Register qwen3_5_moe_text for text-only Qwen3.6 export

Qwen/Qwen3.6-35B-A3B ships model_type=qwen3_5_moe with a vision_config, so
the builder overrides it to qwen3_5_moe_vl and unwraps text_config (whose own
model_type is qwen3_5_moe_text). Unlike every other VL family (gemma3_text,
qwen3_vl_text, qwen3_5_vl_text), the MoE VL had no registered text-only
sibling, so:

  - registry.get("qwen3_5_moe_text") raised KeyError, and
  - build(..., text_only=True) on the VL checkpoint failed (no
    _TEXT_ONLY_MODEL_TYPE entry for qwen3_5_moe_vl).

Qwen35MoECausalLMModel already strips the language_model. prefix, drops
visual./MTP keys, and unpacks fused experts, so it consumes the VL text
weights directly. Register qwen3_5_moe_text -> Qwen35MoECausalLMModel, wire
the text-only override (qwen3_5_moe_vl -> qwen3_5_moe_text, idempotent
self-map), and add default-id/family map entries.

Add an L1 build-graph config (reusing the qwen3_5_moe hybrid MoE tiny config)
so registry-completeness passes and the new key builds. Verified against the
real Qwen3.6 config that both text_only routing and text_config.model_type
now resolve to Qwen35MoECausalLMModel.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: justinchuby <justinchuby@users.noreply.github.com>

Author: justinchuby

src/mobius/_registry.py

@@ -604,6 +604,13 @@ def _detect_fallback_registration(hf_config) -> ModelRegistration | None:
     "qwen2_vl_text": ModelRegistration(Qwen25VLTextModel),
     "qwen3_5": ModelRegistration(Qwen35VL3ModelCausalLMModel, task="hybrid-qwen-vl"),
     "qwen3_5_moe_vl": ModelRegistration(Qwen35MoEVL3ModelCausalLMModel, task="hybrid-qwen-vl"),
+    # Text-only sibling of ``qwen3_5_moe_vl`` (Qwen3.6-35B-A3B). The MoE
+    # backbone ``Qwen35MoECausalLMModel`` already strips ``language_model.``
+    # and drops ``visual.``/MTP keys, so it consumes the VL checkpoint's text
+    # weights directly; ``build(..., text_only=True)`` routes here via
+    # ``_TEXT_ONLY_MODEL_TYPE``. It also matches the VL ``text_config``'s own
+    # ``model_type=qwen3_5_moe_text`` so that config resolves cleanly.
+    "qwen3_5_moe_text": ModelRegistration(Qwen35MoECausalLMModel),
     "qwen3_5_vl": ModelRegistration(Qwen35VL3ModelCausalLMModel, task="hybrid-qwen-vl"),
     "qwen3_5_vl_text": ModelRegistration(Qwen35VLTextModel),
     "qwen3_vl": ModelRegistration(Qwen3VL3ModelCausalLMModel, task="qwen-vl"),
@@ -791,6 +798,12 @@ def _create_default_registry() -> ModelRegistry:
 _TEXT_ONLY_MODEL_TYPE: dict[str, str] = {
     "gemma4_unified": "gemma4_unified_text",
     "gemma4_unified_text": "gemma4_unified_text",
+    # Qwen3.5-MoE-VL (Qwen3.6-35B-A3B): export just the hybrid MoE text
+    # backbone as a standalone decoder-only LLM. The builder overrides
+    # ``qwen3_5_moe`` -> ``qwen3_5_moe_vl`` when a ``vision_config`` is present,
+    # so the text-only override keys off the VL type here.
+    "qwen3_5_moe_vl": "qwen3_5_moe_text",
+    "qwen3_5_moe_text": "qwen3_5_moe_text",
 }
 
 
@@ -894,6 +907,7 @@ def _create_default_registry() -> ModelRegistry:
     "qwen2_moe": "Qwen/Qwen1.5-MoE-A2.7B-Chat",
     "qwen3_moe": "Qwen/Qwen3-30B-A3B",
     "qwen3_5_moe": "Qwen/Qwen3.5-MoE-A3B-128K",
+    "qwen3_5_moe_text": "Qwen/Qwen3.6-35B-A3B",
     "qwen3_next": "Qwen/Qwen3-235B-A22B",
     "granitemoe": "ibm-granite/granite-3.0-1b-a400m-instruct",
     "olmoe": "allenai/OLMoE-1B-7B-0924",
@@ -1141,6 +1155,7 @@ def _create_default_registry() -> ModelRegistry:
     "qwen3_moe": "qwen",
     "qwen3_5_text": "qwen",
     "qwen3_5_moe": "qwen",
+    "qwen3_5_moe_text": "qwen",
     "qwen3_next": "qwen",
     "qwen2_vl": "qwen",
     "qwen2_vl_text": "qwen",

tests/_test_configs.py

@@ -519,6 +519,29 @@ def _base_config(config_cls=None, **overrides) -> ArchitectureConfig:
         },
         True,
     ),
+    # Text-only sibling of the Qwen3.5-MoE-VL (Qwen3.6-35B-A3B) checkpoint,
+    # exported via ``build(..., text_only=True)``. Same hybrid MoE backbone as
+    # ``qwen3_5_moe`` above; registered separately so the VL ``text_config``'s
+    # ``model_type=qwen3_5_moe_text`` and the text-only override both resolve.
+    (
+        "qwen3_5_moe_text",
+        {
+            "hidden_act": "silu",
+            "layer_types": ["linear_attention", "full_attention"],
+            "partial_rotary_factor": 0.25,
+            "mrope_interleaved": True,
+            "num_local_experts": 4,
+            "num_experts_per_tok": 2,
+            "moe_intermediate_size": 32,
+            "shared_expert_intermediate_size": 32,
+            "linear_num_value_heads": 4,
+            "linear_num_key_heads": 2,
+            "linear_key_head_dim": 16,
+            "linear_value_head_dim": 16,
+            "linear_conv_kernel_dim": 4,
+        },
+        True,
+    ),
     # === Falcon and Bloom ===
     # dual_ln=True: Falcon with new_decoder_architecture uses separate ln_attn + ln_mlp.
     # hidden_act="gelu": real Falcon uses GELU (HF FalconConfig.activation default);

tests/build_graph_test.py

@@ -80,6 +80,7 @@
     "minimax",
     "qwen3_5_text",
     "qwen3_5_moe",
+    "qwen3_5_moe_text",
     "qwen3_next",
     # Models using LinearAttention / CausalConvWithState custom ops
     # prevent full shape/type propagation through com.microsoft domain.