Files
enginex-ascend-910-vllm/vllm_ascend/attention/sfa_v1.py
Sun Ruoxi 7f8a1b1f7a init v0.23.0
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00

2043 lines
89 KiB
Python

from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, NamedTuple, TypeVar
import scipy # type: ignore
import torch
import torch_npu
import vllm.envs as envs_vllm
from torch import nn
from vllm.config import VllmConfig, get_current_vllm_config
from vllm.distributed import get_tensor_model_parallel_world_size, get_tp_group
from vllm.logger import logger
from vllm.model_executor.layers.attention.mla_attention import MLACommonMetadataBuilder
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
from vllm.triton_utils import HAS_TRITON
from vllm.v1.attention.backend import (
AttentionBackend, # type: ignore
AttentionCGSupport,
MLAAttentionImpl,
)
from vllm.v1.kv_cache_interface import AttentionSpec
from vllm.v1.worker.utils import select_common_block_size
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.ascend_forward_context import _EXTRA_CTX
from vllm_ascend.attention.attention_mask import AttentionMaskBuilder
from vllm_ascend.attention.attention_v1 import AscendAttentionState
from vllm_ascend.attention.context_parallel.common_cp import AscendPCPMetadata
from vllm_ascend.attention.mla_v1 import MAX_O_PROJ_PREFETCH_SIZE, MLAPO_MAX_SUPPORTED_TOKENS
from vllm_ascend.attention.utils import (
SFA_QSFA_TILE_SIZE,
AscendCommonAttentionMetadata,
ascend_chunked_prefill_workspace_size,
enable_cp,
get_sfa_qsfa_packed_head_dim,
maybe_save_kv_layer_to_connector,
notify_kv_cache_written,
trans_rope_weight,
transdata,
wait_for_kv_layer_from_connector,
)
from vllm_ascend.device.device_op import DeviceOperator
from vllm_ascend.device.mxfp_compat import FLOAT8_E8M0FNU_DTYPE
from vllm_ascend.distributed.utils import all_gather_async
from vllm_ascend.memcache_comm_fence import (
record_attention_compute_start,
)
from vllm_ascend.ops.layer_shard_linear import (
is_hidden_layer,
post_process_after_loading_for_shard_weight_series,
reach_layer_for_shard_weight_series,
register_all_layers_to_shard_weight_series,
)
from vllm_ascend.ops.rotary_embedding import get_cos_and_sin_mla
from vllm_ascend.ops.triton.rope import rope_forward_triton_siso
from vllm_ascend.quantization.methods import (
AscendW8A8DynamicLinearMethod,
AscendW8A8LinearMethod,
AscendW8A8MXFP8DynamicLinearMethod,
)
from vllm_ascend.utils import (
ACL_FORMAT_FRACTAL_ND,
ACL_FORMAT_FRACTAL_NZ,
AscendDeviceType,
_round_up,
dispose_layer,
enable_dsa_cp,
enable_dsa_cp_with_layer_shard,
enable_dsa_cp_with_o_proj_tp,
enable_sfa_dcp_replicated_indexer,
enable_sp,
get_ascend_device_type,
get_weight_prefetch_method,
maybe_trans_nz,
)
from vllm_ascend.worker.npu_input_batch import NPUInputBatch
if TYPE_CHECKING:
from vllm.v1.core.sched.output import SchedulerOutput
# token count limits within bmm_transpose operator
BMM_TRANS_MAX_SUPPORTED_TOKENS = 1024
class _ByteGatherPart(NamedTuple):
name: str
shape: tuple[int, ...]
dtype: torch.dtype
num_bytes_per_row: int
O_PROJ_ACLNN_INPUT_PARAMS = (
"aclnn_input_scale",
"aclnn_input_scale_reciprocal",
"aclnn_input_offset",
)
class DCPGatherContext(NamedTuple):
"""State needed to finish an async fused DCP all-gather."""
# The gathered fused tensor.
gathered: torch.Tensor
# Async all-gather work handle. None means the gather completed synchronously.
handle: torch.distributed.Work | None
# Permutation that restores the original dimension order after dim>0 gather.
restore_perm: tuple[int, ...] | None
# Last-dimension sizes used to split the fused tensor after gather.
split_sizes: tuple[int, ...]
def _get_indexer_types(configs: tuple[Any, ...]) -> Any | None:
for config in configs:
if config is None:
continue
indexer_types = getattr(config, "indexer_types", None)
if indexer_types is not None:
return indexer_types
return None
def _has_shared_indexer_layers(configs: tuple[Any, ...]) -> bool:
indexer_types = _get_indexer_types(configs)
if indexer_types is None:
return False
return any(isinstance(indexer_type, str) and indexer_type.lower() == "shared" for indexer_type in indexer_types)
def _get_config_bool(configs: tuple[Any, ...], attr: str) -> bool:
for config in configs:
if config is not None and hasattr(config, attr):
return bool(getattr(config, attr))
return False
class AscendSFABackend(AttentionBackend):
accept_output_buffer: bool = True
@staticmethod
def get_name() -> str:
# HACK(Ronald1995): vllm `initialize_kv_cache` method in model runner v2 make
# attention name assertion, we just set name to FLASH_ATTN to avoid assertion error.
# rectify this when vllm disable the assertion.
return "ASCEND_SFA" if not envs_vllm.VLLM_USE_V2_MODEL_RUNNER else "FLASH_ATTN"
@staticmethod
def get_builder_cls():
if enable_sfa_dcp_replicated_indexer():
from vllm_ascend.attention.context_parallel.sfa_cp import AscendSFADCPMetadataBuilder
return AscendSFADCPMetadataBuilder
if enable_cp():
from vllm_ascend.attention.context_parallel.sfa_cp import AscendSFACPMetadataBuilder
return AscendSFACPMetadataBuilder
return AscendSFAMetadataBuilder
@staticmethod
def get_kv_cache_shape(
num_blocks: int,
block_size: int,
num_kv_heads: int,
head_size: int,
cache_type: str = "",
) -> tuple[int, ...]:
return (num_blocks, block_size, num_kv_heads, head_size)
@staticmethod
def get_impl_cls() -> type["AscendSFAImpl"]:
if enable_sfa_dcp_replicated_indexer():
from vllm_ascend.attention.context_parallel.sfa_cp import AscendSFADCPImpl
return AscendSFADCPImpl
if enable_cp():
from vllm_ascend.attention.context_parallel.sfa_cp import AscendSFACPImpl
return AscendSFACPImpl
return AscendSFAImpl
@staticmethod
def get_supported_kernel_block_sizes() -> list[int]:
return [128]
@dataclass
class DCPContext:
slot_mapping: torch.Tensor
block_table: torch.Tensor
seq_lens: torch.Tensor
kv_gather_block_ids: torch.Tensor | None = None
kv_gather_block_table: torch.Tensor | None = None
gather_context: DCPGatherContext | None = None
@dataclass
class DSACPContext:
num_tokens: int
num_tokens_pad: int
local_start: int
local_end: int
local_end_with_pad: int
slot_mapping_cp: torch.Tensor
actual_seq_lengths_query: torch.Tensor
actual_seq_lengths_key: torch.Tensor
@dataclass
class AscendSFAMetadata:
"""Metadata for MLACommon.
NOTE: Please read the comment at the top of the file before trying to
understand this class
"""
# NOTE(sang): Definition of context_len, query_len, and seq_len.
# |---------- N-1 iteration --------|
# |---------------- N iteration ---------------------|
# |- tokenA -|......................|-- newTokens ---|
# |---------- context_len ----------|
# |-------------------- seq_len ---------------------|
# |-- query_len ---|
num_actual_tokens: int # Number of tokens excluding padding.
slot_mapping: torch.Tensor
seq_lens: torch.Tensor
seq_lens_cpu: torch.Tensor
cum_query_lens: torch.Tensor
block_table: torch.Tensor
sin: torch.Tensor
cos: torch.Tensor
# For logging.
num_input_tokens: int = 0 # Number of tokens including padding.
# The dimension of the attention heads
head_dim: int | None = None
attn_mask: torch.Tensor = None
# chunked prefill by default if no attn_states passed
attn_state: AscendAttentionState = AscendAttentionState.ChunkedPrefill
dcp_context: DCPContext | None = None
dsa_cp_context: DSACPContext | None = None
reshape_cache_event: torch.npu.Event = None
sfa_cp_metadata: AscendPCPMetadata | None = None
num_decodes: int = 0
num_decode_tokens: int = 0
num_prefills: int = 0
block_size: int = 0
group_len: torch.Tensor | None = None
group_key_idx: torch.Tensor | None = None
group_key_cache_idx: torch.Tensor | None = None
M = TypeVar("M", bound=AscendSFAMetadata)
class AscendSFAMetadataBuilder(MLACommonMetadataBuilder[AscendSFAMetadata]):
"""
NOTE: Please read the comment at the top of the file before trying to
understand this class
"""
def __init__(
self,
kv_cache_spec,
layer_names: list[str],
vllm_config: VllmConfig,
device: torch.device,
metadata_cls: type[AscendSFAMetadata] | None = None,
supports_dcp_with_varlen: bool = False,
):
super().__init__(
kv_cache_spec,
layer_names,
vllm_config,
device,
metadata_cls if metadata_cls is not None else AscendSFAMetadata,
supports_dcp_with_varlen,
)
self.block_size = vllm_config.cache_config.block_size
# Match the logical block size selected for BlockTable.
self.kernel_block_size = select_common_block_size(kv_cache_spec.block_size, [AscendSFABackend])
self.max_blocks = (vllm_config.model_config.max_model_len + self.block_size - 1) // self.block_size
self.speculative_config = vllm_config.speculative_config
self.decode_threshold = 1
max_num_reqs = vllm_config.scheduler_config.max_num_seqs
self.actual_seq_lengths_query = torch.zeros(max_num_reqs + 1, dtype=torch.int32, device=device)
self.actual_seq_lengths_key = torch.empty_like(self.actual_seq_lengths_query)
