diff --git a/.github/workflows/test-docker-gpu.yaml b/.github/workflows/test-docker-gpu.yaml
index b8f57473..985c03aa 100644
--- a/.github/workflows/test-docker-gpu.yaml
+++ b/.github/workflows/test-docker-gpu.yaml
@@ -15,7 +15,7 @@ jobs:
       matrix:
         backend:
           # - llama-cuda
-          # - exllama # https://github.com/sozercan/aikit/issues/94
+          - exllama # https://github.com/sozercan/aikit/issues/94
           - exllama2-gptq
           - exllama2-exl2
           - mamba
diff --git a/pkg/aikit2llb/inference/exllama.go b/pkg/aikit2llb/inference/exllama.go
index 5e20be9e..06b3b93f 100644
--- a/pkg/aikit2llb/inference/exllama.go
+++ b/pkg/aikit2llb/inference/exllama.go
@@ -21,13 +21,9 @@ func installExllama(c *config.InferenceConfig, s llb.State, merge llb.State) llb
 	savedState := s
 	s = s.Run(utils.Sh("apt-get update && apt-get install --no-install-recommends -y bash git ca-certificates python3-pip python3-dev python3-venv python-is-python3 make g++ curl && curl -LsSf https://astral.sh/uv/install.sh | sh && pip install grpcio-tools --break-system-packages && apt-get clean"), llb.IgnoreCache).Root()
 
-	// clone localai exllama backend only
 	s = cloneLocalAI(s)
 
-	s = s.Run(utils.Bashf("source $HOME/.cargo/env && export BUILD_TYPE=cublas && cd /tmp/localai/backend/python/%[1]s && make exllama2", backend)).Root()
-
-	// clone exllama to localai exllama backend path and install python dependencies
-	// s = s.Run(utils.Bashf("git clone --depth 1 %[1]s --branch %[2]s /tmp/%[3]s && mv /tmp/%[3]s/* /tmp/localai/backend/python/%[3]s && rm -rf /tmp/%[3]s && cd /tmp/localai/backend/python/%[3]s && rm -rf .git && source $HOME/.cargo/env && python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto && uv venv && source .venv/bin/activate && ls -al && uv pip install --no-build-isolation --requirement requirements-install.txt && EXLLAMA_NOCOMPILE= uv pip install --no-build-isolation --requirement requirements.txt", exllamaRepo, exllamaTag, backend)).Root()
+	s = s.Run(utils.Bashf("source $HOME/.cargo/env && export BUILD_TYPE=cublas && cd /tmp/localai/backend/python/%[1]s && make %[1]s", backend)).Root()
 
 	diff := llb.Diff(savedState, s)
 	return llb.Merge([]llb.State{merge, diff})
diff --git a/pkg/aikit2llb/inference/image.go b/pkg/aikit2llb/inference/image.go
index 43a923aa..49c716cd 100644
--- a/pkg/aikit2llb/inference/image.go
+++ b/pkg/aikit2llb/inference/image.go
@@ -42,6 +42,7 @@ func emptyImage(c *config.InferenceConfig) *specs.Image {
 		"NVIDIA_DRIVER_CAPABILITIES=compute,utility",
 		"NVIDIA_VISIBLE_DEVICES=all",
 		"LD_LIBRARY_PATH=/usr/local/cuda/lib64",
+		"BUILD_TYPE=cublas",
 	}
 	if c.Runtime == utils.RuntimeNVIDIA {
 		img.Config.Env = append(img.Config.Env, cudaEnv...)
@@ -51,14 +52,13 @@ func emptyImage(c *config.InferenceConfig) *specs.Image {
 		switch c.Backends[b] {
 		case utils.BackendExllama, utils.BackendExllamaV2:
 			exllamaEnv := []string{
-				"BUILD_TYPE=cublas",
-				"EXTERNAL_GRPC_BACKENDS=exllama:/tmp/localai/backend/python/exllama/backend.py,exllama2:/tmp/localai/backend/python/exllama2/run.sh",
+				"EXTERNAL_GRPC_BACKENDS=exllama:/tmp/localai/backend/python/exllama/run.sh,exllama2:/tmp/localai/backend/python/exllama2/run.sh",
 				"CUDA_HOME=/usr/local/cuda",
 			}
 			img.Config.Env = append(img.Config.Env, exllamaEnv...)
 		case utils.BackendMamba:
 			mambaEnv := []string{
-				"EXTERNAL_GRPC_BACKENDS=mamba:/tmp/localai/backend/python/mamba/backend.py",
+				"EXTERNAL_GRPC_BACKENDS=mamba:/tmp/localai/backend/python/mamba/run.sh",
 				"CUDA_HOME=/usr/local/cuda",
 			}
 			img.Config.Env = append(img.Config.Env, mambaEnv...)
diff --git a/pkg/aikit2llb/inference/mamba.go b/pkg/aikit2llb/inference/mamba.go
index 49a41e7f..c8522b6d 100644
--- a/pkg/aikit2llb/inference/mamba.go
+++ b/pkg/aikit2llb/inference/mamba.go
@@ -1,8 +1,6 @@
 package inference
 
 import (
-	"fmt"
-
 	"github.com/moby/buildkit/client/llb"
 	"github.com/sozercan/aikit/pkg/utils"
 )
@@ -10,39 +8,12 @@ import (
 func installMamba(s llb.State, merge llb.State) llb.State {
 	savedState := s
 	// libexpat1 is requirement but git is not. however libexpat1 is a dependency of git
-	s = s.Run(utils.Sh("apt-get install --no-install-recommends -y git python3 python3-dev python3-pip libssl3 openssl && apt-get clean"), llb.IgnoreCache).Root()
+	s = s.Run(utils.Sh("apt-get install --no-install-recommends -y git python3 python3-dev python3-pip python3-venv python-is-python3 libssl3 openssl && apt-get clean"), llb.IgnoreCache).Root()
 
 	s = cloneLocalAI(s)
 
