sglang 0.1.16__py3-none-any.whl → 0.1.18__py3-none-any.whl

sglang/lang/ir.py

@@ -82,6 +82,19 @@ class SglSamplingParams:
             "top_k": self.top_k,
         }
 
+    def to_litellm_kwargs(self):
+        if self.regex is not None:
+            warnings.warn("Regular expression is not supported in the LiteLLM backend.")
+        return {
+            "max_tokens": self.max_new_tokens,
+            "stop": self.stop or None,
+            "temperature": self.temperature,
+            "top_p": self.top_p,
+            "top_k": self.top_k,
+            "frequency_penalty": self.frequency_penalty,
+            "presence_penalty": self.presence_penalty,
+        }
+
     def to_srt_kwargs(self):
         return {
             "max_new_tokens": self.max_new_tokens,
@@ -97,9 +110,9 @@ class SglSamplingParams:
 
 
 class SglFunction:
-    def __init__(self, func, api_num_spec_tokens=None, bind_arguments=None):
+    def __init__(self, func, num_api_spec_tokens=None, bind_arguments=None):
         self.func = func
-        self.api_num_spec_tokens = api_num_spec_tokens
+        self.num_api_spec_tokens = num_api_spec_tokens
         self.bind_arguments = bind_arguments or {}
         self.pin_prefix_rid = None
 
@@ -107,6 +120,7 @@ class SglFunction:
         argspec = inspect.getfullargspec(func)
         assert argspec.args[0] == "s", 'The first argument must be "s"'
         self.arg_names = argspec.args[1:]
+        self.arg_defaults = argspec.defaults if argspec.defaults is not None else []
 
     def bind(self, **kwargs):
         assert all(key in self.arg_names for key in kwargs)
@@ -165,7 +179,18 @@ class SglFunction:
         assert isinstance(batch_kwargs, (list, tuple))
         if len(batch_kwargs) == 0:
             return []
-        assert isinstance(batch_kwargs[0], dict)
+        if not isinstance(batch_kwargs[0], dict):
+            num_programs = len(batch_kwargs)
+            # change the list of argument values to dict of arg_name -> arg_value
+            batch_kwargs = [
+                {self.arg_names[i]: v for i, v in enumerate(arg_values)}
+                for arg_values in batch_kwargs
+                if isinstance(arg_values, (list, tuple)) and
+                   len(self.arg_names) - len(self.arg_defaults) <= len(arg_values) <= len(self.arg_names)
+            ]
+            # Ensure to raise an exception if the number of arguments mismatch
+            if len(batch_kwargs) != num_programs:
+                raise Exception("Given arguments mismatch the SGL function signature")
 
         default_sampling_para = SglSamplingParams(
             max_new_tokens=max_new_tokens,

sglang/launch_server.py

@@ -1,6 +1,9 @@
+"""Launch the inference server."""
+
 import argparse
 
-from sglang.srt.server import ServerArgs, launch_server
+from sglang.srt.server import launch_server
+from sglang.srt.server_args import ServerArgs
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()

sglang/launch_server_llavavid.py

@@ -1,10 +1,11 @@
+"""Launch the inference server for Llava-video model."""
+
 import argparse
 import multiprocessing as mp
 
 from sglang.srt.server import ServerArgs, launch_server
 
 if __name__ == "__main__":
-
     model_overide_args = {}
 
     model_overide_args["mm_spatial_pool_stride"] = 2

sglang/srt/constrained/__init__.py

@@ -1,13 +1,19 @@
 import json
 from typing import Dict, Optional, Union
 
-from outlines.caching import cache as disk_cache
-from outlines.caching import disable_cache
-from outlines.fsm.fsm import RegexFSM
-from outlines.fsm.regex import FSMInfo, make_deterministic_fsm
-from outlines.models.transformers import TransformerTokenizer
 from pydantic import BaseModel
 
+try:
+    from outlines.caching import cache as disk_cache
+    from outlines.fsm.guide import RegexGuide
+    from outlines.caching import disable_cache
+    from outlines.fsm.guide import RegexGuide
+    from outlines.fsm.regex import FSMInfo, make_byte_level_fsm, make_deterministic_fsm
+    from outlines.models.transformers import TransformerTokenizer
+except ImportError as e:
+    print(f'\nError: {e}. Please install a new version of outlines by `pip install "outlines>=0.0.44"`\n')
+    raise
+
 try:
     from outlines.fsm.json_schema import build_regex_from_object
 except ImportError:
@@ -28,11 +34,12 @@ except ImportError:
 