self.spec_actual_seq_lengths_query: list[torch.Tensor] | None = None
self.spec_actual_seq_lengths_key: list[torch.Tensor] | None = None
# Persistent int32 buffers for store_kv_block_metadata inputs, sized to
# max_num_batched_tokens (matches model_runner_v1._make_buffer sizing).
max_num_batched_tokens = vllm_config.scheduler_config.max_num_batched_tokens
self.group_len = torch.zeros(max_num_batched_tokens, dtype=torch.int32, device=device)
self.group_key_idx = torch.zeros(max_num_batched_tokens, dtype=torch.int32, device=device)
self.group_key_cache_idx = torch.zeros(max_num_batched_tokens, dtype=torch.int32, device=device)
self.spec_group_len: list[torch.Tensor] | None = None
self.spec_group_key_idx: list[torch.Tensor] | None = None
self.spec_group_key_cache_idx: list[torch.Tensor] | None = None
if self.speculative_config:
spec_token_num = self.speculative_config.num_speculative_tokens
self.decode_threshold += spec_token_num
assert self.decode_threshold <= 16, (
f"decode_threshold exceeded \
npu_fused_infer_attention_score TND layout's limit of 16, \
got {self.decode_threshold}"
)
self.spec_actual_seq_lengths_query = [
torch.zeros(max_num_reqs * (spec_token_num + 1) + 1, dtype=torch.int32, device=device)
for _ in range(spec_token_num)
]
self.spec_actual_seq_lengths_key = [
torch.zeros(max_num_reqs * (spec_token_num + 1) + 1, dtype=torch.int32, device=device)
for _ in range(spec_token_num)
]
self.spec_group_len = [
torch.zeros(max_num_batched_tokens, dtype=torch.int32, device=device) for _ in range(spec_token_num)
]
self.spec_group_key_idx = [
torch.zeros(max_num_batched_tokens, dtype=torch.int32, device=device) for _ in range(spec_token_num)
]
self.spec_group_key_cache_idx = [
torch.zeros(max_num_batched_tokens, dtype=torch.int32, device=device) for _ in range(spec_token_num)
]
self.reorder_batch_threshold = self.decode_threshold
self.attn_mask_builder = AttentionMaskBuilder(self.device)
self.rope_dim = self.model_config.hf_text_config.qk_rope_head_dim
self.enable_dsa_cp = enable_dsa_cp()
@staticmethod
def determine_chunked_prefill_workspace_size(vllm_config: VllmConfig) -> int:
return ascend_chunked_prefill_workspace_size(vllm_config)
@classmethod
def get_cudagraph_support(
cls: type["AscendSFAMetadataBuilder"],
vllm_config: VllmConfig,
kv_cache_spec: AttentionSpec,
) -> AttentionCGSupport:
# Explicit override in case the underlying builder specialized this getter.
# @override omitted only because of mypy limitation due to type variable.
return AttentionCGSupport.UNIFORM_BATCH
def reorder_batch(self, input_batch: "NPUInputBatch", scheduler_output: "SchedulerOutput") -> bool:
# No need to reorder for Ascend SFA
return False
def build(
self,
common_prefix_len: int,
common_attn_metadata: AscendCommonAttentionMetadata,
fast_build: bool = False,
**kwargs,
) -> AscendSFAMetadata:
# common_prefix_len / fast_build are unused; kept for API compatibility.
return self._build(common_attn_metadata, draft_index=None)
def build_for_drafting(
self,
common_attn_metadata: AscendCommonAttentionMetadata,
draft_index: int,
**kwargs,
) -> AscendSFAMetadata:
return self._build(common_attn_metadata, draft_index=draft_index)
def _build(
self,
common_attn_metadata: AscendCommonAttentionMetadata,
draft_index: int | None = None,
) -> AscendSFAMetadata:
num_reqs = common_attn_metadata.num_reqs
num_actual_tokens = common_attn_metadata.num_actual_tokens
num_input_tokens = common_attn_metadata.num_input_tokens
block_table = common_attn_metadata.block_table_tensor[:num_reqs]
slot_mapping = common_attn_metadata.slot_mapping[:num_input_tokens]
input_positions = common_attn_metadata.positions[:num_input_tokens].long()
block_size = self.kernel_block_size
cum_query_lens = common_attn_metadata.query_start_loc[1 : num_reqs + 1]
seq_lens = common_attn_metadata.seq_lens[:num_reqs]
# Prefer _seq_lens_cpu (always available, updated during draft
# iterations) over seq_lens_cpu (None in async spec decode mode).
if common_attn_metadata._seq_lens_cpu is not None:
seq_lens_cpu = common_attn_metadata._seq_lens_cpu[:num_reqs]
elif common_attn_metadata.seq_lens_cpu is not None:
seq_lens_cpu = common_attn_metadata.seq_lens_cpu[:num_reqs]
else:
seq_lens_cpu = common_attn_metadata.seq_lens[:num_reqs].to("cpu")
cos, sin = get_cos_and_sin_mla(input_positions, use_cache=(draft_index is None))
dsa_cp_context = None
if self.enable_dsa_cp:
global_tp_size = get_tp_group().world_size
num_tokens = num_input_tokens
num_tokens_pad = _round_up(num_tokens, global_tp_size)
num_tokens_per_device = num_tokens_pad // global_tp_size
local_start = get_tp_group().rank_in_group * num_tokens_per_device
local_end_with_pad = local_start + num_tokens_per_device
local_end = min(local_end_with_pad, num_actual_tokens)
pad_size = num_tokens_pad - cos.shape[0]
assert cos.shape == sin.shape, f"cos.shape must be equal to sin.shape, got {cos.shape} and {sin.shape}"
if pad_size > 0:
cos = nn.functional.pad(cos, (0, 0, 0, 0, 0, 0, 0, pad_size))
sin = nn.functional.pad(sin, (0, 0, 0, 0, 0, 0, 0, pad_size))
pad_size_slot = num_tokens_pad - slot_mapping.shape[0]
if pad_size_slot > 0:
slot_mapping = nn.functional.pad(slot_mapping, (0, pad_size_slot), value=-1)
else:
slot_mapping = slot_mapping[:num_tokens_pad]
slot_mapping_cp = slot_mapping[local_start:local_end_with_pad]
cos = cos[local_start:local_end_with_pad]
sin = sin[local_start:local_end_with_pad]
assert cos.shape[0] == num_tokens_per_device, (
f"cos.shape[0] must be equal to num_tokens_per_device, \
got {cos.shape[0]} and {num_tokens_per_device}"
)
assert slot_mapping_cp.shape[0] == num_tokens_per_device, (
f"slot_mapping_cp.shape[0] must be equal to num_tokens_per_device, \
got {slot_mapping_cp.shape[0]} and {num_tokens_per_device}"
)
assert slot_mapping.shape[0] == num_tokens_pad, (
f"slot_mapping.shape[0] must be equal to num_tokens_pad, \
got {slot_mapping.shape[0]} and {num_tokens_pad}"
)
if draft_index is not None:
assert self.spec_actual_seq_lengths_query is not None
assert self.spec_actual_seq_lengths_key is not None
# Per-draft-step buffers: independent, graph-stable storage so
# later draft steps don't clobber earlier ones' metadata.
actual_seq_lengths_query = self.spec_actual_seq_lengths_query[draft_index - 1]
actual_seq_lengths_key = self.spec_actual_seq_lengths_key[draft_index - 1]
else:
actual_seq_lengths_query = self.actual_seq_lengths_query
actual_seq_lengths_key = self.actual_seq_lengths_key
num_segs = cum_query_lens.shape[0]
# Vectorized per-request local query/key lengths for this rank's
# [local_start, local_end_with_pad) slice. Replaces a Python loop
# that did 2 .item() NPU->CPU syncs per request (2 * num_reqs
# syncs/step); now fully on-device with zero syncs.
# global_start[i] = 0 for i==0, else cum_query_lens[i-1]
global_start = common_attn_metadata.query_start_loc[:num_segs]
global_end = cum_query_lens
# Clip each request's [global_start, global_end) to the local range.
# num_local_tokens may be < 0 when the request falls entirely
# outside [local_start, local_end_with_pad); clamp before cumsum.
req_local_start = global_start.clamp(min=local_start)
req_local_end = global_end.clamp(max=local_end_with_pad)
num_local_tokens = req_local_end - req_local_start
local_query_lens = torch.cumsum(num_local_tokens.clamp(min=0), dim=0)
offset = global_end - req_local_end # request tokens on later ranks
valid_local_req = (num_local_tokens > 0) & (seq_lens > 0)
local_key_lens = torch.where(
valid_local_req,
torch.clamp_min(seq_lens - offset, 0),
torch.zeros_like(seq_lens),
)
actual_seq_lengths_query[:num_segs] = local_query_lens
actual_seq_lengths_key[:num_segs] = local_key_lens
actual_seq_lengths_query = actual_seq_lengths_query[:num_reqs]
actual_seq_lengths_key = actual_seq_lengths_key[:num_reqs]
dsa_cp_context = DSACPContext(
num_tokens=num_tokens,
num_tokens_pad=num_tokens_pad,
local_start=local_start,
local_end=local_end,
local_end_with_pad=local_end_with_pad,
slot_mapping_cp=slot_mapping_cp,
actual_seq_lengths_query=actual_seq_lengths_query,
actual_seq_lengths_key=actual_seq_lengths_key,
)
if get_ascend_config().c8_enable_reshape_optim:
if draft_index is not None:
assert self.spec_group_len is not None
assert self.spec_group_key_idx is not None
assert self.spec_group_key_cache_idx is not None
group_len = self.spec_group_len[draft_index - 1]
group_key_idx = self.spec_group_key_idx[draft_index - 1]
group_key_cache_idx = self.spec_group_key_cache_idx[draft_index - 1]
else:
group_len = self.group_len
group_key_idx = self.group_key_idx
group_key_cache_idx = self.group_key_cache_idx
actual_group_len = group_len[:num_input_tokens]
actual_group_key_idx = group_key_idx[:num_input_tokens]
actual_group_key_cache_idx = group_key_cache_idx[:num_input_tokens]
torch.ops._C_ascend.store_kv_block_metadata(
slot_mapping,
actual_group_len,
actual_group_key_idx,
actual_group_key_cache_idx,
block_size,
)
else:
actual_group_len = None
actual_group_key_idx = None
actual_group_key_cache_idx = None
return self.metadata_cls( # type: ignore
num_input_tokens=common_attn_metadata.num_input_tokens,
num_actual_tokens=num_actual_tokens,
cum_query_lens=cum_query_lens,
seq_lens=seq_lens,
seq_lens_cpu=seq_lens_cpu,
slot_mapping=slot_mapping,
head_dim=self.model_config.get_head_size(),
attn_mask=self.attn_mask_builder.get_attention_mask(common_attn_metadata.causal, self.model_config),
attn_state=common_attn_metadata.attn_state,
block_table=block_table,
sin=sin[:num_input_tokens],
cos=cos[:num_input_tokens],
dsa_cp_context=dsa_cp_context,
block_size=block_size,
group_len=actual_group_len,
group_key_idx=actual_group_key_idx,
group_key_cache_idx=actual_group_key_cache_idx,
)
def build_for_graph_capture(
self,
common_attn_metadata: AscendCommonAttentionMetadata,
attn_state: AscendAttentionState = AscendAttentionState.DecodeOnly,
):
if attn_state in {AscendAttentionState.DecodeOnly, AscendAttentionState.SpecDecoding}:
attn_metadata = self.build(
common_prefix_len=0,
common_attn_metadata=common_attn_metadata,
)
else:
raise NotImplementedError("Currently we only support building dummy metadata for DecodeOnly state")
attn_metadata.attn_state = attn_state
return attn_metadata
class AscendSFAImpl(MLAAttentionImpl):
"""
NOTE: Please read the comment at the top of the file before trying to
understand this class
"""
