[Attention] Mamba attention module refactor (#41126)

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2026-05-22 17:13:12 +08:00
committed by GitHub
parent 65b7a812a2
commit 7e1b45a092
10 changed files with 765 additions and 774 deletions
+1 -1
View File
@@ -752,7 +752,7 @@ class CompilationConfig:
"vllm::short_conv",
"vllm::linear_attention",
"vllm::plamo2_mamba_mixer",
"vllm::gdn_attention_core",
"vllm::qwen_gdn_attention_core",
"vllm::gdn_attention_core_xpu",
"vllm::olmo_hybrid_gdn_full_forward",
"vllm::kda_attention",
@@ -0,0 +1,58 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from transformers import PretrainedConfig
from vllm.config import (
VllmConfig,
)
from vllm.distributed import (
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.custom_op import PluggableLayer
from vllm.model_executor.layers.mamba.abstract import MambaBase
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateDtypeCalculator,
)
from vllm.model_executor.models.utils import extract_layer_index
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
class GatedDeltaNetAttention(PluggableLayer, MambaBase):
"""Base class for GatedDeltaNet attention layer."""
def __init__(
self,
config: PretrainedConfig,
vllm_config: VllmConfig,
prefix: str = "",
) -> None:
super().__init__()
self.prefix = prefix
self.tp_size = get_tensor_model_parallel_world_size()
self.tp_rank = get_tensor_model_parallel_rank()
self.layer_idx = extract_layer_index(prefix)
self.hidden_size = config.hidden_size
self.activation = config.hidden_act
self.layer_norm_epsilon = config.rms_norm_eps
self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config
self.quant_config = vllm_config.quant_config
self.speculative_config = vllm_config.speculative_config
self.num_spec = (
self.speculative_config.num_speculative_tokens
if self.speculative_config
else 0
)
@property
def mamba_type(self) -> MambaAttentionBackendEnum:
return MambaAttentionBackendEnum.GDN_ATTN
def get_state_dtype(self) -> tuple[torch.dtype, ...]:
return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
self.model_config.dtype,
self.cache_config.mamba_cache_dtype,
self.cache_config.mamba_ssm_cache_dtype,
)
@@ -5,39 +5,37 @@ import torch
from einops import rearrange
from torch import nn
from vllm.config import CacheConfig, ModelConfig, get_current_vllm_config
from vllm.config import VllmConfig, get_current_vllm_config
from vllm.distributed import (
divide,
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.custom_op import PluggableLayer
from vllm.model_executor.layers.mamba.gdn.base import GatedDeltaNetAttention
from vllm.model_executor.model_loader.weight_utils import sharded_weight_loader
from vllm.model_executor.utils import set_weight_attrs
from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
from .fla.ops.kda import (
from ...fla.ops.kda import (
FusedRMSNormGated,
chunk_kda,
fused_kda_gate,
fused_recurrent_kda,
)
from .linear import (
from ...linear import (
ColumnParallelLinear,
ReplicatedLinear,
RowParallelLinear,
)
from .mamba.abstract import MambaBase
from .mamba.mamba_utils import (
from ..mamba_utils import (
MambaStateDtypeCalculator,
MambaStateShapeCalculator,
is_conv_state_dim_first,
)
from .mamba.ops.causal_conv1d import causal_conv1d_fn, causal_conv1d_update
from .quantization.base_config import QuantizationConfig
from ..ops.causal_conv1d import causal_conv1d_fn, causal_conv1d_update
logger = init_logger(__name__)
@@ -83,11 +81,8 @@ direct_register_custom_op(
)
class KimiDeltaAttention(nn.Module, MambaBase):
@property
def mamba_type(self) -> MambaAttentionBackendEnum:
return MambaAttentionBackendEnum.GDN_ATTN
@PluggableLayer.register("kimi_gated_delta_net_attention")
class KimiGatedDeltaNetAttention(GatedDeltaNetAttention):
def get_state_dtype(
self,
) -> tuple[torch.dtype, torch.dtype, torch.dtype, torch.dtype]:
@@ -106,28 +101,16 @@ class KimiDeltaAttention(nn.Module, MambaBase):
def __init__(
self,
layer_idx: int,
hidden_size: int,
quant_config: QuantizationConfig | None = None,
cache_config: CacheConfig | None = None,
model_config: ModelConfig | None = None,
rms_norm_eps: float = 1e-5,
config: KimiLinearConfig,
vllm_config: VllmConfig,
prefix: str = "",
**kwargs,
) -> None:
super().__init__()
self.tp_size = get_tensor_model_parallel_world_size()
self.tp_rank = get_tensor_model_parallel_rank()
self.hidden_size = hidden_size
self.model_config = model_config
self.cache_config = cache_config
if model_config is None:
raise ValueError("model_config must be provided")
kda_config = model_config.linear_attn_config # type: ignore[attr-defined]
super().__init__(config, vllm_config, prefix)
kda_config = config.linear_attn_config # type: ignore[attr-defined]
assert kda_config is not None, "linear_attn_config must be set"
self.head_dim = kda_config["head_dim"]
self.num_heads = kda_config["num_heads"]
self.layer_idx = layer_idx
self.prefix = prefix
assert self.num_heads % self.tp_size == 0
self.local_num_heads = divide(self.num_heads, self.tp_size)
@@ -138,21 +121,21 @@ class KimiDeltaAttention(nn.Module, MambaBase):
self.hidden_size,
projection_size,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.q_proj",
)
self.k_proj = ColumnParallelLinear(
self.hidden_size,
projection_size,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.k_proj",
)
self.v_proj = ColumnParallelLinear(
self.hidden_size,
projection_size,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.v_proj",
)
@@ -160,7 +143,7 @@ class KimiDeltaAttention(nn.Module, MambaBase):
self.hidden_size,
self.head_dim,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.f_a_proj",
)
@@ -168,7 +151,7 @@ class KimiDeltaAttention(nn.Module, MambaBase):
self.head_dim,
projection_size,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.f_b_proj",
)
self.dt_bias = nn.Parameter(
@@ -181,7 +164,7 @@ class KimiDeltaAttention(nn.Module, MambaBase):
self.hidden_size,
self.num_heads,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.b_proj",
)
@@ -223,24 +206,22 @@ class KimiDeltaAttention(nn.Module, MambaBase):
self.hidden_size,
self.head_dim,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.g_a_proj",
)
self.g_b_proj = ColumnParallelLinear(
self.head_dim,
projection_size,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.g_b_proj",
)
self.o_norm = FusedRMSNormGated(
self.head_dim, eps=rms_norm_eps, activation="sigmoid"
)
self.o_norm = FusedRMSNormGated(self.head_dim, activation="sigmoid")
self.o_proj = RowParallelLinear(
projection_size,
self.hidden_size,
bias=False,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.o_proj",
)
@@ -0,0 +1,634 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from einops import rearrange
from torch import nn
from vllm.config import (
VllmConfig,
get_current_vllm_config,
)
from vllm.distributed import (
divide,
)
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.model_executor.custom_op import PluggableLayer
from vllm.model_executor.layers.fla.ops import (
chunk_gated_delta_rule,
fused_recurrent_gated_delta_rule,
)
from vllm.model_executor.layers.layernorm import RMSNormGated
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
MergedColumnParallelLinear,
RowParallelLinear,
)
from vllm.model_executor.layers.mamba.gdn.base import GatedDeltaNetAttention
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateShapeCalculator,
is_conv_state_dim_first,
)
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
causal_conv1d_fn,
causal_conv1d_update,
)
from vllm.model_executor.model_loader.weight_utils import (
sharded_weight_loader,
)
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.triton_utils.allocation import set_triton_allocator
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
@PluggableLayer.register("olmo_hybrid_gated_delta_net_attention")
class OlmoHybridGatedDeltaNetAttention(GatedDeltaNetAttention):
"""
Gated DeltaNet linear attention layer for OLMo Hybrid.
