[model] fix: restore position_embedding_type and gradient_accumulation_fusion for MLA models

Two regressions caused a ~30% perf drop on DSV3 GB300 MXFP8 (1292 → 914
TFLOPs, March 14 → March 18), introduced by PRs #2599 and #2739.

Fix 1 (model_bridge.py): MLA providers with YaRN rope_scaling were
getting position_embedding_type="learned_absolute". The base
provider_bridge() only set "rope" when rope_type != "yarn", but for MLA
models YaRN scaling is handled via mla_rope_params — position_embedding_type
must always be "rope". Add is_mla_provider check to handle this correctly.

Fix 2 (deepseek_v2/v3_bridge.py): gradient_accumulation_fusion was
hardcoded True before #2739 removed it in favor of the dynamic
can_enable_gradient_accumulation_fusion() check. That check returns False
if fused_weight_gradient_mlp_cuda (APEX CUDA extension) is not importable,
silently disabling the fusion. Restore explicit True for DeepSeek V2/V3.

Signed-off-by: Yu Yao <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Author: yaoyu-33

src/megatron/bridge/models/conversion/model_bridge.py

@@ -448,6 +448,7 @@ def provider_bridge(self, hf_pretrained: HFPreTrained) -> ModelProviderTarget:
             ModelProviderTarget: A configured model provider instance
         """
         from megatron.bridge.models.gpt_provider import GPTModelProvider
+        from megatron.bridge.models.mla_provider import MLAModelProvider
 
         hf_config = hf_pretrained.config
 
@@ -458,22 +459,25 @@ def provider_bridge(self, hf_pretrained: HFPreTrained) -> ModelProviderTarget:
 
         # Use specified provider class, defaulting to GPTModelProvider
         provider_class = self.PROVIDER_CLASS if self.PROVIDER_CLASS is not None else GPTModelProvider
+        is_mla_provider = issubclass(provider_class, MLAModelProvider)
         # Filter kwargs to only fields the provider dataclass accepts, so that MLA-only None
         # values (q_lora_rank, kv_lora_rank, …) are silently dropped for non-MLA providers
         # while still being passed through for MLA providers that declare them as fields.
         valid_fields = provider_class.__dataclass_fields__
         provider = provider_class(**{k: v for k, v in provider_kwargs.items() if k in valid_fields})
 
-        # All models that flow through the base provider_bridge use RoPE.
-        # YARN is already handled above (hf_config_to_provider_kwargs sets
-        # position_embedding_type="yarn"), so we only need to guard against
-        # overwriting it here.  Every other rope_type value (None, "default",
-        # "llama3", "longrope", …) means plain RoPE.
+        # Determine position_embedding_type from HF rope_scaling.
+        # For GPT providers: rope_type=="yarn" → "yarn"; everything else → "rope".
+        # For MLA providers: always "rope" — YaRN scaling parameters are applied
+        # separately via mla_rope_params (rotary_scaling_factor, mscale, etc.) and
+        # position_embedding_type="yarn" is not a valid MLA config value.
         hf_rope_scaling = getattr(hf_config, "rope_scaling", None)
         rope_type = None
         if hf_rope_scaling:
             rope_type = hf_rope_scaling.get("type") or hf_rope_scaling.get("rope_type")
-        if rope_type != "yarn":
+        if rope_type == "yarn" and not is_mla_provider:
+            provider.position_embedding_type = "yarn"
+        else:
             provider.position_embedding_type = "rope"
 
         # Apply MLA rope params via setattr (for MLA models like DeepSeek, Kimi)

src/megatron/bridge/models/deepseek/deepseek_v2_bridge.py

@@ -62,6 +62,7 @@ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> MLAModelProvider
         provider.moe_permute_fusion = True
 
         provider.apply_rope_fusion = False
+        provider.gradient_accumulation_fusion = True
         provider.bias_activation_fusion = True
         provider.bias_dropout_fusion = True
         provider.cross_entropy_fusion_impl = "te"

src/megatron/bridge/models/deepseek/deepseek_v3_bridge.py

@@ -69,6 +69,7 @@ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> MLAModelProvider
         provider.moe_aux_loss_coeff = 0.0001
 
         provider.apply_rope_fusion = False
+        provider.gradient_accumulation_fusion = True
         provider.bias_activation_fusion = True
         provider.bias_dropout_fusion = True
         provider.cross_entropy_fusion_impl = "te"