sglang 0.3.4.post2__py3-none-any.whl → 0.3.5__py3-none-any.whl

sglang/srt/managers/io_struct.py

@@ -56,49 +56,47 @@ class GenerateReqInput:
     # LoRA related
     lora_path: Optional[Union[List[Optional[str]], Optional[str]]] = None
 
-    # Whether it is a single request or a batch request
-    is_single: bool = True
-
-    def post_init(self):
+    def normalize_batch_and_arguments(self):
         if (self.text is None and self.input_ids is None) or (
             self.text is not None and self.input_ids is not None
         ):
             raise ValueError("Either text or input_ids should be provided.")
 
-        self.is_single = False
+        # Derive the batch size
         if self.text is not None:
             if isinstance(self.text, str):
                 self.is_single = True
                 self.batch_size = 1
             else:
+                self.is_single = False
                 self.batch_size = len(self.text)
         else:
             if isinstance(self.input_ids[0], int):
                 self.is_single = True
                 self.batch_size = 1
             else:
+                self.is_single = False
                 self.batch_size = len(self.input_ids)
 
+        # Handle parallel sampling
+        # When parallel sampling is used, we always treat the input as a batch.
         if self.sampling_params is None:
             self.parallel_sample_num = 1
         elif isinstance(self.sampling_params, dict):
             self.parallel_sample_num = self.sampling_params.get("n", 1)
         else:  # isinstance(self.sampling_params, list):
             self.parallel_sample_num = self.sampling_params[0].get("n", 1)
-            for sp in self.sampling_params:
-                # TODO cope with the case that the parallel_sample_num is different for different samples
-                assert self.parallel_sample_num == sp.get(
-                    "n", 1
-                ), "The parallel_sample_num should be the same for all samples in sample params."
-
-        if self.parallel_sample_num > 1:
-            if self.is_single:
-                self.is_single = False
-                if self.text is not None:
-                    self.text = [self.text]
-                if self.input_ids is not None:
-                    self.input_ids = [self.input_ids]
+            assert all(self.parallel_sample_num == sampling_params.get("n", 1) for sampling_params in self.sampling_params), (
+                "The parallel_sample_num should be the same for all samples in sample params.")
 
+        if self.parallel_sample_num > 1 and self.is_single:
+            self.is_single = False
+            if self.text is not None:
+                self.text = [self.text]
+            if self.input_ids is not None:
+                self.input_ids = [self.input_ids]
+
+        # Fill in default arguments
         if self.is_single:
             if self.sampling_params is None:
                 self.sampling_params = {}
@@ -114,9 +112,8 @@ class GenerateReqInput:
             if self.parallel_sample_num == 1:
                 num = self.batch_size
             else:
-                # FIXME support cascade inference
-                # first bs samples are used for caching the prefix for parallel sampling
-                num = self.batch_size + self.parallel_sample_num * self.batch_size
+                # Expand parallel_sample_num
+                num = self.batch_size * self.parallel_sample_num
 
             if self.image_data is None:
                 self.image_data = [None] * num
@@ -129,14 +126,11 @@ class GenerateReqInput:
                 self.sampling_params = [{}] * num
             elif not isinstance(self.sampling_params, list):
                 self.sampling_params = [self.sampling_params] * num
-            else:
-                assert self.parallel_sample_num == 1
 
             if self.rid is None:
                 self.rid = [uuid.uuid4().hex for _ in range(num)]
             else:
                 assert isinstance(self.rid, list), "The rid should be a list."
-                assert self.parallel_sample_num == 1
 
             if self.return_logprob is None:
                 self.return_logprob = [False] * num
@@ -159,6 +153,26 @@ class GenerateReqInput:
             else:
                 assert self.parallel_sample_num == 1
 
+    def regenerate_rid(self):
+        self.rid = uuid.uuid4().hex
+        return self.rid
+
+    def __getitem__(self, i):
+        return GenerateReqInput(
+            text=self.text[i] if self.text is not None else None,
+            input_ids=self.input_ids[i] if self.input_ids is not None else None,
+            image_data=self.image_data[i],
+            sampling_params=self.sampling_params[i],
+            rid=self.rid[i],
+            return_logprob=self.return_logprob[i],
+            logprob_start_len=self.logprob_start_len[i],
+            top_logprobs_num=self.top_logprobs_num[i],
+            return_text_in_logprobs=self.return_text_in_logprobs,
+            stream=self.stream,
+            modalities=self.modalities[i] if self.modalities else None,
+            lora_path=self.lora_path[i] if self.lora_path is not None else None,
+        )
+
 
