[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
+52
View File
@@ -0,0 +1,52 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers import PretrainedConfig
from vllm.config.speculative import MTPModelTypes, SpeculativeConfig
from vllm.transformers_utils.model_arch_config_convertor import (
BailingHybridMTPModelArchConfigConvertor,
)
def _bailing_config() -> PretrainedConfig:
config = PretrainedConfig(
architectures=["BailingMoeV2_5ForCausalLM"],
hidden_size=4096,
kv_lora_rank=512,
num_attention_heads=32,
num_experts=256,
num_hidden_layers=32,
num_key_value_heads=32,
num_nextn_predict_layers=1,
qk_rope_head_dim=64,
vocab_size=157184,
)
config.model_type = "bailing_hybrid"
return config
def test_bailing_hybrid_mtp_hf_config_override():
config = _bailing_config()
overridden = SpeculativeConfig.hf_config_override(config)
assert overridden.model_type == "bailing_hybrid_mtp"
assert overridden.architectures == ["BailingMoeV25MTPModel"]
assert overridden.n_predict == 1
assert "bailing_hybrid_mtp" in MTPModelTypes.__args__
def test_bailing_hybrid_mtp_model_arch_config():
config = _bailing_config()
config.model_type = "bailing_hybrid_mtp"
config.architectures = ["BailingMoeV25MTPModel"]
model_arch_config = BailingHybridMTPModelArchConfigConvertor(
config, config
).convert()
assert model_arch_config.model_type == "bailing_hybrid_mtp"
assert model_arch_config.architectures == ["BailingMoeV25MTPModel"]
assert model_arch_config.total_num_hidden_layers == 1
assert model_arch_config.is_deepseek_mla
+6
View File
@@ -1557,6 +1557,12 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
use_original_num_layers=True,
),
# [MTP]
"BailingMoeV25MTPModel": _HfExamplesInfo(
"inclusionAI/Ring-2.5-1T",
speculative_model="inclusionAI/Ring-2.5-1T",
trust_remote_code=True,
is_available_online=False,
),
"DeepSeekMTPModel": _HfExamplesInfo(
"luccafong/deepseek_mtp_main_random",
speculative_model="luccafong/deepseek_mtp_draft_random",
@@ -0,0 +1,188 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from tests.v1.attention.utils import (
BatchSpec,
create_common_attn_metadata,
create_vllm_config,
)
from vllm.config import CUDAGraphMode, SpeculativeConfig
from vllm.v1.attention.backend import AttentionCGSupport
from vllm.v1.attention.backends.linear_attn import (
BailingLinearAttentionMetadataBuilder,
LinearAttentionMetadataBuilder,
)
from vllm.v1.attention.backends.utils import PAD_SLOT_ID
from vllm.v1.kv_cache_interface import MambaSpec
BLOCK_SIZE = 16
DEVICE = torch.device("cpu")
def _create_mamba_spec(num_speculative_blocks: int = 1) -> MambaSpec:
return MambaSpec(
block_size=BLOCK_SIZE,
shapes=((16, 64),),
dtypes=(torch.float16,),
num_speculative_blocks=num_speculative_blocks,
)
def test_bailing_linear_attention_reports_uniform_batch_cudagraph_support():
vllm_config = create_vllm_config(
hf_config_override={
"architectures": ["BailingMoeV2_5ForCausalLM"],
"model_type": "bailing_hybrid",
}
)
support = BailingLinearAttentionMetadataBuilder.get_cudagraph_support(
vllm_config, _create_mamba_spec()
)
assert support == AttentionCGSupport.UNIFORM_BATCH
def test_non_bailing_linear_attention_keeps_single_token_cudagraph_support():
vllm_config = create_vllm_config(
hf_config_override={
"architectures": ["MiniMaxText01ForCausalLM"],
"model_type": "minimax_text_01",
}
)
support = LinearAttentionMetadataBuilder.get_cudagraph_support(
vllm_config, _create_mamba_spec()
)
