Revert "[II] fix(mtp): skip missing fused indexer shard when EXL3 keeps projections split"

tests/models/test_deepseek_mtp_mxfp8_loader.py

@@ -12,7 +12,6 @@
     dequant_mxfp8_to_bf16,
 )
 from vllm.model_executor.models.deepseek_mtp import (
-    DeepSeekMTP,
     _get_local_model_path,
     _try_load_fp8_linear_as_bf16,
 )
@@ -88,60 +87,6 @@ def test_mtp_fallback_loader_accepts_mxfp8_weight_scale():
     assert pending == {}
 
 
-def test_mtp_indexer_loader_falls_back_to_split_parameters(monkeypatch):
-    prefix = "model.layers.78.mtp_block.self_attn.indexer"
-    wk = torch.nn.Parameter(torch.empty(2, 2, dtype=torch.bfloat16))
-    weights_proj = torch.nn.Parameter(torch.empty(2, 2, dtype=torch.bfloat16))
-
-    def load_parameter(param, loaded_weight):
-        param.data.copy_(loaded_weight)
-
-    wk.weight_loader = load_parameter
-    weights_proj.weight_loader = load_parameter
-    params = {
-        f"{prefix}.wk.weight": wk,
-        f"{prefix}.weights_proj.weight": weights_proj,
-    }
-    model = SimpleNamespace(
-        config=SimpleNamespace(n_routed_experts=0, n_shared_experts=0),
-        quant_config=None,
-        model=SimpleNamespace(mtp_start_layer_idx=78, num_mtp_layers=1),
-        named_parameters=lambda: params.items(),
-        _rewrite_spec_layer_name=lambda _layer, name: name,
-    )
-    monkeypatch.setattr(
-        "vllm.model_executor.models.deepseek_mtp."
-        "rocm_aiter_ops.is_fusion_moe_shared_experts_enabled",
-        lambda: False,
-    )
-    monkeypatch.setattr(
-        "vllm.model_executor.models.deepseek_mtp.fused_moe_make_expert_params_mapping",
-        lambda *args, **kwargs: [],
-    )
-    monkeypatch.setattr(
-        "vllm.model_executor.models.deepseek_mtp.get_pp_missing_layer_names",
-        lambda _model: set(),
-    )
-    monkeypatch.setattr(
-        "vllm.model_executor.models.deepseek_mtp.get_spec_layer_idx_from_weight_name",
-        lambda _config, name: 78 if ".layers.78." in name else None,
-    )
-    wk_weight = torch.full_like(wk, 1)
-    weights_proj_weight = torch.full_like(weights_proj, 2)
-
-    loaded = DeepSeekMTP.load_weights(
-        model,
-        [
-            (f"{prefix}.wk.weight", wk_weight),
-            (f"{prefix}.weights_proj.weight", weights_proj_weight),
-        ],
-    )
-
-    assert torch.equal(wk, wk_weight)
-    assert torch.equal(weights_proj, weights_proj_weight)
-    assert loaded == set(params)
-
-
 def test_mtp_serialized_probe_uses_target_revision_for_same_model(monkeypatch):
     calls: list[tuple[str | None, str | None]] = []
 

vllm/model_executor/models/deepseek_mtp.py

@@ -814,10 +814,6 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
                         )
                     if not has_fused_target:
                         continue
-                # EXL3 may keep the indexer projections split to preserve the
-                # checkpoint's per-projection quantization metadata.
-                if param_name == "wk_weights_proj" and name_mapped not in params_dict:
-                    continue
                 name = name_mapped
 
                 if _try_load_fp8_linear_as_bf16(