[Attention] Abstract the MLA prefill backends and eliminate cuDNN (#32623)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Matthew Bonanni
2026-05-01 13:36:20 -04:00
committed by GitHub
co-authored by Michael Goin Lucas Wilkinson mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 51295793a2
commit f3fef12350
16 changed files with 1629 additions and 708 deletions
@@ -189,12 +189,9 @@ return curr_o @ W_O
import functools
from abc import abstractmethod
from dataclasses import dataclass, field
from dataclasses import dataclass
from enum import Enum
from typing import TYPE_CHECKING, ClassVar, Generic, TypeVar, cast
if TYPE_CHECKING:
from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
from typing import ClassVar, Generic, TypeVar, cast
import torch
import torch.nn as nn
@@ -242,7 +239,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
kNvfp4Dynamic,
)
from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_flashinfer, has_nvidia_artifactory
from vllm.utils.flashinfer import has_flashinfer
from vllm.utils.math_utils import cdiv, round_down
from vllm.utils.torch_utils import (
LayerNameType,
@@ -262,11 +259,9 @@ from vllm.v1.attention.backend import (
MLAAttentionImpl,
SparseMLAAttentionImpl,
)
from vllm.v1.attention.backends.fa_utils import get_flash_attn_version
from vllm.v1.attention.backends.mla.prefill import MLAPrefillBackend
from vllm.v1.attention.backends.utils import (
get_dcp_local_seq_lens,
get_per_layer_parameters,
infer_global_hyperparameters,
split_decodes_and_prefills,
)
from vllm.v1.attention.ops.common import cp_lse_ag_out_rs
@@ -1123,33 +1118,6 @@ class QueryLenSupport(Enum):
VARLEN = "varlen"
try:
from vllm.vllm_flash_attn import ( # type: ignore[attr-defined]
flash_attn_varlen_func,
)
is_vllm_fa = True
except ImportError:
is_vllm_fa = False
flash_attn_varlen_func = None # type: ignore[assignment]
# On ROCm, vllm_flash_attn is not available, try upstream flash_attn instead.
# On CUDA, vllm_flash_attn should always be available (built with vLLM),
# so we don't attempt the fallback there.
if current_platform.is_rocm():
try:
from flash_attn import flash_attn_varlen_func # type: ignore[no-redef]
except ImportError:
logger.debug(
"flash_attn not available on ROCm; "
"MLA models using TRITON_MLA will require flash_attn. "
"AITER_MLA backends use aiter kernels instead."
)
elif current_platform.is_xpu():
from vllm._xpu_ops import xpu_ops
flash_attn_varlen_func = xpu_ops.flash_attn_varlen_func # type: ignore[no-redef,attr-defined,assignment]
def dynamic_per_batched_tensor_quant(
x: torch.Tensor, dtype: torch.dtype = torch.float8_e4m3fn
):
@@ -1161,9 +1129,6 @@ def dynamic_per_batched_tensor_quant(
return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
logger = init_logger(__name__)
@CustomOp.register(
"mla_decode_concat_quant_fp8",
dynamic_arg_dims={"decode_ql_nope": 0, "decode_q_pe": 0},
@@ -1197,9 +1162,6 @@ class _DecodeConcatQuantFP8(QuantFP8):
forward_hip = _make_forward(QuantFP8.forward_hip) # type: ignore[arg-type]
CUDNN_WORKSPACE_SIZE = 12800
class MLACommonBackend(AttentionBackend):
@staticmethod
def get_name() -> str:
@@ -1268,26 +1230,9 @@ class MLACommonPrefillMetadata:
query_start_loc: torch.Tensor
max_query_len: int
chunked_context: ChunkedContextMetadata | None = None
