diff --git a/lib/nnc/ccv_cnnp_model.c b/lib/nnc/ccv_cnnp_model.c
index 9160aafb6..ee506e039 100644
--- a/lib/nnc/ccv_cnnp_model.c
+++ b/lib/nnc/ccv_cnnp_model.c
@@ -1108,10 +1108,10 @@ static void _ccv_cnnp_model_gradient_init(ccv_cnnp_model_t* const model, const i
 			const ccv_nnc_graph_exec_symbol_t* destinations = ccv_nnc_symbolic_graph_destinations(model->graph);
 			const int destination_count = ccv_nnc_symbolic_graph_destination_size(model->graph);
 			int flag = 0;
-			const int outgrad_destination_start = destination_count - i;
-			assert(outgrad_destination_start >= 0);
+			const int outgrad_destination_start = ccv_max(0, destination_count - i);
 			for (j = i - 1; !flag && j >= 0; j--)
-				flag = (destinations[j + outgrad_destination_start].d == outgrad.d);
+				if (j + outgrad_destination_start < destination_count)
+					flag = (destinations[j + outgrad_destination_start].d == outgrad.d);
 			if (!flag) // Only if we cannot find it, we add it.
 				ccv_nnc_symbolic_graph_add_destination(model->graph, outgrad);
 		}
diff --git a/lib/nnc/cmd/blas/mps/ccv_nnc_gemm_mps.m b/lib/nnc/cmd/blas/mps/ccv_nnc_gemm_mps.m
index 2735f1d74..120fa1023 100644
--- a/lib/nnc/cmd/blas/mps/ccv_nnc_gemm_mps.m
+++ b/lib/nnc/cmd/blas/mps/ccv_nnc_gemm_mps.m
@@ -704,7 +704,7 @@ static int _ccv_nnc_gemm_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 
 		ccv_nnc_mfa_context_t* context = ccv_nnc_default_mfa_context();
 		const int is_mfa_supported =
-			ccv_nnc_mfa_context_supported(context) && is_contiguous && is_same_dtype && is_supported_dtype && is_same_batch && !bias && !(ccv_nnc_flags() & CCV_NNC_DISABLE_METAL_FLASH_ATTENTION);
+			ccv_nnc_mfa_context_supported(context) && is_contiguous && is_same_dtype && is_supported_dtype && is_same_batch && !bias && !(ccv_nnc_flags() & CCV_NNC_DISABLE_METAL_FLASH_ATTENTION) && !(ccv_nnc_flags() & CCV_NNC_DISABLE_MFA_GEMM);
 
 		size_t a_data_size = 0;
 		if (a && dw && CCV_GET_DATA_TYPE(a->info.datatype) == CCV_QX)
diff --git a/lib/nnc/mfa/ccv_nnc_mfa_adam.cpp b/lib/nnc/mfa/ccv_nnc_mfa_adam.cpp
index 425057188..68af8fb3e 100644
--- a/lib/nnc/mfa/ccv_nnc_mfa_adam.cpp
+++ b/lib/nnc/mfa/ccv_nnc_mfa_adam.cpp
@@ -4,6 +4,8 @@
 using namespace ccv::nnc;
 
 #include <string>
+#include <sstream>
+#include <iomanip>
 
 // MARK: - C
 
@@ -137,6 +139,12 @@ std::size_t std::hash<mfa::adam::hash>::operator()(const mfa::adam::hash& hash)
   return seed;
 }
 
+static std::string high_precision_to_string(float value) {
+    std::ostringstream oss;
+    oss << std::setprecision(std::numeric_limits<float>::max_digits10) << value;
+    return oss.str();
+}
+
 mfa::adam::pipeline::pipeline(mfa::context* context, mfa::adam::hash hash) {
   // FlashNorm not supported for group adam yet.
   CCV_NNC_MFA_PRECONDITION((hash.data_type == MTL::DataTypeFloat) || (hash.data_type == MTL::DataTypeHalf))
@@ -309,24 +317,24 @@ kernel void adam(
   if (hash.adamw)
   {
     defines += "constant float rate_decay = ";
-    defines += std::to_string(hash.rate * hash.decay) + ";";
+    defines += high_precision_to_string(hash.rate * hash.decay) + ";";
     defines += "\n";
   } else {
     defines += "constant float decay = ";
-    defines += std::to_string(hash.decay) + ";";
+    defines += high_precision_to_string(hash.decay) + ";";
     defines += "\n";
   }
   defines += "constant float scale = ";
-  defines += std::to_string(hash.scale) + ";";
+  defines += high_precision_to_string(hash.scale) + ";";
   defines += "\n";
   defines += "constant float beta1 = ";
-  defines += std::to_string(hash.beta1) + ";";
+  defines += high_precision_to_string(hash.beta1) + ";";
   defines += "\n";
   defines += "constant float beta2 = ";
-  defines += std::to_string(hash.beta2) + ";";
+  defines += high_precision_to_string(hash.beta2) + ";";
   defines += "\n";
   defines += "constant float epsilon = ";
-  defines += std::to_string(hash.epsilon) + ";";
+  defines += high_precision_to_string(hash.epsilon) + ";";
   defines += "\n";
   this->group_size = MTL::Size(threadgroup_size, 1, 1);