Skip to content

GitLab

Commit 4f83cf8f authored by Junxian's avatar Junxian
Browse files

[release] v0.0.1

parents
Hide whitespace changes
Inline Side-by-side
Showing with 4265 additions and 0 deletions
block_sparse_tests/fwd/test_performance/blocksparse.py 0 → 100644
View file @ 4f83cf8f
# Adapted from https://github.com/Dao-AILab/flash-attention/blob/main/benchmarks/benchmark_flash_attention.py
import torch
from block_sparse_attn import (
block_sparse_attn_func,
flash_attn_varlen_func,
)
from utils import (
time_fwd,
flops,
efficiency,
write_to_excel,
)
def generate_base_sparsity_mask(max_seqlen_q, max_seqlen_k, round_base, m_block_dim, n_block_dim, sparsity, causal=False, device="cuda"):
def round_to_multiple(x, base):
return ((x + base - 1) // base) * base
nrow, ncol = round_to_multiple(max_seqlen_q, round_base) // m_block_dim, round_to_multiple(max_seqlen_k, round_base) // n_block_dim
base_mask = torch.zeros(1, nrow, ncol, device=device, dtype=torch.bool)
total_block_num = 0
density = 1.0 - sparsity
if not density == 0.0 and not density == 1.0:
for i in range(nrow): # do in reverse order
idx = nrow - i - 1
if causal:
available_col_num = max(0, ncol - i)
total_block_num += available_col_num
num_one = max(1, int(density * available_col_num))
base_mask[0][idx, torch.randperm(available_col_num)[:num_one]] = True
else:
available_col_num = ncol
total_block_num += available_col_num
num_one = max(1, int(density * available_col_num))
base_mask[0][idx, torch.randperm(available_col_num)[:num_one]] = True
elif density == 1.0:
base_mask[0] = torch.ones_like(base_mask[0])
total_block_num = nrow * ncol
else:
total_block_num = nrow * ncol
calculated_block_num = base_mask.sum().item()
real_sparsity = 1.0 - calculated_block_num / total_block_num
return base_mask, real_sparsity
block_size = 128
def get_sparsity_list(sampling_steps, seqlen, causal):
blockmask_element_num = (seqlen // block_size) ** 2 // (2 if causal else 1)
stride = max(blockmask_element_num // sampling_steps, 1)
actual_steps = (blockmask_element_num + stride - 1) // stride
sparsity_list = []
for i in range(actual_steps):
sparse_rate = (1 + i * stride) / blockmask_element_num
if sparse_rate > 0.95 or sparse_rate < 0.0:
continue
sparsity_list.append(sparse_rate)
return sparsity_list
def profile_blocksparse_fwd():
repeats = 15
block_sparse_repeats = 3
device = 'cuda:0'
dtype = torch.float16
causal = True
batch_size = 8
sparsity_sampling_steps = 20
seqlen_vals = [1024,2048,4096,8192,16384,32768,65536]
headdim = 128
dim = 4096
dropout_p = 0.0
method = ("Block_Sparse_Flash2")
time_f = {}
speed_f = {}
excel_label = ["batch_size", "seqlen", "actual_sparsity", "speed", "latency", "speedup", "base_speed", "base_latency"]
excel_data = []
excel_dir_path = "./excel/blocksparse/"
excel_file_name = f"hdim{headdim}_nheads{dim // headdim}_bts{batch_size}_fwd"
if causal:
excel_file_name += "_causal"
all_results = {}
for seqlen in seqlen_vals:
results = {}
nheads = dim // headdim
shape = (batch_size * seqlen, nheads, headdim)
q = torch.randn(shape, device=device, dtype=dtype)
k = torch.randn(shape, device=device, dtype=dtype)
v = torch.randn(shape, device=device, dtype=dtype)
cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32, device=device)
base_f = time_fwd(flash_attn_varlen_func, q, k, v, cu_seqlens, cu_seqlens, seqlen, seqlen, dropout_p, None, causal, repeats=repeats, verbose=False)
base_speed = efficiency(flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd"), base_f)
results["base"] = [[base_f], [base_speed]]
sparsity_list = get_sparsity_list(sparsity_sampling_steps, seqlen, causal)
print(f"sparsity_list: {sparsity_list}")
for sparsity in sparsity_list:
sum_sparsity, sum_speed, sum_latency = 0, 0, 0
for _ in range(block_sparse_repeats):
cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32, device=device)
head_mask_type = torch.tensor([1] * nheads, device=device, dtype=torch.int32)
base_blockmask, real_sparsity = generate_base_sparsity_mask(seqlen, seqlen, block_size, block_size, block_size, sparsity, causal = causal, device=device)
base_blockmask = base_blockmask.unsqueeze(0).repeat(batch_size, nheads, 1, 1)
config = (causal, headdim, nheads, batch_size, seqlen, sparsity, real_sparsity)
f = time_fwd(block_sparse_attn_func, q, k, v, cu_seqlens, cu_seqlens, head_mask_type, None, base_blockmask, seqlen, seqlen, dropout_p, is_causal=causal, exact_streaming=False, repeats=repeats, verbose=False)
time_f[config, method] = f
print(f"### causal={causal}, headdim={headdim}, nheads = {nheads}, batch_size={batch_size}, seqlen={seqlen}, real_sparsity={real_sparsity} ###")
speed_f[config, method] = efficiency(flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd"), time_f[config, method])
print(
f"{method}"
f"fwd: {speed_f[config, method]:.2f} TFLOPs/s, {(time_f[config, method]*1000):.2f} ms, "
f"fwd base: {base_speed:.2f} TFLOPs/s, {base_f*1000:.2f} ms"
)
sum_sparsity += real_sparsity
sum_speed += speed_f[config, method]
sum_latency += time_f[config, method]
avg_sparsity = sum_sparsity / block_sparse_repeats
avg_speed = sum_speed / block_sparse_repeats
avg_latency = sum_latency / block_sparse_repeats
if avg_sparsity not in results:
results[avg_sparsity] = [[],[]]
results[avg_sparsity][0].append(avg_latency)
results[avg_sparsity][1].append(avg_speed)
excel_data.append([batch_size, seqlen, avg_sparsity, avg_speed, avg_latency, avg_speed / base_speed, base_speed, base_f])
for key in results.keys():
avg_latency = sum(results[key][0]) / len(results[key][0])
avg_speed = sum(results[key][1]) / len(results[key][1])
results[key] = [avg_latency, avg_speed]
all_results[seqlen] = results
import json
with open(f"all_results_{excel_file_name}.json", "w") as f:
json.dump(all_results, f)
write_to_excel(excel_label, excel_data, excel_dir_path, excel_file_name)
profile_blocksparse_fwd()
\ No newline at end of file
block_sparse_tests/fwd/test_performance/token_streaming.py 0 → 100644
View file @ 4f83cf8f
# Adapted from https://github.com/Dao-AILab/flash-attention/blob/main/benchmarks/benchmark_flash_attention.py
import openpyxl
from block_sparse_attn.utils.benchmark import benchmark_forward
import math
import torch
from block_sparse_attn import (
token_streaming_attn_func,
flash_attn_varlen_func,
)
from utils import (
time_fwd,
flops,
efficiency,
write_to_excel,
)
def profile_exact_streaming_fwd():
repeats = 20
block_sparse_repeats = 10
device = 'cuda:0'
dtype = torch.float16
causal = True
batch_size = 8
sink_local_num = [64,256]
seqlen_vals = [4096,8192,16384,32768,65536]
headdim_vals = [128]
dim = 4096
dropout_p = 0.0
methods = (["Flash2"])
time_f = {}
speed_f = {}
for headdim in headdim_vals:
excel_label = ["batch_size", "seqlen", "speed", "latency", "speedup", "base_speed", "base_latency"]
excel_data = []
excel_dir_path = "./excel/streaming/"
excel_file_name = f"hdim{headdim}_nheads{dim // headdim}_bts{batch_size}_sink{sink_local_num[0]}_local{sink_local_num[1]}_fwd"
for seqlen in seqlen_vals:
nheads = dim // headdim
shape = (batch_size * seqlen, nheads, headdim)
q = torch.randn(shape, device=device, dtype=dtype)
k = torch.randn(shape, device=device, dtype=dtype)
v = torch.randn(shape, device=device, dtype=dtype)
cu_seqlens = torch.arange(
0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32, device=device)
base_f = time_fwd(flash_attn_varlen_func, q, k, v, cu_seqlens, cu_seqlens, seqlen, seqlen, dropout_p, None, causal, repeats=repeats, verbose=False)
base_speed = efficiency(flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd"), base_f)
head_mask_type = torch.tensor([-1] * (nheads//2) + [0] * (nheads - nheads//2), device=device, dtype=torch.int32)
streaming_info = torch.tensor([sink_local_num[0], sink_local_num[1]] * nheads, device=device, dtype=torch.int32)
config = (causal, headdim, nheads, batch_size, seqlen, sink_local_num[0], sink_local_num[1])
sum_speed, sum_latency = 0,0
for _ in range(block_sparse_repeats):
f = time_fwd(
token_streaming_attn_func, q, k, v, cu_seqlens, cu_seqlens, head_mask_type, streaming_info, seqlen, seqlen, repeats=repeats, verbose=False
)
time_f[config, "Flash2"] = f
print(f"### causal={causal}, headdim={headdim}, nheads = {nheads}, batch_size={batch_size}, seqlen={seqlen}, sink={sink_local_num[0]}, local={sink_local_num[1]} ###")
for method in methods:
speed_f[config, method] = efficiency(
flops(batch_size, seqlen, headdim,
nheads, causal, mode="fwd"),
time_f[config, method]
)
print(f"{method} fwd: {speed_f[config, method]:.2f} TFLOPs/s, {(time_f[config, method]*1000):.2f} ms, ")
sum_speed += speed_f[config, "Flash2"]
sum_latency += time_f[config, "Flash2"]
excel_data.append([batch_size, seqlen, sum_speed / block_sparse_repeats, sum_latency / block_sparse_repeats, (sum_speed / block_sparse_repeats) / base_speed, base_speed, base_f])
write_to_excel(excel_label, excel_data, excel_dir_path, excel_file_name)
profile_exact_streaming_fwd()
\ No newline at end of file
block_sparse_tests/fwd/test_performance/utils.py 0 → 100644
View file @ 4f83cf8f
# Adapted from https://github.com/Dao-AILab/flash-attention/blob/main/benchmarks/benchmark_flash_attention.py
import openpyxl
from block_sparse_attn.utils.benchmark import benchmark_forward
import math
import torch
import os
def benchmark_fwd(
fn,
*inputs,
grad=None,
repeats=10,
desc="",
verbose=True,
amp=False,
amp_dtype=torch.float16,
**kwinputs,
):
"""Use Pytorch Benchmark on the forward pass of an arbitrary function."""
return benchmark_forward(
fn,
*inputs,
repeats=repeats,
desc=desc,
verbose=verbose,
amp=amp,
amp_dtype=amp_dtype,
**kwinputs,
)
def flops(batch, seqlen, headdim, nheads, causal, mode="fwd"):
assert mode in ["fwd"]
f = 4 * batch * seqlen**2 * nheads * headdim // (2 if causal else 1)
return f if mode == "fwd" else (2.5 * f if mode == "bwd" else 3.5 * f)
def efficiency(flop, time):
return (flop / time / 10**12) if not math.isnan(time) else 0.0
def time_fwd(func, *args, **kwargs):
time_f = benchmark_fwd(func, *args, **kwargs)
return time_f[1].mean
def write_to_excel(label, data, dir_path, file_name):
workbook = openpyxl.Workbook()
sheet = workbook.active
sheet.append(label)
os.makedirs(dir_path, exist_ok=True)
for row in data:
sheet.append(row)
workbook.save(dir_path + file_name + ".xlsx")
block_sparse_tests/fwd_bwd/test_correctness/full_test.py 0 → 100644
View file @ 4f83cf8f
# Adapted from https://github.com/Dao-AILab/flash-attention/blob/main/tests/test_flash_attn.py
import pytest
import torch
from einops import repeat
from block_sparse_attn import (
block_sparse_attn_func,
)
from utils import (
generate_random_padding_mask,
generate_base_sparsity_mask,
generate_qkv,
generate_streaming_mask,
prepare_mixed_exact_mask,
prepare_mixed_mask,
convert_flash_attn_S_to_softmax,
normalize_flash_attn_S,
get_dropout_fraction,
attention_blocksparse_ref
)
MAX_HEADDIM_SM8x = 192
block_size = 128
is_sm75 = torch.cuda.get_device_capability("cuda") == (7, 5)
is_sm8x = torch.cuda.get_device_capability("cuda")[0] == 8
is_sm80 = torch.cuda.get_device_capability("cuda") == (8, 0)
is_sm90 = torch.cuda.get_device_capability("cuda") == (9, 0)
@pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
@pytest.mark.parametrize("mha_type", ["mha", "mqa", "gqa"])
@pytest.mark.parametrize("d", [32, 64, 128])
@pytest.mark.parametrize(
"seqlen_q,seqlen_k",
[
(113, 203),
(128, 217),
(113, 211),
(108, 256),
(256, 512),
(512, 256),
(1024, 1024),
(1023, 1024),
(1024, 1023),
(2048, 2048),
],
)
@pytest.mark.parametrize(
"causal, exact_streaming, sink_num, local_num",
[
# (True, True, 1, 3),
# (True, True, 64, 256),
(True, False, 1, 3),
(False, False, 1, 3),
]
)
@pytest.mark.parametrize("p_dropout", [0.17, 0.0])
@pytest.mark.parametrize("sparsity", [0, 0.1, 0.3, 0.7, 1.0])
@pytest.mark.parametrize("batch_size", [1, 2])
@pytest.mark.parametrize("nheads", [16, 32])
def test_flash_attn_varlen_block_output(
seqlen_q, seqlen_k, d, p_dropout, causal, exact_streaming, sink_num, local_num, mha_type, dtype, sparsity, batch_size, nheads
):
if (
max(seqlen_q, seqlen_k) >= 2048
and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30
):
pytest.skip() # Reference implementation OOM
device = "cuda:0"
# set seed
torch.random.manual_seed(42)
nheads_k = nheads if mha_type == "mha" else (1 if mha_type == "mqa" else 8)
assert nheads % nheads_k == 0
window_size = (-1, -1)
q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
k = torch.randn(batch_size, seqlen_k, nheads_k, d, device=device, dtype=dtype, requires_grad=True)
v = torch.randn(batch_size, seqlen_k, nheads_k, d, device=device, dtype=dtype, requires_grad=True)
query_padding_mask = generate_random_padding_mask(seqlen_q, batch_size, device, mode="random")
key_padding_mask = generate_random_padding_mask(seqlen_k, batch_size, device, mode="random")
alibi_slopes, attn_bias = None, None
(
q_unpad,
k_unpad,
v_unpad,
cu_seqlens_q,
cu_seqlens_k,
max_seqlen_q,
max_seqlen_k,
q,
k,
v,
output_pad_fn,
dq_pad_fn,
dk_pad_fn,
) = generate_qkv(q, k, v, query_padding_mask, key_padding_mask, kvpacked=False)
num_streaming_heads = nheads // 3
num_blocksparse_heads = nheads // 3
num_dense_heads = nheads - num_streaming_heads - num_blocksparse_heads
sparsity_list = [sparsity] * num_blocksparse_heads
head_mask_type = torch.tensor([0] * num_dense_heads + [1] * num_blocksparse_heads + [-1] * num_streaming_heads, device=device, dtype=torch.int32)
base_blockmask = generate_base_sparsity_mask(max_seqlen_q, max_seqlen_k, block_size, block_size, block_size, batch_size, num_blocksparse_heads, sparsity_list, causal = causal, device=device)
streaming_info = torch.tensor([sink_num, local_num] * nheads, device=device, dtype=torch.int32)
streaming_mask = generate_streaming_mask(max_seqlen_q, max_seqlen_k, batch_size, nheads, cu_seqlens_q, cu_seqlens_k, block_size, block_size, block_size, streaming_info, causal=causal, device=device)
if exact_streaming:
assert causal
print(f"exact_streaming: {exact_streaming}")
if exact_streaming:
mixed_mask = prepare_mixed_exact_mask(base_blockmask, streaming_info, head_mask_type, batch_size, nheads, block_size, block_size, block_size, max_seqlen_q, max_seqlen_k, q.shape[1], k.shape[1], query_padding_mask, key_padding_mask, device=device)
else:
mixed_mask = prepare_mixed_mask(base_blockmask, streaming_mask, head_mask_type, batch_size, nheads, block_size, block_size, block_size, max_seqlen_q, max_seqlen_k, q.shape[1], k.shape[1], device=device)
out_unpad, sm_lse, S_dmask = block_sparse_attn_func(
q_unpad, k_unpad, v_unpad,
cu_seqlens_q, cu_seqlens_k,
head_mask_type,
streaming_info,
base_blockmask,
max_seqlen_q, max_seqlen_k,
p_dropout,
deterministic=True,
softmax_scale=None,
is_causal=causal,
exact_streaming=exact_streaming,
return_attn_probs=True,
)
out = output_pad_fn(out_unpad)
if p_dropout > 0.0:
assert S_dmask is not None
S_dmask_converted = convert_flash_attn_S_to_softmax(
S_dmask,
seqlen_q,
seqlen_k,
query_padding_mask,
key_padding_mask,
d,
p_dropout > 0.0,
causal=causal,
window_size=window_size,
)
dropout_mask = S_dmask_converted >= 0
attn_unnorm = S_dmask_converted.abs()
k_rep = repeat(k, "b s h d -> b s (h g) d", g=nheads // nheads_k)
v_rep = repeat(v, "b s h d -> b s (h g) d", g=nheads // nheads_k)
attn = normalize_flash_attn_S(
attn_unnorm,
q,
k_rep,
v_rep,
query_padding_mask,
key_padding_mask,
attn_bias,
p_dropout > 0.0,
causal=causal,
window_size=window_size,
)
dropout_fraction = get_dropout_fraction(
dropout_mask,
mixed_mask,
block_size, block_size,
query_padding_mask,
key_padding_mask,
causal=causal,
window_size=window_size,
).item()
print(f"Actual dropout fraction: {dropout_fraction}")
else:
dropout_mask = None
out_ref, attn_ref = attention_blocksparse_ref(
q,
k,
v,
mixed_mask,
block_size, block_size,
query_padding_mask,
key_padding_mask,
p_dropout,
dropout_mask,
causal=causal,
window_size=window_size,
)
out_pt, attn_pt = attention_blocksparse_ref(
q,
k,
v,
mixed_mask,
block_size, block_size,
query_padding_mask,
key_padding_mask,
p_dropout,
dropout_mask,
causal=causal,
window_size=window_size,
upcast=False,
reorder_ops=True,
)
print(f"Output max diff: {(out - out_ref).abs().max().item()}")
print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")
g = torch.randn_like(out)
# g = torch.zeros_like(out)
if d <= MAX_HEADDIM_SM8x or (is_sm80 or is_sm90):
(
dq_unpad,
dk_unpad,
dv_unpad,
) = torch.autograd.grad(out, (q_unpad, k_unpad, v_unpad), g)
dk = dk_pad_fn(dk_unpad)
dv = dk_pad_fn(dv_unpad)
(
dq_ref,
dk_ref,
dv_ref,
) = torch.autograd.grad(out_ref, (q, k, v), g)
(
dq_pt,
dk_pt,
dv_pt,
) = torch.autograd.grad(out_pt, (q, k, v), g)
dq = dq_pad_fn(dq_unpad)
print(f"dQ max diff: {(dq - dq_ref).abs().max().item()}")
print(f"dK max diff: {(dk - dk_ref).abs().max().item()}")
print(f"dV max diff: {(dv - dv_ref).abs().max().item()}")
print(f"dQ mean diff: {(dq - dq_ref).abs().mean().item()}")
print(f"dK mean diff: {(dk - dk_ref).abs().mean().item()}")
print(f"dV mean diff: {(dv - dv_ref).abs().mean().item()}")
print(f"dQ Pytorch max diff: {(dq_pt - dq_ref).abs().max().item()}")
print(f"dK Pytorch max diff: {(dk_pt - dk_ref).abs().max().item()}")
print(f"dV Pytorch max diff: {(dv_pt - dv_ref).abs().max().item()}")
print(f"dQ Pytorch mean diff: {(dq_pt - dq_ref).abs().mean().item()}")
print(f"dK Pytorch mean diff: {(dk_pt - dk_ref).abs().mean().item()}")
print(f"dV Pytorch mean diff: {(dv_pt - dv_ref).abs().mean().item()}")
assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()
if d <= MAX_HEADDIM_SM8x or (is_sm80 or is_sm90):
assert (dq - dq_ref).abs().max().item() <= 3 * (dq_pt - dq_ref).abs().max().item()
assert (dk - dk_ref).abs().max().item() <= 3 * (dk_pt - dk_ref).abs().max().item()
assert (dv - dv_ref).abs().max().item() <= 3 * (dv_pt - dv_ref).abs().max().item()
\ No newline at end of file
block_sparse_tests/fwd_bwd/test_correctness/utils.py 0 → 100644
View file @ 4f83cf8f
# Adapted from https://github.com/Dao-AILab/flash-attention/blob/main/tests/test_flash_attn.py
import math
import torch
import torch.nn.functional as F
from einops import rearrange, repeat
from block_sparse_attn.bert_padding import pad_input, unpad_input
from block_sparse_attn.flash_attn_interface import _get_block_size
