[Feature] Support MTP speculative decoding for Bailing hybrid models (#44880)

Signed-off-by: zc02384840 <zc02384840@antgroup.com>
Co-authored-by: zc02384840 <zc02384840@antgroup.com>
This commit is contained in:
alex101-ops
2026-07-06 10:38:50 +08:00
committed by GitHub
co-authored by zc02384840
parent 95a248faed
commit f2aaf59151
11 changed files with 1200 additions and 6 deletions
@@ -30,11 +30,10 @@ from vllm.model_executor.layers.mamba.linear.base import LinearAttention
from vllm.model_executor.layers.mamba.linear.minimax_linear_attn import (
MiniMaxText01LinearAttention,
MiniMaxText01LinearKernel,
clear_linear_attention_cache_for_new_sequences,
linear_attention_decode,
linear_attention_prefill_and_mix,
)
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.triton_utils import tl, triton
from vllm.v1.attention.backends.linear_attn import LinearAttentionMetadata
@@ -57,6 +56,290 @@ def _build_rope_parameters(config: PretrainedConfig) -> dict | None:
return rope_parameters or None
def clear_linear_attention_cache_for_new_sequences(
kv_cache: torch.Tensor,
state_indices_tensor: torch.Tensor,
attn_metadata: LinearAttentionMetadata,
) -> None:
num_prefills = getattr(attn_metadata, "num_prefills", 0)
if num_prefills <= 0:
return
num_decodes = getattr(attn_metadata, "num_decodes", 0)
prefill_state_indices = getattr(attn_metadata, "state_indices_tensor_p", None)
for prefill_idx in range(num_prefills):
if num_decodes + prefill_idx + 1 >= len(attn_metadata.query_start_loc):
break
q_start = attn_metadata.query_start_loc[num_decodes + prefill_idx]
q_end = attn_metadata.query_start_loc[num_decodes + prefill_idx + 1]
query_len = q_end - q_start
context_len = attn_metadata.seq_lens[num_decodes + prefill_idx] - query_len
if context_len == 0:
if prefill_state_indices is not None:
block_to_clear = prefill_state_indices[prefill_idx]
else:
block_to_clear = state_indices_tensor[num_decodes + prefill_idx]
kv_cache[block_to_clear, ...] = 0
def linear_attention_prefill_and_mix(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
kv_cache: torch.Tensor,
state_indices_tensor: torch.Tensor,
attn_metadata: LinearAttentionMetadata,
slope_rate: torch.Tensor,
block_size: int,
decode_fn,
prefix_fn,
layer_idx: int | None = None,
) -> torch.Tensor:
hidden = []
req_offset = getattr(attn_metadata, "num_decodes", 0)
prefill_state_indices = getattr(attn_metadata, "state_indices_tensor_p", None)
for _prefill_idx in range(getattr(attn_metadata, "num_prefills", 0)):
if req_offset + _prefill_idx + 1 >= len(attn_metadata.query_start_loc):
break
if prefill_state_indices is not None and _prefill_idx >= len(
prefill_state_indices
):
break
if prefill_state_indices is None and _prefill_idx >= len(state_indices_tensor):
break
_start = attn_metadata.query_start_loc[req_offset + _prefill_idx]
_end = attn_metadata.query_start_loc[req_offset + _prefill_idx + 1]
if prefill_state_indices is not None:
slot_id = prefill_state_indices[_prefill_idx]
else:
slot_id = state_indices_tensor[req_offset + _prefill_idx]
qs = q[_start:_end].transpose(0, 1).contiguous()
ks = k[_start:_end].transpose(0, 1).contiguous()
vs = v[_start:_end].transpose(0, 1).contiguous()
slice_layer_cache = kv_cache[slot_id, ...]