 
 __all__ = [
-    "RegexFSM",
+    "RegexGuide",
     "FSMInfo",
     "make_deterministic_fsm",
     "build_regex_from_object",
     "TransformerTokenizer",
     "disk_cache",
     "disable_cache",
+    "make_byte_level_fsm",
 ]

sglang/srt/constrained/fsm_cache.py

@@ -1,4 +1,6 @@
-from sglang.srt.constrained import RegexFSM, TransformerTokenizer
+"""Cache for the compressed finite state machine."""
+
+from sglang.srt.constrained import RegexGuide, TransformerTokenizer
 from sglang.srt.constrained.base_cache import BaseCache
 
 
@@ -6,6 +8,10 @@ class FSMCache(BaseCache):
     def __init__(self, tokenizer_path, tokenizer_args_dict, enable=True):
         super().__init__(enable=enable)
 
+        if tokenizer_path.endswith(".json") or tokenizer_path.endswith(".model"):
+            # Do not support TiktokenTokenizer or SentencePieceTokenizer
+            return
+
         from importlib.metadata import version
 
         if version("outlines") >= "0.0.35":
@@ -22,4 +28,4 @@ class FSMCache(BaseCache):
             )
 
     def init_value(self, regex):
-        return RegexFSM(regex, self.outlines_tokenizer)
+        return RegexGuide(regex, self.outlines_tokenizer)

sglang/srt/constrained/jump_forward.py

@@ -1,17 +1,43 @@
-import interegular
+"""
+Faster constrained decoding.
+Reference: https://lmsys.org/blog/2024-02-05-compressed-fsm/
+"""
+
+import dataclasses
+from collections import defaultdict
 
-from sglang.srt.constrained import FSMInfo, disk_cache, make_deterministic_fsm
+import interegular
+import outlines.caching
+
+from sglang.srt.constrained import (
+    FSMInfo,
+    disk_cache,
+    make_byte_level_fsm,
+    make_deterministic_fsm,
+)
 from sglang.srt.constrained.base_cache import BaseCache
 
 IP_REGEX = r"((25[0-5]|2[0-4]\d|[01]?\d\d?)\.){3}(25[0-5]|2[0-4]\d|[01]?\d\d?)"
 
 
+@dataclasses.dataclass
+class JumpEdge:
+    symbol: str = None
+    symbol_next_state: int = None
+    byte: int = None
+    byte_next_state: int = None
+
+
 class JumpForwardMap:
     def __init__(self, regex_string):
         @disk_cache()
         def _init_state_to_jump_forward(regex_string):
             regex_pattern = interegular.parse_pattern(regex_string)
-            regex_fsm, _ = make_deterministic_fsm(regex_pattern.to_fsm().reduce())
+
+            byte_fsm = make_byte_level_fsm(
+                regex_pattern.to_fsm().reduce(), keep_utf8=True
+            )
+            regex_fsm, _ = make_deterministic_fsm(byte_fsm)
 
             fsm_info: FSMInfo = regex_fsm.fsm_info
 
@@ -21,40 +47,93 @@ class JumpForwardMap:
                 id_to_symbol.setdefault(id_, []).append(symbol)
 
             transitions = fsm_info.transitions
-            dirty_states = set()
+            outgoings_ct = defaultdict(int)
             state_to_jump_forward = {}
 
