feat(engine): add Qwen3-VL dense and MoE support to Megatron path (#1301)

* feat(engine): add Qwen3-VL dense support to Megatron path

Extend the Megatron engine to train Qwen3-VL dense models end-to-end:
mcore→HF weight conversion for update_weights and HF→mcore loading that
handles Qwen3-VL's nested HF config layout. Without this, GRPO/PPO of
any Qwen3-VL model on the Megatron backend is blocked.

Author: Adiactive

areal/engine/core/model.py

@@ -6,6 +6,7 @@
     "qwen2_vl",
     "qwen2_5_vl",
     "qwen3_vl",
+    "qwen3_vl_moe",
     "gemma3",
 ]
 # This registry is used to check if a model is a vision model that we have checked it works with AReaL.
@@ -23,19 +24,44 @@ def is_qwen2_vl_model(model_type: str) -> bool:
 
 
 def is_qwen3_vl_model(model_type: str) -> bool:
-    return model_type in ["qwen3_vl"]
+    """True for the Qwen3-VL family (dense and MoE).
+
+    Existing call sites in ``fsdp_engine``, ``fsdp_utils/parallel``, and
+    ``awex/fsdp_adapter`` gate family-level behaviour (mRoPE index,
+    attention-mask handling) that is identical for dense and MoE, so this
+    helper covers both. Use ``is_qwen3_vl_moe_model`` when the MoE-vs-dense
+    distinction matters.
+    """
+    return model_type in ("qwen3_vl", "qwen3_vl_moe")
+
+
+def is_qwen3_vl_moe_model(model_type: str) -> bool:
+    return model_type == "qwen3_vl_moe"
 
 
 def is_qwen_vl_model(model_type: str) -> bool:
     return is_qwen2_vl_model(model_type) or is_qwen3_vl_model(model_type)
 
 
+def lang_config(hf_config):
+    """Return the language-model side of a (possibly nested) HF config.
+
+    Qwen3-VL and similar VLMs nest text-model attributes (vocab_size,
+    num_attention_heads, num_key_value_heads, hidden_size, head_dim) under
+    ``hf_config.text_config``. Qwen2.5-VL and pure text models keep them
+    flat. Use this anywhere the caller wants a language-side attribute and
+    doesn't know the model family up front.
+    """
+    return getattr(hf_config, "text_config", hf_config)
+
+
 def is_gemma3_model(model_type: str) -> bool:
     return model_type in ["gemma3"]
 
 
 VALID_MOE_MODELS = [
     "qwen3_moe",
+    "qwen3_vl_moe",
     "qwen3_5_moe",
     "qwen3_5_moe_text",
     "bailing_moe_v2",

areal/engine/megatron_engine.py

@@ -56,7 +56,11 @@
     init_custom_process_group,
     warmup_process_groups,
 )
-from areal.engine.core.model import disable_dropout_in_model, is_valid_vision_model
+from areal.engine.core.model import (
+    disable_dropout_in_model,
+    is_valid_vision_model,
+    lang_config,
+)
 from areal.engine.megatron_utils.checkpointer import MegatronCheckpointManager
 from areal.engine.megatron_utils.deterministic import set_deterministic_algorithms
 from areal.engine.megatron_utils.fp8 import FP8BlockwiseTensorHelper
@@ -1463,7 +1467,7 @@ def _collect_param(
             duplicated_param_names=self._duplicated_param_names,
             gated_linear_unit=is_glu,
         )
-        param = remove_padding(name, param, self.hf_config.vocab_size)
+        param = remove_padding(name, param, lang_config(self.hf_config).vocab_size)
 
         if isinstance(param, FP8BlockwiseTensorHelper):
             # FP8 is stored as uint8, so element_size is 1 byte