Files
sglang/python/sglang/srt/utils/hf_transformers/common.py
T

741 lines
25 KiB
Python

# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Shared helpers used by config, tokenizer, and processor modules."""
import json
import os
from pathlib import Path
from typing import Any, Dict, Optional, Type, Union
import torch
from huggingface_hub import snapshot_download
from sglang.srt.configs import (
AfmoeConfig,
BailingHybridConfig,
BailingMM2Config,
BailingMoeV3VLConfig,
ChatGLMConfig,
Cosmos3Config,
Cosmos3EdgeConfig,
Cosmos3EdgeProjectorConfig,
Cosmos3EdgeTextConfig,
Cosmos3EdgeVisionConfig,
DbrxConfig,
DeepseekVL2Config,
Dots3Config,
DotsOCRConfig,
DotsVLMConfig,
ExaoneConfig,
FalconH1Config,
FalconMambaConfig,
Glm5NextConfig,
Glm5NextTextConfig,
GraniteMoeHybridConfig,
HYV4Config,
InklingAudioConfig,
InklingMMConfig,
InklingModelConfig,
InklingVisionConfig,
InternS2MobiusConfig,
InternS2MobiusTextConfig,
InternS2PreviewConfig,
JetNemotronConfig,
JetVLMConfig,
K2HorizonConfig,
KimiK3Config,
KimiK25Config,
KimiLinearConfig,
KimiVLConfig,
LagunaConfig,
LocateAnythingConfig,
LongcatFlashConfig,
Mamba2Config,
MambaConfig,
MiniCPMHybridConfig,
MiniCPMV4_6Config,
MiniCPMV4_6VisionConfig,
MiniMaxM3VLConfig,
MultiModalityConfig,
MuseGlimmerAssistantConfig,
MuseGlimmerConfig,
NanbeigeConfig,
NemotronH_Nano_Omni_Reasoning_V3_Config,
NemotronH_Nano_VL_V2_Config,
NemotronH_Omni_Reasoning_V3_Config,
NemotronHConfig,
NemotronHPuzzleConfig,
Olmo3Config,
Qwen3_5Config,
Qwen3_5MoeConfig,
Qwen3_5MoeTextConfig,
Qwen3_5TextConfig,
Qwen3NextConfig,
Qwen4ExpConfig,
Qwen4ExpTextConfig,
Spark2_5Config,
Step3p5Config,
Step3p7Config,
Step3VLConfig,
XllmConfig,
)
from sglang.srt.configs.deepseek_ocr import DeepseekVLV2Config
from sglang.srt.configs.internvl import InternVLChatConfig
from sglang.srt.utils import get_bool_env_var, logger, lru_cache_frozenset
from sglang.srt.utils.runai_utils import ObjectStorageModel, is_runai_obj_uri
from ..hf_transformers_patches import normalize_rope_scaling_compat
if get_bool_env_var("SGLANG_USE_MODELSCOPE"):
from modelscope import AutoConfig, GenerationConfig
else:
from transformers import AutoConfig, GenerationConfig
from transformers import PretrainedConfig
# ---------------------------------------------------------------------------
# Config registry
# ---------------------------------------------------------------------------
_CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
cls.model_type: cls
for cls in [
AfmoeConfig,
BailingHybridConfig,
BailingMM2Config,
BailingMoeV3VLConfig,
ChatGLMConfig,
DbrxConfig,
ExaoneConfig,
DeepseekVL2Config,
MultiModalityConfig,
KimiVLConfig,
K2HorizonConfig,
LocateAnythingConfig,
InternVLChatConfig,
LagunaConfig,
Spark2_5Config,
Step3VLConfig,
LongcatFlashConfig,
Olmo3Config,
MuseGlimmerConfig,
MuseGlimmerAssistantConfig,
KimiK3Config,
Glm5NextConfig,
Glm5NextTextConfig,
KimiLinearConfig,
Qwen3NextConfig,
Qwen4ExpConfig,
Qwen4ExpTextConfig,