assert support == AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE
def test_linear_attention_spec_decode_full_graph_metadata_pads_cache_slots():
vllm_config = create_vllm_config(
hf_config_override={
"architectures": ["BailingMoeV2_5ForCausalLM"],
"model_type": "bailing_hybrid",
}
)
vllm_config.speculative_config = SpeculativeConfig(
method="ngram",
num_speculative_tokens=1,
)
vllm_config.compilation_config.cudagraph_mode = CUDAGraphMode.FULL_DECODE_ONLY
builder = BailingLinearAttentionMetadataBuilder(
kv_cache_spec=_create_mamba_spec(),
layer_names=["model.layers.0.self_attn"],
vllm_config=vllm_config,
device=DEVICE,
)
common = create_common_attn_metadata(
BatchSpec(seq_lens=[20, 20, 0], query_lens=[2, 2, 0]),
BLOCK_SIZE,
DEVICE,
)
common.block_table_tensor[2].fill_(-1)
metadata = builder.build(
common_prefix_len=0,
common_attn_metadata=common,
num_accepted_tokens=torch.tensor([1, 2, 1], dtype=torch.int32),
)
assert metadata.num_decodes == 3
assert metadata.num_prefills == 0
assert metadata.num_decode_tokens == 4
assert metadata.state_indices_tensor_d is not None
assert metadata.state_indices_tensor_d.shape == (3, 2)
assert torch.equal(
metadata.state_indices_tensor_d[2],
torch.full((2,), PAD_SLOT_ID, dtype=torch.int32),
)
assert torch.equal(
metadata.state_indices_tensor[2],
torch.tensor(PAD_SLOT_ID, dtype=torch.int32),
)
assert metadata.query_start_loc_d is not None
assert metadata.query_start_loc_d.tolist() == [0, 2, 4, 4]
assert metadata.num_accepted_tokens is not None
assert metadata.num_accepted_tokens.tolist() == [1, 2, 1]
def test_linear_attention_full_graph_metadata_uses_stable_decode_buffers():
vllm_config = create_vllm_config(
hf_config_override={
"architectures": ["BailingMoeV2_5ForCausalLM"],
"model_type": "bailing_hybrid",
}
)
vllm_config.speculative_config = SpeculativeConfig(
method="ngram",
num_speculative_tokens=1,
)
vllm_config.compilation_config.cudagraph_mode = CUDAGraphMode.FULL_DECODE_ONLY
builder = BailingLinearAttentionMetadataBuilder(
kv_cache_spec=_create_mamba_spec(),
layer_names=["model.layers.0.self_attn"],
vllm_config=vllm_config,
device=DEVICE,
)
common = create_common_attn_metadata(
BatchSpec(seq_lens=[20, 20, 0], query_lens=[2, 2, 0]),
BLOCK_SIZE,
DEVICE,
arange_block_indices=True,
)
common.block_table_tensor = torch.tensor(
[[10, 11], [12, 13], [-1, -1]],
dtype=torch.int32,
device=DEVICE,
)
first = builder.build(
common_prefix_len=0,
common_attn_metadata=common,
num_accepted_tokens=torch.tensor([1, 2, 1], dtype=torch.int32),
)
assert first.state_indices_tensor_d is not None
assert first.query_start_loc_d is not None
assert first.num_accepted_tokens is not None
state_ptr = first.state_indices_tensor_d.data_ptr()
query_ptr = first.query_start_loc_d.data_ptr()
accepted_ptr = first.num_accepted_tokens.data_ptr()
common2 = create_common_attn_metadata(
BatchSpec(seq_lens=[36, 0, 0], query_lens=[2, 0, 0]),
BLOCK_SIZE,
DEVICE,
arange_block_indices=True,
)
common2.block_table_tensor = torch.tensor(
[[20, 21], [-1, -1], [-1, -1]],
dtype=torch.int32,
device=DEVICE,
)
second = builder.build(
common_prefix_len=0,
common_attn_metadata=common2,
num_accepted_tokens=torch.tensor([2, 1, 1], dtype=torch.int32),
)
assert second.state_indices_tensor_d is not None
assert second.query_start_loc_d is not None
assert second.num_accepted_tokens is not None
assert second.state_indices_tensor_d.data_ptr() == state_ptr
assert second.query_start_loc_d.data_ptr() == query_ptr