-	s = s.Run(utils.Shf("pip3 install packaging numpy torch==2.1.0 grpcio protobuf --break-system-packages && pip3 install causal-conv1d==1.0.0 mamba-ssm==1.0.1 --break-system-packages")).Root()
+	s = s.Run(utils.Bashf("source $HOME/.cargo/env && export BUILD_TYPE=cublas && cd /tmp/localai/backend/python/%[1]s && make %[1]s", utils.BackendMamba)).Root()
 
 	diff := llb.Diff(savedState, s)
 	return llb.Merge([]llb.State{merge, diff})
 }
-
-func installOpenCV(s llb.State, merge llb.State) llb.State {
-	savedState := s
-	// adding debian 11 (bullseye) repo due to opencv 4.5 requirement
-	s = s.Run(utils.Sh("echo 'deb http://deb.debian.org/debian bullseye main' | tee -a /etc/apt/sources.list")).Root()
-	// pinning libdap packages to bullseye version due to symbol error
-	libdapVersion := "3.20.7-6"
-	libPath := "/usr/lib/x86_64-linux-gnu"
-	s = s.Run(utils.Shf("apt-get update && mkdir -p /tmp/generated/images && apt-get install -y libopencv-imgcodecs4.5 libgomp1 libdap27=%[1]s libdapclient6v5=%[1]s && apt-get clean && ln -s %[2]s/libopencv_core.so.4.5 %[2]s/libopencv_core.so.4.5d && ln -s %[2]s/libopencv_imgcodecs.so.4.5 %[2]s/libopencv_imgcodecs.so.4.5d", libdapVersion, libPath), llb.IgnoreCache).Root()
-	diff := llb.Diff(savedState, s)
-	merge = llb.Merge([]llb.State{merge, diff})
-
-	sdURL := fmt.Sprintf("https://sertaccdn.azureedge.net/localai/%s/stablediffusion", localAIVersion)
-	var opts []llb.HTTPOption
-	opts = append(opts, llb.Filename("stablediffusion"))
-	opts = append(opts, llb.Chmod(0o755))
-	var copyOpts []llb.CopyOption
-	copyOpts = append(copyOpts, &llb.CopyInfo{
-		CreateDestPath: true,
-	})
-	sd := llb.HTTP(sdURL, opts...)
-	merge = merge.File(
-		llb.Copy(sd, "stablediffusion", "/tmp/localai/backend_data/backend-assets/grpc/stablediffusion", copyOpts...),
-		llb.WithCustomName("Copying stable diffusion backend"), //nolint: goconst
-	)
-	return merge
-}
diff --git a/pkg/aikit2llb/inference/stablediffusion.go b/pkg/aikit2llb/inference/stablediffusion.go
new file mode 100644
index 00000000..5a40e537
--- /dev/null
+++ b/pkg/aikit2llb/inference/stablediffusion.go
@@ -0,0 +1,35 @@
+package inference
+
+import (
+	"fmt"
+
+	"github.com/moby/buildkit/client/llb"
+	"github.com/sozercan/aikit/pkg/utils"
+)
+
+func installOpenCV(s llb.State, merge llb.State) llb.State {
+	savedState := s
+	// adding debian 11 (bullseye) repo due to opencv 4.5 requirement
+	s = s.Run(utils.Sh("echo 'deb http://deb.debian.org/debian bullseye main' | tee -a /etc/apt/sources.list")).Root()
+	// pinning libdap packages to bullseye version due to symbol error
+	libdapVersion := "3.20.7-6"
+	libPath := "/usr/lib/x86_64-linux-gnu"
+	s = s.Run(utils.Shf("apt-get update && mkdir -p /tmp/generated/images && apt-get install -y libopencv-imgcodecs4.5 libgomp1 libdap27=%[1]s libdapclient6v5=%[1]s && apt-get clean && ln -s %[2]s/libopencv_core.so.4.5 %[2]s/libopencv_core.so.4.5d && ln -s %[2]s/libopencv_imgcodecs.so.4.5 %[2]s/libopencv_imgcodecs.so.4.5d", libdapVersion, libPath), llb.IgnoreCache).Root()
+	diff := llb.Diff(savedState, s)
+	merge = llb.Merge([]llb.State{merge, diff})
+
+	sdURL := fmt.Sprintf("https://sertaccdn.azureedge.net/localai/%s/stablediffusion", localAIVersion)
+	var opts []llb.HTTPOption
+	opts = append(opts, llb.Filename("stablediffusion"))
+	opts = append(opts, llb.Chmod(0o755))
+	var copyOpts []llb.CopyOption
+	copyOpts = append(copyOpts, &llb.CopyInfo{
+		CreateDestPath: true,
+	})
+	sd := llb.HTTP(sdURL, opts...)
+	merge = merge.File(
+		llb.Copy(sd, "stablediffusion", "/tmp/localai/backend_data/backend-assets/grpc/stablediffusion", copyOpts...),
+		llb.WithCustomName("Copying stable diffusion backend"), //nolint: goconst
+	)
+	return merge
+}