torch.set_printoptions(profile="full")
def generate_random_padding_mask(max_seqlen, batch_size, device, mode="random"):
assert mode in ["full", "random", "third"]
if mode == "full":
lengths = torch.full((batch_size, 1), max_seqlen, device=device, dtype=torch.int32)
elif mode == "random":
lengths = torch.randint(
max(1, max_seqlen - 20), max_seqlen + 1, (batch_size, 1), device=device
)
elif mode == "third":
lengths = torch.randint(max_seqlen // 3, max_seqlen + 1, (batch_size, 1), device=device)
padding_mask = (
repeat(torch.arange(max_seqlen, device=device), "s -> b s", b=batch_size) < lengths
)
return padding_mask
def construct_local_mask(
seqlen_q,
seqlen_k,
window_size=(-1, -1), # -1 means infinite window size
query_padding_mask=None,
key_padding_mask=None,
device=None,
):
row_idx = rearrange(torch.arange(seqlen_q, device=device, dtype=torch.long), "s -> s 1")
col_idx = torch.arange(seqlen_k, device=device, dtype=torch.long)
sk = (
seqlen_k
if key_padding_mask is None
else rearrange(key_padding_mask.sum(-1), "b -> b 1 1 1")
)
sq = (
seqlen_q
if query_padding_mask is None
else rearrange(query_padding_mask.sum(-1), "b -> b 1 1 1")
)
if window_size[0] < 0:
return col_idx > row_idx + sk - sq + window_size[1]
else:
sk = torch.full_like(col_idx, seqlen_k) if key_padding_mask is None else sk
return torch.logical_or(
col_idx > torch.minimum(row_idx + sk - sq + window_size[1], sk),
col_idx < row_idx + sk - sq - window_size[0],
)
def construct_exact_streaming_mask(
seqlen_q,
seqlen_k,
sink_size, # -1 means infinite window size
local_size,
query_padding_mask=None,
key_padding_mask=None,
device=None,
):
assert sink_size >= 0
assert local_size >= 1
row_idx = rearrange(torch.arange(seqlen_q, device=device, dtype=torch.long), "s -> s 1")
col_idx = torch.arange(seqlen_k, device=device, dtype=torch.long)
sk = (
seqlen_k
if key_padding_mask is None
else rearrange(key_padding_mask.sum(-1), "b -> b 1 1 1")
)
sq = (
seqlen_q
if query_padding_mask is None
else rearrange(query_padding_mask.sum(-1), "b -> b 1 1 1")
)
sk = torch.full_like(col_idx, seqlen_k) if key_padding_mask is None else sk
mask = torch.logical_or(
col_idx > torch.minimum(row_idx + sk - sq, sk),
torch.logical_and(
col_idx < row_idx + sk - sq - (local_size-1), col_idx >= sink_size,
)
)
return mask
def generate_qkv(
q, k, v, query_padding_mask=None, key_padding_mask=None, kvpacked=False, qkvpacked=False
):
"""
Arguments:
q: (batch_size, seqlen_q, nheads, d)
k: (batch_size, seqlen_k, nheads_k, d)
v: (batch_size, seqlen_k, nheads_k, d)
query_padding_mask: (batch_size, seqlen), bool
key_padding_mask: (batch_size, seqlen), bool
"""
assert not (kvpacked and qkvpacked)
batch_size, seqlen_q, nheads, d = q.shape
_, seqlen_k, nheads_k, _ = k.shape
assert k.shape == (batch_size, seqlen_k, nheads_k, d)
assert v.shape == (batch_size, seqlen_k, nheads_k, d)
if query_padding_mask is not None:
q_unpad, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, query_padding_mask)
output_pad_fn = lambda output_unpad: pad_input(
output_unpad, indices_q, batch_size, seqlen_q
)
else:
q_unpad = rearrange(q, "b s h d -> (b s) h d")
cu_seqlens_q = torch.arange(
0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32, device=q_unpad.device
)
max_seqlen_q = seqlen_q
output_pad_fn = lambda output_unpad: rearrange(
output_unpad, "(b s) h d -> b s h d", b=batch_size
)
if key_padding_mask is not None:
k_unpad, indices_k, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)
v_unpad, _, _, _ = unpad_input(v, key_padding_mask)
else:
k_unpad = rearrange(k, "b s h d -> (b s) h d")
v_unpad = rearrange(v, "b s h d -> (b s) h d")
cu_seqlens_k = torch.arange(
0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32, device=k_unpad.device
)
max_seqlen_k = seqlen_k
if qkvpacked:
assert (query_padding_mask == key_padding_mask).all()
assert nheads == nheads_k
qkv_unpad = torch.stack([q_unpad, k_unpad, v_unpad], dim=1)
qkv = torch.stack([q, k, v], dim=2)
if query_padding_mask is not None:
dqkv_pad_fn = lambda dqkv_unpad: pad_input(dqkv_unpad, indices_q, batch_size, seqlen_q)
else:
dqkv_pad_fn = lambda dqkv_unpad: rearrange(
dqkv_unpad, "(b s) t h d -> b s t h d", b=batch_size
)
return (
qkv_unpad.detach().requires_grad_(),
cu_seqlens_q,
max_seqlen_q,
qkv.detach().requires_grad_(),
output_pad_fn,
dqkv_pad_fn,
)
elif kvpacked:
kv_unpad = torch.stack([k_unpad, v_unpad], dim=1)
kv = torch.stack([k, v], dim=2)
dq_pad_fn = output_pad_fn
if key_padding_mask is not None:
dkv_pad_fn = lambda dkv_unpad: pad_input(dkv_unpad, indices_k, batch_size, seqlen_k)
else:
dkv_pad_fn = lambda dkv_unpad: rearrange(
dkv_unpad, "(b s) t h d -> b s t h d", b=batch_size
)
return (
q_unpad.detach().requires_grad_(),
kv_unpad.detach().requires_grad_(),
cu_seqlens_q,
cu_seqlens_k,
max_seqlen_q,
max_seqlen_k,
q.detach().requires_grad_(),
kv.detach().requires_grad_(),
output_pad_fn,
dq_pad_fn,
dkv_pad_fn,
)
else:
dq_pad_fn = output_pad_fn
if key_padding_mask is not None:
dk_pad_fn = lambda dk_unpad: pad_input(dk_unpad, indices_k, batch_size, seqlen_k)
else:
dk_pad_fn = lambda dk_unpad: rearrange(dk_unpad, "(b s) h d -> b s h d", b=batch_size)
return (
q_unpad.detach().requires_grad_(),
k_unpad.detach().requires_grad_(),
v_unpad.detach().requires_grad_(),
cu_seqlens_q,
cu_seqlens_k,
max_seqlen_q,
max_seqlen_k,
q.detach().requires_grad_(),
k.detach().requires_grad_(),
v.detach().requires_grad_(),
output_pad_fn,
dq_pad_fn,
dk_pad_fn,
)
def generate_base_sparsity_mask(max_seqlen_q, max_seqlen_k, round_base, m_block_dim, n_block_dim, batch_size, num_blocksparse_heads, sparsity_list, causal=False, device="cuda"):
assert len(sparsity_list) == num_blocksparse_heads
def round_to_multiple(x, base):
return ((x + base - 1) // base) * base
nrow, ncol = round_to_multiple(max_seqlen_q, round_base) // m_block_dim, round_to_multiple(max_seqlen_k, round_base) // n_block_dim
base_mask = torch.zeros(batch_size, num_blocksparse_heads, nrow, ncol, device=device, dtype=torch.bool)
for batch in range(batch_size):
for head_rank in range(num_blocksparse_heads):
sparsity = sparsity_list[head_rank]
if not sparsity == 0.0 and not sparsity == 1.0:
for i in range(nrow):
idx = nrow - i - 1
if causal:
available_col_num = max(0, ncol - i)
num_one = max(1, int(sparsity * available_col_num))
base_mask[batch][head_rank][idx, torch.randperm(available_col_num)[:num_one]] = True
else:
available_col_num = ncol
num_one = max(1, int(sparsity * available_col_num))
base_mask[batch][head_rank][idx, torch.randperm(available_col_num)[:num_one]] = True
elif sparsity == 1.0:
base_mask[batch][head_rank] = torch.ones_like(base_mask[batch][head_rank])
return base_mask
def generate_streaming_mask(max_seqlen_q, max_seqlen_k, batch_size, num_heads,cu_q_len_list, cu_k_len_list, round_base, m_block_dim, n_block_dim, streaming_info, causal=False, device="cuda"):
assert len(streaming_info) == 2 * num_heads
assert len(cu_q_len_list) == batch_size + 1
assert len(cu_k_len_list) == batch_size + 1
assert round_base == m_block_dim == n_block_dim == 128
def round_to_multiple(x, base):
return ((x + base - 1) // base) * base
def ceil_div(x, y):
return (x + y - 1) // y
nrow, ncol = round_to_multiple(max_seqlen_q, round_base) // m_block_dim, round_to_multiple(max_seqlen_k, round_base) // n_block_dim
base_mask = torch.zeros(batch_size, num_heads, nrow, ncol, device=device, dtype=torch.bool)
for batch in range(batch_size):
actual_q = cu_q_len_list[batch + 1] - cu_q_len_list[batch]
actual_k = cu_k_len_list[batch + 1] - cu_k_len_list[batch]
start_row_idx = max((actual_q - actual_k) // m_block_dim, 0) if causal else 0
for head_rank in range(num_heads):
sink_block_num, local_block_num = streaming_info[head_rank * 2], streaming_info[head_rank * 2 + 1]
for i in range(start_row_idx, nrow):
if causal:
max_row_block_num = ceil_div(max(actual_k - actual_q, 0), n_block_dim) + 1 + i - start_row_idx
else:
max_row_block_num = ncol
base_mask[batch, head_rank, i, min(max(max_row_block_num - local_block_num, 0), ncol):min(max_row_block_num, ncol)] = True
base_mask[batch, head_rank, i, :sink_block_num] = True
return base_mask
def replace_ones_with_count(tensor):
ones_mask = tensor == 1
count = torch.cumsum(ones_mask, dim=-1).to(tensor.dtype)
count = count * ones_mask
tensor = tensor.masked_scatter(ones_mask, count[ones_mask])
return tensor
def prepare_mixed_mask(base_blocksparse_mask, base_streaming_mask, head_mask_type, batch_size, num_heads, round_base, m_block_dim, n_block_dim, max_seqlen_q, max_seqlen_k, actual_seqlen_q, actual_seqlen_k, device="cuda"):
def round_to_multiple(x, base):
return ((x + base - 1) // base) * base
nrow, ncol = round_to_multiple(max_seqlen_q, round_base) // m_block_dim, round_to_multiple(max_seqlen_k, round_base) // n_block_dim
mixed_mask = torch.zeros(batch_size, num_heads, nrow, ncol, device=device, dtype=torch.bool)
head_mask_type = replace_ones_with_count(head_mask_type)
for head_rank in range(num_heads):
mask_type = head_mask_type[head_rank]
if mask_type == 0:
mixed_mask[:, head_rank, :, :] = torch.ones_like(mixed_mask[:, head_rank, :, :])
elif mask_type > 0:
for i in range(batch_size):
mixed_mask[i, head_rank, :, :] = base_blocksparse_mask[i][mask_type - 1]
else:
for i in range(batch_size):
mixed_mask[i, head_rank, :, :] = base_streaming_mask[i, head_rank, :, :]
mixed_mask = repeat(mixed_mask, "b h s_m s_n -> b h (s_m d_m) (s_n d_n)", d_m=m_block_dim, d_n=n_block_dim)
mixed_mask = tailor_mixedmask_for_test(mixed_mask, actual_seqlen_q, actual_seqlen_k)
mixed_mask = ~mixed_mask
return mixed_mask
def prepare_mixed_exact_mask(base_blocksparse_mask, streaming_info, head_mask_type, batch_size, num_heads, round_base, m_block_dim, n_block_dim, max_seqlen_q, max_seqlen_k, actual_seqlen_q, actual_seqlen_k, query_padding_mask,
key_padding_mask, device="cuda"):
def round_to_multiple(x, base):
return ((x + base - 1) // base) * base
nrow, ncol = round_to_multiple(max_seqlen_q, round_base) // m_block_dim, round_to_multiple(max_seqlen_k, round_base) // n_block_dim
mixed_mask = torch.zeros(batch_size, num_heads, nrow, ncol, device=device, dtype=torch.bool)
head_mask_type = replace_ones_with_count(head_mask_type)
for head_rank in range(num_heads):
mask_type = head_mask_type[head_rank]
if mask_type == 0:
mixed_mask[:, head_rank, :, :] = torch.ones_like(mixed_mask[:, head_rank, :, :])
elif mask_type > 0:
for i in range(batch_size):
mixed_mask[i, head_rank, :, :] = base_blocksparse_mask[i][mask_type - 1]
mixed_mask = repeat(mixed_mask, "b h s_m s_n -> b h (s_m d_m) (s_n d_n)", d_m=m_block_dim, d_n=n_block_dim)
mixed_mask = tailor_mixedmask_for_test(mixed_mask, actual_seqlen_q, actual_seqlen_k)
mixed_mask = ~mixed_mask
for head_rank in range(num_heads):
mask_type = head_mask_type[head_rank]
if mask_type < 0:
exact_streaming_mask = construct_exact_streaming_mask(
actual_seqlen_q,
actual_seqlen_k,
streaming_info[head_rank * 2],
streaming_info[head_rank * 2 + 1],
query_padding_mask,
key_padding_mask,
device=device,
)
if exact_streaming_mask.dim() == 4:
for i in range(batch_size):
mixed_mask[i, head_rank, :, :] = exact_streaming_mask[i,0,:,:]
else:
for i in range(batch_size):
mixed_mask[i, head_rank, :, :] = exact_streaming_mask
return mixed_mask
def attention_blocksparse_ref(
q, k, v,
mixed_mask,
m_block_dim, n_block_dim,
query_padding_mask=None,
key_padding_mask=None,
p_dropout=0.0,
dropout_mask=None,
causal=False,
window_size=(-1, -1),
upcast=True,
reorder_ops=False,
):
# q, k, v = qkv.float().unbind(dim=2)
if causal:
window_size = (window_size[0], 0)
dtype_og = q.dtype
if upcast:
q, k, v = q.float(), k.float(), v.float()
seqlen_q, seqlen_k = q.shape[1], k.shape[1]
k = repeat(k, "b s h d -> b s (h g) d", g=q.shape[2] // k.shape[2])
v = repeat(v, "b s h d -> b s (h g) d", g=q.shape[2] // v.shape[2])
d = q.shape[-1]
if not reorder_ops:
scores = torch.einsum("bthd,bshd->bhts", q / math.sqrt(d), k)
else:
scores = torch.einsum("bthd,bshd->bhts", q, k / math.sqrt(d))
if key_padding_mask is not None:
scores.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), float("-inf"))
# local mask
if window_size[0] >= 0 or window_size[1] >= 0:
local_mask = construct_local_mask(
seqlen_q,
seqlen_k,
window_size,
query_padding_mask,
key_padding_mask,
q.device,
)
scores.masked_fill_(local_mask, float("-inf"))
scores.masked_fill_(rearrange(mixed_mask, "b h t s -> b h t s"), float("-inf"))
# print("processed blockmask: ", rearrange(~base_blockmask, "h t s -> 1 h t s"))
attention = torch.softmax(scores, dim=-1).to(v.dtype)
if window_size[0] >= 0 or window_size[1] >= 0:
attention = attention.masked_fill(torch.all(torch.bitwise_or(local_mask, rearrange(mixed_mask, "b h t s -> b h t s")), dim=-1, keepdim=True), 0.0)
attention = attention.masked_fill(rearrange(mixed_mask, "b h t s -> b h t s"), 0.0)
if query_padding_mask is not None:
attention = attention.masked_fill(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
dropout_scaling = 1.0 / (1 - p_dropout)
if dropout_mask is not None:
attention_drop = attention.masked_fill(~dropout_mask, 0.0)
else:
attention_drop = attention
output = torch.einsum("bhts,bshd->bthd", attention_drop, v * dropout_scaling)
if query_padding_mask is not None:
output.masked_fill_(rearrange(~query_padding_mask, "b s -> b s 1 1"), 0.0)
return output.to(dtype=dtype_og), attention.to(dtype=dtype_og)
def convert_flash_attn_S_to_softmax(
S,
seqlen_q,
seqlen_k,
query_padding_mask,
key_padding_mask,
head_dim,
is_dropout,
causal=False,
window_size=(-1, -1), # -1 means infinite window size
):
"""FlashAttention stores the S matrix in a different way.
Arguments:
S: (batch_size, nheads, seqlen_q_rounded, seqlen_k_rounded)
query_padding_mask: (batch_size, seqlen_q_rounded)
key_padding_mask: (batch_size, seqlen_k_rounded)
"""
if causal:
window_size = (window_size[0], 0)
seqlen_q_rounded, seqlen_k_rounded = S.shape[-2:]
warps_n = 4
blocksize_m, blocksize_n = _get_block_size(S.device, head_dim, is_dropout, causal)
nblocks_n = (seqlen_k_rounded + blocksize_n - 1) // blocksize_n
nblocks_m = (seqlen_q_rounded + blocksize_m - 1) // blocksize_m
mmas_n = (blocksize_n + 16 - 1) // 16
S_flat = rearrange(
S,
"b h (nblocks_m blocksize_m) (nblocks_n blocksize_n) -> b h nblocks_m nblocks_n (blocksize_m blocksize_n)",
blocksize_m=blocksize_m,
blocksize_n=blocksize_n,
)
S_converted = rearrange(
S_flat,
"b h nblocks_m nblocks_n (mmas_n mmas_m warps_n eight four c2 c1 c0) -> b h (nblocks_m mmas_m warps_n c1 eight) (nblocks_n mmas_n c2 four c0)",
mmas_n=mmas_n,
warps_n=warps_n,
eight=8,
c0=2,
c1=2,
c2=2,
four=4,
)
if window_size[0] >= 0 or window_size[1] >= 0:
local_mask = construct_local_mask(
seqlen_q,
seqlen_k,
window_size,
query_padding_mask,
key_padding_mask,
S.device,
)
local_mask = F.pad(
local_mask,
(0, seqlen_k_rounded - seqlen_k, 0, seqlen_q_rounded - seqlen_q),
value=True,
)
S_converted.masked_fill_(local_mask, 0.0)
# Need to zero out things not in attention_mask in case S was initialized with random values
# and some of those values aren't overwritten.
seqlen_q_og = (
query_padding_mask.shape[-1] if query_padding_mask is not None else seqlen_q_rounded
)
if query_padding_mask is not None:
query_padding_mask = F.pad(query_padding_mask, (0, seqlen_q_rounded - seqlen_q_og))
S_converted = S_converted.masked_fill(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
seqlen_k_og = key_padding_mask.shape[-1] if key_padding_mask is not None else seqlen_k
if key_padding_mask is not None:
key_padding_mask = F.pad(key_padding_mask, (0, seqlen_k_rounded - seqlen_k_og))
S_converted = S_converted.masked_fill(rearrange(~key_padding_mask, "b s -> b 1 1 s"), 0.0)
S_converted = F.pad(S_converted, (0, 0, 0, seqlen_q_og - seqlen_q_rounded))
S_converted = F.pad(S_converted, (0, seqlen_k_og - seqlen_k_rounded))
return S_converted[:, :, :seqlen_q, :seqlen_k]
def normalize_flash_attn_S(
attn_unnorm,
q,
k,
v,
query_padding_mask=None,
key_padding_mask=None,
attn_bias=None,
is_dropout=False,
causal=False,
window_size=(-1, -1), # -1 means infinite window size
):
"""
Arguments:
q: (batch_size, seqlen_q, nheads, head_dim)
k, v: (batch_size, seqlen_k, nheads, head_dim)
key_padding_mask: (batch_size, seqlen_q)
attn_bias: broadcastable to (batch_size, nheads, seqlen_q, seqlen_k)
Output:
softmax_lse: (batch_size, nheads, seqlen_q)
softmax_max: (batch_size, nheads, seqlen_q)
"""
if causal:
window_size = (window_size[0], 0)
q, k, v = q.float(), k.float(), v.float()
_, seqlen_q, _, head_dim = q.shape
seqlen_k = k.shape[1]
scores = torch.einsum("bthd,bshd->bhts", q / math.sqrt(head_dim), k)
if key_padding_mask is not None:
scores.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), float("-inf"))
if window_size[0] >= 0 or window_size[1] >= 0:
local_mask = construct_local_mask(
seqlen_q,
seqlen_k,
window_size,
query_padding_mask,
key_padding_mask,
q.device,
)
scores.masked_fill_(local_mask, float("-inf"))
if attn_bias is not None:
scores = scores + attn_bias.to(dtype=scores.dtype)
_, block_size_n = _get_block_size(scores.device, head_dim, is_dropout, causal)
scores_block = scores.split(block_size_n, dim=-1)
lse_block = torch.stack([torch.logsumexp(s, dim=-1) for s in scores_block], dim=-1)
lse = torch.logsumexp(lse_block, dim=-1)
# lse could be -inf (i.e. all values in scores are -inf), and we want to set those to inf
# so that when we do torch.exp(m - lse), we get 0.0 instead of NaN.
lse[lse == float("-inf")] = float("inf")
scores_max_block = torch.stack([torch.amax(s, dim=-1) for s in scores_block], dim=-1)
cummax_block = torch.cummax(scores_max_block.flip(-1), dim=-1).values.flip(-1).unbind(dim=-1)
attn_unnorm_block = attn_unnorm.split(block_size_n, dim=-1)
attn_norm = torch.cat(
[
a * rearrange(torch.exp(m - lse), "b h s -> b h s 1")
for a, m in zip(attn_unnorm_block, cummax_block)
],
dim=-1,
)
if query_padding_mask is not None:
attn_norm.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
return attn_norm.to(dtype=attn_unnorm.dtype)
def get_dropout_fraction(
dropout_mask,
mixed_mask,
m_block_dim, n_block_dim,
query_padding_mask=None,
key_padding_mask=None,
causal=False,
window_size=(-1, -1), # -1 means infinite window size
):
"""
dropout_mask: (batch_size, nheads, seqlen_q, seqlen_k), bool. True means keep, False means drop.
