[Diffusion] Revert CPU AMX optimizations (#30716)

sgl-kernel/csrc/cpu/torch_extension_cpu.cpp

@@ -235,8 +235,7 @@ at::Tensor flash_attn_varlen_func(
     const at::Tensor& cu_seqlens_k,
     int64_t max_seqlen_q,
     int64_t max_seqlen_k,
-    bool causal,
-    const std::optional<double>& sm_scale);
+    bool causal);
 
 // linear attention
 std::tuple<at::Tensor, at::Tensor> chunk_gated_delta_rule_cpu(
@@ -677,7 +676,7 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
   // flash attn
   m.def(
       "flash_attn_varlen_func(Tensor q, Tensor k, Tensor v, Tensor cu_seqlens_q, Tensor cu_seqlens_k, "
-      "int max_seqlen_q, int max_seqlen_k, bool causal, float? sm_scale) -> Tensor");
+      "int max_seqlen_q, int max_seqlen_k, bool causal) -> Tensor");
   m.impl("flash_attn_varlen_func", torch::kCPU, &flash_attn_varlen_func);
 
   // linear attn