mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
238 lines
7.2 KiB
Python
238 lines
7.2 KiB
Python
# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
#
|
|
#from IPython.core.display import display, HTML
|
|
#display(HTML("<style>.container { width:85% !important; }</style>"))
|
|
#import numpy as np
|
|
#np.set_printoptions(edgeitems=1000, linewidth=1000000)
|
|
|
|
import sys
|
|
|
|
my_path = '/data/projects/fmha_v2/train_ops/build/'
|
|
if my_path not in sys.path:
|
|
sys.path.insert(0, my_path)
|
|
#my_path = '/data/projects/apex_gitlab/apex/contrib/csrc/fmha/build'
|
|
#if my_path not in sys.path:
|
|
# sys.path.insert(0, my_path)
|
|
|
|
import math
|
|
|
|
import apex_mha
|
|
import bert_mha_train as mha
|
|
import numpy as np
|
|
import torch
|
|
|
|
#import fmhalib as mha
|
|
|
|
torch.manual_seed(1234)
|
|
torch.cuda.manual_seed(1234)
|
|
|
|
|
|
class GPUTimer:
|
|
|
|
def __init__(self, stream):
|
|
self.start_ = torch.cuda.Event(enable_timing=True)
|
|
self.stop_ = torch.cuda.Event(enable_timing=True)
|
|
self.stream_ = stream
|
|
|
|
def start(self):
|
|
self.stream_.record_event(self.start_)
|
|
|
|
def stop(self):
|
|
self.stream_.record_event(self.stop_)
|
|
|
|
def sync(self):
|
|
self.stream_.synchronize()
|
|
|
|
def millis(self):
|
|
return self.start_.elapsed_time(self.stop_)
|
|
|
|
|
|
def reshape_softmax(S, b, s, h, d, warps_m, warps_n):
|
|
m = s if s == 128 else 16
|
|
n = s
|
|
m_per_cta = warps_m * 16
|
|
n_per_cta = warps_n * 16
|
|
mmas_m = m // m_per_cta
|
|
mmas_n = n // n_per_cta
|
|
loops = s // (mmas_m * m_per_cta)
|
|
print(loops, )
|
|
assert (loops == 1 and s == 128) or (loops == 16 and s == 256) or (
|
|
loops == 32 and s == 512) or (loops == 24 and s == 384), "no.."
|
|
quads = 8
|
|
lohi = 2
|
|
lr = 2
|
|
vals = 2
|
|
# B x H x LOOPS x MMAS_M x MMAS_N x THREADS_PER_CTA x LOHI x LR x 2
|
|
# B x H x LOOPS x MMAS_M x MMAS_N x WARPS_N x WARPS_M x 32 x LOHI x LR x 2
|
|
# 0 1 2 3 4 5 6 7 8 9 10 11
|
|
# B x H x LOOPS x MMAS_M x MMAS_N x WARPS_N x WARPS_M x QUADS x 4 x LOHI x LR x 2
|
|
# 0 1 2 3 6 9 7 4 5 10 8 11
|
|
# B x H x LOOPS x MMAS_M x WARPS_M x LOHI x QUADS x MMAS_N x WARPS_N x LR x 4 x 2
|
|
|
|
S = S.reshape((b, h, loops, mmas_m, mmas_n, warps_n, warps_m, quads, 4,
|
|
lohi, lr, vals)).permute(0, 1, 2, 3, 6, 9, 7, 4, 5, 10, 8,
|
|
11).reshape((b, h, s, s))
|
|
Snp = S.cpu().numpy()
|
|
dmask = torch.tensor(np.logical_not(np.signbit(Snp)),
|
|
dtype=torch.float32,
|
|
device=device)
|
|
S = S.abs()
|
|
return S, dmask
|
|
|
|
|
|
def mha_ref(qkv, D, b, s, h, d, p_dropout):
|
|
qkv = qkv.view(b, s, h, 3, d)
|
|
q = qkv[:, :, :, 0, :].permute(0, 2, 1, 3)
|
|
k = qkv[:, :, :, 1, :].permute(0, 2, 1, 3)
|
|
v = qkv[:, :, :, 2, :].permute(0, 2, 1, 3)
|
|
p = torch.matmul(q.float(), k.permute(0, 1, 3, 2).float()) / math.sqrt(d)
|
|
s = torch.softmax(p, -1).half()
|
|
d = s * D.half() * (1 / (1 - p_dropout))
|
|
#d = s
|
|
ctx = torch.matmul(d, v)
|
|
|
|
return ctx, p, s
|
|
|
|
|
|
runs = 1
|
|
|
|
s = 512
|
|
b = 32
|
|
|
|
warps_m = 1
|
|
warps_n = 4
|
|
if s == 256:
|
|
runs == 20000
|
|
if s == 384:
|
|
runs == 20000
|
|
elif s == 512:
|
|
warps_n = 8
|
|
runs == 5000
|
|
elif s == 128:
|
|
runs = 20000
|
|
|
|
#runs = 1
|
|
|
|
h = 16
|
|
d = 64
|
|
p_dropout = 0.1
|
|
dtype = torch.float16
|
|
device = torch.device('cuda')
|
|
|
|
if b <= 4: runs *= 10
|
|
|
|
#s_valid = int(s * 0.5)
|
|
#s_valid = int(s * 0.97)
|
|
s_valid = 246 # average per batch
|
|
s_valid = s
|
|
|
|
#slens = [s_valid] * b
|
|
#a = torch.tensor(np.array([0] + slens), dtype=torch.int32)
|
|
#cu_seqlens =torch.cumsum(a, 0).to(dtype=torch.int32, device=device)
|
|
|
|
seqlens = torch.linspace(1, s, b, dtype=torch.int32, device=device)
|
|
seqlens = torch.ones(b) * s
|
|
|
|
cu_seqlens = torch.zeros(b + 1, dtype=torch.int32, device=device)
|
|
cu_seqlens[1:] = torch.cumsum(seqlens, 0)
|
|
total = cu_seqlens[-1].item()
|
|
print(seqlens)
|
|
print(cu_seqlens)
|
|
|
|
assert cu_seqlens.numel() == b + 1, "ahh"
|
|
|
|
qkv = torch.randn((b, s, h, 3, d), device=device, dtype=dtype)
|
|
#qkv_vs = qkv[:, :s_valid, :, : ,:].contiguous().view(total, h, 3, d).permute(0,2,1,3).contiguous()
|
|
qkv_vs = torch.empty((total, h, 3, d), dtype=dtype, device=device)
|
|
for bi in range(b):
|
|
begin = cu_seqlens[bi]
|
|
end = cu_seqlens[bi + 1]
|
|
si = end - begin
|
|
qkv_vs[begin:end, ...] = qkv[bi, :si, ...]
