[Kernel][Performance] Add Triton kernel for Qwen3-VL interleaved MRoPE (#25055)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2025-09-19 10:27:49 +00:00
committed by GitHub
parent a684c0124c
commit cea91a32f2
2 changed files with 85 additions and 43 deletions
@@ -15,7 +15,7 @@ from .common import apply_rotary_emb_dispatch
@triton.jit
def _triton_qwen2vl_mrope_forward(
def _triton_mrope_forward(
q_ptr,
k_ptr,
cos,
@@ -30,12 +30,14 @@ def _triton_qwen2vl_mrope_forward(
pad_hd: tl.constexpr,
mrope_section_t: tl.constexpr,
mrope_section_h: tl.constexpr,
mrope_section_w: tl.constexpr,
is_interleaved: tl.constexpr,
):
# Adapted from
# https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/qwen2vl_mrope.py
# This version supports flatten input tensors from vllm
# and supports cos and sin cache with shape (3, num_tokens, head_dim // 2)
# instead of (3, bsz, seq_len, head_dim)
# instead of (3, bsz, seq_len, head_dim), also supports interleaved rotary
pid = tl.program_id(0)
# locate start address
q_ptr = q_ptr + pid * (n_qh * hd)
@@ -47,9 +49,6 @@ def _triton_qwen2vl_mrope_forward(
# ####################################################################
# Note: cos and sin now have shape (3, num_tokens, head_dim // 2)
t_end = mrope_section_t
h_end = t_end + mrope_section_h
# Updated stride calculation for half head_dim
half_rd = rd // 2
t_cos = cos + pid * half_rd
@@ -61,9 +60,18 @@ def _triton_qwen2vl_mrope_forward(
# Updated offsets for half head_dim
cos_offsets = tl.arange(0, pad_hd // 2)
t_mask = cos_offsets < t_end
h_mask = (t_end <= cos_offsets) & (cos_offsets < h_end)
w_mask = (h_end <= cos_offsets) & (cos_offsets < half_rd)
if is_interleaved:
h_mask = (((cos_offsets % 3) == 1) &
(cos_offsets <= 3 * mrope_section_h))
w_mask = (((cos_offsets % 3) == 2) &
(cos_offsets <= 3 * mrope_section_w))
t_mask = ~(h_mask | w_mask)
else:
t_end = mrope_section_t
h_end = t_end + mrope_section_h
t_mask = cos_offsets < mrope_section_t
h_mask = (t_end <= cos_offsets) & (cos_offsets < h_end)
w_mask = (h_end <= cos_offsets) & (cos_offsets < half_rd)
t_cos_row = tl.load(t_cos + cos_offsets, mask=t_mask, other=0)
h_cos_row = tl.load(h_cos + cos_offsets, mask=h_mask, other=0)
@@ -131,6 +139,7 @@ def triton_mrope(
mrope_section: list[int],
head_size: int,
rotary_dim: int,
mrope_interleaved: bool,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Qwen2VL mrope kernel.
@@ -158,7 +167,7 @@ def triton_mrope(
cos = cos.contiguous()
sin = sin.contiguous()
_triton_qwen2vl_mrope_forward[(n_row, )](
_triton_mrope_forward[(n_row, )](
q,
k,
cos,
@@ -173,6 +182,8 @@ def triton_mrope(
pad_hd,
mrope_section[0],
mrope_section[1],
mrope_section[2],
mrope_interleaved,
)
return q, k
@@ -201,7 +212,7 @@ class MRotaryEmbedding(RotaryEmbedding):
is_neox_style: bool,
dtype: torch.dtype,
mrope_section: Optional[list[int]] = None,
mrope_interleaved: Optional[bool] = False,
mrope_interleaved: bool = False,
) -> None:
# In Qwen2.5-VL, the maximum index value is related to the duration of
# the input video. We enlarge max_position_embeddings to 4 times to get
@@ -282,10 +293,6 @@ class MRotaryEmbedding(RotaryEmbedding):
assert positions.ndim == 1 or positions.ndim == 2
assert key is not None
if self.mrope_interleaved:
# TODO: add triton implementation to support mrope-interleaved
return self.forward_native(positions, query, key)
num_tokens = positions.shape[-1]
cos_sin = self.cos_sin_cache[positions]
cos, sin = cos_sin.chunk(2, dim=-1)
@@ -302,6 +309,7 @@ class MRotaryEmbedding(RotaryEmbedding):
self.mrope_section,
self.head_size,
self.rotary_dim,
self.mrope_interleaved,
)
return q.reshape(query_shape), k.reshape(key_shape)