Skip to content

[Triton] Fix triton launch grids under multi gpu scenarios #1003

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

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

[Triton] Fix triton launch grids under multi gpu scenarios #1003

Show file tree
Hide file tree
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
70 changes: 36 additions & 34 deletions mamba_ssm/ops/triton/mamba3/angle_dt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,22 +194,23 @@ def angle_dt_fwd(
else:
grid = (nheads, batch)

angle_dt_fwd_kernel[grid](
out, output_state,
angle, dt, init_state, cu_seqlens,
out.stride(0), out.stride(1), out.stride(2), out.stride(3),
stride_output_state[0], stride_output_state[1], stride_output_state[2],
angle.stride(0), angle.stride(1), angle.stride(2), angle.stride(3),
dt.stride(0), dt.stride(1), dt.stride(2),
stride_init[0], stride_init[1], stride_init[2],
stride_cu_seqlen,
seqlen, dim,
CHUNK_SIZE=chunk_size,
BLOCK_D=BLOCK_D,
HAS_INIT_STATE=HAS_INIT_STATE,
RETURN_OUTPUT_STATE=return_output_state,
IS_VARLEN=is_varlen,
)
with torch.cuda.device(out.device.index):
angle_dt_fwd_kernel[grid](
out, output_state,
angle, dt, init_state, cu_seqlens,
out.stride(0), out.stride(1), out.stride(2), out.stride(3),
stride_output_state[0], stride_output_state[1], stride_output_state[2],
angle.stride(0), angle.stride(1), angle.stride(2), angle.stride(3),
dt.stride(0), dt.stride(1), dt.stride(2),
stride_init[0], stride_init[1], stride_init[2],
stride_cu_seqlen,
seqlen, dim,
CHUNK_SIZE=chunk_size,
BLOCK_D=BLOCK_D,
HAS_INIT_STATE=HAS_INIT_STATE,
RETURN_OUTPUT_STATE=return_output_state,
IS_VARLEN=is_varlen,
)

if return_output_state:
return out, output_state
Expand Down Expand Up @@ -411,22 +412,23 @@ def angle_dt_bwd(
else:
grid = (nheads, batch)

angle_dt_bwd_kernel[grid](
grad_angle, grad_dt, grad_init_state if has_init_state else grad_init_dummy,
grad_out, grad_output_state, angle, dt, cu_seqlens,
grad_angle.stride(0), grad_angle.stride(1), grad_angle.stride(2), grad_angle.stride(3),
grad_dt.stride(0), grad_dt.stride(1), grad_dt.stride(2),
stride_grad_init[0], stride_grad_init[1], stride_grad_init[2],
grad_out.stride(0), grad_out.stride(1), grad_out.stride(2), grad_out.stride(3),
stride_grad_output_state[0], stride_grad_output_state[1], stride_grad_output_state[2],
angle.stride(0), angle.stride(1), angle.stride(2), angle.stride(3),
dt.stride(0), dt.stride(1), dt.stride(2),
stride_cu_seqlen,
seqlen, dim,
CHUNK_SIZE=chunk_size,
BLOCK_D=BLOCK_D,
HAS_INIT_STATE=has_init_state,
HAS_GRAD_OUTPUT_STATE=HAS_GRAD_OUTPUT_STATE,
IS_VARLEN=is_varlen,
)
with torch.cuda.device(grad_angle.device.index):
angle_dt_bwd_kernel[grid](
grad_angle, grad_dt, grad_init_state if has_init_state else grad_init_dummy,
grad_out, grad_output_state, angle, dt, cu_seqlens,
grad_angle.stride(0), grad_angle.stride(1), grad_angle.stride(2), grad_angle.stride(3),
grad_dt.stride(0), grad_dt.stride(1), grad_dt.stride(2),
stride_grad_init[0], stride_grad_init[1], stride_grad_init[2],
grad_out.stride(0), grad_out.stride(1), grad_out.stride(2), grad_out.stride(3),
stride_grad_output_state[0], stride_grad_output_state[1], stride_grad_output_state[2],
angle.stride(0), angle.stride(1), angle.stride(2), angle.stride(3),
dt.stride(0), dt.stride(1), dt.stride(2),
stride_cu_seqlen,
seqlen, dim,
CHUNK_SIZE=chunk_size,
BLOCK_D=BLOCK_D,
HAS_INIT_STATE=has_init_state,
HAS_GRAD_OUTPUT_STATE=HAS_GRAD_OUTPUT_STATE,
IS_VARLEN=is_varlen,
)
return grad_angle, grad_dt, grad_init_state
39 changes: 20 additions & 19 deletions mamba_ssm/ops/triton/mamba3/grouped_head_reduction.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,25 +189,26 @@ def reduce_grouped_qk_grads_and_bias_triton(