query_padding_mask: (batch_size, seqlen_q)
key_padding_mask: (batch_size, seqlen_k)
"""
if causal:
window_size = (window_size[0], 0)
batch_size, nheads, seqlen_q, seqlen_k = dropout_mask.shape
dropped = ~dropout_mask
valid = torch.ones_like(dropout_mask)
if mixed_mask is not None:
# mixed_mask = repeat(mixed_mask, "b h s_m s_n -> b h (s_m d_m) (s_n d_n)", d_m=m_block_dim, d_n=n_block_dim)
# mixed_mask = tailor_mixedmask_for_test(mixed_mask, seqlen_q, seqlen_k)
dropped.masked_fill_(rearrange(mixed_mask, "b h t s -> b h t s"), False)
valid.masked_fill_(rearrange(mixed_mask, "b h t s -> b h t s"), False)
if query_padding_mask is not None:
dropped.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), False)
valid.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), False)
if key_padding_mask is not None:
dropped.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), False)
valid.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), False)
if window_size[0] >= 0 or window_size[1] >= 0:
local_mask = construct_local_mask(
seqlen_q,
seqlen_k,
window_size,
query_padding_mask,
key_padding_mask,
dropout_mask.device,
)
dropped.masked_fill_(local_mask, False)
valid.masked_fill_(local_mask, False)
dropped_total = dropped.sum()
return dropped.sum() / valid.sum()
def modified_check(a, a_ref, a_pt, rel_paremeter):
assert a.shape == a_ref.shape
assert a.shape == a_pt.shape
left = (a - a_ref).abs().max().item()
right = (a_pt - a_ref).abs().max().item()
rtol = 1e-3
if not right == 0:
assert left < rel_paremeter * right or left < rtol * a_ref.abs().max().item()
else:
assert round(left, 4) == 0 or left < rtol * a_ref.abs().max().item()
def tailor_mixedmask_for_test(spanded_base_mixedmask, seqlen_q, seqlen_k):
batch_size = spanded_base_mixedmask.shape[0]
nheads = spanded_base_mixedmask.shape[1]
spanded_base_mixedmask = spanded_base_mixedmask[:, :, :seqlen_q, :seqlen_k]
pad_blockmask = torch.zeros(batch_size, nheads, seqlen_q, seqlen_k, dtype=torch.bool, device = spanded_base_mixedmask.device)
pad_blockmask[:, :, :spanded_base_mixedmask.shape[2], :spanded_base_mixedmask.shape[3]] = spanded_base_mixedmask
spanded_base_mixedmask = pad_blockmask
spanded_base_mixedmask = spanded_base_mixedmask.contiguous()
return spanded_base_mixedmask
\ No newline at end of file
block_sparse_tests/fwd_bwd/test_performance/block_streaming.py 0 → 100644
View file @ 4f83cf8f
# Adapted from https://github.com/Dao-AILab/flash-attention/blob/main/benchmarks/benchmark_flash_attention.py
import openpyxl
from block_sparse_attn.utils.benchmark import benchmark_forward
import math
import torch
from block_sparse_attn import (
block_streaming_attn_func,
flash_attn_varlen_func,
)
from utils import (
time_fwd_bwd,
flops,
efficiency,
write_to_excel,
)
def profile_block_streaming_fwd_bwd():
repeats = 10
block_sparse_repeats = 5
device = 'cuda:0'
dtype = torch.float16
causal = True
batch_size = 1
sink_local_block_num = [1,3]
seqlen_vals = [1024, 2048, 4096, 8192, 16384, 20480, 24576, 28672, 32768, 65536, 131072]
headdim_vals = [128]
dim = 4096
p_dropout = 0.0
methods = (["Flash2"])
time_f = {}
time_b = {}
time_f_b = {}
speed_f = {}
speed_b = {}
speed_f_b = {}
for headdim in headdim_vals:
excel_label = ["batch_size", "seqlen", "speed", "latency", "speedup", "base_speed", "base_latency"]
excel_data = []
excel_dir_path = "./excel/block_streaming/"
excel_file_name = f"hdim{headdim}_nheads{dim // headdim}_bts{batch_size}_sink_block{sink_local_block_num[0]}_local_block{sink_local_block_num[1]}_fwd_bwd"
for seqlen in seqlen_vals:
nheads = dim // headdim
shape = (batch_size * seqlen, nheads, headdim)
q = torch.randn(shape, device=device, dtype=dtype, requires_grad=True)
k = torch.randn(shape, device=device, dtype=dtype, requires_grad=True)
v = torch.randn(shape, device=device, dtype=dtype, requires_grad=True)
cu_seqlens = torch.arange(
0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32, device=device)
base_f, base_b = time_fwd_bwd(flash_attn_varlen_func, q, k, v, cu_seqlens, cu_seqlens, seqlen, seqlen, p_dropout, None, causal, repeats=repeats, verbose=False)
base_speed = efficiency(flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd_bwd"), base_f + base_b)
head_mask_type = torch.tensor([-1] * (nheads//2) + [0] * (nheads - nheads//2), device=device, dtype=torch.int32)
streaming_info = torch.tensor([sink_local_block_num[0], sink_local_block_num[1]] * nheads, device=device, dtype=torch.int32)
config = (causal, headdim, nheads, batch_size, seqlen, sink_local_block_num[0], sink_local_block_num[1])
sum_speed, sum_latency = 0,0
for _ in range(block_sparse_repeats):
f, b = time_fwd_bwd(
block_streaming_attn_func, q, k, v, cu_seqlens, cu_seqlens, head_mask_type, streaming_info, seqlen, seqlen, p_dropout, False, None, causal, repeats=repeats, verbose=False
)
time_f[config, "Flash2"] = f
time_b[config, "Flash2"] = b
print(f"### causal={causal}, headdim={headdim}, nheads = {nheads}, batch_size={batch_size}, seqlen={seqlen}, sink={sink_local_block_num[0]}, local={sink_local_block_num[1]} ###")
for method in methods:
time_f_b[config, method] = time_f[config, method] + time_b[config, method]
speed_f[config, method] = efficiency(
flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd"),
time_f[config, method]
)
speed_b[config, method] = efficiency(
flops(batch_size, seqlen, headdim, nheads, causal, mode="bwd"),
time_b[config, method]
)
speed_f_b[config, method] = efficiency(
flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd_bwd"),
time_f_b[config, method]
)
print(
f"{method}"
f"fwd: {speed_f[config, method]:.2f} TFLOPs/s, {(time_f[config, method]*1000):.2f} ms, "
f"bwd: {speed_b[config, method]:.2f} TFLOPs/s, {(time_b[config, method]*1000):.2f} ms, "
f"fwd + bwd: {speed_f_b[config, method]:.2f} TFLOPs/s, {(time_f_b[config, method]*1000):.2f} ms, "
f"fwd + bwd base: {base_speed:.2f} TFLOPs/s, {(base_f + base_b)*1000:.2f} ms"
)
sum_speed += speed_f_b[config, "Flash2"]
sum_latency += time_f_b[config, "Flash2"]
excel_data.append([batch_size, seqlen, sum_speed / block_sparse_repeats, sum_latency / block_sparse_repeats, (sum_speed / block_sparse_repeats) / base_speed, base_speed, base_f])
write_to_excel(excel_label, excel_data, excel_dir_path, excel_file_name)
profile_block_streaming_fwd_bwd()
\ No newline at end of file
block_sparse_tests/fwd_bwd/test_performance/blocksparse.py 0 → 100644
View file @ 4f83cf8f
# Adapted from https://github.com/Dao-AILab/flash-attention/blob/main/benchmarks/benchmark_flash_attention.py
import torch
from block_sparse_attn import (
block_sparse_attn_func,
flash_attn_varlen_func,
)
from utils import (
time_fwd_bwd,
flops,
efficiency,
write_to_excel,
)
def generate_base_sparsity_mask(max_seqlen_q, max_seqlen_k, round_base, m_block_dim, n_block_dim, sparsity, causal=False, device="cuda"):
def round_to_multiple(x, base):
return ((x + base - 1) // base) * base
nrow, ncol = round_to_multiple(max_seqlen_q, round_base) // m_block_dim, round_to_multiple(max_seqlen_k, round_base) // n_block_dim
base_mask = torch.zeros(1, nrow, ncol, device=device, dtype=torch.bool)
total_block_num = 0
density = 1.0 - sparsity
if not density == 0.0 and not density == 1.0:
for i in range(nrow): # do in reverse order
idx = nrow - i - 1
if causal:
available_col_num = max(0, ncol - i)
total_block_num += available_col_num
num_one = max(1, int(density * available_col_num))
base_mask[0][idx, torch.randperm(available_col_num)[:num_one]] = True
else:
available_col_num = ncol
total_block_num += available_col_num
num_one = max(1, int(density * available_col_num))
base_mask[0][idx, torch.randperm(available_col_num)[:num_one]] = True
elif density == 1.0:
base_mask[0] = torch.ones_like(base_mask[0])
total_block_num = nrow * ncol
else:
total_block_num = nrow * ncol
calculated_block_num = base_mask.sum().item()
real_sparsity = 1.0 - calculated_block_num / total_block_num
return base_mask, real_sparsity
block_size = 128
def get_sparsity_list(sampling_steps, seqlen, causal):
blockmask_element_num = (seqlen // block_size) ** 2 // (2 if causal else 1)
stride = max(blockmask_element_num // sampling_steps, 1)
actual_steps = (blockmask_element_num + stride - 1) // stride
sparsity_list = []
for i in range(actual_steps):
sparse_rate = (1 + i * stride) / blockmask_element_num
if sparse_rate > 0.95 or sparse_rate < 0.0:
continue
sparsity_list.append(sparse_rate)
return sparsity_list
def profile_blocksparse_fwd_bwd():
repeats = 10
block_sparse_repeats = 5
device = 'cuda:0'
dtype = torch.float16
causal = True
batch_size = 1
sparsity_sampling_steps = 20
seqlen_vals = [8192,16384,32768]
headdim = 128
dim = 4096
dropout_p = 0.0
method = ("Block_Sparse_Attn")
time_f = {}
time_b = {}
time_f_b = {}
speed_f = {}
speed_b = {}
speed_f_b = {}
excel_label = ["batch_size", "seqlen", "actual_sparsity", "speed", "latency", "speedup", "base_speed", "base_latency"]
excel_data = []
excel_dir_path = "./excel/blocksparse/"
excel_file_name = f"hdim{headdim}_nheads{dim // headdim}_bts{batch_size}_fwd_bwd"
if causal:
excel_file_name += "_causal"
all_results = {}
for seqlen in seqlen_vals:
results = {}
nheads = dim // headdim
shape = (batch_size * seqlen, nheads, headdim)
q = torch.randn(shape, device=device, dtype=dtype, requires_grad=True)
k = torch.randn(shape, device=device, dtype=dtype, requires_grad=True)
v = torch.randn(shape, device=device, dtype=dtype, requires_grad=True)
cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32, device=device)
base_f, base_b = time_fwd_bwd(flash_attn_varlen_func, q, k, v, cu_seqlens, cu_seqlens, seqlen, seqlen, dropout_p, None, causal, repeats=repeats, verbose=False)
base_speed = efficiency(flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd_bwd"), base_f + base_b)
results["base"] = [[base_f + base_b], [base_speed]]
sparsity_list = get_sparsity_list(sparsity_sampling_steps, seqlen, causal)
print(f"sparsity_list: {sparsity_list}")
for sparsity in sparsity_list:
sum_sparsity, sum_speed, sum_latency = 0, 0, 0
for _ in range(block_sparse_repeats):
cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32, device=device)
head_mask_type = torch.tensor([1] * nheads, device=device, dtype=torch.int32)
base_blockmask, real_sparsity = generate_base_sparsity_mask(seqlen, seqlen, block_size, block_size, block_size, sparsity, causal = causal, device=device)
base_blockmask = base_blockmask.unsqueeze(0).repeat(batch_size, nheads, 1, 1)
config = (causal, headdim, nheads, batch_size, seqlen, sparsity, real_sparsity)
f, b = time_fwd_bwd(block_sparse_attn_func, q, k, v, cu_seqlens, cu_seqlens, head_mask_type, None, base_blockmask, seqlen, seqlen, dropout_p, is_causal=causal, exact_streaming=False, repeats=repeats, verbose=False)
time_f[config, method] = f
time_b[config, method] = b
print(f"### causal={causal}, headdim={headdim}, nheads = {nheads}, batch_size={batch_size}, seqlen={seqlen}, real_sparsity={real_sparsity} ###")
time_f_b[config, method] = time_f[config, method] + time_b[config, method]
speed_f[config, method] = efficiency(
flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd"),
time_f[config, method]
)
speed_b[config, method] = efficiency(
flops(batch_size, seqlen, headdim, nheads, causal, mode="bwd"),
time_b[config, method]
)
speed_f_b[config, method] = efficiency(
flops(batch_size, seqlen, headdim, nheads, causal, mode="fwd_bwd"),
time_f_b[config, method]
)
print(
f"{method}"
f"fwd: {speed_f[config, method]:.2f} TFLOPs/s, {(time_f[config, method]*1000):.2f} ms, "
f"bwd: {speed_b[config, method]:.2f} TFLOPs/s, {(time_b[config, method]*1000):.2f} ms, "
f"fwd + bwd: {speed_f_b[config, method]:.2f} TFLOPs/s, {(time_f_b[config, method]*1000):.2f} ms, "
f"fwd + bwd base: {base_speed:.2f} TFLOPs/s, {(base_f + base_b)*1000:.2f} ms"
)
sum_sparsity += real_sparsity
sum_speed += speed_f_b[config, method]
sum_latency += time_f_b[config, method]
avg_sparsity = sum_sparsity / block_sparse_repeats
avg_speed = sum_speed / block_sparse_repeats
avg_latency = sum_latency / block_sparse_repeats
if avg_sparsity not in results:
results[avg_sparsity] = [[],[]]
results[avg_sparsity][0].append(avg_latency)
results[avg_sparsity][1].append(avg_speed)
excel_data.append([batch_size, seqlen, avg_sparsity, avg_speed, avg_latency, avg_speed / base_speed, base_speed, base_f + base_b])
for key in results.keys():
avg_latency = sum(results[key][0]) / len(results[key][0])
avg_speed = sum(results[key][1]) / len(results[key][1])
results[key] = [avg_latency, avg_speed]
all_results[seqlen] = results
import json
with open(f"all_results_{excel_file_name}.json", "w") as f:
json.dump(all_results, f)
write_to_excel(excel_label, excel_data, excel_dir_path, excel_file_name)
profile_blocksparse_fwd_bwd()
\ No newline at end of file
block_sparse_tests/fwd_bwd/test_performance/utils.py 0 → 100644
View file @ 4f83cf8f
# Adapted from https://github.com/Dao-AILab/flash-attention/blob/main/benchmarks/benchmark_flash_attention.py
import openpyxl
from block_sparse_attn.utils.benchmark import benchmark_forward, benchmark_backward
import math
import torch
import os
def benchmark_fwd_bwd(
fn,
*inputs,
grad=None,
repeats=10,
desc="",
verbose=True,
amp=False,
amp_dtype=torch.float16,
**kwinputs,
):
"""Use Pytorch Benchmark on the forward+backward pass of an arbitrary function."""
return (
benchmark_forward(
fn,
*inputs,
repeats=repeats,
desc=desc,
verbose=verbose,
amp=amp,
amp_dtype=amp_dtype,
**kwinputs,
),
benchmark_backward(
fn,
*inputs,
grad=grad,
repeats=repeats,
desc=desc,
verbose=verbose,
amp=amp,
amp_dtype=amp_dtype,
**kwinputs,
),
)
def flops(batch, seqlen, headdim, nheads, causal, mode="fwd"):
assert mode in ["fwd", "bwd", "fwd_bwd"]
f = 4 * batch * seqlen**2 * nheads * headdim // (2 if causal else 1)
return f if mode == "fwd" else (2.5 * f if mode == "bwd" else 3.5 * f)
def efficiency(flop, time):
return (flop / time / 10**12) if not math.isnan(time) else 0.0
def time_fwd_bwd(func, *args, **kwargs):
time_f, time_b = benchmark_fwd_bwd(func, *args, **kwargs)
return time_f[1].mean, time_b[1].mean
def write_to_excel(label, data, dir_path, file_name):
workbook = openpyxl.Workbook()
sheet = workbook.active
sheet.append(label)
os.makedirs(dir_path, exist_ok=True)
for row in data:
sheet.append(row)
workbook.save(dir_path + file_name + ".xlsx")
csrc/block_sparse_attn/flash_api.cpp 0 → 100644
View file @ 4f83cf8f
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
/******************************************************************************
* Adapted by Junxian Guo from https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/flash_api.cpp
******************************************************************************/
// Include these 2 headers instead of torch/extension.h since we don't need all of the torch headers.
#include <torch/python.h>
#include <torch/nn/functional.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cutlass/numeric_types.h>
#include "flash.h"
#include "static_switch.h"
#define CHECK_DEVICE(x) TORCH_CHECK(x.is_cuda(), #x " must be on CUDA")
#define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
void set_params_fprop(Flash_fwd_params &params,
// sizes
const size_t b,
const size_t seqlen_q,
const size_t seqlen_k,
const size_t seqlen_q_rounded,
const size_t seqlen_k_rounded,
const size_t h,
const size_t h_k,
const size_t d,
const size_t d_rounded,
// device pointers
const at::Tensor q,
const at::Tensor k,
const at::Tensor v,
at::Tensor out,
void *cu_seqlens_q_d,
void *cu_seqlens_k_d,
void *seqused_k,
void *p_d,
void *softmax_lse_d,
float p_dropout,
float softmax_scale,
int window_size_left,
int window_size_right) {
// Reset the parameters
memset(&params, 0, sizeof(params));
params.is_bf16 = q.dtype() == torch::kBFloat16;
// Set the pointers and strides.
params.q_ptr = q.data_ptr();
params.k_ptr = k.data_ptr();
params.v_ptr = v.data_ptr();
// All stride are in elements, not bytes.
params.q_row_stride = q.stride(-3);
params.k_row_stride = k.stride(-3);
params.v_row_stride = v.stride(-3);
params.q_head_stride = q.stride(-2);
params.k_head_stride = k.stride(-2);
params.v_head_stride = v.stride(-2);
params.o_ptr = out.data_ptr();
params.o_row_stride = out.stride(-3);
params.o_head_stride = out.stride(-2);
if (cu_seqlens_q_d == nullptr) {
params.q_batch_stride = q.stride(0);
params.k_batch_stride = k.stride(0);
params.v_batch_stride = v.stride(0);
params.o_batch_stride = out.stride(0);
}
params.cu_seqlens_q = static_cast<int *>(cu_seqlens_q_d);
params.cu_seqlens_k = static_cast<int *>(cu_seqlens_k_d);
params.seqused_k = static_cast<int *>(seqused_k);
// P = softmax(QK^T)
params.p_ptr = p_d;
// Softmax sum
params.softmax_lse_ptr = softmax_lse_d;
// Set the dimensions.
params.b = b;
params.h = h;
params.h_k = h_k;
params.h_h_k_ratio = h / h_k;
params.seqlen_q = seqlen_q;
params.seqlen_k = seqlen_k;
params.seqlen_q_rounded = seqlen_q_rounded;
params.seqlen_k_rounded = seqlen_k_rounded;
params.d = d;
params.d_rounded = d_rounded;
// Set the different scale values.
params.scale_softmax = softmax_scale;
params.scale_softmax_log2 = softmax_scale * M_LOG2E;
// Set this to probability of keeping an element to simplify things.
params.p_dropout = 1.f - p_dropout;
// Convert p from float to int so we don't have to convert the random uint to float to compare.
// [Minor] We want to round down since when we do the comparison we use <= instead of <
// params.p_dropout_in_uint = uint32_t(std::floor(params.p_dropout * 4294967295.0));
// params.p_dropout_in_uint16_t = uint16_t(std::floor(params.p_dropout * 65535.0));
params.p_dropout_in_uint8_t = uint8_t(std::floor(params.p_dropout * 255.0));
params.rp_dropout = 1.f / params.p_dropout;
params.scale_softmax_rp_dropout = params.rp_dropout * params.scale_softmax;
TORCH_CHECK(p_dropout < 1.f);
// Causal is the special case where window_size_right == 0 and window_size_left < 0.
// Local is the more general case where window_size_right >= 0 or window_size_left >= 0.
params.is_causal = window_size_left < 0 && window_size_right == 0;
if (window_size_left < 0 && window_size_right >= 0) { window_size_left = seqlen_k; }
if (window_size_left >= 0 && window_size_right < 0) { window_size_right = seqlen_k; }
params.window_size_left = window_size_left;
params.window_size_right = window_size_right;
params.is_seqlens_k_cumulative = true;
}
void set_params_dgrad(Flash_bwd_params &params,
// sizes
const size_t b,
const size_t seqlen_q,
const size_t seqlen_k,
const size_t seqlen_q_rounded,
const size_t seqlen_k_rounded,
const size_t h,
const size_t h_k,
const size_t d,
const size_t d_rounded,
// device pointers
const at::Tensor q,
const at::Tensor k,
const at::Tensor v,
const at::Tensor out,
const at::Tensor dout,
at::Tensor dq,
at::Tensor dk,
at::Tensor dv,
void *cu_seqlens_q_d,
void *cu_seqlens_k_d,
void *dq_accum_d,
void *dk_accum_d,
void *dv_accum_d,
void *softmax_lse_d,
void *dsoftmax_sum_d,
float p_dropout,
float softmax_scale,
int window_size_left,
int window_size_right,
bool deterministic) {
set_params_fprop(params,
b, seqlen_q, seqlen_k, seqlen_q_rounded, seqlen_k_rounded, h, h_k, d, d_rounded,
q, k, v, out,
cu_seqlens_q_d,
cu_seqlens_k_d,
nullptr,
nullptr,
softmax_lse_d,
p_dropout,
softmax_scale,
window_size_left,
window_size_right);
// Set the pointers and strides.
params.do_ptr = dout.data_ptr();
params.do_row_stride = dout.stride(-3);
params.do_head_stride = dout.stride(-2);
params.dq_ptr = dq.data_ptr();
params.dk_ptr = dk.data_ptr();
params.dv_ptr = dv.data_ptr();
params.dq_row_stride = dq.stride(-3);
params.dk_row_stride = dk.stride(-3);
params.dv_row_stride = dv.stride(-3);
params.dq_head_stride = dq.stride(-2);
params.dk_head_stride = dk.stride(-2);
params.dv_head_stride = dv.stride(-2);
if (cu_seqlens_q_d == nullptr) {
params.do_batch_stride = dout.stride(0);
params.dq_batch_stride = dq.stride(0);
params.dk_batch_stride = dk.stride(0);
params.dv_batch_stride = dv.stride(0);
}
params.dq_accum_ptr = dq_accum_d;
params.dk_accum_ptr = dk_accum_d;
params.dv_accum_ptr = dv_accum_d;
// Softmax sum
params.dsoftmax_sum = dsoftmax_sum_d;
params.deterministic = deterministic;
}
void run_mha_fwd(Flash_fwd_params &params, cudaStream_t stream, bool force_split_kernel=false) {
FP16_SWITCH(!params.is_bf16, [&] {
FWD_HEADDIM_SWITCH(params.d, [&] {
if (params.num_splits <= 1 && !force_split_kernel) { // If we don't set it num_splits == 0
run_mha_fwd_<elem_type, kHeadDim>(params, stream);
} else {
run_mha_fwd_splitkv_dispatch<elem_type, kHeadDim>(params, stream);
}
});
});
}
void run_mha_fwd_block(Flash_fwd_params &params, cudaStream_t stream, bool force_split_kernel=false) {
FP16_SWITCH(!params.is_bf16, [&] {
FWD_BLOCK_HEADDIM_SWITCH(params.d, [&] {
if (params.num_splits <= 1 && !force_split_kernel) {
run_mha_fwd_block_<elem_type, kHeadDim>(params, stream);
}
});
});
}
// Find the number of splits that maximizes the occupancy. For example, if we have
// batch * n_heads = 48 and we have 108 SMs, having 2 splits (efficiency = 0.89) is
// better than having 3 splits (efficiency = 0.67). However, we also don't want too many
// splits as that would incur more HBM reads/writes.