# Supports forward using the all-gather o_proj weight for decode requests when Sharded CP is enabled.
o_proj_full_pools: dict[tuple[str, int | None, torch.dtype, int, tuple[int, ...]], torch.Tensor] = {}
# q_hadamard and k_hadamard tensor shared when dsa c8 enabled
q_hadamard: torch.Tensor | None = None
k_hadamard: torch.Tensor | None = None
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: int,
alibi_slopes: list[float] | None,
sliding_window: int | None,
kv_cache_dtype: str,
logits_soft_cap: float | None,
attn_type: str,
kv_sharing_target_layer_name: str | None,
**kwargs,
) -> None:
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)
self.num_kv_heads = num_kv_heads
self.kv_cache_dtype = kv_cache_dtype
# MLA Args
self.q_lora_rank = kwargs["q_lora_rank"]
self.kv_lora_rank = kwargs["kv_lora_rank"]
self.qk_nope_head_dim = kwargs["qk_nope_head_dim"]
self.qk_rope_head_dim = kwargs["qk_rope_head_dim"]
self.qk_head_dim = kwargs["qk_head_dim"]
self.v_head_dim = kwargs["v_head_dim"]
self.rotary_emb = kwargs["rotary_emb"]
self.q_proj = kwargs["q_proj"] if self.q_lora_rank is None else kwargs["q_b_proj"]
self.fused_qkv_a_proj = kwargs.get("fused_qkv_a_proj")
self.kv_b_proj = kwargs["kv_b_proj"]
self.o_proj = kwargs["o_proj"]
self.indexer = kwargs["indexer"]
self.kv_a_proj_with_mqa = kwargs.get("kv_a_proj_with_mqa")
self.kv_a_layernorm = kwargs.get("kv_a_layernorm")
self.q_a_layernorm = kwargs.get("q_a_layernorm")
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
self.tp_size = get_tensor_model_parallel_world_size()
self.tp_rank = get_tp_group().rank_in_group
self.q_b_proj = kwargs["q_b_proj"]
self.skip_topk = kwargs.get("skip_topk", False)
self.topk_indices_buffer = kwargs.get("topk_indices_buffer")
ascend_config = get_ascend_config()
self.vllm_config = get_current_vllm_config()
kv_transfer_config = self.vllm_config.kv_transfer_config
self.is_kv_producer = kv_transfer_config is not None and kv_transfer_config.is_kv_producer
self.is_kv_consumer = kv_transfer_config is not None and kv_transfer_config.is_kv_consumer
self.sfa_qsfa_tile_size = SFA_QSFA_TILE_SIZE
self.sfa_qsfa_packed_kv_head_dim = 0
self.sfa_qsfa_k_nope_clip_alpha: torch.Tensor | None = None
self.sfa_qsfa_kr_cache_dummy: torch.Tensor | None = None
self.local_num_heads = self.num_heads
self.layer_name = kwargs.get("layer_name")
hf_config = self.vllm_config.model_config.hf_config
hf_text_config = getattr(self.vllm_config.model_config, "hf_text_config", None)
config_candidates = (hf_config, hf_text_config)
self.index_cache_enabled = _get_config_bool(
config_candidates,
"use_index_cache",
) or _has_shared_indexer_layers(config_candidates)
self.use_index_cache = self.skip_topk or self.index_cache_enabled
self.has_indexer = self.indexer is not None
if not self.has_indexer and not self.skip_topk:
raise ValueError(
"Indexer is required for DSA unless skip_topk is enabled. "
f"Got indexer=None, skip_topk={self.skip_topk}, "
f"layer_name={self.layer_name}."
)
if not self.has_indexer and self.topk_indices_buffer is None:
raise ValueError(
"topk_indices_buffer is required when indexer is None and "
f"skip_topk is enabled. layer_name={self.layer_name}."
)
# indexer param
if self.has_indexer:
self.n_head: int = self.indexer.n_head # 64
self.head_dim: int = self.indexer.head_dim # 128
self.wq_b = self.indexer.wq_b
self.wk_weights_proj = self.indexer.wk_weights_proj
self.k_norm = self.indexer.k_norm
else:
self.n_head = getattr(hf_config, "index_n_heads", 0)
self.head_dim = getattr(hf_config, "index_head_dim", 0)
self.wq_b = None
self.wk_weights_proj = None
self.k_norm = None
self.cp_size = 1
self.is_rope_neox_style = True
self.use_torch_npu_lightning_indexer = False
if self.vllm_config.model_config.hf_config.model_type in ["glm_moe_dsa"]:
self.is_rope_neox_style = False
self.use_torch_npu_lightning_indexer = True
# Sparse C8 has two independent meanings in SFA:
# - SFA packed KV cache for npu_kv_quant_sparse_flash_attention.
# - C8 indexer cache for lightning indexer.
# The user-facing switches control these layouts independently, and
# layers without an indexer only apply the SFA setting.
self.enable_sparse_sfa_c8 = ascend_config.enable_sparse_sfa_c8
self.enable_sparse_li_c8 = self.has_indexer and ascend_config.is_sparse_li_c8_layer(self.layer_name)
if self.enable_sparse_sfa_c8 or self.enable_sparse_li_c8:
if get_ascend_device_type() == AscendDeviceType.A5:
self.c8_k_cache_dtype = torch.float8_e4m3fn
self.c8_k_scale_cache_dtype = torch.float32
else:
self.c8_k_cache_dtype = torch.int8
self.c8_k_scale_cache_dtype = torch.float16
if self.enable_sparse_sfa_c8:
self.sfa_qsfa_packed_kv_head_dim = get_sfa_qsfa_packed_head_dim(
self.kv_lora_rank,
self.qk_rope_head_dim,
self.sfa_qsfa_tile_size,
)
# PD decode consumers with sparse C8 can use mla_prolog_v3 to write the packed KV cache.
# TODO: Re-enable after the community CANN baseline upgrades from 9.0 to 9.1.
# npu_mla_prolog_v3 depends on CANN 9.1 and is not available in the current community CANN 9.0.
cann_version = getattr(torch.version, "cann", None)
if cann_version is not None:
from packaging.version import Version
sfa_prolog_v3_supported = Version(cann_version) >= Version("9.1.0")
else:
sfa_prolog_v3_supported = False
self.enable_sfa_prolog_v3 = (
sfa_prolog_v3_supported
and self.is_kv_consumer
and self.enable_sparse_sfa_c8
and get_ascend_device_type() != AscendDeviceType.A5
)
self.enable_mlapo = ascend_config.enable_mlapo and not (
self.enable_sfa_prolog_v3 or (self.enable_sparse_sfa_c8 and get_ascend_device_type() != AscendDeviceType.A5)
)
# Effective in SFA when FlashComm is enabled.
self.enable_dsa_cp = enable_dsa_cp()
self.enable_sp = enable_sp()
# Enable layer sharding via DSA-CP on the P node in the PD-disaggregated setup.
self.enable_dsa_cp_with_layer_shard = enable_dsa_cp_with_layer_shard()
# SFA DSA-CP mixed deployments keep o_proj in the existing TP layout.
# Decode can use the TP-sharded o_proj directly after an activation
# all-to-all, while prefill/mixed batches temporarily gather the TP
# shards into a full-weight buffer because their SFA output is not
# TP-sharded. This is part of the DSA-CP mixed-mode data path rather
# than an independent user-facing feature switch.
self.enable_dsa_cp_with_o_proj_tp = enable_dsa_cp_with_o_proj_tp()
if self.enable_dsa_cp:
self.local_num_heads = self.num_heads * self.tp_size
if self.enable_dsa_cp_with_layer_shard:
self.layer_sharding_kwargs = []
for layer_name in get_ascend_config().layer_sharding or []:
if layer_name in kwargs:
self.layer_sharding_kwargs.append(kwargs[layer_name])
else:
logger.warning_once(
f"Layer '{layer_name}' not found in kwargs, skipping sharding. "
f"Check layer_sharding config and model layer names."
)
register_all_layers_to_shard_weight_series(self.layer_sharding_kwargs)
@staticmethod
def update_graph_params(
update_stream,
forward_context,
num_tokens,
vllm_config=None,
speculative_config=None,
num_dcp_pcp_tokens=None,
draft_attn_metadatas=None,
):
# sfa does not need to update graph params
pass
def process_weights_after_loading(self, act_dtype: torch.dtype):
# NOTE: We currently do not support quant kv_b_proj.
assert isinstance(self.kv_b_proj.quant_method, UnquantizedLinearMethod)
# NOTE: Weight will be reshaped next, we need to revert and transpose it.
kv_b_proj_weight = torch_npu.npu_format_cast(self.kv_b_proj.weight.data, ACL_FORMAT_FRACTAL_ND).T
assert kv_b_proj_weight.shape == (
self.kv_lora_rank,
self.local_num_heads * (self.qk_nope_head_dim + self.v_head_dim),
), (
f"{kv_b_proj_weight.shape=}, "
f"{self.kv_lora_rank=}, "
f"{self.local_num_heads=}, "
f"{self.qk_nope_head_dim=}, "
f"{self.v_head_dim=}"
)
kv_b_proj_weight = kv_b_proj_weight.view(
self.kv_lora_rank,
self.local_num_heads,
self.qk_nope_head_dim + self.v_head_dim,
)
W_UK, W_UV = kv_b_proj_weight.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
# NOTE: When we make a incontiguous weight contiguous, a new address will be allocated for the weight,
# in graph + RL scenario, we only capture the graph once, and the weight address is expected to be the same
# across iterations, so we need to copy the weight to the original address after making it contiguous.
if not hasattr(self, "W_UV"):
# Convert from (L, N, V) to (N, L, V)
self.W_UV = W_UV.transpose(0, 1).contiguous()
# Convert from (L, N, P) to (N, P, L)
self.W_UK_T = W_UK.permute(1, 2, 0).contiguous()
else:
self.W_UV.copy_(W_UV.transpose(0, 1).contiguous())
self.W_UK_T.copy_(W_UK.permute(1, 2, 0).contiguous())
# TODO(zzzzwwjj): Currently, torch.ops._C_ascend.batch_matmul_transpose cannot support weight nz
# self.W_UV = maybe_trans_nz(self.W_UV)
# Dispose kv_b_proj since it is replaced by W_UV and W_UK_T to save memory
dispose_layer(self.kv_b_proj)
if self.enable_dsa_cp:
if self.enable_dsa_cp_with_layer_shard:
for layer in self.layer_sharding_kwargs or []:
if is_hidden_layer(layer):
post_process_after_loading_for_shard_weight_series(layer)
elif self.enable_dsa_cp_with_o_proj_tp:
self._init_o_proj_tp_full_params()
if self.enable_sfa_prolog_v3:
reasons = self._get_sfa_prolog_v3_unsupported_reasons()
if reasons:
self.enable_sfa_prolog_v3 = False
self.enable_mlapo = False
for msg in reasons:
logger.warning_once(
f"{msg} Disable SFA mla_prolog_v3 for layer {self.layer_name}; "
"fallback to native preprocessing."
