[ROCm][DSV4.1][Perf] Use AITER mHC for the delayed pre block (#56503)

Signed-off-by: Fangzhou Ai <fangzhou@semianalysis.ai>
Co-authored-by: Fangzhou Ai <fangzhou@semianalysis.ai>
This commit is contained in:
Fangzhou Ai
2026-09-12 00:35:48 +00:00
committed by GitHub
co-authored by Fangzhou Ai
parent 9d88ceb026
commit 46d2b23ac5
7 changed files with 620 additions and 19 deletions
@@ -12,6 +12,7 @@ __all__ = [
"mhc_fused_post_pre_cuda",
"hc_head_fused_kernel_cuda",
"mhc_pre_aiter",
"mhc_pre_delayed_aiter",
"mhc_post_aiter",
"mhc_fused_post_pre_aiter",
"hc_head_fused_aiter",
@@ -26,6 +27,7 @@ __all__ = [
"mhc_post_torch",
"mhc_fused_post_pre_torch",
"hc_head_fused_torch",
"mhc_pre_mix_triton",
"mhc_pre_triton",
"mhc_post_triton",
"mhc_fused_post_pre_triton",
+90
View File
@@ -104,6 +104,90 @@ def _mhc_pre_aiter_fake(
return post_mix, comb_mix, layer_input
def mhc_pre_delayed_aiter(
residual: torch.Tensor,
fn: torch.Tensor,
hc_scale: torch.Tensor,
hc_base: torch.Tensor,
rms_eps: float,
hc_pre_eps: float,
hc_sinkhorn_eps: float,
hc_post_mult_value: float,
sinkhorn_repeat: int,
pre_mix: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
"""mHC pre with the pre-mix carried in from the previous sublayer.
Matches ``mhc_pre_delayed_torch``: the stream collapse uses *pre_mix*
rather than the gate computed here, and that gate is returned as the
pre-mix for the next sublayer seam.
Args:
residual: shape (..., hc_mult, hidden_size), dtype torch.bfloat16
fn: shape (hc_mult3, hc_mult * hidden_size), dtype torch.float32
hc_scale: shape (3,), dtype torch.float32
hc_base: shape (hc_mult3,), dtype torch.float32
rms_eps: RMS normalization epsilon
hc_pre_eps: pre-mix epsilon
hc_sinkhorn_eps: sinkhorn epsilon
hc_post_mult_value: post-mix multiplier value
sinkhorn_repeat: number of sinkhorn iterations
pre_mix: shape (..., hc_mult) from the previous sublayer, or None at
model entry to select residual stream zero.
Returns:
post_mix: shape (..., hc_mult, 1), dtype torch.float32
comb_mix: shape (..., hc_mult, hc_mult), dtype torch.float32
layer_input: shape (..., hidden_size), dtype torch.bfloat16
next_pre_mix: shape (..., hc_mult), dtype torch.float32
"""
hidden_size = residual.shape[-1]
assert hidden_size % 256 == 0
from vllm._aiter_ops import rocm_aiter_ops
return rocm_aiter_ops.mhc_pre_delayed(
residual,
fn,
hc_scale,
hc_base,
rms_eps,
hc_pre_eps,
hc_sinkhorn_eps,
hc_post_mult_value,
sinkhorn_repeat,
pre_mix,
)
def _mhc_pre_delayed_aiter_fake(
residual: torch.Tensor,
fn: torch.Tensor,
hc_scale: torch.Tensor,
hc_base: torch.Tensor,
rms_eps: float,
hc_pre_eps: float,
hc_sinkhorn_eps: float,
hc_post_mult_value: float,
sinkhorn_repeat: int,
pre_mix: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
hc_mult = residual.shape[-2]
hidden_size = residual.shape[-1]
outer_shape = residual.shape[:-2]
return (
torch.empty(
*outer_shape, hc_mult, 1, dtype=torch.float32, device=residual.device
),
torch.empty(
*outer_shape, hc_mult, hc_mult, dtype=torch.float32, device=residual.device
),
torch.empty(
*outer_shape, hidden_size, dtype=torch.bfloat16, device=residual.device
),
torch.empty(*outer_shape, hc_mult, dtype=torch.float32, device=residual.device),
)
def mhc_post_aiter(
x: torch.Tensor,
residual: torch.Tensor,
@@ -226,6 +310,12 @@ direct_register_custom_op(
mutates_args=[],
fake_impl=_mhc_pre_aiter_fake,
)
direct_register_custom_op(
op_name="mhc_pre_delayed_aiter",
op_func=mhc_pre_delayed_aiter,
mutates_args=[],
fake_impl=_mhc_pre_delayed_aiter_fake,
)
direct_register_custom_op(
op_name="mhc_post_aiter",
op_func=mhc_post_aiter,
+131
View File
@@ -141,6 +141,129 @@ def _hc_collapse_triton_fake(x: Tensor, pre_mix: Tensor) -> Tensor:
return torch.empty(x.shape[0], x.shape[2], dtype=x.dtype, device=x.device)
@triton.jit
def _mhc_pre_mix_kernel(
gemm_ptr,
sqrsum_ptr,
hc_scale_ptr,
hc_base_ptr,
out_ptr,
splitk,
hc_mult: tl.constexpr,
gemm_stride_k,
gemm_stride_t,
gemm_stride_j,
sqrsum_stride_k,
sqrsum_stride_t,
out_stride_t,
out_stride_j,
inv_hc_hidden,
rms_eps,
hc_pre_eps,
SPLITK_BLOCK: tl.constexpr,
HC_BLOCK: tl.constexpr,
):
"""Recover the pre-mix gate from a split-k mHC pre GEMM output."""
token_idx = tl.program_id(0).to(tl.int64)
ks = tl.arange(0, SPLITK_BLOCK)
js = tl.arange(0, HC_BLOCK)
kmask = ks < splitk
jmask = js < hc_mult
# Only the first hc_mult GEMM columns feed the pre gate.
gemm = tl.load(
gemm_ptr
+ ks[:, None] * gemm_stride_k
+ token_idx * gemm_stride_t
+ js[None, :] * gemm_stride_j,
mask=kmask[:, None] & jmask[None, :],
other=0.0,
).to(tl.float32)
mixes = tl.sum(gemm, 0)
sqrsum = tl.load(
sqrsum_ptr + ks * sqrsum_stride_k + token_idx * sqrsum_stride_t,
mask=kmask,
other=0.0,
).to(tl.float32)
rstd = tl.rsqrt(tl.sum(sqrsum, 0) * inv_hc_hidden + rms_eps)
scale = tl.load(hc_scale_ptr).to(tl.float32)
base = tl.load(hc_base_ptr + js, mask=jmask, other=0.0).to(tl.float32)
pre = tl.sigmoid(mixes * rstd * scale + base) + hc_pre_eps
tl.store(out_ptr + token_idx * out_stride_t + js * out_stride_j, pre, mask=jmask)
def mhc_pre_mix_triton(
gemm_out: Tensor,
sqrsum: Tensor,
hc_scale: Tensor,
hc_base: Tensor,
hc_mult: int,
hc_hidden_size: int,
rms_eps: float,
hc_pre_eps: float,
) -> Tensor:
"""Pre-mix gate for the delayed mHC pre, from AITER's split-k GEMM output.
AITER's ``mhc_pre_big_fuse`` consumes the unreduced ``[splitk, tokens,
hc_mult3]`` GEMM output and the matching row square-sums, but only returns
the post and comb gates. The delayed formulation also needs the pre gate,
to carry into the next sublayer seam. It is the same slice of the same
numbers, so recover it here rather than repeating the projection.
"""
assert gemm_out.ndim == 3 and gemm_out.dtype == torch.float32
assert sqrsum.ndim == 2 and sqrsum.dtype == torch.float32
splitk, num_tokens = gemm_out.shape[0], gemm_out.shape[1]
assert sqrsum.shape == (splitk, num_tokens)
out = torch.empty(num_tokens, hc_mult, dtype=torch.float32, device=gemm_out.device)
if num_tokens == 0:
return out
_mhc_pre_mix_kernel[(num_tokens,)](
gemm_out,
sqrsum,
hc_scale,
hc_base,
out,
splitk,
hc_mult,
gemm_out.stride(0),
gemm_out.stride(1),
gemm_out.stride(2),
sqrsum.stride(0),
sqrsum.stride(1),
out.stride(0),
out.stride(1),
1.0 / hc_hidden_size,
rms_eps,
hc_pre_eps,
SPLITK_BLOCK=triton.next_power_of_2(splitk),
HC_BLOCK=triton.next_power_of_2(hc_mult),
num_warps=1,
# Match the separate FP32 multiply and add in the Torch reference.
enable_fp_fusion=False,
)
return out
def _mhc_pre_mix_triton_fake(
gemm_out: Tensor,
sqrsum: Tensor,
hc_scale: Tensor,
hc_base: Tensor,
hc_mult: int,
hc_hidden_size: int,
rms_eps: float,
hc_pre_eps: float,
) -> Tensor:
return torch.empty(
gemm_out.shape[1], hc_mult, dtype=torch.float32, device=gemm_out.device
)
def hc_head_reduce_triton_kernel(
x: torch.Tensor,
hc_fn: torch.Tensor,
@@ -216,3 +339,11 @@ direct_register_custom_op(
mutates_args=[],
fake_impl=_hc_collapse_triton_fake,
)
direct_register_custom_op(
op_name="mhc_pre_mix_triton",
op_func=mhc_pre_mix_triton,
mutates_args=[],
fake_impl=_mhc_pre_mix_triton_fake,
)