// So we find the best efficiency, then find the smallest number of splits that gets 85%
// of the best efficiency.
inline int num_splits_heuristic(int batch_nheads_mblocks, int num_SMs, int num_n_blocks, int max_splits) {
// If we have enough to almost fill the SMs, then just use 1 split
if (batch_nheads_mblocks >= 0.8f * num_SMs) { return 1; }
max_splits = std::min({max_splits, num_SMs, num_n_blocks});
float max_efficiency = 0.f;
std::vector<float> efficiency;
efficiency.reserve(max_splits);
auto ceildiv = [](int a, int b) { return (a + b - 1) / b; };
// Some splits are not eligible. For example, if we have 64 blocks and choose 11 splits,
// we'll have 6 * 10 + 4 blocks. If we choose 12 splits, we'll have 6 * 11 + (-2) blocks
// (i.e. it's 11 splits anyway).
// So we check if the number of blocks per split is the same as the previous num_splits.
auto is_split_eligible = [&ceildiv, &num_n_blocks](int num_splits) {
return num_splits == 1 || ceildiv(num_n_blocks, num_splits) != ceildiv(num_n_blocks, num_splits - 1);
};
for (int num_splits = 1; num_splits <= max_splits; num_splits++) {
if (!is_split_eligible(num_splits)) {
efficiency.push_back(0.f);
} else {
float n_waves = float(batch_nheads_mblocks * num_splits) / num_SMs;
float eff = n_waves / ceil(n_waves);
// printf("num_splits = %d, eff = %f\n", num_splits, eff);
if (eff > max_efficiency) { max_efficiency = eff; }
efficiency.push_back(eff);
}
}
for (int num_splits = 1; num_splits <= max_splits; num_splits++) {
if (!is_split_eligible(num_splits)) { continue; }
if (efficiency[num_splits - 1] >= 0.85 * max_efficiency) {
// printf("num_splits chosen = %d\n", num_splits);
return num_splits;
}
}
return 1;
}
std::vector<at::Tensor>
mha_fwd(at::Tensor &q, // batch_size x seqlen_q x num_heads x head_size
const at::Tensor &k, // batch_size x seqlen_k x num_heads_k x head_size
const at::Tensor &v, // batch_size x seqlen_k x num_heads_k x head_size
c10::optional<at::Tensor> &out_, // batch_size x seqlen_q x num_heads x head_size
c10::optional<at::Tensor> &alibi_slopes_, // num_heads or batch_size x num_heads
const float p_dropout,
const float softmax_scale,
bool is_causal,
int window_size_left,
int window_size_right,
const bool return_softmax,
c10::optional<at::Generator> gen_) {
auto dprops = at::cuda::getCurrentDeviceProperties();
// bool is_sm75 = dprops->major == 7 && dprops->minor == 5;
bool is_sm8x = dprops->major == 8 && dprops->minor >= 0;
bool is_sm90 = dprops->major == 9 && dprops->minor == 0;
TORCH_CHECK(is_sm90 || is_sm8x, "FlashAttention only supports Ampere GPUs or newer.");
// We will support Turing in the near future
// TORCH_CHECK(is_sm90 || is_sm8x || is_sm75, "FlashAttention only supports Turing GPUs or newer.");
auto q_dtype = q.dtype();
TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kBFloat16,
"FlashAttention only support fp16 and bf16 data type");
if (q_dtype == torch::kBFloat16) {
TORCH_CHECK(is_sm90 || is_sm8x, "bfloat16 is only supported on Ampere GPUs or newer");
}
TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");
CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);
TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
const auto sizes = q.sizes();
const int batch_size = sizes[0];
int seqlen_q = sizes[1];
int num_heads = sizes[2];
const int head_size_og = sizes[3];
const int seqlen_k = k.size(1);
const int num_heads_k = k.size(2);
TORCH_CHECK(batch_size > 0, "batch size must be postive");
TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
if (window_size_left >= seqlen_k) { window_size_left = -1; }
if (window_size_right >= seqlen_k) { window_size_right = -1; }
// causal=true is the same as causal=false in this case
if (seqlen_q == 1 && !alibi_slopes_.has_value()) { is_causal = false; }
if (is_causal) { window_size_right = 0; }
// Faster to transpose q from (b, 1, (nheads_kv ngroups), d) to (b, ngroups, nheads_kv, d) in this case
// H/t Daniel Haziza
const int seqlenq_ngroups_swapped = seqlen_q == 1 && num_heads > num_heads_k && window_size_left < 0 && window_size_right < 0 && p_dropout == 0.f && head_size_og % 8 == 0 && !alibi_slopes_.has_value();
if (seqlenq_ngroups_swapped) {
const int ngroups = num_heads / num_heads_k;
q = q.reshape({batch_size, num_heads_k, ngroups, head_size_og}).transpose(1, 2);
seqlen_q = ngroups;
num_heads = num_heads_k;
}
CHECK_SHAPE(q, batch_size, seqlen_q, num_heads, head_size_og);
CHECK_SHAPE(k, batch_size, seqlen_k, num_heads_k, head_size_og);
CHECK_SHAPE(v, batch_size, seqlen_k, num_heads_k, head_size_og);
at::Tensor q_padded, k_padded, v_padded;
if (head_size_og % 8 != 0) {
q_padded = torch::nn::functional::pad(q, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
k_padded = torch::nn::functional::pad(k, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
v_padded = torch::nn::functional::pad(v, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
} else {
q_padded = q;
k_padded = k;
v_padded = v;
}
at::Tensor out;
if (out_.has_value()) {
out = out_.value();
TORCH_CHECK(out.dtype() == q_dtype, "Output must have the same dtype as inputs");
CHECK_DEVICE(out);
TORCH_CHECK(out.stride(-1) == 1, "Output tensor must have contiguous last dimension");
CHECK_SHAPE(out, batch_size, seqlen_q, num_heads, head_size_og);
if (head_size_og % 8 != 0) { out = torch::empty_like(q_padded); }
} else {
out = torch::empty_like(q_padded);
}
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
const int head_size = round_multiple(head_size_og, 8);
const int head_size_rounded = round_multiple(head_size, 32);
const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
const int seqlen_k_rounded = round_multiple(seqlen_k, 128);
// Otherwise the kernel will be launched from cuda:0 device
// Cast to char to avoid compiler warning about narrowing
at::cuda::CUDAGuard device_guard{(char)q.get_device()};
auto opts = q.options();
auto softmax_lse = torch::empty({batch_size, num_heads, seqlen_q}, opts.dtype(at::kFloat));
at::Tensor p;
// Only return softmax if there's dropout to reduce compilation time
if (return_softmax) {
TORCH_CHECK(p_dropout > 0.0f, "return_softmax is only supported when p_dropout > 0.0");
p = torch::empty({ batch_size, num_heads, seqlen_q_rounded, seqlen_k_rounded }, opts);
}
Flash_fwd_params params;
set_params_fprop(params,
batch_size,
seqlen_q, seqlen_k,
seqlen_q_rounded, seqlen_k_rounded,
num_heads, num_heads_k,
head_size, head_size_rounded,
q_padded, k_padded, v_padded, out,
/*cu_seqlens_q_d=*/nullptr,
/*cu_seqlens_k_d=*/nullptr,
/*seqused_k=*/nullptr,
return_softmax ? p.data_ptr() : nullptr,
softmax_lse.data_ptr(),
p_dropout,
softmax_scale,
window_size_left,
window_size_right);
// This needs to match with run_mha_fwd_splitkv_dispatch
const int block_n = head_size <= 64 ? 256 : (head_size <= 128 ? 128 : 64);
const int num_n_blocks = (seqlen_k + block_n - 1) / block_n;
// Technically kBlockM = 64 only for the splitKV kernels, not the standard kernel.
// In any case we don't expect seqlen_q to be larger than 64 for inference.
const int num_m_blocks = (seqlen_q + 64 - 1) / 64;
params.num_splits = 1;
if (p_dropout == 0.0f) { // SplitKV is not implemented for dropout
params.num_splits = num_splits_heuristic(batch_size * num_heads * num_m_blocks, dprops->multiProcessorCount, num_n_blocks, 128);
if (params.num_splits > 1) {
at::Tensor softmax_lse_accum = torch::empty({params.num_splits, batch_size, num_heads, seqlen_q}, opts.dtype(at::kFloat));
at::Tensor out_accum = torch::empty({params.num_splits, batch_size, num_heads, seqlen_q, head_size_rounded}, opts.dtype(at::kFloat));
params.softmax_lseaccum_ptr = softmax_lse_accum.data_ptr();
params.oaccum_ptr = out_accum.data_ptr();
}
TORCH_CHECK(params.num_splits <= 128, "num_splits > 128 not supported");
}
// number of times random will be generated per thread, to offset philox counter in thc random
// state
// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = params.b * params.h * 32;
auto options = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
auto rng_state = torch::empty({2}, options.dtype(torch::kInt64));
// Forward kernel will populate memory with the seed and offset.
params.rng_state = reinterpret_cast<uint64_t*>(rng_state.data_ptr());
if (p_dropout > 0.0) {
auto gen = at::get_generator_or_default<at::CUDAGeneratorImpl>(
gen_, at::cuda::detail::getDefaultCUDAGenerator());
// See Note [Acquire lock when using random generators]
std::lock_guard<std::mutex> lock(gen->mutex_);
params.philox_args = gen->philox_cuda_state(counter_offset);
}
if (alibi_slopes_.has_value()) {
auto alibi_slopes = alibi_slopes_.value();
TORCH_CHECK(alibi_slopes.dtype() == torch::kFloat32, "ALiBi slopes must have dtype fp32");
CHECK_DEVICE(alibi_slopes);
TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
TORCH_CHECK(alibi_slopes.sizes() == torch::IntArrayRef({num_heads}) || alibi_slopes.sizes() == torch::IntArrayRef({batch_size, num_heads}));
params.alibi_slopes_ptr = alibi_slopes.data_ptr();
params.alibi_slopes_batch_stride = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
} else {
params.alibi_slopes_ptr = nullptr;
}
if (seqlen_k > 0) {
auto stream = at::cuda::getCurrentCUDAStream().stream();
run_mha_fwd(params, stream);
} else {
// If seqlen_k == 0, then we have an empty tensor. We need to set the output to 0.
out.zero_();
softmax_lse.fill_(std::numeric_limits<float>::infinity());
}
at::Tensor out_padded = out;
if (head_size_og % 8 != 0) {
out = out.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
if (out_.has_value()) { out_.value().copy_(out); }
}
if (seqlenq_ngroups_swapped) {
out = out.transpose(1, 2).reshape({batch_size, 1, num_heads_k * seqlen_q, head_size_og});
out_padded = out_padded.transpose(1, 2).reshape({batch_size, 1, num_heads_k * seqlen_q, head_size_og});
q_padded = q_padded.transpose(1, 2).reshape({batch_size, 1, num_heads_k * seqlen_q, head_size_og});
softmax_lse = softmax_lse.reshape({batch_size, num_heads_k * seqlen_q, 1});
}
return {out, q_padded, k_padded, v_padded, out_padded, softmax_lse, p, rng_state};
}
std::vector<at::Tensor>
mha_varlen_fwd(const at::Tensor &q, // total_q x num_heads x head_size, total_q := \sum_{i=0}^{b} s_i
const at::Tensor &k, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &v, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
c10::optional<at::Tensor> &out_, // total_q x num_heads x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &cu_seqlens_q, // b+1
const at::Tensor &cu_seqlens_k, // b+1
c10::optional<at::Tensor> &seqused_k, // b. If given, only this many elements of each batch element's keys are used.
c10::optional<at::Tensor> &alibi_slopes_, // num_heads or b x num_heads
const int max_seqlen_q,
const int max_seqlen_k,
const float p_dropout,
const float softmax_scale,
const bool zero_tensors,
const bool is_causal,
int window_size_left,
int window_size_right,
const bool return_softmax,
c10::optional<at::Generator> gen_) {
if (is_causal) { window_size_right = 0; }
auto dprops = at::cuda::getCurrentDeviceProperties();
// bool is_sm75 = dprops->major == 7 && dprops->minor == 5;
bool is_sm8x = dprops->major == 8 && dprops->minor >= 0;
bool is_sm90 = dprops->major == 9 && dprops->minor == 0;
TORCH_CHECK(is_sm90 || is_sm8x, "FlashAttention only supports Ampere GPUs or newer.");
// We will support Turing in the near future
// TORCH_CHECK(is_sm90 || is_sm8x || is_sm75, "FlashAttention only supports Turing GPUs or newer.");
auto q_dtype = q.dtype();
TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kBFloat16,
"FlashAttention only support fp16 and bf16 data type");
if (q_dtype == torch::kBFloat16) {
TORCH_CHECK(is_sm90 || is_sm8x, "bfloat16 is only supported on Ampere GPUs or newer");
}
TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");
TORCH_CHECK(cu_seqlens_q.dtype() == torch::kInt32, "cu_seqlens_q must have dtype int32");
TORCH_CHECK(cu_seqlens_k.dtype() == torch::kInt32, "cu_seqlens_k must have dtype int32");
CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);
CHECK_DEVICE(cu_seqlens_q);
CHECK_DEVICE(cu_seqlens_k);
TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
CHECK_CONTIGUOUS(cu_seqlens_q);
CHECK_CONTIGUOUS(cu_seqlens_k);
const auto sizes = q.sizes();
const int total_q = sizes[0];
const int batch_size = cu_seqlens_q.numel() - 1;
const int num_heads = sizes[1];
const int head_size_og = sizes[2];
const int total_k = k.size(0);
const int num_heads_k = k.size(1);
TORCH_CHECK(batch_size > 0, "batch size must be positive");
TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
if (window_size_left >= max_seqlen_k) { window_size_left = -1; }
if (window_size_right >= max_seqlen_k) { window_size_right = -1; }
CHECK_SHAPE(q, total_q, num_heads, head_size_og);
CHECK_SHAPE(k, total_k, num_heads_k, head_size_og);
CHECK_SHAPE(v, total_k, num_heads_k, head_size_og);
CHECK_SHAPE(cu_seqlens_q, batch_size + 1);
CHECK_SHAPE(cu_seqlens_k, batch_size + 1);
if (seqused_k.has_value()){
auto seqused_k_ = seqused_k.value();
TORCH_CHECK(seqused_k_.dtype() == torch::kInt32, "seqused_k must have dtype int32");
TORCH_CHECK(seqused_k_.is_cuda(), "seqused_k must be on CUDA device");
TORCH_CHECK(seqused_k_.is_contiguous(), "seqused_k must be contiguous");
CHECK_SHAPE(seqused_k_, batch_size);
}
at::Tensor q_padded, k_padded, v_padded;
if (head_size_og % 8 != 0) {
q_padded = torch::nn::functional::pad(q, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
k_padded = torch::nn::functional::pad(k, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
v_padded = torch::nn::functional::pad(v, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
} else {
q_padded = q;
k_padded = k;
v_padded = v;
}
at::Tensor out;
if (out_.has_value()) {
out = out_.value();
TORCH_CHECK(out.dtype() == q_dtype, "Output must have the same dtype as inputs");
CHECK_DEVICE(out);
TORCH_CHECK(out.stride(-1) == 1, "Output tensor must have contiguous last dimension");
CHECK_SHAPE(out, total_q, num_heads, head_size_og);
if (head_size_og % 8 != 0) { out = torch::empty_like(q_padded); }
} else {
out = torch::empty_like(q_padded);
}
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
const int head_size = round_multiple(head_size_og, 8);
const int head_size_rounded = round_multiple(head_size, 32);
const int seqlen_q_rounded = round_multiple(max_seqlen_q, 128);
const int seqlen_k_rounded = round_multiple(max_seqlen_k, 128);
// Otherwise the kernel will be launched from cuda:0 device
// Cast to char to avoid compiler warning about narrowing
at::cuda::CUDAGuard device_guard{(char)q.get_device()};
auto opts = q.options();
auto softmax_lse = torch::empty({batch_size, num_heads, max_seqlen_q}, opts.dtype(at::kFloat));
at::Tensor p;
// Only return softmax if there's dropout to reduce compilation time
if (return_softmax) {
TORCH_CHECK(p_dropout > 0.0f, "return_softmax is only supported when p_dropout > 0.0");
p = torch::empty({ batch_size, num_heads, seqlen_q_rounded, seqlen_k_rounded }, opts);
}
if (zero_tensors) {
out.zero_();
softmax_lse.fill_(-std::numeric_limits<float>::infinity());
if (return_softmax) {p.zero_();}
}
Flash_fwd_params params;
set_params_fprop(params,
batch_size,
max_seqlen_q, max_seqlen_k,
seqlen_q_rounded, seqlen_k_rounded,
num_heads, num_heads_k,
head_size, head_size_rounded,
q_padded, k_padded, v_padded, out,
cu_seqlens_q.data_ptr(),
cu_seqlens_k.data_ptr(),
seqused_k.has_value() ? seqused_k.value().data_ptr() : nullptr,
return_softmax ? p.data_ptr() : nullptr,
softmax_lse.data_ptr(),
p_dropout,
softmax_scale,
window_size_left,
window_size_right);
// number of times random will be generated per thread, to offset philox counter in thc random
// state
// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = params.b * params.h * 32;
auto options = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
auto rng_state = torch::empty({2}, options.dtype(torch::kInt64));
// Forward kernel will populate memory with the seed and offset.
params.rng_state = reinterpret_cast<uint64_t*>(rng_state.data_ptr());
if (p_dropout > 0.0) {
auto gen = at::get_generator_or_default<at::CUDAGeneratorImpl>(
gen_, at::cuda::detail::getDefaultCUDAGenerator());
// See Note [Acquire lock when using random generators]
std::lock_guard<std::mutex> lock(gen->mutex_);
params.philox_args = gen->philox_cuda_state(counter_offset);
}
if (alibi_slopes_.has_value()) {
auto alibi_slopes = alibi_slopes_.value();
TORCH_CHECK(alibi_slopes.dtype() == torch::kFloat32, "ALiBi slopes must have dtype fp32");
CHECK_DEVICE(alibi_slopes);
TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
TORCH_CHECK(alibi_slopes.sizes() == torch::IntArrayRef({num_heads}) || alibi_slopes.sizes() == torch::IntArrayRef({batch_size, num_heads}));
params.alibi_slopes_ptr = alibi_slopes.data_ptr();
params.alibi_slopes_batch_stride = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
} else {
params.alibi_slopes_ptr = nullptr;
}
if (max_seqlen_k > 0) {
auto stream = at::cuda::getCurrentCUDAStream().stream();
run_mha_fwd(params, stream);
} else {
// If seqlen_k == 0, then we have an empty tensor. We need to set the output to 0.
