[gemma4] Address PR review: hg:// 12B cards, Literal audio_mode, drop getattr

- Point gemma4_12b{,_it} cards at hg://google/gemma-4-12B{,-it} instead of
  the internal /checkpoint/smallomnillm path so the OSS cards don't leak
  cluster-only paths.
- Type Gemma4AudioConfig.audio_mode as Literal["conformer", "linear"] —
  the only two supported pipelines — for static checking.
- Drop defensive getattr(config.audio_config, "audio_mode", ...) calls in
  Gemma4Factory; audio_mode is a proper dataclass field with a default,
  no BC story to preserve.

Author: Yunchao Yang

src/fairseq2/assets/cards/models/gemma4.yaml

@@ -79,15 +79,15 @@ tokenizer_family: gemma4
 name: gemma4_12b
 model_family: gemma4
 model_arch: 12b
-checkpoint: "file:///checkpoint/smallomnillm/shared/models/gemma-4-12B"
-tokenizer: "file:///checkpoint/smallomnillm/shared/models/gemma-4-12B"
+checkpoint: "hg://google/gemma-4-12B"
+tokenizer: "hg://google/gemma-4-12B"
 tokenizer_family: gemma4
 
 ---
 
 name: gemma4_12b_it
 model_family: gemma4
 model_arch: 12b_it
-checkpoint: "file:///checkpoint/smallomnillm/shared/models/gemma-4-12B-it"
-tokenizer: "file:///checkpoint/smallomnillm/shared/models/gemma-4-12B-it"
+checkpoint: "hg://google/gemma-4-12B-it"
+tokenizer: "hg://google/gemma-4-12B-it"
 tokenizer_family: gemma4

src/fairseq2/models/gemma4/audio/config.py

@@ -7,6 +7,7 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
+from typing import Literal
 
 
 @dataclass(kw_only=True)
@@ -29,7 +30,7 @@ class Gemma4AudioConfig:
       and ``rms_norm_eps`` are read.
     """
 
-    audio_mode: str = "conformer"
+    audio_mode: Literal["conformer", "linear"] = "conformer"
     """Audio pipeline selector: ``"conformer"`` or ``"linear"``."""
 
     hidden_size: int = 1024

src/fairseq2/models/gemma4/factory.py

@@ -541,7 +541,7 @@ def create_audio_tower(self) -> Module | None:
             return None
 
         # Unified family: no tower; embedder consumes raw waveform frames.
-        if getattr(config.audio_config, "audio_mode", "conformer") == "linear":
+        if config.audio_config.audio_mode == "linear":
             return None
 
         from fairseq2.models.gemma4.audio.tower import Gemma4AudioTower
@@ -571,9 +571,7 @@ def create_audio_embedder(self) -> Module | None:
         # behaviour where the embedder casts raw inputs to its weight dtype
         # before the norm. The classic conformer path does not need this
         # (its inputs come from the audio tower already in the right dtype).
-        cast_input_dtype = (
-            getattr(config.audio_config, "audio_mode", "conformer") == "linear"
-        )
+        cast_input_dtype = config.audio_config.audio_mode == "linear"
 
         return Gemma4MultimodalAudioEmbedder(
             output_proj_dims=config.audio_config.output_proj_dims,