feat: add stopword checker + iterable generate function

rwkv_pip_package/src/rwkv/utils.py

@@ -7,15 +7,38 @@
 import torch
 from torch.nn import functional as F
 
+
+def end_overlap(a, b):
+    for i in reversed(range(1, len(a) + 1)):
+        if b.startswith(a[-i:]):
+            return i
+    return 0
+
 class PIPELINE_ARGS():
-    def __init__(self, temperature=1.0, top_p=0.85, top_k=0, alpha_frequency=0.2, alpha_presence=0.2, token_ban=[], token_stop=[], chunk_len=256):
+    def __init__(self, 
+                 temperature=1.0, 
+                 top_p=0.85, 
+                 top_k=0, 
+                 alpha_frequency=0.2, 
+                 alpha_presence=0.2, 
+                 token_ban=None,
+                 token_stop=None,
+                 stop_words=None,
+                 chunk_len=256
+                ):
+
+        token_ban = token_ban or []
+        token_stop = token_stop or []
+        stop_words = stop_words or []
+
         self.temperature = temperature
         self.top_p = top_p
         self.top_k = top_k
         self.alpha_frequency = alpha_frequency # Frequency Penalty (as in GPT-3)
         self.alpha_presence = alpha_presence # Presence Penalty (as in GPT-3)
         self.token_ban = token_ban # ban the generation of some tokens
         self.token_stop = token_stop # stop generation whenever you see any token here
+        self.stop_words = stop_words # stop generation whenever you see any token here
         self.chunk_len = chunk_len # split input into chunks to save VRAM (shorter -> slower)
 
 class PIPELINE():
@@ -77,12 +100,23 @@ def sample_logits(self, logits, temperature=1.0, top_p=0.85, top_k=0):
                 probs = probs ** (1.0 / temperature)
             out = torch.multinomial(probs, num_samples=1)[0]
             return int(out)
-
-    def generate(self, ctx, token_count=100, args=PIPELINE_ARGS(), callback=None, state=None):
+
+    def generate(self, *args, callback=None, **kwargs):
+        outstr = []
+        for delta in self.igenerate(*args, **kwargs):
+            outstr += [delta]
+            if callback:
+                callback(delta)
+        return ''.join(outstr)
+
+    def igenerate(self, ctx, token_count=100, args=PIPELINE_ARGS(), state=None):
         all_tokens = []
         out_last = 0
         out_str = ''
         occurrence = {}
+
+        stopword_checker = self.check_stopwords(args.stop_words)
+        next(stopword_checker)
         for i in range(token_count):
 
             # forward & adjust prob.
@@ -108,9 +142,57 @@ def generate(self, ctx, token_count=100, args=PIPELINE_ARGS(), callback=None, st
 
             # output
             tmp = self.decode(all_tokens[out_last:])
+            if len(all_tokens)==1:
+                tmp = tmp[1:] # strip leading space
+            if tmp == '':
+                continue
             if '\ufffd' not in tmp: # is valid utf-8 string?
-                if callback:
-                    callback(tmp)
-                out_str += tmp
+
+                try:
+                    tmp = stopword_checker.send(tmp)
+                except StopIteration:
+                    break
                 out_last = i + 1
-        return out_str
+
+                if tmp is None:
+                    continue
+                yield tmp
+                out_str += tmp
+                out_last = i + 1                
+
+    @staticmethod
+    def check_stopwords(stop_words):
+
+        longest_stopword = 0 if len(stop_words)==0 else max(map(len, stop_words))
+        chunk = ""
+        delta = True
+        # yield
+        to_yield = None
+        while delta:
+            delta = yield to_yield
+            chunk = chunk + delta
+
+            if longest_stopword == 0:
+                # nothing to check just passthrough
+                to_yield = delta
+                continue
+            if chunk == '':
+                to_yield = None
+                continue
+            if any(map(lambda stop_word: chunk.startswith(stop_word), stop_words)):
+                return
+
+            if start_idx := max(map(lambda stop_word: end_overlap(chunk, stop_word), stop_words)):
+                if start_idx > longest_stopword:
+                    start_idx = longest_stopword  # can no longer be a stopword so cut it down
+                good, chunk = chunk[:-start_idx], chunk[-start_idx:]
+                if good:
+                    to_yield = good
+                    continue
+
+                to_yield = None
+                continue
+
+            out = chunk
+            chunk = ""
+            to_yield = out