FalconH1Config,
FalconMambaConfig,
Mamba2Config,
MambaConfig,
GraniteMoeHybridConfig,
HYV4Config,
DotsVLMConfig,
DotsOCRConfig,
Dots3Config,
NemotronH_Nano_VL_V2_Config,
NemotronH_Nano_Omni_Reasoning_V3_Config,
NemotronH_Omni_Reasoning_V3_Config,
NemotronHConfig,
NemotronHPuzzleConfig,
NanbeigeConfig,
DeepseekVLV2Config,
Qwen3_5Config,
Qwen3_5MoeConfig,
Qwen3_5TextConfig,
Qwen3_5MoeTextConfig,
InternS2PreviewConfig,
InternS2MobiusConfig,
InternS2MobiusTextConfig,
JetNemotronConfig,
JetVLMConfig,
KimiK25Config,
Step3p5Config,
Step3p7Config,
MiniCPMHybridConfig,
MiniCPMV4_6Config,
MiniCPMV4_6VisionConfig,
InklingModelConfig,
InklingAudioConfig,
InklingVisionConfig,
InklingMMConfig,
MiniMaxM3VLConfig,
XllmConfig,
]
}
# DeepSeek V3.2 / V4 reuse the V3 config schema. Subclass the upstream
# transformers class with each model_type so AutoConfig.register passes its
# consistency check (which requires class.model_type == registered key).
# Default-value divergences (e.g. V4's topk_group) are handled in
# model_config.py post-load.
try:
from transformers import DeepseekV3Config as _HFDeepseekV3Config
class _DeepseekV32ConfigAlias(_HFDeepseekV3Config):
model_type = "deepseek_v32"
class _DeepseekV4ConfigAlias(_HFDeepseekV3Config):
model_type = "deepseek_v4"
_CONFIG_REGISTRY["deepseek_v32"] = _DeepseekV32ConfigAlias
_CONFIG_REGISTRY["deepseek_v4"] = _DeepseekV4ConfigAlias
# For kimi_k25_eagle3
class _KimiK2ConfigAlias(_HFDeepseekV3Config):
model_type = "kimi_k2"
_CONFIG_REGISTRY["kimi_k2"] = _KimiK2ConfigAlias
except ImportError:
pass
# Newer transformers versions (>=5.10.2) expose MellumConfig directly,
# but fallback to Qwen3MoeConfig for older versions.
try:
import transformers as _hf_transformers
_HFMellumConfig = getattr(_hf_transformers, "MellumConfig", None)
if _HFMellumConfig is not None:
_CONFIG_REGISTRY["mellum"] = _HFMellumConfig
else:
from transformers import Qwen3MoeConfig as _HFQwen3MoeConfig
class _MellumConfigAlias(_HFQwen3MoeConfig):
model_type = "mellum"
def __post_init__(self, **kwargs):
# Qwen3MoeConfig.__post_init__ wipes sliding_window unless
# use_sliding_window=True. Mellum gates sliding attention
# per-layer via layer_types, so preserve sliding_window
# regardless of the legacy use_sliding_window flag.
sliding_window = getattr(self, "sliding_window", None)
super().__post_init__(**kwargs)
self.sliding_window = sliding_window
_CONFIG_REGISTRY["mellum"] = _MellumConfigAlias
except ImportError:
pass
try:
from transformers import Gemma4Config as _HFGemma4Config
class _Gemma4UnifiedConfigAlias(_HFGemma4Config):
model_type = "gemma4_unified"
_CONFIG_REGISTRY["gemma4_unified"] = _Gemma4UnifiedConfigAlias
except ImportError:
pass
for name, cls in _CONFIG_REGISTRY.items():
try:
AutoConfig.register(name, cls)
except ValueError as e:
err = str(e).lower()
if "already registered" not in err and "already used" not in err:
logger.warning("Failed to register config %s: %s", name, e)