assert second.num_accepted_tokens.data_ptr() == accepted_ptr
assert second.state_indices_tensor_d.tolist() == [
[20, 21],
[PAD_SLOT_ID, PAD_SLOT_ID],
[PAD_SLOT_ID, PAD_SLOT_ID],
]
assert second.query_start_loc_d.tolist() == [0, 2, 2, 2]
assert second.num_accepted_tokens.tolist() == [2, 1, 1]
+16
View File
@@ -46,6 +46,7 @@ MTPModelTypes = Literal[
"qwen3_5_mtp",
"longcat_flash_mtp",
"minimax_m3_mtp",
"bailing_hybrid_mtp",
"mtp",
"pangu_ultra_moe_mtp",
"step3p5_mtp",
@@ -463,6 +464,21 @@ class SpeculativeConfig:
{"n_predict": n_predict, "architectures": ["Qwen3NextMTP"]}
)
architectures = getattr(hf_config, "architectures", []) or []
if (
hf_config.model_type == "bailing_hybrid"
or "BailingMoeV2_5ForCausalLM" in architectures
):
hf_config.model_type = "bailing_hybrid_mtp"
if hf_config.model_type == "bailing_hybrid_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update(
{
"n_predict": n_predict,
"architectures": ["BailingMoeV25MTPModel"],
}
)
if hf_config.model_type == "exaone_moe":
hf_config.model_type = "exaone_moe_mtp"
if hf_config.model_type == "exaone_moe_mtp":
@@ -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"),
@@ -271,6 +271,7 @@ class ModelArchConfigConvertorBase:
"pangu_ultra_moe",
"pangu_ultra_moe_mtp",
"bailing_hybrid",
"bailing_hybrid_mtp",
):
# check is deepseek_v4 model
if hasattr(self.hf_text_config, "compress_ratios"):
@@ -539,6 +540,11 @@ class Qwen3NextMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
return getattr(self.hf_text_config, "num_nextn_predict_layers", 0)
class BailingHybridMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def get_num_hidden_layers(self) -> int:
return getattr(self.hf_text_config, "num_nextn_predict_layers", 0)
class Qwen3_5MTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def get_num_hidden_layers(self) -> int:
return getattr(self.hf_text_config, "mtp_num_hidden_layers", 0)
@@ -633,6 +639,7 @@ class MossAudioModelArchConfigConvertor(ModelArchConfigConvertorBase):
# hf_config.model_type -> convertor class
MODEL_ARCH_CONFIG_CONVERTORS = {
"bailing_hybrid_mtp": BailingHybridMTPModelArchConfigConvertor,
"cohere_asr": CohereAsrModelArchConfigConvertor,
"dbrx": DbrxModelArchConfigConvertor,
"deepseek_mtp": DeepSeekMTPModelArchConfigConvertor,
+212 -1
View File
@@ -4,7 +4,7 @@ from dataclasses import dataclass
import torch
from vllm.config import VllmConfig
from vllm.config import CompilationConfig, VllmConfig
from vllm.v1.attention.backend import (
AttentionBackend,
AttentionCGSupport,
@@ -12,6 +12,7 @@ from vllm.v1.attention.backend import (
CommonAttentionMetadata,
)
from vllm.v1.attention.backends.utils import (
PAD_SLOT_ID,
mamba_get_block_table_tensor,
split_decodes_and_prefills,
)
@@ -91,3 +92,213 @@ class LinearAttentionMetadataBuilder(AttentionMetadataBuilder[LinearAttentionMet
state_indices_tensor=state_indices_tensor,
)
return attn_metadata
class BailingLinearAttentionBackend(LinearAttentionBackend):
@staticmethod
def get_name() -> str:
return "BAILING_LINEAR_ATTN"
@staticmethod
def get_builder_cls() -> type["BailingLinearAttentionMetadataBuilder"]:
return BailingLinearAttentionMetadataBuilder
@dataclass
class BailingLinearAttentionMetadata(LinearAttentionMetadata):
state_indices_tensor_d: torch.Tensor | None = None
state_indices_tensor_p: torch.Tensor | None = None
num_accepted_tokens: torch.Tensor | None = None
query_start_loc_d: torch.Tensor | None = None
class BailingLinearAttentionMetadataBuilder(LinearAttentionMetadataBuilder):
