add adapter for Muse-Glimmer

ONBOARDING.md

@@ -202,9 +202,16 @@ registry that every test file imports (`CAUSAL_LM_MODELS` for generative,
 ```
 
 Optional fields:
-- `"dtype": "bfloat16"` / `"float32"` — set the test dtype if fp16 is wrong for
-  the model (bf16-native models, or large multipliers that overflow fp16 on CPU)
-- `"load_fn": True` — use if your adapter has a custom `load_hf_model()` function
+
+- `"is_gated": True` — exclude the checkpoint unless `SPYRE_INCLUDE_GATED=1`.
+- `"always_test": True` — include the checkpoint in addition to the smallest
+  representative selected for its adapter. This does not override gating or
+  test exclusions.
+- `"kind": "dspark_draft"` — identify a causal-LM entry as a speculative-decoding
+  draft model instead of a normal generation model.
+
+For `VISION_MODELS`, `"kind"` is required: use `"tower"` for an encoder-only
+vision tower or `"vlm"` for a complete image-to-text model.
 
 ## Step 6: Run CPU Accuracy Test
 

hf_adapters/auto_spyre_model.py

@@ -77,6 +77,7 @@
     MistralConfig,
     ModernBertConfig,
     MPNetConfig,
+    MuseGlimmerConfig,
     Olmo2Config,
     OlmoConfig,
     OPTConfig,
@@ -128,6 +129,7 @@
     hf_mistral3_vision_mm,
     hf_modernbert,
     hf_mpnet,
+    hf_muse_glimmer_mm,
     hf_olmo,
     hf_olmo2,
     hf_opt,
@@ -207,6 +209,7 @@
     Gemma4UnifiedConfig: hf_gemma4_mm,
     Granite4VisionConfig: hf_granite_vision_mm,
     Mistral3Config: hf_mistral3_vision_mm,
+    MuseGlimmerConfig: hf_muse_glimmer_mm,
 }
 
 # Sequence-classification mapping — used by
@@ -859,19 +862,25 @@ def _generate_image_text_to_text(
     """Run multimodal prefill and text decode through the shared generation loop."""
     # Processor outputs consumed by this adapter rather than by generation config.
     adapter_input_names = module._GENERATION_INPUT_NAMES
-    missing = set(adapter_input_names) - set(kwargs)
+    required_input_names = getattr(
+        module, "_GENERATION_REQUIRED_INPUT_NAMES", adapter_input_names
+    )
+    missing = set(required_input_names) - set(kwargs)
     if missing:
         raise TypeError(
             f"Missing processor inputs for {module.__name__}: {sorted(missing)}"
         )
-    adapter_inputs = {name: kwargs.pop(name) for name in adapter_input_names}
+    adapter_inputs = {
+        name: kwargs.pop(name) for name in adapter_input_names if name in kwargs
+    }
 
     # Inputs keyed by token position need the same compaction/padding as input_ids.
     # The mapping value is the padding value for each tensor.
     aligned_specs = module._GENERATION_TOKEN_ALIGNED_INPUTS
     token_aligned_inputs = {
         name: (adapter_inputs[name], pad_value)
         for name, pad_value in aligned_specs.items()
+        if name in adapter_inputs
     }
     # Image tensors and metadata are passed to prefill unchanged.
     pass_through_inputs = {