Skip to content

GitLab

"...git@developer.sourcefind.cn:OpenDAS/torch-harmonics.git" did not exist on "c7afb5461e6c4c9298ce5afd3ada1f8436cdc15d"
Commit 5a4263f4 authored by Ruff's avatar Ruff Committed by Aarni Koskela
Browse files

Reformat with ruff-format

parent 02e30ca6
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 914 additions and 902 deletions
bitsandbytes/triton/int8_matmul_mixed_dequantize.py
View file @ 5a4263f4
...@@ -3,14 +3,14 @@ import torch ...@@ -3,14 +3,14 @@ import torch
from bitsandbytes.triton.triton_utils import is_triton_available from bitsandbytes.triton.triton_utils import is_triton_available
if not is_triton_available(): if not is_triton_available():
def int8_matmul_mixed_dequantize(a, b, state_x, state_w, bias): return None
else:
def int8_matmul_mixed_dequantize(a, b, state_x, state_w, bias):
return None
else:
import triton import triton
import triton.language as tl import triton.language as tl
from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
# This is a matmul kernel based on triton.ops.matmul # This is a matmul kernel based on triton.ops.matmul
# It is modified to support rowwise quantized input and global quantized weight # It is modified to support rowwise quantized input and global quantized weight
# It's purpose is fused matmul then dequantize # It's purpose is fused matmul then dequantize
...@@ -27,58 +27,83 @@ else: ...@@ -27,58 +27,83 @@ else:
for block_n in [32, 64, 128, 256]: for block_n in [32, 64, 128, 256]:
num_warps = 2 if block_n <= 64 else 4 num_warps = 2 if block_n <= 64 else 4
configs.append( configs.append(
triton.Config({'BLOCK_M': block_m, 'BLOCK_N': block_n, 'BLOCK_K': block_k, 'SPLIT_K': 1}, triton.Config(
num_stages=num_stages, num_warps=num_warps)) {"BLOCK_M": block_m, "BLOCK_N": block_n, "BLOCK_K": block_k, "SPLIT_K": 1},
num_stages=num_stages,
num_warps=num_warps,
),
)
# split_k # split_k
for split_k in [2, 4, 8, 16]: for split_k in [2, 4, 8, 16]:
configs.append(triton.Config({'BLOCK_M': block_m, 'BLOCK_N': block_n, 'BLOCK_K': block_k, 'SPLIT_K': split_k}, configs.append(
num_stages=num_stages, num_warps=num_warps, pre_hook=init_to_zero('C'))) triton.Config(
{"BLOCK_M": block_m, "BLOCK_N": block_n, "BLOCK_K": block_k, "SPLIT_K": split_k},
num_stages=num_stages,
num_warps=num_warps,
pre_hook=init_to_zero("C"),
),
)
return configs return configs
@triton.autotune( @triton.autotune(
configs=[ configs=[
# basic configs for compute-bound matmuls # basic configs for compute-bound matmuls
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=3, num_warps=8), triton.Config({"BLOCK_M": 128, "BLOCK_N": 256, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=3, num_warps=8), triton.Config({"BLOCK_M": 256, "BLOCK_N": 128, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 256, "BLOCK_N": 64, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 64, "BLOCK_N": 256, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 128, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 64, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 64, "BLOCK_N": 128, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 32, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=5, num_warps=2), triton.Config({"BLOCK_M": 64, "BLOCK_N": 32, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=5, num_warps=2),