             for (state, id_), next_state in transitions.items():
-                if state in dirty_states:
-                    continue
-                if state in state_to_jump_forward:
-                    dirty_states.add(state)
-                    del state_to_jump_forward[state]
+                if id_ == fsm_info.alphabet_anything_value:
                     continue
-                if len(id_to_symbol[id_]) > 1:
-                    dirty_states.add(state)
+                symbols = id_to_symbol[id_]
+                for c in symbols:
+                    if len(c) > 1:
+                        # Skip byte level transitions
+                        continue
+
+                    outgoings_ct[state] += 1
+                    if outgoings_ct[state] > 1:
+                        if state in state_to_jump_forward:
+                            del state_to_jump_forward[state]
+                        break
+
+                    state_to_jump_forward[state] = JumpEdge(
+                        symbol=c,
+                        symbol_next_state=next_state,
+                    )
+
+            # Process the byte level jump forward
+            outgoings_ct = defaultdict(int)
+            for (state, id_), next_state in transitions.items():
+                if id_ == fsm_info.alphabet_anything_value:
                     continue
-
-                state_to_jump_forward[state] = (id_to_symbol[id_][0], next_state)
+                symbols = id_to_symbol[id_]
+                for c in symbols:
+                    byte_ = None
+                    if len(c) == 1 and ord(c) < 0x80:
+                        # ASCII character
+                        byte_ = ord(c)
+                    elif len(c) > 1:
+                        # FIXME: This logic is due to the leading \x00
+                        # https://github.com/outlines-dev/outlines/pull/930
+                        byte_ = int(symbols[0][1:], 16)
+
+                    if byte_ is not None:
+                        outgoings_ct[state] += 1
+                        if outgoings_ct[state] > 1:
+                            if state in state_to_jump_forward:
+                                del state_to_jump_forward[state]
+                            break
+                        e = state_to_jump_forward.get(state, JumpEdge())
+                        e.byte = byte_
+                        e.byte_next_state = next_state
+                        state_to_jump_forward[state] = e
 
             return state_to_jump_forward
 
         self.state_to_jump_forward = _init_state_to_jump_forward(regex_string)
 
-    def valid_states(self):
-        return self.state_to_jump_forward.keys()
+    def jump_forward_symbol(self, state):
+        jump_forward_str = ""
+        next_state = state
+        while state in self.state_to_jump_forward:
+            e = self.state_to_jump_forward[state]
+            if e.symbol is None:
+                break
+            jump_forward_str += e.symbol
+            next_state = e.symbol_next_state
+            state = next_state
 
-    def jump_forward(self, state):
+        return jump_forward_str, next_state
+
+    def jump_forward_byte(self, state):
         if state not in self.state_to_jump_forward:
             return None
 
-        jump_forward_str = ""
+        jump_forward_bytes = []
         next_state = None
         while state in self.state_to_jump_forward:
-            symbol, next_state = self.state_to_jump_forward[state]
-            jump_forward_str += symbol
+            e = self.state_to_jump_forward[state]
+            assert e.byte is not None and e.byte_next_state is not None
+            jump_forward_bytes.append((e.byte, e.byte_next_state))
+            next_state = e.byte_next_state
             state = next_state
-        return jump_forward_str, next_state
+
+        return jump_forward_bytes
+
+    def is_jump_forward_symbol_state(self, state):
+        return (
+            state in self.state_to_jump_forward
+            and self.state_to_jump_forward[state].symbol is not None
+        )
 
 
 class JumpForwardCache(BaseCache):
@@ -65,12 +144,21 @@ class JumpForwardCache(BaseCache):
         return JumpForwardMap(regex)
 
 
-def test_main():
-    regex_string = r"The google's DNS sever address is " + IP_REGEX
+def test_main(regex_string):
     jump_forward_map = JumpForwardMap(regex_string)
-    for state in jump_forward_map.valid_states():
-        print(state, f'"{jump_forward_map.jump_forward(state)}"')
+    for state, e in jump_forward_map.state_to_jump_forward.items():
+        if e.symbol is not None:
+            jump_forward_str, next_state = jump_forward_map.jump_forward_symbol(state)
+            print(f"{state} -> {next_state}", jump_forward_str)
+        bytes_ = jump_forward_map.jump_forward_byte(state)
+        print(f"{state} -> {bytes_[-1][1]}", [hex(b) for b, _ in bytes_])
 
 
 if __name__ == "__main__":
-    test_main()
+    import outlines
+
+    outlines.caching.clear_cache()
+    test_main(r"The google's DNS sever address is " + IP_REGEX)
+    test_main(r"霍格沃茨特快列车|霍比特人比尔博")
+    # 霍格: \xe9\x9c\x8d \xe6\xa0\xbc ...
+    # 霍比: \xe9\x9c\x8d \xe6\xaf\x94 ...

sglang/srt/conversation.py

@@ -1,3 +1,5 @@
+"""Conversation templates."""
+
 # Adapted from
 # https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
 import dataclasses

sglang/srt/flush_cache.py

@@ -1,4 +1,6 @@
 """
+Flush the KV cache.
+
 Usage:
 python3 -m sglang.srt.flush_cache --url http://localhost:30000
 """
@@ -13,4 +15,4 @@ if __name__ == "__main__":
     args = parser.parse_args()
 
     response = requests.get(args.url + "/flush_cache")
-    assert response.status_code == 200
+    assert response.status_code == 200

sglang/srt/hf_transformers_utils.py

@@ -1,9 +1,10 @@
 """Utilities for Huggingface Transformers."""
 