query_seq_lens: torch.Tensor | None = None
workspace_buffer: torch.Tensor | None = None
q_data_type: torch.dtype | None = None
output_dtype: torch.dtype | None = None
@dataclass
class FlashInferPrefillMetadata(MLACommonPrefillMetadata):
prefill_main: "BatchPrefillWithRaggedKVCacheWrapper | None" = None
prefill_chunks: "list[BatchPrefillWithRaggedKVCacheWrapper]" = field(
default_factory=list
)
@dataclass
class CudnnPrefillMetadata(MLACommonPrefillMetadata):
class ChunkedContextMetadata(MLACommonPrefillMetadata.ChunkedContextMetadata):
seq_lens: torch.Tensor
cudnn_workspace: torch.Tensor | None = None
prefill_backend: MLAPrefillBackend | None = None
@dataclass
@@ -1333,13 +1278,8 @@ class MLACommonMetadata(AttentionMetadata, Generic[D]):
# The dimension of the attention heads
head_dim: int | None = None
prefill: MLACommonPrefillMetadata | None = None
decode: D | None = None
prefill: (
MLACommonPrefillMetadata
| FlashInferPrefillMetadata
| CudnnPrefillMetadata
| None
) = None
def __post_init__(self):
if self.head_dim is not None and not MLACommonBackend.supports_head_size(
@@ -1352,64 +1292,6 @@ M = TypeVar("M", bound=MLACommonMetadata)
A = TypeVar("A", bound=AttentionMetadata)
def is_deepseek_r1_mla_compatible(vllm_config: VllmConfig) -> bool:
# Check if model has DeepSeek R1 compatible MLA dimensions:
# qk_nope_head_dim = 128, qk_rope_head_dim = 64, v_head_dim = 128
# which results in query/key head dim = 192.
if vllm_config.model_config is None:
return False
hf_text_config = vllm_config.model_config.hf_text_config
qk_nope_head_dim = getattr(hf_text_config, "qk_nope_head_dim", 1)
qk_rope_head_dim = getattr(hf_text_config, "qk_rope_head_dim", 1)
v_head_dim = getattr(hf_text_config, "v_head_dim", 1)
return qk_nope_head_dim == 128 and qk_rope_head_dim == 64 and v_head_dim == 128
@functools.cache
def use_flashinfer_prefill() -> bool:
from vllm.config import get_current_vllm_config
vllm_config = get_current_vllm_config()
if not (
not vllm_config.attention_config.disable_flashinfer_prefill
and has_flashinfer()
and not vllm_config.attention_config.use_cudnn_prefill
and current_platform.is_device_capability_family(100)
):
return False
return is_deepseek_r1_mla_compatible(vllm_config)
@functools.cache
def use_cudnn_prefill() -> bool:
from vllm.config import get_current_vllm_config
vllm_config = get_current_vllm_config()
return (
has_flashinfer()
and vllm_config.attention_config.use_cudnn_prefill
and current_platform.is_device_capability_family(100)
and has_nvidia_artifactory()
)
@functools.cache
def use_trtllm_ragged_deepseek_prefill() -> bool:
"""Check if TRT-LLM ragged DeepSeek prefill should be used."""
from vllm.config import get_current_vllm_config
vllm_config = get_current_vllm_config()
if not (
has_flashinfer()
and vllm_config.attention_config.use_trtllm_ragged_deepseek_prefill
and current_platform.is_device_capability_family(100)
):
return False
return is_deepseek_r1_mla_compatible(vllm_config)
@dataclass
class MLADims:
q_lora_rank: int | None
@@ -1447,15 +1329,14 @@ def get_mla_dims(model_config: ModelConfig) -> MLADims:
@functools.cache
def backend_supports_prefill_query_quantization() -> bool:
"""Check if the selected MLA backend supports prefill query quantization.
"""Check if the selected MLA prefill backend supports query quantization.
