EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Lint]Style: Convert vllm-ascend/ to ruff format(Batch #2) (#5977)

Browse Source 
### What this PR does / why we need it?
**Scope of Changes**:
| File Path |
| :--- |
| `vllm_ascend/attention/attention_mask.py` |
| `vllm_ascend/attention/attention_v1.py` |
| `vllm_ascend/attention/context_parallel/attention_cp.py` |
| `vllm_ascend/attention/context_parallel/common_cp.py` |
| `vllm_ascend/attention/context_parallel/mla_cp.py` |
| `vllm_ascend/attention/utils.py` |
| `vllm_ascend/batch_invariant.py` |
| `vllm_ascend/device/device_op.py` |
| `vllm_ascend/device_allocator/camem.py` |
| `vllm_ascend/envs.py` |


- vLLM version: v0.13.0
- vLLM main:
2c24bc6996

---------

Signed-off-by: MrZ20 <2609716663@qq.com>
This commit is contained in:
SILONG ZENG
2026-01-19 08:59:46 +08:00
committed by GitHub
parent 2b6dc100b5
commit 329961b375
11 changed files with 920 additions and 1045 deletions
Download Patch File Download Diff File Expand all files Collapse all files

708
vllm_ascend/attention/context_parallel/attention_cp.py
View File

File diff suppressed because it is too large Load Diff

59
vllm_ascend/attention/context_parallel/common_cp.py
View File

@@ -1,12 +1,9 @@
from dataclasses import dataclass
from typing import Optional
import torch
import torch.distributed as dist
import torch_npu
from vllm.distributed import (get_dcp_group,
get_decode_context_model_parallel_world_size,
get_pcp_group)
from vllm.distributed import get_dcp_group, get_decode_context_model_parallel_world_size, get_pcp_group
@dataclass
@@ -17,6 +14,7 @@ class AscendPCPMetadata:
Stores index tensors and sequence lengths for routing attention
computations across PCP ranks during long sequence processing.
"""
q_head_idx: torch.Tensor = None
q_tail_idx: torch.Tensor = None
kv_with_q_head_nomask_idx: torch.Tensor = None
@@ -27,7 +25,7 @@ class AscendPCPMetadata:
head_attn_nomask_seqlens: torch.Tensor = None
tail_attn_nomask_seqlens: torch.Tensor = None
q_full_idx: torch.Tensor = None
pcp_allgather_restore_idx: Optional[list[int]] = None
pcp_allgather_restore_idx: list[int] | None = None
@dataclass
@@ -37,6 +35,7 @@ class CPChunkedContextMetadata:
Extends chunked prefill with per-rank chunk information for PCP/DCP.
"""
# For handling chunked prefill
cu_seq_lens: torch.Tensor
starts: torch.Tensor
@@ -47,48 +46,51 @@ class CPChunkedContextMetadata:
chunk_seq_lens_npu: torch.Tensor
# for mla DCP & PCP
padded_chunk_seq_lens_npu: torch.Tensor = None
padded_local_chunk_seq_lens: Optional[list[list[int]]] = None
local_context_lens_allranks: Optional[list[list[int]]] = None
padded_local_chunk_seq_lens: list[list[int]] | None = None
local_context_lens_allranks: list[list[int]] | None = None
padded_local_cu_seq_lens: torch.Tensor = None
cu_seq_lens_lst: Optional[list[list[int]]] = None
chunk_size: Optional[int] = None
cu_seq_lens_lst: list[list[int]] | None = None
chunk_size: int | None = None
@dataclass
class AscendMetadataForPrefill:
""" Prefill-specific metadata for Ascend attention with Context Parallelism."""
"""Prefill-specific metadata for Ascend attention with Context Parallelism."""
@dataclass
class ChunkedContextMetadata:
"""Metadata for chunked context processing within prefill phase."""
actual_chunk_seq_lengths: torch.Tensor
actual_seq_lengths_kv: torch.Tensor
starts: torch.Tensor
chunk_seq_mask_filtered_indices: torch.Tensor
chunked_req_mask: Optional[list[bool]] = None
local_context_lens_allranks: Optional[list[list[int]]] = None
cp_kv_recover_idx_for_chunk: Optional[list[int]] = None
kv_inverse_idx_for_chunk: Optional[list[int]] = None
batch_chunk_seq_mask: Optional[list[bool]] = None
local_total_toks: Optional[int] = None
chunked_req_mask: list[bool] | None = None
local_context_lens_allranks: list[list[int]] | None = None
cp_kv_recover_idx_for_chunk: list[int] | None = None
kv_inverse_idx_for_chunk: list[int] | None = None
batch_chunk_seq_mask: list[bool] | None = None
local_total_toks: int | None = None
""" Prefill Specific Metadata for Ascend"""
pcp_metadata: Optional[AscendPCPMetadata] = None
chunked_context: Optional[ChunkedContextMetadata] = None
pcp_metadata: AscendPCPMetadata | None = None
chunked_context: ChunkedContextMetadata | None = None
block_tables: torch.Tensor = None
actual_seq_lengths_q: torch.Tensor = None
@dataclass
class AscendMetadataForDecode:
""" Decode-specific metadata for Ascend attention with Context Parallelism."""
num_computed_tokens_of_pcp_dcp: Optional[list[list[list[int]]]] = None
"""Decode-specific metadata for Ascend attention with Context Parallelism."""