This implements the linear attention mechanism that replaces sliding window
attention in the hybrid architecture.
"""
def get_state_shape(
self,
) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
return MambaStateShapeCalculator.gated_delta_net_state_shape(
self.tp_size,
self.num_k_heads,
self.num_v_heads,
self.head_k_dim,
self.head_v_dim,
self.conv_kernel_size,
self.num_spec,
)
def __init__(
self,
config,
vllm_config: VllmConfig,
prefix: str = "",
) -> None:
super().__init__(config, vllm_config, prefix=prefix)
assert getattr(config, "linear_use_gate", True), (
"OlmoHybridGatedDeltaNet requires linear_use_gate=True"
)
self.num_k_heads = config.linear_num_key_heads
self.num_v_heads = config.linear_num_value_heads
self.head_k_dim = config.linear_key_head_dim
self.head_v_dim = config.linear_value_head_dim
self.conv_kernel_size = config.linear_conv_kernel_dim
self.key_dim = self.head_k_dim * self.num_k_heads
self.value_dim = self.head_v_dim * self.num_v_heads
self.allow_neg_eigval = getattr(config, "linear_allow_neg_eigval", False)
# Fused QKVG projection: 1 matmul instead of 4
self.in_proj_qkvg = MergedColumnParallelLinear(
input_size=self.hidden_size,
output_sizes=[self.key_dim, self.key_dim, self.value_dim, self.value_dim],
bias=False,
quant_config=self.quant_config,
prefix=f"{prefix}.in_proj_qkvg",
)
# Separate B and A projections to preserve numerical precision.
# Fusing these into one matmul changes FP accumulation order for the
# gating scalars, which compounds through the GDN recurrent state.
self.b_proj = ColumnParallelLinear(
input_size=self.hidden_size,
output_size=self.num_v_heads,
bias=False,
quant_config=self.quant_config,
prefix=f"{prefix}.b_proj",
)
self.a_proj = ColumnParallelLinear(
input_size=self.hidden_size,
output_size=self.num_v_heads,
bias=False,
quant_config=self.quant_config,
prefix=f"{prefix}.a_proj",
)
# Fused conv1d: single parameter instead of 3
self.conv_dim = self.key_dim * 2 + self.value_dim
self.conv1d = ColumnParallelLinear(
input_size=self.conv_kernel_size,
output_size=self.conv_dim,
bias=False,
prefix=f"{prefix}.conv1d",
)
self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1)
delattr(self.conv1d.weight, "weight_loader")
set_weight_attrs(
self.conv1d.weight,
{
"weight_loader": _make_fused_conv1d_weight_loader(
[self.key_dim, self.key_dim, self.value_dim],
self.tp_size,
self.tp_rank,
)
},
)
self.dt_bias = nn.Parameter(
torch.ones(self.num_v_heads // self.tp_size),
)
self.A_log = nn.Parameter(
torch.empty(
divide(self.num_v_heads, self.tp_size),
)
)
set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)})
set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)})
# use eps=1e-5 to match FLA's FusedRMSNormGated
self.o_norm = RMSNormGated(
self.head_v_dim,
eps=1e-5,
group_size=None,
norm_before_gate=True,
device=current_platform.current_device(),
dtype=config.torch_dtype if hasattr(config, "torch_dtype") else None,
)
self.o_proj = RowParallelLinear(
self.value_dim,
self.hidden_size,
bias=False,
input_is_parallel=True,
quant_config=self.quant_config,
prefix=f"{prefix}.o_proj",
)
# FLA triton kernels need a PyTorch-backed allocator for scratch
# memory (required by triton >= 3.x autotuner). Set once at init.
set_triton_allocator(current_platform.current_device())
compilation_config = get_current_vllm_config().compilation_config
if prefix in compilation_config.static_forward_context:
raise ValueError(f"Duplicate layer name: {prefix}")
compilation_config.static_forward_context[prefix] = self
def rearrange_mixed_qkv(self, mixed_qkv):
if mixed_qkv is None:
return None, None, None
query, key, value = torch.split(
mixed_qkv,
[
self.key_dim // self.tp_size,
self.key_dim // self.tp_size,
self.value_dim // self.tp_size,
],
dim=-1,
)
num_k_heads = self.num_k_heads // self.tp_size
num_v_heads = self.num_v_heads // self.tp_size
query = rearrange(query, "l (h d) -> 1 l h d", h=num_k_heads, d=self.head_k_dim)
key = rearrange(key, "l (h d) -> 1 l h d", h=num_k_heads, d=self.head_k_dim)
value = rearrange(value, "l (h d) -> 1 l h d", h=num_v_heads, d=self.head_v_dim)
# GQA expansion if needed
if num_v_heads > num_k_heads:
expand_ratio = num_v_heads // num_k_heads
query = query.unsqueeze(3).expand(-1, -1, -1, expand_ratio, -1)
query = query.reshape(1, query.shape[1], num_v_heads, self.head_k_dim)
key = key.unsqueeze(3).expand(-1, -1, -1, expand_ratio, -1)
key = key.reshape(1, key.shape[1], num_v_heads, self.head_k_dim)
return query.contiguous(), key.contiguous(), value.contiguous()
def forward(
self,
hidden_states: torch.Tensor,
output: torch.Tensor,
):