out_slice = prefix_fn(
qs,
ks,
vs,
slice_layer_cache,
slope_rate,
block_size,
layer_idx=layer_idx,
)
hidden.append(out_slice.contiguous())
if attn_metadata.num_decode_tokens > 0:
hidden_decode = decode_fn(
q, k, v, kv_cache, state_indices_tensor, attn_metadata
)
hidden.insert(0, hidden_decode)
if not hidden:
return torch.empty((0, q.size(1) * q.size(2)), device=q.device, dtype=q.dtype)
hidden = torch.concat(hidden, dim=0).contiguous()
return hidden
@triton.jit
def _bailing_linear_attn_decode_spec_step_kernel(
q_ptr,
k_ptr,
v_ptr,
kv_cache_ptr,
slope_rate_ptr,
state_indices_ptr,
query_start_loc_ptr,
num_accepted_tokens_ptr,
output_ptr,
q_start: tl.constexpr,
D: tl.constexpr,
q_b_stride,
q_h_stride,
q_d_stride,
k_b_stride,
k_h_stride,
k_d_stride,
v_b_stride,
v_h_stride,
v_d_stride,
cache_b_stride,
cache_h_stride,
cache_d0_stride,
cache_d1_stride,
state_indices_b_stride,
state_indices_t_stride,
output_b_stride,
output_d_stride,
DRAFT_IDX: tl.constexpr,
STATE_WIDTH: tl.constexpr,
BLOCK_SIZE: tl.constexpr,
):
req_id = tl.program_id(0)
head_id = tl.program_id(1)
block_id = tl.program_id(2)
req_start = tl.load(query_start_loc_ptr + req_id).to(tl.int64)
req_end = tl.load(query_start_loc_ptr + req_id + 1).to(tl.int64)
query_len = req_end - req_start
if query_len <= DRAFT_IDX:
return
dst_slot = tl.load(
state_indices_ptr
+ req_id * state_indices_b_stride
+ DRAFT_IDX * state_indices_t_stride
).to(tl.int64)
if dst_slot == -1:
return
if DRAFT_IDX == 0:
accepted_offset = tl.load(num_accepted_tokens_ptr + req_id).to(tl.int64) - 1
accepted_offset = tl.maximum(accepted_offset, 0)
accepted_offset = tl.minimum(accepted_offset, STATE_WIDTH - 1)
src_slot = tl.load(
state_indices_ptr
+ req_id * state_indices_b_stride
+ accepted_offset * state_indices_t_stride
).to(tl.int64)
else:
src_slot = tl.load(
state_indices_ptr
+ req_id * state_indices_b_stride
+ (DRAFT_IDX - 1) * state_indices_t_stride
).to(tl.int64)
if src_slot == -1:
return
token_idx = req_start - q_start + DRAFT_IDX
qk_offsets = tl.arange(0, D)
v_offsets = tl.arange(0, BLOCK_SIZE) + block_id * BLOCK_SIZE
qk_mask = qk_offsets < D
v_mask = v_offsets < D
kv_mask = qk_mask[:, None] & v_mask[None, :]
q = tl.load(
q_ptr + token_idx * q_b_stride + head_id * q_h_stride + qk_offsets * q_d_stride,
mask=qk_mask,
other=0.0,
)
k = tl.load(
k_ptr + token_idx * k_b_stride + head_id * k_h_stride + qk_offsets * k_d_stride,
mask=qk_mask,
other=0.0,
)
v = tl.load(
v_ptr + token_idx * v_b_stride + head_id * v_h_stride + v_offsets * v_d_stride,
mask=v_mask,
other=0.0,
)
cache_offsets = (
qk_offsets[:, None] * cache_d0_stride + v_offsets[None, :] * cache_d1_stride
)
src_cache_ptr = (
kv_cache_ptr
+ src_slot * cache_b_stride
+ head_id * cache_h_stride
+ cache_offsets
)
dst_cache_ptr = (
kv_cache_ptr
+ dst_slot * cache_b_stride
+ head_id * cache_h_stride
+ cache_offsets
)
slope = tl.load(slope_rate_ptr + head_id)
decay = tl.exp(-slope)
kv_old = tl.load(src_cache_ptr, mask=kv_mask, other=0.0)
kv_new = k[:, None] * v[None, :] + decay * kv_old
output = tl.sum(q[:, None].to(tl.float32) * kv_new, axis=0)
tl.store(dst_cache_ptr, kv_new, mask=kv_mask)
tl.store(
output_ptr
+ token_idx * output_b_stride
+ (head_id * D + v_offsets) * output_d_stride,
output,
mask=v_mask,
)
def bailing_linear_attention_decode_spec(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
kv_cache: torch.Tensor,
slope_rate: torch.Tensor,
state_indices_tensor: torch.Tensor,
query_start_loc: torch.Tensor,
num_accepted_tokens: torch.Tensor,
q_start: int,
q_end: int | None,
slot_start: int,
slot_end: int | None,
block_size: int,
) -> torch.Tensor:
q_decode = q[q_start:q_end]
k_decode = k[q_start:q_end]
v_decode = v[q_start:q_end]
hidden = torch.empty(
(q_decode.shape[0], q.shape[1] * q.shape[2]),
device=q.device,
dtype=q.dtype,
)
hidden.zero_()
state_indices_tensor = state_indices_tensor[slot_start:slot_end]
query_start_loc = query_start_loc.to(device=q.device)
batch_size = state_indices_tensor.shape[0]
num_heads = q_decode.shape[1]
head_dim = q_decode.shape[2]
assert k_decode.shape == (q_decode.shape[0], num_heads, head_dim)
assert v_decode.shape == (q_decode.shape[0], num_heads, head_dim)
state_width = state_indices_tensor.shape[1]
grid = (batch_size, num_heads, triton.cdiv(head_dim, block_size))
for draft_idx in range(state_width):
_bailing_linear_attn_decode_spec_step_kernel[grid](
q_decode,
k_decode,
v_decode,
kv_cache,
slope_rate,
state_indices_tensor,
query_start_loc,
num_accepted_tokens[:batch_size],
hidden,
q_start,
head_dim,
q_decode.stride(0),