 @dataclass
 class TokenizedGenerateReqInput:
@@ -196,85 +210,61 @@ class EmbeddingReqInput:
     # Dummy sampling params for compatibility
     sampling_params: Union[List[Dict], Dict] = None
 
-    def post_init(self):
+    def normalize_batch_and_arguments(self):
         if (self.text is None and self.input_ids is None) or (
             self.text is not None and self.input_ids is not None
         ):
             raise ValueError("Either text or input_ids should be provided.")
 
+        # Derive the batch size
         if self.text is not None:
-            self.is_single = isinstance(self.text, str)
+            if isinstance(self.text, str):
+                self.is_single = True
+                self.batch_size = 1
+            else:
+                self.is_single = False
+                self.batch_size = len(self.text)
         else:
-            self.is_single = isinstance(self.input_ids[0], int)
+            if isinstance(self.input_ids[0], int):
+                self.is_single = True
+                self.batch_size = 1
+            else:
+                self.is_single = False
+                self.batch_size = len(self.input_ids)
 
+        # Fill in default arguments
         if self.is_single:
             if self.rid is None:
                 self.rid = uuid.uuid4().hex
             if self.sampling_params is None:
                 self.sampling_params = {}
-            self.sampling_params["max_new_tokens"] = 1
+            self.sampling_params["max_new_tokens"] = 0
         else:
-            # support select operation
-            self.batch_size = (
-                len(self.text) if self.text is not None else len(self.input_ids)
-            )
             if self.rid is None:
                 self.rid = [uuid.uuid4().hex for _ in range(self.batch_size)]
             else:
-                if not isinstance(self.rid, list):
-                    raise ValueError("The rid should be a list.")
+                assert isinstance(self.rid, list), "The rid should be a list."
+
             if self.sampling_params is None:
                 self.sampling_params = [{}] * self.batch_size
             for i in range(self.batch_size):
-                self.sampling_params[i]["max_new_tokens"] = 1
+                self.sampling_params[i]["max_new_tokens"] = 0
 
+    def regenerate_rid(self):
+        self.rid = uuid.uuid4().hex
+        return self.rid
 
-@dataclass
-class TokenizedEmbeddingReqInput:
-    # The request id
-    rid: str
-    # The input text
-    input_text: str
-    # The input token ids
-    input_ids: List[int]
-    # Dummy sampling params for compatibility
-    sampling_params: SamplingParams
+    def __getitem__(self, i):
+        return EmbeddingReqInput(
+            text=self.text[i] if self.text is not None else None,
+            input_ids=self.input_ids[i] if self.input_ids is not None else None,
+            sampling_params=self.sampling_params[i],
+            rid=self.rid[i],
+        )
 
 
 @dataclass
-class RewardReqInput:
-    # The input prompt in the chat format. It can be a single prompt or a batch of prompts.
-    conv: Union[List[List[Dict]], List[Dict]]
-    # The request id.
-    rid: Optional[Union[List[str], str]] = None
-    # Dummy sampling params for compatibility
-    sampling_params: Union[List[Dict], Dict] = None
-
-    def post_init(self):
-        self.is_single = isinstance(self.conv[0], dict)
-
-        if self.is_single:
-            if self.rid is None:
-                self.rid = uuid.uuid4().hex
-            if self.sampling_params is None:
-                self.sampling_params = {}
-            self.sampling_params["max_new_tokens"] = 1
-        else:
-            # support select operation
-            self.batch_size = len(self.conv)
-            if self.rid is None:
-                self.rid = [uuid.uuid4().hex for _ in range(self.batch_size)]
-            else:
-                if not isinstance(self.rid, list):
-                    raise ValueError("The rid should be a list.")
-            if self.sampling_params is None:
-                self.sampling_params = [{}] * self.batch_size
-            for i in range(self.batch_size):
-                self.sampling_params[i]["max_new_tokens"] = 1
-
-
-@dataclass
-class TokenizedRewardReqInput:
+class TokenizedEmbeddingReqInput:
     # The request id
     rid: str
     # The input text
@@ -294,6 +284,8 @@ class BatchTokenIDOut:
     decoded_texts: List[str]
     decode_ids: List[int]
     read_offsets: List[int]
+    # Only used when `--skip-tokenizer-init`
+    output_ids: Optional[List[int]]
     skip_special_tokens: List[bool]
     spaces_between_special_tokens: List[bool]
     meta_info: List[Dict]