# NOTE: We wrap the ENTIRE linear attention forward (projections +
# core recurrence + output norm + output projection) in a single
# custom op, rather than just wrapping the recurrent core like
# other GDN models (e.g. Qwen3Next) do.
#
# Why: torch.compile with inductor generates fused kernels for
# matmuls and pointwise ops. These fused kernels can differ in
# floating-point accumulation order from eager-mode cuBLAS,
# introducing small numerical differences (~1e-7 per op). For
# standard transformer attention this is harmless because each
# position is computed independently. But for the GDN recurrent
# state, these tiny input differences compound at every timestep
# across the full sequence length, causing severe logprob
# divergence (e.g. ~15% top-1 agreement with eager baseline).
#
# By making the full forward opaque to inductor, the projections
# and output norm run with eager-mode kernels (cuBLAS, triton),
# preserving numerical consistency. The tradeoff is reduced
# compilation speedup (~1.5x vs ~3x), but logprob agreement
# improves from ~15% to ~83% top-1 vs eager.
#
# The remaining ~17% divergence comes from inductor compiling
# the MLP and transformer attention layers that are NOT wrapped
# in custom ops -- their small precision differences propagate
# as inputs to the GDN layers from outside.
torch.ops.vllm.olmo_hybrid_gdn_full_forward(
hidden_states,
output,
self.prefix,
)
def _full_forward(
self,
hidden_states: torch.Tensor,
output: torch.Tensor,
):
num_tokens = hidden_states.size(0)
# ============================================================
# Part 1: Input Projection (2 fused matmuls instead of 6)
# ============================================================
projected_qkvg, _ = self.in_proj_qkvg(hidden_states)
conv_dim_sharded = (self.key_dim * 2 + self.value_dim) // self.tp_size
mixed_qkv = projected_qkvg[..., :conv_dim_sharded]
gate = projected_qkvg[..., conv_dim_sharded:]
b, _ = self.b_proj(hidden_states)
a, _ = self.a_proj(hidden_states)
# ============================================================
# Part 2: Core Attention
# ============================================================
core_attn_out = torch.zeros(
(num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim),
dtype=hidden_states.dtype,
device=hidden_states.device,
)
self._forward_core(
mixed_qkv=mixed_qkv,
b=b,
a=a,
core_attn_out=core_attn_out,
)
# ============================================================
# Part 3: Output Projection
# ============================================================
gate = gate.view(num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim)
core_attn_out_flat = core_attn_out.reshape(-1, core_attn_out.shape[-1])
gate_flat = gate.reshape(-1, gate.shape[-1])
core_attn_out_normed = self.o_norm(core_attn_out_flat, gate_flat)
core_attn_out = core_attn_out_normed.view(
num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim
)
core_attn_out = rearrange(core_attn_out, "l h d -> l (h d)")
output[:num_tokens], _ = self.o_proj(core_attn_out)
def _forward_core(
self,
mixed_qkv: torch.Tensor,
b: torch.Tensor,
a: torch.Tensor,
core_attn_out: torch.Tensor,
):
"""
Core attention computation (called by custom op).
"""
forward_context = get_forward_context()
attn_metadata = forward_context.attn_metadata
if attn_metadata is None:
# V1 profile run
return
assert isinstance(attn_metadata, dict)
attn_metadata = attn_metadata[self.prefix] # type: ignore[assignment]
assert isinstance(attn_metadata, GDNAttentionMetadata)
has_initial_state = attn_metadata.has_initial_state
spec_query_start_loc = attn_metadata.spec_query_start_loc
non_spec_query_start_loc = attn_metadata.non_spec_query_start_loc
spec_sequence_masks = attn_metadata.spec_sequence_masks
spec_token_indx = attn_metadata.spec_token_indx
non_spec_token_indx = attn_metadata.non_spec_token_indx
spec_state_indices_tensor = attn_metadata.spec_state_indices_tensor
non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor
self_kv_cache = self.kv_cache
# conv_state must be (..., dim, width-1) for the conv kernels.
# DS layout stores it that way directly; SD layout needs a transpose.
conv_state = (
self_kv_cache[0]
if is_conv_state_dim_first()
else self_kv_cache[0].transpose(-1, -2)
)
ssm_state = self_kv_cache[1]
num_actual_tokens = attn_metadata.num_actual_tokens
num_accepted_tokens = attn_metadata.num_accepted_tokens
mixed_qkv = mixed_qkv[:num_actual_tokens]
b = b[:num_actual_tokens]
a = a[:num_actual_tokens]
conv_weights = self.conv1d.weight.view(
self.conv1d.weight.size(0), self.conv1d.weight.size(2)
)
if spec_sequence_masks is not None:
if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
mixed_qkv_spec = mixed_qkv
mixed_qkv_non_spec = None
else:
mixed_qkv_spec = mixed_qkv.index_select(0, spec_token_indx)
mixed_qkv_non_spec = mixed_qkv.index_select(0, non_spec_token_indx)
else:
mixed_qkv_spec = None
mixed_qkv_non_spec = mixed_qkv
if spec_sequence_masks is not None:
assert spec_query_start_loc is not None
assert spec_state_indices_tensor is not None
assert num_accepted_tokens is not None
mixed_qkv_spec = causal_conv1d_update(
mixed_qkv_spec,
conv_state,
conv_weights,
None, # no bias
self.activation,
conv_state_indices=spec_state_indices_tensor[:, 0][
: attn_metadata.num_spec_decodes
],
num_accepted_tokens=num_accepted_tokens,
query_start_loc=spec_query_start_loc,
max_query_len=spec_state_indices_tensor.size(-1),
validate_data=False,
)
if attn_metadata.num_prefills > 0:
assert mixed_qkv_non_spec is not None
mixed_qkv_non_spec_T = mixed_qkv_non_spec.transpose(0, 1)
mixed_qkv_non_spec = causal_conv1d_fn(
mixed_qkv_non_spec_T,
conv_weights,
None,
activation=self.activation,
conv_states=conv_state,
has_initial_state=has_initial_state,
cache_indices=non_spec_state_indices_tensor,
query_start_loc=non_spec_query_start_loc,
metadata=attn_metadata,
).transpose(0, 1)
elif attn_metadata.num_decodes > 0:
assert non_spec_state_indices_tensor is not None
mixed_qkv_non_spec = causal_conv1d_update(
mixed_qkv_non_spec,
conv_state,
conv_weights,
None,
self.activation,
conv_state_indices=non_spec_state_indices_tensor[
: attn_metadata.num_decodes
],
validate_data=True,
)
else:
mixed_qkv_non_spec = None
query_spec, key_spec, value_spec = self.rearrange_mixed_qkv(mixed_qkv_spec)
query_non_spec, key_non_spec, value_non_spec = self.rearrange_mixed_qkv(
mixed_qkv_non_spec
)
g, beta = fused_olmo_hybrid_gdn_gating(
self.A_log, a, b, self.dt_bias, self.allow_neg_eigval
)
if spec_sequence_masks is not None:
assert spec_token_indx is not None
assert non_spec_token_indx is not None
if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
g_spec = g
beta_spec = beta
g_non_spec = None
beta_non_spec = None
else:
g_spec = g.index_select(1, spec_token_indx)
beta_spec = beta.index_select(1, spec_token_indx)
g_non_spec = g.index_select(1, non_spec_token_indx)
beta_non_spec = beta.index_select(1, non_spec_token_indx)
else:
g_spec = None
beta_spec = None
g_non_spec = g
beta_non_spec = beta
if spec_sequence_masks is not None:
assert spec_query_start_loc is not None
assert spec_state_indices_tensor is not None
assert num_accepted_tokens is not None
core_attn_out_spec, last_recurrent_state = fused_recurrent_gated_delta_rule(
q=query_spec,
k=key_spec,
v=value_spec,
g=g_spec,
beta=beta_spec,
initial_state=ssm_state,
inplace_final_state=True,
cu_seqlens=spec_query_start_loc[: attn_metadata.num_spec_decodes + 1],
ssm_state_indices=spec_state_indices_tensor,
num_accepted_tokens=num_accepted_tokens,
use_qk_l2norm_in_kernel=True,
)
else:
core_attn_out_spec, last_recurrent_state = None, None
if attn_metadata.num_prefills > 0:
assert non_spec_state_indices_tensor is not None
assert has_initial_state is not None
assert non_spec_query_start_loc is not None
initial_state = ssm_state[non_spec_state_indices_tensor].contiguous()
initial_state[~has_initial_state, ...] = 0
(
core_attn_out_non_spec,
last_recurrent_state,
) = chunk_gated_delta_rule(
q=query_non_spec,
k=key_non_spec,
v=value_non_spec,
g=g_non_spec,
beta=beta_non_spec,
initial_state=initial_state,
output_final_state=True,
cu_seqlens=non_spec_query_start_loc,
use_qk_l2norm_in_kernel=True,
)
ssm_state[non_spec_state_indices_tensor] = last_recurrent_state.to(
ssm_state.dtype
)
elif attn_metadata.num_decodes > 0:
assert non_spec_query_start_loc is not None
assert non_spec_state_indices_tensor is not None
core_attn_out_non_spec, last_recurrent_state = (
fused_recurrent_gated_delta_rule(
q=query_non_spec,
k=key_non_spec,
v=value_non_spec,
g=g_non_spec,
beta=beta_non_spec,
initial_state=ssm_state,
inplace_final_state=True,
cu_seqlens=non_spec_query_start_loc[
: attn_metadata.num_decodes + 1
],
ssm_state_indices=non_spec_state_indices_tensor,
use_qk_l2norm_in_kernel=True,
)
)
else:
core_attn_out_non_spec, last_recurrent_state = None, None
if spec_sequence_masks is not None and core_attn_out_non_spec is not None:
merged_out = torch.empty(
(1, num_actual_tokens, *core_attn_out_spec.shape[2:]),
dtype=core_attn_out_non_spec.dtype,
device=core_attn_out_non_spec.device,
)
merged_out.index_copy_(1, spec_token_indx, core_attn_out_spec)
merged_out.index_copy_(1, non_spec_token_indx, core_attn_out_non_spec)
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
elif spec_sequence_masks is not None:
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)
def _make_fused_conv1d_weight_loader(dims, tp_size, tp_rank):
"""Weight loader for loading separate HF conv weights into a fused conv1d.