# good for int8 # good for int8
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=3, num_warps=8), triton.Config({"BLOCK_M": 128, "BLOCK_N": 256, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=3, num_warps=8), triton.Config({"BLOCK_M": 256, "BLOCK_N": 128, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 256, "BLOCK_N": 64, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 64, "BLOCK_N": 256, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 128, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 64, "BLOCK_K": 64, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 128, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 64, "BLOCK_N": 128, "BLOCK_K": 64, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 32, "BLOCK_K": 64, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=5, num_warps=2), triton.Config({"BLOCK_M": 64, "BLOCK_N": 32, "BLOCK_K": 64, "SPLIT_K": 1}, num_stages=5, num_warps=2),
*get_configs_io_bound(), *get_configs_io_bound(),
], ],
key=['M', 'N', 'K'], key=["M", "N", "K"],
prune_configs_by={ prune_configs_by={"early_config_prune": early_config_prune, "perf_model": estimate_matmul_time, "top_k": 10},
'early_config_prune': early_config_prune, )
'perf_model': estimate_matmul_time, @triton.heuristics(
'top_k': 10 {
"EVEN_K": lambda args: args["K"] % (args["BLOCK_K"] * args["SPLIT_K"]) == 0,
}, },
) )
@triton.heuristics({
'EVEN_K': lambda args: args['K'] % (args['BLOCK_K'] * args['SPLIT_K']) == 0,
})
@triton.jit @triton.jit
def _int8_matmul_mixed_dequantize(A, B, C, bias, state_x_ptr, state_w_ptr, M, N, K, divfactor: tl.constexpr, has_bias : tl.constexpr, def _int8_matmul_mixed_dequantize(
stride_am, stride_ak, A,
stride_bk, stride_bn, B,
stride_cm, stride_cn, C,
BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr, BLOCK_K: tl.constexpr, bias,
GROUP_M: tl.constexpr, SPLIT_K: tl.constexpr, EVEN_K: tl.constexpr, state_x_ptr,
ACC_TYPE: tl.constexpr state_w_ptr,
): M,
N,
K,
divfactor: tl.constexpr,
has_bias: tl.constexpr,
stride_am,
stride_ak,
stride_bk,
stride_bn,
stride_cm,
stride_cn,
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr,
GROUP_M: tl.constexpr,
SPLIT_K: tl.constexpr,
EVEN_K: tl.constexpr,
ACC_TYPE: tl.constexpr,
):
# matrix multiplication # matrix multiplication
pid = tl.program_id(0) pid = tl.program_id(0)
pid_z = tl.program_id(1) pid_z = tl.program_id(1)
...@@ -115,13 +140,13 @@ else: ...@@ -115,13 +140,13 @@ else:
b = tl.load(B) b = tl.load(B)
else: else:
k_remaining = K - k * (BLOCK_K * SPLIT_K) k_remaining = K - k * (BLOCK_K * SPLIT_K)
a = tl.load(A, mask=rk[None, :] < k_remaining, other=0.) a = tl.load(A, mask=rk[None, :] < k_remaining, other=0.0)
b = tl.load(B, mask=rk[:, None] < k_remaining, other=0.) b = tl.load(B, mask=rk[:, None] < k_remaining, other=0.0)
acc += tl.dot(a, b) acc += tl.dot(a, b)
A += BLOCK_K * SPLIT_K * stride_ak A += BLOCK_K * SPLIT_K * stride_ak
B += BLOCK_K * SPLIT_K * stride_bk B += BLOCK_K * SPLIT_K * stride_bk
acc = (w_factor * (x_factor * (acc * divfactor))) acc = w_factor * (x_factor * (acc * divfactor))
acc = acc.to(C.dtype.element_ty) acc = acc.to(C.dtype.element_ty)
# conditionally add bias # conditionally add bias
...@@ -137,10 +162,9 @@ else: ...@@ -137,10 +162,9 @@ else:
else: else:
tl.atomic_add(C, acc, mask=mask) tl.atomic_add(C, acc, mask=mask)
