Commit: bb8a464e79a4ccd7e0ef517bbdd226ac6446c5d5 - zirpu (git)

Emergent generative agents

Commit MetaInfo

• Revision: bb8a464e79a4ccd7e0ef517bbdd226ac6446c5d5 (tree)
• Zeit: 2023-06-12 03:14:32
• Autor: Corbin <cds@corb...>
• Commiter: Corbin

Log Message

Bump RWKV and LLaMA dependencies.

Ändern Zusammenfassung

• modified: flake.lock
• modified: flake.nix
• modified: src/agent.py
• modified: src/gens/camelid.py
• modified: src/gens/mawrkov.py

Diff

--- a/flake.lock
+++ b/flake.lock
@@ -5,11 +5,11 @@
         "systems": "systems"
       },
       "locked": {
-        "lastModified": 1681202837,
-        "narHash": "sha256-H+Rh19JDwRtpVPAWp64F+rlEtxUWBAQW28eAi3SRSzg=",
+        "lastModified": 1685518550,
+        "narHash": "sha256-o2d0KcvaXzTrPRIo0kOLV0/QXHhDQ5DTi+OxcjO8xqY=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "cfacdce06f30d2b68473a46042957675eebb3401",
+        "rev": "a1720a10a6cfe8234c0e93907ffe81be440f4cef",
         "type": "github"
       },
       "original": {
@@ -39,11 +39,11 @@
         "nixpkgs": "nixpkgs"
       },
       "locked": {
-        "lastModified": 1682476640,
-        "narHash": "sha256-mLVO3T86AaXg3CXJNEjxjAaUGRhB0+8J1yGRV9LN/8M=",
+        "lastModified": 1685666615,
+        "narHash": "sha256-+hL7AHJja9EyGbwKOIB4yK/6QsMkSgyi/Dv0YpWl9wA=",
         "owner": "MostAwesomeDude",
         "repo": "llama.cpp",
-        "rev": "8a2a3e5098187a70a3949aa8a9351f2f26478d84",
+        "rev": "a48b63971f6e4229c00cf23da4cc09881a8f6874",
         "type": "github"
       },
       "original": {
@@ -70,11 +70,11 @@
     },
     "nixpkgs_2": {
       "locked": {
-        "lastModified": 1682453498,
-        "narHash": "sha256-WoWiAd7KZt5Eh6n+qojcivaVpnXKqBsVgpixpV2L9CE=",
+        "lastModified": 1685564631,
+        "narHash": "sha256-8ywr3AkblY4++3lIVxmrWZFzac7+f32ZEhH/A8pNscI=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "c8018361fa1d1650ee8d4b96294783cf564e8a7f",
+        "rev": "4f53efe34b3a8877ac923b9350c874e3dcd5dc0a",
         "type": "github"
       },
       "original": {

--- a/flake.nix
+++ b/flake.nix
@@ -13,13 +13,13 @@
         llama-lib = llama-cpp-lib.packages.${system}.default;
         llama-cpp-python = pkgs.python310.pkgs.buildPythonPackage rec {
           pname = "llama-cpp-python";
-          version = "0.1.38";
+          version = "0.1.57";
 
           src = pkgs.fetchFromGitHub {
             owner = "abetlen";
             repo = pname;
             rev = "v${version}";
-            sha256 = "sha256-/Ykndsp6puFxa+FSHNln9M2frS7/sMMBJSNJ/mU/CSI=";
+            sha256 = "sha256-BrR3N+3KRu96j0MIydyrvFb2BN3COeBPISac+ixq3XM=";
           };
           format = "setuptools";
 
@@ -28,8 +28,13 @@
             sed -i -e "s,_load_shared_library(_lib_base_name),ctypes.CDLL('${llama-lib}/lib/libllama.so')," llama_cpp/llama_cpp.py
           '';
 
+          # Imports server.app, which needs fancy Starlette packages I'm not
+          # willing to deal with right now. ~ C.
+          doCheck = false;
+
           propagatedBuildInputs = with pkgs.python310.pkgs; [
-            typing-extensions
+            numpy typing-extensions
+            # anyio fastapi uvicorn
           ];
         };
         sentence-transformers = pkgs.python310.pkgs.buildPythonPackage rec {
@@ -57,12 +62,9 @@
 
           src = pkgs.fetchFromGitHub {
             owner = "saharNooby";
-            # owner = "iacore";
             repo = "rwkv.cpp";
-            rev = "c736ef5411606b529d3a74c139ee111ef1a28bb9";
-            sha256 = "sha256-zJFmuhyY2kT/WVStBpHSnlmwclXZmVoiFvsurCDHW4E=";
-            # rev = "ae390c6";
-            # sha256 = "sha256-ojDsZgXwd3+E6AGtB/KANGz3Y0W5l9CWGjfhjJEefDQ=";
+            rev = "363dfb1a061507aee661300fc8e2e153b6e99dc2";
+            sha256 = "sha256-HlJmXMXSUNgPJN6TSGnNeeBeY3/9HmRH9Qa2d4jPEu4=";
             fetchSubmodules = true;
           };
 
@@ -119,7 +121,7 @@
         devShells.default = pkgs.mkShell {
           name = "zirpu-env";
           packages = with pkgs; [
-            git
+            git gdb
             # our Python
             py
             # catching Python mistakes

--- a/src/agent.py
+++ b/src/agent.py
@@ -17,7 +17,8 @@ from twisted.internet.threads import deferToThread
 from twisted.words.protocols.irc import IRCClient
 
 from common import irc_line, Timer, SentenceIndex, breakAt
-from gens.mawrkov import MawrkovGen, force
+from gens.camelid import CamelidGen
+from gens.mawrkov import MawrkovGen
 from gens.trans import SentenceEmbed
 
 build_traits = " + ".join
@@ -29,9 +30,16 @@ def load_character(path):
         return json.load(handle)
 