Currently supported backends:
- FlashInfer prefill
- TRT-LLM ragged DeepSeek prefill
- FlashInfer
- TRT-LLM Ragged
Not supported:
- cuDNN Prefill
- FlashAttention
- FlashAttention (FA3/FA4)
- Non-GB200 devices (FP8 prefill requires device capability 100)
"""
# FP8 prefill query quantization requires GB200 (device capability 100)
@@ -1463,7 +1344,15 @@ def backend_supports_prefill_query_quantization() -> bool:
if not current_platform.is_device_capability_family(100):
return False
return use_flashinfer_prefill() or use_trtllm_ragged_deepseek_prefill()
from vllm.config import get_current_vllm_config
from vllm.v1.attention.backends.mla.prefill import get_mla_prefill_backend
vllm_config = get_current_vllm_config()
backend_cls = get_mla_prefill_backend(vllm_config)
return backend_cls.get_name() in (
"FLASHINFER",
"TRTLLM_RAGGED",
)
class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
@@ -1574,7 +1463,6 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
metadata_cls if metadata_cls is not None else MLACommonMetadata
)
self.kv_cache_spec = kv_cache_spec
scheduler_config = vllm_config.scheduler_config
self.model_config = vllm_config.model_config
parallel_config = vllm_config.parallel_config
self.compilation_config = vllm_config.compilation_config
@@ -1634,139 +1522,32 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
device=device,
)
self._use_cudnn_prefill = use_cudnn_prefill()
self._use_fi_prefill = use_flashinfer_prefill()
self._use_trtllm_ragged_prefill = use_trtllm_ragged_deepseek_prefill()
self.prefill_metadata_cls = (
FlashInferPrefillMetadata
if self._use_fi_prefill
else CudnnPrefillMetadata
if self._use_cudnn_prefill
else MLACommonPrefillMetadata
from vllm.v1.attention.backends.mla.prefill import get_mla_prefill_backend
prefill_backend_cls = get_mla_prefill_backend(vllm_config)
self._prefill_backend = prefill_backend_cls(
num_heads=self.num_heads,
scale=self.model_config.get_head_size() ** -0.5,
kv_lora_rank=self.mla_dims.kv_lora_rank,
qk_nope_head_dim=self.mla_dims.qk_nope_head_dim,
qk_rope_head_dim=self.mla_dims.qk_rope_head_dim,
v_head_dim=self.mla_dims.v_head_dim,
vllm_config=vllm_config,
device=device,
layer_names=layer_names,
)
if self._use_fi_prefill:
self._workspace_buffer = torch.empty(
envs.VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE,
dtype=torch.uint8,
device=device,
)
self._fi_prefill_main: BatchPrefillWithRaggedKVCacheWrapper | None = None
self._fi_prefill_chunks: list[BatchPrefillWithRaggedKVCacheWrapper] = []
self._global_hyperparameters = infer_global_hyperparameters(
get_per_layer_parameters(vllm_config, layer_names, MLACommonImpl) # type: ignore[type-abstract]
)
if self._use_trtllm_ragged_prefill:
self._workspace_buffer = torch.empty(
envs.VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE,
dtype=torch.uint8,
device=device,
)
if self._use_cudnn_prefill:
self.cudnn_workspace = torch.empty(
CUDNN_WORKSPACE_SIZE * scheduler_config.max_num_seqs,
dtype=torch.int8,
device=device,
)
supports_spec_decode = self.query_len_support != QueryLenSupport.SINGLE_ONLY
self._init_reorder_batch_threshold(
self.reorder_batch_threshold, supports_spec_decode, supports_dcp_with_varlen
)
# Validate consistency between query_len_support and reorder_batch_threshold
if self.query_len_support == QueryLenSupport.SINGLE_ONLY:
assert self.reorder_batch_threshold == 1, (
f"reorder_batch_threshold must be 1 when query_len_support is "
f"SINGLE_ONLY, got {self.reorder_batch_threshold}"
)
