use TabbyML/NeoX-70M for minimal e2e deployment (#10)

* use TabbyML/NeoX-70M for minimal e2e deployment

* Use python3 of triton image

.dockerignore

@@ -1,4 +1,3 @@
-testdata
 deployment
 
 **/.git

Makefile

@@ -1,6 +1,5 @@
 POETRY_EXISTS := $(shell which poetry &> /dev/null)
 PRE_COMMIT_HOOK := .git/hooks/pre-commit
-LOCAL_MODEL := testdata/tiny-70M/models/fastertransformer/1
 
 pre-commit:
 	poetry run pre-commit
@@ -14,12 +13,6 @@ endif
 $(PRE_COMMIT_HOOK):
 	poetry run pre-commit install --install-hooks
 
-$(LOCAL_MODEL):
-	poetry run python -m tabby.tools.huggingface_gptneox_convert \
-		-in_file EleutherAI/pythia-70m-deduped \
-		-o $@ \
-		-i_g 1 -m_n tiny-70M -p 1 -w fp16
-
 setup-development-environment: install-poetry $(PRE_COMMIT_HOOK)
 
 
@@ -29,11 +22,11 @@ DEV_FLAGS := $(UP_FLAGS) --build
 up:
 	docker-compose -f deployment/docker-compose.yml $(UP_FLAGS)
 
-up-triton: $(LOCAL_MODEL)
+up-triton:
 	docker-compose -f deployment/docker-compose.yml -f deployment/docker-compose.triton.yml $(UP_FLAGS)
 
 dev:
 	docker-compose -f deployment/docker-compose.yml -f deployment/docker-compose.dev.yml $(DEV_FLAGS)
 
-dev-triton: $(LOCAL_MODEL)
+dev-triton:
 	docker-compose -f deployment/docker-compose.yml -f deployment/docker-compose.triton.yml -f deployment/docker-compose.dev.yml $(DEV_FLAGS)

deployment/docker-compose.dev.yml

@@ -1,11 +1,13 @@
 version: '3.3'
 
 services:
-  server:
+  init:
     build:
       context: ..
       args:
         PYPI_INDEX_URL: https://pypi.tuna.tsinghua.edu.cn/simple
+
+  server:
     command: uvicorn tabby.server:app --host 0.0.0.0 --port 5000 --reload
     volumes:
       - ../:/app

deployment/docker-compose.triton.yml

@@ -5,23 +5,18 @@ services:
     image: tabbyml/tabby
     environment:
       - MODEL_BACKEND=triton
-      - TRITON_TOKENIZER_NAME=EleutherAI/pythia-70m-deduped
-    links:
-      - triton
-
-  admin:
-    links:
-      - triton
-
-
+    depends_on:
+      triton:
+        condition: service_healthy
 
   triton:
     image: tabbyml/fastertransformer_backend
     container_name: tabby-triton
-    command: mpirun -n 1 --allow-run-as-root /opt/tritonserver/bin/tritonserver --model-repository=/model
+    command: /scripts/triton.sh
     shm_size: 1gb
     volumes:
-      - ../testdata/tiny-70M/models:/model
+      - ./scripts:/scripts
+      - ./data/hf_cache:/root/.cache/huggingface
     deploy:
       resources:
         reservations:
@@ -29,3 +24,13 @@ services:
             - driver: nvidia
               count: all
               capabilities: [gpu]
+    environment:
+      - MODEL_NAME=TabbyML/NeoX-70M
+    depends_on:
+      init:
+        condition: service_completed_successfully
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8002/metrics"]
+      interval: 2s
+      timeout: 2s
+      start_period: 2s

deployment/docker-compose.yml

@@ -1,18 +1,28 @@
 version: '3.3'
 
 services:
+  init:
+    image: tabbyml/tabby
+    container_name: tabby-init
+    command: python -m tabby.tools.model_preload --repo_id TabbyML/NeoX-70M --backend python
+    volumes:
+      - ./data/hf_cache:/root/.cache/huggingface
+
   server:
     image: tabbyml/tabby
     container_name: tabby-server
     command: uvicorn tabby.server:app --host 0.0.0.0 --port 5000
     environment:
-      - PYTHON_MODEL_NAME=EleutherAI/pythia-70m-deduped
+      - MODEL_NAME=TabbyML/NeoX-70M
       - EVENTS_LOG_DIR=/logs
     ports:
       - "5000:5000"
     volumes:
       - ./data/logs/tabby-server:/logs
       - ./data/hf_cache:/root/.cache/huggingface
+    depends_on:
+      init:
+        condition: service_completed_successfully
 
   admin:
     image: tabbyml/tabby
@@ -20,9 +30,6 @@ services:
     command: streamlit run tabby/admin/Home.py
     ports:
       - "8501:8501"
-    links:
-      - server
-      - vector
 
   vector:
     image: timberio/vector:0.28.1-alpine

deployment/scripts/triton.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+set -e
+
+# Get model dir.
+MODEL_DIR=$(python3 <<EOF
+from huggingface_hub import snapshot_download
+
+print(snapshot_download(repo_id='$MODEL_NAME', allow_patterns='triton/**/*', local_files_only=True))
+EOF
+)
+
+# Set model dir in triton config.
+sed -i 's@${MODEL_DIR}@'$MODEL_DIR'@g' $MODEL_DIR/triton/fastertransformer/config.pbtxt
+
+# Start triton server.
+mpirun -n 1 \
+  --allow-run-as-root /opt/tritonserver/bin/tritonserver \
+  --model-repository=$MODEL_DIR/triton

tabby/server/__init__.py

@@ -16,16 +16,17 @@ app = FastAPI(
     docs_url="/",
 )
 