+import functools
 import json
 import os
 import warnings
-from typing import List, Optional, Tuple, Union
+from typing import AbstractSet, Collection, Literal, Optional, Union
 
 from huggingface_hub import snapshot_download
 from transformers import (
@@ -84,6 +85,12 @@ def get_tokenizer(
     tokenizer_revision: Optional[str] = None,
     **kwargs,
 ) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
+    if tokenizer_name.endswith(".json"):
+        return TiktokenTokenizer(tokenizer_name)
+
+    if tokenizer_name.endswith(".model"):
+        return SentencePieceTokenizer(tokenizer_name)
+
     """Gets a tokenizer for the given model name via Huggingface."""
     if is_multimodal_model(tokenizer_name):
         processor = get_processor(
@@ -170,3 +177,125 @@ def get_processor(
         **kwargs,
     )
     return processor
+
+
+class TiktokenTokenizer:
+    def __init__(self, tokenizer_path):
+        import tiktoken
+        from jinja2 import Template
+
+        PAT_STR_B = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
+
+        # Read JSON
+        name = "tmp-json"
+        with open(tokenizer_path, "rb") as fin:
+            tok_dict = json.load(fin)
+
+        mergeable_ranks = {
+            bytes(item["bytes"]): item["token"] for item in tok_dict["regular_tokens"]
+        }
+        special_tokens = {
+            bytes(item["bytes"]).decode(): item["token"]
+            for item in tok_dict["special_tokens"]
+        }
+        assert tok_dict["word_split"] == "V1"
+
+        kwargs = {
+            "name": name,
+            "pat_str": tok_dict.get("pat_str", PAT_STR_B),
+            "mergeable_ranks": mergeable_ranks,
+            "special_tokens": special_tokens,
+        }
+        if "default_allowed_special" in tok_dict:
+            default_allowed_special = set(
+                [
+                    bytes(bytes_list).decode()
+                    for bytes_list in tok_dict["default_allowed_special"]
+                ]
+            )
+        else:
+            default_allowed_special = None
+        if "vocab_size" in tok_dict:
+            kwargs["explicit_n_vocab"] = tok_dict["vocab_size"]
+
+        tokenizer = tiktoken.Encoding(**kwargs)
+        tokenizer._default_allowed_special = default_allowed_special or set()
+        tokenizer._default_allowed_special |= {"<|separator|>"}
+
+        def encode_patched(
+            self,
+            text: str,
+            *,
+            allowed_special: Union[
+                Literal["all"], AbstractSet[str]
+            ] = set(),  # noqa: B006
+            disallowed_special: Union[Literal["all"], Collection[str]] = "all",
+        ) -> list[int]:
+            if isinstance(allowed_special, set):
+                allowed_special |= self._default_allowed_special
+            return tiktoken.Encoding.encode(
+                self,
+                text,
+                allowed_special=allowed_special,
+                disallowed_special=disallowed_special,
+            )
+
+        tokenizer.encode = functools.partial(encode_patched, tokenizer)
+
+        # Convert to HF interface
+        self.tokenizer = tokenizer
+        self.eos_token_id = tokenizer._special_tokens["<|eos|>"]
+        self.vocab_size = tokenizer.n_vocab
+        self.chat_template = Template(
+            "{% for message in messages %}{% if message['role'] == 'user' %}{{ 'Human: ' + message['content'].strip() + '<|separator|>\n\n' }}{% elif message['role'] == 'system' %}{{ 'System: ' + message['content'].strip() + '<|separator|>\n\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: '  + message['content'] + '<|separator|>\n\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}"
+        )
+
+    def encode(self, x, add_special_tokens=False):
+        return self.tokenizer.encode(x)
+
+    def decode(self, x):
+        return self.tokenizer.decode(x)
+
+    def batch_decode(
+        self, batch, skip_special_tokens=True, spaces_between_special_tokens=False
+    ):
+        if isinstance(batch[0], int):
+            batch = [[x] for x in batch]
+        return self.tokenizer.decode_batch(batch)
+
+    def apply_chat_template(self, messages, tokenize, add_generation_prompt):
+        ret = self.chat_template.render(messages=messages, add_generation_prompt=add_generation_prompt)
+        return self.encode(ret) if tokenize else ret
+
+
+class SentencePieceTokenizer:
+    def __init__(self, tokenizer_path):
+        import sentencepiece as spm
+        from jinja2 import Template
+
+        tokenizer = spm.SentencePieceProcessor(model_file=tokenizer_path)
+
+        # Convert to HF interface
+        self.tokenizer = tokenizer
+        self.eos_token_id = tokenizer.eos_id()
+        self.vocab_size = tokenizer.vocab_size()
+        self.chat_template = Template(
+            "{% for message in messages %}{% if message['role'] == 'user' %}{{ 'Human: ' + message['content'].strip() + '<|separator|>\n\n' }}{% elif message['role'] == 'system' %}{{ 'System: ' + message['content'].strip() + '<|separator|>\n\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: '  + message['content'] + '<|separator|>\n\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}"
+        )
+
+    def encode(self, x, add_special_tokens=False):
+        return self.tokenizer.encode(x)
+
+    def decode(self, x):
+        return self.tokenizer.decode(x)
+
+    def batch_decode(
+        self, batch, skip_special_tokens=True, spaces_between_special_tokens=False
+    ):
+        if isinstance(batch[0], int):
+            batch = [[x] for x in batch]
+        return self.tokenizer.decode(batch)
+
+    def apply_chat_template(self, messages, tokenize, add_generation_prompt):
+        ret = self.chat_template.render(messages=messages, add_generation_prompt=add_generation_prompt)
+        return self.encode(ret) if tokenize else ret