# Cosmos3 (understanding tower) reuses the Qwen3-VL config schema. Register it
# with AutoConfig only (not `_CONFIG_REGISTRY`), so the nested `text_config` is
# flattened onto the top-level config in `get_config` — the same path the base
# Qwen3-VL config relies on. Adding it to `_CONFIG_REGISTRY` would trigger a
# `from_pretrained` reload that drops that flattening.
try:
AutoConfig.register(Cosmos3Config.model_type, Cosmos3Config)
except ValueError as e:
err = str(e).lower()
if "already registered" not in err and "already used" not in err:
logger.warning("Failed to register config %s: %s", Cosmos3Config.model_type, e)
# Cosmos3-Edge native text support starts from the checkpoint root config, then
# consumes ``text_config`` in ``sglang.srt.models.cosmos3_edge``. Keep it out of
# `_CONFIG_REGISTRY` so the generic parser can flatten text attributes onto the
# root config after `AutoConfig.from_pretrained`, matching other multimodal
# configs that use a text sub-config.
for _cosmos3_edge_config_cls in (
Cosmos3EdgeTextConfig,
Cosmos3EdgeVisionConfig,
Cosmos3EdgeProjectorConfig,
Cosmos3EdgeConfig,
):
try:
AutoConfig.register(
_cosmos3_edge_config_cls.model_type, _cosmos3_edge_config_cls
)
except ValueError as e:
err = str(e).lower()
if "already registered" not in err and "already used" not in err:
logger.warning(
"Failed to register config %s: %s",
_cosmos3_edge_config_cls.model_type,
e,
)
# ---------------------------------------------------------------------------
# Download / path helpers
# ---------------------------------------------------------------------------
def download_from_hf(
model_path: str,
allow_patterns: Optional[Union[str, list]] = None,
):
if os.path.exists(model_path):
return model_path
if not allow_patterns:
allow_patterns = ["*.json", "*.bin", "*.model"]
return snapshot_download(model_path, allow_patterns=allow_patterns)
def resolve_runai_obj_uri(model_name_or_path: str) -> str:
if is_runai_obj_uri(model_name_or_path):
return ObjectStorageModel.get_path(model_name_or_path)
return model_name_or_path
def _resolve_local_or_cached_file(model_name_or_path, filename, revision=None):
"""Resolve a file from a local directory or HF hub cache (no network)."""
local_path = Path(model_name_or_path) / filename
if local_path.is_file():
return str(local_path)
from huggingface_hub import hf_hub_download
return hf_hub_download(
model_name_or_path, filename, revision=revision, local_files_only=True
)
def _cached_file_exists(model_name_or_path, filename, revision=None) -> bool:
"""Whether *filename* is available locally or in the HF cache (no network)."""
try:
_resolve_local_or_cached_file(model_name_or_path, filename, revision)
return True
except Exception:
return False
def _remote_file_exists(repo_id, filename, revision=None) -> bool:
"""Whether *filename* exists on the HF hub (HEAD request only, no download).
Returns False on any error (offline, gated, network, invalid id) so callers
fall back to their default path instead of crashing.
"""
from huggingface_hub.constants import HF_HUB_OFFLINE
if HF_HUB_OFFLINE:
return False
try:
from huggingface_hub import HfApi
return HfApi().file_exists(repo_id, filename, revision=revision)
except Exception:
return False
def check_gguf_file(model: Union[str, os.PathLike]) -> bool:
model = Path(model)
if not model.is_file():
return False
elif model.suffix == ".gguf":
return True
with open(model, "rb") as f:
header = f.read(4)
return header == b"GGUF"
def resolve_hf_gguf_reference(
model: str, revision: Optional[str] = None
) -> Optional[str]:
"""Download a .gguf named by Hub reference and return its local path.
