iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[Model][Frontend] Add Ling 3.0 Flash BF16, MTP, and parser support (#51045)

Browse Source 
Signed-off-by: zexplorerhj <zhjoneson@163.com>
This commit is contained in:
zexplorerhj
2026-08-06 00:38:28 +08:00
committed by GitHub
parent 66b3c0e61f
commit d4da0c55af
15 changed files with 2070 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
benchmarks/kernels/benchmark_moe.py
+1
View File
@@ -777,6 +777,7 @@ def get_model_params(config):
intermediate_size = config.moe_intermediate_size
hidden_size = config.hidden_size
elif architecture in (
"BailingMoeV3ForCausalLM",
"Qwen2MoeForCausalLM",
"Qwen3MoeForCausalLM",
"Qwen3NextForCausalLM",
docs/models/supported_models.md
+1
View File
@@ -348,6 +348,7 @@ th {
| `BailingMoeForCausalLM` | Ling | `inclusionAI/Ling-lite-1.5`, `inclusionAI/Ling-plus`, etc. | ✅︎ | ✅︎ |
| `BailingMoeV2ForCausalLM` | Ling | `inclusionAI/Ling-mini-2.0`, etc. | ✅︎ | ✅︎ |
| `BailingMoeV2_5ForCausalLM` | Ling | `inclusionAI/Ling-2.5-1T`, `inclusionAI/Ring-2.5-1T` | | ✅︎ |
| `BailingMoeV3ForCausalLM` | Ling | `inclusionAI/Ling-3.0-flash` | | ✅︎ |
| `BloomForCausalLM` | BLOOM, BLOOMZ, BLOOMChat | `bigscience/bloom`, `bigscience/bloomz`, etc. | | ✅︎ |
| `ChatGLMModel`, `ChatGLMForConditionalGeneration` | ChatGLM | `zai-org/chatglm2-6b`, `zai-org/chatglm3-6b`, `thu-coai/ShieldLM-6B-chatglm3`, etc. | ✅︎ | ✅︎ |
| `CohereForCausalLM`, `Cohere2ForCausalLM` | Command-R, Command-A | `CohereLabs/c4ai-command-r-v01`, `CohereLabs/c4ai-command-r7b-12-2024`, `CohereLabs/c4ai-command-a-03-2025`, `CohereLabs/command-a-reasoning-08-2025`, etc. | ✅︎ | ✅︎ |
tests/models/registry.py
+8
View File
@@ -213,6 +213,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"BailingMoeV2_5ForCausalLM": _HfExamplesInfo(
"inclusionAI/Ring-2.5-1T", trust_remote_code=True
),
"BailingMoeV3ForCausalLM": _HfExamplesInfo(
"inclusionAI/Ling-3.0-flash", trust_remote_code=True
),
"BloomForCausalLM": _HfExamplesInfo(
"bigscience/bloom-560m", {"1b": "bigscience/bloomz-1b1"}
),
@@ -1614,6 +1617,11 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
trust_remote_code=True,
is_available_online=False,
),
"BailingMoeV3MTPModel": _HfExamplesInfo(
"inclusionAI/Ling-3.0-flash",
speculative_model="inclusionAI/Ling-3.0-flash",
trust_remote_code=True,
),
"DeepSeekMTPModel": _HfExamplesInfo(
"luccafong/deepseek_mtp_main_random",
speculative_model="luccafong/deepseek_mtp_draft_random",
tests/parser/engine/test_ling3.py
+105
View File
@@ -0,0 +1,105 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import json
from tests.parser.engine.replay_harness import MockTokenizer, _test_request
from vllm.reasoning import ReasoningParserManager
from vllm.tool_parsers import ToolParserManager
def _tokenizer() -> MockTokenizer:
vocab = {
"<think>": 1,
"</think>": 2,
"<tool_call>": 3,
"</tool_call>": 4,
"<arg_key>": 5,
"</arg_key>": 6,
"<arg_value>": 7,
"</arg_value>": 8,
}
return MockTokenizer(vocab=vocab, tokens=[])
def test_ling3_registered():
assert ReasoningParserManager.get_reasoning_parser("ling3") is not None
assert ToolParserManager.get_tool_parser("ling3") is not None
def test_ling3_defaults_thinking_on():
parser_cls = ReasoningParserManager.get_reasoning_parser("ling3")
parser = parser_cls(_tokenizer())
assert parser.reasoning_start_str == "<think>"
