class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
"""ROCm sparse MLA attention layer for DeepSeek V4."""
backend_cls = DeepseekV4ROCMAiterMLASparseBackend
def __init__(self, *args, **kwargs):
vllm_config = args[0] if args else kwargs["vllm_config"]
super().__init__(*args, **kwargs)
self._has_kv_transfer = vllm_config.kv_transfer_config is not None
# Block scale for the preshuffled weight; None = not preshuffled.
self._wqa_wkv_scale: torch.Tensor | None = None
self._wo_b_scale: torch.Tensor | None = None
@classmethod
def get_padded_num_q_heads(cls, num_heads: int) -> int:
return num_heads
def prepare_attn_preshuffle(self) -> None:
from vllm._aiter_ops import rocm_aiter_ops
if not rocm_aiter_ops.is_enabled():
return
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
_upcast_e8m0_to_fp32,
)
from vllm.model_executor.utils import replace_parameter
def _prep(linear) -> torch.Tensor | None:
w = getattr(linear, "weight", None)
if w is None or w.dim() != 2:
return None
# K % 128 (group-128 quant) and N % 16 (shuffle_weight) must hold.
if w.shape[-1] % 128 != 0 or w.shape[0] % 16 != 0:
return None
ws = getattr(linear, "weight_scale_inv", None) # per-block scale
if ws is None:
return None
if ws.dtype == torch.float8_e8m0fnu:
ws = _upcast_e8m0_to_fp32(ws).contiguous()
# Shuffle the weight in place (single weight, no unshuffled copy).
replace_parameter(
linear,
"weight",
rocm_aiter_ops.shuffle_weight(w.data, layout=(16, 16)),
)
return ws
self._wqa_wkv_scale = _prep(self.fused_wqa_wkv)
self._wo_b_scale = _prep(self.wo_b)
def _bpre_attn_gemm(
self,
weight: torch.Tensor,
scale: torch.Tensor,
x: torch.Tensor,
reduce_tp: bool,
) -> torch.Tensor:
from vllm._aiter_ops import rocm_aiter_ops
x_fp8, x_scale = rocm_aiter_ops.group_fp8_quant(x, transpose_scale=True)
out = rocm_aiter_ops.gemm_a8w8_blockscale_bpreshuffle(
x_fp8, weight, x_scale, scale, output_dtype=x.dtype
)
if reduce_tp and get_tensor_model_parallel_world_size() > 1:
out = tensor_model_parallel_all_reduce(out)
return out
def _fused_wqa_wkv_gemm(self, hidden_states: torch.Tensor) -> torch.Tensor:
if self._wqa_wkv_scale is not None and hidden_states.dim() == 2:
return self._bpre_attn_gemm(
self.fused_wqa_wkv.weight, self._wqa_wkv_scale, hidden_states, False
)
return super()._fused_wqa_wkv_gemm(hidden_states)
def _o_proj(self, o: torch.Tensor, positions: torch.Tensor) -> torch.Tensor:
# ROCm BF16 reference wo_a path (inverse RoPE + einsum) + wo_b.
z = rocm_inv_rope_einsum(
self.rotary_emb,
o,
positions,
self.rope_head_dim,
self.n_local_groups,
self.o_lora_rank,
self.wo_a,
)
zf = z.flatten(1)
if self._wo_b_scale is not None and zf.dim() == 2:
return self._bpre_attn_gemm(self.wo_b.weight, self._wo_b_scale, zf, True)
return self.wo_b(zf)
def forward_mqa(
self,
q: torch.Tensor,
kv: torch.Tensor,
positions: torch.Tensor,
output: torch.Tensor,
) -> None:
assert output.shape == q.shape, (
f"output buffer shape {output.shape} must match q shape {q.shape}"
)
assert output.dtype == q.dtype, (
f"output buffer dtype {output.dtype} must match q dtype {q.dtype}"
)
forward_context = get_forward_context()
attn_metadata = forward_context.attn_metadata
if attn_metadata is None:
# Warmup dummy run: no real metadata. Reserve the same bf16
# gather workspace _forward_prefill would; the dequantize / topk
# / sparse_fwd kernels are skipped this step.
swa_only = self.compress_ratio <= 1
N = (
0
if swa_only
else (self.max_model_len + self.compress_ratio - 1)
// self.compress_ratio
)
M = N + self.window_size + self.max_num_batched_tokens
current_workspace_manager().get_simultaneous(
((self.PREFILL_CHUNK_SIZE, M, q.shape[-1]), torch.bfloat16),
)
output.zero_()
return
assert isinstance(attn_metadata, dict)
rocm_metadata = cast(
DeepseekV4ROCMAiterMLASparseMetadata | None,
attn_metadata.get(self.prefix),
)
swa_metadata = cast(
DeepseekV4ROCMAiterSparseSWAMetadata | None,
attn_metadata.get(self.swa_cache_layer.prefix),
)
assert swa_metadata is not None
swa_only = self.compress_ratio <= 1
self_kv_cache = self.kv_cache if not swa_only else None
swa_kv_cache = self.swa_cache_layer.kv_cache
num_decodes = swa_metadata.num_decodes
num_prefills = swa_metadata.num_prefills
num_decode_tokens = swa_metadata.num_decode_tokens
if num_prefills > 0:
self._forward_prefill(
q=q[num_decode_tokens:],
positions=positions[num_decode_tokens:],
compressed_k_cache=self_kv_cache,
swa_k_cache=swa_kv_cache,
output=output[num_decode_tokens:],
attn_metadata=rocm_metadata,
swa_metadata=swa_metadata,
)
if num_decodes > 0:
self._forward_decode(
q=q[:num_decode_tokens],
kv_cache=self_kv_cache,
swa_metadata=swa_metadata,
attn_metadata=rocm_metadata,
swa_only=swa_only,
output=output[:num_decode_tokens],
adaptive_splits=(
_ON_GFX950
and not swa_only
and self.compress_ratio == 128
and rocm_metadata is not None
and rocm_metadata.for_cudagraph_capture
),
)
def _forward_decode(
self,
q: torch.Tensor,
kv_cache: torch.Tensor | None,
swa_metadata: DeepseekV4ROCMAiterSparseSWAMetadata,
attn_metadata: DeepseekV4ROCMAiterMLASparseMetadata | None,
swa_only: bool,
output: torch.Tensor,
adaptive_splits: bool,
) -> None:
num_decodes = swa_metadata.num_decodes
num_decode_tokens = swa_metadata.num_decode_tokens
topk_indices = None
topk_lens = None
topk_ragged_indices = None
topk_ragged_indptr = None
if not swa_only:
assert attn_metadata is not None
assert swa_metadata.is_valid_token is not None
block_size = attn_metadata.block_size // self.compress_ratio
is_valid = swa_metadata.is_valid_token[:num_decode_tokens]
if self.compress_ratio == 4:
assert self.topk_indices_buffer is not None
(
topk_ragged_indices,
topk_ragged_indptr,
topk_lens,
) = compute_global_topk_ragged_indices_and_indptr(
self.topk_indices_buffer[:num_decode_tokens],
swa_metadata.token_to_req_indices,
attn_metadata.block_table[:num_decodes],
block_size,
is_valid,
)
else:
topk_indices = attn_metadata.c128a_global_decode_topk_indices
topk_lens = attn_metadata.c128a_decode_topk_lens
topk_ragged_indices = attn_metadata.c128a_decode_topk_ragged_indices
topk_ragged_indptr = attn_metadata.c128a_decode_topk_ragged_indptr
rocm_sparse_attn_decode(
q=q,
kv_cache=kv_cache,
swa_k_cache=self.swa_cache_layer.kv_cache,
swa_only=swa_only,
topk_indices=topk_indices,
topk_lens=topk_lens,
swa_indices=swa_metadata.decode_swa_indices,
swa_lens=swa_metadata.decode_swa_lens,
swa_ragged_indices=swa_metadata.decode_swa_ragged_indices,
swa_ragged_indptr=swa_metadata.decode_swa_ragged_indptr,
topk_ragged_indices=topk_ragged_indices,
topk_ragged_indptr=topk_ragged_indptr,
attn_sink=self.attn_sink,
scale=self.scale,
head_dim=self.head_dim,
nope_head_dim=self.nope_head_dim,
rope_head_dim=self.rope_head_dim,
output=output,
adaptive_splits=adaptive_splits,
extra_cache_nan_free=_trust_dsv4_extra_cache_nan_free(
self.kv_cache_dtype,
