Fix workspace_shapes: output dim is hidden_dim, not K*2

K comes from hidden_states.size(-1) which is the full BF16 dimension
(7168), not the packed weight dimension. K*2=14336 is wrong.
The MoE output is always hidden_dim (7168).

vllm/patches/fused_moe/experts/cutedsl_moe.py

@@ -271,13 +271,9 @@ class CuTeDSLMoEExperts(mk.FusedMoEExpertsModular):
         # Our runner manages its own workspace internally (pre-allocated buffers)
         workspace1 = (0,)
         workspace2 = (0,)
-        # K is the packed dimension from w1.shape[-1].
-        # For NVFP4 uint8 packed weights, K_packed = K_logical // 2.
-        # The output of the L2 GEMM is hidden_dim (unpacked).
-        # If K == hidden_dim, weights are BF16 (not packed).
-        # If K == hidden_dim // 2, weights are NVFP4 packed.
-        output_dim = max(K * 2, self.hidden_dim)
-        output = (M, output_dim)
+        # The output of the full 2-stage MoE pipeline is hidden_dim.
+        # K comes from hidden_states.size(-1) (full BF16 dimension, not packed).
+        output = (M, self.hidden_dim)
         return (workspace1, workspace2, output)
 
     def apply(
@@ -313,9 +309,4 @@ class CuTeDSLMoEExperts(mk.FusedMoEExpertsModular):
         )
 
         # Copy result into output tensor
-        if result.shape != output.shape:
-            import sys
-            print(f"[CuTeDSL MoE] SHAPE MISMATCH: result={result.shape} output={output.shape} "
-                  f"hidden_dim={self.hidden_dim} w1={w1.shape if w1 is not None else None} "
-                  f"hs={hidden_states.shape}", file=sys.stderr, flush=True)
         output.copy_(result)