feat: use model state cache to achieve 5x - 50x faster preparation time for generation

backend-python/dep_check.py

@@ -1,3 +1,4 @@
+import cyac
 import GPUtil
 import torch
 import rwkv

backend-python/main.py

@@ -11,7 +11,7 @@ import uvicorn
 from utils.rwkv import *
 from utils.torch import *
 from utils.ngrok import *
-from routes import completion, config
+from routes import completion, config, state_cache
 import global_var
 
 app = FastAPI()
@@ -26,11 +26,13 @@ app.add_middleware(
 
 app.include_router(completion.router)
 app.include_router(config.router)
+app.include_router(state_cache.router)
 
 
 @app.on_event("startup")
 def init():
     global_var.init()
+    state_cache.init()
 
     set_torch()
 

backend-python/routes/config.py

@@ -14,7 +14,7 @@ router = APIRouter()
 def get_tokens_path(model_path: str):
     model_path = model_path.lower()
     default_tokens_path = (
-        f"{pathlib.Path(__file__).parent.parent.resolve()}/20B_tokenizer.json"
+        f"{pathlib.Path(__file__).parent.parent.resolve()}/rwkv_pip/20B_tokenizer.json"
     )
     if "raven" in model_path:
         return default_tokens_path

backend-python/routes/state_cache.py

@@ -0,0 +1,98 @@
+from typing import Any, Dict
+from fastapi import APIRouter, HTTPException, Response, status
+from pydantic import BaseModel
+import gc
+import copy
+
+router = APIRouter()
+
+trie = None
+dtrie: Dict = {}
+
+
+def init():
+    global trie
+    try:
+        import cyac
+        import mmap
+        import os
+
+        if os.path.exists("state_cache.trie"):
+            with open("state_cache.trie", "r") as bf:
+                buff_object = mmap.mmap(bf.fileno(), 0, access=mmap.ACCESS_READ)
+            trie = cyac.Trie.from_buff(buff_object, copy=False)
+        else:
+            trie = cyac.Trie()
+    except ModuleNotFoundError:
+        print("cyac not found")
+
+
+class AddStateBody(BaseModel):
+    prompt: str
+    tokens: list[str]
+    state: Any
+    logits: Any
+
+
+@router.post("/add-state")
+def add_state(body: AddStateBody):
+    global trie, dtrie
+    if trie is None:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")
+
+    id = trie.insert(body.prompt)
+    dtrie[id] = {
+        "tokens": copy.deepcopy(body.tokens),
+        "state": copy.deepcopy(body.state),
+        "logits": copy.deepcopy(body.logits),
+    }
+
+    return "success"
+
+
+@router.post("/reset-state")
+def reset_state():
+    global trie
+    if trie is None:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")
+
+    trie = cyac.Trie()
+    gc.collect()
+
+    return "success"
+
+
+class LongestPrefixStateBody(BaseModel):
+    prompt: str
+
+
+@router.post("/longest-prefix-state")
+def longest_prefix_state(body: LongestPrefixStateBody):
+    global trie
+    if trie is None:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")
+
+    id = -1
+    for id, len in trie.prefix(body.prompt):
+        pass
+    if id != -1:
+        v = dtrie[id]
+        return {
+            "prompt": trie[id],
+            "tokens": v["tokens"],
+            "state": v["state"],
+            "logits": v["logits"],
+        }
+    else:
+        return {"prompt": "", "tokens": [], "state": None, "logits": None}
+
+
+@router.post("/save-state")
+def save_state():
+    global trie
+    if trie is None:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")
+
+    trie.save("state_cache.trie")
+
+    return "success"

backend-python/utils/rwkv.py

@@ -1,8 +1,11 @@
 import os
 import pathlib
+import copy
 from typing import Dict, List
+from fastapi import HTTPException
 from pydantic import BaseModel
 from rwkv_pip.utils import PIPELINE
+from routes import state_cache
 
 
 END_OF_TEXT = 0
@@ -61,9 +64,37 @@ class RWKV:
         return out
 
     def generate(self, prompt: str, stop: str = None):
-        self.model_state = None
-        self.model_tokens = []
-        logits = self.run_rnn(self.pipeline.encode(prompt))
+        cache = None
+        delta_prompt = prompt
+        try:
+            cache = state_cache.longest_prefix_state(
+                state_cache.LongestPrefixStateBody(prompt=prompt)
+            )
+        except HTTPException:
+            pass
+        if cache is None or cache["prompt"] == "":
+            self.model_state = None
+            self.model_tokens = []
+        else:
+            delta_prompt = prompt[len(cache["prompt"]) :]
+            self.model_state = copy.deepcopy(cache["state"])
+            self.model_tokens = copy.deepcopy(cache["tokens"])
+            logits = copy.deepcopy(cache["logits"])
+
+        if delta_prompt != "":
+            logits = self.run_rnn(self.pipeline.encode(delta_prompt))
+            try:
+                state_cache.add_state(
+                    state_cache.AddStateBody(
+                        prompt=prompt,
+                        tokens=self.model_tokens,
+                        state=self.model_state,
+                        logits=logits,
+                    )
+                )
+            except HTTPException:
+                pass
+
         begin = len(self.model_tokens)
         out_last = begin
 
@@ -94,9 +125,32 @@ class RWKV:
                 if stop is not None:
                     if stop in response:
                         response = response.split(stop)[0]
+                        try:
+                            state_cache.add_state(
+                                state_cache.AddStateBody(
+                                    prompt=prompt + response,
+                                    tokens=self.model_tokens,
+                                    state=self.model_state,
+                                    logits=logits,
+                                )
+                            )
+                        except HTTPException:
+                            pass
                         yield response, ""
                         break
                 out_last = begin + i + 1
+                if i == self.max_tokens_per_generation - 1:
+                    try:
+                        state_cache.add_state(
+                            state_cache.AddStateBody(
+                                prompt=prompt + response,
+                                tokens=self.model_tokens,
+                                state=self.model_state,
+                                logits=logits,
+                            )
+                        )
+                    except HTTPException:
+                        pass
                 yield response, delta