)
else:
self._process_weights_for_fused_prolog_v3()
if not self.enable_sfa_prolog_v3 and self.enable_mlapo:
quant_method = getattr(
getattr(self.fused_qkv_a_proj, "quant_method", None),
"quant_method",
None,
)
reasons = []
is_quantized = isinstance(quant_method, (AscendW8A8LinearMethod, AscendW8A8MXFP8DynamicLinearMethod))
if self.fused_qkv_a_proj is None:
reasons.append("fused_qkv_a_proj is None, mlapo is disabled.")
if not is_quantized and get_ascend_device_type() != AscendDeviceType.A5:
reasons.append(
"Currently mlapo only supports W8A8 quantization in SFA scenario on non-A5 devices."
"Some layers in your model are not quantized with W8A8,"
"thus mlapo is disabled for these layers."
)
if self.enable_dsa_cp:
reasons.append("Currently mlapo does not support SFA with CP,thus mlapo is disabled for these layers.")
if reasons:
self.enable_mlapo = False
for msg in reasons:
logger.warning_once(msg)
else:
self.mlapo_is_quantized = is_quantized
if get_ascend_device_type() == AscendDeviceType.A5:
if is_quantized:
self._process_weights_for_fused_mlapo_a5(act_dtype)
else:
self._process_weights_for_fused_mlapo_a5_float(act_dtype)
else:
self._process_weights_for_fused_mlapo(act_dtype)
if self.enable_sparse_li_c8 and get_ascend_device_type() == AscendDeviceType.A5:
if hasattr(self, "mlapo_is_quantized") and not self.mlapo_is_quantized:
self.c8_k_cache_dtype = act_dtype
self.c8_k_scale_cache_dtype = act_dtype
if not self.enable_mlapo and not self.enable_sfa_prolog_v3:
# if mlapo, W_UK_T can't trans nz
self.W_UK_T = maybe_trans_nz(self.W_UK_T)
if self.has_indexer and self.enable_sparse_li_c8 and AscendSFAImpl.q_hadamard is None:
AscendSFAImpl.q_hadamard = torch.tensor(scipy.linalg.hadamard(128), dtype=torch.bfloat16, device="npu") / (
128**0.5
)
if self.has_indexer and self.enable_sparse_li_c8 and AscendSFAImpl.k_hadamard is None:
AscendSFAImpl.k_hadamard = torch.tensor(scipy.linalg.hadamard(128), dtype=torch.bfloat16, device="npu") / (
128**0.5
)
@staticmethod
def _is_w8a8_dynamic_linear(layer: torch.nn.Module | None) -> bool:
quant_method = getattr(getattr(layer, "quant_method", None), "quant_method", None)
return isinstance(quant_method, AscendW8A8DynamicLinearMethod)
def _get_sfa_prolog_v3_unsupported_reasons(self) -> list[str]:
reasons = []
for name, layer in (
("fused_qkv_a_proj", self.fused_qkv_a_proj),
("q_proj", self.q_proj),
):
if not self._is_w8a8_dynamic_linear(layer):
reasons.append(f"Currently SFA mla_prolog_v3 only supports W8A8 dynamic quantization for {name}.")
if self.kv_a_layernorm is None or self.q_a_layernorm is None:
reasons.append("SFA mla_prolog_v3 requires q_a_layernorm and kv_a_layernorm.")
if getattr(self.q_proj, "_chunk_size", 0):
reasons.append("SFA mla_prolog_v3 does not support chunked q_proj weights yet.")
if self.enable_dsa_cp:
reasons.append("SFA mla_prolog_v3 does not support DSA-CP; DSA-CP takes precedence.")
if self.is_kv_producer:
reasons.append("SFA mla_prolog_v3 is disabled on KV producer workers.")
return reasons
def _process_weights_for_fused_prolog_v3(self) -> None:
assert self.fused_qkv_a_proj is not None
assert self.q_proj is not None
fused_weight = self.fused_qkv_a_proj.weight.data
weight_dq = fused_weight[..., : self.q_lora_rank].contiguous()
weight_dkv_kr = fused_weight[..., self.q_lora_rank :].contiguous()
weight_uq_qr = self.q_proj.weight.data.contiguous()
self.weight_dq = torch_npu.npu_format_cast(weight_dq, ACL_FORMAT_FRACTAL_NZ)
self.weight_dkv_kr = torch_npu.npu_format_cast(weight_dkv_kr, ACL_FORMAT_FRACTAL_NZ)
self.weight_uq_qr = torch_npu.npu_format_cast(weight_uq_qr, ACL_FORMAT_FRACTAL_NZ)
q_a_proj_deq_scl = self.fused_qkv_a_proj.weight_scale[: self.q_lora_rank].contiguous()
kv_a_proj_deq_scl = self.fused_qkv_a_proj.weight_scale[self.q_lora_rank :].contiguous()
self.dequant_scale_w_dq = q_a_proj_deq_scl.view(1, -1).to(torch.float)
self.dequant_scale_w_dkv_kr = kv_a_proj_deq_scl.view(1, -1).to(torch.float)
self.dequant_scale_w_uq_qr = self.q_proj.weight_scale.data.view(1, -1).to(torch.float)
if self.enable_sparse_sfa_c8:
self.sfa_qsfa_k_nope_clip_alpha = torch.ones(
1,
dtype=torch.float32,
device=self.weight_dq.device,
)
if self.sfa_qsfa_kr_cache_dummy is None:
# ckvkr_repo_mode=1 stores rope in the packed KV cache, but the
# operator still requires kr_cache. Keep a stable, non-aliased
# dummy so first-run tiling/graph capture cannot alias kv_cache.
self.sfa_qsfa_kr_cache_dummy = torch.empty(
0,
dtype=torch.bfloat16,
device=self.weight_dq.device,
)
if self.is_kv_consumer:
# Decode-only workers only execute Prolog. Drop the native Linear
# weights after their Prolog layouts and scales have been copied.
dispose_layer(self.fused_qkv_a_proj)
dispose_layer(self.q_proj)
torch.npu.empty_cache()
# Processing the input parameters for MLAPO by reordering and transposing
# QKV(and part of Q) weight, applying RoPE-related dimension transformations,
# and handling quantization parameters.
def _process_weights_for_fused_mlapo(self, act_dtype: torch.dtype):
assert self.kv_a_proj_with_mqa is None
assert self.fused_qkv_a_proj is not None
kv_a_proj_wt = self.fused_qkv_a_proj.weight.data[..., self.q_lora_rank :].contiguous()
q_a_proj_wt = self.fused_qkv_a_proj.weight.data[..., : self.q_lora_rank].contiguous()
kv_a_proj_wt = kv_a_proj_wt.t().contiguous()
kv_a_proj_wt = trans_rope_weight(kv_a_proj_wt, self.qk_rope_head_dim)
kv_a_proj_wt = kv_a_proj_wt.t().contiguous()
wd_qkv = torch.cat((kv_a_proj_wt, q_a_proj_wt), dim=-1)
wd_qkv = wd_qkv.t().contiguous()
wd_qkv = transdata(wd_qkv, block_size=(16, 32)).unsqueeze(0).contiguous()
self.wd_qkv = torch_npu.npu_format_cast(wd_qkv, 29)
kv_a_proj_deq_scl = self.fused_qkv_a_proj.deq_scale[self.q_lora_rank :].contiguous()
q_a_proj_deq_scl = self.fused_qkv_a_proj.deq_scale[: self.q_lora_rank].contiguous()
kv_a_proj_deq_scl = kv_a_proj_deq_scl.reshape(self.kv_lora_rank + self.qk_rope_head_dim, -1).contiguous()
kv_a_proj_deq_scl = trans_rope_weight(kv_a_proj_deq_scl, self.qk_rope_head_dim)
kv_a_proj_deq_scl = kv_a_proj_deq_scl.view(self.kv_lora_rank + self.qk_rope_head_dim).contiguous()
self.deq_scale_qkv = torch.cat((kv_a_proj_deq_scl, q_a_proj_deq_scl), dim=-1).contiguous()
kv_a_proj_qt_bias = self.fused_qkv_a_proj.quant_bias[self.q_lora_rank :].contiguous()
q_a_proj_qt_bias = self.fused_qkv_a_proj.quant_bias[: self.q_lora_rank].contiguous()
kv_a_proj_qt_bias = kv_a_proj_qt_bias.reshape(self.kv_lora_rank + self.qk_rope_head_dim, -1).contiguous()
kv_a_proj_qt_bias = trans_rope_weight(kv_a_proj_qt_bias, self.qk_rope_head_dim)
kv_a_proj_qt_bias = kv_a_proj_qt_bias.view(self.kv_lora_rank + self.qk_rope_head_dim).contiguous()
self.quant_bias_qkv = torch.cat((kv_a_proj_qt_bias, q_a_proj_qt_bias), dim=-1).contiguous()
wu_q = self.q_proj.weight.data
wu_q = wu_q.t().reshape(self.num_heads, self.qk_nope_head_dim + self.qk_rope_head_dim, -1)
wu_q = trans_rope_weight(wu_q, self.qk_rope_head_dim)
wu_q = wu_q.reshape(self.num_heads * (self.qk_nope_head_dim + self.qk_rope_head_dim), -1)
wu_q = transdata(wu_q, block_size=(16, 32)).unsqueeze(0).contiguous()
self.wu_q = torch_npu.npu_format_cast(wu_q, 29)
qb_deq_scl = self.q_proj.deq_scale.data
qb_deq_scl = qb_deq_scl.reshape(self.num_heads, self.qk_nope_head_dim + self.qk_rope_head_dim, -1)
qb_deq_scl = trans_rope_weight(qb_deq_scl, self.qk_rope_head_dim)
self.qb_deq_scl = qb_deq_scl.reshape(self.num_heads * (self.qk_nope_head_dim + self.qk_rope_head_dim))
qb_qt_bias = self.q_proj.quant_bias.data
qb_qt_bias = qb_qt_bias.reshape(self.num_heads, self.qk_nope_head_dim + self.qk_rope_head_dim, -1)
qb_qt_bias = trans_rope_weight(qb_qt_bias, self.qk_rope_head_dim)
self.qb_qt_bias = qb_qt_bias.reshape(self.num_heads * (self.qk_nope_head_dim + self.qk_rope_head_dim))
device = self.q_proj.weight.device
self.gamma1 = self.q_a_layernorm.weight.data # type: ignore[union-attr]
self.beta1 = self.q_a_layernorm.bias.data # type: ignore[union-attr]
self.gamma2 = self.kv_a_layernorm.weight.data # type: ignore[union-attr]
self.quant_scale0 = self.fused_qkv_a_proj.input_scale.data
self.quant_offset0 = self.fused_qkv_a_proj.input_offset.data
self.quant_scale1 = self.q_proj.input_scale.data
self.quant_offset1 = self.q_proj.input_offset.data
self.ctkv_scale = torch.tensor([1], dtype=act_dtype, device=device)
self.q_nope_scale = torch.tensor([1], dtype=act_dtype, device=device)
