Defer building top-token objects to Rust

proto/generate.proto

@@ -143,6 +143,17 @@ message PrefillTokens {
     repeated string texts = 3;
 }
 
+message TopTokens {
+    /// Top Token IDs
+    repeated uint32 ids = 1;
+    /// Top Logprobs
+    repeated float logprobs = 2;
+    /// Top Token Texts
+    repeated string texts = 3;
+    /// If the tokens are special
+    repeated bool is_special = 6;
+}
+
 message TopToken {
     /// Token ID
     uint32 token_id = 3;
@@ -170,7 +181,8 @@ message Generation {
     /// Complete generated text
     optional GeneratedText generated_text = 7;
     /// Top tokens
-    repeated TopToken top_tokens = 8;
+    // repeated TopToken top_tokens = 8;
+    TopTokens top_tokens = 8;
 }
 
 message FilterBatchRequest {

router/src/infer.rs

@@ -167,10 +167,7 @@ impl Infer {
                         .collect();
                 }
                 // Push last token
-                InferStreamResponse::Intermediate{
+                InferStreamResponse::Intermediate { token, top_tokens } => {
-                    token,
-                    top_tokens,
-                } => {
                     result_tokens.push(token);
                     result_top_tokens.push(top_tokens);
                 }
@@ -182,7 +179,6 @@ impl Infer {
                     start,
                     queued,
                     top_tokens,
-
                 } => {
                     result_tokens.push(token);
                     result_top_tokens.push(top_tokens);
@@ -203,7 +199,11 @@ impl Infer {
                 generated_text,
                 queued,
                 start,
-                top_tokens: if use_top_tokens { result_top_tokens } else { Vec::new() },
+                top_tokens: if use_top_tokens {
+                    result_top_tokens
+                } else {
+                    Vec::new()
+                },
             })
         } else {
             let err = InferError::IncompleteGeneration;
@@ -533,16 +533,24 @@ fn send_responses(
         special: generation.token_is_special,
     };
 
-
     // generation.top_tokens
+
     let mut top_tokens = Vec::new();
-    for top_token in generation.top_tokens {
+    if let Some(top_tokens_) = generation.top_tokens {
-        top_tokens.push(Token{
+        top_tokens.extend(
-            id: top_token.token_id,
+            top_tokens_
-            text: top_token.token_text,
+                .ids
-            logprob: top_token.token_logprob,
+                .into_iter()
-            special: top_token.token_is_special,
+                .zip(top_tokens_.logprobs.into_iter())
-        })
+                .zip(top_tokens_.texts.into_iter())
+                .zip(top_tokens_.is_special.into_iter())
+                .map(|(((id, logprob), text), special)| Token {
+                    id,
+                    text,
+                    logprob,
+                    special,
+                })
+        )
     }
 
     if let Some(generated_text) = generation.generated_text {
@@ -562,7 +570,7 @@ fn send_responses(
     } else {
         // Send message
         entry.response_tx.send_timeout(
-            Ok(InferStreamResponse::Intermediate{token, top_tokens}),
+            Ok(InferStreamResponse::Intermediate { token, top_tokens }),
             Duration::from_millis(10),
         )?;
     }

server/text_generation_server/models/causal_lm.py

@@ -13,6 +13,7 @@ from text_generation_server.models.types import (
     PrefillTokens,
     Generation,
     GeneratedText,
+    TopTokens,
 )
 from text_generation_server.pb import generate_pb2
 from text_generation_server.utils import NextTokenChooser, StoppingCriteria, Sampling
@@ -126,7 +127,9 @@ class CausalLMBatch(Batch):
         position_ids = tokenized_inputs["attention_mask"].long().cumsum(-1) - 1
         position_ids.masked_fill_(tokenized_inputs["attention_mask"] == 0, 1)
         all_input_ids = tokenized_inputs["input_ids"].T.split(1, dim=1)
-        top_n_tokens_tensor = torch.tensor(top_n_tokens, device=device, dtype=torch.int64)
+        top_n_tokens_tensor = torch.tensor(
+            top_n_tokens, device=device, dtype=torch.int64
+        )
 
         max_tokens = len(inputs) * (max_input_length + max_decode_tokens)
 
@@ -574,7 +577,9 @@ class CausalLM(Model):
         stopped = True
 
         batch_top_token_ids, batch_top_token_logprobs = batch_top_tokens(
-            batch.top_n_tokens, batch.top_n_tokens_tensor, torch.softmax(logits[:, -1], -1)
+            batch.top_n_tokens,
+            batch.top_n_tokens_tensor,
+            torch.softmax(logits[:, -1], -1),
         )
 
         # Zipped iterator
@@ -652,8 +657,7 @@ class CausalLM(Model):
                     generated_text = None
 
                 # Prefill
-                prefill = stopping_criteria.current_tokens == 1
+                if stopping_criteria.current_tokens == 1 and request.prefill_logprobs:
-                if prefill and request.prefill_logprobs:
                     # Remove generated token to only have prefill and add nan for first prompt token
                     prefill_logprobs = [float("nan")] + torch.log_softmax(
                         logits, -1
@@ -672,15 +676,20 @@ class CausalLM(Model):
                 else:
                     prefill_tokens = None
 
-                # Todo: Make optional for prefill
+                if top_n_tokens > 0:
-                if not prefill and top_n_tokens > 0:
+                    toptoken_texts = self.tokenizer.batch_decode(
-                    top_tokens = self.decode_top_tokens(
+                        top_token_ids,
-                        input_ids=all_input_ids[:-1].view(-1).tolist(),
+                        clean_up_tokenization_spaces=False,
-                        top_n_tokens=top_n_tokens,
+                        skip_special_tokens=False,
-                        top_token_ids=top_token_ids,
+                    )
-                        top_token_logprobs=top_token_logprobs,
+                    special_toptokens = [
-                        prefix_offset=prefix_offset,
+                        token_id in self.all_special_ids for token_id in top_token_ids
-                        read_offset=read_offset,
+                    ]
+                    top_tokens = TopTokens(
+                        top_token_ids,
+                        top_token_logprobs,
+                        toptoken_texts,
+                        special_toptokens,
                     )
                 else:
                     top_tokens = None

server/text_generation_server/models/flash_causal_lm.py

@@ -17,7 +17,7 @@ from text_generation_server.models.types import (
     PrefillTokens,
     Generation,
     GeneratedText,
-    TopToken,
+    TopTokens,
 )
 from text_generation_server.pb import generate_pb2
 from text_generation_server.utils import StoppingCriteria, HeterogeneousNextTokenChooser
@@ -358,7 +358,9 @@ class FlashCausalLMBatch(Batch):
             prefill_next_token_indices = torch.tensor(
                 prefill_next_token_indices, dtype=torch.int64, device=device
             )
-        top_n_tokens_tensor = torch.tensor(top_n_tokens, device=device, dtype=torch.int64)
+        top_n_tokens_tensor = torch.tensor(
+            top_n_tokens, device=device, dtype=torch.int64
+        )
 
         return cls(
             batch_id=pb.id,
@@ -1039,15 +1041,20 @@ class FlashCausalLM(Model):
                 else:
                     prefill_tokens = None
 
-                # Todo: Make optional for prefill
+                if top_n_tokens > 0:
-                if not prefill and top_n_tokens > 0:
+                    toptoken_texts = self.tokenizer.batch_decode(
-                    top_tokens = self.decode_top_tokens(
+                        top_token_ids,
-                        input_ids=all_input_ids[:-1],
+                        clean_up_tokenization_spaces=False,
-                        top_n_tokens=top_n_tokens,
+                        skip_special_tokens=False,
-                        top_token_ids=top_token_ids,
+                    )
-                        top_token_logprobs=top_token_logprobs,
+                    special_toptokens = [
-                        prefix_offset=prefix_offset,
+                        token_id in self.all_special_ids for token_id in top_token_ids
-                        read_offset=read_offset,
+                    ]
+                    top_tokens = TopTokens(
+                        top_token_ids,
+                        top_token_logprobs,
+                        toptoken_texts,
+                        special_toptokens,
                     )
                 else:
                     top_tokens = None

server/text_generation_server/models/model.py

@@ -5,7 +5,7 @@ from abc import ABC, abstractmethod
 from typing import List, Tuple, Optional, TypeVar, Type
 from transformers import PreTrainedTokenizerBase, PretrainedConfig
 
-from text_generation_server.models.types import Batch, GeneratedText, TopToken
+from text_generation_server.models.types import Batch, GeneratedText
 from text_generation_server.pb.generate_pb2 import InfoResponse
 
 B = TypeVar("B", bound=Batch)
@@ -86,76 +86,6 @@ class Model(ABC):
         else:
             return "", prefix_offset, read_offset
 
-    def decode_tokens(
-        self,
-        input_ids: List[int],
-        new_input_ids: List[int],
-        prefix_offset: int = 0,
-        read_offset: int = 0,
-    ) -> Tuple[str, int, int]:
-        """Version of decode_token that supports multiple new tokens for the same prefix."""
-
-        # The prefix text is necessary only to defeat cleanup algorithms in the decode
-        # which decide to add a space or not depending on the surrounding ids.
-        prefix_text = self.tokenizer.decode(
-            input_ids[prefix_offset:read_offset], skip_special_tokens=False
-        )
-
-        new_sequences = [
-            input_ids[prefix_offset:] + [new_id] for new_id in new_input_ids
-        ]
-        new_texts = self.tokenizer.batch_decode(
-            new_sequences, skip_special_tokens=False
-        )
-
-        prefix_len = len(prefix_text)
-        results = []
-        for new_text in new_texts:
-            if len(new_text) > prefix_len and not new_text.endswith("<EFBFBD>"):
-                # utf-8 char at the end means it's a potential unfinished byte sequence
-                # from byte fallback tokenization.
-                # If it's in the middle, it's probably a real invalid id generated
-                # by the model
-                new_text = new_text[prefix_len:]
-                results.append((new_text, read_offset, len(input_ids) + 1))
-            else:
-                results.append(("", prefix_offset, read_offset))
-        return results
-
-    def decode_top_tokens(
-        self,
-        input_ids,
-        top_n_tokens,
-        top_token_ids,
-        top_token_logprobs,
-        prefix_offset,
-        read_offset,
-    ):
-        if top_n_tokens == 0:
-            return []
-
-        top_token_texts = self.decode_tokens(
-            input_ids=input_ids,
-            new_input_ids=top_token_ids,
-            prefix_offset=prefix_offset,
-            read_offset=read_offset,
-        )
-
-        top_tokens = []
-        for token_id, (top_token_text, _, _), token_logprob in zip(
-            top_token_ids, top_token_texts, top_token_logprobs
-        ):
-            tok_itm = token_id
-            top_tokens.append(
-                TopToken(
-                    token_id=token_id,
-                    token_logprob=token_logprob,
-                    token_text=top_token_text,
-                    token_is_special=tok_itm in self.all_special_ids,
-                )
-            )
-        return top_tokens
-
     def check_initialized(self):
         uninitialized_parameters = []
         for n, p in self.model.named_parameters():

server/text_generation_server/models/seq2seq_lm.py

@@ -12,6 +12,7 @@ from text_generation_server.models.types import (
     Batch,
     Generation,
     PrefillTokens,
+    TopTokens,
 )
 from text_generation_server.pb import generate_pb2
 from text_generation_server.utils import NextTokenChooser, StoppingCriteria, Sampling
@@ -130,7 +131,9 @@ class Seq2SeqLMBatch(Batch):
             prefix_offsets.append(0)
             read_offsets.append(1)
         all_decoder_input_ids = decoder_input_ids.view(-1).split(1)
-        top_n_tokens_tensor = torch.tensor(top_n_tokens, device=device, dtype=torch.int64)
+        top_n_tokens_tensor = torch.tensor(
+            top_n_tokens, device=device, dtype=torch.int64
+        )
 
         max_tokens = len(inputs) * (max_input_length + max_decode_tokens)
 
@@ -637,7 +640,9 @@ class Seq2SeqLM(Model):
         )
 
         batch_top_token_ids, batch_top_token_logprobs = batch_top_tokens(
-            batch.top_n_tokens, batch.top_n_tokens_tensor, torch.softmax(logits[:, -1], -1)
+            batch.top_n_tokens,
+            batch.top_n_tokens_tensor,
+            torch.softmax(logits[:, -1], -1),
         )
 
         # Finished requests
@@ -722,8 +727,7 @@ class Seq2SeqLM(Model):
                     generated_text = None
 
                 # Prefill
-                prefill = stopping_criteria.current_tokens == 1
+                if stopping_criteria.current_tokens == 1 and request.prefill_logprobs:
-                if prefill and request.prefill_logprobs:
                     prefill_tokens = PrefillTokens(
                         [self.tokenizer.bos_token_id],
                         [float("nan")],
@@ -732,15 +736,20 @@ class Seq2SeqLM(Model):
                 else:
                     prefill_tokens = None
 
-                # Todo: Make optional for prefill. How to implement in API?
+                if top_n_tokens > 0:
-                if not prefill and top_n_tokens > 0:
+                    toptoken_texts = self.tokenizer.batch_decode(
-                    top_tokens = self.decode_top_tokens(
+                        top_token_ids,
-                        input_ids=all_decoder_input_ids[:-1].view(-1).tolist(),
+                        clean_up_tokenization_spaces=False,
-                        top_n_tokens=top_n_tokens,
+                        skip_special_tokens=False,
-                        top_token_ids=top_token_ids,
+                    )
-                        top_token_logprobs=top_token_logprobs,
+                    special_toptokens = [
-                        prefix_offset=prefix_offset,
+                        token_id in self.all_special_ids for token_id in top_token_ids
-                        read_offset=read_offset,
+                    ]
+                    top_tokens = TopTokens(
+                        top_token_ids,
+                        top_token_logprobs,
+                        toptoken_texts,
+                        special_toptokens,
                     )
                 else:
                     top_tokens = None

server/text_generation_server/models/types.py

@@ -72,28 +72,23 @@ class PrefillTokens:
         return len(self.token_ids)
 
 
-@dataclass(eq=True)
+@dataclass
-@total_ordering
+class TopTokens:
-class TopToken:
+    token_ids: List[int]
-    token_id: int
+    logprobs: List[float]
-    token_logprob: float
+    texts: List[str]
-    token_text: str
+    is_special: List[bool]
-    token_is_special: bool
 
-    def __gt__(self, other):
+    def to_pb(self) -> generate_pb2.TopTokens:
-        # We tiebreak equal logprobs with the _lower_ token_id to align with
+        return generate_pb2.TopTokens(
-        # greedy ordering (torch.argmax)
+            ids=self.token_ids,
-        return self.token_logprob > other.token_logprob or (
+            logprobs=self.logprobs,
-            self.token_logprob == other.token_logprob and self.token_id < other.token_id
+            texts=self.texts,
+            is_special=self.is_special,
         )
 
-    def to_pb(self) -> generate_pb2.TopToken:
+    def __len__(self):
-        return generate_pb2.TopToken(
+        return len(self.token_ids)
-            token_id=self.token_id,
-            token_logprob=self.token_logprob,
-            token_text=self.token_text,
-            token_is_special=self.token_is_special,
-        )
 
 
 @dataclass
@@ -106,7 +101,7 @@ class Generation:
     token_is_special: bool
     generated_text: Optional[GeneratedText]
     # Optional for now, since it's not yet supported for every model.
-    top_tokens: Optional[List[TopToken]]
+    top_tokens: Optional[TopTokens]
 
     def to_pb(self) -> generate_pb2.Generation:
         return generate_pb2.Generation(
@@ -121,7 +116,5 @@ class Generation:
             generated_text=self.generated_text.to_pb()
             if self.generated_text is not None
             else None,
-            top_tokens=[toptoken.to_pb() for toptoken in self.top_tokens]
+            top_tokens=self.top_tokens.to_pb() if self.top_tokens is not None else None,
-            if self.top_tokens
-            else None,
         )