sglang/srt/layers/extend_attention.py

@@ -8,6 +8,12 @@ from sglang.srt.utils import wrap_kernel_launcher
 CUDA_CAPABILITY = torch.cuda.get_device_capability()
 
 
+@triton.jit
+def tanh(x):
+    # Tanh is just a scaled sigmoid
+    return 2 * tl.sigmoid(2 * x) - 1
+
+
 @triton.jit
 def _fwd_kernel(
     Q_Extend,
@@ -39,6 +45,7 @@ def _fwd_kernel(
     BLOCK_DMODEL: tl.constexpr,
     BLOCK_M: tl.constexpr,
     BLOCK_N: tl.constexpr,
+    logit_cap: tl.constexpr,
 ):
     cur_seq = tl.program_id(0)
     cur_head = tl.program_id(1)
@@ -90,6 +97,10 @@ def _fwd_kernel(
         qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
         qk += tl.dot(q, k)
         qk *= sm_scale
+
+        if logit_cap > 0:
+            qk = logit_cap * tanh(qk / logit_cap)
+
         qk = tl.where(mask_m[:, None] & mask_n[None, :], qk, float("-inf"))
 
         n_e_max = tl.maximum(tl.max(qk, 1), e_max)
@@ -126,6 +137,10 @@ def _fwd_kernel(
         qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
         qk += tl.dot(q, k)
         qk *= sm_scale
+
+        if logit_cap > 0:
+            qk = logit_cap * tanh(qk / logit_cap)
+
         mask_causual = (cur_block_m * BLOCK_M + offs_m[:, None]) >= (
             start_n + offs_n[None, :]
         )
@@ -176,6 +191,7 @@ def extend_attention_fwd(
     b_seq_len_extend,
     max_len_in_batch,
     max_len_extend,
+    logit_cap=-1,
 ):
     """
     q_extend, k_extend, v_extend, o_extend: contiguous tensors
@@ -271,6 +287,7 @@ def extend_attention_fwd(
         BLOCK_N=BLOCK_N,
         num_warps=num_warps,
         num_stages=num_stages,
+        logit_cap=logit_cap,
     )
     cached_kernel = wrap_kernel_launcher(_fwd_kernel)