|
|
qkv_vs = qkv_vs.contiguous().view(total, h, 3, d).permute(0, 2, 1,
|
|
3).contiguous()
|
|
qkvt = qkv.view((b, s, h, 3, d)).permute(1, 0, 2, 3, 4).contiguous()
|
|
mask = torch.ones((b, s), device=device, dtype=dtype)
|
|
|
|
stream = torch.cuda.Stream()
|
|
with torch.cuda.stream(stream):
|
|
assert torch.cuda.current_stream() == stream
|
|
|
|
timer = GPUTimer(stream)
|
|
|
|
for it in range(runs):
|
|
Pa, Sa, Da, Ca = apex_mha.fwd(qkvt, mask, p_dropout)
|
|
is_nl = b < 4
|
|
is_training = True
|
|
|
|
timer.start()
|
|
for it in range(runs):
|
|
ctx, S = mha.fwd(qkv_vs, cu_seqlens, p_dropout, s, is_training, is_nl,
|
|
None)
|
|
timer.stop()
|
|
timer.sync()
|
|
ms_fused = timer.millis() / runs
|
|
# S will be overwritten in the backward pass, so reshape it here already.
|
|
Snew, D = reshape_softmax(S, b, s, h, d, warps_m, warps_n)
|
|
|
|
timer.start()
|
|
for it in range(runs):
|
|
Pa, Sa, Da, Ca = apex_mha.fwd(qkvt, mask, p_dropout)
|
|
timer.stop()
|
|
timer.sync()
|
|
ms_apex = timer.millis() / runs
|
|
|
|
print(Ca.shape, qkvt.shape)
|
|
timer.start()
|
|
for it in range(runs):
|
|
dqkv_a, dU_a = apex_mha.bwd(h, Ca, Ca, Sa, Pa, mask, qkvt, Da,
|
|
p_dropout)
|
|
timer.stop()
|
|
timer.sync()
|
|
ms_apex_bwd = timer.millis() / runs
|
|
|
|
timer.start()
|
|
for it in range(runs):
|
|
if b < 4 and b > 1:
|
|
_ = mha.bwd_nl(ctx, qkv_vs, S, cu_seqlens, p_dropout, s)
|
|
else:
|
|
dqkv2, dp_mma = mha.bwd(ctx, qkv_vs, S, cu_seqlens, p_dropout, s)
|
|
timer.stop()
|
|
timer.sync()
|
|
ms_fused_bwd = timer.millis() / runs
|
|
|
|
ctx_ref, Pref, Sref = mha_ref(qkv, D, b, s, h, d, p_dropout)
|
|
ctx_ref = ctx_ref.permute(0, 2, 1, 3)
|
|
#ctx = ctx.view((b,s,h,d))
|
|
|
|
ctx_pad = torch.zeros((b, s, h, d), dtype=dtype, device=device)
|
|
for bi in range(b):
|
|
begin = cu_seqlens[bi]
|
|
end = cu_seqlens[bi + 1]
|
|
si = end - begin
|
|
ctx_pad[bi, :si, ...] = ctx[begin:end, ...]
|
|
|
|
print(torch.allclose(Snew.float(), Sref.float(), atol=1e-4))
|
|
print(torch.allclose(ctx_ref.float(), ctx_pad.float(), atol=1e-3))
|
|
|
|
print(
|
|
'[FWD s={:d}, b={:d}] Fused {:.3f}ms Apex {:.3f}ms Diff {:.3f}ms Speedup {:.2f}x'
|
|
.format(s, b, ms_fused, ms_apex, ms_apex - ms_fused, ms_apex / ms_fused))
|
|
print(
|
|
'[BWD s={:d}, b={:d}] Fused {:.3f}ms Apex {:.3f}ms Diff {:.3f}ms Speedup {:.2f}x'
|
|
.format(s, b, ms_fused_bwd, ms_apex_bwd, ms_apex_bwd - ms_fused_bwd,
|
|
ms_apex_bwd / ms_fused_bwd))
|