Skip to content

Fix triton rotary kernel for training on context lengths > 65k #386

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

Fix triton rotary kernel for training on context lengths > 65k #386

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 8 additions & 4 deletions fast_llm/functional/triton/rotary.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,12 +17,14 @@ def triton_rotary_kernel(
num_heads: tl_constexpr,
rotary_block_size: tl_constexpr,
head_block_size: tl_constexpr,
seq_len: tl_constexpr,
backward: tl_constexpr,
):
# TODO: Int64 ptr if needed?
pid_0 = tl.program_id(axis=0)
pid_1 = tl.program_id(axis=1)
pid_2 = tl.program_id(axis=2)
pid_0_1 = tl.program_id(axis=0) # Folded (batch * seq) index
pid_2 = tl.program_id(axis=1) # Head index
pid_0 = pid_0_1 // seq_len
pid_1 = pid_0_1 - pid_0 * seq_len

offsets = tl.arange(0, rotary_block_size)
head_offsets = pid_2 * head_block_size + tl.arange(0, head_block_size)[:, None]
Expand Down Expand Up @@ -76,7 +78,8 @@ def triton_rotary_(
if head_block_size > num_heads:
head_block_size = triton.next_power_of_2(num_heads)

triton_rotary_kernel[(batch_size, seq_len, triton.cdiv(num_heads, head_block_size))](
# Folded the large y dim into the x dim as gridDim.x is 32 bit while gridDim.y and gridDim.z are 16 bit registers
triton_rotary_kernel[(batch_size * seq_len, triton.cdiv(num_heads, head_block_size))](
input_,
frequencies,
input_.stride(0),
Expand All @@ -86,6 +89,7 @@ def triton_rotary_(
num_heads,
rotary_block_size,
head_block_size,
seq_len,
backward, # noqa
)
return input_
Expand Down