# On KV consumers (decode-only) MLAPO uses the transformed weights built above;
# the original fused_qkv_a_proj/q_proj weights and quant params are no longer
# referenced, so drop them to save memory.
if (
self.vllm_config.kv_transfer_config is not None
and self.vllm_config.kv_transfer_config.is_kv_consumer
and self.vllm_config.scheduler_config.max_num_batched_tokens <= MLAPO_MAX_SUPPORTED_TOKENS
):
self.fused_qkv_a_proj.weight = None
self.fused_qkv_a_proj.deq_scale = None
self.fused_qkv_a_proj.quant_bias = None
self.q_proj.weight = None
self.q_proj.deq_scale = None
self.q_proj.quant_bias = None
torch.npu.empty_cache()
def _process_weights_for_fused_mlapo_a5(self, act_dtype: torch.dtype):
assert self.fused_qkv_a_proj is not None
assert self.q_proj is not None
weight_dq = self.fused_qkv_a_proj.weight.data[..., : self.q_lora_rank].contiguous()
self.weight_dq = torch_npu.npu_format_cast(weight_dq, 29)
weight_uq_qr = self.q_proj.weight.data.contiguous()
self.weight_uq_qr_scale = self.q_proj.weight_scale.data.transpose(0, 1)
self.weight_uq_qr_scale = self.weight_uq_qr_scale.reshape(
-1, self.weight_uq_qr_scale.shape[1] * self.weight_uq_qr_scale.shape[2]
)
self.weight_uq_qr = torch_npu.npu_format_cast(weight_uq_qr, 29)
weight_dkv_kr = self.fused_qkv_a_proj.weight.data[..., self.q_lora_rank :].contiguous()
self.weight_dkv_kr = torch_npu.npu_format_cast(weight_dkv_kr, 29)
weight_scale = self.fused_qkv_a_proj.weight_scale
weight_scale = weight_scale.transpose(0, 1)
weight_scale = weight_scale.reshape(-1, weight_scale.shape[1] * weight_scale.shape[2])
self.weight_dq_scale = weight_scale[: self.q_lora_rank, ...]
self.weight_dkv_kr_scale = weight_scale[self.q_lora_rank :, ...]
def _process_weights_for_fused_mlapo_a5_float(self, act_dtype: torch.dtype):
assert self.fused_qkv_a_proj is not None
assert self.q_proj is not None
self.fused_qkv_a_proj.weight.data = self.fused_qkv_a_proj.weight.data.T
weight_dq = self.fused_qkv_a_proj.weight.data[..., : self.q_lora_rank].contiguous()
self.weight_dq_cpu = weight_dq.cpu()
self.weight_dq = torch_npu.npu_format_cast(weight_dq, 29)
weight_uq_qr = self.q_proj.weight.data.T
weight_uq_qr = weight_uq_qr.contiguous()
self.weight_uq_qr_cpu = weight_uq_qr.cpu()
self.weight_uq_qr = torch_npu.npu_format_cast(weight_uq_qr, 29)
weight_dkv_kr = self.fused_qkv_a_proj.weight.data[..., self.q_lora_rank :].contiguous()
self.weight_dkv_kr_cpu = weight_dkv_kr.cpu()
self.weight_dkv_kr = torch_npu.npu_format_cast(weight_dkv_kr, 29)
def forward_mha(
self,
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
kv_c_and_k_pe_cache: torch.Tensor,
attn_metadata: M,
k_scale: torch.Tensor,
output: torch.Tensor,
) -> None:
raise NotImplementedError("forward_mha is not supported for SFA attention. Use forward() instead.")
def forward_mqa(
self,
q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
kv_c_and_k_pe_cache: torch.Tensor,
attn_metadata: M,
layer,
) -> tuple[torch.Tensor, torch.Tensor | None]:
raise NotImplementedError("forward_mqa is not supported for SFA attention. Use forward() instead.")
def rope_single(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
) -> torch.Tensor:
B, N, D = x.shape
S = 1
x = x.view(B, N, S, D)
x = torch_npu.npu_interleave_rope(x, cos, sin)
return x.view(B, N, D)
def _init_o_proj_tp_full_params(self):
"""
Initialize TP-mode aliases and Full-mode buffers for DSA-CP o_proj.
In SFA DSA-CP mixed execution, the same model instance can run both
decode-only and prefill/mixed batches:
- Decode-only batches all-to-all the SFA output in the TP group, then
run the original TP-sharded o_proj.
- Prefill/mixed batches produce SFA output that is not directly
compatible with TP-sharded o_proj, so each rank all-gathers the TP
o_proj shards and input-sharded quant params before running o_proj.
The original TP parameter storage remains the persistent source of
truth. The o_proj_tp_* tensors below alias that storage, while the
o_proj_full_* tensors are temporary gather destinations reused across
forwards. They are not a second persistent copy of the TP weight.
"""
sample = self.o_proj.weight
self.o_proj_full_weight_gather_dim = 1 if self._is_o_proj_unquantized() else 0
if self.o_proj_full_weight_gather_dim == 0:
full_shape = (sample.shape[0] * self.tp_size, sample.shape[1])
gather_shape = full_shape
else:
full_shape = (sample.shape[0], sample.shape[1] * self.tp_size)
gather_shape = (sample.shape[1] * self.tp_size, sample.shape[0])
# Main and MTP layers can use different quantized o_proj weight layouts,
# so key the shared full-gather pool by gather dimension, dtype, and shape.
pool_key = (
sample.device.type,
sample.device.index,
sample.dtype,
self.o_proj_full_weight_gather_dim,
full_shape,
)
if pool_key not in AscendSFAImpl.o_proj_full_pools:
AscendSFAImpl.o_proj_full_pools[pool_key] = torch.empty(
gather_shape, dtype=sample.dtype, device=sample.device
)
self.o_proj_full_gather_pool = AscendSFAImpl.o_proj_full_pools[pool_key]
if self.o_proj_full_weight_gather_dim == 0:
self.o_proj_full_pool = self.o_proj_full_gather_pool
else:
self.o_proj_full_pool = self.o_proj_full_gather_pool.transpose(0, 1)
# TP tensors alias the original parameter storage. The TP shard remains
# the single source of truth; full-weight tensors below are temporary
# gather destinations only.
self.o_proj_tp_weight = self.o_proj.weight.detach()
if self.o_proj_full_weight_gather_dim == 0:
self.o_proj_tp_weight_gather_input = self.o_proj_tp_weight
else:
# Communication scratch only: all_gather_into_tensor concatenates on
# dim0, while unquantized row-parallel o_proj is sharded on dim1.
self.o_proj_tp_weight_gather_input = self.o_proj_tp_weight.transpose(0, 1).contiguous()
self.o_proj_tp_aclnn_input_params = {}
self.o_proj_full_aclnn_input_params = {}
for param_name in O_PROJ_ACLNN_INPUT_PARAMS:
param = getattr(self.o_proj, param_name, None)
if param is None:
continue
self.o_proj_tp_aclnn_input_params[param_name] = param.detach()
self.o_proj_full_aclnn_input_params[param_name] = param.repeat(self.tp_size)
self.o_proj_tp_input_sharded_quant_params = {}
self.o_proj_full_input_sharded_quant_params = {}
for param_name, param in self._iter_o_proj_input_sharded_quant_params():
self.o_proj_tp_input_sharded_quant_params[param_name] = param.detach()
self.o_proj_full_input_sharded_quant_params[param_name] = torch.empty(
(param.shape[0] * self.tp_size, *param.shape[1:]), dtype=param.dtype, device=param.device
)
def _iter_o_proj_input_sharded_quant_params(self):
if not isinstance(self.o_proj, nn.Module):
return
for param_name, param in self.o_proj.named_parameters(recurse=False):
if param_name == "weight" or param_name in O_PROJ_ACLNN_INPUT_PARAMS:
continue
if getattr(param, "input_dim", None) == 1:
yield param_name, param
def _switch_o_proj_params(self, params: dict[str, torch.Tensor]):
for param_name, param in params.items():
getattr(self.o_proj, param_name).set_(param)
def _get_o_proj_linear_method(self):
quant_method = self.o_proj.quant_method
return getattr(quant_method, "quant_method", quant_method)
def _is_o_proj_unquantized(self) -> bool:
return isinstance(self._get_o_proj_linear_method(), UnquantizedLinearMethod)
def _apply_o_proj_full_weight(self, attn_output: torch.Tensor) -> torch.Tensor:
return self._get_o_proj_linear_method().apply(self.o_proj, attn_output)
def _handle_o_proj_weight_switch_and_forward(
self,
attn_output: torch.Tensor,
output: torch.Tensor,
o_proj_full_handle: torch.distributed.Work | None,
o_proj_full_param_handles: list[torch.distributed.Work | None] | None,
should_shard_weight: bool,
) -> tuple[torch.Tensor, bool]:
"""
Handle o_proj weight switching between TP-mode and Full-mode, and execute forward computation.
"""
# Gather o_proj weight from all TP ranks for Full-mode computation
if should_shard_weight:
# Wait for the completion of o_proj weight all-gather operation
if o_proj_full_handle is not None:
o_proj_full_handle.wait()
for handle in o_proj_full_param_handles or []:
if handle is not None:
handle.wait()
# Temporarily switch o_proj to the gathered full-weight view for
# prefill/mixed DSA-CP, whose attention output is not TP-sharded.
self.o_proj.weight.set_(self.o_proj_full_pool)
self._switch_o_proj_params(self.o_proj_full_aclnn_input_params)
self._switch_o_proj_params(self.o_proj_full_input_sharded_quant_params)
output[...] = self._apply_o_proj_full_weight(attn_output)
# Restore TP aliases so later decode batches keep using TP storage.
self.o_proj.weight.set_(self.o_proj_tp_weight)
self._switch_o_proj_params(self.o_proj_tp_aclnn_input_params)
self._switch_o_proj_params(self.o_proj_tp_input_sharded_quant_params)
return output, False
else:
# For decode scenario: perform all-to-all communication on o_proj input activations
# Reshape for all-to-all: [batch * seq, tp_size, head_dim] -> [tp_size, batch * seq, head_dim]
send = (
attn_output.view(-1, self.tp_size, self.num_heads * self.v_head_dim)
.permute(1, 0, 2)
.reshape(-1, self.num_heads * self.v_head_dim)
)
attn_output = torch.empty_like(send)
torch.distributed.all_to_all_single(attn_output, send, group=get_tp_group().device_group)
return attn_output, True
@staticmethod
def _flatten_for_byte_gather(tensor: torch.Tensor) -> tuple[torch.Tensor, int]:
if tensor.dim() == 0:
raise RuntimeError("Byte-packed all-gather requires tensors with a token dimension.")
tensor = tensor.contiguous()
num_rows = tensor.shape[0]
num_bytes_per_row = tensor.element_size()
for dim in tensor.shape[1:]:
num_bytes_per_row *= dim
return tensor.view(torch.int8).view(num_rows, num_bytes_per_row), num_bytes_per_row
@classmethod
def _all_gather_byte_packed_async(
cls,
parts: list[tuple[str, torch.Tensor]],
async_op: bool,
) -> tuple[torch.Tensor, torch.distributed.Work | None, tuple[_ByteGatherPart, ...]]:
if not parts:
raise RuntimeError("Byte-packed all-gather requires at least one tensor.")
packed_parts = []
metadata = []
expected_num_rows: int | None = None
for name, tensor in parts:
num_rows = tensor.shape[0]
if expected_num_rows is None:
expected_num_rows = num_rows
elif num_rows != expected_num_rows:
raise RuntimeError(
"Cannot byte-pack KV tensors with different token counts: "
f"expected {expected_num_rows}, got {num_rows} for {name}."