num_computed_tokens_of_pcp_dcp: list[list[list[int]]] | None = None
batch_seq_mask: torch.Tensor = None
block_tables: torch.Tensor = None
def _process_attn_out_lse(attn_output: torch.Tensor, softmax_lse: torch.Tensor,
batch_seq_mask: torch.Tensor) -> torch.Tensor:
def _process_attn_out_lse(
attn_output: torch.Tensor, softmax_lse: torch.Tensor, batch_seq_mask: torch.Tensor
) -> torch.Tensor:
pcp_size = get_pcp_group().world_size
dcp_size = get_decode_context_model_parallel_world_size()
dcp_group = get_dcp_group().device_group if dcp_size > 1 else None
@@ -104,21 +106,17 @@ def _process_attn_out_lse(attn_output: torch.Tensor, softmax_lse: torch.Tensor,
# permute: [bs, num_heads, v_head_dim+1] -> [num_heads, v_head_dim+1, bs]
attn_out_lse = attn_out_lse.permute([1, 2, 0]).contiguous()
attn_out_lse_all2all = torch.empty_like(attn_out_lse)
dist.all_to_all_single(attn_out_lse_all2all,
attn_out_lse,
group=dcp_group)
dist.all_to_all_single(attn_out_lse_all2all, attn_out_lse, group=dcp_group)
attn_out_lse = attn_out_lse_all2all.permute([2, 0, 1])
if pcp_size > 1:
# AllGather out&lse within CP group
attn_out_lse = get_pcp_group().all_gather(attn_out_lse.contiguous(),
dim=0)
attn_out_lse = get_pcp_group().all_gather(attn_out_lse.contiguous(), dim=0)
return attn_out_lse
def _npu_attention_update(head_size,
attn_out_lse: torch.Tensor) -> torch.Tensor:
def _npu_attention_update(head_size, attn_out_lse: torch.Tensor) -> torch.Tensor:
pcp_size = get_pcp_group().world_size
dcp_size = get_decode_context_model_parallel_world_size()
# [PCP * S, DCP * H, D+1]
@@ -134,8 +132,7 @@ def _npu_attention_update(head_size,
# Flatten [N, S, H, D+1], N = pcp_size * dcp_size
x = x.view(-1, S, H, D_plus_1)
# Split out lse
out_flat, lse_flat = torch.split(x, [D, 1],
dim=-1) # [N, S, H, D], [N, S, H, 1]
out_flat, lse_flat = torch.split(x, [D, 1], dim=-1) # [N, S, H, D], [N, S, H, 1]
# out: [N, S, H, D] -> [N, S*H, D]
# lse: [N, S, H, 1] -> [N, S*H]
out_flat = out_flat.flatten(1, 2) # [N, S*H, D]

502
vllm_ascend/attention/context_parallel/mla_cp.py
View File

@@ -1,35 +1,43 @@
from typing import Optional, Tuple, TypeVar
from typing import TypeVar
import numpy as np
import torch
import torch_npu
from vllm.config import VllmConfig
from vllm.distributed import (get_dcp_group,
get_decode_context_model_parallel_rank,
get_decode_context_model_parallel_world_size,
get_pcp_group)
from vllm.distributed import (
get_dcp_group,
get_decode_context_model_parallel_rank,
get_decode_context_model_parallel_world_size,
get_pcp_group,
)
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.utils.math_utils import cdiv
from vllm.v1.kv_cache_interface import AttentionSpec, MLAAttentionSpec
# isort: off
from vllm_ascend.attention.mla_v1 import (
AscendMLADecodeMetadata, AscendMLAImpl, AscendMLAMetadata,
AscendMLAMetadataBuilder, AscendMLAPrefillMetadata,
DecodeMLAPreprocessResult, PrefillMLAPreprocessResult,
BUILD_METADATA_STEP_PREFILL)
#isort: on
AscendMLADecodeMetadata,
AscendMLAImpl,
AscendMLAMetadata,
AscendMLAMetadataBuilder,
AscendMLAPrefillMetadata,
DecodeMLAPreprocessResult,
PrefillMLAPreprocessResult,
BUILD_METADATA_STEP_PREFILL,
)
# isort: on
from vllm_ascend.attention.utils import (AscendCommonAttentionMetadata)
from vllm_ascend.attention.context_parallel.common_cp import (
AscendPCPMetadata, CPChunkedContextMetadata, _process_attn_out_lse,
_npu_attention_update)
from vllm_ascend.compilation.acl_graph import (get_draft_graph_params,
get_graph_params,
update_graph_params_workspaces)
from vllm_ascend.utils import weak_ref_tensors, vllm_version_is
AscendPCPMetadata,
CPChunkedContextMetadata,
_npu_attention_update,
_process_attn_out_lse,
)
from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
from vllm_ascend.compilation.acl_graph import get_draft_graph_params, get_graph_params, update_graph_params_workspaces
from vllm_ascend.utils import vllm_version_is, weak_ref_tensors
if vllm_version_is('0.13.0'):
if vllm_version_is("0.13.0"):
from vllm.v1.attention.backends.utils import AttentionCGSupport
else:
from vllm.v1.attention.backend import AttentionCGSupport
@@ -54,28 +62,21 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
metadata_cls: type[AscendMLAMetadata] | None = None,
supports_dcp_with_varlen: bool = False,
):
super().__init__(kv_cache_spec, layer_names, vllm_config, device,
metadata_cls, supports_dcp_with_varlen)
super().__init__(kv_cache_spec, layer_names, vllm_config, device, metadata_cls, supports_dcp_with_varlen)
self.pcp_size = get_pcp_group().world_size
self.pcp_rank = get_pcp_group(
).rank_in_group if self.pcp_size > 1 else 0
self.pcp_rank = get_pcp_group().rank_in_group if self.pcp_size > 1 else 0
self.dcp_size = get_decode_context_model_parallel_world_size()
self.dcp_rank = get_decode_context_model_parallel_rank(
) if self.dcp_size > 1 else 0
self.dcp_rank = get_decode_context_model_parallel_rank() if self.dcp_size > 1 else 0
self.cp_local_block_size = vllm_config.parallel_config.cp_kv_cache_interleave_size
self.cp_virtual_block_size = self.cp_local_block_size * self.dcp_size * self.pcp_size
scheduler_config = vllm_config.scheduler_config
decode_max_num_seqs = getattr(scheduler_config, 'decode_max_num_seqs',
0)
decode_max_num_seqs = getattr(scheduler_config, "decode_max_num_seqs", 0)
max_num_seqs = max(scheduler_config.max_num_seqs, decode_max_num_seqs)
self.batch_seq_mask_buf = torch.empty(max_num_seqs *
self.decode_threshold,
dtype=torch.uint8,
device=device)
self.block_size = (self.block_size *
self.cp_virtual_block_size) // np.gcd(
self.block_size, self.cp_virtual_block_size)
self.batch_seq_mask_buf = torch.empty(max_num_seqs * self.decode_threshold, dtype=torch.uint8, device=device)
self.block_size = (self.block_size * self.cp_virtual_block_size) // np.gcd(
self.block_size, self.cp_virtual_block_size
)
def build(
self,
@@ -85,15 +86,10 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
) -> AscendMLAMetadata:
metadata_cls = super().build(common_prefix_len, common_attn_metadata)
if self.num_prefills == 0 and self.pcp_size > 1:
self.slot_mapping[:self.