q_decode.stride(1),
q_decode.stride(2),
k_decode.stride(0),
k_decode.stride(1),
k_decode.stride(2),
v_decode.stride(0),
v_decode.stride(1),
v_decode.stride(2),
kv_cache.stride(0),
kv_cache.stride(1),
kv_cache.stride(2),
kv_cache.stride(3),
state_indices_tensor.stride(0),
state_indices_tensor.stride(1),
hidden.stride(0),
hidden.stride(1),
DRAFT_IDX=draft_idx,
STATE_WIDTH=state_width,
BLOCK_SIZE=block_size,
)
return hidden
class BailingGroupRMSNormGate(RMSNormGated):
def __init__(
self,
@@ -208,6 +491,13 @@ class BailingMoELinearAttention(LinearAttention):
raise ValueError(f"Duplicate layer name: {prefix}")
compilation_config.static_forward_context[prefix] = self
def get_attn_backend(self):
from vllm.v1.attention.backends.linear_attn import (
BailingLinearAttentionBackend,
)
return BailingLinearAttentionBackend
@staticmethod
def weight_direct_load(param: torch.Tensor, loaded_weight: torch.Tensor) -> None:
"""Load weight for linear attention layers.
@@ -368,13 +658,42 @@ class BailingMoELinearAttention(LinearAttention):
def _decode_infer(self, q, k, v, kv_cache, state_indices_tensor, attn_metadata):
"""Handle decode (single token per sequence)."""
decode_state_indices = getattr(attn_metadata, "state_indices_tensor_d", None)
num_accepted_tokens = getattr(attn_metadata, "num_accepted_tokens", None)
query_start_loc = getattr(attn_metadata, "query_start_loc_d", None)
if (
decode_state_indices is not None
and decode_state_indices.dim() > 1
and num_accepted_tokens is not None
and query_start_loc is not None
):
return bailing_linear_attention_decode_spec(
q,
k,
v,
kv_cache,
self.tp_slope,
decode_state_indices,
query_start_loc,
num_accepted_tokens,
q_start=0,
q_end=attn_metadata.num_decode_tokens,
slot_start=0,
slot_end=attn_metadata.num_decodes,
block_size=32,
)
decode_state_indices = (
state_indices_tensor
if decode_state_indices is None
else decode_state_indices
)
hidden = linear_attention_decode(
q,
k,
v,
kv_cache,
self.tp_slope,
state_indices_tensor,
decode_state_indices,
q_start=0,
q_end=attn_metadata.num_decode_tokens,
slot_start=0,
@@ -533,6 +533,10 @@ class BailingMoeV25Model(nn.Module):
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.word_embeddings(input_ids)
@property
def embed_tokens(self) -> nn.Module:
return self.word_embeddings
def forward(
self,
input_ids: torch.Tensor | None,
@@ -0,0 +1,380 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Inference-only Bailing MoE v2.5 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_linear import (
BailingMoeV25,
BailingMoeV25MLAAttention,
)
from vllm.sequence import IntermediateTensors
from .utils import PPMissingLayer, is_pp_missing_parameter, maybe_prefix
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 BailingMTPSharedHead(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 BailingMoeV25MultiTokenPredictorLayer(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 = BailingMoeV25MLAAttention(
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 = BailingMoeV25(
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 = BailingMTPSharedHead(
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:
assert inputs_embeds is not None
inputs_embeds = torch.where(positions.unsqueeze(-1) == 0, 0, inputs_embeds)
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 BailingMoeV25MultiTokenPredictor(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): BailingMoeV25MultiTokenPredictorLayer(
vllm_config,
f"{prefix}.layers.{idx}",
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 BailingMoeV25MTPModel(nn.Module):
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.lm_head: nn.Module | None = None
self.model = BailingMoeV25MultiTokenPredictor(
vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"),
)
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:
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,
num_redundant_experts=0,
)
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.dense", ".self_attn.o_proj")
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 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
+1
View File
@@ -612,6 +612,7 @@ _SPECULATIVE_DECODING_MODELS = {
"DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
"DeepSeekV4MTPModel": ("vllm.models.deepseek_v4", "DeepSeekV4MTP"),
"MiniMaxM3MTP": ("vllm.models.minimax_m3", "MiniMaxM3MTP"),
"BailingMoeV25MTPModel": ("bailing_moe_mtp", "BailingMoeV25MTPModel"),
"Gemma4MTPModel": ("gemma4_mtp", "Gemma4MTP"),
"ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
"ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),