feat: cuda transformers (#1401)

* Use cuda in transformers if available tensorflow probably needs a different check. Signed-off-by: Erich Schubert <kno10@users.noreply.github.com> * feat: expose CUDA at top level Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * tests: add to tests and create workflow for py extra backends * doc: update note on how to use core images --------- Signed-off-by: Erich Schubert <kno10@users.noreply.github.com> Signed-off-by: Ettore Di Giacinto <mudler@localai.io> Co-authored-by: Erich Schubert <kno10@users.noreply.github.com>
2024-12-18 20:27:57 +00:00 · 2023-12-08 15:45:04 +01:00 · 2023-12-08 15:45:04 +01:00 · 887b3dff04
commit 887b3dff04
parent 3822bd2369
9 changed files with 163 additions and 11 deletions
--- a/.github/workflows/test-extra.yml
+++ b/.github/workflows/test-extra.yml
@ -0,0 +1,75 @@
+---
+name: 'Tests extras backends'
+
+on:
+  pull_request:
+  push:
+    branches:
+      - master
+    tags:
+      - '*'
+
+concurrency:
+  group: ci-tests-extra-${{ github.head_ref || github.ref }}-${{ github.repository }}
+  cancel-in-progress: true
+
+jobs:
+  tests-linux:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Release space from worker
+        run: |
+          echo "Listing top largest packages"
+          pkgs=$(dpkg-query -Wf '${Installed-Size}\t${Package}\t${Status}\n' | awk '$NF == "installed"{print $1 "\t" $2}' | sort -nr)
+          head -n 30 <<< "${pkgs}"
+          echo
+          df -h
+          echo
+          sudo apt-get remove -y '^llvm-.*|^libllvm.*' || true
+          sudo apt-get remove --auto-remove android-sdk-platform-tools || true
+          sudo apt-get purge --auto-remove android-sdk-platform-tools || true
+          sudo rm -rf /usr/local/lib/android
+          sudo apt-get remove -y '^dotnet-.*|^aspnetcore-.*' || true
+          sudo rm -rf /usr/share/dotnet
+          sudo apt-get remove -y '^mono-.*' || true
+          sudo apt-get remove -y '^ghc-.*' || true
+          sudo apt-get remove -y '.*jdk.*|.*jre.*' || true
+          sudo apt-get remove -y 'php.*' || true
+          sudo apt-get remove -y hhvm powershell firefox monodoc-manual msbuild || true
+          sudo apt-get remove -y '^google-.*' || true
+          sudo apt-get remove -y azure-cli || true
+          sudo apt-get remove -y '^mongo.*-.*|^postgresql-.*|^mysql-.*|^mssql-.*' || true
+          sudo apt-get remove -y '^gfortran-.*' || true
+          sudo apt-get autoremove -y
+          sudo apt-get clean
+          echo
+          echo "Listing top largest packages"
+          pkgs=$(dpkg-query -Wf '${Installed-Size}\t${Package}\t${Status}\n' | awk '$NF == "installed"{print $1 "\t" $2}' | sort -nr)
+          head -n 30 <<< "${pkgs}"
+          echo
+          sudo rm -rfv build || true
+          df -h
+      - name: Clone
+        uses: actions/checkout@v4
+        with: 
+          submodules: true
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential ffmpeg
+          curl https://repo.anaconda.com/pkgs/misc/gpgkeys/anaconda.asc | gpg --dearmor > conda.gpg && \
+             sudo install -o root -g root -m 644 conda.gpg /usr/share/keyrings/conda-archive-keyring.gpg && \
+              gpg --keyring /usr/share/keyrings/conda-archive-keyring.gpg --no-default-keyring --fingerprint 34161F5BF5EB1D4BFBBB8F0A8AEB4F8B29D82806 && \
+             sudo /bin/bash -c 'echo "deb [arch=amd64 signed-by=/usr/share/keyrings/conda-archive-keyring.gpg] https://repo.anaconda.com/pkgs/misc/debrepo/conda stable main" > /etc/apt/sources.list.d/conda.list' && \
+             sudo /bin/bash -c 'echo "deb [arch=amd64 signed-by=/usr/share/keyrings/conda-archive-keyring.gpg] https://repo.anaconda.com/pkgs/misc/debrepo/conda stable main" | tee -a /etc/apt/sources.list.d/conda.list' && \
+             sudo apt-get update && \
+             sudo apt-get install -y conda
+          sudo apt-get install -y ca-certificates cmake curl patch
+          sudo apt-get install -y libopencv-dev && sudo ln -s /usr/include/opencv4/opencv2 /usr/include/opencv2
+          
+          sudo rm -rfv /usr/bin/conda || true
+
+      - name: Test
+        run: |
+           PATH=$PATH:/opt/conda/bin make test-extra
+ 
--- a/5
+++ b/5
@ -414,6 +414,11 @@ prepare-extra-conda-environments:
 	$(MAKE) -C backend/python/petals
 	$(MAKE) -C backend/python/exllama2