assert parser.reasoning_end_str == "</think>"
reasoning, content = parser.extract_reasoning(
"<think>reason</think>answer", _test_request()
)
assert reasoning == "reason"
assert content == "answer"
def test_ling3_disable_thinking_keeps_reasoning_as_content():
parser_cls = ReasoningParserManager.get_reasoning_parser("ling3")
parser = parser_cls(
_tokenizer(),
chat_template_kwargs={"enable_thinking": False},
)
reasoning, content = parser.extract_reasoning(
"<think>reason</think>answer", _test_request()
)
assert reasoning is None
assert content == "<think>reason</think>answer"
def test_ling3_enable_thinking_keeps_open_reasoning_as_content():
parser_cls = ReasoningParserManager.get_reasoning_parser("ling3")
parser = parser_cls(
_tokenizer(),
chat_template_kwargs={"enable_thinking": True},
)
reasoning, content = parser.extract_reasoning(
"<think>only reasoning", _test_request()
)
assert reasoning is None
assert content == "only reasoning"
def test_ling3_tool_call_without_newline():
parser_cls = ToolParserManager.get_tool_parser("ling3")
parser = parser_cls(_tokenizer())
result = parser.extract_tool_calls(
"<tool_call>get_weather<arg_key>city</arg_key>"
"<arg_value>Beijing</arg_value></tool_call>",
request=_test_request(),
)
assert result.tools_called
assert result.content is None
assert result.tool_calls[0].function.name == "get_weather"
assert json.loads(result.tool_calls[0].function.arguments) == {
"city": "Beijing",
}
def test_ling3_tool_call_without_args():
parser_cls = ToolParserManager.get_tool_parser("ling3")
parser = parser_cls(_tokenizer())
result = parser.extract_tool_calls(
"<tool_call>ping</tool_call>",
request=_test_request(),
)
assert result.tools_called
assert result.content is None
assert result.tool_calls[0].function.name == "ping"
assert json.loads(result.tool_calls[0].function.arguments) == {}
vllm/config/speculative.py
+12 -1
View File
@@ -48,6 +48,7 @@ MTPModelTypes = Literal[
"qwen3_next_mtp",
"qwen3_5_mtp",
"longcat_flash_mtp",
"bailing_hybrid_v3_mtp",
"minimax_m3_mtp",
"bailing_hybrid_mtp",
"mtp",
@@ -489,7 +490,9 @@ class SpeculativeConfig:
)
architectures = getattr(hf_config, "architectures", []) or []
if (
if initial_architecture == "BailingMoeV3ForCausalLM":
hf_config.model_type = "bailing_hybrid_v3_mtp"
elif (
hf_config.model_type == "bailing_hybrid"
or "BailingMoeV2_5ForCausalLM" in architectures
):
@@ -502,6 +505,14 @@ class SpeculativeConfig:
"architectures": ["BailingMoeV25MTPModel"],
}
)
if hf_config.model_type == "bailing_hybrid_v3_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update(
{
"n_predict": n_predict,
"architectures": ["BailingMoeV3MTPModel"],
}
)
if hf_config.model_type == "exaone_moe":
hf_config.model_type = "exaone_moe_mtp"
vllm/model_executor/layers/fused_moe/configs/E=512,N=192,device_name=NVIDIA_H20-3e.json
+147
View File
@@ -0,0 +1,147 @@
{
"triton_version": "3.6.0",
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"512": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
}
}
vllm/model_executor/models/bailing_moe_v3.py
+1289
View File
File diff suppressed because it is too large Load Diff
vllm/model_executor/models/bailing_moe_v3_mtp.py
+389
View File
@@ -0,0 +1,389 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Inference-only Bailing MoE v3 MTP model."""
