Detokenize incrementally when streaming (#653)

python/sglang/srt/layers/radix_attention.py

@@ -136,7 +136,33 @@ class RadixAttention(nn.Module):
             return self.decode_forward(q, k, v, input_metadata)
 
     def store_kv_cache(self, cache_k, cache_v, input_metadata: InputMetadata):
-        key_buffer = input_metadata.token_to_kv_pool.get_key_buffer(self.layer_id)
-        key_buffer[input_metadata.out_cache_loc] = cache_k
-        value_buffer = input_metadata.token_to_kv_pool.get_value_buffer(self.layer_id)
-        value_buffer[input_metadata.out_cache_loc] = cache_v
+        kv_cache = input_metadata.token_to_kv_pool.kv_data[self.layer_id]
+        _store_kv_cache(cache_k, cache_v, kv_cache, input_metadata.out_cache_loc)
+
+
+try:
+
+    @torch.library.custom_op("mylib::store_kv_cache", mutates_args={"kv_cache"})
+    def _store_kv_cache(
+        k: torch.Tensor,
+        v: torch.Tensor,
+        kv_cache: torch.Tensor,
+        cache_loc: torch.Tensor,
+    ) -> None:
+        kv_cache[cache_loc, 0] = k
+        kv_cache[cache_loc, 1] = v
+
+    @_store_kv_cache.register_fake
+    def _(k, v, kv_cache, cache_loc):
+        pass
+
+except:
+
+    def _store_kv_cache(
+        k: torch.Tensor,
+        v: torch.Tensor,
+        kv_cache: torch.Tensor,
+        cache_loc: torch.Tensor,
+    ) -> None:
+        kv_cache[cache_loc, 0] = k
+        kv_cache[cache_loc, 1] = v

python/sglang/srt/managers/controller/infer_batch.py

@@ -82,6 +82,14 @@ class Req:
         self.input_ids = None  # input_ids = origin_input_ids + output_ids
 
         # For incremental decoding
+        # ----- | --------- read_ids -------|
+        # ----- |   surr_ids  |
+        # xxxxx | xxxxxxxxxxx | xxxxxxxxxxx |
+        # ----- ^ ----------- ^ ----------- ^
+        # ----- 1 ----------- 2 ----------- 3
+        # 1: surr_offset
+        # 2: read_offset
+        # 3: last token
         self.decoded_text = ""
         self.surr_offset = None  # Surrounding offset to defeat the cleanup algorithm
         self.read_offset = None
@@ -132,7 +140,7 @@ class Req:
         return self.finished_reason is not None
 
     # Based on https://github.com/vllm-project/vllm/blob/7a64d24aad69e4d2548aa0bf528d9fe63428ab01/vllm/transformers_utils/detokenizer.py#L194-L313
-    def init_detokenize_incrementally(self):
+    def init_incremental_detokenize(self):
         first_iter = self.surr_offset is None or self.read_offset is None
 
         if first_iter:
@@ -142,13 +150,11 @@ class Req:
             )
 
         all_ids = self.origin_input_ids_unpadded + self.output_ids
-        surr_ids = all_ids[self.surr_offset : self.read_offset]
-        read_ids = all_ids[self.surr_offset :]
+        return all_ids[self.surr_offset :], self.read_offset - self.surr_offset
 
-        return surr_ids, read_ids, len(all_ids)
-
-    def detokenize_incrementally(self, inplace: bool = True):
-        surr_ids, read_ids, num_all_tokens = self.init_detokenize_incrementally()
+    def get_next_inc_detokenization(self):
+        read_ids, read_offset = self.init_incremental_detokenize()
+        surr_ids = read_ids[:read_offset]
 
         surr_text = self.tokenizer.decode(
             surr_ids,
@@ -162,13 +168,7 @@ class Req:
         )
 
         if len(new_text) > len(surr_text) and not new_text.endswith("�"):
-            new_text = new_text[len(surr_text) :]
-            if inplace:
-                self.decoded_text += new_text
-                self.surr_offset = self.read_offset
-                self.read_offset = num_all_tokens
-
-            return True, new_text
+            return True, new_text[len(surr_text) :]
 
         return False, ""
 
@@ -501,7 +501,7 @@ class Batch:
                     cur_output_ids = req.output_ids
 
                     req.output_ids.extend(suffix_ids)
-                    decode_res, new_text = req.detokenize_incrementally(inplace=False)
+                    decode_res, new_text = req.get_next_inc_detokenization()
                     if not decode_res:
                         req.output_ids = cur_output_ids
                         continue

python/sglang/srt/managers/controller/tp_worker.py

@@ -590,8 +590,8 @@ class ModelTpServer:
     def handle_finished_requests(self, batch: Batch):
         output_rids = []
         decoded_texts = []
-        surr_output_ids = []
-        read_output_ids = []
+        output_read_ids = []
+        output_read_offsets = []
         output_skip_special_tokens = []
         output_spaces_between_special_tokens = []
         output_meta_info = []
@@ -615,9 +615,9 @@ class ModelTpServer:
             ):
                 output_rids.append(req.rid)
                 decoded_texts.append(req.decoded_text)
-                surr_ids, read_ids, _ = req.init_detokenize_incrementally()
-                surr_output_ids.append(surr_ids)
-                read_output_ids.append(read_ids)
+                read_ids, read_offset = req.init_incremental_detokenize()
+                output_read_ids.append(read_ids)
+                output_read_offsets.append(read_offset)
                 output_skip_special_tokens.append(
                     req.sampling_params.skip_special_tokens
                 )
@@ -654,8 +654,8 @@ class ModelTpServer:
                 BatchTokenIDOut(
                     output_rids,
                     decoded_texts,
-                    surr_output_ids,
-                    read_output_ids,
+                    output_read_ids,
+                    output_read_offsets,
                     output_skip_special_tokens,
                     output_spaces_between_special_tokens,
                     output_meta_info,

python/sglang/srt/managers/detokenizer_manager.py

 """DetokenizerManager is a process that detokenizes the token ids."""
 
 import asyncio
+import dataclasses
 import inspect
+from typing import List
 
 import uvloop
 import zmq
@@ -16,6 +18,14 @@ from sglang.utils import find_printable_text, get_exception_traceback, graceful_
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 
 
+@dataclasses.dataclass
+class DecodeStatus:
+    decoded_text: str
+    decode_ids: List[int]
+    surr_offset: int
+    read_offset: int
+
+
 class DetokenizerManager:
     def __init__(
         self,
@@ -35,19 +45,42 @@ class DetokenizerManager:
             trust_remote_code=server_args.trust_remote_code,
         )
 
+        self.decode_status = {}
+
     async def handle_loop(self):
         while True:
             recv_obj: BatchTokenIDOut = await self.recv_from_router.recv_pyobj()
             assert isinstance(recv_obj, BatchTokenIDOut)
+            bs = len(recv_obj.rids)
+
+            # FIXME: incremental detokenize is not compatible with jump forward
+            # Initialize decode status
+            read_ids, surr_ids = [], []
+            for i in range(bs):
+                rid = recv_obj.rids[i]
+                if rid not in self.decode_status:
+                    s = DecodeStatus(
+                        decoded_text=recv_obj.decoded_texts[i],
+                        decode_ids=recv_obj.decode_ids[i],
+                        surr_offset=0,
+                        read_offset=recv_obj.read_offsets[i],
+                    )
+                    self.decode_status[rid] = s
+                else:
+                    s = self.decode_status[rid]
+                    s.decode_ids = recv_obj.decode_ids[i]
+
+                read_ids.append(s.decode_ids[s.surr_offset :])
+                surr_ids.append(s.decode_ids[s.surr_offset : s.read_offset])
 
             # TODO(lmzheng): handle skip_special_tokens/spaces_between_special_tokens per request
             surr_texts = self.tokenizer.batch_decode(
-                recv_obj.surr_output_ids,
+                surr_ids,
                 skip_special_tokens=recv_obj.skip_special_tokens[0],
                 spaces_between_special_tokens=recv_obj.spaces_between_special_tokens[0],
             )
             read_texts = self.tokenizer.batch_decode(
-                recv_obj.read_output_ids,
+                read_ids,
                 skip_special_tokens=recv_obj.skip_special_tokens[0],
                 spaces_between_special_tokens=recv_obj.spaces_between_special_tokens[0],
             )
@@ -55,11 +88,20 @@ class DetokenizerManager:
             # Trim stop str
             # TODO(lmzheng): handle the case where multiple stop strs are hit
             output_strs = []
-            for i in range(len(recv_obj.rids)):
+            for i in range(bs):
+                s = self.decode_status[recv_obj.rids[i]]
                 new_text = read_texts[i][len(surr_texts[i]) :]
                 if recv_obj.finished_reason[i] is None:
-                    new_text = find_printable_text(new_text)
-                output_strs.append(recv_obj.decoded_texts[i] + new_text)
+                    # Streaming chunk: update the decode status
+                    if len(new_text) > 0 and not new_text.endswith("�"):
+                        s.decoded_text = s.decoded_text + new_text
+                        s.surr_offset = s.read_offset
+                        s.read_offset = len(s.decode_ids)
+                        new_text = ""
+                    else:
+                        new_text = find_printable_text(new_text)
+
+                output_strs.append(s.decoded_text + new_text)
 
                 if isinstance(recv_obj.finished_reason[i], FINISH_MATCHED_STR):
                     pos = output_strs[i].find(recv_obj.finished_reason[i].matched)

python/sglang/srt/managers/io_struct.py

@@ -111,8 +111,8 @@ class TokenizedGenerateReqInput:
 class BatchTokenIDOut:
     rids: List[str]
     decoded_texts: List[str]
-    surr_output_ids: List[List[int]]
-    read_output_ids: List[List[int]]
+    decode_ids: List[int]
+    read_offsets: List[int]
     skip_special_tokens: List[bool]
     spaces_between_special_tokens: List[bool]
     meta_info: List[Dict]