out.zero_();
softmax_lse.fill_(std::numeric_limits<float>::infinity());
}
at::Tensor out_padded = out;
if (head_size_og % 8 != 0) {
out = out.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
if (out_.has_value()) { out_.value().copy_(out); }
}
return {out, q_padded, k_padded, v_padded, out_padded, softmax_lse, p, rng_state};
}
void run_mha_bwd(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
FP16_SWITCH(!params.is_bf16, [&] {
if (params.d <= 32) {
run_mha_bwd_<elem_type, 32>(params, stream, configure);
} else if (params.d <= 64) {
run_mha_bwd_<elem_type, 64>(params, stream, configure);
} else if (params.d <= 96) {
run_mha_bwd_<elem_type, 96>(params, stream, configure);
} else if (params.d <= 128) {
run_mha_bwd_<elem_type, 128>(params, stream, configure);
} else if (params.d <= 160) {
run_mha_bwd_<elem_type, 160>(params, stream, configure);
} else if (params.d <= 192) {
run_mha_bwd_<elem_type, 192>(params, stream, configure);
} else if (params.d <= 224) {
run_mha_bwd_<elem_type, 224>(params, stream, configure);
} else if (params.d <= 256) {
run_mha_bwd_<elem_type, 256>(params, stream, configure);
}
});
}
void run_mha_bwd_block(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
FP16_SWITCH(!params.is_bf16, [&] {
BWD_BLOCK_HEADDIM_SWITCH(params.d, [&] {
run_mha_bwd_block_<elem_type, kHeadDim>(params, stream, configure);
});
});
}
std::vector<at::Tensor>
mha_bwd(const at::Tensor &dout, // batch_size x seqlen_q x num_heads, x head_size_og
const at::Tensor &q, // batch_size x seqlen_q x num_heads x head_size
const at::Tensor &k, // batch_size x seqlen_k x num_heads_k x head_size
const at::Tensor &v, // batch_size x seqlen_k x num_heads_k x head_size
const at::Tensor &out, // batch_size x seqlen_q x num_heads x head_size
const at::Tensor &softmax_lse, // b x h x seqlen_q
c10::optional<at::Tensor> &dq_, // batch_size x seqlen_q x num_heads x head_size
c10::optional<at::Tensor> &dk_, // batch_size x seqlen_k x num_heads_k x head_size
c10::optional<at::Tensor> &dv_, // batch_size x seqlen_k x num_heads_k x head_size
c10::optional<at::Tensor> &alibi_slopes_, // num_heads or batch_size x num_heads
const float p_dropout, // probability to drop
const float softmax_scale,
const bool is_causal,
int window_size_left,
int window_size_right,
const bool deterministic,
c10::optional<at::Generator> gen_,
c10::optional<at::Tensor> &rng_state) {
if (is_causal) { window_size_right = 0; }
auto dprops = at::cuda::getCurrentDeviceProperties();
// bool is_sm75 = dprops->major == 7 && dprops->minor == 5;
bool is_sm8x = dprops->major == 8 && dprops->minor >= 0;
bool is_sm80 = dprops->major == 8 && dprops->minor == 0;
bool is_sm90 = dprops->major == 9 && dprops->minor == 0;
TORCH_CHECK(is_sm90 || is_sm8x, "FlashAttention only supports Ampere GPUs or newer.");
// We will support Turing in the near future
// TORCH_CHECK(is_sm90 || is_sm8x || is_sm75, "FlashAttention only supports Turing GPUs or newer.");
bool is_dropout = p_dropout > 0.0;
auto stream = at::cuda::getCurrentCUDAStream().stream();
auto q_dtype = q.dtype();
TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kBFloat16,
"FlashAttention only support fp16 and bf16 data type");
if (q_dtype == torch::kBFloat16) {
TORCH_CHECK(is_sm90 || is_sm8x, "bfloat16 is only supported on Ampere GPUs or newer");
}
TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");
TORCH_CHECK(out.dtype() == q_dtype, "query and out must have the same dtype");
TORCH_CHECK(dout.dtype() == q_dtype, "query and dout must have the same dtype");
CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);
CHECK_DEVICE(out); CHECK_DEVICE(dout); CHECK_DEVICE(softmax_lse);
TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(out.stride(-1) == 1, "out tensor must have contiguous last dimension");
TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
const auto sizes = q.sizes();
const int batch_size = sizes[0];
const int seqlen_q = sizes[1];
const int num_heads = sizes[2];
const int head_size_og = dout.size(3);
const int head_size = sizes[3];
const int seqlen_k = k.size(1);
const int num_heads_k = k.size(2);
TORCH_CHECK(batch_size > 0, "batch size must be positive");
TORCH_CHECK(head_size % 8 == 0, "head_size should be a multiple of 8");
TORCH_CHECK(head_size <= 256, "FlashAttention backward only supports head dimension at most 256");
if (head_size > 192) {
TORCH_CHECK(is_sm80 || is_sm90, "FlashAttention backward for head dim > 192 requires A100/A800 or H100/H800");
}
TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
const int head_size_rounded = round_multiple(head_size, 32);
const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
const int seqlen_k_rounded = round_multiple(seqlen_k, 128);
TORCH_CHECK(head_size == round_multiple(head_size_og, 8), "head_size must be head_size_og rounded to a multiple of 8");
if (window_size_left >= seqlen_k) { window_size_left = -1; }
if (window_size_right >= seqlen_k) { window_size_right = -1; }
CHECK_SHAPE(q, batch_size, seqlen_q, num_heads, head_size);
CHECK_SHAPE(k, batch_size, seqlen_k, num_heads_k, head_size);
CHECK_SHAPE(v, batch_size, seqlen_k, num_heads_k, head_size);
CHECK_SHAPE(out, batch_size, seqlen_q, num_heads, head_size);
CHECK_SHAPE(dout, batch_size, seqlen_q, num_heads, head_size_og);
at::Tensor dq, dk, dv;
if (dq_.has_value()) {
dq = dq_.value();
TORCH_CHECK(dq.dtype() == q_dtype, "dq must have the same dtype as q");
CHECK_DEVICE(dq);
TORCH_CHECK(dq.stride(-1) == 1, "dq must have contiguous last dimension");
CHECK_SHAPE(dq, batch_size, seqlen_q, num_heads, head_size);
} else {
dq = torch::empty_like(q);
}
if (dk_.has_value()) {
dk = dk_.value();
TORCH_CHECK(dk.dtype() == q_dtype, "dk must have the same dtype as q");
CHECK_DEVICE(dk);
TORCH_CHECK(dk.stride(-1) == 1, "dk must have contiguous last dimension");
CHECK_SHAPE(dk, batch_size, seqlen_k, num_heads_k, head_size);
} else {
dk = torch::empty_like(k);
}
if (dv_.has_value()) {
dv = dv_.value();
TORCH_CHECK(dv.dtype() == q_dtype, "dv must have the same dtype as q");
CHECK_DEVICE(dv);
TORCH_CHECK(dv.stride(-1) == 1, "dv must have contiguous last dimension");
CHECK_SHAPE(dv, batch_size, seqlen_k, num_heads_k, head_size);
} else {
dv = torch::empty_like(k);
}
at::Tensor dout_padded;
if (head_size_og % 8 != 0) {
dout_padded = torch::nn::functional::pad(dout, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
} else {
dout_padded = dout;
}
// bool loop = seqlen_k > blocksize_c;
// TODO: change later, for now set to true for simplicity
bool loop = true;
// Otherwise the kernel will be launched from cuda:0 device
// Cast to char to avoid compiler warning about narrowing
at::cuda::CUDAGuard device_guard{(char)q.get_device()};
auto opts = q.options();
auto softmax_d = torch::empty({batch_size, num_heads, seqlen_q_rounded}, opts.dtype(at::kFloat));
at::Tensor dq_accum;
at::Tensor dk_accum, dv_accum;
if (loop) {
if (!deterministic) {
dq_accum = torch::empty({batch_size, seqlen_q_rounded, num_heads, head_size_rounded}, opts.dtype(at::kFloat));
} else {
const int nsplits = (dprops->multiProcessorCount + batch_size * num_heads - 1) / (batch_size * num_heads);
dq_accum = torch::zeros({nsplits, batch_size, seqlen_q_rounded, num_heads, head_size_rounded}, opts.dtype(at::kFloat));
}
// dk_accum = torch::empty({batch_size, num_heads_k, seqlen_k_rounded, head_size_rounded}, opts.dtype(at::kFloat));
// dv_accum = torch::empty({batch_size, num_heads_k, seqlen_k_rounded, head_size_rounded}, opts.dtype(at::kFloat));
}
at::Tensor dk_expanded, dv_expanded;
if (num_heads_k != num_heads) { // MQA / GQA
dk_expanded = torch::empty({batch_size, seqlen_k, num_heads, head_size}, opts);
dv_expanded = torch::empty({batch_size, seqlen_k, num_heads, head_size}, opts);
} else {
dk_expanded = dk;
dv_expanded = dv;
}
Flash_bwd_params params;
set_params_dgrad(params,
batch_size,
seqlen_q, seqlen_k,
seqlen_q_rounded, seqlen_k_rounded,
num_heads, num_heads_k,
head_size, head_size_rounded,
q, k, v, out,
dout_padded, dq, dk_expanded, dv_expanded,
nullptr,
nullptr,
loop ? dq_accum.data_ptr() : nullptr,
// loop ? dk_accum.data_ptr() : nullptr,
// loop ? dv_accum.data_ptr() : nullptr,
nullptr,
nullptr,
softmax_lse.data_ptr(),
softmax_d.data_ptr(),
p_dropout,
softmax_scale,
window_size_left,
window_size_right,
deterministic);
params.dq_accum_split_stride = !deterministic ? 0 : dq_accum.stride(0);
auto launch = &run_mha_bwd;
// launch(params, stream, /*configure=*/true);
auto gen = at::get_generator_or_default<at::CUDAGeneratorImpl>(
gen_, at::cuda::detail::getDefaultCUDAGenerator());
// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = params.b * params.h * 32;
if ( rng_state.has_value() ) {
params.rng_state = reinterpret_cast<uint64_t*>(rng_state.value().data_ptr());
} else if( is_dropout ) {
// See Note [Acquire lock when using random generators]
std::lock_guard<std::mutex> lock(gen->mutex_);
params.philox_args = gen->philox_cuda_state(counter_offset);
auto seeds = at::cuda::philox::unpack(params.philox_args);
params.rng_state[0] = std::get<0>(seeds);
params.rng_state[1] = std::get<1>(seeds);
}
if (alibi_slopes_.has_value()) {
auto alibi_slopes = alibi_slopes_.value();
TORCH_CHECK(alibi_slopes.dtype() == torch::kFloat32, "ALiBi slopes must have dtype fp32");
CHECK_DEVICE(alibi_slopes);
TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
TORCH_CHECK(alibi_slopes.sizes() == torch::IntArrayRef({num_heads}) || alibi_slopes.sizes() == torch::IntArrayRef({batch_size, num_heads}));
params.alibi_slopes_ptr = alibi_slopes.data_ptr();
params.alibi_slopes_batch_stride = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
} else {
params.alibi_slopes_ptr = nullptr;
}
if (seqlen_q > 0) {
launch(params, stream, /*configure=*/false);
} else {
// If seqlen_q == 0, then we have an empty tensor. We need to set the output to 0.
dk_expanded.zero_();
dv_expanded.zero_();
softmax_d.zero_();
}
// For MQA/GQA we need to sum dK and dV across the groups
if (num_heads_k != num_heads) {
at::sum_out(dk, at::reshape(dk_expanded, {batch_size, seqlen_k, num_heads_k, num_heads / num_heads_k, head_size}), {3});
at::sum_out(dv, at::reshape(dv_expanded, {batch_size, seqlen_k, num_heads_k, num_heads / num_heads_k, head_size}), {3});
}
if (head_size_og % 8 != 0) {
dq = dq.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
dk = dk.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
dv = dv.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
}
return { dq, dk, dv, softmax_d };
}
std::vector<at::Tensor>
mha_varlen_bwd(const at::Tensor &dout, // total_q x num_heads, x head_size
const at::Tensor &q, // total_q x num_heads x head_size, total_q := \sum_{i=0}^{b} s_i
const at::Tensor &k, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &v, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &out, // total_q x num_heads x head_size
const at::Tensor &softmax_lse, // b x h x s softmax logsumexp
c10::optional<at::Tensor> &dq_, // total_q x num_heads x head_size, total_q := \sum_{i=0}^{b} s_i
c10::optional<at::Tensor> &dk_, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
c10::optional<at::Tensor> &dv_, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &cu_seqlens_q, // b+1
const at::Tensor &cu_seqlens_k, // b+1
c10::optional<at::Tensor> &alibi_slopes_, // num_heads or b x num_heads
const int max_seqlen_q,
const int max_seqlen_k, // max sequence length to choose the kernel
const float p_dropout, // probability to drop
const float softmax_scale,
const bool zero_tensors,
const bool is_causal,
int window_size_left,
int window_size_right,
const bool deterministic,
c10::optional<at::Generator> gen_,
c10::optional<at::Tensor> &rng_state) {
if (is_causal) { window_size_right = 0; }
auto dprops = at::cuda::getCurrentDeviceProperties();
// bool is_sm75 = dprops->major == 7 && dprops->minor == 5;
bool is_sm8x = dprops->major == 8 && dprops->minor >= 0;
bool is_sm80 = dprops->major == 8 && dprops->minor == 0;
bool is_sm90 = dprops->major == 9 && dprops->minor == 0;
TORCH_CHECK(is_sm90 || is_sm8x, "FlashAttention only supports Ampere GPUs or newer.");
// We will support Turing in the near future
// TORCH_CHECK(is_sm90 || is_sm8x || is_sm75, "FlashAttention only supports Turing GPUs or newer.");
bool is_dropout = p_dropout > 0.0;
auto stream = at::cuda::getCurrentCUDAStream().stream();
auto q_dtype = q.dtype();
TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kBFloat16,
"FlashAttention only support fp16 and bf16 data type");
if (q_dtype == torch::kBFloat16) {
TORCH_CHECK(is_sm90 || is_sm8x, "bfloat16 is only supported on Ampere GPUs or newer");
}
TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");
TORCH_CHECK(out.dtype() == q_dtype, "query and out must have the same dtype");
TORCH_CHECK(dout.dtype() == q_dtype, "query and dout must have the same dtype");
TORCH_CHECK(cu_seqlens_q.dtype() == torch::kInt32, "cu_seqlens_q must have dtype int32");
TORCH_CHECK(cu_seqlens_k.dtype() == torch::kInt32, "cu_seqlens_k must have dtype int32");
CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);
CHECK_DEVICE(out); CHECK_DEVICE(dout); CHECK_DEVICE(softmax_lse);
CHECK_DEVICE(cu_seqlens_q); CHECK_DEVICE(cu_seqlens_k);
TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(out.stride(-1) == 1, "out tensor must have contiguous last dimension");
TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
CHECK_CONTIGUOUS(cu_seqlens_q);
CHECK_CONTIGUOUS(cu_seqlens_k);
const auto sizes = q.sizes();
const int total_q = sizes[0];
const int batch_size = cu_seqlens_q.numel() - 1;
const int num_heads = sizes[1];
const int head_size_og = dout.size(2);
const int head_size = sizes[2];
const int total_k = k.size(0);
const int num_heads_k = k.size(1);
TORCH_CHECK(batch_size > 0, "batch size must be positive");
TORCH_CHECK(head_size % 8 == 0, "head_size should be a multiple of 8");
TORCH_CHECK(head_size <= 256, "FlashAttention backward only supports head dimension at most 256");
if (head_size > 192) {
TORCH_CHECK(is_sm80 || is_sm90, "FlashAttention backward for head dim > 192 requires A100/A800 or H100/H800");
}
TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
const int head_size_rounded = round_multiple(head_size, 32);
const int seqlen_q_rounded = round_multiple(max_seqlen_q, 128);
const int seqlen_k_rounded = round_multiple(max_seqlen_k, 128);
TORCH_CHECK(head_size == round_multiple(head_size_og, 8), "head_size must be head_size_og rounded to a multiple of 8");
if (window_size_left >= max_seqlen_k) { window_size_left = -1; }
if (window_size_right >= max_seqlen_k) { window_size_right = -1; }
CHECK_SHAPE(q, total_q, num_heads, head_size);
CHECK_SHAPE(k, total_k, num_heads_k, head_size);
CHECK_SHAPE(v, total_k, num_heads_k, head_size);
CHECK_SHAPE(out, total_q, num_heads, head_size);
CHECK_SHAPE(dout, total_q, num_heads, head_size_og);
CHECK_SHAPE(cu_seqlens_q, batch_size + 1);
CHECK_SHAPE(cu_seqlens_k, batch_size + 1);
at::Tensor dq, dk, dv;
if (dq_.has_value()) {
dq = dq_.value();
TORCH_CHECK(dq.dtype() == q_dtype, "dq must have the same dtype as q");
CHECK_DEVICE(dq);
TORCH_CHECK(dq.stride(-1) == 1, "dq must have contiguous last dimension");
CHECK_SHAPE(dq, total_q, num_heads, head_size);
} else {
dq = torch::empty_like(q);
}
if (dk_.has_value()) {
dk = dk_.value();
TORCH_CHECK(dk.dtype() == q_dtype, "dk must have the same dtype as q");
CHECK_DEVICE(dk);
TORCH_CHECK(dk.stride(-1) == 1, "dk must have contiguous last dimension");
CHECK_SHAPE(dk, total_k, num_heads_k, head_size);
} else {
dk = torch::empty_like(k);
}
if (dv_.has_value()) {
dv = dv_.value();
TORCH_CHECK(dv.dtype() == q_dtype, "dv must have the same dtype as q");
CHECK_DEVICE(dv);
TORCH_CHECK(dv.stride(-1) == 1, "dv must have contiguous last dimension");
CHECK_SHAPE(dv, total_k, num_heads_k, head_size);
} else {
dv = torch::empty_like(v);
}
at::Tensor dout_padded;
if (head_size_og % 8 != 0) {
dout_padded = torch::nn::functional::pad(dout, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
} else {
dout_padded = dout;
}
// bool loop = max_seqlen_k > blocksize_c;
// TODO: change later, for now set to true for simplicity
bool loop = true;
// Otherwise the kernel will be launched from cuda:0 device
// Cast to char to avoid compiler warning about narrowing
at::cuda::CUDAGuard device_guard{(char)q.get_device()};
auto opts = q.options();
auto softmax_d = torch::empty({batch_size, num_heads, seqlen_q_rounded}, opts.dtype(at::kFloat));
at::Tensor dq_accum;
if (loop) {
// We don't want to allocate dq_accum of size (batch, seqlen_q_rounded, num_heads, head_size_rounded)
// because that would be too large if there is a very long sequence and the rest of the sequences are short.
// Instead, we allocate dq_accum of size (total_q + 128 * batch, num_heads, head_size_rounded).
// Note that 128 is the max block size on the seqlen_q dimension.
// For dQ, the i-th sequence is stored in indices from cu_seqlens[i] + 128 * i to
// cu_seqlens[i + 1] * 128 * i - 1. This ensures that the i-th sequence and (i + 1)-th sequence will
// be at least 128 apart. It's ok for us to do atomicAdds up to 128 rows beyond what we're normally
// allowed to do. So we won't have to do any bound checking, and performance should stay the same.
if (!deterministic) {
dq_accum = torch::empty({total_q + 128 * batch_size, num_heads, head_size_rounded}, opts.dtype(at::kFloat));
} else {
const int nsplits = (dprops->multiProcessorCount + batch_size * num_heads - 1) / (batch_size * num_heads);
dq_accum = torch::zeros({nsplits, total_q + 128 * batch_size, num_heads, head_size_rounded}, opts.dtype(at::kFloat));
}
}
at::Tensor dk_expanded, dv_expanded;
if (num_heads_k != num_heads) { // MQA / GQA
dk_expanded = torch::empty({total_k, num_heads, head_size}, opts);
dv_expanded = torch::empty({total_k, num_heads, head_size}, opts);
} else {
dk_expanded = dk;
dv_expanded = dv;
}
if( zero_tensors ) {
dq.zero_();
dk_expanded.zero_();
dv_expanded.zero_();
softmax_d.zero_();
}
Flash_bwd_params params;
set_params_dgrad(params,
batch_size,
max_seqlen_q, max_seqlen_k,
seqlen_q_rounded, seqlen_k_rounded,
num_heads, num_heads_k,
head_size, head_size_rounded,
q, k, v, out,
dout_padded, dq, dk_expanded, dv_expanded,
cu_seqlens_q.data_ptr(),
cu_seqlens_k.data_ptr(),
loop ? dq_accum.data_ptr() : nullptr,
nullptr,
nullptr,
softmax_lse.data_ptr(),
softmax_d.data_ptr(),
p_dropout,
softmax_scale,
window_size_left,
window_size_right,
deterministic);
params.dq_accum_split_stride = !deterministic ? 0 : dq_accum.stride(0);
auto launch = &run_mha_bwd;
// launch(params, stream, /*configure=*/true);
auto gen = at::get_generator_or_default<at::CUDAGeneratorImpl>(
gen_, at::cuda::detail::getDefaultCUDAGenerator());
// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = params.b * params.h * 32;
if ( rng_state.has_value() ) {
params.rng_state = reinterpret_cast<uint64_t*>(rng_state.value().data_ptr());
} else if( is_dropout ) {
// See Note [Acquire lock when using random generators]
std::lock_guard<std::mutex> lock(gen->mutex_);
params.philox_args = gen->philox_cuda_state(counter_offset);
auto seeds = at::cuda::philox::unpack(params.philox_args);
params.rng_state[0] = std::get<0>(seeds);
params.rng_state[1] = std::get<1>(seeds);
}
if (alibi_slopes_.has_value()) {
auto alibi_slopes = alibi_slopes_.value();
TORCH_CHECK(alibi_slopes.dtype() == torch::kFloat32, "ALiBi slopes must have dtype fp32");
CHECK_DEVICE(alibi_slopes);
TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
TORCH_CHECK(alibi_slopes.sizes() == torch::IntArrayRef({num_heads}) || alibi_slopes.sizes() == torch::IntArrayRef({batch_size, num_heads}));
params.alibi_slopes_ptr = alibi_slopes.data_ptr();
params.alibi_slopes_batch_stride = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
} else {
params.alibi_slopes_ptr = nullptr;
}
if (max_seqlen_q > 0) {
launch(params, stream, /*configure=*/false);
} else {
// If seqlen_q == 0, then we have an empty tensor. We need to set the output to 0.