dims: list of original (un-sharded) dims per section,
e.g. [key_dim, key_dim, value_dim]
"""
sharded_dims = [d // tp_size for d in dims]
def weight_loader(param, loaded_weight, loaded_shard_id=None):
if loaded_weight.dim() == 2:
loaded_weight = loaded_weight.unsqueeze(1)
dim = dims[loaded_shard_id]
shard_size = dim // tp_size
tp_start = tp_rank * shard_size
sharded_weight = loaded_weight[tp_start : tp_start + shard_size]
offset = sum(sharded_dims[:loaded_shard_id])
param.data[offset : offset + shard_size].copy_(sharded_weight)
return weight_loader
def olmo_hybrid_gdn_full_forward(
hidden_states: torch.Tensor,
output: torch.Tensor,
layer_name: str,
) -> None:
"""Full linear attention forward wrapped as a custom op.
Prevents inductor from compiling the projections around the GDN core,
which would introduce numerical divergence that compounds through
the recurrent state.
"""
forward_context: ForwardContext = get_forward_context()
self = forward_context.no_compile_layers[layer_name]
self._full_forward(
hidden_states=hidden_states,
output=output,
)
def olmo_hybrid_gdn_full_forward_fake(
hidden_states: torch.Tensor,
output: torch.Tensor,
layer_name: str,
) -> None:
"""Fake implementation for torch.compile."""
return
direct_register_custom_op(
op_name="olmo_hybrid_gdn_full_forward",
op_func=olmo_hybrid_gdn_full_forward,
mutates_args=["output"],
fake_impl=olmo_hybrid_gdn_full_forward_fake,
)
@triton.jit
def fused_olmo_hybrid_gdn_gating_kernel(
g,
beta_output,
A_log,
a,
b,
dt_bias,
seq_len,
allow_neg_eigval: tl.constexpr,
NUM_HEADS: tl.constexpr,
beta: tl.constexpr,
threshold: tl.constexpr,
BLK_HEADS: tl.constexpr,
):
i_b, i_s, i_d = tl.program_id(0), tl.program_id(1), tl.program_id(2)
head_off = i_d * BLK_HEADS + tl.arange(0, BLK_HEADS)
off = i_b * seq_len * NUM_HEADS + i_s * NUM_HEADS + head_off
mask = head_off < NUM_HEADS
blk_A_log = tl.load(A_log + head_off, mask=mask)
blk_a = tl.load(a + off, mask=mask)
blk_b = tl.load(b + off, mask=mask)
blk_bias = tl.load(dt_bias + head_off, mask=mask)
# g = -self.A_log.float().exp() * F.softplus(a.float() + self.dt_bias)
x = blk_a.to(tl.float32) + blk_bias.to(tl.float32)
softplus_x = tl.where(
beta * x <= threshold, (1 / beta) * tl.log(1 + tl.exp(beta * x)), x
)
blk_g = -tl.exp(blk_A_log.to(tl.float32)) * softplus_x
tl.store(g + off, blk_g.to(g.dtype.element_ty), mask=mask)
# beta = self.b_proj(hidden_states).sigmoid()
# if self.allow_neg_eigval: beta = beta * 2.0
blk_beta_output = tl.sigmoid(blk_b.to(tl.float32))
if allow_neg_eigval:
blk_beta_output = blk_beta_output * 2.0
tl.store(
beta_output + off, blk_beta_output.to(beta_output.dtype.element_ty), mask=mask
)
def fused_olmo_hybrid_gdn_gating(
A_log: torch.Tensor,
a: torch.Tensor,
b: torch.Tensor,
dt_bias: torch.Tensor,
allow_neg_eigval: bool = False,
beta: float = 1.0,
threshold: float = 20.0,
) -> tuple[torch.Tensor, torch.Tensor]:
batch, num_heads = a.shape
seq_len = 1
grid = (batch, seq_len, triton.cdiv(num_heads, 8))
g = torch.empty(1, batch, num_heads, dtype=torch.float32, device=a.device)
beta_output = torch.empty(1, batch, num_heads, dtype=torch.float32, device=b.device)
fused_olmo_hybrid_gdn_gating_kernel[grid](
g,
beta_output,
A_log,
a,
b,
dt_bias,
seq_len,
allow_neg_eigval,
num_heads,
beta,
threshold,
8,
num_warps=1,
)
return g, beta_output
@@ -5,7 +5,6 @@
import torch
from einops import rearrange
from torch import nn
from transformers.activations import ACT2FN
from vllm import envs
from vllm._aiter_ops import rocm_aiter_ops
@@ -15,8 +14,6 @@ from vllm.config import (
)
from vllm.distributed import (
divide,
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
@@ -37,10 +34,9 @@ from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
RowParallelLinear,
)
from vllm.model_executor.layers.mamba.abstract import MambaBase
from vllm.model_executor.layers.mamba.gdn.base import GatedDeltaNetAttention
from vllm.model_executor.layers.mamba.mamba_mixer2 import mamba_v2_sharded_weight_loader
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateDtypeCalculator,
MambaStateShapeCalculator,
is_conv_state_dim_first,
)
@@ -55,7 +51,6 @@ from vllm.model_executor.layers.quantization.inc import INCConfig
from vllm.model_executor.model_loader.weight_utils import (
sharded_weight_loader,
)
from vllm.model_executor.models.utils import extract_layer_index
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.transformers_utils.configs.qwen3_next import Qwen3NextConfig
@@ -67,7 +62,6 @@ from vllm.utils.torch_utils import (
direct_register_custom_op,
)
