Files
enginex-ascend-910-vllm/vllm_ascend/ops/gdn.py
Sun Ruoxi 7f8a1b1f7a init v0.23.0
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00

455 lines
21 KiB
Python

#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# This file is a part of the vllm-ascend project.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import torch
from einops import rearrange
from vllm.distributed import get_pcp_group
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.fla.ops.l2norm import l2norm_fwd
from vllm.model_executor.layers.mamba.gdn.base import GatedDeltaNetAttention
from vllm.model_executor.layers.mamba.mamba_utils import MambaStateShapeCalculator
from vllm.triton_utils import triton
from vllm.v1.attention.backend import AttentionBackend, AttentionMetadata # type: ignore
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
from vllm.v1.attention.backends.utils import PAD_SLOT_ID
from vllm_ascend.attention.utils import maybe_save_kv_layer_to_connector
from vllm_ascend.device.device_op import DeviceOperator
from vllm_ascend.ops.gdn_attn_builder import AscendGDNAttentionBackend
from vllm_ascend.ops.triton.fla.chunk import chunk_gated_delta_rule
from vllm_ascend.ops.triton.fla.fused_qkvzba_split_reshape import fused_qkvzba_split_reshape_cat
from vllm_ascend.ops.triton.fla.utils import clear_ssm_states
from vllm_ascend.ops.triton.mamba.causal_conv1d import extract_last_width
class AscendGatedDeltaNetAttention(GatedDeltaNetAttention):
def _split_ba_for_tp(self, ba: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
if hasattr(self, "split_ba"):
return self.split_ba(ba)
return ba.chunk(2, dim=-1)
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 _warmup_prefill_kernels(self, qkv_or_qkvz: torch.Tensor, v_dim: int) -> None:
return
def _warmup_prefill_kernels_v0202(self, mixed_qkv: torch.Tensor) -> None:
return
def get_attn_backend(self) -> type[AttentionBackend]:
return AscendGDNAttentionBackend
def forward(
self,
hidden_states: torch.Tensor,
output: torch.Tensor,
):
"""
Forward pass with three parts:
1. Input projection
2. Core attention (custom op)
3. Output projection
"""
num_tokens = hidden_states.size(0)
if hasattr(self, "in_proj_qkv"):
mixed_qkv, _ = self.in_proj_qkv(hidden_states)
ba, _ = self.in_proj_ba(hidden_states)
z, _ = self.in_proj_z(hidden_states)
z = z.reshape(z.size(0), -1, self.head_v_dim)
b, a = self._split_ba_for_tp(ba)
b = b.contiguous()
a = a.contiguous()
else:
if not self.gqa_interleaved_layout:
mixed_qkvz, _ = self.in_proj_qkvz(hidden_states)
num_tokens = mixed_qkvz.size(0)
qkv_size = (self.key_dim * 2 + self.value_dim) // self.tp_size
z_size = self.value_dim // self.tp_size
mixed_qkv, z = mixed_qkvz.split([qkv_size, z_size], dim=-1)
z = z.reshape(z.size(0), -1, self.head_v_dim)
ba, _ = self.in_proj_ba(hidden_states)
b, a = self._split_ba_for_tp(ba)
b = b.contiguous()
a = a.contiguous()
else:
projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states)
projected_states_ba, _ = self.in_proj_ba(hidden_states)
num_tokens = projected_states_qkvz.size(0)
mixed_qkv, z, b, a = fused_qkvzba_split_reshape_cat(
projected_states_qkvz,
projected_states_ba,
triton.cdiv(self.num_k_heads, self.tp_size),
triton.cdiv(self.num_v_heads, self.tp_size),
self.head_k_dim,
self.head_v_dim,
)
# ============================================================
# Part 2: Core Attention (Custom Op)
# ============================================================
# Note: we should not use torch.empty here like other attention backends,
# see discussions in https://github.com/vllm-project/vllm/pull/28182
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,