dk_expanded.zero_();
dv_expanded.zero_();
softmax_d.zero_();
}
// For MQA/GQA we need to sum dK and dV across the groups
if (num_heads_k != num_heads) {
at::sum_out(dk, at::reshape(dk_expanded, {total_k, num_heads_k, num_heads / num_heads_k, head_size}), {2});
at::sum_out(dv, at::reshape(dv_expanded, {total_k, num_heads_k, num_heads / num_heads_k, head_size}), {2});
}
if (head_size_og % 8 != 0) {
dq = dq.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
dk = dk.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
dv = dv.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
}
return { dq, dk, dv, softmax_d };
}
std::vector<at::Tensor>
mha_fwd_kvcache(at::Tensor &q, // batch_size x seqlen_q x num_heads x head_size
const at::Tensor &kcache, // batch_size_c x seqlen_k x num_heads_k x head_size
const at::Tensor &vcache, // batch_size_c x seqlen_k x num_heads_k x head_size
c10::optional<const at::Tensor> &k_, // batch_size x seqlen_knew x num_heads_k x head_size
c10::optional<const at::Tensor> &v_, // batch_size x seqlen_knew x num_heads_k x head_size
c10::optional<const at::Tensor> &seqlens_k_, // batch_size
c10::optional<const at::Tensor> &rotary_cos_, // seqlen_ro x (rotary_dim / 2)
c10::optional<const at::Tensor> &rotary_sin_, // seqlen_ro x (rotary_dim / 2)
c10::optional<const at::Tensor> &cache_batch_idx_, // indices to index into the KV cache
c10::optional<at::Tensor> &alibi_slopes_, // num_heads or batch_size x num_heads
c10::optional<at::Tensor> &out_, // batch_size x seqlen_q x num_heads x head_size
const float softmax_scale,
bool is_causal,
int window_size_left,
int window_size_right,
bool is_rotary_interleaved, // if true, rotary combines indices 0 & 1, else indices 0 & rotary_dim / 2
int num_splits
) {
auto dprops = at::cuda::getCurrentDeviceProperties();
// bool is_sm75 = dprops->major == 7 && dprops->minor == 5;
bool is_sm8x = dprops->major == 8 && dprops->minor >= 0;
bool is_sm90 = dprops->major == 9 && dprops->minor == 0;
TORCH_CHECK(is_sm90 || is_sm8x, "FlashAttention only supports Ampere GPUs or newer.");
// We will support Turing in the near future
// TORCH_CHECK(is_sm90 || is_sm8x || is_sm75, "FlashAttention only supports Turing GPUs or newer.");
auto q_dtype = q.dtype();
TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kBFloat16,
"FlashAttention only support fp16 and bf16 data type");
if (q_dtype == torch::kBFloat16) {
TORCH_CHECK(is_sm90 || is_sm8x, "bfloat16 is only supported on Ampere GPUs or newer");
}
TORCH_CHECK(kcache.dtype() == q_dtype, "query and key must have the same dtype");
TORCH_CHECK(vcache.dtype() == q_dtype, "query and value must have the same dtype");
CHECK_DEVICE(q); CHECK_DEVICE(kcache); CHECK_DEVICE(vcache);
TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(kcache.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(vcache.stride(-1) == 1, "Input tensor must have contiguous last dimension");
const auto sizes = q.sizes();
const int batch_size = sizes[0];
int seqlen_q = sizes[1];
int num_heads = sizes[2];
const int head_size_og = sizes[3];
const int seqlen_k = kcache.size(1);
const int num_heads_k = kcache.size(2);
const int batch_size_c = kcache.size(0);
TORCH_CHECK(batch_size > 0, "batch size must be postive");
TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
// causal=true is the same as causal=false in this case
if (seqlen_q == 1 && !alibi_slopes_.has_value()) { is_causal = false; }
if (is_causal) { window_size_right = 0; }
// Faster to transpose q from (b, 1, (nheads_kv ngroups), d) to (b, ngroups, nheads_kv, d) in this case
// H/t Daniel Haziza
const int seqlenq_ngroups_swapped = seqlen_q == 1 && num_heads > num_heads_k && window_size_left < 0 && window_size_right < 0 && head_size_og % 8 == 0 && !alibi_slopes_.has_value();
if (seqlenq_ngroups_swapped) {
const int ngroups = num_heads / num_heads_k;
q = q.reshape({batch_size, num_heads_k, ngroups, head_size_og}).transpose(1, 2);
seqlen_q = ngroups;
num_heads = num_heads_k;
}
if (window_size_left >= seqlen_k) { window_size_left = -1; }
if (window_size_right >= seqlen_k) { window_size_right = -1; }
CHECK_SHAPE(q, batch_size, seqlen_q, num_heads, head_size_og);
CHECK_SHAPE(kcache, batch_size_c, seqlen_k, num_heads_k, head_size_og);
CHECK_SHAPE(vcache, batch_size_c, seqlen_k, num_heads_k, head_size_og);
at::Tensor q_padded, kcache_padded, vcache_padded;
if (head_size_og % 8 != 0) {
q_padded = torch::nn::functional::pad(q, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
kcache_padded = torch::nn::functional::pad(kcache, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
vcache_padded = torch::nn::functional::pad(vcache, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
} else {
q_padded = q;
kcache_padded = kcache;
vcache_padded = vcache;
}
at::Tensor out;
if (out_.has_value()) {
out = out_.value();
TORCH_CHECK(out.dtype() == q_dtype, "Output must have the same dtype as inputs");
CHECK_DEVICE(out);
TORCH_CHECK(out.stride(-1) == 1, "Output tensor must have contiguous last dimension");
CHECK_SHAPE(out, batch_size, seqlen_q, num_heads, head_size_og);
if (head_size_og % 8 != 0) { out = torch::empty_like(q_padded); }
} else {
out = torch::empty_like(q_padded);
}
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
const int head_size = round_multiple(head_size_og, 8);
const int head_size_rounded = round_multiple(head_size, 32);
const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
const int seqlen_k_rounded = round_multiple(seqlen_k, 128);
// Otherwise the kernel will be launched from cuda:0 device
// Cast to char to avoid compiler warning about narrowing
at::cuda::CUDAGuard device_guard{(char)q.get_device()};
auto opts = q.options();
auto softmax_lse = torch::empty({batch_size, num_heads, seqlen_q}, opts.dtype(at::kFloat));
Flash_fwd_params params;
set_params_fprop(params,
batch_size,
seqlen_q, seqlen_k,
seqlen_q_rounded, seqlen_k_rounded,
num_heads, num_heads_k,
head_size, head_size_rounded,
q_padded, kcache_padded, vcache_padded, out,
/*cu_seqlens_q_d=*/nullptr,
/*cu_seqlens_k_d=*/nullptr,
/*seqused_k=*/nullptr,
/*p_ptr=*/nullptr,
softmax_lse.data_ptr(),
/*p_dropout=*/0.f,
softmax_scale,
window_size_left,
window_size_right);
at::Tensor k, v, k_padded, v_padded;
if (k_.has_value()) {
TORCH_CHECK(v_.has_value(), "If key is supplied, value must also be passed in");
TORCH_CHECK(seqlens_k_.has_value(), "If key is supplied, seqlens_k must also be passed in");
TORCH_CHECK(seqlen_q <= seqlen_k, "If key is supplied, it must have seqlen <= the seqlen of the KV cache");
k = k_.value();
v = v_.value();
TORCH_CHECK(k.dtype() == q_dtype, "Key must have the same dtype as query");
TORCH_CHECK(v.dtype() == q_dtype, "Value must have the same dtype as query");
CHECK_DEVICE(k); CHECK_DEVICE(v);
TORCH_CHECK(k.stride(-1) == 1, "Key tensor must have contiguous last dimension");
TORCH_CHECK(v.stride(-1) == 1, "Value tensor must have contiguous last dimension");
int seqlen_knew = k.size(1);
CHECK_SHAPE(k, batch_size, seqlen_knew, num_heads_k, head_size_og);
CHECK_SHAPE(v, batch_size, seqlen_knew, num_heads_k, head_size_og);
if (head_size_og % 8 != 0) {
k_padded = torch::nn::functional::pad(k, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
v_padded = torch::nn::functional::pad(v, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
} else {
k_padded = k;
v_padded = v;
}
params.seqlen_knew = seqlen_knew;
params.knew_ptr = k_padded.data_ptr();
params.vnew_ptr = v_padded.data_ptr();
// All stride are in elements, not bytes.
params.knew_batch_stride = k_padded.stride(0);
params.vnew_batch_stride = v_padded.stride(0);
params.knew_row_stride = k_padded.stride(-3);
params.vnew_row_stride = v_padded.stride(-3);
params.knew_head_stride = k_padded.stride(-2);
params.vnew_head_stride = v_padded.stride(-2);
}
if (seqlens_k_.has_value()) {
auto seqlens_k = seqlens_k_.value();
TORCH_CHECK(seqlens_k.dtype() == torch::kInt32, "seqlens_k must have dtype int32");
CHECK_DEVICE(seqlens_k);
CHECK_CONTIGUOUS(seqlens_k);
CHECK_SHAPE(seqlens_k, batch_size);
params.cu_seqlens_k = static_cast<int *>(seqlens_k.data_ptr());
}
params.is_seqlens_k_cumulative = !(seqlens_k_.has_value());
if (rotary_cos_.has_value()) {
TORCH_CHECK(k_.has_value(), "If rotary cos/sin are provided, new key / value to be appended to KV cache must also be provided");
auto rotary_cos = rotary_cos_.value();
CHECK_DEVICE(rotary_cos);
params.rotary_dim = rotary_cos.size(1) * 2;
TORCH_CHECK(params.rotary_dim <= head_size, "rotary_dim must be <= headdim");
TORCH_CHECK(params.rotary_dim % 16 == 0, "Only rotary dimensions divisible by 16 are currently supported");
const int seqlen_ro = rotary_cos.size(0);
TORCH_CHECK(seqlen_ro >= seqlen_k, "cos/sin seqlen must be at least the seqlen of KV cache");
CHECK_SHAPE(rotary_cos, seqlen_ro, params.rotary_dim / 2);
CHECK_CONTIGUOUS(rotary_cos);
TORCH_CHECK(rotary_cos.scalar_type() == q_dtype, "rotary_cos must have the same dtype as query");
TORCH_CHECK(rotary_sin_.has_value(), "If rotary cos is provided, rotary sin must also be provided");
auto rotary_sin = rotary_sin_.value();
CHECK_DEVICE(rotary_sin);
CHECK_SHAPE(rotary_sin, seqlen_ro, params.rotary_dim / 2);
CHECK_CONTIGUOUS(rotary_sin);
TORCH_CHECK(rotary_sin.scalar_type() == q_dtype, "rotary_cos must have the same dtype as query");
params.rotary_cos_ptr = rotary_cos.data_ptr();
params.rotary_sin_ptr = rotary_sin.data_ptr();
params.is_rotary_interleaved = is_rotary_interleaved;
} else {
params.rotary_dim = 0;
}
if (cache_batch_idx_.has_value()) {
auto cache_batch_idx = cache_batch_idx_.value();
CHECK_DEVICE(cache_batch_idx);
CHECK_CONTIGUOUS(cache_batch_idx);
TORCH_CHECK(cache_batch_idx.scalar_type() == torch::kInt32, "cache_batch_idx must have dtype int32");
params.cache_batch_idx = reinterpret_cast<int *>(cache_batch_idx.data_ptr());
}
// This needs to match with run_mha_fwd_splitkv_dispatch
const int block_n = head_size <= 64 ? 256 : (head_size <= 128 ? 128 : 64);
const int num_n_blocks = (seqlen_k + block_n - 1) / block_n;
// Technically kBlockM = 64 only for the splitKV kernels, not the standard kernel.
// In any case we don't expect seqlen_q to be larger than 64 for inference.
const int num_m_blocks = (seqlen_q + 64 - 1) / 64;
params.num_splits = num_splits;
if (num_splits < 1) {
params.num_splits = num_splits_heuristic(batch_size * num_heads * num_m_blocks, dprops->multiProcessorCount, num_n_blocks, 128);
}
TORCH_CHECK(params.num_splits <= 128, "num_splits > 128 not supported");
if (params.num_splits > 1) {
at::Tensor softmax_lse_accum = torch::empty({params.num_splits, batch_size, num_heads, seqlen_q}, opts.dtype(at::kFloat));
at::Tensor out_accum = torch::empty({params.num_splits, batch_size, num_heads, seqlen_q, head_size_rounded}, opts.dtype(at::kFloat));
params.softmax_lseaccum_ptr = softmax_lse_accum.data_ptr();
params.oaccum_ptr = out_accum.data_ptr();
}
if (alibi_slopes_.has_value()) {
auto alibi_slopes = alibi_slopes_.value();
TORCH_CHECK(alibi_slopes.dtype() == torch::kFloat32, "ALiBi slopes must have dtype fp32");
CHECK_DEVICE(alibi_slopes);
TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
TORCH_CHECK(alibi_slopes.sizes() == torch::IntArrayRef({num_heads}) || alibi_slopes.sizes() == torch::IntArrayRef({batch_size, num_heads}));
params.alibi_slopes_ptr = alibi_slopes.data_ptr();
params.alibi_slopes_batch_stride = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
} else {
params.alibi_slopes_ptr = nullptr;
}
auto stream = at::cuda::getCurrentCUDAStream().stream();
// Only split kernel supports appending to KV cache, or indexing to the cache with cache_batch_idx
run_mha_fwd(params, stream, /*force_split_kernel=*/k_.has_value() || cache_batch_idx_.has_value());
if (head_size_og % 8 != 0) {
out = out.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
if (out_.has_value()) { out_.value().copy_(out); }
if (k_.has_value()) {
// It's expensive to copy the KV cache here for the case where head size not divisible by 8,
// but we don't expect to get this case in practice. This is just so that the code works for that case.
kcache.copy_(kcache_padded.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)}));
vcache.copy_(vcache_padded.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)}));
}
}
if (seqlenq_ngroups_swapped) {
out = out.transpose(1, 2).reshape({batch_size, 1, num_heads_k * seqlen_q, head_size_og});
softmax_lse = softmax_lse.reshape({batch_size, num_heads_k * seqlen_q, 1});
}
return {out, softmax_lse};
}
// add by JXGuo
const int SPARSE_SIZE = 128;
std::vector<at::Tensor>
mha_fwd_block(const at::Tensor &q,
// total_q x num_heads x head_size, total := \sum_{i=0}^{b} s_i
const at::Tensor &k,
// total_k x num_heads x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &v,
// total_k x num_heads x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &cu_seqlens_q, // b+1
const at::Tensor &cu_seqlens_k, // b+1
// const bool is_blocksparse,
const int m_block_dim,
const int n_block_dim,
const at::Tensor &head_mask_type, // (num_heads)
c10::optional<at::Tensor> &streaming_info_, // (num_heads, 2)
c10::optional<at::Tensor> &row_blockmask_, // (batch_size, num_blocksparse_heads, seqlen_m / m_block_dim, seqlen_n / n_block_dim)
const int max_seqlen_q_,
const int max_seqlen_k_,
const float p_dropout,
const float softmax_scale,
const bool is_causal,
const bool exact_streaming,
const bool return_softmax,
int window_size_left,
int window_size_right,
c10::optional<at::Generator> gen_)
{
auto dprops = at::cuda::getCurrentDeviceProperties();
// bool is_sm75 = dprops->major == 7 && dprops->minor == 5;
bool is_sm8x = dprops->major == 8 && dprops->minor >= 0;
bool is_sm90 = dprops->major == 9 && dprops->minor == 0;
const bool has_blockmask = row_blockmask_.has_value();
const bool has_streaming_info = streaming_info_.has_value();
at::Tensor row_blockmask, streaming_info;
if (has_blockmask){
row_blockmask = row_blockmask_.value();
}
if (has_streaming_info){
streaming_info = streaming_info_.value();
}
TORCH_CHECK(is_sm90 || is_sm8x, "FlashAttention only supports Ampere GPUs or newer.");
auto q_dtype = q.dtype();
TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kBFloat16,
"FlashAttention only support fp16 and bf16 data type");
if (q_dtype == torch::kBFloat16) {
TORCH_CHECK(is_sm90 || is_sm8x, "bfloat16 is only supported on Ampere GPUs or newer");
}
TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");
TORCH_CHECK(cu_seqlens_q.dtype() == torch::kInt32, "cu_seqlens_q must have dtype int32");
TORCH_CHECK(cu_seqlens_k.dtype() == torch::kInt32, "cu_seqlens_k must have dtype int32");
TORCH_CHECK(head_mask_type.dtype() == torch::kInt32, "head_mask_type must have dtype int32");
if(has_blockmask){
TORCH_CHECK(row_blockmask.dtype() == torch::kInt32, "row_blockmask must have dtype int32");
TORCH_CHECK(m_block_dim % SPARSE_SIZE == 0, "m_block_dim must be a multiple of 128");
TORCH_CHECK(n_block_dim % SPARSE_SIZE == 0, "n_block_dim must be a multiple of 128");
}
if(has_streaming_info){
TORCH_CHECK(streaming_info.dtype() == torch::kInt32, "streaming_info must have dtype int32");
TORCH_CHECK(m_block_dim % SPARSE_SIZE == 0, "m_block_dim must be a multiple of 128");
TORCH_CHECK(n_block_dim % SPARSE_SIZE == 0, "n_block_dim must be a multiple of 128");
}
CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);
CHECK_DEVICE(cu_seqlens_q);
CHECK_DEVICE(cu_seqlens_k);
CHECK_DEVICE(head_mask_type);
if(has_blockmask){
CHECK_DEVICE(row_blockmask);
}
if(has_streaming_info){
CHECK_DEVICE(streaming_info);
}
TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
CHECK_CONTIGUOUS(cu_seqlens_q);
CHECK_CONTIGUOUS(cu_seqlens_k);
CHECK_CONTIGUOUS(head_mask_type);
if(has_blockmask){
CHECK_CONTIGUOUS(row_blockmask);
}
if(has_streaming_info){
CHECK_CONTIGUOUS(streaming_info);
}
const auto sizes = q.sizes();
const int total_q = sizes[0];
const int batch_size = cu_seqlens_q.numel() - 1;
const int num_heads = sizes[1];
const int head_size_og = sizes[2];
const int total_k = k.size(0);
const int num_heads_k = k.size(1);
int num_blocksparse_heads = 0;
TORCH_CHECK(batch_size > 0, "batch size must be positive");
TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
if (window_size_left >= max_seqlen_k_) { window_size_left = -1; }
if (window_size_right >= max_seqlen_k_) { window_size_right = -1; }
CHECK_SHAPE(q, total_q, num_heads, head_size_og);
CHECK_SHAPE(k, total_k, num_heads_k, head_size_og);
CHECK_SHAPE(v, total_k, num_heads_k, head_size_og);
CHECK_SHAPE(cu_seqlens_q, batch_size + 1);
CHECK_SHAPE(cu_seqlens_k, batch_size + 1);
at::Tensor q_padded, k_padded, v_padded;
if (head_size_og % 8 != 0) {
q_padded = torch::nn::functional::pad(q, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
k_padded = torch::nn::functional::pad(k, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
v_padded = torch::nn::functional::pad(v, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
} else {
q_padded = q;
k_padded = k;
v_padded = v;
}
at::Tensor out;
out = torch::empty_like(q_padded);
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
const int head_size = round_multiple(head_size_og, 8);
const int head_size_rounded = round_multiple(head_size, 32);
const int max_seqlen_q_rounded = round_multiple(max_seqlen_q_, SPARSE_SIZE);
const int max_seqlen_k_rounded = round_multiple(max_seqlen_k_, SPARSE_SIZE);
CHECK_SHAPE(head_mask_type, num_heads);
if(has_blockmask){
num_blocksparse_heads = row_blockmask.size(1);
CHECK_SHAPE(row_blockmask, batch_size, num_blocksparse_heads, max_seqlen_q_rounded / m_block_dim, max_seqlen_k_rounded / n_block_dim);
}
if(has_streaming_info){
CHECK_SHAPE(streaming_info, num_heads * 2);
}
// Otherwise the kernel will be launched from cuda:0 device
// Cast to char to avoid compiler warning about narrowing
at::cuda::CUDAGuard device_guard{(char)q.get_device()};
auto opts = q.options();
auto softmax_lse = torch::empty({batch_size, num_heads, max_seqlen_q_}, opts.dtype(at::kFloat));
at::Tensor p;
// Only return softmax if there's dropout to reduce compilation time
if (return_softmax) {
TORCH_CHECK(p_dropout > 0.0f, "return_softmax is only supported when p_dropout > 0.0");
p = torch::zeros({ batch_size, num_heads, max_seqlen_q_rounded, max_seqlen_k_rounded }, opts);
}
if(is_causal){
window_size_right = 0;
}
Flash_fwd_params params;
set_params_fprop(params,
batch_size,
max_seqlen_q_, max_seqlen_k_,
max_seqlen_q_rounded, max_seqlen_k_rounded,
num_heads, num_heads_k,
head_size, head_size_rounded,
q_padded, k_padded, v_padded, out,
cu_seqlens_q.data_ptr(),
cu_seqlens_k.data_ptr(),
nullptr,
return_softmax ? p.data_ptr() : nullptr,
softmax_lse.data_ptr(),
p_dropout,
softmax_scale,
window_size_left,
window_size_right);
params.head_mask_type = static_cast<int *>(head_mask_type.data_ptr());
if(has_blockmask){
params.blockmask = static_cast<int *>(row_blockmask.data_ptr());
params.m_block_dim = m_block_dim;
params.n_block_dim = n_block_dim;
params.num_blocksparse_heads = num_blocksparse_heads;
}else{
params.blockmask = nullptr;
}
if(has_streaming_info){
params.streaming_info = static_cast<int *>(streaming_info.data_ptr());
params.is_exact_streaming = exact_streaming;
params.m_block_dim = m_block_dim;
params.n_block_dim = n_block_dim;
}else{
params.streaming_info = nullptr;
params.is_exact_streaming = false;
}
// number of times random will be generated per thread, to offset philox counter in thc random
// state
// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = params.b * params.h * 32;
auto options = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
auto rng_state = torch::empty({2}, options.dtype(torch::kInt64));
// Forward kernel will populate memory with the seed and offset.
