Skip to content

[KDA] Speed up chunk_kda by introducing lowerbound gate #703

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

Merged
zhiyuan1i merged 10 commits into from
Dec 30, 2025
Merged

[KDA] Speed up chunk_kda by introducing lowerbound gate #703

Show file tree
Hide file tree
Changes from 9 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
0c9799c
[KDA] Speed up chunk_kda by introducing lowerbound gate
zhiyuan1i Dec 30, 2025
97c085f
add doc strings
zhiyuan1i Dec 30, 2025
e9a0605
fix typos
zhiyuan1i Dec 30, 2025
21a9a0b
remove lines
zhiyuan1i Dec 30, 2025
e71d48f
clean
zhiyuan1i Dec 30, 2025
970f78c
Add check
zhiyuan1i Dec 30, 2025
3896ef4
update bmk
zhiyuan1i Dec 30, 2025
2e003e3
add dA warning
zhiyuan1i Dec 30, 2025
86867ca
[KDA]: Support FP16 Gate (train/flash-linear-attention!43)
zhiyuan1i Dec 30, 2025
990535b
更新 chunk_delta_h.py
zhiyuan1i Dec 30, 2025
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
24 changes: 6 additions & 18 deletions benchmarks/ops/benchmark_kda.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,4 @@

import os

import torch
import triton
Expand Down Expand Up @@ -38,21 +37,11 @@ def benchmark(T, provider):
dtype = torch.bfloat16
B, H, D = 1, 16, 128

# Set TMA environment variable based on provider
original_tma_env = os.environ.get('FLA_USE_TMA', '0')

if provider.endswith('_no_tma'):
os.environ['FLA_USE_TMA'] = '0'
provider_base = provider.replace('_no_tma', '')
else:
os.environ['FLA_USE_TMA'] = '1'
provider_base = provider

quantiles = [0.5, 0.2, 0.8]
results = 0, 0, 0

do = torch.randn(B, T, H, D, dtype=dtype, device=device)
if provider_base == 'gdn':
if provider == 'gdn':
q = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
k = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
v = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
Expand All @@ -69,7 +58,7 @@ def benchmark(T, provider):
)[0].backward(do),
quantiles=quantiles,
)
elif provider_base == 'attn':
elif provider == 'attn':
q = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
k = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
v = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
Expand All @@ -81,7 +70,7 @@ def benchmark(T, provider):
).backward(do),
quantiles=quantiles,
)
elif provider_base == 'comba':
elif provider == 'comba':
q = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
k = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
p = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
Expand All @@ -100,7 +89,7 @@ def benchmark(T, provider):
)[0].backward(do),
quantiles=quantiles,
)
elif provider_base == 'kda':
elif provider == 'kda':
q = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
k = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
v = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
Expand All @@ -114,10 +103,11 @@ def benchmark(T, provider):
g=g,
beta=beta,
use_qk_l2norm_in_kernel=True,
safe_gate=True, # Plese Carefully read doc strings
Comment thread
zhiyuan1i marked this conversation as resolved.
)[0].backward(do),
quantiles=quantiles,
)
elif provider_base == 'dplr':
elif provider == 'dplr':
q = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
k = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
a = torch.randn(B, T, H, D, dtype=dtype, device=device).requires_grad_(True)
Expand All @@ -137,8 +127,6 @@ def benchmark(T, provider):
quantiles=quantiles,
)

# Restore original TMA environment variable
os.environ['FLA_USE_TMA'] = original_tma_env
return results


Expand Down
20 changes: 10 additions & 10 deletions fla/ops/common/chunk_delta_h.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,7 +146,7 @@ def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
m_t = (i_t * BT + tl.arange(0, BT)) < T
b_g_last = tl.load(g + bos * H + last_idx * H + i_h)
p_g = tl.make_block_ptr(g + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,))
b_g = tl.load(p_g, boundary_check=(0,))
b_g = tl.load(p_g, boundary_check=(0,)).to(tl.float32)
if USE_EXP2:
b_v = b_v * tl.where(m_t, exp2(b_g_last - b_g), 0)[:, None]
b_g_last = exp2(b_g_last)
Expand All @@ -163,28 +163,28 @@ def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(

if USE_GK:
o_k1 = tl.arange(0, 64)
b_gk_last1 = tl.load(gk + (bos + last_idx) * H*K + i_h * K + o_k1, mask=(o_k1 < K), other=0.)
b_gk_last1 = tl.load(gk + (bos + last_idx) * H*K + i_h * K + o_k1, mask=(o_k1 < K), other=0.).to(tl.float32)
if USE_EXP2:
b_h1 *= exp2(b_gk_last1)[:, None]
else:
b_h1 *= exp(b_gk_last1)[:, None]
if K > 64:
o_k2 = 64 + o_k1
b_gk_last2 = tl.load(gk + (bos + last_idx) * H*K + i_h * K + o_k2, mask=(o_k2 < K), other=0.)
b_gk_last2 = tl.load(gk + (bos + last_idx) * H*K + i_h * K + o_k2, mask=(o_k2 < K), other=0.).to(tl.float32)
if USE_EXP2:
b_h2 *= exp2(b_gk_last2)[:, None]
else:
b_h2 *= exp(b_gk_last2)[:, None]
if K > 128:
o_k3 = 128 + o_k1
b_gk_last3 = tl.load(gk + (bos + last_idx) * H*K + i_h * K + o_k3, mask=(o_k3 < K), other=0.)
b_gk_last3 = tl.load(gk + (bos + last_idx) * H*K + i_h * K + o_k3, mask=(o_k3 < K), other=0.).to(tl.float32)
if USE_EXP2:
b_h3 *= exp2(b_gk_last3)[:, None]
else:
b_h3 *= exp(b_gk_last3)[:, None]
if K > 192:
o_k4 = 192 + o_k1
b_gk_last4 = tl.load(gk + (bos + last_idx) * H*K + i_h * K + o_k4, mask=(o_k4 < K), other=0.)
b_gk_last4 = tl.load(gk + (bos + last_idx) * H*K + i_h * K + o_k4, mask=(o_k4 < K), other=0.).to(tl.float32)
if USE_EXP2:
b_h4 *= exp2(b_gk_last4)[:, None]
else:
Expand Down Expand Up @@ -335,7 +335,7 @@ def chunk_gated_delta_rule_bwd_kernel_dhu_blockdim64(
if USE_G:
bg_last = tl.load(g + (bos + last_idx) * H + i_h)
p_g = tl.make_block_ptr(g + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,))
b_g = tl.load(p_g, boundary_check=(0,))
b_g = tl.load(p_g, boundary_check=(0,)).to(tl.float32)
if USE_EXP2:
bg_last_exp = exp2(bg_last)
b_g_exp = exp2(b_g)
Expand All @@ -354,31 +354,31 @@ def chunk_gated_delta_rule_bwd_kernel_dhu_blockdim64(
b_k = tl.load(p_k, boundary_check=(0, 1))
if USE_GK:
o_k1 = tl.arange(0, 64)
b_gk_last1 = tl.load(gk + last_idx * H*K + o_k1, mask=(o_k1 < K), other=0.)
b_gk_last1 = tl.load(gk + last_idx * H*K + o_k1, mask=(o_k1 < K), other=0.).to(tl.float32)
b_dv = tl.dot(b_k, b_dh1.to(b_k.dtype))

if K > 64:
p_k = tl.make_block_ptr(k, (T, K), (H*K, 1), (i_t * BT, 64), (BT, 64), (1, 0))
b_k = tl.load(p_k, boundary_check=(0, 1))
if USE_GK:
o_k2 = 64 + o_k1
b_gk_last2 = tl.load(gk + last_idx * H*K + o_k2, mask=(o_k2 < K), other=0.)
b_gk_last2 = tl.load(gk + last_idx * H*K + o_k2, mask=(o_k2 < K), other=0.).to(tl.float32)
b_dv += tl.dot(b_k, b_dh2.to(b_k.dtype))

if K > 128:
p_k = tl.make_block_ptr(k, (T, K), (H*K, 1), (i_t * BT, 128), (BT, 64), (1, 0))
b_k = tl.load(p_k, boundary_check=(0, 1))
if USE_GK:
o_k3 = 128 + o_k1
b_gk_last3 = tl.load(gk + last_idx * H*K + o_k3, mask=(o_k3 < K), other=0.)
b_gk_last3 = tl.load(gk + last_idx * H*K + o_k3, mask=(o_k3 < K), other=0.).to(tl.float32)
b_dv += tl.dot(b_k, b_dh3.to(b_k.dtype))

if K > 192:
p_k = tl.make_block_ptr(k, (T, K), (H*K, 1), (i_t * BT, 192), (BT, 64), (1, 0))
b_k = tl.load(p_k, boundary_check=(0, 1))
if USE_GK:
o_k4 = 192 + o_k1
b_gk_last4 = tl.load(gk + last_idx * H*K + o_k4, mask=(o_k4 < K), other=0.)
b_gk_last4 = tl.load(gk + last_idx * H*K + o_k4, mask=(o_k4 < K), other=0.).to(tl.float32)
b_dv += tl.dot(b_k, b_dh4.to(b_k.dtype))

if USE_G:
Expand Down
2 changes: 1 addition & 1 deletion fla/ops/gla/chunk.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -349,7 +349,7 @@ def chunk_gla_fwd_kernel_o(
# [BT, BK]
b_q = tl.load(p_q, boundary_check=(0, 1))
# [BT, BK]
b_g = tl.load(p_g, boundary_check=(0, 1))
b_g = tl.load(p_g, boundary_check=(0, 1)).to(tl.float32)
# [BT, BK]
if USE_EXP2:
b_qg = (b_q * exp2(b_g)).to(b_q.dtype)
Expand Down
Loading