from collections.abc import Iterable
import torch
import torch.nn as nn
from transformers.configuration_utils import PretrainedConfig
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
from vllm.distributed import get_pp_group
from vllm.model_executor.layers.fused_moe import (
fused_moe_make_expert_params_mapping,
)
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.bailing_moe_v3 import (
BailingMoeV3MLAAttention,
BailingMoeV3MoE,
)
from vllm.model_executor.models.interfaces import SupportsPP
from vllm.model_executor.models.utils import (
PPMissingLayer,
is_pp_missing_parameter,
maybe_prefix,
)
from vllm.sequence import IntermediateTensors
def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:
speculative_config = vllm_config.speculative_config
if speculative_config is not None:
draft_model_config = speculative_config.draft_model_config
if draft_model_config is not None:
return draft_model_config.hf_config
return vllm_config.model_config.hf_config
class BailingMoeV3MTPSharedHead(nn.Module):
def __init__(
self,
config: PretrainedConfig,
prefix: str,
vllm_config: VllmConfig,
) -> None:
super().__init__()
self.head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=vllm_config.quant_config,
prefix=maybe_prefix(prefix, "head"),
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
return hidden_states
class BailingMoeV3MultiTokenPredictorLayer(nn.Module):
def __init__(
self,
*,
vllm_config: VllmConfig,
prefix: str,
layer_id: int,
) -> None:
super().__init__()
config = _get_draft_hf_config(vllm_config)
self.config = config
self.layer_id = layer_id
self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.self_attn = BailingMoeV3MLAAttention(
config,
quant_config=vllm_config.quant_config,
layer_id=layer_id,
prefix=maybe_prefix(prefix, "self_attn"),
cache_config=vllm_config.cache_config,
)
self.post_attention_layernorm = RMSNorm(
config.hidden_size, eps=config.rms_norm_eps
)
self.mlp = BailingMoeV3MoE(
config,
quant_config=vllm_config.quant_config,
layer_id=layer_id,
prefix=maybe_prefix(prefix, "mlp"),
)
self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.shared_head = BailingMoeV3MTPSharedHead(
config,
maybe_prefix(prefix, "shared_head"),
vllm_config,
)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
previous_hidden_states: torch.Tensor,
inputs_embeds: torch.Tensor | None = None,
spec_step_index: int = 0,
) -> torch.Tensor:
del input_ids, spec_step_index
assert inputs_embeds is not None
inputs_embeds = self.enorm(inputs_embeds)
previous_hidden_states = self.hnorm(previous_hidden_states)
hidden_states = self.eh_proj(
torch.cat([inputs_embeds, previous_hidden_states], dim=-1)
)
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
hidden_states = self.self_attn(hidden_states, positions)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states.to(residual.device)
return self.final_layernorm(hidden_states)
class BailingMoeV3MultiTokenPredictor(nn.Module):
def __init__(
self,
*,
vllm_config: VllmConfig,
prefix: str = "",
) -> None:
super().__init__()
config = _get_draft_hf_config(vllm_config)
self.mtp_start_layer_idx = config.num_hidden_layers
self.num_mtp_layers = config.num_nextn_predict_layers
self.layers = nn.ModuleDict(
{
str(idx): BailingMoeV3MultiTokenPredictorLayer(
vllm_config=vllm_config,
prefix=f"{prefix}.layers.{idx}",
layer_id=idx,
)
for idx in range(
self.mtp_start_layer_idx,
self.mtp_start_layer_idx + self.num_mtp_layers,
)
}
)
if get_pp_group().is_first_rank:
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
prefix=maybe_prefix(prefix, "embed_tokens"),
)
else:
self.embed_tokens = PPMissingLayer()
self.logits_processor = LogitsProcessor(config.vocab_size)
