Commit: 710ae6e7774980d003893e18f50bdabc95a00f6f - zirpu (git)

Emergent generative agents

Commit MetaInfo

• Revision: 710ae6e7774980d003893e18f50bdabc95a00f6f (tree)
• Time: 2023-05-12 04:28:12
• Author: Corbin <cds@corb...>
• Commiter: Corbin

Log Message

Parameterize model path.

So that it can run on other peoples' machines!

Change Summary

• modified: src/agent.py
• modified: src/gens/camelid.py
• modified: src/gens/mawrkov.py

Incremental Difference

--- a/src/agent.py
+++ b/src/agent.py
@@ -17,9 +17,7 @@ from twisted.internet.threads import deferToThread
 from twisted.words.protocols.irc import IRCClient
 
 from common import irc_line, Timer, SentenceIndex, breakAt
-# from gens.camelid import CamelidGen
 from gens.mawrkov import MawrkovGen
-# from gens.trans import Flavor, HFGen, SentenceEmbed
 from gens.trans import SentenceEmbed
 
 build_traits = " + ".join
@@ -32,7 +30,8 @@ def load_character(path):
 
 MAX_NEW_TOKENS = 128
 print("~ Initializing mawrkov adapter…")
-gen = MawrkovGen(MAX_NEW_TOKENS)
+model_path = sys.argv[1]
+gen = MawrkovGen(model_path, MAX_NEW_TOKENS)
 # Need to protect per-gen data structures in C.
 genLock = Lock()
 GiB = 1024 ** 3
@@ -72,16 +71,17 @@ class Mind:
             # Newlines are added here.
             self.logits, self.state = gen.feedForward(gen.tokenize(s + "\n"),
                                                       self.logits, self.state)
-            print("~ Write:", s)
+            print(s)
 
     def complete(self, s):
         with genLock:
             completion, self.logits, self.state = gen.complete(s, self.logits, self.state)
-            print("~ Predicted:", completion)
+            print("«", completion, "»")
             return completion
 
     def infer(self, tag, prefix):
         def cb():
+            print(prefix)
             d = self.switchTag(tag)
             d.addCallback(lambda _: deferToThread(self.complete, prefix))
             return d
@@ -126,7 +126,6 @@ class ChainOfThoughts(Agent):
         thoughts = self.index.search(s, 2)
         for thought in thoughts:
             if thought not in self.recentThoughts:
-                print("~ New relevant thought:", thought)
                 self.recentThoughts.append(thought)
                 self.broadcast(thought)
 
@@ -178,7 +177,6 @@ class IRCAgent(Agent, IRCClient):
         user = user.split("!", 1)[0]
         self.broadcast(irc_line(datetime.now(), channel, user, line))
         if self.nickname in line:
-            print("~ Ping on IRC:", self.nickname)
             d = self.mind.infer("irc", self.prefix(channel))
 
             @d.addCallback
@@ -205,7 +203,7 @@ def go():
     clock = Clock()
     LoopingCall(clock.go).start(300.0, now=False)
 
-    for logpath in sys.argv[1:]:
+    for logpath in sys.argv[2:]:
         character = load_character(logpath)
         title = character["title"]
         firstStatement = f"I am {title}."

--- a/src/gens/camelid.py
+++ b/src/gens/camelid.py
@@ -4,17 +4,15 @@ from common import Timer
 
 from llama_cpp import Llama, llama_cpp
 
-MODEL = "/home/simpson/models/export/llama/30b-4bit.bin"
-MODEL_SIZE = os.stat(MODEL).st_size * 3 // 2
-
 class CamelidGen:
     model_name = "LLaMA?"
     model_arch = "LLaMA"
-    def __init__(self, max_new_tokens):
-        self.llama = Llama(MODEL, n_ctx=1024)
+    def __init__(self, model_path, max_new_tokens):
+        self.llama = Llama(model_path, n_ctx=1024)
+        self.model_size = os.stat(model_path).st_size * 3 // 2
         self.max_new_tokens = max_new_tokens
 
-    def footprint(self): return MODEL_SIZE
+    def footprint(self): return self.model_size
     def contextLength(self): return llama_cpp.llama_n_ctx(self.llama.ctx)
 
     # XXX doesn't work?
@@ -31,7 +29,8 @@ class CamelidGen:
 class CamelidEmbed:
     # XXX is that still the case?
     embedding_width = 4096
-    def __init__(self): self.llama = Llama(MODEL, embedding=True)
+    def __init__(self, model_path):
+        self.llama = Llama(model_path, embedding=True)
 
     def embed(self, s):
         with Timer("embedding"):

--- a/src/gens/mawrkov.py
+++ b/src/gens/mawrkov.py
@@ -4,8 +4,6 @@ import sys
 
 import tokenizers
 
-from common import Timer
-
 # Monkey-patch to get rwkv available.
 RWKV = "@RWKV@"
 RWKV_PATH = os.path.join(RWKV, "bin")
@@ -22,8 +20,6 @@ rwkv_cpp_shared_library = bare_import(RWKV_PATH, "rwkv_cpp_shared_library")
 rwkv_cpp_model = bare_import(RWKV_PATH, "rwkv_cpp_model")
 sampling = bare_import(RWKV_PATH, "sampling")
 
-MODEL = "/home/simpson/models/export/rwkv/rwkv-pile-14b-Q4_3.bin"
-MODEL_SIZE = os.stat(MODEL).st_size * 3 // 2
 TOKENIZER_PATH = os.path.join(RWKV, "share", "20B_tokenizer.json")
 
 # Upstream recommends temp 0.7, top_p 0.5
@@ -31,15 +27,18 @@ TEMPERATURE = 0.8
 TOP_P = 0.8
 
 class MawrkovGen:
+    # XXX might be wrong
     model_name = "The Pile (14B params, 4-bit quantized)"
     model_arch = "RWKV"
-    def __init__(self, max_new_tokens):
+    def __init__(self, model_path, max_new_tokens):
         self.max_new_tokens = max_new_tokens
+        self.model_size = os.stat(model_path).st_size * 3 // 2
         self.tokenizer = tokenizers.Tokenizer.from_file(TOKENIZER_PATH)
         self.lib = rwkv_cpp_shared_library.load_rwkv_shared_library()
-        self.model = rwkv_cpp_model.RWKVModel(self.lib, MODEL)
+        self.model = rwkv_cpp_model.RWKVModel(self.lib, model_path)
 
-    def footprint(self): return MODEL_SIZE
+    # XXX wrong
+    def footprint(self): return self.model_size
     def contextLength(self): return 8192
     def tokenize(self, s): return self.tokenizer.encode(s).ids
     def countTokens(self, s): return len(self.tokenize(s))
@@ -52,10 +51,9 @@ class MawrkovGen:
     def complete(self, s, logits, state):
         logits, state = self.feedForward(self.tokenize(s), logits, state)
         tokens = []
-        with Timer("completion"):
-            for i in range(self.max_new_tokens):
-                token = sampling.sample_logits(logits, TEMPERATURE, TOP_P)
-                tokens.append(token)
-                logits, state = self.feedForward([token], logits, state)
-                if "\n" in self.tokenizer.decode([token]): break
+        for i in range(self.max_new_tokens):
+            token = sampling.sample_logits(logits, TEMPERATURE, TOP_P)
+            tokens.append(token)
+            logits, state = self.feedForward([token], logits, state)
+            if "\n" in self.tokenizer.decode([token]): break
         return self.tokenizer.decode(tokens).split("\n", 1)[0], logits, state