)
packed_tensor, num_bytes_per_row = cls._flatten_for_byte_gather(tensor)
packed_parts.append(packed_tensor)
metadata.append(
_ByteGatherPart(
name=name,
shape=tuple(tensor.shape),
dtype=tensor.dtype,
num_bytes_per_row=num_bytes_per_row,
)
)
packed_input = torch.cat(packed_parts, dim=1) if len(packed_parts) > 1 else packed_parts[0]
gathered, handle = all_gather_async(
packed_input,
get_tp_group(),
async_op=async_op,
)
return gathered, handle, tuple(metadata)
@staticmethod
def _restore_byte_gathered_tensors(
gathered: torch.Tensor,
metadata: tuple[_ByteGatherPart, ...],
) -> dict[str, torch.Tensor]:
chunks = torch.split(gathered, [part.num_bytes_per_row for part in metadata], dim=1)
num_rows = gathered.shape[0]
restored = {}
for part, chunk in zip(metadata, chunks):
restored[part.name] = chunk.contiguous().view(part.dtype).view(num_rows, *part.shape[1:])
return restored
def _get_full_kv(self, k, attn_metadata):
return k
def exec_kv(
self,
kv_no_split: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
kv_cache: tuple,
slots: torch.Tensor,
attn_metadata: M,
):
B = kv_no_split.shape[0]
N = self.num_kv_heads
S = 1
# npu_kv_rmsnorm_rope_cache needs [B, N, S, D]
kv_no_split = kv_no_split.view(B, N, S, self.kv_lora_rank + self.qk_rope_head_dim)
cache_mode = "PA"
use_custom_kv = self.enable_sparse_sfa_c8 and (
get_ascend_device_type() != AscendDeviceType.A5 or self.enable_dsa_cp or not self.has_indexer
)
if use_custom_kv:
assert self.kv_a_layernorm is not None
return custom_kv_rmsnorm_rope(
kv_no_split,
self.kv_a_layernorm.weight,
cos,
sin,
self.kv_lora_rank,
self.qk_rope_head_dim,
epsilon=self.kv_a_layernorm.variance_epsilon,
dst_type=(torch.float8_e4m3fn if get_ascend_device_type() == AscendDeviceType.A5 else 1),
tile_size=self.sfa_qsfa_tile_size,
)
if self.enable_dsa_cp:
_, _, k_pe, k_nope = torch_npu.npu_kv_rmsnorm_rope_cache(
kv_no_split,
self.kv_a_layernorm.weight, # type: ignore[union-attr]
cos,
sin,
slots.to(torch.int64),
kv_cache[1],
kv_cache[0],
epsilon=self.kv_a_layernorm.variance_epsilon, # type: ignore[union-attr]
cache_mode=cache_mode,
is_output_kv=True,
)
return k_pe, k_nope, None
else:
torch_npu.npu_kv_rmsnorm_rope_cache(
kv_no_split,
self.kv_a_layernorm.weight, # type: ignore[union-attr]
cos,
sin,
slots.to(torch.int64),
kv_cache[1],
kv_cache[0],
epsilon=self.kv_a_layernorm.variance_epsilon, # type: ignore[union-attr]
cache_mode=cache_mode,
)
return None, None
# Return `ql_nope`, `q_pe`
def _q_proj_and_k_up_proj(self, x):
q_nope, q_pe = (
self.q_proj(x)[0]
.view(-1, self.local_num_heads, self.qk_head_dim)
.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
)
# Convert from (B, N, P) to (N, B, P)
q_nope = q_nope.transpose(0, 1)
# Multiply (N, B, P) x (N, P, L) -> (N, B, L)
ql_nope = torch.bmm(q_nope, self.W_UK_T)
# Convert from (N, B, L) to (B, N, L)
return ql_nope.transpose(0, 1), q_pe
def _v_up_proj(self, x):
num_input_tokens, _, _ = x.shape
if (
x.dtype in [torch.float16, torch.bfloat16]
and hasattr(torch.ops._C_ascend, "batch_matmul_transpose")
and num_input_tokens <= BMM_TRANS_MAX_SUPPORTED_TOKENS
):
x = x.view(-1, self.local_num_heads, self.kv_lora_rank)
res = torch.empty((num_input_tokens, self.local_num_heads, self.v_head_dim), dtype=x.dtype, device=x.device)
torch.ops._C_ascend.batch_matmul_transpose(x, self.W_UV, res)
x = res.reshape(-1, self.local_num_heads * self.v_head_dim)
elif hasattr(torch_npu, "npu_transpose_batchmatmul"):
# Convert from (N, B, L)/(N, B, 1, L) to (N, B, L)
x = x.view(-1, self.local_num_heads, self.kv_lora_rank)
# Multiply (N, B, L) x (N, L, V) -> (B, N, V)
x = torch_npu.npu_transpose_batchmatmul(x, self.W_UV, perm_x1=(1, 0, 2), perm_y=(1, 0, 2))
# Convert from (N, B, V) to (B, N * V)
x = x.reshape(-1, self.local_num_heads * self.v_head_dim)
else:
# Convert from (B, N, L) to (N, B, L)
x = x.view(-1, self.local_num_heads, self.kv_lora_rank).transpose(0, 1)
# # Multiply (N, B, L) x (N, L, V) -> (N, B, V)
x = torch.bmm(x, self.W_UV)
# # Convert from (N, B, V) to (B, N * V)
x = x.transpose(0, 1).reshape(-1, self.local_num_heads * self.v_head_dim)
return x
def _sfa_preprocess_with_mlapo(
self,
hidden_states: torch.Tensor,
kv_cache: tuple[torch.Tensor, ...],
cos: torch.Tensor,
sin: torch.Tensor,
slot_mapping: torch.Tensor,
num_input_tokens: int,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
return DeviceOperator.sfa_preprocess_with_mlapo(
self,
hidden_states,
kv_cache,
cos,
sin,
slot_mapping,
num_input_tokens,
)
def _sfa_preprocess_with_prolog_v3(
self,
hidden_states: torch.Tensor,
kv_cache: tuple[torch.Tensor, ...],
cos: torch.Tensor,
sin: torch.Tensor,
slot_mapping: torch.Tensor,
cache_mode: str,
) -> tuple[
torch.Tensor,
torch.Tensor,
torch.Tensor,
torch.Tensor | tuple[torch.Tensor, torch.Tensor] | None,
torch.Tensor | None,
torch.Tensor | None,
]:
ql_nope, q_pe, _, q_c, q_c_scale = DeviceOperator.execute_sfa_mla_prolog_v3(
self,
hidden_states=hidden_states,
rope_sin=sin,
rope_cos=cos,
kv_cache=kv_cache,
slot_mapping=slot_mapping,
cache_mode=cache_mode,
)
ql_nope = ql_nope.view(-1, self.local_num_heads, self.kv_lora_rank)
q_pe = q_pe.view(-1, self.local_num_heads, self.qk_rope_head_dim)
if self.has_indexer:
if q_c is None:
raise RuntimeError("npu_mla_prolog_v3 did not return query_norm for SFA indexer.")
q_c = q_c.view(-1, self.q_lora_rank)
if q_c_scale is not None and self.wq_b is not None and self._is_w8a8_dynamic_linear(self.wq_b):
q_c = (q_c, q_c_scale.view(-1))
else:
q_c = None
k_nope = kv_cache[0] if cache_mode == "TND" else None
k_pe = kv_cache[1] if cache_mode == "TND" and not self.enable_sparse_sfa_c8 else None
return hidden_states, ql_nope, q_pe, q_c, k_nope, k_pe
def indexer_select_pre_process(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
):
if not self.has_indexer:
raise RuntimeError(
f"indexer_select_pre_process should not be called when indexer is None. layer_name={self.layer_name}."
)
assert self.wk_weights_proj is not None
assert self.k_norm is not None
kw, _ = self.wk_weights_proj(x)
k_li = kw[:, : self.head_dim]
k_li = self.k_norm(k_li).unsqueeze(1)
k_li = k_li.view(-1, 1, self.head_dim)
if HAS_TRITON:
cos = cos.view(-1, self.qk_rope_head_dim)
sin = sin.view(-1, self.qk_rope_head_dim)
k_li = rope_forward_triton_siso(
k_li, cos, sin, rope_dim=self.qk_rope_head_dim, is_neox_style=self.is_rope_neox_style
)
else:
k_li_pe, k_li_nope = torch.split(
k_li, [self.qk_rope_head_dim, self.head_dim - self.qk_rope_head_dim], dim=-1
)
cos = cos.view(-1, 1, 1, self.qk_rope_head_dim)
sin = sin.view(-1, 1, 1, self.qk_rope_head_dim)
k_li_pe = k_li_pe.unsqueeze(2)
k_li_pe = torch_npu.npu_rotary_mul(k_li_pe, cos, sin)
k_li_pe = k_li_pe.squeeze(2)
k_li = torch.cat([k_li_pe, k_li_nope], dim=-1) # [b*s,128]
if self.enable_sparse_li_c8:
k_li = k_li @ AscendSFAImpl.k_hadamard
k_li, k_li_scale = torch_npu.npu_dynamic_quant(k_li.view(-1, self.head_dim), dst_type=self.c8_k_cache_dtype)
k_li_scale = k_li_scale.to(self.c8_k_scale_cache_dtype) # [b*s,]
k_li_scale = k_li_scale.unsqueeze(-1) # [b*s,1]
else:
k_li_scale = None
return k_li, k_li_scale
def indexer_select_post_process(
self,
x: torch.Tensor,
q_c: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
kv_cache: tuple[torch.Tensor, ...],
attn_metadata: M,
cos: torch.Tensor,
sin: torch.Tensor,
actual_seq_lengths_query: torch.Tensor,
actual_seq_lengths_key: torch.Tensor,
):
if not self.has_indexer:
raise RuntimeError(
f"indexer_select_post_process should not be called when indexer is None. layer_name={self.layer_name}."
