Merge pull request #38 from modal-labs/erikbern/inference-web-endpoint

Add web endpoint to the inference server

README.md

@@ -45,7 +45,7 @@ modal run --detach src.train --config=config/mistral.yml --data=data/sqlqa.jsonl
 4. Try the model from a completed training run. You can select a folder via `modal volume ls example-runs-vol`, and then specify the training folder with the `--run-folder` flag (something like `/runs/axo-2023-11-24-17-26-66e8`) for inference:
 
 ```bash
-modal run -q src.inference --run-folder /runs/<run_tag>
+modal run -q src.inference --run-name <run_tag>
 ```
 
 Our quickstart example trains a 7B model on a text-to-SQL dataset as a proof of concept (it takes just a few minutes). It uses DeepSpeed ZeRO-3 to shard the model state across 2 A100s. Inference on the fine-tuned model displays conformity to the output structure (`[SQL] ... [/SQL]`). To achieve better results, you would need to use more data! Refer to the full development section below.

ci/check_inference.py

@@ -10,7 +10,7 @@
 CREATE TABLE head (age INTEGER)
 How many heads of the departments are older than 56 ? [/INST] """
 
-    p = subprocess.Popen(["modal", "run", "src.inference", "--run-folder", f"/runs/{run_name}", "--prompt", prompt], stdout=subprocess.PIPE)
+    p = subprocess.Popen(["modal", "run", "src.inference", "--run-name", run_name, "--prompt", prompt], stdout=subprocess.PIPE)
     output = ""
 
     for line in iter(p.stdout.readline, b''):

src/inference.py

@@ -1,12 +1,20 @@
 import time
 from pathlib import Path
+import yaml
 
 import modal
+from fastapi.responses import StreamingResponse
 
 from .common import stub, vllm_image, VOLUME_CONFIG
 
 N_INFERENCE_GPU = 2
 
+with vllm_image.imports():
+    from vllm.engine.arg_utils import AsyncEngineArgs
+    from vllm.engine.async_llm_engine import AsyncLLMEngine
+    from vllm.sampling_params import SamplingParams
+    from vllm.utils import random_uuid
+
 
 @stub.cls(
     gpu=modal.gpu.H100(count=N_INFERENCE_GPU),
@@ -16,16 +24,24 @@
     container_idle_timeout=120,
 )
 class Inference:
-    def __init__(self, run_folder: str) -> None:
-        import yaml
-
-        with open(f"{run_folder}/config.yml") as f:
-            config = yaml.safe_load(f.read())
-        model_path = (Path(run_folder) / config["output_dir"] / "merged").resolve()
-
+    def __init__(self, run_name: str = "", run_dir: str = "/runs") -> None:
+        self.run_name = run_name
+        self.run_dir = run_dir
+
+    @modal.enter()
+    def init(self):
+        if self.run_name:
+            run_name = self.run_name
+        else:
+            # Pick the last run automatically
+            run_name = VOLUME_CONFIG[self.run_dir].listdir("/")[-1].path
+
+        # Grab the output dir (usually "lora-out")
+        with open(f"{self.run_dir}/{run_name}/config.yml") as f:
+            output_dir = yaml.safe_load(f.read())["output_dir"]
+
+        model_path = f"{self.run_dir}/{run_name}/{output_dir}/merged"
         print("Initializing vLLM engine on:", model_path)
-        from vllm.engine.arg_utils import AsyncEngineArgs
-        from vllm.engine.async_llm_engine import AsyncLLMEngine
 
         engine_args = AsyncEngineArgs(
             model=model_path,
@@ -34,14 +50,10 @@ def __init__(self, run_folder: str) -> None:
         )
         self.engine = AsyncLLMEngine.from_engine_args(engine_args)
 
-    @modal.method()
-    async def completion(self, input: str):
+    async def _stream(self, input: str):
         if not input:
             return
 
-        from vllm.sampling_params import SamplingParams
-        from vllm.utils import random_uuid
-
         sampling_params = SamplingParams(
             repetition_penalty=1.1,
             temperature=0.2,
@@ -71,16 +83,25 @@ async def completion(self, input: str):
         print(f"Request completed: {throughput:.4f} tokens/s")
         print(request_output.outputs[0].text)
 
+    @modal.method()
+    async def completion(self, input: str):
+        async for text in self._stream(input):
+            yield text
+
+    @modal.web_endpoint()
+    async def web(self, input: str):
+        return StreamingResponse(self._stream(input), media_type="text/event-stream")
+
 
 @stub.local_entrypoint()
-def inference_main(run_folder: str, prompt: str = ""):
+def inference_main(run_name: str = "", prompt: str = ""):
     if prompt:
-        for chunk in Inference(run_folder).completion.remote_gen(prompt):
+        for chunk in Inference(run_name).completion.remote_gen(prompt):
             print(chunk, end="")
     else:
         prompt = input(
             "Enter a prompt (including the prompt template, e.g. [INST] ... [/INST]):\n"
         )
         print("Loading model ...")
-        for chunk in Inference(run_folder).completion.remote_gen(prompt):
+        for chunk in Inference(run_name).completion.remote_gen(prompt):
             print(chunk, end="")

src/train.py

@@ -150,4 +150,4 @@ def main(
 
     print(f"Training complete. Run tag: {run_name}")
     print(f"To inspect weights, run `modal volume ls example-runs-vol {run_name}`")
-    print(f"To run sample inference, run `modal run -q src.inference --run-folder /runs/{run_name}`")
+    print(f"To run sample inference, run `modal run -q src.inference --run-name {run_name}`")