more fixes5

vllm/patches/staging_kernel.py

@@ -150,9 +150,8 @@ def _deepseek_v4_stage_mega_moe_inputs_kernel(
 
     # Pack E2M1 pairs into single bytes (2 per byte, low nibble first)
     # mxf4nvf4 reads FP4 packed from SMEM — must match kernel's TMA layout
-    e2m1_flat = tl.reshape(e2m1_4bit, [BLOCK_K])
-    e2m1_lo = e2m1_flat[0::2]  # even indices → low nibble
-    e2m1_hi = e2m1_flat[1::2]  # odd indices → high nibble
+    e2m1_lo = e2m1_4bit[0::2]  # even indices → low nibble
+    e2m1_hi = e2m1_4bit[1::2]  # odd indices → high nibble
     e2m1_packed = (e2m1_hi << 4 | e2m1_lo).to(tl.uint8)  # [BLOCK_K // 2]
 
     k_offsets_out = k_block_id * (BLOCK_K // 2) + tl.arange(0, BLOCK_K // 2)