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
# Optional ROCm AITER Triton kernels for the GDN decode fast-path.
# Availability is checked centrally via rocm_aiter_ops; the actual function
@@ -263,20 +257,11 @@ class ChunkGatedDeltaRule(CustomOp):
)
@PluggableLayer.register("gated_delta_net_attention")
class GatedDeltaNetAttention(PluggableLayer, MambaBase):
@property
def mamba_type(self) -> MambaAttentionBackendEnum:
return MambaAttentionBackendEnum.GDN_ATTN
def get_state_dtype(self) -> tuple[torch.dtype, torch.dtype]:
return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
self.model_config.dtype,
self.cache_config.mamba_cache_dtype,
self.cache_config.mamba_ssm_cache_dtype,
)
def get_state_shape(self) -> tuple[tuple[int, ...], tuple[int, ...]]:
@PluggableLayer.register("qwen_gated_delta_net_attention")
class QwenGatedDeltaNetAttention(GatedDeltaNetAttention):
def get_state_shape(
self,
) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
return MambaStateShapeCalculator.gated_delta_net_state_shape(
self.tp_size,
self.num_k_heads,
@@ -294,33 +279,15 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
prefix: str = "",
gqa_interleaved_layout=False,
) -> None:
super().__init__()
self.tp_size = get_tensor_model_parallel_world_size()
self.tp_rank = get_tensor_model_parallel_rank()
self.hidden_size = config.hidden_size
self.num_v_heads = config.linear_num_value_heads
super().__init__(config, vllm_config, prefix)
self.num_k_heads = config.linear_num_key_heads
self.num_v_heads = config.linear_num_value_heads
self.head_k_dim = config.linear_key_head_dim
self.head_v_dim = config.linear_value_head_dim
self.conv_kernel_size = config.linear_conv_kernel_dim
self.key_dim = self.head_k_dim * self.num_k_heads
self.value_dim = self.head_v_dim * self.num_v_heads
self.conv_kernel_size = config.linear_conv_kernel_dim
self.layer_idx = extract_layer_index(prefix)
self.activation = config.hidden_act
self.act = ACT2FN[config.hidden_act]
self.layer_norm_epsilon = config.rms_norm_eps
self.prefix = prefix
self.config = config
self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
self.speculative_config = vllm_config.speculative_config
self.num_spec = (
self.speculative_config.num_speculative_tokens
if self.speculative_config
else 0
)
self.gqa_interleaved_layout = gqa_interleaved_layout
if current_platform.is_xpu():
self._forward_method = self.forward_xpu
@@ -355,7 +322,7 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
hidden_size=self.hidden_size,
key_dim=self.key_dim,
value_dim=self.value_dim,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.in_proj_qkvz",
)
@@ -365,10 +332,10 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
self.in_proj_ba = self.create_ba_proj(
hidden_size=self.hidden_size,
num_v_heads=self.num_v_heads,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.in_proj_ba",
)
self.disable_tp_for_ba_proj = self.maybe_disable_tp(quant_config)
self.disable_tp_for_ba_proj = self.maybe_disable_tp(self.quant_config)
query_key_settings = (self.key_dim, 0, False)
value_settings = (self.value_dim, 0, False)
@@ -421,7 +388,7 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
self.hidden_size,
bias=False,
input_is_parallel=True,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}.out_proj",
)
@@ -765,7 +732,7 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
device=projected_states_qkvz.device,
)
torch.ops.vllm.gdn_attention_core(
torch.ops.vllm.qwen_gdn_attention_core(
projected_states_qkvz,
projected_states_ba,
z,
@@ -826,7 +793,7 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
device=hidden_states.device,
)
torch.ops.vllm.gdn_attention_core(
torch.ops.vllm.qwen_gdn_attention_core(
mixed_qkv,
b,
a,
@@ -1492,7 +1459,7 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
return
def gdn_attention_core(
def qwen_gdn_attention_core(
qkv_or_qkvz: torch.Tensor,
b_or_ba: torch.Tensor,
a_or_z_out: torch.Tensor,
@@ -1545,8 +1512,8 @@ def gdn_attention_core_fake(
direct_register_custom_op(
op_name="gdn_attention_core",
op_func=gdn_attention_core,
op_name="qwen_gdn_attention_core",
op_func=qwen_gdn_attention_core,
mutates_args=["a_or_z_out", "core_attn_out"],
fake_impl=gdn_attention_core_fake,
)
+13 -27
View File
@@ -7,7 +7,7 @@ import torch
from torch import nn
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, ParallelConfig, VllmConfig
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
get_pp_group,
get_tensor_model_parallel_world_size,
@@ -18,7 +18,6 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoE,
fused_moe_make_expert_params_mapping,
)
from vllm.model_executor.layers.kda import KimiDeltaAttention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
@@ -27,6 +26,9 @@ from vllm.model_executor.layers.linear import (
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.gdn.kimi_gdn_linear_attn import (
KimiGatedDeltaNetAttention,
)
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateCopyFunc,
MambaStateCopyFuncCalculator,
@@ -286,26 +288,22 @@ class KimiDecoderLayer(nn.Module):
def __init__(
self,
config: KimiLinearConfig,
layer_idx: int,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
parallel_config: ParallelConfig | None = None,
model_config: ModelConfig | None = None,
vllm_config: VllmConfig,
prefix: str = "",
**kwargs,
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
self.is_moe = config.is_moe
layer_idx = int(prefix.rsplit(".", 1)[1])
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
if config.is_kda_layer(layer_idx):
self.self_attn = KimiDeltaAttention(
layer_idx=layer_idx,
hidden_size=config.hidden_size,
quant_config=quant_config,
cache_config=cache_config,
model_config=config,
self.self_attn = KimiGatedDeltaNetAttention(
config,
vllm_config,
prefix=f"{prefix}.self_attn",
)
else:
@@ -385,10 +383,6 @@ class KimiLinearModel(nn.Module):
super().__init__()
config = vllm_config.model_config.hf_text_config
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
parallel_config = vllm_config.parallel_config
self.config = config
self.vocab_size = config.vocab_size
@@ -402,19 +396,11 @@ class KimiLinearModel(nn.Module):
else:
self.embed_tokens = PPMissingLayer()
extra_kwargs = {}
def get_layer(prefix: str):
layer_idx = int(prefix.rsplit(".", 1)[1])
return KimiDecoderLayer(
config,
layer_idx,
cache_config,
quant_config,
parallel_config,
model_config,
vllm_config,
prefix,
**extra_kwargs,
)
self.start_layer, self.end_layer, self.layers = make_layers(
+6 -645
View File
@@ -26,55 +26,38 @@ from functools import partial
from itertools import islice
import torch
from einops import rearrange
from torch import nn
from transformers.activations import ACT2FN
from vllm.compilation.decorators import support_torch_compile
from vllm.config import (
CacheConfig,
ModelConfig,
SpeculativeConfig,
VllmConfig,
get_current_vllm_config,
)
from vllm.distributed import (
divide,
get_pp_group,
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather,
)
from vllm.distributed.utils import split_tensor_along_last_dim
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fla.ops import (
chunk_gated_delta_rule,
fused_recurrent_gated_delta_rule,
)
from vllm.model_executor.layers.layernorm import RMSNorm, RMSNormGated
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.abstract import MambaBase
from vllm.model_executor.layers.mamba.gdn.olmo_gdn_linear_attn import (
OlmoHybridGatedDeltaNetAttention,
)
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateCopyFunc,
MambaStateCopyFuncCalculator,
MambaStateDtypeCalculator,
MambaStateShapeCalculator,
is_conv_state_dim_first,
)
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
causal_conv1d_fn,
causal_conv1d_update,
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead,
@@ -82,17 +65,8 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
)
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader,
sharded_weight_loader,
)
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.triton_utils import tl, triton
from vllm.triton_utils.allocation import set_triton_allocator
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backend import AttentionMetadata
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
from .interfaces import HasInnerState, IsHybrid, SupportsLoRA, SupportsPP
from .utils import (
@@ -107,502 +81,6 @@ from .utils import (
logger = init_logger(__name__)
def _make_fused_conv1d_weight_loader(dims, tp_size, tp_rank):
"""Weight loader for loading separate HF conv weights into a fused conv1d.
dims: list of original (un-sharded) dims per section,
e.g. [key_dim, key_dim, value_dim]
"""
sharded_dims = [d // tp_size for d in dims]
def weight_loader(param, loaded_weight, loaded_shard_id=None):
if loaded_weight.dim() == 2:
loaded_weight = loaded_weight.unsqueeze(1)
dim = dims[loaded_shard_id]
shard_size = dim // tp_size
tp_start = tp_rank * shard_size
sharded_weight = loaded_weight[tp_start : tp_start + shard_size]
offset = sum(sharded_dims[:loaded_shard_id])
param.data[offset : offset + shard_size].copy_(sharded_weight)
return weight_loader
class OlmoHybridGatedDeltaNet(nn.Module, MambaBase):
"""
Gated DeltaNet linear attention layer for OLMo Hybrid.
