Merge pull request #38 from VE-FORBRYDERNE/warp

Move TFS warper code into aiserver.py

aiserver.py

@@ -568,6 +568,83 @@ if(not vars.model in ["InferKit", "Colab", "OAI", "ReadOnly", "TPUMeshTransforme
         except:
             pass
 
+
+        # Patch transformers to use our custom logit warpers
+        from transformers import LogitsProcessorList, LogitsWarper, TopKLogitsWarper, TopPLogitsWarper, TemperatureLogitsWarper
+        class TailFreeLogitsWarper(LogitsWarper):
+
+            def __init__(self, tfs: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+                tfs = float(tfs)
+                if tfs < 0 or tfs > 1.0:
+                    raise ValueError(f"`tfs` has to be a float > 0 and < 1, but is {tfs}")
+                self.tfs = tfs
+                self.filter_value = filter_value
+                self.min_tokens_to_keep = min_tokens_to_keep
+
+            def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+                if self.filter_value >= 1.0:
+                    return scores
+                sorted_logits, sorted_indices = torch.sort(scores, descending=True)
+                probs = sorted_logits.softmax(dim=-1)
+
+                # Compute second derivative normalized CDF
+                d2 = probs.diff().diff().abs()
+                normalized_d2 = d2 / d2.sum(dim=-1, keepdim=True)
+                normalized_d2_cdf = normalized_d2.cumsum(dim=-1)
+
+                # Remove tokens with CDF value above the threshold (token with 0 are kept)
+                sorted_indices_to_remove = normalized_d2_cdf > self.tfs
+
+                # Centre the distribution around the cutoff as in the original implementation of the algorithm
+                sorted_indices_to_remove = torch.cat(
+                    (
+                        torch.zeros(scores.shape[0], 1, dtype=torch.bool, device=scores.device),
+                        sorted_indices_to_remove,
+                        torch.ones(scores.shape[0], 1, dtype=torch.bool, device=scores.device),
+                    ),
+                    dim=-1,
+                )
+
+                if self.min_tokens_to_keep > 1:
+                    # Keep at least min_tokens_to_keep
+                    sorted_indices_to_remove[..., : self.min_tokens_to_keep] = 0
+
+                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+                scores = scores.masked_fill(indices_to_remove, self.filter_value)
+                return scores
+
+        def new_get_logits_warper(
+            top_k: int = None,
+            top_p: float = None,
+            tfs: float = None,
+            temp: float = None,
+            beams: int = 1,
+        ) -> LogitsProcessorList:
+            warper_list = LogitsProcessorList()
+            if(top_k is not None and top_k > 0):
+                warper_list.append(TopKLogitsWarper(top_k=top_k, min_tokens_to_keep=1 + (beams > 1)))
+            if(top_p is not None and top_p < 1.0):
+                warper_list.append(TopPLogitsWarper(top_p=top_p, min_tokens_to_keep=1 + (beams > 1)))
+            if(tfs is not None and tfs < 1.0):
+                warper_list.append(TailFreeLogitsWarper(tfs=tfs, min_tokens_to_keep=1 + (beams > 1)))
+            if(temp is not None and temp != 1.0):
+                warper_list.append(TemperatureLogitsWarper(temperature=temp))
+            return warper_list
+        
+        def new_sample(self, *args, **kwargs):
+            assert kwargs.pop("logits_warper", None) is not None
+            kwargs["logits_warper"] = new_get_logits_warper(
+                vars.top_k,
+                vars.top_p,
+                vars.tfs,
+                vars.temp,
+                1,
+            )
+            return new_sample.old_sample(self, *args, **kwargs)
+        new_sample.old_sample = transformers.generation_utils.GenerationMixin.sample
+        transformers.generation_utils.GenerationMixin.sample = new_sample
+
+
         # Sets up dynamic world info scanner
         class DynamicWorldInfoScanCriteria(StoppingCriteria):
             def __init__(
@@ -1463,10 +1540,6 @@ def generate(txt, minimum, maximum, found_entries=None):
     
     # Submit input text to generator
     try:
-        top_p = vars.top_p if vars.top_p > 0.0 else None
-        top_k = vars.top_k if vars.top_k > 0 else None
-        tfs = vars.tfs if vars.tfs > 0.0 else None
-
         gen_in = tokenizer.encode(txt, return_tensors="pt", truncation=True).long()
         if(vars.sp is not None):
             soft_tokens = torch.arange(
@@ -1499,10 +1572,6 @@ def generate(txt, minimum, maximum, found_entries=None):
                     min_length=minimum, 
                     max_length=maximum-already_generated,
                     repetition_penalty=vars.rep_pen,
-                    top_p=top_p,
-                    top_k=top_k,
-                    tfs=tfs,
-                    temperature=vars.temp,
                     bad_words_ids=vars.badwordsids,
                     use_cache=True,
                     num_return_sequences=numseqs