sglang/srt/managers/schedule_batch.py

@@ -37,8 +37,7 @@ import torch
 
 from sglang.global_config import global_config
 from sglang.srt.configs.model_config import ModelConfig
-from sglang.srt.constrained import RegexGuide
-from sglang.srt.constrained.jump_forward import JumpForwardMap
+from sglang.srt.constrained.grammar import Grammar
 from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 from sglang.srt.mem_cache.chunk_cache import ChunkCache
 from sglang.srt.mem_cache.memory_pool import BaseTokenToKVPool, ReqToTokenPool
@@ -212,9 +211,6 @@ class Req:
         # this does not include the jump forward tokens.
         self.completion_tokens_wo_jump_forward = 0
 
-        # The number of cached tokens, that were already cached in the KV store
-        self.cached_tokens = 0
-
         # For vision inputs
         self.image_inputs: Optional[ImageInputs] = None
 
@@ -222,7 +218,10 @@ class Req:
         self.prefix_indices = []
         self.extend_input_len = 0
         self.last_node = None
-        self.is_inflight_req = 0
+        self.is_being_chunked = 0
+
+        # For retraction
+        self.is_retracted = False
 
         # Logprobs (arguments)
         self.return_logprob = False
@@ -243,13 +242,14 @@ class Req:
         # The relative logprob_start_len in an extend batch
         self.extend_logprob_start_len = 0
 
-        # Embedding
+        # Embedding (return values)
         self.embedding = None
 
         # Constrained decoding
-        self.regex_fsm: RegexGuide = None
-        self.regex_fsm_state: int = 0
-        self.jump_forward_map: JumpForwardMap = None
+        self.grammar: Optional[Grammar] = None
+
+        # The number of cached tokens, that were already cached in the KV cache
+        self.cached_tokens = 0
 
         # For Qwen2-VL
         self.mrope_position_delta = []  # use mutable object
@@ -359,6 +359,8 @@ class Req:
                     return
 
     def jump_forward_and_retokenize(self, jump_forward_str, next_state):
+        assert self.grammar is not None and self.tokenizer is not None
+
         if self.origin_input_text is None:
             # Recovering text can only use unpadded ids
             self.origin_input_text = self.tokenizer.decode(
@@ -398,7 +400,8 @@ class Req:
                 self.surr_offset = self.read_offset - i
                 break
 
-        self.regex_fsm_state = next_state
+        # update the inner state of the grammar
+        self.grammar.jump_and_retokenize(old_output_ids, self.output_ids, next_state)
 
         if self.return_logprob:
             # For fast-forward part's logprobs
@@ -468,8 +471,8 @@ class ScheduleBatch:
     # Stream
     has_stream: bool = False
 
-    # Has regex
-    has_regex: bool = False
+    # Has grammar
+    has_grammar: bool = False
 
     # device
     device: str = "cuda"
@@ -477,7 +480,7 @@ class ScheduleBatch:
     @classmethod
     def init_new(
         cls,
-        reqs,
+        reqs: List[Req],
         req_to_token_pool,
         token_to_kv_pool,
         tree_cache,
@@ -491,7 +494,7 @@ class ScheduleBatch:
             model_config=model_config,
             return_logprob=any(req.return_logprob for req in reqs),
             has_stream=any(req.stream for req in reqs),
-            has_regex=any(req.regex_fsm for req in reqs),
+            has_grammar=any(req.grammar for req in reqs),
             device=req_to_token_pool.device,
         )
 
@@ -561,7 +564,7 @@ class ScheduleBatch:
             seq_lens[i] -= encoder_len
 
             if len(req.prefix_indices) < encoder_len:
-                # NOTE: the encoder part should considered as a whole
+                # NOTE: the encoder part should be considered as a whole
                 assert len(req.prefix_indices) == 0
                 input_ids[i] = input_ids[i][encoder_len:]
                 encoder_out_cache_loc.append(self.out_cache_loc[pt : pt + encoder_len])
@@ -648,6 +651,7 @@ class ScheduleBatch:
 
             req.extend_logprob_start_len = extend_logprob_start_len
             pt += req.extend_input_len
+            req.is_retracted = False
 