def _build_fi_prefill_wrappers(self, prefill: FlashInferPrefillMetadata):
qo_indptr = prefill.query_start_loc
has_context = False
if prefill.chunked_context is not None:
chunked_context = prefill.chunked_context
has_context = True
if self._fi_prefill_main is None:
from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
self._fi_prefill_main = BatchPrefillWithRaggedKVCacheWrapper(
self._workspace_buffer, "NHD", backend="cutlass"
)
if has_context:
num_chunks = chunked_context.cu_seq_lens.shape[0]
# Allocate more prefill chunk wrappers if needed
if len(self._fi_prefill_chunks) < num_chunks:
from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
for _ in range(len(self._fi_prefill_chunks), num_chunks):
self._fi_prefill_chunks.append(
BatchPrefillWithRaggedKVCacheWrapper(
self._workspace_buffer, "NHD", backend="cutlass"
)
)
assert num_chunks <= len(self._fi_prefill_chunks)
# In MLA, the non-latent num_qo_heads == num_kv_heads
num_qo_heads = self.num_heads
num_kv_heads = num_qo_heads
# Sanity: Verify that num_kv_heads == 1 since it is latent space
assert self.kv_cache_spec.num_kv_heads == 1
# Get non-latent head_dim_qk and head_dim_vo
head_dim_qk = self.mla_dims.qk_nope_head_dim + self.mla_dims.qk_rope_head_dim
head_dim_vo = self.mla_dims.v_head_dim
# For main run, qo_indptr == kv_indptr
kv_indptr = qo_indptr.clone()
# Prepare main prefill
self._fi_prefill_main.plan(
qo_indptr=qo_indptr,
kv_indptr=kv_indptr,
num_qo_heads=num_qo_heads,
num_kv_heads=num_kv_heads,
head_dim_qk=head_dim_qk,
head_dim_vo=head_dim_vo,
causal=True, # This is main run
sm_scale=self._global_hyperparameters.sm_scale,
window_left=self._global_hyperparameters.window_left,
logits_soft_cap=self._global_hyperparameters.logits_soft_cap,
q_data_type=self.q_data_type,
o_data_type=prefill.output_dtype,
)
# Prepare context prefills
if has_context:
for i in range(num_chunks):
kv_indptr_chunk = chunked_context.cu_seq_lens[i]
self._fi_prefill_chunks[i].plan(
qo_indptr=qo_indptr,
kv_indptr=kv_indptr_chunk,
num_qo_heads=num_qo_heads,
num_kv_heads=num_kv_heads,
head_dim_qk=head_dim_qk,
head_dim_vo=head_dim_vo,
causal=False, # This is context run
sm_scale=self._global_hyperparameters.sm_scale,
window_left=self._global_hyperparameters.window_left,
logits_soft_cap=self._global_hyperparameters.logits_soft_cap,
q_data_type=self.q_data_type,
o_data_type=prefill.output_dtype,
)
prefill.prefill_main = self._fi_prefill_main
prefill.prefill_chunks = self._fi_prefill_chunks
def _build_decode(
self,
block_table_tensor: torch.Tensor,
@@ -1972,18 +1753,14 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
dtype=torch.int32,
)
chunked_context_metadata_cls = (
CudnnPrefillMetadata.ChunkedContextMetadata
if self._use_cudnn_prefill
else MLACommonPrefillMetadata.ChunkedContextMetadata
)
prefill_tokens_with_context = None
if num_prefills_with_context_cpu > 0:
prefill_tokens_with_context = prefill_query_start_loc_cpu[
num_prefills_with_context_cpu
].item()
_ChunkedMetadata = MLACommonPrefillMetadata.ChunkedContextMetadata
if self.dcp_world_size > 1:
chunked_context_metadata = chunked_context_metadata_cls(
chunked_context_metadata = _ChunkedMetadata(
cu_seq_lens=cu_seq_lens_cpu.to(device, non_blocking=True),
starts=local_chunk_starts.to(device, non_blocking=True),
seq_tot=padded_local_chunk_seq_lens.sum(dim=1).tolist(),
@@ -2004,7 +1781,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
prefill_tokens_with_context=prefill_tokens_with_context,
)
else:
chunked_context_metadata = chunked_context_metadata_cls(
