class KimiK3MultiHeadLatentAttentionWrapper(MultiHeadLatentAttentionWrapper):
"""Kimi-K3 MLA wrapper with eager AITER q/kv RMSNorm and output gate
fusions and a fused decode Q-prep path."""
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
self._use_eager_qk_rmsnorm_fusion = bool(rocm_aiter_ops.is_enabled())
self._fused_qk_prep = self._fused_qk_prep_supported()
self._identity_rope: tuple[torch.Tensor, torch.Tensor] | None = None
self._qkv_gate_fused = isinstance(
self.fused_qkv_a_proj, KimiK3MergedQKVGateLinear
)
self._fused_q_heads = self.num_heads
if self._fused_qk_prep:
from vllm.v1.attention.backends.mla.rocm_aiter_mla import AiterMLAHelper
self._fused_q_heads = AiterMLAHelper.get_actual_mla_num_heads(
self.num_heads
)
# Positions fed to the fused kernel's identity RoPE; see _fused_decode.
# Allocated once here so every captured CUDA graph shares one zeroed
# buffer instead of capturing a memset per layer.
zero_positions = None
if self._fused_qk_prep:
vllm_config = get_current_vllm_config()
max_tokens = max(
vllm_config.scheduler_config.max_num_batched_tokens,
vllm_config.compilation_config.max_cudagraph_capture_size or 0,
)
zero_positions = torch.zeros(max_tokens, dtype=torch.int64)
self.register_buffer("_zero_positions", zero_positions, persistent=False)
# The fold reads q_b_proj.weight as plain bf16/fp16 rows, so a quantized
# q_b_proj keeps the unfused path.
self._fold_qb_uk = (
self._fused_qk_prep
and self.q_b_proj is not None
and isinstance(self.q_b_proj.quant_method, UnquantizedLinearMethod)
and self.rotary_emb is None
and not self.dcp_q_replicate
and self.indexer is None
)
self._w_fold: torch.Tensor | None = None
def _fused_qk_prep_supported(self) -> bool:
attn = self.mla_attn
if not (
rocm_aiter_ops.is_mla_enabled()
and attn.kv_cache_dtype.startswith("fp8")
and self.kv_lora_rank == _OPT_KV_LORA_RANK
and self.qk_rope_head_dim == _OPT_ROT_DIM
and self.kv_lora_rank * self.num_heads >= _OPT_MIN_SIZE
):
return False
# The fused kernel writes q_out only for slot >= 0, so padded rows would
# be garbage after a DCP query all-gather. PCP and HiSparse rewrite the
# KV inputs/slots before the standalone write, which this path bypasses.
if attn.impl.dcp_world_size > 1 or attn.use_pcp:
return False
if attn.hisparse_cache is not None or attn.q_pad_num_heads is not None:
return False
# Only the W_UK variants reproduced in _fused_decode are supported.
return (
attn.is_aiter_triton_fp4_bmm_enabled
or not attn.is_aiter_triton_fp8_bmm_enabled
)
def get_input_quant_key(self) -> QuantKey | None:
if self.indexer is not None:
return None
consumers = (
[self.fused_qkv_a_proj]
if self.q_lora_rank is not None
else [self.kv_a_proj_with_mqa, self.q_proj]
)
if self.g_proj is not None:
consumers.append(self.g_proj)
if any(layer is None for layer in consumers):
return None
keys = [get_input_quant_key(layer) for layer in consumers if layer is not None]
return keys[0] if all(key == keys[0] for key in keys) else None
def _normalize_q_kv(
self,
q_c: torch.Tensor,
kv_c: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
q_layernorm = cast(RMSNorm, self.q_a_layernorm)
kv_layernorm = cast(RMSNorm, self.kv_a_layernorm)
if self._use_eager_qk_rmsnorm_fusion and not torch.compiler.is_compiling():
return torch.ops.vllm.fused_mla_dual_rms_norm(
q_c,
q_layernorm.weight,
kv_c,
kv_layernorm.weight,
q_layernorm.variance_epsilon,
kv_layernorm.variance_epsilon,
)
return q_layernorm(q_c), kv_layernorm(kv_c)
@staticmethod
def _compute_w_fold(
w_qb: torch.Tensor,
w_uk: torch.Tensor,
num_heads: int,
qk_nope_head_dim: int,
qk_rope_head_dim: int,
) -> torch.Tensor:
"""W_fold = [W_UK @ W_qb_nope ; W_qb_pe]: one GEMM gives [ql_nope | q_pe].
``w_qb`` is the q_b_proj weight [N * (P + R), Lq], ``w_uk`` is [L, N, P].
Rows hold all heads' nope rows, then all pe rows, so the two output
halves of ``x @ W_fold.T`` are contiguous: [B, N * L] and [B, N * R]."""