def int8_matmul_mixed_dequantize(a, b, state_x, state_w, bias): def int8_matmul_mixed_dequantize(a, b, state_x, state_w, bias):
device = a.device device = a.device
divfactor = 1. / (127. * 127.) divfactor = 1.0 / (127.0 * 127.0)
has_bias = 0 if bias is None else 1 has_bias = 0 if bias is None else 1
# handle non-contiguous inputs if necessary # handle non-contiguous inputs if necessary
if a.stride(0) > 1 and a.stride(1) > 1: if a.stride(0) > 1 and a.stride(1) > 1:
...@@ -154,12 +178,28 @@ else: ...@@ -154,12 +178,28 @@ else:
# allocates output # allocates output
c = torch.empty((M, N), device=device, dtype=torch.float16) c = torch.empty((M, N), device=device, dtype=torch.float16)
# accumulator types # accumulator types
ACC_TYPE = tl.float32 #if a.dtype in [torch.float16, torch.bfloat16, torch.float32] else tl.int32 ACC_TYPE = tl.float32 # if a.dtype in [torch.float16, torch.bfloat16, torch.float32] else tl.int32
# launch int8_matmul_mixed_dequantize kernel # launch int8_matmul_mixed_dequantize kernel
grid = lambda META: (triton.cdiv(M, META['BLOCK_M']) * triton.cdiv(N, META['BLOCK_N']), META['SPLIT_K']) grid = lambda META: (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]), META["SPLIT_K"])
_int8_matmul_mixed_dequantize[grid](a, b, c, bias, state_x, state_w, M, N, K, divfactor, has_bias, _int8_matmul_mixed_dequantize[grid](
a.stride(0), a.stride(1), a,
b.stride(0), b.stride(1), b,
c.stride(0), c.stride(1), c,
GROUP_M=8, ACC_TYPE=ACC_TYPE) bias,
state_x,
state_w,
M,
N,
K,
divfactor,
has_bias,
a.stride(0),
a.stride(1),
b.stride(0),
b.stride(1),
c.stride(0),
c.stride(1),
GROUP_M=8,
ACC_TYPE=ACC_TYPE,
)
return c return c
bitsandbytes/triton/int8_matmul_rowwise_dequantize.py
View file @ 5a4263f4
...@@ -3,7 +3,9 @@ import torch ...@@ -3,7 +3,9 @@ import torch
from bitsandbytes.triton.triton_utils import is_triton_available from bitsandbytes.triton.triton_utils import is_triton_available
if not is_triton_available(): if not is_triton_available():
def int8_matmul_rowwise_dequantize(a, b, state_x, state_w, bias): return None
def int8_matmul_rowwise_dequantize(a, b, state_x, state_w, bias):
return None
else: else:
import triton import triton
import triton.language as tl import triton.language as tl
...@@ -17,7 +19,6 @@ else: ...@@ -17,7 +19,6 @@ else:
def init_to_zero(name): def init_to_zero(name):
return lambda nargs: nargs[name].zero_() return lambda nargs: nargs[name].zero_()
def get_configs_io_bound(): def get_configs_io_bound():
configs = [] configs = []
for num_stages in [2, 3, 4, 5, 6]: for num_stages in [2, 3, 4, 5, 6]:
...@@ -26,58 +27,83 @@ else: ...@@ -26,58 +27,83 @@ else:
for block_n in [32, 64, 128, 256]: for block_n in [32, 64, 128, 256]:
num_warps = 2 if block_n <= 64 else 4 num_warps = 2 if block_n <= 64 else 4
configs.append( configs.append(
triton.Config({'BLOCK_M': block_m, 'BLOCK_N': block_n, 'BLOCK_K': block_k, 'SPLIT_K': 1}, triton.Config(
num_stages=num_stages, num_warps=num_warps)) {"BLOCK_M": block_m, "BLOCK_N": block_n, "BLOCK_K": block_k, "SPLIT_K": 1},
num_stages=num_stages,
num_warps=num_warps,
),
)
# split_k # split_k
for split_k in [2, 4, 8, 16]: for split_k in [2, 4, 8, 16]:
configs.append(triton.Config({'BLOCK_M': block_m, 'BLOCK_N': block_n, 'BLOCK_K': block_k, 'SPLIT_K': split_k}, configs.append(
num_stages=num_stages, num_warps=num_warps, pre_hook=init_to_zero('C'))) triton.Config(
