diff --git a/include/ircd/gpt/model.h b/include/ircd/gpt/model.h
index cbbd72cec..e28daa538 100644
--- a/include/ircd/gpt/model.h
+++ b/include/ircd/gpt/model.h
@@ -30,6 +30,7 @@ namespace ircd::gpt::model
 	extern std::vector<json::object> default_data;
 
 	constexpr auto alignment {4096};
+	extern conf::item<bool> cache_shared;
 }
 
 /// Layer normalization
diff --git a/ircd/gpt.cc b/ircd/gpt.cc
index b71317047..49e91c9ee 100644
--- a/ircd/gpt.cc
+++ b/ircd/gpt.cc
@@ -256,10 +256,15 @@ try
 }
 ,model
 {
-	std::make_unique<pipe::model>
-	(
-		*const_cast<const gpt::model::decoder *>(gpt::model::default_model)
-	)
+	!gpt::model::cache_shared?
+		std::make_unique<pipe::model>
+		(
+			*const_cast<const gpt::model::decoder *>(gpt::model::default_model)
+		):
+		std::make_unique<pipe::model>
+		(
+			*const_cast<gpt::model::decoder *>(gpt::model::default_model)
+		)
 }
 ,desc
 {
diff --git a/ircd/gpt_gpu.cl b/ircd/gpt_gpu.cl
index 7290cdf20..d38b82e50 100644
--- a/ircd/gpt_gpu.cl
+++ b/ircd/gpt_gpu.cl
@@ -1167,7 +1167,7 @@ ircd_gpt_prop_elem(__global const struct ircd_gpt_ctrl *const ctrl,
 
 	const float4
 	param = param_[li],
-	grad = ctrl->label[0].loss.mean,
+	grad = ctrl->target.loss.mean,
 	alpha[2] = { 1.0f - opts->beta[0], 1.0f - opts->beta[1], },
 	exp_avg = ts? exp_avg_[li]: 0.0f,
 	exp_avg_sqr = ts? exp_avg_sqr_[li]: 0.0f,
@@ -1179,13 +1179,9 @@ ircd_gpt_prop_elem(__global const struct ircd_gpt_ctrl *const ctrl,
 	delta = opts->alpha * (exp_avg_dot / denom),
 	update = param - delta;
 
-	param_[li] = param + FLT_EPSILON;
-	exp_avg_[li] = exp_avg + FLT_EPSILON;
-	exp_avg_sqr_[li] = exp_avg_sqr + FLT_EPSILON;
-
-	//param_[li] = update;
-	//exp_avg_[li] = exp_avg_dot;
-	//exp_avg_sqr_[li] = exp_avg_sqr_dot;
+	param_[li] = update;
+	exp_avg_[li] = exp_avg_dot;
+	exp_avg_sqr_[li] = exp_avg_sqr_dot;
 }
 
 //
diff --git a/ircd/gpt_pipe.cc b/ircd/gpt_pipe.cc
index a5c01ebdb..3791974e6 100644
--- a/ircd/gpt_pipe.cc
+++ b/ircd/gpt_pipe.cc
@@ -238,7 +238,7 @@ ircd::gpt::pipe::cycle::cycle(gpt::samp &samp)
 	tokens,
 	cached,
 	true,
-	false,
+	((false) && gpt::model::cache_shared)
 }
 ,stage
 {