adapt to dsv32 on dcu

python/sglang/srt/layers/elementwise.py

@@ -187,9 +187,7 @@ fused_dual_residual_rmsnorm_kernel_autotune = rmsnorm_autotune(
 
 def fused_dual_residual_rmsnorm(x, residual, weight1, weight2, eps, autotune=False):
     assert len(x.shape) == 2
-    assert (
-        x.shape == residual.shape and x.dtype == residual.dtype
-    ), f"{x.shape=} {residual.shape=} {x.dtype=} {residual.dtype=}"
+    assert x.shape == residual.shape and x.dtype == residual.dtype
     output, mid = torch.empty_like(x), torch.empty_like(x)
     bs, hidden_dim = x.shape
     if autotune:

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -575,10 +575,7 @@ class FusedMoE(torch.nn.Module):
             )
 
         # Flashinfer assumes w31 format for w13_weight. Same for the scales.
-        if (
-            should_use_flashinfer_trtllm_moe()
-            and self.quant_method.__class__.__name__ == "ModelOptNvFp4FusedMoEMethod"
-        ):
+        if should_use_flashinfer_trtllm_moe():
             shard_id = {"w1": "w3", "w3": "w1", "w2": "w2"}[shard_id]
 
         WEIGHT_SCALE_SUPPORTED = [e.value for e in FusedMoeWeightScaleSupported]