supports_spec_decode_metadata = True
supports_update_block_table: bool = False
@classmethod
def get_cudagraph_support(
cls,
vllm_config: VllmConfig,
kv_cache_spec: AttentionSpec,
) -> AttentionCGSupport:
return AttentionCGSupport.UNIFORM_BATCH
def __init__(
self,
kv_cache_spec: AttentionSpec,
layer_names: list[str],
vllm_config: VllmConfig,
device: torch.device,
):
super().__init__(kv_cache_spec, layer_names, vllm_config, device)
self.compilation_config: CompilationConfig = vllm_config.compilation_config
self.num_spec_tokens: int = vllm_config.num_speculative_tokens
self.use_spec_decode: bool = self.num_spec_tokens > 0
self.decode_cudagraph_max_bs: int = vllm_config.scheduler_config.max_num_seqs
if self.compilation_config.max_cudagraph_capture_size is not None:
self.decode_cudagraph_max_bs = min(
self.decode_cudagraph_max_bs,
self.compilation_config.max_cudagraph_capture_size,
)
self.decode_state_indices_tensor: torch.Tensor = torch.empty(
(self.decode_cudagraph_max_bs, 1 + self.num_spec_tokens),
dtype=torch.int32,
device=device,
)
self.decode_legacy_state_indices_tensor: torch.Tensor = torch.empty(
(self.decode_cudagraph_max_bs,),
dtype=torch.int32,
device=device,
)
self.decode_query_start_loc: torch.Tensor = torch.empty(
(self.decode_cudagraph_max_bs + 1,),
dtype=torch.int32,
device=device,
)
self.decode_num_accepted_tokens: torch.Tensor = torch.empty(
(self.decode_cudagraph_max_bs,),
dtype=torch.int32,
device=device,
)
self._init_reorder_batch_threshold(1, self.use_spec_decode)
def build_for_cudagraph_capture(
self,
common_attn_metadata: CommonAttentionMetadata,
) -> BailingLinearAttentionMetadata:
num_accepted_tokens = None
if self.use_spec_decode:
assert common_attn_metadata.max_query_len <= 1 + self.num_spec_tokens, (
"Bailing linear attention only supports speculative decoding "
"with query length <= 1 + number of speculative tokens."
)
num_accepted_tokens = torch.diff(common_attn_metadata.query_start_loc)
return self.build(
common_prefix_len=0,
common_attn_metadata=common_attn_metadata,
num_accepted_tokens=num_accepted_tokens,
)
def build( # type: ignore[override]
self,
common_prefix_len: int,
common_attn_metadata: CommonAttentionMetadata,
fast_build: bool = False,
*,
num_accepted_tokens: torch.Tensor | None = None,
num_decode_draft_tokens_cpu: torch.Tensor | None = None,
) -> BailingLinearAttentionMetadata:
query_start_loc = common_attn_metadata.query_start_loc
seq_lens = common_attn_metadata.seq_lens
num_reqs = common_attn_metadata.num_reqs
use_spec_decode = self.use_spec_decode and num_accepted_tokens is not None
state_indices_tensor = mamba_get_block_table_tensor(
common_attn_metadata.block_table_tensor,
common_attn_metadata.seq_lens,
self.kv_cache_spec,
self.vllm_config.cache_config.mamba_cache_mode,
)
if state_indices_tensor.dim() == 1:
state_indices_tensor = state_indices_tensor.unsqueeze(-1)
decode_threshold = self.reorder_batch_threshold if use_spec_decode else 1
num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = (
split_decodes_and_prefills(
common_attn_metadata,
decode_threshold=decode_threshold,
)
)
state_indices_tensor_d, state_indices_tensor_p = torch.split(
state_indices_tensor,
[num_decodes, num_prefills],
dim=0,
)
state_indices_tensor_p = state_indices_tensor_p[:, 0]
query_start_loc_d = None
if use_spec_decode:
assert num_accepted_tokens is not None
state_indices_tensor_d = state_indices_tensor_d[