num_decode_tokens] = self.slot_mapping[:self.
num_decode_tokens
* self.
pcp_size:
self.
pcp_size]
self.slot_mapping[self.num_decode_tokens:self.num_decode_tokens *
self.pcp_size].fill_(-1)
self.slot_mapping[: self.num_decode_tokens] = self.slot_mapping[
: self.num_decode_tokens * self.pcp_size : self.pcp_size
]
self.slot_mapping[self.num_decode_tokens : self.num_decode_tokens * self.pcp_size].fill_(-1)
metadata_cls.slot_mapping = self.slot_mapping
return metadata_cls
@@ -118,8 +114,8 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
# In dcp only spec decode graph padding case,
# num_actual_tokens_pcp_padded may be less than num_actual_tokens
self.num_actual_tokens = max(
long_seq_metadata.num_actual_tokens_pcp_padded,
common_attn_metadata.num_actual_tokens)
long_seq_metadata.num_actual_tokens_pcp_padded, common_attn_metadata.num_actual_tokens
)
def build_cp_metadata(
self,
@@ -131,30 +127,23 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
return AscendPCPMetadata(
q_head_idx=common_long_seq_metadata.q_head_idx_tensor,
q_tail_idx=common_long_seq_metadata.q_tail_idx_tensor,
kv_with_q_head_nomask_idx=common_long_seq_metadata.
kv_with_q_head_nomask_idx_tensor,
kv_with_q_head_mask_idx=common_long_seq_metadata.
kv_with_q_head_mask_idx_tensor,
kv_with_q_tail_nomask_idx=common_long_seq_metadata.
kv_with_q_tail_nomask_idx_tensor,
kv_with_q_tail_mask_idx=common_long_seq_metadata.
kv_with_q_tail_mask_idx_tensor,
kv_with_q_head_nomask_idx=common_long_seq_metadata.kv_with_q_head_nomask_idx_tensor,
kv_with_q_head_mask_idx=common_long_seq_metadata.kv_with_q_head_mask_idx_tensor,
kv_with_q_tail_nomask_idx=common_long_seq_metadata.kv_with_q_tail_nomask_idx_tensor,
kv_with_q_tail_mask_idx=common_long_seq_metadata.kv_with_q_tail_mask_idx_tensor,
attn_mask_seqlens=common_long_seq_metadata.attn_mask_seqlens,
head_attn_nomask_seqlens=common_long_seq_metadata.
head_attn_nomask_seqlens,
tail_attn_nomask_seqlens=common_long_seq_metadata.
tail_attn_nomask_seqlens,
head_attn_nomask_seqlens=common_long_seq_metadata.head_attn_nomask_seqlens,
tail_attn_nomask_seqlens=common_long_seq_metadata.tail_attn_nomask_seqlens,
q_full_idx=common_long_seq_metadata.q_full_idx,
pcp_allgather_restore_idx=common_long_seq_metadata.
pcp_allgather_restore_idx)
pcp_allgather_restore_idx=common_long_seq_metadata.pcp_allgather_restore_idx,
)
def build_chunked_metadata(
self,
common_prefix_len: int,
common_attn_metadata: AscendCommonAttentionMetadata,
):
chunked_context_metadata = super().build_chunked_metadata(
common_prefix_len, common_attn_metadata)
chunked_context_metadata = super().build_chunked_metadata(common_prefix_len, common_attn_metadata)
if chunked_context_metadata is None:
return None
@@ -162,33 +151,37 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
assert long_seq_metadata is not None
num_computed_tokens_of_pcp_dcp = long_seq_metadata.num_computed_tokens_of_pcp_dcp
assert num_computed_tokens_of_pcp_dcp is not None
local_context_lens_allranks = torch.tensor(
num_computed_tokens_of_pcp_dcp[self.num_decodes_flatten:]).reshape(
-1, self.dcp_size * self.pcp_size)
local_context_lens_allranks = torch.tensor(num_computed_tokens_of_pcp_dcp[self.num_decodes_flatten :]).reshape(
-1, self.dcp_size * self.pcp_size
)