params.rng_state = reinterpret_cast<uint64_t*>(rng_state.data_ptr());
if (p_dropout > 0.0) {
auto gen = at::get_generator_or_default<at::CUDAGeneratorImpl>(
gen_, at::cuda::detail::getDefaultCUDAGenerator());
// See Note [Acquire lock when using random generators]
std::lock_guard<std::mutex> lock(gen->mutex_);
params.philox_args = gen->philox_cuda_state(counter_offset);
}
auto stream = at::cuda::getCurrentCUDAStream().stream();
run_mha_fwd_block(params, stream);
at::Tensor out_padded = out;
if (head_size_og % 8 != 0) {
out = out.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
}
return {out, q_padded, k_padded, v_padded, out_padded, softmax_lse, p, rng_state};
}
std::vector<at::Tensor>
mha_bwd_block(const at::Tensor &dout, // total_q x num_heads, x head_size
const at::Tensor &q, // total_q x num_heads x head_size, total_q := \sum_{i=0}^{b} s_i
const at::Tensor &k, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &v, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &out, // total_q x num_heads x head_size
const at::Tensor &softmax_lse, // b x h x s softmax logsumexp
c10::optional<at::Tensor> &dq_, // total_q x num_heads x head_size, total_q := \sum_{i=0}^{b} s_i
c10::optional<at::Tensor> &dk_, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
c10::optional<at::Tensor> &dv_, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
const at::Tensor &cu_seqlens_q, // b+1
const at::Tensor &cu_seqlens_k, // b+1
const int m_block_dim,
const int n_block_dim,
const at::Tensor &head_mask_type, // (num_heads)
c10::optional<at::Tensor> &streaming_info_,
c10::optional<at::Tensor> &col_blockmask_, // (batch_size, num_blocksparse_heads, seqlen_n / n_block_dim, seqlen_m / m_block_dim)
const int max_seqlen_q_,
const int max_seqlen_k_, // max sequence length to choose the kernel
const float p_dropout, // probability to drop
const float softmax_scale,
const bool zero_tensors,
const bool is_causal,
int window_size_left,
int window_size_right,
const bool deterministic,
c10::optional<at::Generator> gen_,
c10::optional<at::Tensor> &rng_state)
{
if (is_causal) { window_size_right = 0; }
auto dprops = at::cuda::getCurrentDeviceProperties();
bool is_sm8x = dprops->major == 8 && dprops->minor >= 0;
bool is_sm80 = dprops->major == 8 && dprops->minor == 0;
bool is_sm90 = dprops->major == 9 && dprops->minor == 0;
const bool has_blockmask = col_blockmask_.has_value();
const bool has_streaming_info = streaming_info_.has_value();
at::Tensor col_blockmask, streaming_info;
if (has_blockmask){
col_blockmask = col_blockmask_.value();
}
if (has_streaming_info){
streaming_info = streaming_info_.value();
}
TORCH_CHECK(is_sm90 || is_sm8x, "FlashAttention only supports Ampere GPUs or newer.");
bool is_dropout = p_dropout > 0.0;
auto stream = at::cuda::getCurrentCUDAStream().stream();
auto q_dtype = q.dtype();
TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kBFloat16,
"FlashAttention only support fp16 and bf16 data type");
if (q_dtype == torch::kBFloat16) {
TORCH_CHECK(is_sm90 || is_sm8x, "bfloat16 is only supported on Ampere GPUs or newer");
}
TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");
TORCH_CHECK(out.dtype() == q_dtype, "query and out must have the same dtype");
TORCH_CHECK(dout.dtype() == q_dtype, "query and dout must have the same dtype");
TORCH_CHECK(cu_seqlens_q.dtype() == torch::kInt32, "cu_seqlens_q must have dtype int32");
TORCH_CHECK(cu_seqlens_k.dtype() == torch::kInt32, "cu_seqlens_k must have dtype int32");
TORCH_CHECK(head_mask_type.dtype() == torch::kInt32, "head_mask_type must have dtype int32");
if(has_blockmask){
TORCH_CHECK(col_blockmask.dtype() == torch::kInt32, "col_blockmask must have dtype int32");
TORCH_CHECK(m_block_dim % SPARSE_SIZE == 0, "m_block_dim must be a multiple of 128");
TORCH_CHECK(n_block_dim % SPARSE_SIZE == 0, "n_block_dim must be a multiple of 128");
}
if(has_streaming_info){
TORCH_CHECK(streaming_info.dtype() == torch::kInt32, "streaming_info must have dtype int32");
TORCH_CHECK(m_block_dim % SPARSE_SIZE == 0, "m_block_dim must be a multiple of 128");
TORCH_CHECK(n_block_dim % SPARSE_SIZE == 0, "n_block_dim must be a multiple of 128");
}
CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);
CHECK_DEVICE(out); CHECK_DEVICE(dout); CHECK_DEVICE(softmax_lse);
CHECK_DEVICE(cu_seqlens_q); CHECK_DEVICE(cu_seqlens_k);
CHECK_DEVICE(head_mask_type);
if(has_blockmask){
CHECK_DEVICE(col_blockmask);
}
if(has_streaming_info){
CHECK_DEVICE(streaming_info);
}
TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
TORCH_CHECK(out.stride(-1) == 1, "out tensor must have contiguous last dimension");
TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
CHECK_CONTIGUOUS(cu_seqlens_q);
CHECK_CONTIGUOUS(cu_seqlens_k);
CHECK_CONTIGUOUS(head_mask_type);
if(has_blockmask){
CHECK_CONTIGUOUS(col_blockmask);
}
if(has_streaming_info){
CHECK_CONTIGUOUS(streaming_info);
}
const auto sizes = q.sizes();
const int total_q = sizes[0];
const int batch_size = cu_seqlens_q.numel() - 1;
const int num_heads = sizes[1];
const int head_size_og = dout.size(2);
const int head_size = sizes[2];
const int total_k = k.size(0);
const int num_heads_k = k.size(1);
int num_blocksparse_heads = 0;
TORCH_CHECK(batch_size > 0, "batch size must be positive");
TORCH_CHECK(head_size % 8 == 0, "head_size should be a multiple of 8");
TORCH_CHECK(head_size <= 256, "FlashAttention backward only supports head dimension at most 256");
if (head_size > 192) {
TORCH_CHECK(is_sm80 || is_sm90, "FlashAttention backward for head dim > 192 requires A100/A800 or H100/H800");
}
TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
const int head_size_rounded = round_multiple(head_size, 32);
const int max_seqlen_q_rounded = round_multiple(max_seqlen_q_, SPARSE_SIZE);
const int max_seqlen_k_rounded = round_multiple(max_seqlen_k_, SPARSE_SIZE);
TORCH_CHECK(head_size == round_multiple(head_size_og, 8), "head_size must be head_size_og rounded to a multiple of 8");
if (window_size_left >= max_seqlen_k_) { window_size_left = -1; }
if (window_size_right >= max_seqlen_k_) { window_size_right = -1; }
CHECK_SHAPE(q, total_q, num_heads, head_size);
CHECK_SHAPE(k, total_k, num_heads_k, head_size);
CHECK_SHAPE(v, total_k, num_heads_k, head_size);
CHECK_SHAPE(out, total_q, num_heads, head_size);
CHECK_SHAPE(dout, total_q, num_heads, head_size_og);
CHECK_SHAPE(cu_seqlens_q, batch_size + 1);
CHECK_SHAPE(cu_seqlens_k, batch_size + 1);
CHECK_SHAPE(head_mask_type, num_heads);
if(has_blockmask){
num_blocksparse_heads = col_blockmask.size(1);
CHECK_SHAPE(col_blockmask, batch_size, num_blocksparse_heads, max_seqlen_k_rounded / n_block_dim, max_seqlen_q_rounded / m_block_dim);// todo: check the shape
}
if(has_streaming_info){
CHECK_SHAPE(streaming_info, num_heads * 2);
}
at::Tensor dq, dk, dv;
if (dq_.has_value()) {
dq = dq_.value();
TORCH_CHECK(dq.dtype() == q_dtype, "dq must have the same dtype as q");
CHECK_DEVICE(dq);
TORCH_CHECK(dq.stride(-1) == 1, "dq must have contiguous last dimension");
CHECK_SHAPE(dq, total_q, num_heads, head_size);
} else {
dq = torch::empty_like(q);
}
if (dk_.has_value()) {
dk = dk_.value();
TORCH_CHECK(dk.dtype() == q_dtype, "dk must have the same dtype as q");
CHECK_DEVICE(dk);
TORCH_CHECK(dk.stride(-1) == 1, "dk must have contiguous last dimension");
CHECK_SHAPE(dk, total_k, num_heads_k, head_size);
} else {
dk = torch::empty_like(k);
}
if (dv_.has_value()) {
dv = dv_.value();
TORCH_CHECK(dv.dtype() == q_dtype, "dv must have the same dtype as q");
CHECK_DEVICE(dv);
TORCH_CHECK(dv.stride(-1) == 1, "dv must have contiguous last dimension");
CHECK_SHAPE(dv, total_k, num_heads_k, head_size);
} else {
dv = torch::empty_like(v);
}
at::Tensor dout_padded;
if (head_size_og % 8 != 0) {
dout_padded = torch::nn::functional::pad(dout, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
} else {
dout_padded = dout;
}
// bool loop = max_seqlen_k > blocksize_c;
// TODO: change later, for now set to true for simplicity
bool loop = true;
// Otherwise the kernel will be launched from cuda:0 device
// Cast to char to avoid compiler warning about narrowing
at::cuda::CUDAGuard device_guard{(char)q.get_device()};
auto opts = q.options();
auto softmax_d = torch::empty({batch_size, num_heads, max_seqlen_q_rounded}, opts.dtype(at::kFloat));
at::Tensor dq_accum;
if (loop) {
// We don't want to allocate dq_accum of size (batch, seqlen_q_rounded, num_heads, head_size_rounded)
// because that would be too large if there is a very long sequence and the rest of the sequences are short.
// Instead, we allocate dq_accum of size (total_q + 128 * batch, num_heads, head_size_rounded).
// Note that 128 is the max block size on the seqlen_q dimension.
// For dQ, the i-th sequence is stored in indices from cu_seqlens[i] + 128 * i to
// cu_seqlens[i + 1] * 128 * i - 1. This ensures that the i-th sequence and (i + 1)-th sequence will
// be at least 128 apart. It's ok for us to do atomicAdds up to 128 rows beyond what we're normally
// allowed to do. So we won't have to do any bound checking, and performance should stay the same.
if (!deterministic) {
dq_accum = torch::empty({total_q + 128 * batch_size, num_heads, head_size_rounded}, opts.dtype(at::kFloat));
} else {
const int nsplits = (dprops->multiProcessorCount + batch_size * num_heads - 1) / (batch_size * num_heads);
dq_accum = torch::zeros({nsplits, total_q + 128 * batch_size, num_heads, head_size_rounded}, opts.dtype(at::kFloat));
}
}
at::Tensor dk_expanded, dv_expanded;
if (num_heads_k != num_heads) { // MQA / GQA
dk_expanded = torch::empty({total_k, num_heads, head_size}, opts);
dv_expanded = torch::empty({total_k, num_heads, head_size}, opts);
} else {
dk_expanded = dk;
dv_expanded = dv;
}
if( zero_tensors ) {
dq.zero_();
dk_expanded.zero_();
dv_expanded.zero_();
softmax_d.zero_();
}
Flash_bwd_params params;
set_params_dgrad(
params,
batch_size,
max_seqlen_q_, max_seqlen_k_,
max_seqlen_q_rounded, max_seqlen_k_rounded,
num_heads, num_heads_k,
head_size, head_size_rounded,
q, k, v, out,
dout_padded, dq, dk_expanded, dv_expanded,
cu_seqlens_q.data_ptr(),
cu_seqlens_k.data_ptr(),
loop ? dq_accum.data_ptr() : nullptr,
nullptr,
nullptr,
softmax_lse.data_ptr(),
softmax_d.data_ptr(),
p_dropout,
softmax_scale,
window_size_left, window_size_right, deterministic
);
params.dq_accum_split_stride = !deterministic ? 0 : dq_accum.stride(0);
params.head_mask_type = static_cast<int *>(head_mask_type.data_ptr());
if(has_blockmask){
params.blockmask = static_cast<int *>(col_blockmask.data_ptr());
params.m_block_dim = m_block_dim;
params.n_block_dim = n_block_dim;
params.num_blocksparse_heads = num_blocksparse_heads;
}else{
params.blockmask = nullptr;
}
if(has_streaming_info){
params.streaming_info = static_cast<int *>(streaming_info.data_ptr());
params.m_block_dim = m_block_dim;
params.n_block_dim = n_block_dim;
}else{
params.streaming_info = nullptr;
}
auto launch = &run_mha_bwd_block;
// launch(params, stream, /*configure=*/true);
auto gen = at::get_generator_or_default<at::CUDAGeneratorImpl>(
gen_, at::cuda::detail::getDefaultCUDAGenerator());
// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = params.b * params.h * 32;
if ( rng_state.has_value() ) {
params.rng_state = reinterpret_cast<uint64_t*>(rng_state.value().data_ptr());
} else if(is_dropout) {
// See Note [Acquire lock when using random generators]
std::lock_guard<std::mutex> lock(gen->mutex_);
params.philox_args = gen->philox_cuda_state(counter_offset);
auto seeds = at::cuda::philox::unpack(params.philox_args);
params.rng_state[0] = std::get<0>(seeds);
params.rng_state[1] = std::get<1>(seeds);
}
if (max_seqlen_q_ > 0) {
launch(params, stream, /*configure=*/false);
} else {
// If seqlen_q == 0, then we have an empty tensor. We need to set the output to 0.
dk_expanded.zero_();
dv_expanded.zero_();
softmax_d.zero_();
}
// For MQA/GQA we need to sum dK and dV across the groups
if (num_heads_k != num_heads) {
at::sum_out(dk, at::reshape(dk_expanded, {total_k, num_heads_k, num_heads / num_heads_k, head_size}), {2});
at::sum_out(dv, at::reshape(dv_expanded, {total_k, num_heads_k, num_heads / num_heads_k, head_size}), {2});
}
if (head_size_og % 8 != 0) {
dq = dq.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
dk = dk.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
dv = dv.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
}
return { dq, dk, dv, softmax_d };
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.doc() = "FlashAttention";
m.def("fwd", &mha_fwd, "Forward pass");
m.def("varlen_fwd", &mha_varlen_fwd, "Forward pass (variable length)");
m.def("bwd", &mha_bwd, "Backward pass");
m.def("varlen_bwd", &mha_varlen_bwd, "Backward pass (variable length)");
m.def("fwd_kvcache", &mha_fwd_kvcache, "Forward pass, with KV-cache");
m.def("fwd_block", &mha_fwd_block, "Forward pass, with blockmask");
m.def("bwd_block", &mha_bwd_block, "Forward pass, with blockmask");
}
csrc/block_sparse_attn/src/alibi.h 0 → 100644
View file @ 4f83cf8f
#include <cmath>
#include <cute/tensor.hpp>
#include <cutlass/cutlass.h>
#include <cutlass/array.h>
#include "utils.h"
namespace flash {
using namespace cute;
////////////////////////////////////////////////////////////////////////////////////////////////////
template <bool Is_causal, typename Engine, typename Layout>
inline __device__ void apply_alibi(Tensor<Engine, Layout> &tensor,
const int col_idx_offset_,
const int max_seqlen_k,
const int row_idx_offset,
const int max_seqlen_q,
const int warp_row_stride,
const float alibi_slope) {
// tensor has shape (ncol=(2, MMA_M), nrow=(2, MMA_N))
static_assert(Layout::rank == 2, "Only support 2D Tensor");
const int lane_id = threadIdx.x % 32;
const int col_idx_offset = col_idx_offset_ + (lane_id % 4) * 2;
if constexpr (Is_causal) { // Simpler, we add the same bias vector to all rows
#pragma unroll
for (int nj = 0; nj < size<1, 1>(tensor); ++nj) {
const int col_idx_base = col_idx_offset + nj * 8;
#pragma unroll
for (int j = 0; j < size<1, 0>(tensor); ++j) {
const int col_idx = col_idx_base + j;
#pragma unroll
for (int mi = 0; mi < size<0>(tensor); ++mi) {
tensor(mi, make_coord(j, nj)) += alibi_slope * col_idx;
}
}
}
} else { // Bias depends on both row_idx and col_idx
#pragma unroll
for (int mi = 0; mi < size<0, 1>(tensor); ++mi) {
const int row_idx_base = row_idx_offset + mi * warp_row_stride;
#pragma unroll
for (int i = 0; i < size<0, 0>(tensor); ++i) {
const int row_idx = row_idx_base + i * 8;
#pragma unroll
for (int nj = 0; nj < size<1, 1>(tensor); ++nj) {
const int col_idx_base = col_idx_offset + nj * 8;
#pragma unroll
for (int j = 0; j < size<1, 0>(tensor); ++j) {
const int col_idx = col_idx_base + j;
tensor(make_coord(i, mi), make_coord(j, nj)) -= alibi_slope * abs(row_idx + max_seqlen_k - max_seqlen_q - col_idx);
}
}
}
}
}
}
} // namespace flash
csrc/block_sparse_attn/src/block_info.h 0 → 100644
View file @ 4f83cf8f
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
#pragma once
namespace flash {
////////////////////////////////////////////////////////////////////////////////////////////////////
template<bool Varlen=true>
struct BlockInfo {
template<typename Params>
__device__ BlockInfo(const Params &params, const int bidb)
: sum_s_q(!Varlen || params.cu_seqlens_q == nullptr ? -1 : params.cu_seqlens_q[bidb])
, sum_s_k(!Varlen || params.cu_seqlens_k == nullptr || !params.is_seqlens_k_cumulative ? -1 : params.cu_seqlens_k[bidb])
, actual_seqlen_q(!Varlen || params.cu_seqlens_q == nullptr ? params.seqlen_q : params.cu_seqlens_q[bidb + 1] - sum_s_q)
// If is_seqlens_k_cumulative, then seqlen_k is cu_seqlens_k[bidb + 1] - cu_seqlens_k[bidb].
// Otherwise it's cu_seqlens_k[bidb], i.e., we use cu_seqlens_k to store the sequence lengths of K.
, seqlen_k_cache(!Varlen || params.cu_seqlens_k == nullptr ? params.seqlen_k : (params.is_seqlens_k_cumulative ? params.cu_seqlens_k[bidb + 1] - sum_s_k : params.cu_seqlens_k[bidb]))
, actual_seqlen_k(params.seqused_k ? params.seqused_k[bidb] : seqlen_k_cache + (params.knew_ptr == nullptr ? 0 : params.seqlen_knew))
{
}
template <typename index_t>
inline __device__ index_t q_offset(const index_t batch_stride, const index_t row_stride, const int bidb) const {
return sum_s_q == -1 ? bidb * batch_stride : uint32_t(sum_s_q) * row_stride;
}
template <typename index_t>
inline __device__ index_t k_offset(const index_t batch_stride, const index_t row_stride, const int bidb) const {
return sum_s_k == -1 ? bidb * batch_stride : uint32_t(sum_s_k) * row_stride;
}
const int sum_s_q;
const int sum_s_k;
const int actual_seqlen_q;
// We have to have seqlen_k_cache declared before actual_seqlen_k, otherwise actual_seqlen_k is set to 0.
const int seqlen_k_cache;
const int actual_seqlen_k;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace flash
csrc/block_sparse_attn/src/flash.h 0 → 100644
View file @ 4f83cf8f
/******************************************************************************
* Copyright (c) 2023, Tri Dao.
******************************************************************************/
/******************************************************************************
* Adapted by Junxian Guo from https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/src/flash.h
******************************************************************************/
#pragma once
#include <cuda.h>
#include <vector>
#ifdef OLD_GENERATOR_PATH
#include <ATen/CUDAGeneratorImpl.h>
#else
#include <ATen/cuda/CUDAGeneratorImpl.h>
#endif
#include <ATen/cuda/CUDAGraphsUtils.cuh> // For at::cuda::philox::unpack
constexpr int TOTAL_DIM = 0;
constexpr int H_DIM = 1;
constexpr int D_DIM = 2;
////////////////////////////////////////////////////////////////////////////////////////////////////
struct Qkv_params {
using index_t = int64_t;
// The QKV matrices.
void *__restrict__ q_ptr;
void *__restrict__ k_ptr;
void *__restrict__ v_ptr;
// The stride between rows of the Q, K and V matrices.
index_t q_batch_stride;
index_t k_batch_stride;
index_t v_batch_stride;
index_t q_row_stride;
index_t k_row_stride;
index_t v_row_stride;
index_t q_head_stride;
index_t k_head_stride;
index_t v_head_stride;
// The number of heads.
int h, h_k;
// In the case of multi-query and grouped-query attention (MQA/GQA), nheads_k could be
// different from nheads (query).
int h_h_k_ratio; // precompute h / h_k,
};
////////////////////////////////////////////////////////////////////////////////////////////////////
struct Flash_fwd_params : public Qkv_params {
// The O matrix (output).
void * __restrict__ o_ptr;
void * __restrict__ oaccum_ptr;
// The stride between rows of O.
index_t o_batch_stride;
index_t o_row_stride;
index_t o_head_stride;
// The pointer to the P matrix.
void * __restrict__ p_ptr;
// The pointer to the softmax sum.
void * __restrict__ softmax_lse_ptr;
void * __restrict__ softmax_lseaccum_ptr;
// The dimensions.
int b, seqlen_q, seqlen_k, seqlen_knew, d, seqlen_q_rounded, seqlen_k_rounded, d_rounded, rotary_dim;
// The scaling factors for the kernel.
float scale_softmax;
float scale_softmax_log2;
// array of length b+1 holding starting offset of each sequence.
int * __restrict__ cu_seqlens_q;
int * __restrict__ cu_seqlens_k;
// If provided, the actual length of each k sequence.
int * __restrict__ seqused_k;
int *__restrict__ blockmask;
int *__restrict__ streaming_info;
int *__restrict__ head_mask_type;
// add by JXGuo
int m_block_dim, n_block_dim, num_blocksparse_heads;
// The K_new and V_new matrices.
void * __restrict__ knew_ptr;
void * __restrict__ vnew_ptr;
// The stride between rows of the Q, K and V matrices.
index_t knew_batch_stride;
index_t vnew_batch_stride;
index_t knew_row_stride;
index_t vnew_row_stride;
index_t knew_head_stride;
index_t vnew_head_stride;
// The cos and sin matrices for rotary embedding.
void * __restrict__ rotary_cos_ptr;
void * __restrict__ rotary_sin_ptr;
// The indices to index into the KV cache.
int *__restrict__ cache_batch_idx;
// The dropout probability (probability of keeping an activation).
float p_dropout;
// uint32_t p_dropout_in_uint;
// uint16_t p_dropout_in_uint16_t;
uint8_t p_dropout_in_uint8_t;
// Scale factor of 1 / (1 - p_dropout).
float rp_dropout;
float scale_softmax_rp_dropout;
// Local window size
int window_size_left, window_size_right;
// Random state.
at::PhiloxCudaState philox_args;
// Pointer to the RNG seed (idx 0) and offset (idx 1).
uint64_t * rng_state;
bool is_bf16;
bool is_causal;
bool is_exact_streaming;
// If is_seqlens_k_cumulative, then seqlen_k is cu_seqlens_k[bidb + 1] - cu_seqlens_k[bidb].
