemma4 hunks, document limitations

- Rebase onto latest main, dropping Gemma4 commits (already in PR #239)
- Document batch=1 limitation in FunASREmbeddingModel
- Document audio_token_id=0 collision workaround
- Add explicit shape-unknown comment in SkipLayerNorm rank guard

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Justin Chu <justinchu@microsoft.com>

Author: justinchuby

src/mobius/models/fun_asr.py

@@ -338,6 +338,17 @@ class FunASREmbeddingModel(nn.Module):
 
     Output:
         inputs_embeds: ``(batch, seq_len, hidden_size)``
+
+    Note:
+        This model only supports ``batch=1``. The CumSum-based audio scatter
+        uses a flat ``audio_features`` table shared across the batch, so with
+        ``batch > 1`` each row's cumsum restarts at 1 and indexes the same
+        prefix of ``audio_features``.
+
+        When ``audio_token_id`` is 0, generated token 0 during autoregressive
+        decoding would collide with audio placeholders. Callers should bypass
+        this model for decode steps and use the embed_tokens weight table
+        directly (see ``examples/fun_asr.py``).
     """
 
     def __init__(self, config: ArchitectureConfig):

src/mobius/rewrite_rules/_skip_layer_norm.py

@@ -74,7 +74,8 @@ def check(self, context, add_out, norm_out, **_):
         # Both Add inputs must have at least 2 dimensions.  This prevents
         # fusing a bias-Add (e.g. MatMul + 1D bias → LayerNorm) which
         # would produce a SkipLayerNormalization with a 1D skip input
-        # that ORT rejects.
+        # that ORT rejects.  Unknown shapes are allowed through since
+        # most intermediate values lack static shape info.
         for i, inp in enumerate(producer.inputs):
             if inp is not None and inp.shape is not None:
                 rank = len(inp.shape)
@@ -157,6 +158,8 @@ def check(self, context, add_out, norm_out, **_):
 
         # Both Add inputs must have at least 2 dimensions — reject
         # bias-Add patterns (e.g. MatMul + 1D bias → LayerNorm).
+        # Unknown shapes are allowed through since most intermediate
+        # values lack static shape info.
         for i, inp in enumerate(producer.inputs):
             if inp is not None and inp.shape is not None:
                 rank = len(inp.shape)