n_blocks = triton.cdiv(N, block_n)
grid = (num_row_blocks, R, num_qk_groups * n_blocks)
_reduce_grouped_qk_grads_and_bias_partial_kernel[grid](
dq_raw,
dk_raw,
dq_out,
dk_out,
dq_bias_partial,
dk_bias_partial,
total_rows,
R,
H,
num_qk_groups,
N,
group_size,
n_blocks,
block_m,
block_n,
store_grouped,
num_warps=8,
)
with torch.cuda.device(dq_raw.device.index):
_reduce_grouped_qk_grads_and_bias_partial_kernel[grid](
dq_raw,
dk_raw,
dq_out,
dk_out,
dq_bias_partial,
dk_bias_partial,
total_rows,
R,
H,
num_qk_groups,
N,
group_size,
n_blocks,
block_m,
block_n,
store_grouped,
num_warps=8,
)

dq_bias = torch.empty((H, R, N), dtype=torch.float32, device=dq_raw.device)
dk_bias = torch.empty_like(dq_bias)
Expand Down
151 changes: 78 additions & 73 deletions mamba_ssm/ops/triton/mamba3/mamba3_mimo_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -593,18 +593,19 @@ def bwd_dtrap_ddt_triton(
dtrap = torch.zeros_like(trap)

grid = (B, H, nchunks)
bwd_dtrap_ddt_kernel[grid](
trap, dt, dfactor, dgamma_diag,
ddt, dtrap,
trap.stride(0), trap.stride(1), trap.stride(2),
dt.stride(0), dt.stride(1), dt.stride(2),
dfactor.stride(0), dfactor.stride(1), dfactor.stride(2),
dgamma_diag.stride(0), dgamma_diag.stride(1), dgamma_diag.stride(2),
ddt.stride(0), ddt.stride(1), ddt.stride(2),
dtrap.stride(0), dtrap.stride(1), dtrap.stride(2),
S,
chunk_size,
)
with torch.cuda.device(trap.device.index):
bwd_dtrap_ddt_kernel[grid](
trap, dt, dfactor, dgamma_diag,
ddt, dtrap,
trap.stride(0), trap.stride(1), trap.stride(2),
dt.stride(0), dt.stride(1), dt.stride(2),
dfactor.stride(0), dfactor.stride(1), dfactor.stride(2),
dgamma_diag.stride(0), dgamma_diag.stride(1), dgamma_diag.stride(2),
ddt.stride(0), ddt.stride(1), ddt.stride(2),
dtrap.stride(0), dtrap.stride(1), dtrap.stride(2),
S,
chunk_size,
)
return ddt, dtrap

def compute_dacs_segsum_triton_varlen(
Expand Down Expand Up @@ -672,18 +673,19 @@ def compute_dacs_segsum_triton_varlen(
state_chunk_in_seq[chunk_start:chunk_end] = torch.arange(n, dtype=torch.int32, device=da.device)

grid = (B, H, nchunks)
dacs_segsum_kernel_varlen[grid](
da, da_cs, da_cs_rev, segsum, cu_seqlens, state_seq_mapping, state_chunk_in_seq,
da.stride(0), da.stride(1), da.stride(2),
da_cs.stride(0), da_cs.stride(1), da_cs.stride(2),
da_cs_rev.stride(0), da_cs_rev.stride(1), da_cs_rev.stride(2),
segsum.stride(0), segsum.stride(1), segsum.stride(2),
segsum.stride(3), segsum.stride(4),
cu_seqlens.stride(0), state_seq_mapping.stride(0), state_chunk_in_seq.stride(0),
S,
num_sequences,
chunk_size,
)
with torch.cuda.device(da.device.index):
dacs_segsum_kernel_varlen[grid](
da, da_cs, da_cs_rev, segsum, cu_seqlens, state_seq_mapping, state_chunk_in_seq,
da.stride(0), da.stride(1), da.stride(2),
da_cs.stride(0), da_cs.stride(1), da_cs.stride(2),
da_cs_rev.stride(0), da_cs_rev.stride(1), da_cs_rev.stride(2),
segsum.stride(0), segsum.stride(1), segsum.stride(2),
segsum.stride(3), segsum.stride(4),
cu_seqlens.stride(0), state_seq_mapping.stride(0), state_chunk_in_seq.stride(0),
S,
num_sequences,
chunk_size,
)

return da_cs, da_cs_rev, segsum

Expand All @@ -708,16 +710,17 @@ def compute_dacs_segsum_triton(
segsum = torch.empty(B, H, nchunks, chunk_size, chunk_size, device=da.device, dtype=da.dtype)

grid = (B, H, nchunks)
dacs_segsum_kernel[grid](
da, da_cs, da_cs_rev, segsum,
da.stride(0), da.stride(1), da.stride(2),
da_cs.stride(0), da_cs.stride(1), da_cs.stride(2),
da_cs_rev.stride(0), da_cs_rev.stride(1), da_cs_rev.stride(2),
segsum.stride(0), segsum.stride(1), segsum.stride(2),
segsum.stride(3), segsum.stride(4),
S,
chunk_size,
)
with torch.cuda.device(da.device.index):
dacs_segsum_kernel[grid](
da, da_cs, da_cs_rev, segsum,
da.stride(0), da.stride(1), da.stride(2),
da_cs.stride(0), da_cs.stride(1), da_cs.stride(2),
da_cs_rev.stride(0), da_cs_rev.stride(1), da_cs_rev.stride(2),
segsum.stride(0), segsum.stride(1), segsum.stride(2),
segsum.stride(3), segsum.stride(4),
S,
chunk_size,
)

