chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Upgrade transformers==5.3.0 (#17784)

Browse Source 
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>
Co-authored-by: Alison Shao <alisonshao@mac.lan>
Co-authored-by: Mick <mickjagger19@icloud.com>
This commit is contained in:
Xinyuan Tong
2026-03-18 20:50:43 +00:00
committed by GitHub
parent e5750a572c
commit d1e95af282
95 changed files with 1134 additions and 341 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
python/sglang/srt/configs/internvl.py
View File

@@ -593,7 +593,6 @@ class InternLM2Tokenizer(PreTrainedTokenizer):
current_sub_tokens.append(token)
prev_is_special = False
out_string += self.sp_model.decode(current_sub_tokens)
out_string = self.clean_up_tokenization(out_string)
out_string = self._maybe_add_prefix_space(tokens=tokens, decoded=out_string)
return out_string[1:]

56
python/sglang/srt/configs/model_config.py
View File

@@ -51,10 +51,20 @@ class ModelImpl(str, Enum):
MINDSPORE = "mindspore"
def is_deepseek_nsa(config: PretrainedConfig) -> bool:
def is_deepseek_nsa(config) -> bool:
architectures = (
config.get("architectures")
if isinstance(config, dict)
else getattr(config, "architectures", None)
)
index_topk = (
config.get("index_topk")
if isinstance(config, dict)
else getattr(config, "index_topk", None)
)
return (
config.architectures is not None
and config.architectures[0]
architectures is not None
and architectures[0]
in [
"DeepseekV3ForCausalLM",
"DeepseekV32ForCausalLM",
@@ -63,7 +73,7 @@ def is_deepseek_nsa(config: PretrainedConfig) -> bool:
"PixtralForConditionalGeneration",
"GlmMoeDsaForCausalLM",
]
and getattr(config, "index_topk", None) is not None
and index_topk is not None
)
@@ -457,10 +467,9 @@ class ModelConfig:
or "default"
)
if rope_type != "default":
mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
scaling_factor = rope_scaling["factor"]
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
self.scaling = self.scaling * mscale * mscale
self.scaling = compute_mla_mscale_scaling(
rope_scaling, self.scaling
)
elif "MiniCPM3ForCausalLM" in self.hf_config.architectures:
self.head_dim = 128
self.attention_arch = AttentionArch.MLA
@@ -501,12 +510,9 @@ class ModelConfig:
# Handle rope scaling with yarn
self.scaling = 1 / math.sqrt(self.qk_nope_head_dim + self.qk_rope_head_dim)
if self.hf_config.rope_scaling:
mscale_all_dim = self.hf_config.rope_scaling.get(
"mscale_all_dim", False
self.scaling = compute_mla_mscale_scaling(
self.hf_config.rope_scaling, self.scaling
)
scaling_factor = self.hf_config.rope_scaling["factor"]
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
self.scaling = self.scaling * mscale * mscale
elif "SarvamMLAForCausalLM" in self.hf_config.architectures:
self.head_dim = (
self.hf_config.qk_nope_head_dim + self.hf_config.qk_rope_head_dim
@@ -518,12 +524,9 @@ class ModelConfig:
self.v_head_dim = self.hf_config.v_head_dim
self.scaling = 1 / math.sqrt(self.qk_nope_head_dim + self.qk_rope_head_dim)
if self.hf_config.rope_scaling:
mscale_all_dim = self.hf_config.rope_scaling.get(
"mscale_all_dim", False
self.scaling = compute_mla_mscale_scaling(
self.hf_config.rope_scaling, self.scaling
)
scaling_factor = self.hf_config.rope_scaling["factor"]
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
self.scaling = self.scaling * mscale * mscale
else:
if (
"MistralModel" in self.hf_config.architectures
@@ -1410,6 +1413,23 @@ def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
return 0.1 * mscale * math.log(scale) + 1.0
def compute_mla_mscale_scaling(rope_scaling: dict, base_scaling: float) -> float:
"""Compute MLA attention scaling factor from rope_scaling with mscale.
Used by DeepSeek, BailingMoe, SarvamMLA and similar MLA models.
Warns if 'factor' is missing from rope_scaling (common in v5 configs).
"""
mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
if "factor" not in rope_scaling:
logger.warning(
"rope_scaling missing 'factor', defaulting to 1.0. "
"Check model accuracy.",
)
scaling_factor = rope_scaling.get("factor", 1.0)
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
return base_scaling * mscale * mscale
def is_hybrid_swa_model(model_architectures: List[str]):
hybrid_swa_archs = {