# Note(qcs): The max local context lengths
# padded to `cp_local_block_size`.
padded_local_context_lens_cpu = (cdiv(
self.context_lens_cpu,
self.cp_virtual_block_size,
) * self.cp_local_block_size)
padded_local_max_context_chunk_across_ranks = (cdiv(
self.max_context_chunk,
self.cp_virtual_block_size,
) * self.cp_local_block_size)
local_chunk_starts = (torch.arange(
self.num_chunks, dtype=torch.int32).unsqueeze(1).expand(
-1, self.num_prefills) *
padded_local_max_context_chunk_across_ranks)
padded_local_context_lens_cpu = (
cdiv(
self.context_lens_cpu,
self.cp_virtual_block_size,
)
* self.cp_local_block_size
)
padded_local_max_context_chunk_across_ranks = (
cdiv(
self.max_context_chunk,
self.cp_virtual_block_size,
)
* self.cp_local_block_size
)
local_chunk_starts = (
torch.arange(self.num_chunks, dtype=torch.int32).unsqueeze(1).expand(-1, self.num_prefills)
* padded_local_max_context_chunk_across_ranks
)
local_chunk_ends = torch.min(
padded_local_context_lens_cpu.unsqueeze(0),
local_chunk_starts + padded_local_max_context_chunk_across_ranks,
)
padded_local_chunk_seq_lens = (local_chunk_ends -
local_chunk_starts).clamp(min=0)
padded_local_cu_chunk_seq_lens_cpu = torch.zeros(self.num_chunks,
self.num_prefills + 1,
dtype=torch.int32,
pin_memory=True)
padded_local_chunk_seq_lens = (local_chunk_ends - local_chunk_starts).clamp(min=0)
padded_local_cu_chunk_seq_lens_cpu = torch.zeros(
self.num_chunks, self.num_prefills + 1, dtype=torch.int32, pin_memory=True
)
torch.cumsum(
padded_local_chunk_seq_lens,
dim=1,
@@ -197,8 +190,7 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
)
chunked_metadata = CPChunkedContextMetadata(
cu_seq_lens=chunked_context_metadata.cu_seq_lens,
starts=local_chunk_starts.pin_memory().to(self.device,
non_blocking=True),
starts=local_chunk_starts.pin_memory().to(self.device, non_blocking=True),
seq_tot=padded_local_chunk_seq_lens.sum(dim=1).tolist(),
max_seq_lens=chunked_context_metadata.max_seq_lens,
chunk_seq_lens=self.chunk_seq_lens,
@@ -207,18 +199,14 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
padded_chunk_seq_lens_npu=padded_local_chunk_seq_lens.npu(),
padded_local_chunk_seq_lens=padded_local_chunk_seq_lens.tolist(),
local_context_lens_allranks=local_context_lens_allranks.tolist(),
padded_local_cu_seq_lens=padded_local_cu_chunk_seq_lens_cpu.
pin_memory().to(self.device, non_blocking=True),
padded_local_cu_seq_lens=padded_local_cu_chunk_seq_lens_cpu.pin_memory().to(self.device, non_blocking=True),
cu_seq_lens_lst=self.cu_seq_lens_cpu.tolist(),
chunk_size=padded_local_max_context_chunk_across_ranks,
)
return chunked_metadata
def get_block_table_size(
self, common_attn_metadata: AscendCommonAttentionMetadata,
build_metadata_step: int):
self.num_decodes_flatten = self.query_lens[:self.num_decodes].sum(
).item()
def get_block_table_size(self, common_attn_metadata: AscendCommonAttentionMetadata, build_metadata_step: int):
self.num_decodes_flatten = self.query_lens[: self.num_decodes].sum().item()
if build_metadata_step == BUILD_METADATA_STEP_PREFILL:
# For pcp + spec decode, we flatten seq_lens and block_table
# to avoid irregular attn_mask shape
@@ -231,12 +219,9 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
common_prefix_len: int,
common_attn_metadata: AscendCommonAttentionMetadata,
) -> AscendMLAPrefillMetadata:
prefill_metadata = super().build_prefill_metadata(
common_prefix_len, common_attn_metadata)
prefill_metadata.pcp_metadata = self.build_cp_metadata(
common_prefix_len, common_attn_metadata)
prefill_metadata.block_table = self.block_table[
self.num_decodes_flatten:, ...]
prefill_metadata = super().build_prefill_metadata(common_prefix_len, common_attn_metadata)
prefill_metadata.pcp_metadata = self.build_cp_metadata(common_prefix_len, common_attn_metadata)
prefill_metadata.block_table = self.block_table[self.num_decodes_flatten :, ...]
return prefill_metadata
def build_decode_metadata(
@@ -244,24 +229,20 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
common_prefix_len: int,
common_attn_metadata: AscendCommonAttentionMetadata,
) -> AscendMLADecodeMetadata:
decode_metadata = super().build_decode_metadata(
common_prefix_len, common_attn_metadata)
decode_metadata = super().build_decode_metadata(common_prefix_len, common_attn_metadata)
long_seq_metadata = common_attn_metadata.prefill_context_parallel_metadata
assert long_seq_metadata is not None
num_computed_tokens_of_pcp_dcp = long_seq_metadata.num_computed_tokens_of_pcp_dcp
assert num_computed_tokens_of_pcp_dcp is not None
# [bs, pcp_size, dcp_size]
num_computed_tokens_of_cp_dcp_array = np.array(
num_computed_tokens_of_pcp_dcp)[:self.num_decodes_flatten]
num_computed_tokens_of_cp_dcp_array = np.array(num_computed_tokens_of_pcp_dcp)[: self.num_decodes_flatten]
cp_seq_len = num_computed_tokens_of_cp_dcp_array[:, self.pcp_rank,
self.dcp_rank]
cp_seq_len = num_computed_tokens_of_cp_dcp_array[:, self.pcp_rank, self.dcp_rank]
cp_seq_len = torch.tensor(cp_seq_len, dtype=torch.int32)
batch_seq_mask = (cp_seq_len == 0)
self.batch_seq_mask_buf[:batch_seq_mask.shape[0]].copy_(
batch_seq_mask, non_blocking=True)
batch_seq_mask = self.batch_seq_mask_buf[:batch_seq_mask.shape[0]]
batch_seq_mask = cp_seq_len == 0
self.batch_seq_mask_buf[: batch_seq_mask.shape[0]].copy_(batch_seq_mask, non_blocking=True)
batch_seq_mask = self.batch_seq_mask_buf[: batch_seq_mask.shape[0]]
cp_seq_len = torch.where(cp_seq_len == 0, 1, cp_seq_len)
decode_metadata.cp_seq_len = cp_seq_len
decode_metadata.batch_seq_mask = batch_seq_mask
@@ -280,30 +261,35 @@ class AscendMlaCPImpl(AscendMLAImpl):
head_size: int,
scale: float,
num_kv_heads: int,
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
alibi_slopes: list[float] | None,
sliding_window: int | None,
kv_cache_dtype: str,
logits_soft_cap: Optional[float],
logits_soft_cap: float | None,
attn_type: str,
kv_sharing_target_layer_name: Optional[str],
kv_sharing_target_layer_name: str | None,
**kwargs,
):
super().__init__(num_heads, head_size, scale, num_kv_heads,
alibi_slopes, sliding_window, kv_cache_dtype,
logits_soft_cap, attn_type,
kv_sharing_target_layer_name, **kwargs)
super().__init__(
num_heads,
head_size,
scale,
num_kv_heads,
alibi_slopes,
sliding_window,
kv_cache_dtype,
logits_soft_cap,
attn_type,
kv_sharing_target_layer_name,
**kwargs,
)
self.pcp_size = get_pcp_group().world_size
self.pcp_rank = get_pcp_group(
).rank_in_group if self.pcp_size > 1 else 0
self.pcp_group = get_pcp_group(
).device_group if self.pcp_size > 1 else None
self.pcp_rank = get_pcp_group().rank_in_group if self.pcp_size > 1 else 0
self.pcp_group = get_pcp_group().device_group if self.pcp_size > 1 else None
self.dcp_size = get_decode_context_model_parallel_world_size()
self.dcp_rank = get_decode_context_model_parallel_rank(
) if self.dcp_size > 1 else 0
self.dcp_group = get_dcp_group(
).device_group if self.dcp_size > 1 else None
self.dcp_rank = get_decode_context_model_parallel_rank() if self.dcp_size > 1 else 0
self.dcp_group = get_dcp_group().device_group if self.dcp_size > 1 else None
def get_num_actual_tokens(self, attn_metadata: M):
if self.pcp_size > 1:
@@ -320,103 +306,80 @@ class AscendMlaCPImpl(AscendMLAImpl):
x = x.transpose(0, 1).reshape(-1, self.num_heads * self.v_head_dim)
return x
def mla_preprocess_prefill(self, q_c, kv_no_split, kv_cache,
attn_metadata):
def mla_preprocess_prefill(self, q_c, kv_no_split, kv_cache, attn_metadata):
if not self.pcp_size > 1:
return super().mla_preprocess_prefill(q_c, kv_no_split, kv_cache,
attn_metadata)
return super().mla_preprocess_prefill(q_c, kv_no_split, kv_cache, attn_metadata)
num_decode_tokens = attn_metadata.num_decode_tokens
num_actual_tokens = (attn_metadata.num_actual_tokens_pcp_padded -
self.pcp_size * num_decode_tokens
) // self.pcp_size + num_decode_tokens
num_actual_tokens = (
attn_metadata.num_actual_tokens_pcp_padded - self.pcp_size * num_decode_tokens
) // self.pcp_size + num_decode_tokens
prefill_q_c = q_c[num_decode_tokens:num_actual_tokens]
prefill_q = self.q_proj(prefill_q_c)[0] \
.view(-1, self.num_heads, self.qk_head_dim)
prefill_q_pe = prefill_q[..., self.qk_nope_head_dim:]
prefill_q_nope = prefill_q[..., :self.qk_nope_head_dim]
cos = attn_metadata.prefill.cos[:num_actual_tokens - num_decode_tokens]
sin = attn_metadata.prefill.sin[:num_actual_tokens - num_decode_tokens]
prefill_q = self.q_proj(prefill_q_c)[0].view(-1, self.num_heads, self.qk_head_dim)
prefill_q_pe = prefill_q[..., self.qk_nope_head_dim :]
prefill_q_nope = prefill_q[..., : self.qk_nope_head_dim]
cos = attn_metadata.prefill.cos[: num_actual_tokens - num_decode_tokens]
sin = attn_metadata.prefill.sin[: num_actual_tokens - num_decode_tokens]
prefill_q_pe = self.rope_single(prefill_q_pe, cos, sin)
prefill_kv_no_split = kv_no_split[:num_actual_tokens]
kv_c, k_pe = prefill_kv_no_split.split(
[self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
kv_c, k_pe = prefill_kv_no_split.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
assert len(
kv_cache
) > 1, "the number of kv cache should be greater than 1, namely (nope_cache and rope_cache)"
kv_c_normed = kv_c_normed.view(
[num_actual_tokens, self.num_kv_heads, -1])
assert len(kv_cache) > 1, "the number of kv cache should be greater than 1, namely (nope_cache and rope_cache)"
kv_c_normed = kv_c_normed.view([num_actual_tokens, self.num_kv_heads, -1])
k_pe = k_pe.unsqueeze(1)
prefill_k_pe = k_pe
prefill_k_pe[num_decode_tokens:num_actual_tokens] = self.rope_single(
prefill_k_pe[num_decode_tokens:num_actual_tokens], cos, sin)