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_tokens(input_ids)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
previous_hidden_states: torch.Tensor,
inputs_embeds: torch.Tensor | None = None,
spec_step_idx: int = 0,
) -> torch.Tensor:
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
current_step_idx = spec_step_idx % self.num_mtp_layers
return self.layers[str(self.mtp_start_layer_idx + current_step_idx)](
input_ids,
positions,
previous_hidden_states,
inputs_embeds,
current_step_idx,
)
def compute_logits(
self,
hidden_states: torch.Tensor,
spec_step_idx: int = 0,
lm_head: nn.Module | None = None,
) -> torch.Tensor:
current_step_idx = spec_step_idx % self.num_mtp_layers
mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)]
head = lm_head if lm_head is not None else mtp_layer.shared_head.head
return self.logits_processor(head, mtp_layer.shared_head(hidden_states))
@support_torch_compile
class BailingMoeV3MTPModel(nn.Module, SupportsPP):
packed_modules_mapping = {
"gate_up_proj": ["gate_proj", "up_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
}
def __init__(
self,
*,
vllm_config: VllmConfig,
prefix: str = "",
) -> None:
super().__init__()
self.config = _get_draft_hf_config(vllm_config)
self.model = BailingMoeV3MultiTokenPredictor(
vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"),
)
self.lm_head: nn.Module | None = None
def share_lm_head(self, lm_head: nn.Module) -> None:
self.lm_head = lm_head
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.embed_input_ids(input_ids)
def forward(
self,
input_ids: torch.Tensor | None,
positions: torch.Tensor,
hidden_states: torch.Tensor,
intermediate_tensors: IntermediateTensors | None = None,
inputs_embeds: torch.Tensor | None = None,
spec_step_idx: int = 0,
) -> torch.Tensor:
del intermediate_tensors
return self.model(
input_ids,
positions,
hidden_states,
inputs_embeds,
spec_step_idx,
)
def compute_logits(
self,
hidden_states: torch.Tensor,
spec_step_idx: int = 0,
) -> torch.Tensor:
return self.model.compute_logits(hidden_states, spec_step_idx, self.lm_head)
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
return fused_moe_make_expert_params_mapping(
self,
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.num_experts,
)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
stacked_params_mapping = [
(".fused_qkv_a_proj", ".q_a_proj", 0),
(".fused_qkv_a_proj", ".kv_a_proj_with_mqa", 1),
(".gate_up_proj", ".gate_proj", 0),
(".gate_up_proj", ".up_proj", 1),
]
expert_params_mapping = list(self.get_expert_mapping())
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: set[str] = set()
loaded_mtp_layers: set[int] = set()
def load_param(
name: str,
loaded_weight: torch.Tensor,
shard_id=None,
) -> bool:
name = maybe_remap_kv_scale_name(name, params_dict)
if name is None:
return False
if name not in params_dict or is_pp_missing_parameter(name, self):
return False
param = params_dict[name]
weight_loader = getattr(param, "weight_loader", default_weight_loader)
if shard_id is None:
weight_loader(param, loaded_weight)
elif isinstance(shard_id, int):
weight_loader(param, loaded_weight, shard_id)
else:
weight_loader(
param,
loaded_weight,
name,
expert_id=shard_id[0],
shard_id=shard_id[1],
)
loaded_params.add(name)
return True
def get_spec_layer_idx(name: str) -> int | None:
if not name.startswith("model.layers."):
return None
try:
layer_idx = int(name.split("model.layers.", 1)[1].split(".", 1)[0])
except (IndexError, ValueError):
return None
mtp_idx = layer_idx - self.config.num_hidden_layers
if 0 <= mtp_idx < self.config.num_nextn_predict_layers:
return layer_idx
return None
def normalize_name(name: str) -> str:
name = name.replace(".attention.", ".self_attn.")
