Model: Switch logprobs to use post-sampling

Previously, pre-sampling logprobs were used from the raw logits,
but newer versions of exl2 allow for returning token probs post-sampling.
Convert these to logprobs and send to the user.

Signed-off-by: kingbri <[email protected]>

backends/exllamav2/model.py

@@ -466,11 +466,15 @@ def unload(self, loras_only: bool = False):
     def encode_tokens(self, text: str, **kwargs):
         """Wrapper to encode tokens from a text string"""
 
-        return self.tokenizer.encode(
-            text,
-            add_bos=unwrap(kwargs.get("add_bos_token"), True),
-            encode_special_tokens=unwrap(kwargs.get("encode_special_tokens"), True),
-        )[0].tolist()
+        return (
+            self.tokenizer.encode(
+                text,
+                add_bos=unwrap(kwargs.get("add_bos_token"), True),
+                encode_special_tokens=unwrap(kwargs.get("encode_special_tokens"), True),
+            )
+            .flatten()
+            .tolist()
+        )
 
     def decode_tokens(self, ids: List[int], **kwargs):
         """Wrapper to decode tokens from a list of IDs"""
@@ -489,35 +493,24 @@ def get_special_tokens(self, add_bos_token: bool, ban_eos_token: bool):
             "unk_token": self.tokenizer.unk_token,
         }
 
-    def get_logprobs(self, logits: torch.Tensor, max_logprobs: int):
-        normalized_logits = torch.log_softmax(logits, dim=-1)
-        top_values, top_ids = torch.topk(normalized_logits, max_logprobs, dim=-1)
-
+    def get_logprobs(self, token_ids: torch.Tensor, token_probs: torch.Tensor):
         top_tokens = list(
             map(
                 lambda index: self.tokenizer.extended_id_to_piece.get(
                     index, self.tokenizer.id_to_piece[index]
                 ),
-                top_ids[0].tolist(),
+                token_ids.flatten().tolist(),
             )
         )
-        top_values = top_values[0].tolist()
-
-        return dict(zip_longest(top_tokens, top_values))
 
-    def get_token_probs(self, token_ids: torch.tensor, token_probs: torch.Tensor):
-        normalized_probs = torch.log(token_probs)
+        top_values = torch.log(token_probs).flatten().tolist()
 
-        tokens = list(
-            map(
-                lambda index: self.tokenizer.extended_id_to_piece.get(
-                    index, self.tokenizer.id_to_piece[index]
-                ),
-                token_ids[0].tolist(),
-            )
+        # Cannot return -inf in JSON
+        cleaned_values = list(
+            map(lambda value: -1000 if value == float("-inf") else value, top_values)
         )
 
-        return dict(zip_longest(tokens, normalized_probs[0].tolist()))
+        return dict(zip_longest(top_tokens, cleaned_values))
 
     def generate(self, prompt: str, **kwargs):
         """Generate a response to a prompt"""
@@ -712,7 +705,10 @@ def generate_gen(self, prompt: str, **kwargs):
         add_bos_token = unwrap(kwargs.get("add_bos_token"), True)
         ban_eos_token = unwrap(kwargs.get("ban_eos_token"), False)
         logit_bias = kwargs.get("logit_bias")
+
+        # Logprobs
         request_logprobs = unwrap(kwargs.get("logprobs"), 0)
+        self.generator.return_top_tokens = request_logprobs
 
         # Override sampler settings for temp = 0
         if gen_settings.temperature == 0:
@@ -832,14 +828,20 @@ def generate_gen(self, prompt: str, **kwargs):
             if auto_scale_penalty_range:
                 gen_settings.token_repetition_range = generated_tokens
 
-            # Generate
-            chunk, eos, tokens, token_probs, logits = self.generator.stream()
+            # Run dict generation
+            # Guarantees return of chunk, eos, and chunk_token_ids
+            raw_generation = self.generator.stream_ex()
 
             if token_healing:
                 # Extract healed token
                 ids[:, -1] = self.generator.sequence_ids[:, -2]
                 token_healing = False
 
+            # Get parameters that will always exist
+            chunk = raw_generation["chunk"]
+            eos = raw_generation["eos"]
+            tokens = raw_generation["chunk_token_ids"]
+
             save_tokens = torch.cat(
                 (save_tokens, tokens.expand(save_tokens.shape[0], -1)), dim=-1
             )
@@ -863,17 +865,26 @@ def generate_gen(self, prompt: str, **kwargs):
                 }
 
                 if request_logprobs > 0:
-                    # Get sampled token probs
-                    if token_probs.numel() > 0 and tokens.numel() > 0:
-                        generation["token_probs"] = self.get_token_probs(
-                            tokens, token_probs
-                        )
-
-                    # Get logprob choices
-                    if logits.numel() > 0:
-                        generation["logprobs"] = self.get_logprobs(
-                            logits, request_logprobs
-                        )
+                    # Get top tokens and probs
+                    top_tokens = unwrap(
+                        raw_generation.get("top_tokens"),
+                        torch.empty((1, 0, 1), dtype=torch.long),
+                    )
+
+                    top_probs = unwrap(
+                        raw_generation.get("top_probs"),
+                        torch.empty((1, 0, 1), dtype=torch.float),
+                    )
+
+                    if top_tokens.numel() > 0 and top_probs.numel() > 0:
+                        logprobs = self.get_logprobs(top_tokens, top_probs)
+                        generation["logprobs"] = logprobs
+
+                        # The first logprob is the selected token prob
+                        generation["token_probs"] = {
+                            token: logprobs[token]
+                            for token in list(logprobs.keys())[:1]
+                        }
 
                 yield generation
                 full_response += chunk_buffer