This implements the linear attention mechanism that replaces sliding window
attention in the hybrid architecture.
"""
@property
def mamba_type(self) -> MambaAttentionBackendEnum:
return MambaAttentionBackendEnum.GDN_ATTN
def get_state_dtype(self) -> tuple[torch.dtype, torch.dtype]:
return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
self.model_config.dtype,
self.cache_config.mamba_cache_dtype,
self.cache_config.mamba_ssm_cache_dtype,
)
def get_state_shape(self) -> tuple[tuple[int, ...], tuple[int, ...]]:
return MambaStateShapeCalculator.gated_delta_net_state_shape(
self.tp_size,
self.num_k_heads,
self.num_v_heads,
self.head_k_dim,
self.head_v_dim,
self.conv_kernel_size,
self.num_spec,
)
def __init__(
self,
config,
model_config: ModelConfig | None = None,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
speculative_config: SpeculativeConfig | None = None,
prefix: str = "",
) -> None:
super().__init__()
self.tp_size = get_tensor_model_parallel_world_size()
self.tp_rank = get_tensor_model_parallel_rank()
self.hidden_size = config.hidden_size
self.num_v_heads = config.linear_num_value_heads
self.num_k_heads = config.linear_num_key_heads
self.head_k_dim = config.linear_key_head_dim
self.head_v_dim = config.linear_value_head_dim
self.key_dim = self.head_k_dim * self.num_k_heads
self.value_dim = self.head_v_dim * self.num_v_heads
self.conv_kernel_size = config.linear_conv_kernel_dim
self.layer_idx = extract_layer_index(prefix)
self.activation = config.hidden_act
self.act = ACT2FN[config.hidden_act]
self.layer_norm_epsilon = config.rms_norm_eps
assert getattr(config, "linear_use_gate", True), (
"OlmoHybridGatedDeltaNet requires linear_use_gate=True"
)
self.allow_neg_eigval = getattr(config, "linear_allow_neg_eigval", False)
self.prefix = prefix
self.config = config
self.model_config = model_config
self.cache_config = cache_config
self.quant_config = quant_config
self.speculative_config = speculative_config
self.num_spec = (
self.speculative_config.num_speculative_tokens
if self.speculative_config
else 0
)
# Fused QKVG projection: 1 matmul instead of 4
self.in_proj_qkvg = MergedColumnParallelLinear(
input_size=self.hidden_size,
output_sizes=[self.key_dim, self.key_dim, self.value_dim, self.value_dim],
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.in_proj_qkvg",
)
# Separate B and A projections to preserve numerical precision.
# Fusing these into one matmul changes FP accumulation order for the
# gating scalars, which compounds through the GDN recurrent state.
self.b_proj = ColumnParallelLinear(
input_size=self.hidden_size,
output_size=self.num_v_heads,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.b_proj",
)
self.a_proj = ColumnParallelLinear(
input_size=self.hidden_size,
output_size=self.num_v_heads,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.a_proj",
)
# Fused conv1d: single parameter instead of 3
self.conv_dim = self.key_dim * 2 + self.value_dim
self.conv1d = ColumnParallelLinear(
input_size=self.conv_kernel_size,
output_size=self.conv_dim,
bias=False,
prefix=f"{prefix}.conv1d",
)
self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1)
delattr(self.conv1d.weight, "weight_loader")
set_weight_attrs(
self.conv1d.weight,
{
"weight_loader": _make_fused_conv1d_weight_loader(
[self.key_dim, self.key_dim, self.value_dim],
self.tp_size,
self.tp_rank,
)
},
)
self.dt_bias = nn.Parameter(
torch.ones(self.num_v_heads // self.tp_size),
)
self.A_log = nn.Parameter(
torch.empty(
divide(self.num_v_heads, self.tp_size),
)
)
set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)})
set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)})
# use eps=1e-5 to match FLA's FusedRMSNormGated
self.o_norm = RMSNormGated(
self.head_v_dim,
eps=1e-5,
group_size=None,
norm_before_gate=True,
device=current_platform.current_device(),
dtype=config.torch_dtype if hasattr(config, "torch_dtype") else None,
)
self.o_proj = RowParallelLinear(
self.value_dim,
self.hidden_size,
bias=False,
input_is_parallel=True,
quant_config=quant_config,
prefix=f"{prefix}.o_proj",
)
# FLA triton kernels need a PyTorch-backed allocator for scratch
# memory (required by triton >= 3.x autotuner). Set once at init.
set_triton_allocator(current_platform.current_device())
compilation_config = get_current_vllm_config().compilation_config
if prefix in compilation_config.static_forward_context:
raise ValueError(f"Duplicate layer name: {prefix}")
compilation_config.static_forward_context[prefix] = self
def rearrange_mixed_qkv(self, mixed_qkv):
if mixed_qkv is None:
return None, None, None
query, key, value = torch.split(
mixed_qkv,
[
self.key_dim // self.tp_size,
self.key_dim // self.tp_size,
self.value_dim // self.tp_size,
],
dim=-1,
)
num_k_heads = self.num_k_heads // self.tp_size
num_v_heads = self.num_v_heads // self.tp_size
query = rearrange(query, "l (h d) -> 1 l h d", h=num_k_heads, d=self.head_k_dim)
key = rearrange(key, "l (h d) -> 1 l h d", h=num_k_heads, d=self.head_k_dim)
value = rearrange(value, "l (h d) -> 1 l h d", h=num_v_heads, d=self.head_v_dim)
# GQA expansion if needed
if num_v_heads > num_k_heads:
expand_ratio = num_v_heads // num_k_heads
query = query.unsqueeze(3).expand(-1, -1, -1, expand_ratio, -1)
query = query.reshape(1, query.shape[1], num_v_heads, self.head_k_dim)
key = key.unsqueeze(3).expand(-1, -1, -1, expand_ratio, -1)
key = key.reshape(1, key.shape[1], num_v_heads, self.head_k_dim)
return query.contiguous(), key.contiguous(), value.contiguous()
def forward(
self,
hidden_states: torch.Tensor,
output: torch.Tensor,
):