:, : 1 + self.num_spec_tokens
]
query_start_loc_d = query_start_loc[: num_decodes + 1]
num_accepted_tokens = num_accepted_tokens[:num_decodes]
else:
state_indices_tensor_d = state_indices_tensor_d[:, 0]
num_accepted_tokens = None
legacy_state_indices_tensor = state_indices_tensor[:, 0]
cudagraph_mode = self.compilation_config.cudagraph_mode
use_full_cudagraph = (
cudagraph_mode is not None and cudagraph_mode.has_full_cudagraphs()
)
if (
num_prefills == 0
and num_decodes <= self.decode_cudagraph_max_bs
and use_full_cudagraph
):
padded_bs = num_reqs
is_padded_decode = seq_lens[:num_decodes] == 0
if state_indices_tensor_d.dim() > 1:
state_indices_tensor_d = torch.where(
is_padded_decode.unsqueeze(1),
torch.full_like(state_indices_tensor_d, PAD_SLOT_ID),
state_indices_tensor_d,
)
self.decode_state_indices_tensor[:num_decodes].copy_(
state_indices_tensor_d,
non_blocking=True,
)
state_indices_tensor_d = self.decode_state_indices_tensor[:padded_bs]
state_indices_tensor_d[num_decodes:] = PAD_SLOT_ID
self.decode_legacy_state_indices_tensor[:num_decodes].copy_(
torch.where(
is_padded_decode,
torch.full_like(
legacy_state_indices_tensor[:num_decodes],
PAD_SLOT_ID,
),
legacy_state_indices_tensor[:num_decodes],
),
non_blocking=True,
)
legacy_state_indices_tensor = self.decode_legacy_state_indices_tensor[
:padded_bs
]
legacy_state_indices_tensor[num_decodes:] = PAD_SLOT_ID
if state_indices_tensor_d.dim() == 1:
state_indices_tensor_d = legacy_state_indices_tensor
if use_spec_decode and num_accepted_tokens is not None:
assert query_start_loc_d is not None
self.decode_query_start_loc[: num_decodes + 1].copy_(
query_start_loc_d,
non_blocking=True,
)
decode_num_query_tokens = query_start_loc_d[-1]
query_start_loc_d = self.decode_query_start_loc[: padded_bs + 1]
query_start_loc_d[num_decodes + 1 :] = decode_num_query_tokens
self.decode_num_accepted_tokens[:num_decodes].copy_(
num_accepted_tokens,
non_blocking=True,
)
num_accepted_tokens = self.decode_num_accepted_tokens[:padded_bs]
num_accepted_tokens[num_decodes:] = 1
return BailingLinearAttentionMetadata(
num_prefills=num_prefills,
num_prefill_tokens=num_prefill_tokens,
num_decodes=num_decodes,
num_decode_tokens=num_decode_tokens,
query_start_loc=query_start_loc,
seq_lens=seq_lens,
state_indices_tensor=legacy_state_indices_tensor,
state_indices_tensor_d=state_indices_tensor_d,
state_indices_tensor_p=state_indices_tensor_p,
num_accepted_tokens=num_accepted_tokens,
query_start_loc_d=query_start_loc_d,
)
+12 -2
View File
@@ -132,6 +132,9 @@ from vllm.v1.attention.backend import (
CommonAttentionMetadata,
)
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
from vllm.v1.attention.backends.linear_attn import (
BailingLinearAttentionMetadataBuilder,
)
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
from vllm.v1.attention.backends.utils import (
NULL_BLOCK_ID,
@@ -2442,9 +2445,16 @@ class GPUModelRunner(
extra_attn_metadata_args = {}
if use_spec_decode and isinstance(
builder, (Mamba2AttentionMetadataBuilder, GDNAttentionMetadataBuilder)
builder,
(
Mamba2AttentionMetadataBuilder,
GDNAttentionMetadataBuilder,
BailingLinearAttentionMetadataBuilder,
),
):
assert ubid is None, "UBatching not supported with GDN yet"
assert ubid is None, (
"UBatching not supported with GDN or linear attn yet"
)
extra_attn_metadata_args = dict(
num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs_padded],
num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[