)
assert self.wk_weights_proj is not None
assert self.wq_b is not None
kw, _ = self.wk_weights_proj(x)
weights = kw[:, self.head_dim :]
if isinstance(q_c, tuple):
q_c_tensor, q_c_scale = q_c
q_c_tensor = q_c_tensor.view(-1, q_c_tensor.shape[-1])
quant_matmul_kwargs = dict(
bias=None,
output_dtype=x.dtype,
)
if q_c_tensor.dtype == torch.float8_e4m3fn:
if q_c_scale.dim() == 2:
q_c_scale = q_c_scale.view(q_c_scale.shape[0], -1, 2)
quant_matmul_kwargs.update(
scale_dtype=FLOAT8_E8M0FNU_DTYPE,
pertoken_scale_dtype=FLOAT8_E8M0FNU_DTYPE,
group_sizes=[1, 1, getattr(self.wq_b.quant_method.quant_method, "group_size", 32)],
)
elif q_c_scale.dim() > 1 and q_c_scale.shape[-1] == 1:
q_c_scale = q_c_scale.squeeze(dim=-1)
q_li = torch_npu.npu_quant_matmul(
q_c_tensor,
self.wq_b.weight,
self.wq_b.weight_scale,
pertoken_scale=q_c_scale,
**quant_matmul_kwargs,
)
else:
q_li, _ = self.wq_b(q_c)
q_li = q_li.view(-1, self.n_head, self.head_dim)
if HAS_TRITON:
q_li = rope_forward_triton_siso(
q_li, cos, sin, rope_dim=self.qk_rope_head_dim, is_neox_style=self.is_rope_neox_style
)
else:
q_li_pe, q_li_nope = torch.split(
q_li, [self.qk_rope_head_dim, self.head_dim - self.qk_rope_head_dim], dim=-1
)
q_li_pe = q_li_pe.unsqueeze(2)
q_li_pe = torch_npu.npu_rotary_mul(q_li_pe, cos, sin)
q_li_pe = q_li_pe.squeeze(2)
q_li = torch.cat([q_li_pe, q_li_nope], dim=-1)
q_li_scale = None
q_li_shape_ori = None
if self.enable_sparse_li_c8:
q_li_shape_ori = q_li.shape
q_li = q_li @ AscendSFAImpl.q_hadamard
q_li, q_li_scale = torch_npu.npu_dynamic_quant(q_li.view(-1, self.head_dim), dst_type=self.c8_k_cache_dtype)
q_li_scale = q_li_scale.to(self.c8_k_scale_cache_dtype) # [b*s,]
record_attention_compute_start()
return DeviceOperator.indexer_select_post_process(
self,
q_li,
q_li_scale,
q_li_shape_ori,
weights,
kv_cache,
attn_metadata,
actual_seq_lengths_query,
actual_seq_lengths_key,
self.enable_sparse_li_c8,
self.use_torch_npu_lightning_indexer,
)
def _get_indexcache_topk_indices(self, num_tokens: int) -> torch.Tensor:
if self.topk_indices_buffer is None:
raise RuntimeError("IndexCache requires topk_indices_buffer when skip_topk is enabled.")
topk_indices = self.topk_indices_buffer[:num_tokens]
if topk_indices.dim() == 2:
topk_indices = topk_indices.unsqueeze(1)
return topk_indices
def _update_indexcache_topk_indices(self, topk_indices: torch.Tensor) -> None:
if self.topk_indices_buffer is None:
return
num_tokens = topk_indices.shape[0]
topk_tokens = topk_indices.shape[-1]
topk_indices_to_cache = topk_indices
topk_indices_buffer = self.topk_indices_buffer[:num_tokens, :topk_tokens]
if topk_indices_to_cache.dim() == 3 and topk_indices_buffer.dim() == 2:
assert topk_indices_to_cache.shape[1] == 1
topk_indices_to_cache = topk_indices_to_cache.squeeze(1)
topk_indices_buffer.copy_(topk_indices_to_cache)
def _execute_sparse_flash_attention_process(
self, ql_nope, q_pe, kv_cache, topk_indices, attn_metadata, actual_seq_lengths_query, actual_seq_lengths_key
):
return DeviceOperator.execute_sparse_flash_attention_process(
self,
ql_nope,
q_pe,
kv_cache,
topk_indices,
attn_metadata,
actual_seq_lengths_query,
actual_seq_lengths_key,
)
def _record_dcp_query_gather_context(
self,
ql_nope: torch.Tensor,
q_pe: torch.Tensor,
attn_metadata: M,
) -> None:
return
def _record_dcp_kv_gather_context(
self,
kv_cache: tuple[torch.Tensor, ...],
attn_metadata: M,
) -> None:
"""Start a DCP KV gather after this layer has populated its cache.
The base implementation deliberately does nothing. The replicated-indexer
DCP implementation overrides it for batches containing prefill requests.
"""
return
def forward(
self,
layer_name,
hidden_states: torch.Tensor, # query in unified attn
kv_cache: tuple[torch.Tensor, ...],
attn_metadata: M,
need_gather_q_kv: bool = False,
output: torch.Tensor | None = None,
) -> torch.Tensor:
assert output is not None, "Output tensor must be provided."
if attn_metadata is None:
# Profiling run.
if self.enable_dsa_cp_with_layer_shard and not _EXTRA_CTX.in_profile_run:
for layer in self.layer_sharding_kwargs or []:
if is_hidden_layer(layer):
reach_layer_for_shard_weight_series(layer)
return output.fill_(0)
cos = attn_metadata.cos
sin = attn_metadata.sin
slot_mapping = attn_metadata.slot_mapping
slot_mapping_cp = None
if self.enable_dsa_cp:
assert attn_metadata.dsa_cp_context is not None
slot_mapping_cp = attn_metadata.dsa_cp_context.slot_mapping_cp
actual_seq_lengths_query = attn_metadata.dsa_cp_context.actual_seq_lengths_query
actual_seq_lengths_key = attn_metadata.dsa_cp_context.actual_seq_lengths_key
else:
actual_seq_lengths_query = attn_metadata.cum_query_lens
actual_seq_lengths_key = attn_metadata.seq_lens
# DCP replicated indexer stores LI cache with the full/no-CP metadata, while
# SFA KV remains stored with the DCP-sharded slot mapping.
slot_mapping_sfa = (
attn_metadata.dcp_context.slot_mapping
if attn_metadata.dcp_context is not None
else attn_metadata.slot_mapping
)
# Inputs and outputs may be padded for CUDA graphs
num_input_tokens = attn_metadata.num_input_tokens
output_padded = output
# all-gather o_proj weight for prefill stage of PD mix node
o_proj_full_handle = None
o_proj_full_param_handles = None
# Prefill/mixed DSA-CP computes o_proj with a temporary full weight.
# Decode keeps the original TP path and only exchanges activations.
full_gather_o_proj_enabled = self.enable_dsa_cp_with_o_proj_tp and attn_metadata.attn_state not in {
AscendAttentionState.DecodeOnly,
AscendAttentionState.SpecDecoding,
}
if self.enable_sfa_prolog_v3 and attn_metadata.attn_state in (
AscendAttentionState.DecodeOnly,
AscendAttentionState.SpecDecoding,
):
if self.enable_sp:
hidden_states = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
hidden_states.contiguous(), need_gather_q_kv
)
assert slot_mapping.numel() == hidden_states.shape[0], (
"SFA Prolog V3 requires one cache index per input token, "
f"got token_x={hidden_states.shape[0]} and cache_index={slot_mapping.numel()}."
)
if self.has_indexer:
k_li, k_li_scale = self.indexer_select_pre_process(x=hidden_states, cos=cos, sin=sin)
else:
k_li, k_li_scale = None, None
# Prolog updates the paged KV cache in place. Wait for the prompt
# blocks before writing the first Decode token into their tail block.
wait_for_kv_layer_from_connector(layer_name)
hidden_states, ql_nope, q_pe, q_c, _, _ = self._sfa_preprocess_with_prolog_v3(
hidden_states=hidden_states,
kv_cache=kv_cache,
cos=cos,
sin=sin,
slot_mapping=slot_mapping,
cache_mode="PA_BSND",
)
# run mlapo ops when dsa-cp is disabled, and ensure that num_tokens satisfies the count limitation
elif self.enable_mlapo and (
get_ascend_device_type() == AscendDeviceType.A5 or num_input_tokens <= MLAPO_MAX_SUPPORTED_TOKENS
):
hidden_states = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
hidden_states.contiguous(), need_gather_q_kv
)
hidden_states, ql_nope, q_pe, q_c = self._sfa_preprocess_with_mlapo(
hidden_states=hidden_states,
kv_cache=kv_cache,
cos=cos,
sin=sin,
slot_mapping=slot_mapping,
num_input_tokens=num_input_tokens,
)
if self.has_indexer:
k_li, k_li_scale = self.indexer_select_pre_process(
x=hidden_states,
cos=cos,
sin=sin,
)
else:
k_li, k_li_scale = None, None
wait_for_kv_layer_from_connector(layer_name)
# native
else:
assert self.fused_qkv_a_proj is not None, "q lora is required for DSA."