return da_cs, da_cs_rev, segsum

Expand Down Expand Up @@ -1336,31 +1339,32 @@ def bwd_dadt_fused_triton_varlen(
dadt_out = torch.zeros(B, H, S, device=ddA_cs.device, dtype=torch.float32)
grid = (B, H, nchunks_global)

bwd_dadt_cumsum_fused_kernel_varlen[grid](
ddA_cs, ddA_cs_rev, dA_cs, dA_cs_rev, dadt_out,
cu_seqlens, state_seq_mapping, state_chunk_in_seq,
ddA_cs.stride(0), ddA_cs.stride(1), ddA_cs.stride(2),
cu_seqlens.stride(0),
state_seq_mapping.stride(0),
state_chunk_in_seq.stride(0),
S=S,
CHUNK_SIZE=chunk_size,
)
with torch.cuda.device(dSSdA.device.index):
bwd_dadt_cumsum_fused_kernel_varlen[grid](
ddA_cs, ddA_cs_rev, dA_cs, dA_cs_rev, dadt_out,
cu_seqlens, state_seq_mapping, state_chunk_in_seq,
ddA_cs.stride(0), ddA_cs.stride(1), ddA_cs.stride(2),
cu_seqlens.stride(0),
state_seq_mapping.stride(0),
state_chunk_in_seq.stride(0),
S=S,
CHUNK_SIZE=chunk_size,
)

bwd_segsum_dadt_kernel_varlen[grid](
dSSdA, SSdA, dadt_out,
cu_seqlens, state_seq_mapping, state_chunk_in_seq,
dSSdA.stride(0), dSSdA.stride(1), dSSdA.stride(2),
dSSdA.stride(3), dSSdA.stride(4),
SSdA.stride(0), SSdA.stride(1), SSdA.stride(2),
SSdA.stride(3), SSdA.stride(4),
dadt_out.stride(0), dadt_out.stride(1), dadt_out.stride(2),
cu_seqlens.stride(0),
state_seq_mapping.stride(0),
state_chunk_in_seq.stride(0),
S=S,
CHUNK_SIZE=chunk_size,
)
bwd_segsum_dadt_kernel_varlen[grid](
dSSdA, SSdA, dadt_out,
cu_seqlens, state_seq_mapping, state_chunk_in_seq,
dSSdA.stride(0), dSSdA.stride(1), dSSdA.stride(2),
dSSdA.stride(3), dSSdA.stride(4),
SSdA.stride(0), SSdA.stride(1), SSdA.stride(2),
SSdA.stride(3), SSdA.stride(4),
dadt_out.stride(0), dadt_out.stride(1), dadt_out.stride(2),
cu_seqlens.stride(0),
state_seq_mapping.stride(0),
state_chunk_in_seq.stride(0),
S=S,
CHUNK_SIZE=chunk_size,
)
return dadt_out


Expand Down Expand Up @@ -1388,21 +1392,22 @@ def bwd_dtrap_ddt_triton_varlen(
dtrap = torch.zeros_like(trap)
grid = (B, H, nchunks_global)

bwd_dtrap_ddt_kernel_varlen[grid](
trap, dt, dfactor, dgamma_diag, ddt, dtrap,
cu_seqlens, state_seq_mapping, state_chunk_in_seq,
trap.stride(0), trap.stride(1), trap.stride(2),
dt.stride(0), dt.stride(1), dt.stride(2),
dfactor.stride(0), dfactor.stride(1), dfactor.stride(2),
dgamma_diag.stride(0), dgamma_diag.stride(1), dgamma_diag.stride(2),
ddt.stride(0), ddt.stride(1), ddt.stride(2),
dtrap.stride(0), dtrap.stride(1), dtrap.stride(2),
cu_seqlens.stride(0),
state_seq_mapping.stride(0),
state_chunk_in_seq.stride(0),
S=S,
CHUNK_SIZE=chunk_size,
)
with torch.cuda.device(trap.device.index):
bwd_dtrap_ddt_kernel_varlen[grid](
trap, dt, dfactor, dgamma_diag, ddt, dtrap,
cu_seqlens, state_seq_mapping, state_chunk_in_seq,
trap.stride(0), trap.stride(1), trap.stride(2),
dt.stride(0), dt.stride(1), dt.stride(2),
dfactor.stride(0), dfactor.stride(1), dfactor.stride(2),
dgamma_diag.stride(0), dgamma_diag.stride(1), dgamma_diag.stride(2),
ddt.stride(0), ddt.stride(1), ddt.stride(2),
dtrap.stride(0), dtrap.stride(1), dtrap.stride(2),
cu_seqlens.stride(0),
state_seq_mapping.stride(0),
state_chunk_in_seq.stride(0),
S=S,
CHUNK_SIZE=chunk_size,
)
return ddt, dtrap


Expand Down
Loading