{"BLOCK_M": block_m, "BLOCK_N": block_n, "BLOCK_K": block_k, "SPLIT_K": split_k},
num_stages=num_stages,
num_warps=num_warps,
pre_hook=init_to_zero("C"),
),
)
return configs return configs
@triton.autotune( @triton.autotune(
configs=[ configs=[
# basic configs for compute-bound matmuls # basic configs for compute-bound matmuls
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=3, num_warps=8), triton.Config({"BLOCK_M": 128, "BLOCK_N": 256, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=3, num_warps=8), triton.Config({"BLOCK_M": 256, "BLOCK_N": 128, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 256, "BLOCK_N": 64, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 64, "BLOCK_N": 256, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 128, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 64, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 64, "BLOCK_N": 128, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 32, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=5, num_warps=2), triton.Config({"BLOCK_M": 64, "BLOCK_N": 32, "BLOCK_K": 32, "SPLIT_K": 1}, num_stages=5, num_warps=2),
# good for int8 # good for int8
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=3, num_warps=8), triton.Config({"BLOCK_M": 128, "BLOCK_N": 256, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=3, num_warps=8), triton.Config({"BLOCK_M": 256, "BLOCK_N": 128, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=3, num_warps=8),
triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 256, "BLOCK_N": 64, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 64, "BLOCK_N": 256, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 128, "BLOCK_K": 128, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 64, "BLOCK_K": 64, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 128, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 64, "BLOCK_N": 128, "BLOCK_K": 64, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4), triton.Config({"BLOCK_M": 128, "BLOCK_N": 32, "BLOCK_K": 64, "SPLIT_K": 1}, num_stages=4, num_warps=4),
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=5, num_warps=2), triton.Config({"BLOCK_M": 64, "BLOCK_N": 32, "BLOCK_K": 64, "SPLIT_K": 1}, num_stages=5, num_warps=2),
*get_configs_io_bound(), *get_configs_io_bound(),
], ],
key=['M', 'N', 'K'], key=["M", "N", "K"],
prune_configs_by={ prune_configs_by={"early_config_prune": early_config_prune, "perf_model": estimate_matmul_time, "top_k": 10},
'early_config_prune': early_config_prune, )
'perf_model': estimate_matmul_time, @triton.heuristics(
'top_k': 10 {
"EVEN_K": lambda args: args["K"] % (args["BLOCK_K"] * args["SPLIT_K"]) == 0,
}, },
) )
@triton.heuristics({
'EVEN_K': lambda args: args['K'] % (args['BLOCK_K'] * args['SPLIT_K']) == 0,
})
@triton.jit @triton.jit
def _int8_matmul_rowwise_dequantize(A, B, C, bias, state_x_ptr, state_w_ptr, M, N, K, divfactor, has_bias : tl.constexpr, def _int8_matmul_rowwise_dequantize(
stride_am, stride_ak, A,
stride_bk, stride_bn, B,
stride_cm, stride_cn, C,
BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr, BLOCK_K: tl.constexpr, bias,
GROUP_M: tl.constexpr, SPLIT_K: tl.constexpr, EVEN_K: tl.constexpr, state_x_ptr,
ACC_TYPE: tl.constexpr state_w_ptr,
): M,
N,
K,
divfactor,
has_bias: tl.constexpr,
stride_am,
stride_ak,
stride_bk,
stride_bn,