# NOTE: We wrap the ENTIRE linear attention forward (projections +
# core recurrence + output norm + output projection) in a single
# custom op, rather than just wrapping the recurrent core like
# other GDN models (e.g. Qwen3Next) do.
#
# Why: torch.compile with inductor generates fused kernels for
# matmuls and pointwise ops. These fused kernels can differ in
# floating-point accumulation order from eager-mode cuBLAS,
# introducing small numerical differences (~1e-7 per op). For
# standard transformer attention this is harmless because each
# position is computed independently. But for the GDN recurrent
# state, these tiny input differences compound at every timestep
# across the full sequence length, causing severe logprob
# divergence (e.g. ~15% top-1 agreement with eager baseline).
#
# By making the full forward opaque to inductor, the projections
# and output norm run with eager-mode kernels (cuBLAS, triton),
# preserving numerical consistency. The tradeoff is reduced
# compilation speedup (~1.5x vs ~3x), but logprob agreement
# improves from ~15% to ~83% top-1 vs eager.
#
# The remaining ~17% divergence comes from inductor compiling
# the MLP and transformer attention layers that are NOT wrapped
# in custom ops -- their small precision differences propagate
# as inputs to the GDN layers from outside.
torch.ops.vllm.olmo_hybrid_gdn_full_forward(
hidden_states,
output,
self.prefix,
)
def _full_forward(
self,
hidden_states: torch.Tensor,
output: torch.Tensor,
):
num_tokens = hidden_states.size(0)
# ============================================================
# Part 1: Input Projection (2 fused matmuls instead of 6)
# ============================================================
projected_qkvg, _ = self.in_proj_qkvg(hidden_states)
conv_dim_sharded = (self.key_dim * 2 + self.value_dim) // self.tp_size
mixed_qkv = projected_qkvg[..., :conv_dim_sharded]
gate = projected_qkvg[..., conv_dim_sharded:]
b, _ = self.b_proj(hidden_states)
a, _ = self.a_proj(hidden_states)
# ============================================================
# Part 2: Core Attention
# ============================================================
core_attn_out = torch.zeros(
(num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim),
dtype=hidden_states.dtype,
device=hidden_states.device,
)
self._forward_core(
mixed_qkv=mixed_qkv,
b=b,
a=a,
core_attn_out=core_attn_out,
)
# ============================================================
# Part 3: Output Projection
# ============================================================
gate = gate.view(num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim)
core_attn_out_flat = core_attn_out.reshape(-1, core_attn_out.shape[-1])
gate_flat = gate.reshape(-1, gate.shape[-1])
core_attn_out_normed = self.o_norm(core_attn_out_flat, gate_flat)
core_attn_out = core_attn_out_normed.view(
num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim
)
core_attn_out = rearrange(core_attn_out, "l h d -> l (h d)")
output[:num_tokens], _ = self.o_proj(core_attn_out)
def _forward_core(
self,
mixed_qkv: torch.Tensor,
b: torch.Tensor,
a: torch.Tensor,
core_attn_out: torch.Tensor,
):
"""
Core attention computation (called by custom op).
"""
forward_context = get_forward_context()
attn_metadata: AttentionMetadata = forward_context.attn_metadata
if attn_metadata is None:
# V1 profile run
return
assert isinstance(attn_metadata, dict)
attn_metadata = attn_metadata[self.prefix]
assert isinstance(attn_metadata, GDNAttentionMetadata)
has_initial_state = attn_metadata.has_initial_state
spec_query_start_loc = attn_metadata.spec_query_start_loc
non_spec_query_start_loc = attn_metadata.non_spec_query_start_loc
spec_sequence_masks = attn_metadata.spec_sequence_masks
spec_token_indx = attn_metadata.spec_token_indx
non_spec_token_indx = attn_metadata.non_spec_token_indx
spec_state_indices_tensor = attn_metadata.spec_state_indices_tensor
non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor
self_kv_cache = self.kv_cache
# conv_state must be (..., dim, width-1) for the conv kernels.
# DS layout stores it that way directly; SD layout needs a transpose.
conv_state = (
self_kv_cache[0]
if is_conv_state_dim_first()
else self_kv_cache[0].transpose(-1, -2)
)
ssm_state = self_kv_cache[1]
num_actual_tokens = attn_metadata.num_actual_tokens
num_accepted_tokens = attn_metadata.num_accepted_tokens
mixed_qkv = mixed_qkv[:num_actual_tokens]
b = b[:num_actual_tokens]
a = a[:num_actual_tokens]
conv_weights = self.conv1d.weight.view(
self.conv1d.weight.size(0), self.conv1d.weight.size(2)
)
if spec_sequence_masks is not None:
if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
mixed_qkv_spec = mixed_qkv
mixed_qkv_non_spec = None
else:
mixed_qkv_spec = mixed_qkv.index_select(0, spec_token_indx)
mixed_qkv_non_spec = mixed_qkv.index_select(0, non_spec_token_indx)
else:
mixed_qkv_spec = None
mixed_qkv_non_spec = mixed_qkv
if spec_sequence_masks is not None:
mixed_qkv_spec = causal_conv1d_update(
mixed_qkv_spec,
conv_state,
conv_weights,
None, # no bias
self.activation,
conv_state_indices=spec_state_indices_tensor[:, 0][
: attn_metadata.num_spec_decodes
],
num_accepted_tokens=num_accepted_tokens,
query_start_loc=spec_query_start_loc,
max_query_len=spec_state_indices_tensor.size(-1),
validate_data=False,
)
if attn_metadata.num_prefills > 0:
mixed_qkv_non_spec_T = mixed_qkv_non_spec.transpose(0, 1)
mixed_qkv_non_spec = causal_conv1d_fn(
mixed_qkv_non_spec_T,
conv_weights,
None,
activation=self.activation,
conv_states=conv_state,
has_initial_state=has_initial_state,
cache_indices=non_spec_state_indices_tensor,
query_start_loc=non_spec_query_start_loc,
metadata=attn_metadata,
).transpose(0, 1)
elif attn_metadata.num_decodes > 0:
mixed_qkv_non_spec = causal_conv1d_update(
mixed_qkv_non_spec,
conv_state,
conv_weights,
None,
self.activation,
conv_state_indices=non_spec_state_indices_tensor[
: attn_metadata.num_decodes
],
validate_data=True,
)
else:
mixed_qkv_non_spec = None
query_spec, key_spec, value_spec = self.rearrange_mixed_qkv(mixed_qkv_spec)
query_non_spec, key_non_spec, value_non_spec = self.rearrange_mixed_qkv(
mixed_qkv_non_spec
)
g, beta = fused_olmo_hybrid_gdn_gating(
self.A_log, a, b, self.dt_bias, self.allow_neg_eigval
)
if spec_sequence_masks is not None:
if attn_metadata.num_prefills == 0 and attn_metadata.num_decodes == 0:
g_spec = g
beta_spec = beta
g_non_spec = None
beta_non_spec = None
else:
g_spec = g.index_select(1, spec_token_indx)
beta_spec = beta.index_select(1, spec_token_indx)
g_non_spec = g.index_select(1, non_spec_token_indx)
beta_non_spec = beta.index_select(1, non_spec_token_indx)
else:
g_spec = None
beta_spec = None
g_non_spec = g
beta_non_spec = beta
if spec_sequence_masks is not None:
core_attn_out_spec, last_recurrent_state = fused_recurrent_gated_delta_rule(
q=query_spec,
k=key_spec,
v=value_spec,
g=g_spec,
beta=beta_spec,
initial_state=ssm_state,
inplace_final_state=True,