+prepare-test-extra:
+	$(MAKE) -C backend/python/transformers
+
+test-extra: prepare-test-extra
+	$(MAKE) -C backend/python/transformers test

 backend-assets/grpc:
 	mkdir -p backend-assets/grpc
--- a/api/backend/image.go
+++ b/api/backend/image.go
@ -16,7 +16,7 @@ func ImageGeneration(height, width, mode, step, seed int, positive_prompt, negat
 		model.WithContext(o.Context),
 		model.WithModel(c.Model),
 		model.WithLoadGRPCLoadModelOpts(&proto.ModelOptions{
-			CUDA:          c.Diffusers.CUDA,
+			CUDA:          c.CUDA,
 			SchedulerType: c.Diffusers.SchedulerType,
 			PipelineType:  c.Diffusers.PipelineType,
 			CFGScale:      c.Diffusers.CFGScale,
--- a/api/backend/options.go
+++ b/api/backend/options.go
@ -46,6 +46,7 @@ func gRPCModelOpts(c config.Config) *pb.ModelOptions {
 		Seed:           int32(c.Seed),
 		NBatch:         int32(b),
 		NoMulMatQ:      c.NoMulMatQ,
+		CUDA:           c.CUDA, // diffusers, transformers
 		DraftModel:     c.DraftModel,
 		AudioPath:      c.VallE.AudioPath,
 		Quantization:   c.Quantization,
--- a/api/config/config.go
+++ b/api/config/config.go
@ -46,6 +46,10 @@ type Config struct {

 	// Vall-e-x
 	VallE VallE `yaml:"vall-e"`
+
+	// CUDA
+	// Explicitly enable CUDA or not (some backends might need it)
+	CUDA bool `yaml:"cuda"`
 }

 type VallE struct {
@ -67,7 +71,6 @@ type GRPC struct {
 type Diffusers struct {
 	PipelineType     string  `yaml:"pipeline_type"`
 	SchedulerType    string  `yaml:"scheduler_type"`
-	CUDA             bool    `yaml:"cuda"`
 	EnableParameters string  `yaml:"enable_parameters"` // A list of comma separated parameters to specify
 	CFGScale         float32 `yaml:"cfg_scale"`         // Classifier-Free Guidance Scale
 	IMG2IMG          bool    `yaml:"img2img"`           // Image to Image Diffuser
--- a/backend/python/transformers/test_transformers_server.py
+++ b/backend/python/transformers/test_transformers_server.py
@ -31,7 +31,7 @@ class TestBackendServicer(unittest.TestCase):
        """
        This method tests if the server starts up successfully
        """
-        time.sleep(2)
+        time.sleep(10)
        try:
            self.setUp()
            with grpc.insecure_channel("localhost:50051") as channel:
@ -48,11 +48,12 @@ class TestBackendServicer(unittest.TestCase):
        """
        This method tests if the model is loaded successfully
        """
+        time.sleep(10)
        try:
            self.setUp()
            with grpc.insecure_channel("localhost:50051") as channel:
                stub = backend_pb2_grpc.BackendStub(channel)
-                response = stub.LoadModel(backend_pb2.ModelOptions(Model="bert-base-nli-mean-tokens"))
+                response = stub.LoadModel(backend_pb2.ModelOptions(Model="bert-base-cased"))
                self.assertTrue(response.success)
                self.assertEqual(response.message, "Model loaded successfully")
        except Exception as err:
@ -65,11 +66,13 @@ class TestBackendServicer(unittest.TestCase):
        """
        This method tests if the embeddings are generated successfully
        """
+        time.sleep(10)
        try:
            self.setUp()
            with grpc.insecure_channel("localhost:50051") as channel:
                stub = backend_pb2_grpc.BackendStub(channel)
-                response = stub.LoadModel(backend_pb2.ModelOptions(Model="bert-base-nli-mean-tokens"))
+                response = stub.LoadModel(backend_pb2.ModelOptions(Model="bert-base-cased"))
+                print(response.message)
                self.assertTrue(response.success)
                embedding_request = backend_pb2.PredictOptions(Embeddings="This is a test sentence.")
                embedding_response = stub.Embedding(embedding_request)
--- a/backend/python/transformers/transformers_server.py
+++ b/backend/python/transformers/transformers_server.py
@ -14,14 +14,27 @@ import backend_pb2
 import backend_pb2_grpc

 import grpc
+import torch

-from transformers import AutoModel
+from transformers import AutoTokenizer, AutoModel

 _ONE_DAY_IN_SECONDS = 60 * 60 * 24

 # If MAX_WORKERS are specified in the environment use it, otherwise default to 1
 MAX_WORKERS = int(os.environ.get('PYTHON_GRPC_MAX_WORKERS', '1'))

+
+def mean_pooling(model_output, attention_mask):
+    """
+    Mean pooling to get sentence embeddings. See:
+    https://huggingface.co/sentence-transformers/paraphrase-distilroberta-base-v1
+    """
+    token_embeddings = model_output[0]
+    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+    sum_embeddings = torch.sum(token_embeddings * input_mask_expanded, 1) # Sum columns
+    sum_mask = torch.clamp(input_mask_expanded.sum(1), min=1e-9)
+    return sum_embeddings / sum_mask
+
 # Implement the BackendServicer class with the service methods
 class BackendServicer(backend_pb2_grpc.BackendServicer):
    """
@ -56,9 +69,19 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
        model_name = request.Model
        try:
            self.model = AutoModel.from_pretrained(model_name, trust_remote_code=True) # trust_remote_code is needed to use the encode method with embeddings models like jinai-v2
+            self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+            if request.CUDA:
+                try:
+                    # TODO: also tensorflow, make configurable
+                    import torch.cuda
+                    if torch.cuda.is_available():
+                        print("Loading model", model_name, "to CUDA.", file=sys.stderr)
+                        self.model = self.model.to("cuda")
+                except Exception as err:
+                    print("Not using CUDA:", err, file=sys.stderr)
        except Exception as err:
            return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
-
        # Implement your logic here for the LoadModel service
        # Replace this with your desired response
        return backend_pb2.Result(message="Model loaded successfully", success=True)
@ -74,10 +97,20 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
        Returns:
            An EmbeddingResult object that contains the calculated embeddings.
        """
-        # Implement your logic here for the Embedding service
-        # Replace this with your desired response
+
+        # Tokenize input
+        max_length = 512
+        if request.Tokens != 0:
+            max_length = request.Tokens
+        encoded_input = self.tokenizer(request.Embeddings, padding=True, truncation=True, max_length=max_length, return_tensors="pt")    
+
+        # Create word embeddings
+        model_output = self.model(**encoded_input)
+
+        # Pool to get sentence embeddings; i.e. generate one 1024 vector for the entire sentence
+        sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask']).detach().numpy()
        print("Calculated embeddings for: " + request.Embeddings, file=sys.stderr)
-        sentence_embeddings = self.model.encode(request.Embeddings)
+        print("Embeddings:", sentence_embeddings, file=sys.stderr)
        return backend_pb2.EmbeddingResult(embeddings=sentence_embeddings)


--- a/docs/content/advanced/_index.en.md
+++ b/docs/content/advanced/_index.en.md
@ -207,6 +207,9 @@ lora_adapter: "/path/to/lora/adapter"
 lora_base: "/path/to/lora/base"
 # Disable mulmatq (CUDA)
 no_mulmatq: true
+
+# Diffusers/transformers
+cuda: true
 ```

 ### Prompt templates 
@ -363,4 +366,32 @@ You can control the backends that are built by setting the `GRPC_BACKENDS` envir
 make GRPC_BACKENDS=backend-assets/grpc/llama-cpp build
 ```

-By default, all the backends are built.
+By default, all the backends are built.
+
+### Extra backends
+
+LocalAI can be extended with extra backends. The backends are implemented as `gRPC` services and can be written in any language. The container images that are built and published on [quay.io](https://quay.io/repository/go-skynet/local-ai?tab=tags) contain a set of images split in core and extra. By default Images bring all the dependencies and backends supported by LocalAI (we call those `extra` images). The `-core` images instead bring only the strictly necessary dependencies to run LocalAI without only a core set of backends.
+
+If you wish to build a custom container image with extra backends, you can use the core images and build only the backends you are interested into. For instance, to use the diffusers backend:
+
+```Dockerfile
+FROM quay.io/go-skynet/local-ai:master-ffmpeg-core
+
+RUN PATH=$PATH:/opt/conda/bin make -C backend/python/diffusers
+```
+
+Remember also to set the `EXTERNAL_GRPC_BACKENDS` environment variable (or `--external-grpc-backends` as CLI flag) to point to the backends you are using (`EXTERNAL_GRPC_BACKENDS="backend_name:/path/to/backend"`), for example with diffusers:
+
+```Dockerfile
+FROM quay.io/go-skynet/local-ai:master-ffmpeg-core
+
+RUN PATH=$PATH:/opt/conda/bin make -C backend/python/diffusers
+
+ENV EXTERNAL_GRPC_BACKENDS="diffusers:/build/backend/python/diffusers/run.sh"
+```
+
+{{% notice note %}}
+
+You can specify remote external backends or path to local files. The syntax is `backend-name:/path/to/backend` or `backend-name:host:port`.
+
+{{% /notice %}}
--- a/docs/content/getting_started/_index.en.md
+++ b/docs/content/getting_started/_index.en.md
@ -178,6 +178,7 @@ You can control LocalAI with command line arguments, to specify a binding addres
 | --watchdog-busy-timeout value | $WATCHDOG_BUSY_TIMEOUT | 5m | Watchdog timeout. This will restart the backend if it crashes.  |
 | --watchdog-idle-timeout value | $WATCHDOG_IDLE_TIMEOUT | 15m | Watchdog idle timeout. This will restart the backend if it crashes. |
 | --preload-backend-only | $PRELOAD_BACKEND_ONLY | false | If set, the api is NOT launched, and only the preloaded models / backends are started. This is intended for multi-node setups. |
+| --external-grpc-backends | EXTERNAL_GRPC_BACKENDS | none | Comma separated list of external gRPC backends to use. Format: `name:host:port` or `name:/path/to/file` |

 ### Container images