 MAX_NEW_TOKENS = 128
-print("~ Initializing mawrkov adapter…")
-model_path = sys.argv[1]
-gen = MawrkovGen(model_path, MAX_NEW_TOKENS)
+gens = {
+    "llama": CamelidGen,
+    "rwkv": MawrkovGen,
+}
+model_cls = sys.argv[1]
+if model_cls not in gens:
+    raise ValueError("must be one of %r" % tuple(gens.keys()))
+print("~ Initializing adapter:", model_cls)
+model_path = sys.argv[2]
+gen = gens[model_cls](model_path, MAX_NEW_TOKENS)
 # Need to protect per-gen data structures in C.
 genLock = Lock()
 GiB = 1024 ** 3
@@ -50,7 +58,6 @@ prologues = {
 
 class Mind:
     currentTag = None
-    logits = state = None
 
     def __init__(self, yarn, name):
         self.yarn = yarn
@@ -103,7 +110,7 @@ class Mind:
             # Breaking the newline invariant...
             d.addCallback(lambda _: deferToThread(self.writeRaw, prefix))
             # ...so that this inference happens before the newline...
-            d.addCallback(lambda _: force(tokens, self.logits))
+            d.addCallback(lambda _: self.yarn.force(tokens))
             # XXX decode should be on gen ABC
             d.addCallback(lambda t: gen.tokenizer.decode([t]))
 
@@ -279,7 +286,7 @@ def go():
     clock = Clock()
     LoopingCall(clock.go).start(60 * 30, now=True)
 
-    for logpath in sys.argv[2:]:
+    for logpath in sys.argv[3:]:
         character = load_character(logpath)
         title = character["title"]
         firstStatement = f"I am {title}."

--- a/src/gens/camelid.py
+++ b/src/gens/camelid.py
@@ -8,41 +8,59 @@ class CamelidGen(Gen):
     model_name = "LLaMA?"
     model_arch = "LLaMA"
     def __init__(self, model_path, max_new_tokens):
-        self.llama = Llama(model_path, n_ctx=1024)
+        self.llama = Llama(model_path)
         self.model_size = os.stat(model_path).st_size * 3 // 2
         self.max_new_tokens = max_new_tokens
 
     def footprint(self): return self.model_size
     def contextLength(self): return llama_cpp.llama_n_ctx(self.llama.ctx)
 
-    # XXX doesn't work?
     def tokenize(self, s): return self.llama.tokenize(s.encode("utf-8"))
-    def decode(self, ts): return self.llama.detokenize(ts)
+    def decode(self, ts): return self.llama.detokenize(ts).decode("utf-8")
 
     def fork(self):
         return CamelidYarn(self.max_new_tokens, self.llama, self.llama.save_state())
 
+yarn_cache = [None]
+
 class CamelidYarn(Yarn):
     def __init__(self, max_new_tokens, llama, state):
         self.max_new_tokens = max_new_tokens
         self.llama = llama
         self.state = state
 
+    def activate(self):
+        if yarn_cache[0] is not self:
+            if yarn_cache[0]: yarn_cache[0].deactivate()
+            yarn_cache[0] = self
+            self.llama.load_state(self.state)
+
+    def deactivate(self): self.state = self.llama.save_state()
+
     def feedForward(self, tokens):
-        self.llama.load_state(self.state)
+        self.activate()
         self.llama.eval(tokens)
-        self.state = self.llama.save_state()
 
     def complete(self):
-        self.llama.load_state(self.state)
+        self.activate()
         tokens = []
         for _ in range(self.max_new_tokens):
             token = self.llama.sample()
-            if "\n" in self.llama.detokenize([token]): break
+            if b"\n" in self.llama.detokenize([token]): break
             tokens.append(token)
-        self.state = self.llama.save_state()
         return tokens
 
+    def force(self, options):
+        self.activate()
+        return self.llama.sample(logits_processor=Force(options))
+
+class Force:
+    def __init__(self, options): self.options = options
+    def __call__(self, input_ids, scores):
+        # +10 is hopefully not too much.
+        for option in self.options: scores[option] += 10
+        return scores
+
 class CamelidEmbed:
     def __init__(self, model_path):
         self.llama = Llama(model_path, embedding=True)

--- a/src/gens/mawrkov.py
+++ b/src/gens/mawrkov.py
@@ -25,8 +25,10 @@ sampling = bare_import(RWKV_PATH, "sampling")
 TOKENIZER_PATH = os.path.join(RWKV, "share", "20B_tokenizer.json")
 
 # Upstream recommends temp 0.7, top_p 0.5
-TEMPERATURE = 0.8
-TOP_P = 0.8
+# TEMPERATURE = 0.8
+# TOP_P = 0.8
+TEMPERATURE = 0.9
+TOP_P = 0.9
 
 class MawrkovGen(Gen):
     model_name = "The Pile"
@@ -68,7 +70,7 @@ class MawrkovYarn(Yarn):
             self.feedForward([token])
         return tokens
 
-def force(options, logits):
-    # +10 is hopefully not too much.
-    biases = {opt: 10 for opt in options}
-    return sampling.sample_logits(logits, TEMPERATURE, TOP_P, biases)
+    def force(self, options):
+        # +10 is hopefully not too much.
+        biases = {opt: 10 for opt in options}
+        return sampling.sample_logits(self.logits, TEMPERATURE, TOP_P, biases)