cu_seqlens=spec_query_start_loc[: attn_metadata.num_spec_decodes + 1],
ssm_state_indices=spec_state_indices_tensor,
num_accepted_tokens=num_accepted_tokens,
use_qk_l2norm_in_kernel=True,
)
else:
core_attn_out_spec, last_recurrent_state = None, None
if attn_metadata.num_prefills > 0:
initial_state = ssm_state[non_spec_state_indices_tensor].contiguous()
initial_state[~has_initial_state, ...] = 0
(
core_attn_out_non_spec,
last_recurrent_state,
) = chunk_gated_delta_rule(
q=query_non_spec,
k=key_non_spec,
v=value_non_spec,
g=g_non_spec,
beta=beta_non_spec,
initial_state=initial_state,
output_final_state=True,
cu_seqlens=non_spec_query_start_loc,
use_qk_l2norm_in_kernel=True,
)
ssm_state[non_spec_state_indices_tensor] = last_recurrent_state.to(
ssm_state.dtype
)
elif attn_metadata.num_decodes > 0:
core_attn_out_non_spec, last_recurrent_state = (
fused_recurrent_gated_delta_rule(
q=query_non_spec,
k=key_non_spec,
v=value_non_spec,
g=g_non_spec,
beta=beta_non_spec,
initial_state=ssm_state,
inplace_final_state=True,
cu_seqlens=non_spec_query_start_loc[
: attn_metadata.num_decodes + 1
],
ssm_state_indices=non_spec_state_indices_tensor,
use_qk_l2norm_in_kernel=True,
)
)
else:
core_attn_out_non_spec, last_recurrent_state = None, None
if spec_sequence_masks is not None and core_attn_out_non_spec is not None:
merged_out = torch.empty(
(1, num_actual_tokens, *core_attn_out_spec.shape[2:]),
dtype=core_attn_out_non_spec.dtype,
device=core_attn_out_non_spec.device,
)
merged_out.index_copy_(1, spec_token_indx, core_attn_out_spec)
merged_out.index_copy_(1, non_spec_token_indx, core_attn_out_non_spec)
core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
elif spec_sequence_masks is not None:
core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
else:
core_attn_out[:num_actual_tokens] = core_attn_out_non_spec.squeeze(0)
class OlmoHybridAttention(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -750,22 +228,15 @@ class OlmoHybridDecoderLayer(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
super().__init__()
config = vllm_config.model_config.hf_config
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
speculative_config = vllm_config.speculative_config
layer_idx = extract_layer_index(prefix)
self.layer_type = config.layer_types[layer_idx]
self.layer_idx = layer_idx
if self.layer_type == "linear_attention":
self.linear_attn = OlmoHybridGatedDeltaNet(
self.linear_attn = OlmoHybridGatedDeltaNetAttention(
config,
model_config=model_config,
cache_config=cache_config,
quant_config=quant_config,
speculative_config=speculative_config,
vllm_config,
prefix=f"{prefix}.linear_attn",
)
self.input_layernorm = RMSNorm(
@@ -1068,113 +539,3 @@ class OlmoHybridForCausalLM(
),
)
return loader.load_weights(weights)
def olmo_hybrid_gdn_full_forward(
hidden_states: torch.Tensor,
output: torch.Tensor,
layer_name: str,
) -> None:
"""Full linear attention forward wrapped as a custom op.
Prevents inductor from compiling the projections around the GDN core,
which would introduce numerical divergence that compounds through
the recurrent state.
"""
forward_context: ForwardContext = get_forward_context()
self = forward_context.no_compile_layers[layer_name]
self._full_forward(
hidden_states=hidden_states,
output=output,
)
def olmo_hybrid_gdn_full_forward_fake(
hidden_states: torch.Tensor,
output: torch.Tensor,
layer_name: str,
) -> None:
"""Fake implementation for torch.compile."""
return
direct_register_custom_op(
op_name="olmo_hybrid_gdn_full_forward",
op_func=olmo_hybrid_gdn_full_forward,
mutates_args=["output"],
fake_impl=olmo_hybrid_gdn_full_forward_fake,
)
@triton.jit
def fused_olmo_hybrid_gdn_gating_kernel(
g,
beta_output,
A_log,
a,
b,
dt_bias,
seq_len,
allow_neg_eigval: tl.constexpr,
NUM_HEADS: tl.constexpr,
beta: tl.constexpr,
threshold: tl.constexpr,
BLK_HEADS: tl.constexpr,
):
i_b, i_s, i_d = tl.program_id(0), tl.program_id(1), tl.program_id(2)
head_off = i_d * BLK_HEADS + tl.arange(0, BLK_HEADS)
off = i_b * seq_len * NUM_HEADS + i_s * NUM_HEADS + head_off
mask = head_off < NUM_HEADS
blk_A_log = tl.load(A_log + head_off, mask=mask)
blk_a = tl.load(a + off, mask=mask)
blk_b = tl.load(b + off, mask=mask)
blk_bias = tl.load(dt_bias + head_off, mask=mask)
# g = -self.A_log.float().exp() * F.softplus(a.float() + self.dt_bias)
x = blk_a.to(tl.float32) + blk_bias.to(tl.float32)
softplus_x = tl.where(
beta * x <= threshold, (1 / beta) * tl.log(1 + tl.exp(beta * x)), x
)
blk_g = -tl.exp(blk_A_log.to(tl.float32)) * softplus_x
tl.store(g + off, blk_g.to(g.dtype.element_ty), mask=mask)
# beta = self.b_proj(hidden_states).sigmoid()
# if self.allow_neg_eigval: beta = beta * 2.0
blk_beta_output = tl.sigmoid(blk_b.to(tl.float32))
if allow_neg_eigval:
blk_beta_output = blk_beta_output * 2.0
tl.store(
beta_output + off, blk_beta_output.to(beta_output.dtype.element_ty), mask=mask
)
def fused_olmo_hybrid_gdn_gating(
A_log: torch.Tensor,
a: torch.Tensor,
b: torch.Tensor,
dt_bias: torch.Tensor,
allow_neg_eigval: bool = False,
beta: float = 1.0,
threshold: float = 20.0,
) -> tuple[torch.Tensor, torch.Tensor]:
batch, num_heads = a.shape
seq_len = 1
grid = (batch, seq_len, triton.cdiv(num_heads, 8))
g = torch.empty(1, batch, num_heads, dtype=torch.float32, device=a.device)
beta_output = torch.empty(1, batch, num_heads, dtype=torch.float32, device=b.device)
fused_olmo_hybrid_gdn_gating_kernel[grid](
g,
beta_output,
A_log,
a,
b,
dt_bias,
seq_len,
allow_neg_eigval,
num_heads,
beta,
threshold,
8,
num_warps=1,
)
return g, beta_output
+4 -2
View File
@@ -40,7 +40,9 @@ from vllm.model_executor.layers.layernorm import (
GemmaRMSNorm as Qwen3_5RMSNorm,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.gdn_linear_attn import GatedDeltaNetAttention
from vllm.model_executor.layers.mamba.gdn.qwen_gdn_linear_attn import (
QwenGatedDeltaNetAttention,
)
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateCopyFunc,
MambaStateCopyFuncCalculator,
@@ -133,7 +135,7 @@ class Qwen3_5DecoderLayer(Qwen3NextDecoderLayer):
self.layer_idx = extract_layer_index(prefix)
if self.layer_type == "linear_attention":
self.linear_attn = GatedDeltaNetAttention(
self.linear_attn = QwenGatedDeltaNetAttention(
config=config,
vllm_config=vllm_config,
prefix=f"{prefix}.linear_attn",
+4 -2
View File
@@ -37,7 +37,9 @@ from vllm.model_executor.layers.linear import (
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.gdn_linear_attn import GatedDeltaNetAttention
from vllm.model_executor.layers.mamba.gdn.qwen_gdn_linear_attn import (
QwenGatedDeltaNetAttention,
)
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateCopyFunc,
MambaStateCopyFuncCalculator,
@@ -337,7 +339,7 @@ class Qwen3NextDecoderLayer(nn.Module):
self.layer_idx = extract_layer_index(prefix)
if self.layer_type == "linear_attention":
self.linear_attn = GatedDeltaNetAttention(
self.linear_attn = QwenGatedDeltaNetAttention(
config,
vllm_config=vllm_config,
prefix=f"{prefix}.linear_attn",