// Otherwise it's cu_seqlens_k[bidb], i.e., we use cu_seqlens_k to store the sequence lengths of K.
bool is_seqlens_k_cumulative;
bool is_rotary_interleaved;
int num_splits; // For split-KV version
void * __restrict__ alibi_slopes_ptr;
index_t alibi_slopes_batch_stride;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
struct Flash_bwd_params : public Flash_fwd_params {
// The dO and dQKV matrices.
void *__restrict__ do_ptr;
void *__restrict__ dq_ptr;
void *__restrict__ dk_ptr;
void *__restrict__ dv_ptr;
// To accumulate dQ
void *__restrict__ dq_accum_ptr;
void *__restrict__ dk_accum_ptr;
void *__restrict__ dv_accum_ptr;
// // To accumulate dK and dV in case we're splitting the bwd along seqlen_q
// dimension void *__restrict__ dk_accum_ptr; void *__restrict__
// dv_accum_ptr;
// The stride between rows of the dO, dQ, dK and dV matrices.
// TD [2022-04-16]: We're using 32-bit indexing to save registers.
// The code probably won't work for arrays larger than 2GB.
index_t do_batch_stride;
index_t do_row_stride;
index_t do_head_stride;
index_t dq_batch_stride;
index_t dk_batch_stride;
index_t dv_batch_stride;
index_t dq_row_stride;
index_t dk_row_stride;
index_t dv_row_stride;
index_t dq_head_stride;
index_t dk_head_stride;
index_t dv_head_stride;
// The pointer to the softmax d sum.
void *__restrict__ dsoftmax_sum;
bool deterministic;
index_t dq_accum_split_stride;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
template<typename T, int Headdim> void run_mha_fwd_(Flash_fwd_params &params, cudaStream_t stream);
template<typename T, int Headdim> void run_mha_fwd_splitkv_dispatch(Flash_fwd_params &params, cudaStream_t stream);
template<typename T, int Headdim> void run_mha_bwd_(Flash_bwd_params &params, cudaStream_t stream, const bool configure);
template<typename T, int Headdim> void run_mha_fwd_block_(Flash_fwd_params &params, cudaStream_t stream);
template<typename T, int Headdim> void run_mha_bwd_block_(Flash_bwd_params &params, cudaStream_t stream, const bool configure);
csrc/block_sparse_attn/src/flash_blockmask.h 0 → 100644
View file @ 4f83cf8f
/******************************************************************************
* Copyright (c) 2024, Junxian Guo.
******************************************************************************/
#pragma once
namespace flash {
class fwdIteratorBase{
};
// ////////////////////////////////////////////////////////////////////////////////////////////////////
class fwdStreaming: public fwdIteratorBase{
public:
template<typename Params, typename BlockInfo>
__device__ fwdStreaming(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int n_block_min, int n_block_max) {//row first
this -> row_factor = params.m_block_dim / kBlockM;
this -> col_factor = params.n_block_dim / kBlockN;
this -> sink_block_num = params.streaming_info[head_idx * 2] * col_factor;
this -> local_block_num = params.streaming_info[head_idx * 2 + 1] * col_factor;
this -> m_block_dim = params.m_block_dim;
this -> n_block_dim = params.n_block_dim;
this -> mask_type = params.head_mask_type[head_idx];
this -> n_block_min = n_block_min;
this -> n_block_max = n_block_max;
int act_k = binfo.actual_seqlen_k;
int act_q = binfo.actual_seqlen_q;
bool causal = params.is_causal;
if (causal){
int start_row_idx = max(int((act_q-act_k)/m_block_dim), 0);
this -> start_block_val = (cute::ceil_div(max(act_k - act_q, 0), n_block_dim) + 1 + loop_step_idx/row_factor - start_row_idx) * col_factor;
}else{
this -> start_block_val = max(cute::ceil_div(n_block_max * kBlockN, n_block_dim) * col_factor, 0);
};
this -> no_gap = start_block_val - n_block_min < sink_block_num + local_block_num;
this -> max_block_idx = min(sink_block_num + local_block_num, start_block_val - n_block_min);
assert(mask_type < 0);
assert(params.m_block_dim % kBlockM == 0);
assert(params.n_block_dim % kBlockN == 0);
};
__device__ int mask_val(int block_col_idx) const {
if (block_col_idx > max_block_idx || block_col_idx < 0){
return -1;
};
int ret = 0;
if (no_gap){
ret = start_block_val - 1 - block_col_idx;
return ret >= n_block_min ? ret : -1;
}else{
if (block_col_idx < local_block_num){
return start_block_val - 1 - block_col_idx;
}else{
ret = sink_block_num - 1 - (block_col_idx - local_block_num);
return ret >= n_block_min ? ret : -1;
};
};
};
__device__ int max_no_larger(int target) const {
if(max_block_idx == 0){
return -1;
};
int left = 0;
int right = max_block_idx - 1;
while (left <= right) {
int mid = left + (right - left) / 2;
if (mask_val(mid) > target) {
left = mid + 1;
} else {
right = mid - 1;
};
};
return (left < max_block_idx && mask_val(left) <= target) ? left : -1;
};
int sink_block_num, local_block_num;
int start_block_val;
bool no_gap;
int max_block_idx;
int m_block_dim, n_block_dim;
int mask_type;
int n_block_min, n_block_max;
int row_factor, col_factor;
};
class fwdExactStreaming: public fwdIteratorBase{
public:
template<typename Params, typename BlockInfo>
__device__ fwdExactStreaming(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int n_block_min, int n_block_max) {//row first
this -> row_factor = params.m_block_dim / kBlockM;
this -> col_factor = params.n_block_dim / kBlockN;
int sink_num = params.streaming_info[head_idx * 2];
int local_num = params.streaming_info[head_idx * 2 + 1];
this -> m_block_dim = params.m_block_dim;
this -> n_block_dim = params.n_block_dim;
this -> sink_block_num = cute::ceil_div(sink_num, n_block_dim) * col_factor;
this -> local_block_num = (cute::ceil_div(local_num, n_block_dim)+2) * col_factor;
this -> mask_type = params.head_mask_type[head_idx];
this -> n_block_min = n_block_min;
this -> n_block_max = n_block_max;
int act_k = binfo.actual_seqlen_k;
int act_q = binfo.actual_seqlen_q;
bool causal = params.is_causal;
if (causal){
int start_row_idx = max(int((act_q-act_k)/m_block_dim), 0);
this -> start_block_val = (cute::ceil_div(max(act_k - act_q, 0), n_block_dim) + 1 + loop_step_idx/row_factor - start_row_idx) * col_factor;
}else{
this -> start_block_val = max(cute::ceil_div(n_block_max * kBlockN, n_block_dim) * col_factor, 0);
};
this -> no_gap = start_block_val - n_block_min < sink_block_num + local_block_num;
this -> max_block_idx = min(sink_block_num + local_block_num, start_block_val - n_block_min);
assert(mask_type < 0);
assert(params.m_block_dim % kBlockM == 0);
assert(params.n_block_dim % kBlockN == 0);
};
__device__ int mask_val(int block_col_idx) const {
if (block_col_idx > max_block_idx || block_col_idx < 0){
return -1;
};
int ret = 0;
if (no_gap){
ret = start_block_val - 1 - block_col_idx;
return ret >= n_block_min ? ret : -1;
}else{
if (block_col_idx < local_block_num){
return start_block_val - 1 - block_col_idx;
}else{
ret = sink_block_num - 1 - (block_col_idx - local_block_num);
return ret >= n_block_min ? ret : -1;
};
};
};
__device__ int max_no_larger(int target) const {
if(max_block_idx == 0){
return -1;
};
int left = 0;
int right = max_block_idx - 1;
while (left <= right) {
int mid = left + (right - left) / 2;
if (mask_val(mid) > target) {
left = mid + 1;
} else {
right = mid - 1;
};
};
return (left < max_block_idx && mask_val(left) <= target) ? left : -1;
};
int sink_block_num, local_block_num;
int start_block_val;
bool no_gap;
int max_block_idx;
int m_block_dim, n_block_dim;
int mask_type;
int n_block_min, n_block_max;
int row_factor, col_factor;
};
// ////////////////////////////////////////////////////////////////////////////////////////////////////
class fwdBlockmask: public fwdIteratorBase{
public:
template<typename Params, typename BlockInfo>
__device__ fwdBlockmask(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int n_block_min, int n_block_max) {//row first
this -> row_factor = params.m_block_dim / kBlockM;
this -> col_factor = params.n_block_dim / kBlockN;
this -> max_block_idx = cute::ceil_div(binfo.actual_seqlen_k, params.n_block_dim) * col_factor;
this -> m_block_dim = params.m_block_dim;
this -> n_block_dim = params.n_block_dim;
this -> mask_type = params.head_mask_type[head_idx];
this -> n_block_min = n_block_min;
this -> n_block_max = n_block_max;
assert(mask_type > 0);
assert(params.m_block_dim % kBlockM == 0);
assert(params.n_block_dim % kBlockN == 0);
blockmask_ptr = params.blockmask + (batch_idx * params.num_blocksparse_heads + mask_type - 1) * int(params.seqlen_q_rounded / m_block_dim) * int(params.seqlen_k_rounded / n_block_dim) + int(loop_step_idx / row_factor) * int(params.seqlen_k_rounded / n_block_dim);
};
__device__ int mask_val(int block_col_idx) const {
if (block_col_idx > max_block_idx || block_col_idx < 0){
return -1;
};
int real_block_idx = block_col_idx / col_factor;
int block_col_offset = block_col_idx % col_factor;
int mask_val = blockmask_ptr[real_block_idx];
return mask_val == -1 ? -1 : col_factor * mask_val + col_factor - 1 - block_col_offset;
};
__device__ int max_no_larger(int target) const {
if(max_block_idx == 0){
return -1;
};
int left = 0;
int right = max_block_idx - 1;
while (left <= right) {
int mid = left + (right - left) / 2;
if (mask_val(mid) > target) {
left = mid + 1;
} else {
right = mid - 1;
};
};
return (left < max_block_idx && mask_val(left) <= target) ? left : -1;
};
int *blockmask_ptr;
int max_block_idx;
int m_block_dim, n_block_dim;
int mask_type;
int n_block_min, n_block_max;
int row_factor, col_factor;
};
// ////////////////////////////////////////////////////////////////////////////////////////////////////
template<bool Is_streaming, bool Is_exact_streaming>
class fwdIterator{};
template<>
struct fwdIterator<false, false>: public fwdBlockmask{
template<typename Params, typename BlockInfo>
__device__ fwdIterator(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int n_block_min, int n_block_max): fwdBlockmask(params, binfo, kBlockM, kBlockN, batch_idx, head_idx, loop_step_idx, n_block_min, n_block_max) {};
};
template<>
struct fwdIterator<true, false>: public fwdStreaming{
template<typename Params, typename BlockInfo>
__device__ fwdIterator(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int n_block_min, int n_block_max): fwdStreaming(params, binfo, kBlockM, kBlockN, batch_idx, head_idx, loop_step_idx, n_block_min, n_block_max) {};
};
template<>
struct fwdIterator<true, true>: public fwdExactStreaming{
template<typename Params, typename BlockInfo>
__device__ fwdIterator(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int n_block_min, int n_block_max): fwdExactStreaming(params, binfo, kBlockM, kBlockN, batch_idx, head_idx, loop_step_idx, n_block_min, n_block_max) {};
};
////////////////////////////////////////////////////////////////////////////////////////////////////
class bwdIteratorBase{
};
struct bwdStreaming: public bwdIteratorBase{
public:
template<typename Params, typename BlockInfo>
__device__ bwdStreaming(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int m_block_min, int m_block_max) {// col first
this -> row_factor = params.m_block_dim / kBlockM;
this -> col_factor = params.n_block_dim / kBlockN;
this -> m_block_dim = params.m_block_dim;
this -> n_block_dim = params.n_block_dim;
this -> mask_type = params.head_mask_type[head_idx];
this -> m_block_min = m_block_min;
this -> m_block_max = m_block_max;
int mask_block_col = cute::ceil_div(loop_step_idx+1, col_factor);
int sink = (this -> mask_type) < 0 ? params.streaming_info[head_idx * 2]: cute::ceil_div(binfo.actual_seqlen_k, this -> n_block_dim);
int local = (this -> mask_type) < 0 ? params.streaming_info[head_idx * 2 + 1]: 0;
this -> sink_block_num = sink * col_factor;
this -> local_block_num = local * col_factor;
int act_q = binfo.actual_seqlen_q;
int act_k = binfo.actual_seqlen_k;
bool causal = params.is_causal;
if(mask_block_col <= sink){
this -> start_block_val = m_block_max;
this -> max_block_idx = m_block_max - m_block_min;
}else{
if (causal){
int free_token_num = act_q - min(act_q, act_k - loop_step_idx * kBlockN);
int end_mask_block_row_idx = free_token_num / params.m_block_dim;//zero based
int num_mask_block_in_end_row = max(0, cute::ceil_div(act_k - act_q + (end_mask_block_row_idx + 1) * params.m_block_dim, params.n_block_dim));
int local_col_mask_block_num = max(0, local - (num_mask_block_in_end_row - mask_block_col));
if(local_col_mask_block_num > 0){
this -> start_block_val = min((end_mask_block_row_idx + local_col_mask_block_num) * row_factor, m_block_max);
this -> max_block_idx = min(local_col_mask_block_num * row_factor, m_block_max - m_block_min);
}else{
this -> start_block_val = 0;
this -> max_block_idx = 0;
};
}else{
int n_mask_block_col = max(cute::ceil_div(act_k, n_block_dim), 0);
bool in_none_causal_local = !causal && mask_block_col <= n_mask_block_col && mask_block_col > n_mask_block_col - local;
if(in_none_causal_local){
this -> start_block_val = m_block_max;
this -> max_block_idx = m_block_max - m_block_min;
}else{
this -> start_block_val = 0;
this -> max_block_idx = 0;
};
};
}
assert(mask_type <= 0); //for blocksparse, mask_type > 0; for streaming, mask_type < 0; for dense, mask_type = 0
assert(params.m_block_dim % kBlockM == 0);
assert(params.n_block_dim % kBlockN == 0);
};
__device__ int mask_val(int block_row_idx) const {
if (block_row_idx > max_block_idx || block_row_idx < 0){
return -1;
};
int ret = start_block_val - 1 - block_row_idx;
return ret >= m_block_min ? ret : -1;
};
__device__ int max_no_larger(int target) const {
if(max_block_idx == 0){
return -1;
};
int left = 0;
int right = max_block_idx - 1;
while (left <= right) {
int mid = left + (right - left) / 2;
if (mask_val(mid) > target) {
left = mid + 1;
} else {
right = mid - 1;
};
};
return (left < max_block_idx && mask_val(left) <= target) ? left : -1;
};
int sink_block_num, local_block_num;
int start_block_val;
int max_block_idx;
int m_block_dim, n_block_dim;
int mask_type;
int m_block_min, m_block_max;
int row_factor, col_factor;
};
struct bwdBlockmask: public bwdIteratorBase{
public:
template<typename Params, typename BlockInfo>
__device__ bwdBlockmask(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int m_block_min, int m_block_max) {
this -> row_factor = params.m_block_dim / kBlockM;
this -> col_factor = params.n_block_dim / kBlockN;
this -> max_block_idx = cute::ceil_div(binfo.actual_seqlen_q, params.m_block_dim) * row_factor;
this -> m_block_dim = params.m_block_dim;
this -> n_block_dim = params.n_block_dim;
this -> mask_type = params.head_mask_type[head_idx];
this -> m_block_min = m_block_min;
this -> m_block_max = m_block_max;
assert(mask_type > 0);
assert(params.m_block_dim % kBlockM == 0);
assert(params.n_block_dim % kBlockN == 0);
blockmask_ptr = params.blockmask + (batch_idx * params.num_blocksparse_heads + mask_type - 1) * int(params.seqlen_k_rounded / n_block_dim) * int(params.seqlen_q_rounded / m_block_dim) + int(loop_step_idx / col_factor) * int(params.seqlen_q_rounded / m_block_dim);
};
__device__ int mask_val(int block_row_idx) const {
if (block_row_idx > max_block_idx || block_row_idx < 0){
return -1;
};
int real_block_idx = block_row_idx / row_factor;
int block_row_offset = block_row_idx % row_factor;
int mask_val = blockmask_ptr[real_block_idx];
return mask_val == -1 ? -1 : row_factor * mask_val + row_factor - 1 - block_row_offset;
};
__device__ int max_no_larger(int target) const {
if(max_block_idx == 0){
return -1;
};
int left = 0;
int right = max_block_idx - 1;
while (left <= right) {
int mid = left + (right - left) / 2;
if (mask_val(mid) > target) {
left = mid + 1;
} else {
right = mid - 1;
};
};
return (left < max_block_idx && mask_val(left) <= target) ? left : -1;
};
int *blockmask_ptr;
int max_block_idx;
int m_block_dim, n_block_dim;
int mask_type;
int m_block_min, m_block_max;
int row_factor, col_factor;
};
template<bool Is_streaming>
class bwdIterator{};
template<>
struct bwdIterator<false>: public bwdBlockmask{
template<typename Params, typename BlockInfo>
__device__ bwdIterator(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int m_block_min, int m_block_max): bwdBlockmask(params, binfo, kBlockM, kBlockN, batch_idx, head_idx, loop_step_idx, m_block_min, m_block_max) {};
};
template<>
struct bwdIterator<true>: public bwdStreaming{
template<typename Params, typename BlockInfo>
__device__ bwdIterator(const Params &params, const BlockInfo &binfo, const int kBlockM, const int kBlockN, const int batch_idx, const int head_idx, const int loop_step_idx, int m_block_min, int m_block_max): bwdStreaming(params, binfo, kBlockM, kBlockN, batch_idx, head_idx, loop_step_idx, m_block_min, m_block_max) {};
};
////////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace flash
\ No newline at end of file
csrc/block_sparse_attn/src/flash_bwd_block_hdim128_bf16_sm80.cu 0 → 100644
View file @ 4f83cf8f
// Copyright (c) 2023, Tri Dao.
// Adapted by Junxian Guo from https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/src/flash_bwd_hdim128_bf16_sm80.cu
// Splitting the different head dimensions to different files to speed up compilation.
// This file is auto-generated. See "generate_kernels.py"
#include "flash_bwd_launch_template.h"
template<>
void run_mha_bwd_block_<cutlass::bfloat16_t, 128>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
run_mha_bwd_block_hdim128<cutlass::bfloat16_t>(params, stream, configure);
}
csrc/block_sparse_attn/src/flash_bwd_block_hdim128_fp16_sm80.cu 0 → 100644
View file @ 4f83cf8f
// Copyright (c) 2023, Tri Dao.
// Adapted by Junxian Guo from https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/src/flash_bwd_hdim128_fp16_sm80.cu
// Splitting the different head dimensions to different files to speed up compilation.
// This file is auto-generated. See "generate_kernels.py"
#include "flash_bwd_launch_template.h"
template<>
void run_mha_bwd_block_<cutlass::half_t, 128>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
run_mha_bwd_block_hdim128<cutlass::half_t>(params, stream, configure);
}
csrc/block_sparse_attn/src/flash_bwd_block_hdim32_bf16_sm80.cu 0 → 100644
View file @ 4f83cf8f
// Copyright (c) 2023, Tri Dao.
// Adapted by Junxian Guo from https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/src/flash_bwd_hdim32_bf16_sm80.cu
// Splitting the different head dimensions to different files to speed up compilation.
// This file is auto-generated. See "generate_kernels.py"
#include "flash_bwd_launch_template.h"
template<>
void run_mha_bwd_block_<cutlass::bfloat16_t, 32>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
run_mha_bwd_block_hdim32<cutlass::bfloat16_t>(params, stream, configure);
}
\ No newline at end of file
csrc/block_sparse_attn/src/flash_bwd_block_hdim32_fp16_sm80.cu 0 → 100644
View file @ 4f83cf8f
// Copyright (c) 2023, Tri Dao.
// Adapted by Junxian Guo from https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/src/flash_bwd_hdim32_fp16_sm80.cu
// Splitting the different head dimensions to different files to speed up compilation.
// This file is auto-generated. See "generate_kernels.py"
#include "flash_bwd_launch_template.h"
template<>
void run_mha_bwd_block_<cutlass::half_t, 32>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
run_mha_bwd_block_hdim32<cutlass::half_t>(params, stream, configure);
}
csrc/block_sparse_attn/src/flash_bwd_block_hdim64_bf16_sm80.cu 0 → 100644
View file @ 4f83cf8f
// Copyright (c) 2023, Tri Dao.
// Adapted by Junxian Guo from https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/src/flash_bwd_hdim64_bf16_sm80.cu
// Splitting the different head dimensions to different files to speed up compilation.
// This file is auto-generated. See "generate_kernels.py"
#include "flash_bwd_launch_template.h"
template<>
void run_mha_bwd_block_<cutlass::bfloat16_t, 64>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
run_mha_bwd_block_hdim64<cutlass::bfloat16_t>(params, stream, configure);
}
csrc/block_sparse_attn/src/flash_bwd_block_hdim64_fp16_sm80.cu 0 → 100644
View file @ 4f83cf8f
// Copyright (c) 2023, Tri Dao.
// Adapted by Junxian Guo from https://github.com/Dao-AILab/flash-attention/blob/main/csrc/flash_attn/src/flash_bwd_hdim64_fp16_sm80.cu
// Splitting the different head dimensions to different files to speed up compilation.
// This file is auto-generated. See "generate_kernels.py"
#include "flash_bwd_launch_template.h"
template<>
void run_mha_bwd_block_<cutlass::half_t, 64>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
run_mha_bwd_block_hdim64<cutlass::half_t>(params, stream, configure);
}
csrc/block_sparse_attn/src/flash_bwd_hdim128_bf16_sm80.cu 0 → 100644
View file @ 4f83cf8f
// Copyright (c) 2023, Tri Dao.
// Splitting the different head dimensions to different files to speed up compilation.
// This file is auto-generated. See "generate_kernels.py"
#include "flash_bwd_launch_template.h"
template<>
void run_mha_bwd_<cutlass::bfloat16_t, 128>(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
run_mha_bwd_hdim128<cutlass::bfloat16_t>(params, stream, configure);
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment