[ROCBLAS_GEMM] Default use of hipMallocAsync

Default use of hipMallocAsync rather than hipMalloc in rocblas_gemm and
add support of fp16_fp16_fp32 in rocblas_gemm.
Signed-off-by: wenjh <wenjh@sugon.com>

tests/pytorch/test_numerics.py

@@ -2088,6 +2088,8 @@ def test_transformer_layer_hidden_states_format(dtype, bs, model):
 def test_kv_cache_accuracy(dtype, bs, model_key, use_RoPE, input_format, module, backend, is_paged):
     reset_rng_states()
     
+    if backend in ["FusedAttention"]:
+        pytest.skip("Not support FusedAttention")
     if backend in ["FusedAttention", "FlashAttention"] and dtype == torch.float32:
         pytest.skip("FusedAttention and FlashAttention do not support FP32")
     if use_RoPE:

transformer_engine/common/gemm/rocm_gemm.cu

@@ -1459,7 +1459,7 @@ void rocblas_gemm(const Tensor *inputA,
   // extract the stream order alloc env
   bool stream_order_alloc = false;
   if (const char* env_p = std::getenv("ROCBLAS_STREAM_ORDER_ALLOC") ) {
-    if (env_p != nullptr && std::string(env_p) == "1")
+    if (env_p == nullptr || std::string(env_p) == "1")
       stream_order_alloc = true;
   }
 
@@ -1467,6 +1467,7 @@ void rocblas_gemm(const Tensor *inputA,
 
 
   NVTE_CHECK((A_type==rocblas_datatype_f16_r && B_type==rocblas_datatype_f16_r && D_type==rocblas_datatype_f16_r) || 
+  (A_type==rocblas_datatype_f16_r && B_type==rocblas_datatype_f16_r && D_type==rocblas_datatype_f32_r) ||
        (A_type==rocblas_datatype_bf16_r && B_type==rocblas_datatype_bf16_r && D_type==rocblas_datatype_bf16_r) || 
        (A_type==rocblas_datatype_bf16_r && B_type==rocblas_datatype_bf16_r && D_type==rocblas_datatype_f32_r) ||
        (A_type==rocblas_datatype_f32_r && B_type==rocblas_datatype_f32_r && D_type==rocblas_datatype_f32_r) ||