)
torch.ops.vllm.qwen_gdn_attention_core(
mixed_qkv,
b,
a,
core_attn_out,
self.prefix,
False,
)
# ============================================================
# Part 3: Output Projection
# ============================================================
maybe_save_kv_layer_to_connector("", [])
z_shape_og = z.shape
# Reshape input data into 2D tensor
core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
z = z.reshape(-1, z.shape[-1])
core_attn_out = self.norm(core_attn_out, z)
core_attn_out = core_attn_out.reshape(z_shape_og)
core_attn_out = rearrange(core_attn_out, "... h d -> ... (h d)")
output[:num_tokens], _ = self.out_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)
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 # noqa: E501
non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor # noqa: E501
self_kv_cache = self.kv_cache
ssm_state = self_kv_cache[1]
num_actual_tokens = attn_metadata.num_actual_tokens
mixed_qkv = mixed_qkv[:num_actual_tokens]
b = b[:num_actual_tokens]
a = a[:num_actual_tokens]
# 1. Convolution sequence transformation
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
# 1.1: Process the multi-query part
if spec_sequence_masks is not None:
conv_weights_T = conv_weights.transpose(0, 1)
activation_num = 1 if self.activation else 0
spec_causal_conv1d_meta = attn_metadata.spec_decode_metadata.spec_causal_conv1d
spec_query_start_loc_device = spec_causal_conv1d_meta.query_start_loc
output_spec = torch.empty_like(mixed_qkv_spec)
torch.ops._C_ascend.npu_causal_conv1d_custom(
output_spec,
mixed_qkv_spec,
conv_weights_T,
conv_state=self_kv_cache[0],
bias_opt=self.conv1d.bias,
query_start_loc_opt=spec_query_start_loc_device,
cache_indices_opt=spec_causal_conv1d_meta.cache_indices,
initial_state_mode_opt=None,
num_accepted_tokens_opt=spec_causal_conv1d_meta.num_accepted_tokens,
activation_mode=activation_num,
pad_slot_id=PAD_SLOT_ID,
run_mode=1,
)
mixed_qkv_spec = output_spec
# 1.2: Process the remaining part
if attn_metadata.num_prefills > 0:
if mixed_qkv_non_spec is not None:
non_spec_causal_conv1d_meta = attn_metadata.non_spec_prefill_metadata.causal_conv1d
query_start_loc_opt = non_spec_causal_conv1d_meta.query_start_loc
cache_indices_opt = non_spec_causal_conv1d_meta.cache_indices
initial_state_mode_opt = non_spec_causal_conv1d_meta.initial_state_mode
if get_pcp_group().world_size > 1:
conv_weights_T = conv_weights.transpose(0, 1)
activation_num = 1 if self.activation else 0
non_spec_query_start_loc = attn_metadata.non_spec_query_start_loc
assert non_spec_query_start_loc is not None
non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor
width = conv_weights.shape[1]
state_len = width - 1
num_seqs = non_spec_query_start_loc.shape[0] - 1
prefill_seq_offset = max(0, num_seqs - attn_metadata.num_prefills)
prefill_cache_indices = non_spec_state_indices_tensor[prefill_seq_offset:]
mixed_qkv_non_spec_T = mixed_qkv_non_spec.transpose(0, 1)
last_width_prefill_x = extract_last_width(
mixed_qkv_non_spec_T, non_spec_query_start_loc[prefill_seq_offset:], state_len
)
pcp_rank = get_pcp_group().rank_in_group
all_last_width_prefill_x = get_pcp_group().all_gather(
last_width_prefill_x.unsqueeze(0).contiguous(), 0
)
if pcp_rank > 0 and prefill_cache_indices.shape[0] > 0:
self_kv_cache[0][prefill_cache_indices, :state_len, :] = all_last_width_prefill_x[
pcp_rank - 1, ...