+MODEL_NAME = os.environ.get("MODEL_NAME")
 MODEL_BACKEND = os.environ.get("MODEL_BACKEND", "python")
 
 if MODEL_BACKEND == "triton":
     model_backend = TritonService(
-        tokenizer_name=os.environ.get("TRITON_TOKENIZER_NAME", None),
+        tokenizer_name=MODEL_NAME,
         host=os.environ.get("TRITON_HOST", "triton"),
         port=os.environ.get("TRITON_PORT", "8001"),
     )
 else:
-    model_backend = PythonModelService(os.environ["PYTHON_MODEL_NAME"])
+    model_backend = PythonModelService(MODEL_NAME)
 
 
 @app.post("/v1/completions")

tabby/server/python.py

@@ -13,8 +13,12 @@ class PythonModelService:
         self,
         model_name,
     ):
-        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
-        self.model = AutoModelForCausalLM.from_pretrained(model_name)
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            model_name, local_files_only=True
+        )
+        self.model = AutoModelForCausalLM.from_pretrained(
+            model_name, local_files_only=True
+        )
 
     def generate(self, request: CompletionRequest) -> List[Choice]:
         input_ids = self.tokenizer.encode(request.prompt, return_tensors="pt")

tabby/tools/model_preload.py

@@ -0,0 +1,43 @@
+from dataclasses import dataclass, field
+
+from huggingface_hub import snapshot_download
+from transformers import AutoModelForCausalLM, AutoTokenizer, HfArgumentParser
+
+
+@dataclass
+class Arguments:
+    repo_id: str = field(
+        metadata={"help": "Huggingface model repository id, e.g TabbyML/NeoX-160M"}
+    )
+    backend: str = "python"
+    prefer_local_files: bool = True
+
+
+def parse_args():
+    parser = HfArgumentParser(Arguments)
+    return parser.parse_args()
+
+
+def preload(local_files_only=False):
+    AutoTokenizer.from_pretrained(args.repo_id, local_files_only=local_files_only)
+    AutoModelForCausalLM.from_pretrained(
+        args.repo_id, local_files_only=local_files_only
+    )
+    snapshot_download(
+        repo_id=args.repo_id,
+        allow_patterns="triton/**/*",
+        local_files_only=local_files_only,
+    )
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    print(f"Loading {args.repo_id} ...")
+    try:
+        preload(local_files_only=args.prefer_local_files)
+    except Exception as e:
+        if "offline" in str(e):
+            preload(local_files_only=False)
+        else:
+            raise e
+    print(f"Loaded {args.repo_id} !")

testdata/tiny-70M/models/fastertransformer/.gitignore

@@ -1 +0,0 @@
-1

testdata/tiny-70M/models/fastertransformer/config.pbtxt

@@ -1,208 +0,0 @@
-name: "fastertransformer"
-backend: "fastertransformer"
-default_model_filename: "tiny-70M"
-max_batch_size: 1024
-
-model_transaction_policy {
-  decoupled: False
-}
-
-input [
-  {
-    name: "input_ids"
-    data_type: TYPE_UINT32
-    dims: [ -1 ]
-  },
-  {
-    name: "start_id"
-    data_type: TYPE_UINT32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "end_id"
-    data_type: TYPE_UINT32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "input_lengths"
-    data_type: TYPE_UINT32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-  },
-  {
-    name: "request_output_len"
-    data_type: TYPE_UINT32
-    dims: [ -1 ]
-  },
-  {
-    name: "runtime_top_k"
-    data_type: TYPE_UINT32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "runtime_top_p"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "beam_search_diversity_rate"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "temperature"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "len_penalty"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "repetition_penalty"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "random_seed"
-    data_type: TYPE_UINT64
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "is_return_log_probs"
-    data_type: TYPE_BOOL
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "beam_width"
-    data_type: TYPE_UINT32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "bad_words_list"
-    data_type: TYPE_INT32
-    dims: [ 2, -1 ]
-    optional: true
-  },
-  {
-    name: "stop_words_list"
-    data_type: TYPE_INT32
-    dims: [ 2, -1 ]
-    optional: true
-  },
-  {
-    name: "prompt_learning_task_name_ids"
-    data_type: TYPE_UINT32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "top_p_decay"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "top_p_min"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  },
-  {
-    name: "top_p_reset_ids"
-    data_type: TYPE_UINT32
-    dims: [ 1 ]
-    reshape: { shape: [ ] }
-    optional: true
-  }
-]
-output [
-  {
-    name: "output_ids"
-    data_type: TYPE_UINT32
-    dims: [ -1, -1 ]
-  },
-  {
-    name: "sequence_length"
-    data_type: TYPE_UINT32
-    dims: [ -1 ]
-  },
-  {
-    name: "cum_log_probs"
-    data_type: TYPE_FP32
-    dims: [ -1 ]
-  },
-  {
-    name: "output_log_probs"
-    data_type: TYPE_FP32
-    dims: [ -1, -1 ]
-  }
-]
-instance_group [
-  {
-    count: 1
-    kind: KIND_CPU
-  }
-]
-parameters {
-  key: "tensor_para_size"
-  value: {
-    string_value: "1"
-  }
-}
-parameters {
-  key: "pipeline_para_size"
-  value: {
-    string_value: "1"
-  }
-}
-parameters {
-  key: "data_type"
-  value: {
-    string_value: "fp32"
-  }
-}
-parameters {
-  key: "model_type"
-  value: {
-    string_value: "GPT-NeoX"
-  }
-}
-parameters {
-  key: "model_checkpoint_path"
-  value: {
-    string_value: "/model/fastertransformer/1/1-gpu"
-  }
-}
-parameters {
-  key: "enable_custom_all_reduce"
-  value: {
-    string_value: "0"
-  }
-}