weight_prefetch_method = get_weight_prefetch_method()
weight_prefetch_method.maybe_prefetch_mla_or_sla_weight_in_current_stream(
inputs=self.fused_qkv_a_proj.weight, dependency=hidden_states
)
if self.enable_sp and not self.enable_dsa_cp:
hidden_states = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
hidden_states.contiguous(), need_gather_q_kv
)
qkv_lora = self.fused_qkv_a_proj(hidden_states)[0]
q_c, kv_no_split = qkv_lora.split(
[self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
dim=-1,
)
assert self.q_a_layernorm is not None, "q_a_layernorm must be initialized"
q_c = self.q_a_layernorm(q_c)
if self.has_indexer:
k_li, k_li_scale = self.indexer_select_pre_process(
x=hidden_states,
cos=cos,
sin=sin,
)
else:
k_li, k_li_scale = None, None
wait_for_kv_layer_from_connector(layer_name)
if self.enable_dsa_cp:
assert slot_mapping_cp is not None
kv_slots = slot_mapping_cp
else:
kv_slots = slot_mapping_sfa
kv_outputs = self.exec_kv(kv_no_split, cos, sin, kv_cache, kv_slots, attn_metadata)
k_pe, k_nope = kv_outputs[:2]
knope_scale = kv_outputs[2] if len(kv_outputs) == 3 else None
if (
self.enable_sparse_sfa_c8
and not self.enable_dsa_cp
and (get_ascend_device_type() != AscendDeviceType.A5 or not self.has_indexer)
):
assert k_pe is not None
assert k_nope is not None
assert knope_scale is not None
packed_kv = torch.cat([k_nope, k_pe, knope_scale], dim=-1)
packed_head_dim = self.sfa_qsfa_packed_kv_head_dim
assert packed_kv.shape[-1] == packed_head_dim
torch_npu.npu_scatter_nd_update_(
kv_cache[0].view(-1, packed_head_dim),
slot_mapping_sfa.view(-1, 1),
packed_kv.view(-1, packed_head_dim),
)
if self.enable_dsa_cp:
assert k_pe is not None
assert k_nope is not None
async_op = self.enable_dsa_cp_with_layer_shard or full_gather_o_proj_enabled
kv_ag_handles = []
# Pack all KV-related tensors into one byte stream so DSA-CP only
# submits one KV all-gather while still preserving original dtypes.
if self.enable_sparse_sfa_c8:
assert knope_scale is not None
fused_kv_parts = [
k_nope.view(-1, k_nope.shape[-1]),
k_pe.view(-1, k_pe.shape[-1]),
knope_scale.view(-1, knope_scale.shape[-1]),
]
else:
fused_kv_parts = [
k_pe.view(-1, k_pe.shape[-1]),
k_nope.view(-1, k_nope.shape[-1]),
]
fused_kv_input = torch.cat(fused_kv_parts, dim=1)
kv_gather_parts = [("sfa_kv", fused_kv_input)]
if self.has_indexer:
assert k_li is not None
k_li_gather_input = k_li
if not self.enable_sparse_sfa_c8 and not self.enable_sparse_li_c8:
k_li_gather_input = k_li.view(-1, k_li.shape[-1])
kv_gather_parts.append(("k_li", k_li_gather_input))
if self.has_indexer and self.enable_sparse_li_c8:
assert k_li_scale is not None
kv_gather_parts.append(("k_li_scale", k_li_scale))
kv_gathered_bytes, kv_ag_handle, kv_gather_metadata = self._all_gather_byte_packed_async(
kv_gather_parts,
async_op=async_op,
)
if kv_ag_handle is not None:
kv_ag_handles.append(kv_ag_handle)
ql_nope, q_pe = self._q_proj_and_k_up_proj(q_c)
q_pe = self.rope_single(q_pe, cos, sin)
self._record_dcp_query_gather_context(ql_nope, q_pe, attn_metadata)
if self.enable_dsa_cp:
for kv_ag_handle in kv_ag_handles:
kv_ag_handle.wait()
kv_gather_outputs = self._restore_byte_gathered_tensors(kv_gathered_bytes, kv_gather_metadata)
fused_kv_no_split = kv_gather_outputs["sfa_kv"]
if self.has_indexer:
k_li = kv_gather_outputs["k_li"]
if self.enable_sparse_li_c8:
k_li_scale = kv_gather_outputs["k_li_scale"]
if self.enable_dsa_cp_with_layer_shard:
for layer in self.layer_sharding_kwargs or []:
if is_hidden_layer(layer):
reach_layer_for_shard_weight_series(layer)
elif full_gather_o_proj_enabled:
_, o_proj_full_handle = all_gather_async(
self.o_proj_tp_weight_gather_input,
get_tp_group(),
output=self.o_proj_full_gather_pool,
)
o_proj_full_param_handles = []
for param_name, param in self.o_proj_tp_input_sharded_quant_params.items():
_, param_handle = all_gather_async(
param,
get_tp_group(),
output=self.o_proj_full_input_sharded_quant_params[param_name],
)
o_proj_full_param_handles.append(param_handle)
if kv_cache is not None:
assert fused_kv_no_split is not None
if self.enable_sparse_sfa_c8:
torch_npu.npu_scatter_nd_update_(
kv_cache[0].view(-1, fused_kv_no_split.shape[-1]),
slot_mapping_sfa[: attn_metadata.num_actual_tokens].view(-1, 1),
fused_kv_no_split[: attn_metadata.num_actual_tokens],
)
k_pe = None
k_nope = None
else:
k_pe, k_nope = fused_kv_no_split.split(
[self.qk_rope_head_dim, self.kv_lora_rank],
dim=-1,
)
if not self.enable_sparse_sfa_c8:
assert k_pe is not None
assert k_nope is not None
k_nope = k_nope.view(k_nope.shape[0], 1, -1)
k_pe = k_pe.view(k_pe.shape[0], 1, -1)
DeviceOperator.reshape_and_cache(
key=k_nope[: attn_metadata.num_actual_tokens],
value=k_pe[: attn_metadata.num_actual_tokens],
key_cache=kv_cache[0],
value_cache=kv_cache[1],
slot_mapping=slot_mapping_sfa[: attn_metadata.num_actual_tokens],
)
# DCP's prefill path may all-gather only the blocks referenced by
# this batch. It must start after the current layer's SFA KV write,
# but before the indexer/top-k work so communication can overlap it.
if kv_cache is not None:
self._record_dcp_kv_gather_context(kv_cache, attn_metadata)
if self.has_indexer:
assert k_li is not None
k_li = self._get_full_kv(k_li, attn_metadata)
if kv_cache is not None and self.is_kv_producer:
attn_metadata.reshape_cache_event = torch.npu.Event()
if kv_cache is not None and self.has_indexer:
assert k_li is not None
if self.enable_sparse_sfa_c8:
dsa_k_cache_idx = 1
dsa_k_scale_cache_idx = 2
else:
dsa_k_cache_idx = 2
dsa_k_scale_cache_idx = 3
if get_ascend_config().c8_enable_reshape_optim:
torch.ops._C_ascend.store_kv_block(
k_li,
kv_cache[dsa_k_cache_idx],
attn_metadata.group_len,
attn_metadata.group_key_idx,
attn_metadata.group_key_cache_idx,
attn_metadata.block_size,
)
else:
torch_npu.npu_scatter_nd_update_(
kv_cache[dsa_k_cache_idx].view(-1, k_li.shape[-1]),
slot_mapping.view(-1, 1),
k_li.view(-1, k_li.shape[-1]),
) # b, s, n, d
if self.enable_sparse_li_c8:
assert len(kv_cache) == (3 if self.enable_sparse_sfa_c8 else 4)
if k_li_scale is not None:
if get_ascend_config().c8_enable_reshape_optim:
torch.ops._C_ascend.store_kv_block(
k_li_scale,
kv_cache[dsa_k_scale_cache_idx],
attn_metadata.group_len,
attn_metadata.group_key_idx,
attn_metadata.group_key_cache_idx,
attn_metadata.block_size,
)
else:
torch_npu.npu_scatter_nd_update_(
kv_cache[dsa_k_scale_cache_idx].view(-1, k_li_scale.shape[-1]),
slot_mapping.view(-1, 1),
k_li_scale.view(-1, k_li_scale.shape[-1]),
)
if kv_cache is not None and self.is_kv_producer:
attn_metadata.reshape_cache_event.record()
notify_kv_cache_written(self.layer_name or "")
if self.enable_dsa_cp and attn_metadata.dsa_cp_context is not None:
topk_num_tokens = attn_metadata.dsa_cp_context.local_end_with_pad - attn_metadata.dsa_cp_context.local_start
else:
topk_num_tokens = num_input_tokens or hidden_states.shape[0]
if self.skip_topk:
topk_indices = self._get_indexcache_topk_indices(topk_num_tokens)
else:
if not self.has_indexer:
raise RuntimeError(f"skip_topk is False but indexer is None. layer_name={self.layer_name}.")
assert q_c is not None
topk_indices = self.indexer_select_post_process(
x=hidden_states,
q_c=q_c,
kv_cache=kv_cache,
attn_metadata=attn_metadata,
cos=cos,
sin=sin,
actual_seq_lengths_query=actual_seq_lengths_query,
actual_seq_lengths_key=actual_seq_lengths_key,
)
if self.use_index_cache:
self._update_indexcache_topk_indices(topk_indices)
attn_output = self._execute_sparse_flash_attention_process(
ql_nope,
q_pe,
kv_cache,
topk_indices,
attn_metadata,
actual_seq_lengths_query,
actual_seq_lengths_key,
)
attn_output = self._v_up_proj(attn_output)
weight_prefetch_method = get_weight_prefetch_method()
weight_prefetch_method.maybe_prefetch_mla_or_sla_weight_in_current_stream(
inputs=self.o_proj.weight,
dependency=attn_output,
max_size=MAX_O_PROJ_PREFETCH_SIZE,
linear_layer=self.o_proj,
)
if self.enable_dsa_cp_with_o_proj_tp:
# SFA DSA-CP mixed mode keeps o_proj weight sharded in the TP domain:
# 1. prefill/mixed: gather TP shards into a temporary full weight.
# 2. decode-only: all-to-all hidden states, then run TP o_proj.
result, require_o_proj_forward = self._handle_o_proj_weight_switch_and_forward(
attn_output=attn_output,
output=output,
o_proj_full_handle=o_proj_full_handle,
o_proj_full_param_handles=o_proj_full_param_handles,
should_shard_weight=full_gather_o_proj_enabled,
)
if not require_o_proj_forward:
return result
attn_output = result
output[...] = self.o_proj(attn_output)[0]
maybe_save_kv_layer_to_connector(layer_name, list(kv_cache))
return output_padded
def custom_kv_rmsnorm_rope(
kv: torch.Tensor,
gamma: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
kv_lora_rank: int,
qk_rope_head_dim: int,
*,
epsilon: float = 1e-5,
dst_type: torch.dtype | int = torch.float8_e4m3fn,
tile_size: int = SFA_QSFA_TILE_SIZE,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
rms_in, rope_in = kv.split([kv_lora_rank, qk_rope_head_dim], dim=-1)
k_nope, _ = torch_npu.npu_rms_norm(rms_in, gamma, epsilon=epsilon)
k_rope = torch_npu.npu_interleave_rope(rope_in, cos, sin)
prefix_shape = k_nope.shape[:-1]
k_nope, knope_scale = torch_npu.npu_dynamic_block_quant(
k_nope.contiguous().view(-1, 1, kv_lora_rank),
dst_type=dst_type,
row_block_size=1,
col_block_size=tile_size,
)
if dst_type == 1 or dst_type == torch.int8:
# Return byte views so the caller can concatenate all three components.
return (
k_rope.contiguous().view(torch.int8),
k_nope.view(*prefix_shape, kv_lora_rank),
knope_scale.to(torch.float32).view(*prefix_shape, -1).contiguous().view(torch.int8),
)
# A5 transports the BF16 rope and scale bytes through FP8-typed tensors.
return (
k_rope.view(torch.float8_e4m3fn),
k_nope,
knope_scale.view(knope_scale.shape[0], -1).view(torch.float8_e4m3fn),
)