         # Set fields
         self.input_ids = torch.tensor(sum(input_ids, []), dtype=torch.int32).to(
@@ -780,6 +784,7 @@ class ScheduleBatch:
             req.prefix_indices = []
             req.last_node = None
             req.extend_input_len = 0
+            req.is_retracted = True
 
             # For incremental logprobs
             req.last_update_decode_tokens = 0
@@ -803,26 +808,10 @@ class ScheduleBatch:
         keep_indices = set(i for i in range(len(self.reqs)))
 
         for i, req in enumerate(self.reqs):
-            if req.jump_forward_map is not None:
-                jump_forward_bytes = req.jump_forward_map.jump_forward_byte(
-                    req.regex_fsm_state
-                )
-                if jump_forward_bytes is not None and len(jump_forward_bytes) > 1:
-                    suffix_bytes = []
-                    continuation_range = range(0x80, 0xC0)
-                    cur_state = req.regex_fsm_state
-                    while (
-                        len(jump_forward_bytes)
-                        and jump_forward_bytes[0][0] in continuation_range
-                    ):
-                        # continuation bytes
-                        byte_edge = jump_forward_bytes.pop(0)
-                        suffix_bytes.append(byte_edge[0])
-                        cur_state = byte_edge[1]
-
-                    suffix_tokens = [f"<0x{hex(b)[2:].upper()}>" for b in suffix_bytes]
-                    suffix_ids = req.tokenizer.convert_tokens_to_ids(suffix_tokens)
-
+            if req.grammar is not None:
+                jump_helper = req.grammar.try_jump(req.tokenizer)
+                if jump_helper.can_jump():
+                    suffix_ids = jump_helper.suffix_ids
                     # Current ids, for cache and revert
                     cur_all_ids = tuple(req.origin_input_ids + req.output_ids)[:-1]
                     cur_output_ids = req.output_ids
@@ -836,10 +825,8 @@ class ScheduleBatch:
                     (
                         jump_forward_str,
                         next_state,
-                    ) = req.jump_forward_map.jump_forward_symbol(cur_state)
+                    ) = req.grammar.jump_forward_str_state(jump_helper)
 
-                    # Make the incrementally decoded text part of jump_forward_str
-                    # so that the UTF-8 will not corrupt
                     jump_forward_str = new_text + jump_forward_str
                     if not req.jump_forward_and_retokenize(
                         jump_forward_str, next_state
@@ -906,7 +893,7 @@ class ScheduleBatch:
 
     def filter_batch(
         self,
-        current_inflight_req: Optional[Req] = None,
+        being_chunked_req: Optional[Req] = None,
         keep_indices: Optional[List[int]] = None,
     ):
         if keep_indices is None:
@@ -914,7 +901,7 @@ class ScheduleBatch:
                 i
                 for i in range(len(self.reqs))
                 if not self.reqs[i].finished()
-                and self.reqs[i] is not current_inflight_req
+                and self.reqs[i] is not being_chunked_req
             ]
 
         if keep_indices is None or len(keep_indices) == 0:
@@ -946,7 +933,7 @@ class ScheduleBatch:
             self.top_logprobs_nums = None
 
         self.has_stream = any(req.stream for req in self.reqs)
-        self.has_regex = any(req.regex_fsm for req in self.reqs)
+        self.has_grammar = any(req.grammar for req in self.reqs)
 
         self.sampling_info.filter_batch(keep_indices, new_indices)
 
@@ -979,7 +966,7 @@ class ScheduleBatch:
 
         self.return_logprob = self.return_logprob or other.return_logprob
         self.has_stream = self.has_stream or other.has_stream
-        self.has_regex = self.has_regex or other.has_regex
+        self.has_grammar = self.has_grammar or other.has_grammar
 
     def get_model_worker_batch(self):
         if self.forward_mode.is_decode():
@@ -989,13 +976,10 @@ class ScheduleBatch:
             extend_prefix_lens = self.prefix_lens
             extend_logprob_start_lens = self.extend_logprob_start_lens
 
-        if self.has_regex:
-            self.sampling_info.regex_fsms = [req.regex_fsm for req in self.reqs]
-            self.sampling_info.regex_fsm_states = [
-                req.regex_fsm_state for req in self.reqs
-            ]
+        if self.has_grammar:
+            self.sampling_info.grammars = [req.grammar for req in self.reqs]
         else:
-            self.sampling_info.regex_fsms = None
+            self.sampling_info.grammars = None
 
         global bid
         bid += 1

sglang/srt/managers/schedule_policy.py

@@ -30,7 +30,9 @@ from sglang.srt.mem_cache.radix_cache import TreeNode
 # This can prevent the server from being too conservative.
 # Note that this only clips the estimation in the scheduler but does not change the stop
 # condition. The request can still generate tokens until it hits the unclipped max_new_tokens.
-CLIP_MAX_NEW_TOKENS = int(os.environ.get("SGLANG_CLIP_MAX_NEW_TOKENS", "4096"))
+CLIP_MAX_NEW_TOKENS_ESTIMATION = int(
+    os.environ.get("SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION", "4096")
+)
 
 
 class SchedulePolicy:
@@ -43,9 +45,15 @@ class SchedulePolicy:
         self.tree_cache = tree_cache
 
     def calc_priority(self, waiting_queue: List[Req]):
+        if len(waiting_queue) > 128 and self.policy == "lpm":
+            # Turn off the expensive prefix matching and sorting when the #queue is large.
+            policy = "fcfs"
+        else:
+            policy = self.policy
+
         # Compute matched prefix length
         prefix_computed = False
-        if self.policy == "lpm" or self.policy == "dfs-weight":
+        if policy == "lpm" or policy == "dfs-weight":
             for r in waiting_queue:
                 # NOTE: the prefix_indices must always be aligned with last_node
                 r.prefix_indices, r.last_node = self.tree_cache.match_prefix(
@@ -54,18 +62,18 @@ class SchedulePolicy:
 
             prefix_computed = True
 
-        if self.policy == "lpm":
+        if policy == "lpm":
             # Longest Prefix Match
             waiting_queue.sort(key=lambda x: -len(x.prefix_indices))
-        elif self.policy == "fcfs":
+        elif policy == "fcfs":
             # first come first serve
             pass
-        elif self.policy == "lof":
+        elif policy == "lof":
             # longest output first
             waiting_queue.sort(key=lambda x: -x.sampling_params.max_new_tokens)
-        elif self.policy == "random":
+        elif policy == "random":
             random.shuffle(waiting_queue)
-        elif self.policy == "dfs-weight":
+        elif policy == "dfs-weight":
             last_node_to_reqs = defaultdict(list)
             for req in waiting_queue:
                 last_node_to_reqs[req.last_node].append(req)
@@ -83,7 +91,7 @@ class SchedulePolicy:
                 waiting_queue,
             )
         else:
-            raise ValueError(f"Unknown schedule_policy: {self.policy}")
+            raise ValueError(f"Unknown schedule_policy: {policy=}")
 
         return prefix_computed
 
@@ -146,7 +154,7 @@ class PrefillAdder:
                 [
                     min(
                         (r.sampling_params.max_new_tokens - len(r.output_ids)),
-                        CLIP_MAX_NEW_TOKENS,
+                        CLIP_MAX_NEW_TOKENS_ESTIMATION,
                     )
                     * self.new_token_ratio
                     for r in running_batch.reqs
@@ -186,7 +194,7 @@ class PrefillAdder:
             len(req.prefix_indices),
             req.extend_input_len,
             (
-                min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS)
+                min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS_ESTIMATION)
                 if not truncated
                 else 0
             ),
@@ -258,7 +266,7 @@ class PrefillAdder:
             self._prefill_one_req(
                 0,
                 req.extend_input_len,
-                min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS),
+                min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS_ESTIMATION),
             )
         else:
             # Chunked prefill
@@ -276,7 +284,7 @@ class PrefillAdder:
             return self.add_one_req_ignore_eos(req)
 
         total_tokens = req.extend_input_len + min(
-            req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS
+            req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS_ESTIMATION
         )
         input_tokens = req.extend_input_len
         prefix_len = len(req.prefix_indices)
@@ -302,7 +310,10 @@ class PrefillAdder:
                 self._prefill_one_req(
                     prefix_len,
                     input_tokens,
-                    min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS),
+                    min(
+                        req.sampling_params.max_new_tokens,
+                        CLIP_MAX_NEW_TOKENS_ESTIMATION,
+                    ),
                 )
             else:
                 # Chunked prefill