self._has_kv_transfer,
not swa_only and kv_cache is not None,
),
)
def _forward_prefill(
self,
q: torch.Tensor,
positions: torch.Tensor,
compressed_k_cache: torch.Tensor | None,
swa_k_cache: torch.Tensor,
output: torch.Tensor,
attn_metadata: DeepseekV4ROCMAiterMLASparseMetadata | None,
swa_metadata: DeepseekV4ROCMAiterSparseSWAMetadata,
) -> None:
swa_only = attn_metadata is None
num_prefills = swa_metadata.num_prefills
num_prefill_tokens = swa_metadata.num_prefill_tokens
num_decodes = swa_metadata.num_decodes
num_decode_tokens = swa_metadata.num_decode_tokens
seq_lens = swa_metadata.prefill_seq_lens
gather_lens = swa_metadata.prefill_gather_lens
assert seq_lens is not None
assert gather_lens is not None
query_start_loc_cpu = swa_metadata.query_start_loc_cpu
query_start_loc = swa_metadata.query_start_loc
assert query_start_loc_cpu is not None
assert query_start_loc is not None
prefill_token_base = query_start_loc_cpu[num_decodes]
if not swa_only:
if self.compress_ratio == 4:
assert self.topk_indices_buffer is not None
topk_indices = self.topk_indices_buffer[num_decode_tokens:]
topk_indices = topk_indices[:num_prefill_tokens]
else:
assert attn_metadata is not None
topk_indices = attn_metadata.c128a_prefill_topk_indices
assert topk_indices is not None
top_k = topk_indices.shape[-1]
N = (self.max_model_len + self.compress_ratio - 1) // self.compress_ratio
else:
assert self.topk_indices_buffer is not None
topk_indices = self.topk_indices_buffer[num_decode_tokens:]
top_k = 0
N = 0
M = N + self.window_size + self.max_num_batched_tokens
num_chunks = (num_prefills + self.PREFILL_CHUNK_SIZE - 1) // (
self.PREFILL_CHUNK_SIZE
)
workspace_manager = current_workspace_manager()
kv = workspace_manager.get_simultaneous(
((self.PREFILL_CHUNK_SIZE, M, q.shape[-1]), torch.bfloat16),
)[0]
for chunk_idx in range(num_chunks):
chunk_start = chunk_idx * self.PREFILL_CHUNK_SIZE
chunk_end = min(chunk_start + self.PREFILL_CHUNK_SIZE, num_prefills)
chunk_size = chunk_end - chunk_start
if not swa_only:
assert attn_metadata is not None
assert compressed_k_cache is not None
block_table = attn_metadata.block_table[num_decodes:]
# compressed_k_cache is OCP on every platform (Triton encoder).
dequantize_and_gather_k_cache(
kv[:chunk_size],
compressed_k_cache,
seq_lens=seq_lens[chunk_start:chunk_end] // self.compress_ratio,
gather_lens=None,
block_table=block_table[chunk_start:chunk_end],
block_size=attn_metadata.block_size // self.compress_ratio,
offset=0,
use_fnuz=False,
)
swa_block_table = swa_metadata.block_table[num_decodes:]
dequantize_and_gather_k_cache(
kv[:chunk_size],
swa_k_cache,
seq_lens=seq_lens[chunk_start:chunk_end],
gather_lens=gather_lens[chunk_start:chunk_end],
block_table=swa_block_table[chunk_start:chunk_end],
block_size=swa_metadata.block_size,
offset=N,
use_fnuz=current_platform.is_fp8_fnuz(),
)
query_start = (
query_start_loc_cpu[num_decodes + chunk_start] - prefill_token_base
)
query_end = (
query_start_loc_cpu[num_decodes + chunk_end] - prefill_token_base
)
combined_indices, combined_lens = combine_topk_swa_indices(
topk_indices[query_start:query_end],
query_start_loc[
num_decodes + chunk_start : num_decodes + chunk_end + 1
],
seq_lens[chunk_start:chunk_end],
gather_lens[chunk_start:chunk_end],
self.window_size,
self.compress_ratio,
top_k,
M,
N,
)
rocm_sparse_attn_prefill(
q=q[query_start:query_end],
kv=kv.view(-1, 1, q.shape[-1]),
indices=combined_indices,
topk_length=combined_lens,
scale=self.scale,
head_dim=self.head_dim,
nope_head_dim=self.nope_head_dim,
rope_head_dim=self.rope_head_dim,
attn_sink=self.attn_sink,
output=output[query_start:query_end],
)