stride_cm,
stride_cn,
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr,
GROUP_M: tl.constexpr,
SPLIT_K: tl.constexpr,
EVEN_K: tl.constexpr,
ACC_TYPE: tl.constexpr,
):
# matrix multiplication # matrix multiplication
pid = tl.program_id(0) pid = tl.program_id(0)
pid_z = tl.program_id(1) pid_z = tl.program_id(1)
...@@ -114,13 +140,13 @@ else: ...@@ -114,13 +140,13 @@ else:
b = tl.load(B) b = tl.load(B)
else: else:
k_remaining = K - k * (BLOCK_K * SPLIT_K) k_remaining = K - k * (BLOCK_K * SPLIT_K)
a = tl.load(A, mask=rk[None, :] < k_remaining, other=0.) a = tl.load(A, mask=rk[None, :] < k_remaining, other=0.0)
b = tl.load(B, mask=rk[:, None] < k_remaining, other=0.) b = tl.load(B, mask=rk[:, None] < k_remaining, other=0.0)
acc += tl.dot(a, b) acc += tl.dot(a, b)
A += BLOCK_K * SPLIT_K * stride_ak A += BLOCK_K * SPLIT_K * stride_ak
B += BLOCK_K * SPLIT_K * stride_bk B += BLOCK_K * SPLIT_K * stride_bk
acc = (w_factor * (x_factor * (acc * divfactor))) acc = w_factor * (x_factor * (acc * divfactor))
acc = acc.to(C.dtype.element_ty) acc = acc.to(C.dtype.element_ty)
if has_bias: if has_bias:
...@@ -135,9 +161,8 @@ else: ...@@ -135,9 +161,8 @@ else:
else: else:
tl.atomic_add(C, acc, mask=mask) tl.atomic_add(C, acc, mask=mask)
def int8_matmul_rowwise_dequantize(a, b, state_x, state_w, bias): def int8_matmul_rowwise_dequantize(a, b, state_x, state_w, bias):
divfactor = 1. / (127. * 127.) divfactor = 1.0 / (127.0 * 127.0)
has_bias = 0 if bias is None else 1 has_bias = 0 if bias is None else 1
...@@ -154,12 +179,28 @@ else: ...@@ -154,12 +179,28 @@ else:
# allocates output # allocates output
c = torch.empty((M, N), device=device, dtype=torch.float16) c = torch.empty((M, N), device=device, dtype=torch.float16)
# accumulator types # accumulator types
ACC_TYPE = tl.float32 #if a.dtype in [torch.float16, torch.bfloat16, torch.float32] else tl.int32 ACC_TYPE = tl.float32 # if a.dtype in [torch.float16, torch.bfloat16, torch.float32] else tl.int32
# launch int8_matmul_rowwise_dequantize kernel # launch int8_matmul_rowwise_dequantize kernel
grid = lambda META: (triton.cdiv(M, META['BLOCK_M']) * triton.cdiv(N, META['BLOCK_N']), META['SPLIT_K']) grid = lambda META: (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]), META["SPLIT_K"])
_int8_matmul_rowwise_dequantize[grid](a, b, c, bias, state_x, state_w, M, N, K, divfactor, has_bias, _int8_matmul_rowwise_dequantize[grid](
a.stride(0), a.stride(1), a,
b.stride(0), b.stride(1), b,
c.stride(0), c.stride(1), c,
GROUP_M=8, ACC_TYPE=ACC_TYPE) bias,
state_x,
state_w,
M,
N,
K,
divfactor,
has_bias,
a.stride(0),
a.stride(1),
b.stride(0),
b.stride(1),
c.stride(0),
c.stride(1),
GROUP_M=8,
ACC_TYPE=ACC_TYPE,
)
return c return c
bitsandbytes/triton/quantize_columnwise_and_transpose.py
View file @ 5a4263f4
...@@ -5,9 +5,10 @@ import torch ...@@ -5,9 +5,10 @@ import torch
from bitsandbytes.triton.triton_utils import is_triton_available from bitsandbytes.triton.triton_utils import is_triton_available
if not is_triton_available(): if not is_triton_available():
def quantize_columnwise_and_transpose(x: torch.Tensor): return None
else:
def quantize_columnwise_and_transpose(x: torch.Tensor):
return None
else:
import triton import triton
import triton.language as tl import triton.language as tl
...@@ -15,23 +16,23 @@ else: ...@@ -15,23 +16,23 @@ else:
# TODO: autotune this better. # TODO: autotune this better.
@triton.autotune( @triton.autotune(
configs=[ configs=[
triton.Config({}, num_stages=1), triton.Config({}, num_stages=1),
triton.Config({}, num_stages=2), triton.Config({}, num_stages=2),
triton.Config({}, num_stages=4), triton.Config({}, num_stages=4),
triton.Config({}, num_stages=8), triton.Config({}, num_stages=8),
triton.Config({}, num_stages=16), triton.Config({}, num_stages=16),
triton.Config({}, num_stages=1, num_warps=8), triton.Config({}, num_stages=1, num_warps=8),
triton.Config({}, num_stages=2, num_warps=8), triton.Config({}, num_stages=2, num_warps=8),
triton.Config({}, num_stages=4, num_warps=8), triton.Config({}, num_stages=4, num_warps=8),
triton.Config({}, num_stages=8, num_warps=8), triton.Config({}, num_stages=8, num_warps=8),
triton.Config({}, num_stages=16, num_warps=8), triton.Config({}, num_stages=16, num_warps=8),
triton.Config({}, num_warps=1), triton.Config({}, num_warps=1),
triton.Config({}, num_warps=2), triton.Config({}, num_warps=2),
triton.Config({}, num_warps=4), triton.Config({}, num_warps=4),
triton.Config({}, num_warps=8), triton.Config({}, num_warps=8),
], ],
key=['n_elements'] key=["n_elements"],
) )
@triton.jit @triton.jit
def _quantize_columnwise_and_transpose( def _quantize_columnwise_and_transpose(
...@@ -39,7 +40,8 @@ else: ...@@ -39,7 +40,8 @@ else:
output_ptr, output_ptr,
output_maxs, output_maxs,
n_elements, n_elements,
M : tl.constexpr, N : tl.constexpr, M: tl.constexpr,
N: tl.constexpr,
BLOCK_SIZE: tl.constexpr, BLOCK_SIZE: tl.constexpr,
P2: tl.constexpr, P2: tl.constexpr,
): ):
...@@ -47,12 +49,12 @@ else: ...@@ -47,12 +49,12 @@ else:
block_start = pid block_start = pid
p2_arange = tl.arange(0, P2) p2_arange = tl.arange(0, P2)
p2_arange_mask = p2_arange < M p2_arange_mask = p2_arange < M
arange = p2_arange * N arange = p2_arange * N
offsets = block_start + arange offsets = block_start + arange
x = tl.load(x_ptr + offsets, mask=p2_arange_mask) x = tl.load(x_ptr + offsets, mask=p2_arange_mask)
abs_x = tl.abs(x) abs_x = tl.abs(x)
max_val = tl.max(tl.where(p2_arange_mask, abs_x, 0), axis=0) max_val = tl.max(tl.where(p2_arange_mask, abs_x, 0), axis=0)
output = tl.libdevice.llrint(127. * (x / max_val)) output = tl.libdevice.llrint(127.0 * (x / max_val))
new_start = pid * M new_start = pid * M
new_offsets = new_start + p2_arange new_offsets = new_start + p2_arange
...@@ -68,6 +70,6 @@ else: ...@@ -68,6 +70,6 @@ else:
assert x.is_cuda and output.is_cuda assert x.is_cuda and output.is_cuda
n_elements = output.numel() n_elements = output.numel()
grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),) grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),)
_quantize_columnwise_and_transpose[grid](x, output, output_maxs, n_elements, M, N, BLOCK_SIZE=M, P2=P2) _quantize_columnwise_and_transpose[grid](x, output, output_maxs, n_elements, M, N, BLOCK_SIZE=M, P2=P2)
return output, output_maxs return output, output_maxs
bitsandbytes/triton/quantize_global.py
View file @ 5a4263f4
This diff is collapsed.
bitsandbytes/triton/quantize_rowwise.py
View file @ 5a4263f4
...@@ -5,9 +5,10 @@ import torch ...@@ -5,9 +5,10 @@ import torch
from bitsandbytes.triton.triton_utils import is_triton_available from bitsandbytes.triton.triton_utils import is_triton_available
if not is_triton_available(): if not is_triton_available():
def quantize_rowwise(x: torch.Tensor): return None
else:
def quantize_rowwise(x: torch.Tensor):
return None
else:
import triton import triton
import triton.language as tl import triton.language as tl
...@@ -15,21 +16,21 @@ else: ...@@ -15,21 +16,21 @@ else:
# TODO: autotune this better. # TODO: autotune this better.
@triton.autotune( @triton.autotune(
configs=[ configs=[
triton.Config({}, num_stages=1, num_warps=8), triton.Config({}, num_stages=1, num_warps=8),
triton.Config({}, num_stages=2, num_warps=8), triton.Config({}, num_stages=2, num_warps=8),
triton.Config({}, num_stages=4, num_warps=8), triton.Config({}, num_stages=4, num_warps=8),
triton.Config({}, num_stages=8, num_warps=8), triton.Config({}, num_stages=8, num_warps=8),
triton.Config({}, num_stages=1), triton.Config({}, num_stages=1),
triton.Config({}, num_stages=2), triton.Config({}, num_stages=2),
triton.Config({}, num_stages=4), triton.Config({}, num_stages=4),
triton.Config({}, num_stages=8), triton.Config({}, num_stages=8),
triton.Config({}, num_warps=1), triton.Config({}, num_warps=1),
triton.Config({}, num_warps=2), triton.Config({}, num_warps=2),
triton.Config({}, num_warps=4), triton.Config({}, num_warps=4),
triton.Config({}, num_warps=8), triton.Config({}, num_warps=8),
], ],
key=['n_elements'] key=["n_elements"],
) )
@triton.jit @triton.jit
def _quantize_rowwise( def _quantize_rowwise(
...@@ -49,7 +50,7 @@ else: ...@@ -49,7 +50,7 @@ else:
abs_x = tl.abs(x) abs_x = tl.abs(x)
max_val = tl.max(tl.where(row_mask, abs_x, 0), axis=0) max_val = tl.max(tl.where(row_mask, abs_x, 0), axis=0)
output = tl.libdevice.llrint(127. * (x / max_val)) output = tl.libdevice.llrint(127.0 * (x / max_val))
tl.store(output_ptr + offsets, output, mask=row_mask) tl.store(output_ptr + offsets, output, mask=row_mask)
tl.store(output_maxs + pid, max_val) tl.store(output_maxs + pid, max_val)
......
bitsandbytes/utils.py
View file @ 5a4263f4
This diff is collapsed.
check_bnb_install.py
View file @ 5a4263f4
This diff is collapsed.
examples/int8_inference_huggingface.py
View file @ 5a4263f4
This diff is collapsed.
install_cuda.py
View file @ 5a4263f4
This diff is collapsed.
scripts/stale.py
View file @ 5a4263f4
This diff is collapsed.
tests/test_autograd.py
View file @ 5a4263f4
This diff is collapsed.
tests/test_cuda_setup_evaluator.py
View file @ 5a4263f4
...@@ -35,7 +35,4 @@ def test_get_cuda_bnb_library_path_override(monkeypatch, cuda120_spec, caplog): ...@@ -35,7 +35,4 @@ def test_get_cuda_bnb_library_path_override(monkeypatch, cuda120_spec, caplog):
def test_get_cuda_bnb_library_path_nocublaslt(monkeypatch, cuda111_noblas_spec): def test_get_cuda_bnb_library_path_nocublaslt(monkeypatch, cuda111_noblas_spec):
monkeypatch.delenv("BNB_CUDA_VERSION", raising=False) monkeypatch.delenv("BNB_CUDA_VERSION", raising=False)
assert ( assert get_cuda_bnb_library_path(cuda111_noblas_spec).stem == "libbitsandbytes_cuda111_nocublaslt"
get_cuda_bnb_library_path(cuda111_noblas_spec).stem
== "libbitsandbytes_cuda111_nocublaslt"
)
tests/test_functional.py
View file @ 5a4263f4
This diff is collapsed.
tests/test_generation.py
View file @ 5a4263f4
This diff is collapsed.
tests/test_linear4bit.py
View file @ 5a4263f4
This diff is collapsed.
tests/test_linear8bitlt.py
View file @ 5a4263f4
This diff is collapsed.
tests/test_modules.py
View file @ 5a4263f4
This diff is collapsed.
tests/test_optim.py
View file @ 5a4263f4
This diff is collapsed.
tests/test_triton.py
View file @ 5a4263f4
This diff is collapsed.
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