Cotran/onnx conversion (#145)

* onnx replacement

* onnx conversion for nli and embedding model

* fix naming

* fix naming

* fix naming

* pin version

arch/src/consts.rs

@@ -1,5 +1,5 @@
-pub const DEFAULT_EMBEDDING_MODEL: &str = "BAAI/bge-large-en-v1.5";
-pub const DEFAULT_INTENT_MODEL: &str = "tasksource/deberta-base-long-nli";
+pub const DEFAULT_EMBEDDING_MODEL: &str = "katanemo/bge-large-en-v1.5-onnx";
+pub const DEFAULT_INTENT_MODEL: &str = "katanemo/deberta-base-nli-onnx";
 pub const DEFAULT_PROMPT_TARGET_THRESHOLD: f64 = 0.8;
 pub const DEFAULT_HALLUCINATED_THRESHOLD: f64 = 0.1;
 pub const RATELIMIT_SELECTOR_HEADER_KEY: &str = "x-arch-ratelimit-selector";

model_server/Dockerfile

@@ -15,7 +15,7 @@ WORKDIR /src
 # specify list of models that will go into the image as a comma separated list
 # following models have been tested to work with this image
 # "sentence-transformers/all-MiniLM-L6-v2,sentence-transformers/all-mpnet-base-v2,thenlper/gte-base,thenlper/gte-large,thenlper/gte-small"
-ENV MODELS="BAAI/bge-large-en-v1.5"
+ENV MODELS="katanemo/bge-large-en-v1.5-onnx"
 
 COPY ./app ./app
 COPY ./app/guard_model_config.yaml .

model_server/Dockerfile.gpu

@@ -45,7 +45,7 @@ RUN if command -v nvcc >/dev/null 2>&1; then \
 COPY . /src
 
 # Specify list of models that will go into the image as a comma separated list
-ENV MODELS="BAAI/bge-large-en-v1.5"
+ENV MODELS="katanemo/bge-large-en-v1.5-onnx"
 ENV DEBIAN_FRONTEND=noninteractive
 
 COPY /app /app

model_server/app/guard_model_config.yaml

@@ -1,3 +1,3 @@
 jailbreak:
-  cpu: "katanemolabs/Arch-Guard-cpu"
-  gpu: "katanemolabs/Arch-Guard-gpu"
+  cpu: "katanemo/Arch-Guard-cpu"
+  gpu: "katanemo/Arch-Guard-gpu"

model_server/app/load_models.py

@@ -3,6 +3,7 @@
 from transformers import AutoTokenizer, pipeline
 import sqlite3
 import torch
+from optimum.onnxruntime import ORTModelForFeatureExtraction, ORTModelForSequenceClassification  # type: ignore
 
 
 def get_device():
@@ -16,13 +17,14 @@ def get_device():
     return device
 
 
-def load_transformers(models=os.getenv("MODELS", "BAAI/bge-large-en-v1.5")):
+def load_transformers(model_name=os.getenv("MODELS", "katanemo/bge-large-en-v1.5-onnx")):
     transformers = {}
     device = get_device()
-    for model in models.split(","):
-        transformers[model] = sentence_transformers.SentenceTransformer(
-            model, device=device
-        )
+    transformers["tokenizer"] = AutoTokenizer.from_pretrained(model_name)
+    transformers["model"] = ORTModelForFeatureExtraction.from_pretrained(
+        model_name, device_map = device
+    )
+    transformers["model_name"] = model_name
 
     return transformers
 
@@ -31,42 +33,51 @@ def load_guard_model(
     model_name,
     hardware_config="cpu",
 ):
-    guard_mode = {}
-    guard_mode["tokenizer"] = AutoTokenizer.from_pretrained(
+    guard_model = {}
+    guard_model["tokenizer"] = AutoTokenizer.from_pretrained(
         model_name, trust_remote_code=True
     )
-    guard_mode["model_name"] = model_name
+    guard_model["model_name"] = model_name
     if hardware_config == "cpu":
         from optimum.intel import OVModelForSequenceClassification
 
         device = "cpu"
-        guard_mode["model"] = OVModelForSequenceClassification.from_pretrained(
+        guard_model["model"] = OVModelForSequenceClassification.from_pretrained(
             model_name, device_map=device, low_cpu_mem_usage=True
         )
     elif hardware_config == "gpu":
         from transformers import AutoModelForSequenceClassification
         import torch
 
         device = "cuda" if torch.cuda.is_available() else "cpu"
-        guard_mode["model"] = AutoModelForSequenceClassification.from_pretrained(
+        guard_model["model"] = AutoModelForSequenceClassification.from_pretrained(
             model_name, device_map=device, low_cpu_mem_usage=True
         )
-    guard_mode["device"] = device
-    guard_mode["hardware_config"] = hardware_config
-    return guard_mode
+    guard_model["device"] = device
+    guard_model["hardware_config"] = hardware_config
+    return guard_model
 
 
 def load_zero_shot_models(
-    models=os.getenv("ZERO_SHOT_MODELS", "tasksource/deberta-base-long-nli")
+    model_name=os.getenv("ZERO_SHOT_MODELS", "katanemo/deberta-base-nli-onnx")
 ):
-    zero_shot_models = {}
+    zero_shot_model = {}
     device = get_device()
-    for model in models.split(","):
-        zero_shot_models[model] = pipeline(
-            "zero-shot-classification", model=model, device=device
-        )
+    zero_shot_model["model"] = ORTModelForSequenceClassification.from_pretrained(
+        model_name
+    )
+    zero_shot_model["tokenizer"] = AutoTokenizer.from_pretrained(model_name)
+
+    # create pipeline
+    zero_shot_model["pipeline"] = pipeline(
+        "zero-shot-classification",
+        model=zero_shot_model["model"],
+        tokenizer=zero_shot_model["tokenizer"],
+        device=device,
+    )
+    zero_shot_model["model_name"] = model_name
 
-    return zero_shot_models
+    return zero_shot_model
 
 
 if __name__ == "__main__":

model_server/app/main.py

@@ -1,4 +1,3 @@
-import os
 from fastapi import FastAPI, Response, HTTPException
 from pydantic import BaseModel
 from app.load_models import (
@@ -53,21 +52,25 @@ async def healthz():
 async def models():
     models = []
 
-    for model in transformers.keys():
-        models.append({"id": model, "object": "model"})
+    models.append({"id": transformers["model_name"], "object": "model"})
 
     return {"data": models, "object": "list"}
 
 
 @app.post("/embeddings")
 async def embedding(req: EmbeddingRequest, res: Response):
-
-    if req.model not in transformers:
+    if req.model != transformers["model_name"]:
         raise HTTPException(status_code=400, detail="unknown model: " + req.model)
 
     start = time.time()
-    embeddings = transformers[req.model].encode([req.input])
-    logger.info(f"Embedding Call Complete Time: {time.time()-start}")
+    encoded_input = transformers["tokenizer"](
+        req.input, padding=True, truncation=True, return_tensors="pt"
+    )
+    embeddings = transformers["model"](**encoded_input)
+    embeddings = embeddings[0][:, 0]
+    # normalize embeddings
+    embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1).detach().numpy()
+    print(f"Embedding Call Complete Time: {time.time()-start}")
     data = []
 
     for embedding in embeddings.tolist():
@@ -165,11 +168,13 @@ def remove_punctuations(s, lower=True):
 
 @app.post("/zeroshot")
 async def zeroshot(req: ZeroShotRequest, res: Response):
-    if req.model not in zero_shot_models:
+    logger.info(f"zero-shot request: {req}")
+    if req.model != zero_shot_models["model_name"]:
         raise HTTPException(status_code=400, detail="unknown model: " + req.model)
 
-    classifier = zero_shot_models[req.model]
+    classifier = zero_shot_models["pipeline"]
     labels_without_punctuations = [remove_punctuations(label) for label in req.labels]
+    start = time.time()
     predicted_classes = classifier(
         req.input, candidate_labels=labels_without_punctuations, multi_label=True
     )
@@ -178,6 +183,7 @@ async def zeroshot(req: ZeroShotRequest, res: Response):
     orig_map = [label_map[label] for label in predicted_classes["labels"]]
     final_scores = dict(zip(orig_map, predicted_classes["scores"]))
     predicted_class = label_map[predicted_classes["labels"][0]]
+    logger.info(f"zero-shot taking {time.time()-start} seconds")
 
     return {
         "predicted_class": predicted_class,
@@ -201,10 +207,11 @@ async def hallucination(req: HallucinationRequest, res: Response):
         example     {"name": "John", "age": "25"}
     prompt: input prompt from the user
     """
-    if req.model not in zero_shot_models:
+    if req.model != zero_shot_models["model_name"]:
         raise HTTPException(status_code=400, detail="unknown model: " + req.model)
 
-    classifier = zero_shot_models[req.model]
+    start = time.time()
+    classifier = zero_shot_models["pipeline"]
     candidate_labels = [f"{k} is {v}" for k, v in req.parameters.items()]
     hypothesis_template = "{}"
     result = classifier(
@@ -215,7 +222,9 @@ async def hallucination(req: HallucinationRequest, res: Response):
     )
     result_score = result["scores"]
     result_params = {k[0]: s for k, s in zip(req.parameters.items(), result_score)}
-    logger.info(f"hallucination result: {result_params}")
+    logger.info(
+        f"hallucination result: {result_params}, taking {time.time()-start} seconds"
+    )
 
     return {
         "params_scores": result_params,

model_server/requirements.txt

@@ -8,13 +8,12 @@ pyyaml==6.0.2
 accelerate
 psutil==6.0.0
 # guard inference packages
-optimum-intel
-openvino
+optimum-intel==1.19.0
+openvino==2024.4.0
 psutil
-pandas
 dateparser
 openai==1.50.2
 pandas
 tf-keras
-onnx
+onnx==1.17.0
 pytest