owner/repo/path/inside/repo.gguf -> exactly that file
owner/repo:QUANT_TYPE -> the only matching quantization
owner/repo -> the only .gguf in the repo
"""
from sglang.srt.utils import is_remote_url
if not model or os.path.exists(model) or is_remote_url(model):
return None
from huggingface_hub import hf_hub_download
if ":" in model:
repo_id, _, quant_type = model.rpartition(":")
if repo_id.count("/") != 1 or not quant_type:
return None
from huggingface_hub import HfApi
files = [
sibling.rfilename
for sibling in HfApi().repo_info(repo_id, revision=revision).siblings
]
suffix = f"-{quant_type}.gguf"
candidates = [filename for filename in files if filename.endswith(suffix)]
if not candidates:
available = sorted(
filename for filename in files if filename.endswith(".gguf")
)
raise ValueError(
f"No file matching quant type {quant_type!r} in {repo_id}. "
f"Available GGUF files: {available}"
)
if len(candidates) > 1:
raise ValueError(
f"Quant type {quant_type!r} is ambiguous in {repo_id}: "
f"{sorted(candidates)}. Pass the full owner/repo/path/file.gguf "
"reference instead."
)
return hf_hub_download(repo_id, candidates[0], revision=revision)
parts = model.strip("/").split("/")
if len(parts) < 2:
return None
if len(parts) > 2 and model.endswith(".gguf"):
repo_id = "/".join(parts[:2])
filename = "/".join(parts[2:])
return hf_hub_download(repo_id, filename, revision=revision)
if len(parts) != 2:
return None
from huggingface_hub import HfApi
try:
files = [
s.rfilename for s in HfApi().repo_info(model, revision=revision).siblings
]
except Exception:
return None
if any(f == "config.json" for f in files):
return None
candidates = [f for f in files if f.endswith(".gguf")]
if not candidates:
return None
if len(candidates) > 1:
listing = "\n ".join(f"{model}/{f}" for f in sorted(candidates))
raise ValueError(
f"{model} contains {len(candidates)} .gguf files; name the one to "
f"serve:\n {listing}"
)
return hf_hub_download(model, candidates[0], revision=revision)
def gguf_sidecar_dir(
gguf_path: Union[str, os.PathLike], sentinel: str
) -> Optional[Path]:
"""Directory containing *sentinel* next to a .gguf file, if there is one."""
directory = Path(gguf_path).parent
return directory if (directory / sentinel).is_file() else None
# ---------------------------------------------------------------------------
# Rope / text config helpers
# ---------------------------------------------------------------------------
def get_rope_config(config):
"""Get (rope_theta, rope_params) from config, supporting both v4 and v5.
Trust-remote-code configs or parent configs passed to sub-models may not
have the v5 ``rope_parameters`` property, so we fall back to the v4-style
``config.rope_theta`` / ``config.rope_scaling`` attributes.
Returns:
(rope_theta, rope_params): In v5, rope_params is the full
rope_parameters dict (which subsumes rope_scaling and includes
rope_theta). In v4, rope_params is the rope_scaling dict or None.
"""
rope_params = getattr(config, "rope_parameters", None)
if rope_params is not None:
rope_theta = rope_params.get("rope_theta", getattr(config, "rope_theta", 10000))
return rope_theta, rope_params
return getattr(config, "rope_theta", 10000), getattr(config, "rope_scaling", None)
def _patch_text_config(parent_config: PretrainedConfig, text_config):
"""Synchronize standard attributes between parent config and text sub-config.
In transformers v5, the "untangle config" refactor removed automatic
inheritance of top-level PretrainedConfig attributes (pad_token_id,
tie_word_embeddings, etc.) from sub-configs. Downstream code expects
these attributes to be present on both configs (some models pass the
parent directly to the language model, others pass the text sub-config),
so we propagate in both directions when an attribute is missing.
(See https://github.com/huggingface/transformers/pull/41541)
"""
_ATTRS_TO_PROPAGATE = [
"pad_token_id",
"bos_token_id",
"eos_token_id",
"tie_word_embeddings",
]
for attr in _ATTRS_TO_PROPAGATE:
parent_has = hasattr(parent_config, attr)
text_has = hasattr(text_config, attr)
if parent_has and not text_has:
setattr(text_config, attr, getattr(parent_config, attr))
elif text_has and not parent_has:
setattr(parent_config, attr, getattr(text_config, attr))
return text_config
def get_hf_text_config(config: PretrainedConfig):
"""Get the "sub" config relevant to llm for multi modal models.
No op for pure text models.
"""
if config.architectures is not None:
class_name = config.architectures[0]
if class_name.startswith("Llava") and class_name.endswith("ForCausalLM"):
# We support non-hf version of llava models, so we do not want to
# read the wrong values from the unused default text_config.
# NOTE(HandH1998): We set `torch_dtype` of config to `torch.float16` for the weights, as
# `torch.float16` is default used for image features in `python/sglang/srt/models/llava.py`.
setattr(config, "dtype", torch.float16)
return config
text_config = None
# Some models (e.g. DeepSeek-OCR) store sub-configs as plain dicts.
# Convert to PretrainedConfig early so hasattr() checks and asserts work.
parent_dtype = getattr(config, "dtype", None)
for _attr in ("text_config", "llm_config", "language_config", "thinker_config"):
_sub = getattr(config, _attr, None)
if isinstance(_sub, dict):
_converted = PretrainedConfig(**_sub)
if getattr(_converted, "dtype", None) is None and parent_dtype is not None:
_converted.dtype = parent_dtype
setattr(config, _attr, _converted)
elif _sub is not None and parent_dtype is not None:
# transformers v5 multimodal configs (e.g. Mistral3Config) carry
# `dtype` only on the top-level config, leaving the sub-configs at
# None. Without this, _get_and_verify_dtype falls back to float32
# and then "auto" downcasts to float16, which overflows the Pixtral
# vision tower on real images and produces NaN features.
if getattr(_sub, "dtype", None) is None:
_sub.dtype = parent_dtype
# Priority: thinker_config > llm_config > language_config > text_config
if hasattr(config, "thinker_config"):
# qwen2.5 omni
thinker_config = config.thinker_config
if hasattr(thinker_config, "text_config"):
setattr(
thinker_config.text_config,
"dtype",
getattr(thinker_config, "dtype", None),
)
text_config = thinker_config.text_config
else:
text_config = thinker_config
elif hasattr(config, "llm_config"):
# PointsV1.5 Chat Model
assert hasattr(config.llm_config, "num_attention_heads")
text_config = config.llm_config
elif hasattr(config, "language_config"):
text_config = config.language_config
elif hasattr(config, "text_config"):
# The code operates under the assumption that text_config should have
# `num_attention_heads` (among others). Assert here to fail early
# if transformers config doesn't align with this assumption.
assert hasattr(config.text_config, "num_attention_heads")
text_config = config.text_config
# Ensure rope_scaling dicts have "type" for remote-code compat (v5).
normalize_rope_scaling_compat(config)
if text_config is not None:
return _patch_text_config(config, text_config)
return config
# ---------------------------------------------------------------------------
# Model-specific helpers
# ---------------------------------------------------------------------------
def _ensure_sub_configs(config: PretrainedConfig, *attr_names: str) -> None:
"""Convert dict-valued sub-configs to proper AutoConfig objects in-place."""
for attr in attr_names:
sub = getattr(config, attr, None)
if sub is not None and isinstance(sub, dict):
setattr(config, attr, AutoConfig.for_model(**sub))
def _is_deepseek_ocr_model(config: PretrainedConfig) -> bool:
# TODO: Remove this workaround once AutoConfig correctly identifies deepseek-ocr.
# Hugging Face's AutoConfig currently misidentifies it as deepseekvl2.
auto_map = getattr(config, "auto_map", None) or {}
return auto_map.get("AutoModel") == "modeling_deepseekocr.DeepseekOCRForCausalLM"
def _is_deepseek_ocr2_model(config: PretrainedConfig) -> bool:
auto_map = getattr(config, "auto_map", None) or {}
return auto_map.get("AutoModel") == "modeling_deepseekocr2.DeepseekOCR2ForCausalLM"
def _override_v_head_dim_if_zero(config: PretrainedConfig, patch: int = 128) -> None:
patched = False
for attr in ("text_config", "language_config"):
sub = getattr(config, attr, None)
if sub is None:
continue
if isinstance(sub, dict):
if sub.get("v_head_dim") == 0:
sub["v_head_dim"] = patch
patched = True
elif getattr(sub, "v_head_dim", None) == 0:
sub.v_head_dim = patch
patched = True
if patched:
logger.warning(
f"Overriding v_head_dim from 0 to {patch} to avoid potential issues."
)
# ---------------------------------------------------------------------------
# Context length / generation config / sparse attention
# ---------------------------------------------------------------------------
# Models don't use the same configuration key for determining the maximum
# context length. Store them here so we can sanely check them.
# NOTE: The ordering here is important. Some models have two of these and we
# have a preference for which value gets used.
CONTEXT_LENGTH_KEYS = [
"max_sequence_length",
"seq_length",
"max_seq_len",
"model_max_length",
"max_position_embeddings",
]
def get_context_length(config):
"""Get the context length of a model from a huggingface model configs."""
text_config = config
rope_scaling = getattr(text_config, "rope_scaling", None)
if rope_scaling:
rope_scaling_factor = rope_scaling.get("factor", 1)
if "original_max_position_embeddings" in rope_scaling:
rope_scaling_factor = 1
if rope_scaling.get("rope_type", None) == "llama3":
rope_scaling_factor = 1
else:
rope_scaling_factor = 1
for key in CONTEXT_LENGTH_KEYS:
val = getattr(text_config, key, None)
if val is not None:
return int(rope_scaling_factor * val)
return 2048
@lru_cache_frozenset(maxsize=32)
def get_generation_config(
model: str,
trust_remote_code: bool,
revision: Optional[str] = None,
**kwargs,
):
if check_gguf_file(model):
sidecar = gguf_sidecar_dir(model, "generation_config.json")
if sidecar is not None:
model = str(sidecar)
else:
from .gguf_native import (
build_gguf_generation_config,
has_native_gguf_support,
)
if has_native_gguf_support(model):
return build_gguf_generation_config(model)
try:
return GenerationConfig.from_pretrained(
model, trust_remote_code=trust_remote_code, revision=revision, **kwargs
)
except (FileNotFoundError, OSError) as e:
# A missing generation_config.json is normal for many checkpoints and
# is surfaced by HF as a generic OSError (not FileNotFoundError). Treat
# it as benign — proceed without a generation config, at DEBUG level so
# normal startup logs stay quiet.
logger.debug(
"No generation config for %s: %s. Proceeding without it.",
model,
e,
)
return None
# Qwen-1M related
def get_sparse_attention_config(
model: str,
sparse_attention_config_filename: str = "sparse_attention_config.json",
) -> Dict[str, Any]:
is_local = os.path.isdir(model)
if not is_local:
model = download_from_hf(model, allow_patterns=["*.json"])
config_file = os.path.join(model, sparse_attention_config_filename)
if not os.path.exists(config_file):
return {}
with open(config_file) as f:
config = json.load(f)
return config
# ---------------------------------------------------------------------------
# Tokenizer / processor helpers
# ---------------------------------------------------------------------------
# Some models don't have an available processor, e.g.: InternVL
def get_tokenizer_from_processor(processor):
from transformers import PreTrainedTokenizerBase
if isinstance(processor, PreTrainedTokenizerBase):
return processor
return processor.tokenizer
# Turn-final markers that some checkpoints ship without EOS metadata:
# <|eom_id|> (Llama-3 tool use), <|content_model_end_sampling|> (Inkling,
# whose bundled tokenizer config leaves eos_token unset), and
# <|ifm|im_end|> (some K2 Horizon checkpoints, notably 0.9B, name only
# <|endoftext|> as EOS).
_ADDITIONAL_STOP_TOKEN_TEXTS = (
"<|eom_id|>",
"<|content_model_end_sampling|>",
"<|ifm|im_end|>",
)
def attach_additional_stop_token_ids(tokenizer):
added = tokenizer.get_added_vocab()
stop_ids = {added[text] for text in _ADDITIONAL_STOP_TOKEN_TEXTS if text in added}
tokenizer.additional_stop_token_ids = stop_ids or None