N, P, R = num_heads, qk_nope_head_dim, qk_rope_head_dim
L = w_uk.shape[0]
w = w_qb.view(N, P + R, -1).float()
nope = torch.einsum("lnp,npk->nlk", w_uk.float(), w[:, :P]) # [N, L, Lq]
return (
torch.cat([nope.reshape(N * L, -1), w[:, P:].reshape(N * R, -1)])
.to(w_qb.dtype)
.contiguous()
)
@staticmethod
def _split_fold_output(
y: torch.Tensor,
num_heads: int,
kv_lora_rank: int,
qk_rope_head_dim: int,
) -> tuple[torch.Tensor, torch.Tensor]:
B, N, L = y.shape[0], num_heads, kv_lora_rank
return y[:, : N * L].view(B, N, L), y[:, N * L :].view(B, N, qk_rope_head_dim)
@torch.no_grad()
def _build_w_fold(self) -> None:
assert self.q_b_proj is not None
w_qb = self.q_b_proj.weight # [N * (P + R), q_lora_rank]
W_UK, _ = split_kv_b_proj(
self.mla_attn.kv_b_proj,
w_qb.dtype,
self.kv_lora_rank,
self.num_heads,
self.qk_nope_head_dim,
self.v_head_dim,
) # [L, N, P]
self._w_fold = self._compute_w_fold(
w_qb, W_UK, self.num_heads, self.qk_nope_head_dim, self.qk_rope_head_dim
)
def process_weights_after_loading(self) -> None:
if self._fold_qb_uk:
self._build_w_fold()
def _fused_decode(
self,
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
output_shape: tuple[int, int],
attn_metadata,
layer,
kv_cache: torch.Tensor,
slot_mapping: torch.Tensor,
) -> torch.Tensor:
"""Decode-only MLA through one fused AITER launch.
Replaces the fp8 KV-cache write plus the ``[ql_nope | q_pe]`` concat and
that query's static fp8 quant with a single
``fused_qk_rope_concat_and_cache_mla``. On an untouched upstream build
the concat and quant already arrive as one inductor-fused kernel, so
this collapses two launches into one.
"""
from aiter.ops.cache import fused_qk_rope_concat_and_cache_mla
attn = self.mla_attn
if q.dim() == 2:
from aiter.ops.triton.gemm.basic.gemm_a16w16 import gemm_a16w16
assert self._w_fold is not None
y = gemm_a16w16(q, self._w_fold)
ql_nope, q_pe = self._split_fold_output(
y, self.num_heads, self.kv_lora_rank, self.qk_rope_head_dim
)
else:
q_nope, q_pe = q.split(
[self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
)
q_nope_t = q_nope.transpose(0, 1)
if attn.is_aiter_triton_fp4_bmm_enabled:
from aiter.ops.triton.batched_gemm_a16wfp4 import (
batched_gemm_a16wfp4,
)
ql_nope = batched_gemm_a16wfp4(
q_nope_t,
attn.W_K,
attn.W_K_scale,
transpose_bm=True,
prequant=True,
y_scale=layer._q_scale,
)
else:
assert attn.W_UK_T is not None # guaranteed by _fused_qk_prep_supported
B, N = q_nope.shape[0], q_nope.shape[1]
L = attn.W_UK_T.shape[-1]
ql_nope = q_nope_t.new_empty((B, N, L))
torch.bmm(q_nope_t, attn.W_UK_T, out=ql_nope.transpose(0, 1))
if self._identity_rope is None:
# cos = 1, sin = 0 makes the kernel's RoPE the identity, which is
# what a NoPE model needs. A full-length cache would cost ~268 MB at
# K3's 1M max_position_embeddings, so it holds one row and the
# kernel is given all-zero positions below.
half = self.qk_rope_head_dim // 2
opts = {"dtype": kv_c_normed.dtype, "device": kv_c_normed.device}
self._identity_rope = (
torch.ones(1, half, **opts),
torch.zeros(1, half, **opts),
)
cos_cache, sin_cache = self._identity_rope
# An fp8 KV cache is allocated as uint8 and re-viewed as fp8 before use;
# the AITER kernel rejects the raw uint8 dtype outright.
fp8_dtype = current_platform.fp8_dtype()
if kv_cache.dtype != fp8_dtype:
kv_cache = kv_cache.view(fp8_dtype)
head_size = attn.kv_lora_rank + self.qk_rope_head_dim
num_tokens, num_heads = ql_nope.shape[:2]
q_out_padded = torch.empty(
(num_tokens, self._fused_q_heads, head_size),
dtype=fp8_dtype,
device=ql_nope.device,
)
q_out = q_out_padded[:, :num_heads]