chunked_context_metadata = _ChunkedMetadata(
cu_seq_lens=cu_seq_lens_cpu.to(device, non_blocking=True),
starts=chunk_starts.to(device, non_blocking=True),
seq_tot=chunk_seq_lens.sum(dim=1).tolist(),
@@ -2018,35 +1795,22 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
prefill_tokens_with_context=prefill_tokens_with_context,
)
if self._use_cudnn_prefill:
chunked_context_metadata.seq_lens = chunk_seq_lens
assert (
max(chunked_context_metadata.max_seq_lens)
<= self.chunked_prefill_workspace_size
)
prefill_metadata = self.prefill_metadata_cls(
prefill_metadata = MLACommonPrefillMetadata(
block_table=block_table_tensor[reqs_start:, ...],
query_start_loc=prefill_query_start_loc,
max_query_len=max_query_len,
chunked_context=chunked_context_metadata,
output_dtype=self.model_config.dtype,
q_data_type=self.q_data_type,
prefill_backend=self._prefill_backend,
)
if self._use_cudnn_prefill:
assert isinstance(prefill_metadata, CudnnPrefillMetadata)
prefill_metadata.query_seq_lens = (
prefill_query_start_loc[1:] - prefill_query_start_loc[:-1]
)
prefill_metadata.cudnn_workspace = self.cudnn_workspace
if self._use_trtllm_ragged_prefill:
prefill_metadata.query_seq_lens = (
prefill_query_start_loc[1:] - prefill_query_start_loc[:-1]
)
prefill_metadata.workspace_buffer = self._workspace_buffer
self._prefill_backend.prepare_metadata(prefill_metadata)
decode_metadata = None
if num_decodes > 0:
@@ -2091,10 +1855,6 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
decode=decode_metadata,
)
if self._use_fi_prefill and num_prefills > 0:
assert isinstance(attn_metadata.prefill, FlashInferPrefillMetadata)
self._build_fi_prefill_wrappers(attn_metadata.prefill)
return attn_metadata # type: ignore[return-value]
@@ -2240,308 +2000,12 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
and (self.qk_rope_head_dim == 64)
)
if use_trtllm_ragged_deepseek_prefill():
logger.info_once("Using TRT-LLM ragged DeepSeek prefill for MLA")
self._run_prefill_context_chunk = (
self._run_prefill_context_chunk_trtllm_ragged
)
self._run_prefill_new_tokens = self._run_prefill_new_tokens_trtllm_ragged
self._pad_v = False
elif use_flashinfer_prefill():
logger.info_once("Using FlashInfer prefill for MLA")
self._run_prefill_context_chunk = self._run_prefill_context_chunk_fi
self._run_prefill_new_tokens = self._run_prefill_new_tokens_fi
self._pad_v = False
elif use_cudnn_prefill():
logger.info_once("Using CUDNN prefill for MLA")
self._run_prefill_context_chunk = self._run_prefill_context_chunk_cudnn
self._run_prefill_new_tokens = self._run_prefill_new_tokens_cudnn
self._pad_v = False
else: # Use FlashAttention
if flash_attn_varlen_func is None:
raise RuntimeError(
"MLA attention requires FlashAttention but it is not "
"available. Please install flash_attn or use "
"--attention-backend ROCM_AITER_MLA."
)
logger.info_once("Using FlashAttention prefill for MLA")
self._run_prefill_context_chunk = self._run_prefill_context_chunk_fa
self._run_prefill_new_tokens = self._run_prefill_new_tokens_fa
# Handle the differences between the flash_attn_varlen from
# flash_attn and the one from vllm_flash_attn. The former is used on
# RoCM and the latter has an additional parameter to control
# FA2 vs FA3
self.flash_attn_varlen_func = flash_attn_varlen_func
self.vllm_flash_attn_version = get_flash_attn_version(
head_size=self.qk_head_dim
)
if self.vllm_flash_attn_version is not None:
self.flash_attn_varlen_func = functools.partial(
flash_attn_varlen_func, fa_version=self.vllm_flash_attn_version
)
# For MLA the v head dim is smaller than qk head dim so we pad out
# v with 0s to match the qk head dim for attention backends that do
# not support different headdims.
# FA3 on Hopper (SM90) and FA4 natively handle diff headdims.
device_capability = current_platform.get_device_capability()
self._pad_v = self.vllm_flash_attn_version is None or not (
(
self.vllm_flash_attn_version == 3
and device_capability is not None
and device_capability[0] == 9
)
or self.vllm_flash_attn_version == 4
)
self.dcp_world_size: int = -1
self.cp_kv_cache_interleave_size: int = (
get_current_vllm_config().parallel_config.cp_kv_cache_interleave_size
)
def _flash_attn_varlen_diff_headdims(
self, q, k, v, return_softmax_lse=False, softmax_scale=None, **kwargs
):
maybe_padded_v = v
if self._pad_v:
maybe_padded_v = torch.nn.functional.pad(
v, [0, q.shape[-1] - v.shape[-1]], value=0
)
if is_vllm_fa:
kwargs["return_softmax_lse"] = return_softmax_lse
else:
# ROCm leverages the upstream flash_attn, which takes a parameter
# called "return_attn_probs" instead of return_softmax_lse
kwargs["return_attn_probs"] = return_softmax_lse
if envs.VLLM_BATCH_INVARIANT:
kwargs["num_splits"] = 1
attn_out = self.flash_attn_varlen_func(
q=q,
k=k,
v=maybe_padded_v,
softmax_scale=softmax_scale,
**kwargs,
)
# Unpack the output if there is multiple results
lse = None
if isinstance(attn_out, tuple):
attn_out, lse = attn_out[0], attn_out[1]
# Remain consistent with old `flash_attn_varlen_func` where there
# is only one output tensor if `return_softmax_lse` is False.
if return_softmax_lse:
return attn_out, lse
return attn_out
def _run_prefill_new_tokens_fa(
self, prefill: MLACommonPrefillMetadata, q, k, v, return_softmax_lse
):
return self._flash_attn_varlen_diff_headdims(
q=q,
k=k,
v=v,
cu_seqlens_q=prefill.query_start_loc,
cu_seqlens_k=prefill.query_start_loc,
max_seqlen_q=prefill.max_query_len,
max_seqlen_k=prefill.max_query_len,
softmax_scale=self.scale,
causal=True,
return_softmax_lse=return_softmax_lse,
)
def _run_prefill_new_tokens_fi(
self, prefill: MLACommonPrefillMetadata, q, k, v, return_softmax_lse
):
assert isinstance(prefill, FlashInferPrefillMetadata)
assert prefill.prefill_main is not None
ret = prefill.prefill_main.run(
q=q,
k=k,
v=v,
return_lse=return_softmax_lse,
)
if isinstance(ret, tuple):
return ret[0], ret[1].transpose(0, 1).contiguous()
return ret
def _run_prefill_new_tokens_cudnn(
self, prefill: MLACommonPrefillMetadata, q, k, v, return_softmax_lse
):
assert isinstance(prefill, CudnnPrefillMetadata)