return name.replace(
"mlp.gate.e_score_correction_bias",
"mlp.gate.expert_bias",
)
def load_lm_head(loaded_weight: torch.Tensor) -> None:
for layer_idx in range(
self.model.mtp_start_layer_idx,
self.model.mtp_start_layer_idx + self.model.num_mtp_layers,
):
name = f"model.layers.{layer_idx}.shared_head.head.weight"
load_param(name, loaded_weight)
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
if name == "model.word_embeddings.weight":
load_param("model.embed_tokens.weight", loaded_weight)
continue
if name == "lm_head.weight":
load_lm_head(loaded_weight)
continue
spec_layer = get_spec_layer_idx(name)
if spec_layer is None:
continue
name = normalize_name(name)
loaded = False
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
if "mlp.experts." in name and name not in params_dict:
continue
mapped_name = name.replace(weight_name, param_name)
if load_param(mapped_name, loaded_weight, shard_id):
loaded = True
break
if loaded:
loaded_mtp_layers.add(spec_layer)
continue
if "mlp.experts" in name:
for mapping in expert_params_mapping:
param_name, weight_name, expert_id, shard_id = mapping
if weight_name not in name:
continue
mapped_name = name.replace(weight_name, param_name)
if load_param(mapped_name, loaded_weight, (expert_id, shard_id)):
loaded = True
break
if loaded:
loaded_mtp_layers.add(spec_layer)
continue
if load_param(name, loaded_weight):
loaded_mtp_layers.add(spec_layer)
for layer_idx in range(
self.model.mtp_start_layer_idx,
self.model.mtp_start_layer_idx + self.model.num_mtp_layers,
):
if layer_idx not in loaded_mtp_layers:
raise ValueError(
f"Bailing V3 MTP speculative decoding layer {layer_idx} "
"weights are missing from checkpoint. Use a checkpoint that "
"includes MTP layer weights, or disable speculative decoding."
)
return loaded_params
vllm/model_executor/models/registry.py
+2
View File
@@ -79,6 +79,7 @@ _TEXT_GENERATION_MODELS = {
"BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
"BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
"BailingMoeV2_5ForCausalLM": ("bailing_moe_linear", "BailingMoeV25ForCausalLM"),
"BailingMoeV3ForCausalLM": ("bailing_moe_v3", "BailingMoeV3ForCausalLM"),
"BloomForCausalLM": ("bloom", "BloomForCausalLM"),
"ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
"ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
@@ -641,6 +642,7 @@ _SPECULATIVE_DECODING_MODELS = {
"EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
"DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
"DeepSeekV4MTPModel": ("vllm.models.deepseek_v4", "DeepSeekV4MTP"),
"BailingMoeV3MTPModel": ("bailing_moe_v3_mtp", "BailingMoeV3MTPModel"),
"MiniMaxM3MTP": ("vllm.models.minimax_m3", "MiniMaxM3MTP"),
"BailingMoeV25MTPModel": ("bailing_moe_mtp", "BailingMoeV25MTPModel"),
"InklingMTPModel": ("vllm.models.inkling", "InklingMTP"),
vllm/parser/ling3.py
+83
View File
@@ -0,0 +1,83 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Ling3 parser for reasoning and tool calls.
Ling3 uses the same XML tool-call format as GLM-4.7 and defaults thinking on.
It follows the GLM-style ``<think>`` / ``</think>`` reasoning format and
treats ``<tool_call>`` as an implicit reasoning terminator.
"""
from __future__ import annotations
from dataclasses import replace
from typing import TYPE_CHECKING
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.parser.engine.adapters import make_adapters
from vllm.parser.engine.parser_engine import ParserEngine
from vllm.parser.glm47_moe import (
THINK_END,
THINK_START,
Glm47MoeParser,
glm47_moe_config,
)
if TYPE_CHECKING:
from vllm.tokenizers import TokenizerLike
from vllm.tool_parsers.abstract_tool_parser import Tool
class Ling3Parser(Glm47MoeParser):
"""Ling3 parser backed by the GLM XML parser engine."""
def __init__(
self,
tokenizer: TokenizerLike,
tools: list[Tool] | None = None,
**kwargs,
) -> None:
chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {}
thinking = chat_kwargs.get("thinking", None)
enable_thinking = chat_kwargs.get("enable_thinking", None)
self.thinking_enabled = (
True
if thinking is None and enable_thinking is None
else bool(thinking) or bool(enable_thinking)
)
parser_config = replace(
glm47_moe_config(thinking=self.thinking_enabled),
name="ling3",
)
kwargs.setdefault(
"parser_engine_config",
parser_config,
)
ParserEngine.__init__(self, tokenizer, tools, **kwargs)
@property
def reasoning_start_str(self) -> str:
return THINK_START
@property
def reasoning_end_str(self) -> str:
return THINK_END
def extract_reasoning(
self,
model_output: str,
request: ChatCompletionRequest | ResponsesRequest,
) -> tuple[str | None, str | None]:
if not self.thinking_enabled:
return None, model_output
reasoning, content = super().extract_reasoning(model_output, request)
if reasoning and not content and "<tool_call>" not in model_output:
return None, reasoning
return reasoning, content
(
Ling3ParserReasoningAdapter,
Ling3ParserToolAdapter,
) = make_adapters(Ling3Parser)
vllm/reasoning/__init__.py
+4
View File
@@ -60,6 +60,10 @@ _REASONING_PARSERS_TO_REGISTER = {
"glm47_moe_reasoning_parser",
"Glm47MoeParserReasoningAdapter",
),
"ling3": (
"ling3_reasoning_parser",
"Ling3ParserReasoningAdapter",
),
"openai_gptoss": (
"gptoss_reasoning_parser",
"GptOssReasoningParser",
vllm/reasoning/ling3_reasoning_parser.py
+6
View File
@@ -0,0 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.parser.ling3 import Ling3ParserReasoningAdapter
__all__ = ["Ling3ParserReasoningAdapter"]
vllm/tool_parsers/__init__.py
+4
View File
@@ -58,6 +58,10 @@ _TOOL_PARSERS_TO_REGISTER = {
"glm47_moe_tool_parser",
"Glm47MoeModelToolParser",
),
"ling3": (
"ling3_tool_parser",
"Ling3ToolParser",
),
"granite-20b-fc": (
"granite_20b_fc_tool_parser",
"Granite20bFCToolParser",
vllm/tool_parsers/ling3_tool_parser.py
+11
View File
@@ -0,0 +1,11 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
from vllm.parser.ling3 import Ling3ParserToolAdapter
class Ling3ToolParser(Ling3ParserToolAdapter): # type: ignore[valid-type, misc]
supports_required_and_named = False
structural_tag_model = "glm_4_7"
vllm/transformers_utils/model_arch_config_convertor.py
+8
View File
@@ -289,6 +289,7 @@ class ModelArchConfigConvertorBase:
"pangu_ultra_moe_mtp",
"bailing_hybrid",
"bailing_hybrid_mtp",
"bailing_hybrid_v3_mtp",
):
# check is deepseek_v4 model
if hasattr(self.hf_text_config, "compress_ratios"):
@@ -571,6 +572,12 @@ class BailingHybridMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
return getattr(self.hf_text_config, "num_nextn_predict_layers", 0)
class BailingHybridV3MTPModelArchConfigConvertor(
BailingHybridMTPModelArchConfigConvertor
):
pass
class Qwen3_5MTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def get_num_hidden_layers(self) -> int:
return getattr(self.hf_text_config, "mtp_num_hidden_layers", 0)
@@ -693,6 +700,7 @@ MODEL_ARCH_CONFIG_CONVERTORS = {
"moss_audio": MossAudioModelArchConfigConvertor,
"mpt": MPTModelArchConfigConvertor,
"nemotron-nas": NemotronNasModelArchConfigConvertor,
"bailing_hybrid_v3_mtp": BailingHybridV3MTPModelArchConfigConvertor,
"pangu_ultra_moe_mtp": PanguUltraMoeMTPModelArchConfigConvertor,
"qwen3_5_mtp": Qwen3_5MTPModelArchConfigConvertor,
"qwen3_next_mtp": Qwen3NextMTPModelArchConfigConvertor,