[diffusion] kernels: stop specializing on per-request sequence lengths (#41710)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Mick
2026-10-05 22:21:01 +08:00
committed by GitHub
co-authored by Mick Qian Claude Opus 5.5
parent a977e3b9d5
commit 150f568900
6 changed files with 202 additions and 10 deletions
@@ -171,7 +171,7 @@ def _push_vec4(
return mean, m2, cnt
@triton.jit
@triton.jit(do_not_specialize=["seq_len", "n_rows"])
def _layernorm_modulate_kernel(
y_ptr,
y_q_ptr,
@@ -353,7 +353,7 @@ def _qk_ln_head_one(
tl.store(dst + row_base[:, None] + cols2[None, :], y, mask=out_mask)
@triton.jit
@triton.jit(do_not_specialize=["n_rows"])
def _qk_ln_head_kernel(
q_out_ptr,
k_out_ptr,
@@ -5,7 +5,7 @@ import triton
import triton.language as tl
@triton.jit
@triton.jit(do_not_specialize=["rows"])
def _pack_qkv_destination_major_kernel(
output_ptr,
q_ptr,
@@ -22,7 +22,10 @@ import triton.language as tl # type: ignore
_MAX_INT32 = 2**31 - 1
@triton.jit
# The per-request sizes only index. The channel count, ``total`` and the
# strides stay specialized for the channel vectorization, and the per-layer
# padding and cache length for their constant folding (1.5% of kernel time).
@triton.jit(do_not_specialize=["T", "H", "W", "out_t", "out_h", "out_w"])
def _cat_pad_cl3d_kernel(
x_ptr,
cache_ptr,
@@ -207,7 +210,9 @@ def cat_pad_channels_last_3d(
return out
@triton.jit
# ``out_w`` stays specialized: it is the contiguous dimension of the
# channels-first traversal.
@triton.jit(do_not_specialize=["out_t", "out_h", "t_offset"])
def _dup_up3d_add_kernel(
main_ptr,
src_ptr,
@@ -71,7 +71,7 @@ def try_fused_scaled_residual_add_exact(
return output
@triton.jit
@triton.jit(do_not_specialize=["seq_len", "stride_i_b"])
def _fused_layernorm_scale_shift_gate_select01_kernel(
output_ptr,
gate_out_ptr,
@@ -165,7 +165,7 @@ def _fused_layernorm_scale_shift_gate_select01_kernel(
tl.store(gate_row_ptr + cols, gate, mask=mask)
@triton.jit
@triton.jit(do_not_specialize=["seq_len", "stride_i_b"])
def _fused_residual_layernorm_scale_shift_gate_select01_kernel(
output_ptr,
residual_out_ptr,
@@ -273,7 +273,7 @@ def _fused_residual_layernorm_scale_shift_gate_select01_kernel(
tl.store(gate_row_ptr + cols, gate, mask=mask)
@triton.jit
@triton.jit(do_not_specialize=["seq_len", "num_frames", "frame_seqlen"])
def _fused_scale_shift_4d_kernel(
output_ptr,
normalized_ptr,
@@ -322,7 +322,7 @@ def _fused_scale_shift_4d_kernel(
tl.store(out_ptrs, output, mask=mask)
@triton.jit
@triton.jit(do_not_specialize=["B", "L"])
def fuse_scale_shift_kernel_blc_opt(
x_ptr,
shift_ptr,
@@ -27,7 +27,7 @@ import triton.language as tl # type: ignore
from sglang.srt.utils.custom_op import register_custom_op
@triton.jit
@triton.jit(do_not_specialize=["rows"])
def _temb_table_slices_kernel(
out_ptr,
temb_ptr,