assert prefill.query_seq_lens is not None
from flashinfer.prefill import cudnn_batch_prefill_with_kv_cache
output, lse = cudnn_batch_prefill_with_kv_cache(
q=q,
k_cache=k,
v_cache=v,
scale=self.scale,
workspace_buffer=prefill.cudnn_workspace,
max_token_per_sequence=prefill.max_query_len,
max_sequence_kv=prefill.max_query_len,
actual_seq_lens_q=prefill.query_seq_lens.view(-1, 1, 1, 1),
actual_seq_lens_kv=prefill.query_seq_lens.view(-1, 1, 1, 1),
causal=True,
# Do not support False for now
return_lse=True,
# Indicates actual_seq_lens are on GPU or CPU.
is_cuda_graph_compatible=True,
)
if return_softmax_lse:
return output, lse
return output
def _run_prefill_context_chunk_fa(
self, prefill: MLACommonPrefillMetadata, chunk_idx: int, q, k, v
):
assert prefill.chunked_context is not None
return self._flash_attn_varlen_diff_headdims(
q=q,
k=k,
v=v,
cu_seqlens_q=prefill.query_start_loc,
cu_seqlens_k=prefill.chunked_context.cu_seq_lens[chunk_idx],
max_seqlen_q=prefill.max_query_len,
max_seqlen_k=prefill.chunked_context.max_seq_lens[chunk_idx],
softmax_scale=self.scale,
causal=False, # Context is unmasked
return_softmax_lse=True,
)
def _run_prefill_context_chunk_fi(
self, prefill: MLACommonPrefillMetadata, chunk_idx: int, q, k, v
):
assert isinstance(prefill, FlashInferPrefillMetadata)
attn_out, lse = prefill.prefill_chunks[chunk_idx].run(
q=q,
k=k,
v=v,
return_lse=True,
)
# Convert from (q_len, num_heads) to (num_heads, q_len)
return attn_out, lse.transpose(0, 1).contiguous()
def _run_prefill_context_chunk_cudnn(
self, prefill: MLACommonPrefillMetadata, chunk_idx: int, q, k, v
):
assert isinstance(prefill, CudnnPrefillMetadata)
assert prefill.chunked_context is not None
assert prefill.chunked_context.seq_lens[chunk_idx] is not None
assert prefill.query_seq_lens is not None
from flashinfer.prefill import cudnn_batch_prefill_with_kv_cache
return cudnn_batch_prefill_with_kv_cache(
q=q,
k_cache=k,
v_cache=v,
scale=self.scale,
workspace_buffer=prefill.cudnn_workspace,
max_token_per_sequence=prefill.max_query_len,
max_sequence_kv=prefill.chunked_context.max_seq_lens[chunk_idx],
actual_seq_lens_q=prefill.query_seq_lens.view(-1, 1, 1, 1),
actual_seq_lens_kv=prefill.chunked_context.seq_lens[chunk_idx].view(
-1, 1, 1, 1
),
causal=False,
return_lse=True,
# Indicates actual_seq_lens are on GPU or CPU.
is_cuda_graph_compatible=True,
)
def _run_prefill_new_tokens_trtllm_ragged(
self, prefill: MLACommonPrefillMetadata, q, k, v, return_softmax_lse
):
"""TRT-LLM ragged attention for new tokens (causal)."""
from flashinfer.prefill import trtllm_ragged_attention_deepseek
assert prefill.query_seq_lens is not None
assert prefill.workspace_buffer is not None
# allocate BF16 / FP16 output tensor for TRT-LLM ragged attention
out = torch.empty(
q.shape[0],
q.shape[1],
v.shape[2],
device=q.device,
dtype=prefill.output_dtype,
)
ret = trtllm_ragged_attention_deepseek(
query=q,
key=k,
value=v,
workspace_buffer=prefill.workspace_buffer,
seq_lens=prefill.query_seq_lens,
max_q_len=prefill.max_query_len,
max_kv_len=prefill.max_query_len,
bmm1_scale=self.scale,
bmm2_scale=1.0,
o_sf_scale=1.0,
batch_size=prefill.query_seq_lens.shape[0],
window_left=-1,
cum_seq_lens_q=prefill.query_start_loc,
cum_seq_lens_kv=prefill.query_start_loc,
enable_pdl=False,
is_causal=True,
return_lse=return_softmax_lse,
out=out,
)
if isinstance(ret, tuple):
# Convert from (q_len, num_heads) to (num_heads, q_len)
return ret[0], ret[1].transpose(0, 1).contiguous()
return ret
def _run_prefill_context_chunk_trtllm_ragged(
self, prefill: MLACommonPrefillMetadata, chunk_idx: int, q, k, v
):
"""TRT-LLM ragged attention for context chunks (non-causal)."""
