osmarks/nanogpt-experiments
1
0
Fork 0
mirror of https://github.com/osmarks/nanogpt-experiments.git synced 2024-11-10 20:09:58 +00:00
Code Issues Releases Wiki Activity

add flash attention support, resolving last few issues but for now seems to work ok

Browse Source
This commit is contained in:
Andrej Karpathy 2023-01-30 23:18:26 +00:00
parent 0e90ee9d48
commit ae06d0b15a
1 changed files with 19 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
model.py
View File

@ -45,11 +45,16 @@ class CausalSelfAttention(nn.Module):
# regularization
self.attn_dropout = nn.Dropout(config.dropout)
self.resid_dropout = nn.Dropout(config.dropout)
# causal mask to ensure that attention is only applied to the left in the input sequence
self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
.view(1, 1, config.block_size, config.block_size))
self.n_head = config.n_head
self.n_embd = config.n_embd
self.dropout = config.dropout
# flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
if not self.flash:
print("WARNING: using slow attention, install PyTorch nightly for fast Flash Attention")
# causal mask to ensure that attention is only applied to the left in the input sequence
self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
.view(1, 1, config.block_size, config.block_size))
def forward(self, x):
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
@ -61,11 +66,17 @@ class CausalSelfAttention(nn.Module):
v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
# causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
att = att.masked_fill(self.bias[:,:,:T,:T] == 0, float('-inf'))
att = F.softmax(att, dim=-1)
att = self.attn_dropout(att)
y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
if self.flash:
# efficient attention using Flash Attention CUDA kernels
assert self.dropout == 0.0, "need dropout=0.0 for now, PyTorch team is working on fix in #92917"
y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=self.dropout, is_causal=True)
else:
# manual implementation of attention
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
att = att.masked_fill(self.bias[:,:,:T,:T] == 0, float('-inf'))
att = F.softmax(att, dim=-1)
att = self.attn_dropout(att)
y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
# output projection