prefill_k_pe[num_decode_tokens:num_actual_tokens], cos, sin
)
prefill_k_c_normed = kv_c_normed[:num_actual_tokens]
prefill_kv_c_k_pe = torch.cat([prefill_k_c_normed, prefill_k_pe],
dim=-1)
prefill_kv_c_k_pe = torch.cat([prefill_k_c_normed, prefill_k_pe], dim=-1)
prefill_kv_c_k_pe = get_pcp_group().all_gather(prefill_kv_c_k_pe, 0)
prefill_kv_c_k_pe = torch.index_select(
prefill_kv_c_k_pe, 0,
attn_metadata.prefill.pcp_metadata.pcp_allgather_restore_idx)
prefill_kv_c_k_pe = prefill_kv_c_k_pe[num_decode_tokens *
self.pcp_size:]
prefill_k_c_normed, prefill_k_pe = prefill_kv_c_k_pe.split(
[self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
prefill_kv_c_k_pe, 0, attn_metadata.prefill.pcp_metadata.pcp_allgather_restore_idx
)
prefill_kv_c_k_pe = prefill_kv_c_k_pe[num_decode_tokens * self.pcp_size :]
prefill_k_c_normed, prefill_k_pe = prefill_kv_c_k_pe.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
kv_c_normed, k_pe = prefill_k_c_normed, prefill_k_pe
prefill_k_c_normed = prefill_k_c_normed.squeeze()
slot_mapping = attn_metadata.slot_mapping[self.pcp_size *
num_decode_tokens:]
torch_npu._npu_reshape_and_cache(key=kv_c_normed,
value=k_pe,
key_cache=kv_cache[0],
value_cache=kv_cache[1],
slot_indices=slot_mapping)
prefill_k_nope, prefill_value = self.kv_b_proj(
prefill_k_c_normed)[0].view(
-1, self.num_heads,
self.qk_nope_head_dim + self.v_head_dim).split(
[self.qk_nope_head_dim, self.v_head_dim], dim=-1)
slot_mapping = attn_metadata.slot_mapping[self.pcp_size * num_decode_tokens :]
torch_npu._npu_reshape_and_cache(
key=kv_c_normed, value=k_pe, key_cache=kv_cache[0], value_cache=kv_cache[1], slot_indices=slot_mapping
)
prefill_k_nope, prefill_value = (
self.kv_b_proj(prefill_k_c_normed)[0]
.view(-1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
)
prefill_k_pe = prefill_k_pe.expand((*prefill_k_nope.shape[:-1], -1))
return PrefillMLAPreprocessResult(prefill_q_nope, prefill_q_pe,
prefill_k_nope, prefill_k_pe,
prefill_value)
return PrefillMLAPreprocessResult(prefill_q_nope, prefill_q_pe, prefill_k_nope, prefill_k_pe, prefill_value)
def mla_preprocess_decode(self, q_c, kv_no_split, kv_cache, attn_metadata):
num_decode_tokens = attn_metadata.num_decode_tokens
decode_q_c = q_c[:num_decode_tokens]
cos = attn_metadata.decode.cos
sin = attn_metadata.decode.sin
decode_ql_nope, decode_q_pe = \
self._q_proj_and_k_up_proj(decode_q_c)
decode_ql_nope, decode_q_pe = self.reorg_decode_q(
decode_ql_nope, decode_q_pe)
decode_ql_nope, decode_q_pe = self._q_proj_and_k_up_proj(decode_q_c)
decode_ql_nope, decode_q_pe = self.reorg_decode_q(decode_ql_nope, decode_q_pe)
decode_q_pe = self.rope_single(decode_q_pe, cos, sin)
decode_slots = attn_metadata.slot_mapping[:num_decode_tokens]
decode_kv_no_split = kv_no_split[:num_decode_tokens]
decode_k_pe, decode_k_nope = self.exec_kv_decode(
decode_kv_no_split, cos, sin, kv_cache, decode_slots)
return DecodeMLAPreprocessResult(decode_ql_nope, decode_q_pe,
decode_k_nope, decode_k_pe)
decode_k_pe, decode_k_nope = self.exec_kv_decode(decode_kv_no_split, cos, sin, kv_cache, decode_slots)
return DecodeMLAPreprocessResult(decode_ql_nope, decode_q_pe, decode_k_nope, decode_k_pe)
def get_context_seq_len_npu(self, index: int,
attn_metadata: AscendMLAMetadata):
def get_context_seq_len_npu(self, index: int, attn_metadata: AscendMLAMetadata):
prefill_metadata = attn_metadata.prefill
assert prefill_metadata is not None
assert prefill_metadata.chunked_context is not None
assert isinstance(prefill_metadata.chunked_context,
CPChunkedContextMetadata)
assert isinstance(prefill_metadata.chunked_context, CPChunkedContextMetadata)
assert prefill_metadata.chunked_context.padded_chunk_seq_lens_npu is not None
iters = len(prefill_metadata.chunked_context.seq_tot)
assert 0 <= index < iters
return prefill_metadata.chunked_context.padded_chunk_seq_lens_npu[
index]
return prefill_metadata.chunked_context.padded_chunk_seq_lens_npu[index]
def reorg_decode_q(self, decode_q_nope, decode_q_pe):
if self.dcp_size > 1:
decode_q_no_split = torch.cat([decode_q_nope, decode_q_pe], dim=-1)
decode_q_no_split = get_dcp_group().all_gather(
decode_q_no_split, 1)
decode_q_nope, decode_q_pe = decode_q_no_split.split(
[self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
decode_q_no_split = get_dcp_group().all_gather(decode_q_no_split, 1)
decode_q_nope, decode_q_pe = decode_q_no_split.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
return decode_q_nope, decode_q_pe
def _forward_prefill(
@@ -426,12 +389,11 @@ class AscendMlaCPImpl(AscendMLAImpl):
k_nope: torch.Tensor,
k_pe: torch.Tensor,
value: torch.Tensor,
kv_c_and_k_pe_cache: Tuple[torch.Tensor],
kv_c_and_k_pe_cache: tuple[torch.Tensor],
attn_metadata: AscendMLAMetadata,
) -> torch.Tensor:
if not self.pcp_size > 1:
return super()._forward_prefill(q_nope, q_pe, k_nope, k_pe, value,
kv_c_and_k_pe_cache, attn_metadata)
return super()._forward_prefill(q_nope, q_pe, k_nope, k_pe, value, kv_c_and_k_pe_cache, attn_metadata)
assert attn_metadata.prefill is not None
assert attn_metadata.prefill.pcp_metadata is not None
num_tokens = q_nope.size(0)
@@ -455,7 +417,8 @@ class AscendMlaCPImpl(AscendMLAImpl):
kv_nomask_idx=kv_with_q_head_nomask_idx,
attn_mask_seqlens=attn_mask_seqlens,
attn_nomask_seqlens=head_attn_nomask_seqlens,
mask=attn_metadata.attn_mask)
mask=attn_metadata.attn_mask,
)
output_tail, lse_tail = self._attention_with_mask_and_nomask(
q_nope=torch.index_select(q_nope, 0, q_tail_idx),
@@ -467,19 +430,16 @@ class AscendMlaCPImpl(AscendMLAImpl):
kv_nomask_idx=kv_with_q_tail_nomask_idx,
attn_mask_seqlens=attn_mask_seqlens,
attn_nomask_seqlens=tail_attn_nomask_seqlens,
mask=attn_metadata.attn_mask)
mask=attn_metadata.attn_mask,
)
q_full_idx = attn_metadata.prefill.pcp_metadata.q_full_idx
attn_output = torch.index_select(
torch.cat([output_head, output_tail], dim=0), 0, q_full_idx)
attn_lse = torch.index_select(torch.cat([lse_head, lse_tail], dim=1),
1, q_full_idx)
attn_output = torch.index_select(torch.cat([output_head, output_tail], dim=0), 0, q_full_idx)
attn_lse = torch.index_select(torch.cat([lse_head, lse_tail], dim=1), 1, q_full_idx)
output, _ = self._compute_prefill_context(q_nope, q_pe,
kv_c_and_k_pe_cache,
self.qk_rope_head_dim,
attn_metadata, attn_output,
attn_lse)
output, _ = self._compute_prefill_context(
q_nope, q_pe, kv_c_and_k_pe_cache, self.qk_rope_head_dim, attn_metadata, attn_output, attn_lse
)
output = output.reshape([num_tokens, self.num_heads * self.v_head_dim])
@@ -498,44 +458,40 @@ class AscendMlaCPImpl(AscendMLAImpl):
attn_nomask_seqlens: list[torch.Tensor],
mask: torch.Tensor,
):
attn_output = torch.empty(q_nope.shape[0],
self.num_heads,
self.v_head_dim,
dtype=k_pe.dtype,
device=k_pe.device)
attn_lse = torch.empty(self.num_heads,
q_pe.shape[0],
dtype=torch.float32,
device=k_pe.device)
attn_output = torch.empty(
q_nope.shape[0], self.num_heads, self.v_head_dim, dtype=k_pe.dtype, device=k_pe.device
)
attn_lse = torch.empty(self.num_heads, q_pe.shape[0], dtype=torch.float32, device=k_pe.device)
# mask
k_nope_mask = torch.index_select(k_nope, 0, kv_mask_idx)
value_mask = torch.index_select(value, 0, kv_mask_idx)
k_pe_mask = torch.index_select(k_pe, 0, kv_mask_idx)
torch_npu.atb.npu_ring_mla(q_nope=q_nope,
q_rope=q_pe,
k_nope=k_nope_mask,
k_rope=k_pe_mask,
value=value_mask,
mask=mask,
seqlen=attn_mask_seqlens,
head_num=self.num_heads,
kv_head_num=self.num_heads,
pre_out=None,
prev_lse=None,
qk_scale=self.scale,
kernel_type="kernel_type_high_precision",
mask_type="mask_type_triu",
input_layout="type_bsnd",
calc_type="calc_type_first_ring",
output=attn_output,
softmax_lse=attn_lse)
torch_npu.atb.npu_ring_mla(
q_nope=q_nope,
q_rope=q_pe,
k_nope=k_nope_mask,
k_rope=k_pe_mask,
value=value_mask,
mask=mask,
seqlen=attn_mask_seqlens,
head_num=self.num_heads,
kv_head_num=self.num_heads,
pre_out=None,
prev_lse=None,
qk_scale=self.scale,
kernel_type="kernel_type_high_precision",
mask_type="mask_type_triu",
input_layout="type_bsnd",
calc_type="calc_type_first_ring",
output=attn_output,
softmax_lse=attn_lse,
)
# nomask
if not kv_nomask_idx or len(kv_nomask_idx[0]) == 0:
return attn_output, attn_lse
for kv_nomask_idx_split, attn_nomask_seqlens_split in zip(
kv_nomask_idx, attn_nomask_seqlens):
for kv_nomask_idx_split, attn_nomask_seqlens_split in zip(kv_nomask_idx, attn_nomask_seqlens):
k_nope_nomask = torch.index_select(k_nope, 0, kv_nomask_idx_split)
value_nomask = torch.index_select(value, 0, kv_nomask_idx_split)
k_pe_nomask = torch.index_select(k_pe, 0, kv_nomask_idx_split)
@@ -557,7 +513,8 @@ class AscendMlaCPImpl(AscendMLAImpl):
input_layout="type_bsnd",
calc_type="calc_type_default",
output=attn_output,
softmax_lse=attn_lse)
softmax_lse=attn_lse,
)
return attn_output, attn_lse
def _forward_decode(
@@ -579,10 +536,8 @@ class AscendMlaCPImpl(AscendMLAImpl):
else:
num_heads = self.num_heads
k_nope = k_nope.view(-1, block_size, self.num_kv_heads,
self.kv_lora_rank)
k_pe = k_pe.view(-1, block_size, self.num_kv_heads,
self.qk_rope_head_dim)
k_nope = k_nope.view(-1, block_size, self.num_kv_heads, self.kv_lora_rank)
k_pe = k_pe.view(-1, block_size, self.num_kv_heads, self.qk_rope_head_dim)
q_nope = q_nope.view(num_tokens, num_heads, -1)
q_pe = q_pe.view(num_tokens, num_heads, -1)
# use pcp & dcp split computed token nums from scheduler to compute actual seq_len and seq_mask
@@ -606,20 +561,35 @@ class AscendMlaCPImpl(AscendMLAImpl):
workspace = graph_params.workspaces.get(num_tokens)
if workspace is None:
workspace = torch_npu.atb._npu_multi_head_latent_attention_get_workspace(
q_nope, q_pe, k_nope, k_pe, decode_meta.block_table,
seq_len, num_heads, self.scale, self.num_kv_heads,
**common_kwargs)
q_nope,
q_pe,
k_nope,
k_pe,
decode_meta.block_table,
seq_len,
num_heads,
self.scale,
self.num_kv_heads,
**common_kwargs,
)
update_graph_params_workspaces(num_tokens, workspace)
attn_output = torch.empty_like(q_nope)
softmax_lse = torch.empty((num_tokens, num_heads, 1),
dtype=q_nope.dtype,
device=q_nope.device)
softmax_lse = torch.empty((num_tokens, num_heads, 1), dtype=q_nope.dtype, device=q_nope.device)
graph_params.attn_params[num_tokens].append(
(weak_ref_tensors(q_nope), weak_ref_tensors(q_pe),
weak_ref_tensors(k_nope), weak_ref_tensors(k_pe),
decode_meta.block_table, seq_len, num_heads, self.scale,
self.num_kv_heads, weak_ref_tensors(attn_output),
weak_ref_tensors(softmax_lse)))
(
weak_ref_tensors(q_nope),
weak_ref_tensors(q_pe),
weak_ref_tensors(k_nope),
weak_ref_tensors(k_pe),
decode_meta.block_table,
seq_len,
num_heads,
self.scale,
self.num_kv_heads,
weak_ref_tensors(attn_output),
weak_ref_tensors(softmax_lse),
)
)
torch.npu.graph_task_group_begin(stream)
torch_npu.atb.npu_multi_head_latent_attention(
q_nope,
@@ -634,14 +604,13 @@ class AscendMlaCPImpl(AscendMLAImpl):
**common_kwargs,
workspace=workspace,
output=attn_output,
lse=softmax_lse)
lse=softmax_lse,
)
handle = torch.npu.graph_task_group_end(stream)
graph_params.handles[num_tokens].append(handle)
else:
attn_output = torch.empty_like(q_nope)
softmax_lse = torch.empty((num_tokens, num_heads, 1),
dtype=q_nope.dtype,
device=q_nope.device)
softmax_lse = torch.empty((num_tokens, num_heads, 1), dtype=q_nope.dtype, device=q_nope.device)
torch_npu.atb.npu_multi_head_latent_attention(
q_nope,
q_pe,
@@ -655,20 +624,17 @@ class AscendMlaCPImpl(AscendMLAImpl):
return_lse=True,
calc_type="calc_type_ring",
output=attn_output,
lse=softmax_lse)
lse=softmax_lse,
)
# Update out&lse
attn_out_lse = _process_attn_out_lse(attn_output, softmax_lse,
decode_meta.batch_seq_mask)
attn_out_lse = _process_attn_out_lse(attn_output, softmax_lse, decode_meta.batch_seq_mask)
attn_output = _npu_attention_update(self.kv_lora_rank, attn_out_lse)
return self._v_up_proj(attn_output)
def _out_lse_reshape(self, attn_out: torch.Tensor,
attn_lse: torch.Tensor) -> torch.Tensor:
attn_out = attn_out.contiguous().view(
attn_out.shape[0] * attn_out.shape[1], attn_out.shape[2])
attn_lse = attn_lse.contiguous().view(
attn_lse.shape[0] * attn_lse.shape[1] * attn_lse.shape[2])
def _out_lse_reshape(self, attn_out: torch.Tensor, attn_lse: torch.Tensor) -> torch.Tensor:
attn_out = attn_out.contiguous().view(attn_out.shape[0] * attn_out.shape[1], attn_out.shape[2])
attn_lse = attn_lse.contiguous().view(attn_lse.shape[0] * attn_lse.shape[1] * attn_lse.shape[2])
return attn_out, attn_lse
def _reorg_kvcache(
@@ -706,8 +672,7 @@ class AscendMlaCPImpl(AscendMLAImpl):
assert chunked_context.max_seq_lens is not None
assert chunked_context.chunk_size is not None
padded_local_chunk_seq_lens_lst = chunked_context.padded_local_chunk_seq_lens[
chunk_idx]
padded_local_chunk_seq_lens_lst = chunked_context.padded_local_chunk_seq_lens[chunk_idx]
local_context_lens_allranks = chunked_context.local_context_lens_allranks
sum_seq_len = chunked_context.cu_seq_lens_lst[chunk_idx][-1]
max_seq_len = chunked_context.max_seq_lens[chunk_idx]
@@ -720,14 +685,16 @@ class AscendMlaCPImpl(AscendMLAImpl):
cache_kv_c_k_pe = get_pcp_group().all_gather(cache_kv_c_k_pe, 0)
allgatered_kv_c_normed, allgatered_k_pe = cache_kv_c_k_pe.split(
[self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
[self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
)
kv_c_segments = []
k_pe_segments = []
src_token_idx = 0
max_seq_len_check = 0
for padded_local_chunk_seq_len, local_context_lens in zip(
padded_local_chunk_seq_lens_lst, local_context_lens_allranks):
padded_local_chunk_seq_lens_lst, local_context_lens_allranks
):
cur_seq_len = 0
for rank, local_context_len in enumerate(local_context_lens):
# Note(qcs): We split the context into multiple chunks,
@@ -742,15 +709,12 @@ class AscendMlaCPImpl(AscendMLAImpl):
padded_local_chunk_seq_len,
)
if local_chunk_len != 0:
kv_c_segment = allgatered_kv_c_normed[rank * toks +
src_token_idx:rank *
toks +
src_token_idx +
local_chunk_len]
k_pe_segment = allgatered_k_pe[rank * toks +
src_token_idx:rank * toks +
src_token_idx +
local_chunk_len]
kv_c_segment = allgatered_kv_c_normed[
rank * toks + src_token_idx : rank * toks + src_token_idx + local_chunk_len
]
k_pe_segment = allgatered_k_pe[
rank * toks + src_token_idx : rank * toks + src_token_idx + local_chunk_len
]
kv_c_segments.append(kv_c_segment)
k_pe_segments.append(k_pe_segment)
cur_seq_len += local_chunk_len