].transpose(-1, -2)
mixed_qkv_non_spec_output = torch.empty_like(mixed_qkv_non_spec)
torch.ops._C_ascend.npu_causal_conv1d_custom(
mixed_qkv_non_spec_output,
mixed_qkv_non_spec,
conv_weights_T,
conv_state=self_kv_cache[0],
bias_opt=self.conv1d.bias,
query_start_loc_opt=query_start_loc_opt,
cache_indices_opt=cache_indices_opt,
initial_state_mode_opt=initial_state_mode_opt,
num_accepted_tokens_opt=None,
activation_mode=activation_num,
pad_slot_id=PAD_SLOT_ID,
run_mode=0,
)
mixed_qkv_non_spec = mixed_qkv_non_spec_output
if prefill_cache_indices.shape[0] > 0:
self_kv_cache[0][prefill_cache_indices, :state_len, :] = all_last_width_prefill_x[
-1, ...
].transpose(-1, -2)
else:
conv_weights_T = conv_weights.transpose(0, 1)
activation_num = 1 if self.activation else 0
mixed_qkv_non_spec_output = torch.empty_like(mixed_qkv_non_spec)
torch.ops._C_ascend.npu_causal_conv1d_custom(
mixed_qkv_non_spec_output,
mixed_qkv_non_spec,
conv_weights_T,
conv_state=self_kv_cache[0],
bias_opt=self.conv1d.bias,
query_start_loc_opt=query_start_loc_opt,
cache_indices_opt=cache_indices_opt,
initial_state_mode_opt=initial_state_mode_opt,
num_accepted_tokens_opt=None,
activation_mode=activation_num,
pad_slot_id=PAD_SLOT_ID,
run_mode=0,
)
mixed_qkv_non_spec = mixed_qkv_non_spec_output
elif attn_metadata.num_decodes > 0:
conv_weights_T = conv_weights.transpose(0, 1)
activation_num = 1 if self.activation else 0
non_spec_causal_conv1d_meta = attn_metadata.non_spec_decode_metadata.causal_conv1d
non_spec_query_start_loc_device = non_spec_causal_conv1d_meta.query_start_loc
output_non_spec = torch.empty_like(mixed_qkv_non_spec)
torch.ops._C_ascend.npu_causal_conv1d_custom(
output_non_spec,
mixed_qkv_non_spec,
conv_weights_T,
conv_state=self_kv_cache[0],
bias_opt=self.conv1d.bias,
query_start_loc_opt=non_spec_query_start_loc_device,
cache_indices_opt=non_spec_causal_conv1d_meta.cache_indices,
initial_state_mode_opt=None,
num_accepted_tokens_opt=None,
activation_mode=activation_num,
pad_slot_id=PAD_SLOT_ID,
run_mode=1,
)
mixed_qkv_non_spec = output_non_spec
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)
# 2. Recurrent attention
g, beta = DeviceOperator.fused_gdn_gating(self.A_log, a, b, self.dt_bias)
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
split_non_spec = (
spec_sequence_masks is None and attn_metadata.num_prefills > 0 and attn_metadata.num_decodes > 0
)
num_decode_tokens = attn_metadata.num_decode_tokens
# 2.1: Process the multi-query part
if spec_sequence_masks is not None:
actual_seq_lengths = attn_metadata.spec_decode_metadata.actual_seq_lengths
query_spec = l2norm_fwd(query_spec)
key_spec = l2norm_fwd(key_spec)