from flashinfer.prefill import trtllm_ragged_attention_deepseek
assert prefill.chunked_context is not None
assert prefill.chunked_context.seq_lens[chunk_idx] is not None
assert prefill.workspace_buffer is not None
out = torch.empty(
q.shape[0],
q.shape[1],
v.shape[2],
device=q.device,
dtype=prefill.output_dtype,
)
attn_out, lse = trtllm_ragged_attention_deepseek(
query=q,
key=k,
value=v,
workspace_buffer=prefill.workspace_buffer,
seq_lens=prefill.chunked_context.seq_lens[chunk_idx],
max_q_len=prefill.max_query_len,
max_kv_len=prefill.chunked_context.max_seq_lens[chunk_idx],
bmm1_scale=self.scale,
bmm2_scale=1.0,
o_sf_scale=1.0,
batch_size=prefill.chunked_context.seq_lens[chunk_idx].shape[0],
window_left=-1,
cum_seq_lens_q=prefill.query_start_loc,
cum_seq_lens_kv=prefill.chunked_context.cu_seq_lens[chunk_idx],
enable_pdl=False,
is_causal=False,
return_lse=True,
out=out,
)
# Convert from (q_len, num_heads) to (num_heads, q_len)
return attn_out, lse.transpose(0, 1).contiguous()
def _concat_k_nope_k_pe(
self, k_nope: torch.Tensor, k_pe: torch.Tensor
) -> torch.Tensor:
@@ -2582,6 +2046,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
):
assert attn_metadata.prefill is not None
prefill_metadata = attn_metadata.prefill
assert prefill_metadata.prefill_backend is not None
assert prefill_metadata.chunked_context is not None
use_fp8_prefill = prefill_metadata.q_data_type == current_platform.fp8_dtype()
@@ -2649,12 +2114,13 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
k = self._concat_k_nope_k_pe(k_nope, k_pe)
attn_output, attn_softmax_lse = self._run_prefill_context_chunk(
prefill=prefill_metadata,
chunk_idx=i,
q=q,
k=k,
v=v,
attn_output, attn_softmax_lse = (
prefill_metadata.prefill_backend.run_prefill_context_chunk(
chunk_idx=i,
q=q,
k=k,
v=v,
)
)
if output is None:
@@ -2687,6 +2153,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
assert k_scale is None, "DCP not support scaled kvcache now."
assert attn_metadata.prefill is not None
prefill_metadata = attn_metadata.prefill
assert prefill_metadata.prefill_backend is not None
assert prefill_metadata.chunked_context is not None
assert prefill_metadata.chunked_context.padded_local_chunk_seq_lens is not None
assert prefill_metadata.chunked_context.local_context_lens_allranks is not None
@@ -2753,12 +2220,13 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
k_nope, v = kv_nope.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
k = self._concat_k_nope_k_pe(k_nope, k_pe)
attn_output, attn_softmax_lse = self._run_prefill_context_chunk(
prefill=prefill_metadata,
chunk_idx=i,
q=q,
k=k,
v=v,
attn_output, attn_softmax_lse = (
prefill_metadata.prefill_backend.run_prefill_context_chunk(
chunk_idx=i,
q=q,
k=k,
v=v,
)
)
if output is None:
@@ -2790,11 +2258,11 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
k_scale: torch.Tensor,
output: torch.Tensor,
) -> None:
# TODO (zyongye): Prefill function here
assert attn_metadata.prefill is not None
assert self.dcp_world_size != -1
prefill_metadata = attn_metadata.prefill
assert prefill_metadata.prefill_backend is not None
use_fp8_prefill = prefill_metadata.q_data_type == current_platform.fp8_dtype()
# Convert q to FP8 if FP8 prefill attention is enabled
@@ -2813,8 +2281,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
k = k.to(prefill_metadata.q_data_type)
v = v.to(prefill_metadata.q_data_type)
output_prefill = self._run_prefill_new_tokens(
prefill=prefill_metadata,
output_prefill = prefill_metadata.prefill_backend.run_prefill_new_tokens(
q=q,
k=k,
v=v,
@@ -2839,11 +2306,6 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
q, kv_c_and_k_pe_cache, attn_metadata, k_scale
)
# unpad if necessary
if self._pad_v:
context_output = context_output[..., : v.shape[-1]]
suffix_output = suffix_output[..., : v.shape[-1]]
output = output.view(-1, self.num_heads, self.v_head_dim)
merge_attn_states(
output=output,
@@ -2854,7 +2316,8 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
prefill_tokens_with_context=prefill_metadata.chunked_context.prefill_tokens_with_context,
)
else:
output_prefill = output_prefill[..., : v.shape[-1]].flatten(start_dim=-2)
assert isinstance(output_prefill, torch.Tensor)
output_prefill = output_prefill.flatten(start_dim=-2)
output.copy_(output_prefill)
@abstractmethod