# The kernel must not see the real positions: not every AITER variant
# clamps them into the one-row cos/sin cache. The head-grouped kernel
# (AITER v0.1.26) reads cos = sin = 0 past the cache at one head per
# block, zeroing q_pe and the cached k_pe, and faults at more heads per
# block. RoPE is the identity, so position 0 is exact for every token.
# The kernel takes its token count from slot_mapping.
slot_mapping = slot_mapping.flatten()
num_slots = slot_mapping.shape[0]
assert num_slots <= self._zero_positions.shape[0]
fused_qk_rope_concat_and_cache_mla(
ql_nope,
q_pe,
kv_c_normed,
k_pe.squeeze(1),
kv_cache.view(kv_cache.shape[0], -1, head_size),
q_out,
slot_mapping,
layer._k_scale,
layer._q_scale,
self._zero_positions[:num_slots],
cos_cache,
sin_cache,
is_neox=True,
is_nope_first=True,
)
# forward_mqa is MLA-specific and not on AttentionImpl; MLAAttention
# annotates its own call to it the same way.
attn_out, _ = attn.impl.forward_mqa(q_out, kv_cache, attn_metadata, attn) # type: ignore[attr-defined]
output = ql_nope.new_empty(output_shape)
attn._v_up_proj(attn_out, out=output)
return output
def forward(
self,
positions: torch.Tensor,
hidden_states: torch.Tensor | QuantizedActivation,
llama_4_scaling: torch.Tensor | None = None,
skip_o_proj: bool = False,
) -> torch.Tensor:
q_c = None
num_tokens = (
hidden_states.orig_shape[0]
if isinstance(hidden_states, QuantizedActivation)
else hidden_states.shape[0]
)
gate = None
if self.q_lora_rank is not None:
assert self.fused_qkv_a_proj is not None, (
"fused_qkv_a_proj is required when q_lora_rank is not None"
)
assert self.q_a_layernorm is not None, (
"q_a_layernorm is required when q_lora_rank is not None"
)
assert self.q_b_proj is not None, (
"q_b_proj is required when q_lora_rank is not None"
)
qkv_lora = self.fused_qkv_a_proj(hidden_states)[0]
if self._qkv_gate_fused:
qkv_lora, gate = qkv_lora.split(
[
self.q_lora_rank + self.kv_lora_rank + self.qk_rope_head_dim,
self.num_heads * self.v_head_dim,
],
dim=-1,
)
q_c, kv_lora = qkv_lora.split(
[self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
dim=-1,
)
kv_c, k_pe = kv_lora.split(
[self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
)
q_proj_input, kv_c_normed = self._normalize_q_kv(q_c, kv_c)
q_proj_layer = self.q_b_proj
else:
assert self.kv_a_proj_with_mqa is not None, (
"kv_a_proj_with_mqa is required when q_lora_rank is None"
)
assert self.q_proj is not None, (
"q_proj is required when q_lora_rank is None"
)
kv_lora = self.kv_a_proj_with_mqa(hidden_states)[0]
kv_c, k_pe = kv_lora.split(
[self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
)
kv_c_normed = self.kv_a_layernorm(kv_c)
q_proj_layer = self.q_proj
q_proj_input = hidden_states
# Add head dim of 1 to k_pe.
k_pe = k_pe.unsqueeze(1)
output_shape = (num_tokens, self.num_heads * self.v_head_dim)
attn_metadata = layer = kv_cache = slot_mapping = None
fuse = self._fused_qk_prep and not self.dcp_q_replicate
if fuse:
attn_metadata, layer, kv_cache, slot_mapping = get_attention_context(
self.mla_attn.layer_name
)
fuse = (
attn_metadata is not None
and slot_mapping is not None
and kv_cache.numel() > 0
and attn_metadata.num_actual_tokens == attn_metadata.num_decode_tokens
)
if fuse and self._fold_qb_uk and llama_4_scaling is None:
q = q_proj_input
else:
q = q_proj_layer(q_proj_input)[0]
heads = self.num_heads
if self.dcp_q_replicate:
heads *= q_proj_layer.group_size
q = q.view(-1, heads, self.qk_head_dim)
if self.rotary_emb is not None:
q[..., self.qk_nope_head_dim :], k_pe = self.rotary_emb(
positions, q[..., self.qk_nope_head_dim :], k_pe
)
if self.indexer and self.is_sparse and not self.skip_topk:
self.indexer(hidden_states, q_c, positions, self.indexer_rope_emb)
if llama_4_scaling is not None:
q *= llama_4_scaling
q_dcp_replicated = None
if self.dcp_q_replicate:
q_dcp_replicated, q = q, q_proj_layer._local_view(q)
if fuse:
attn_out = self._fused_decode(
q,
kv_c_normed,
k_pe,
output_shape,
attn_metadata,
layer,
kv_cache,
slot_mapping,
)
else:
attn_out = self.mla_attn(
q,
kv_c_normed,
k_pe,
output_shape=output_shape,
q_dcp_replicated=q_dcp_replicated,
)
if self.g_proj is not None:
gate = self.g_proj(hidden_states)[0]
if gate is not None:
if rocm_aiter_ops.is_enabled():
from aiter.ops.triton.fusions.fused_sigmoid_mul import fused_sigmoid_mul
attn_out = fused_sigmoid_mul(attn_out, gate)
else:
attn_out = attn_out * gate.sigmoid()
if skip_o_proj: # the specialized decode runs o_proj in its kernel
return attn_out
return self.o_proj(attn_out)[0]