# Dispatches to the vllm-ascend AscendC custom operator
# (csrc/recurrent_gated_delta_rule), NOT the built-in CANN operator.
# The custom op extends dtype support (e.g. float32 state) and is
# loaded at runtime via ASCEND_CUSTOM_OPP_PATH.
core_attn_out_spec = torch.ops._C_ascend.npu_recurrent_gated_delta_rule(
query=query_spec.squeeze(0),
key=key_spec.squeeze(0),
value=value_spec.squeeze(0),
g=g_spec.squeeze(0),
beta=beta_spec.squeeze(0),
state=ssm_state,
scale=key_spec.shape[-1] ** -0.5,
actual_seq_lengths=actual_seq_lengths,
ssm_state_indices=spec_state_indices_tensor.flatten(),
num_accepted_tokens=spec_causal_conv1d_meta.num_accepted_tokens.to(torch.int32),
).unsqueeze(0)
else:
core_attn_out_spec, last_recurrent_state = None, None
# 2.2: Process non-spec-decode part in mixed non-spec batches
if split_non_spec:
assert mixed_qkv_non_spec is not None
assert g_non_spec is not None
assert beta_non_spec is not None
query_decode, key_decode, value_decode = self.rearrange_mixed_qkv(mixed_qkv_non_spec[:num_decode_tokens])
actual_seq_lengths = attn_metadata.non_spec_decode_metadata.actual_seq_lengths
query_decode = l2norm_fwd(query_decode)
key_decode = l2norm_fwd(key_decode)
core_attn_out_decode = torch.ops._C_ascend.npu_recurrent_gated_delta_rule(
query=query_decode.squeeze(0),
key=key_decode.squeeze(0),
value=value_decode.squeeze(0),
g=g_non_spec[:, :num_decode_tokens].squeeze(0),
beta=beta_non_spec[:, :num_decode_tokens].squeeze(0),
state=ssm_state,
scale=key_decode.shape[-1] ** -0.5,
actual_seq_lengths=actual_seq_lengths,
ssm_state_indices=non_spec_state_indices_tensor[: attn_metadata.num_decodes],
).unsqueeze(0)
else:
core_attn_out_decode = None
# 2.3: Process the remaining part
if attn_metadata.num_prefills > 0:
prefill_query_start_loc = attn_metadata.prefill_query_start_loc
prefill_state_indices = attn_metadata.prefill_state_indices
prefill_has_initial_state = attn_metadata.prefill_has_initial_state
assert prefill_query_start_loc is not None
assert prefill_state_indices is not None
assert prefill_has_initial_state is not None
assert g_non_spec is not None
assert beta_non_spec is not None
if split_non_spec:
query_non_spec = query_non_spec[:, num_decode_tokens:]
key_non_spec = key_non_spec[:, num_decode_tokens:]
value_non_spec = value_non_spec[:, num_decode_tokens:]
g_non_spec = g_non_spec[:, num_decode_tokens:]
beta_non_spec = beta_non_spec[:, num_decode_tokens:]
initial_state = ssm_state[prefill_state_indices].transpose(-1, -2).contiguous()
clear_ssm_states(initial_state, prefill_has_initial_state)
(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=prefill_query_start_loc,
prebuilt_meta=attn_metadata.non_spec_prefill_metadata.chunk,
head_first=False,
use_qk_l2norm_in_kernel=True,
)
ssm_state[prefill_state_indices] = last_recurrent_state.transpose(-1, -2).contiguous().to(ssm_state.dtype)
if split_non_spec:
core_attn_out_non_spec = torch.cat(
[core_attn_out_decode, core_attn_out_non_spec],
dim=1,
)
elif attn_metadata.num_decodes > 0:
actual_seq_lengths = attn_metadata.non_spec_decode_metadata.actual_seq_lengths
query_non_spec = l2norm_fwd(query_non_spec)
key_non_spec = l2norm_fwd(key_non_spec)
# Dispatches to the vllm-ascend AscendC custom operator
# (csrc/recurrent_gated_delta_rule), NOT the built-in CANN operator.
core_attn_out_non_spec = torch.ops._C_ascend.npu_recurrent_gated_delta_rule(
query=query_non_spec.squeeze(0),
key=key_non_spec.squeeze(0),
value=value_non_spec.squeeze(0),
g=g_non_spec.squeeze(0) if g_non_spec is not None else g_non_spec,
beta=beta_non_spec.squeeze(0) if beta_non_spec is not None else beta_non_spec,
state=ssm_state,
scale=key_non_spec.shape[-1] ** -0.5,
actual_seq_lengths=actual_seq_lengths,
ssm_state_indices=non_spec_state_indices_tensor,
).unsqueeze(0)
else:
core_attn_out_non_spec, last_recurrent_state = None, None
# 3. Merge core attention output
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)