HGB
commited on
Commit
•
9c8bb9e
1
Parent(s):
e9fd7b3
add bert padding + modify internVit classes
Browse files- modeling_intern_vit.py +83 -51
- triton.py → triton-test.py +0 -0
- triton_bert_pading.py +224 -0
- triton_flash_atn.py +363 -236
modeling_intern_vit.py
CHANGED
@@ -20,14 +20,9 @@ from transformers.utils import logging
|
|
20 |
from .configuration_intern_vit import InternVisionConfig
|
21 |
|
22 |
try:
|
23 |
-
|
24 |
-
from flash_attn.flash_attn_interface import \
|
25 |
-
flash_attn_unpadded_qkvpacked_func
|
26 |
-
except: # v2
|
27 |
-
from flash_attn.flash_attn_interface import \
|
28 |
-
flash_attn_varlen_qkvpacked_func as flash_attn_unpadded_qkvpacked_func
|
29 |
|
30 |
-
from
|
31 |
|
32 |
has_flash_attn = True
|
33 |
except:
|
@@ -74,28 +69,31 @@ class FlashAttention(nn.Module):
|
|
74 |
max_s = seqlen
|
75 |
cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32,
|
76 |
device=qkv.device)
|
77 |
-
output =
|
78 |
qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
79 |
-
|
80 |
)
|
81 |
-
output = rearrange(
|
|
|
82 |
else:
|
83 |
nheads = qkv.shape[-2]
|
84 |
x = rearrange(qkv, 'b s three h d -> b s (three h d)')
|
85 |
-
x_unpad, indices, cu_seqlens, max_s = unpad_input(
|
86 |
-
|
87 |
-
|
|
|
|
|
88 |
x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
89 |
-
|
90 |
)
|
91 |
output = rearrange(pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'),
|
92 |
indices, batch_size, seqlen),
|
93 |
'b s (h d) -> b s h d', h=nheads)
|
94 |
else:
|
95 |
assert max_s is not None
|
96 |
-
output =
|
97 |
qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
98 |
-
|
99 |
)
|
100 |
|
101 |
return output, None
|
@@ -111,7 +109,8 @@ class InternRMSNorm(nn.Module):
|
|
111 |
input_dtype = hidden_states.dtype
|
112 |
hidden_states = hidden_states.to(torch.float32)
|
113 |
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
114 |
-
hidden_states = hidden_states *
|
|
|
115 |
return self.weight * hidden_states.to(input_dtype)
|
116 |
|
117 |
|
@@ -120,12 +119,14 @@ try:
|
|
120 |
|
121 |
InternRMSNorm = FusedRMSNorm # noqa
|
122 |
|
123 |
-
logger.info(
|
|
|
124 |
except ImportError:
|
125 |
# using the normal InternRMSNorm
|
126 |
pass
|
127 |
except Exception:
|
128 |
-
logger.warning(
|
|
|
129 |
pass
|
130 |
|
131 |
|
@@ -154,7 +155,8 @@ class InternVisionEmbeddings(nn.Module):
|
|
154 |
self.num_patches = (self.image_size // self.patch_size) ** 2
|
155 |
self.num_positions = self.num_patches + 1
|
156 |
|
157 |
-
self.position_embedding = nn.Parameter(
|
|
|
158 |
|
159 |
def _get_pos_embed(self, pos_embed, H, W):
|
160 |
target_dtype = pos_embed.dtype
|
@@ -166,14 +168,17 @@ class InternVisionEmbeddings(nn.Module):
|
|
166 |
|
167 |
def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
|
168 |
target_dtype = self.patch_embedding.weight.dtype
|
169 |
-
|
|
|
170 |
batch_size, _, height, width = patch_embeds.shape
|
171 |
patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
|
172 |
-
class_embeds = self.class_embedding.expand(
|
|
|
173 |
embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
|
174 |
position_embedding = torch.cat([
|
175 |
self.position_embedding[:, :1, :],
|
176 |
-
self._get_pos_embed(
|
|
|
177 |
], dim=1)
|
178 |
embeddings = embeddings + position_embedding.to(target_dtype)
|
179 |
return embeddings
|
@@ -189,38 +194,48 @@ class InternAttention(nn.Module):
|
|
189 |
self.num_heads = config.num_attention_heads
|
190 |
self.use_flash_attn = config.use_flash_attn and has_flash_attn
|
191 |
if config.use_flash_attn and not has_flash_attn:
|
192 |
-
print(
|
|
|
193 |
self.head_dim = self.embed_dim // self.num_heads
|
194 |
if self.head_dim * self.num_heads != self.embed_dim:
|
195 |
raise ValueError(
|
196 |
-
f'embed_dim must be divisible by num_heads (got `embed_dim`: {
|
|
|
197 |
f' {self.num_heads}).'
|
198 |
)
|
199 |
|
200 |
self.scale = self.head_dim ** -0.5
|
201 |
-
self.qkv = nn.Linear(self.embed_dim, 3 *
|
|
|
202 |
self.attn_drop = nn.Dropout(config.attention_dropout)
|
203 |
self.proj_drop = nn.Dropout(config.dropout)
|
204 |
|
205 |
self.qk_normalization = config.qk_normalization
|
206 |
|
207 |
if self.qk_normalization:
|
208 |
-
self.q_norm = InternRMSNorm(
|
209 |
-
|
|
|
|
|
210 |
|
211 |
if self.use_flash_attn:
|
212 |
-
self.inner_attn = FlashAttention(
|
|
|
213 |
self.proj = nn.Linear(self.embed_dim, self.embed_dim)
|
214 |
|
215 |
def _naive_attn(self, x):
|
216 |
B, N, C = x.shape
|
217 |
-
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C //
|
218 |
-
|
|
|
|
|
219 |
|
220 |
if self.qk_normalization:
|
221 |
B_, H_, N_, D_ = q.shape
|
222 |
-
q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)
|
223 |
-
|
|
|
|
|
224 |
|
225 |
attn = ((q * self.scale) @ k.transpose(-2, -1))
|
226 |
attn = attn.softmax(dim=-1)
|
@@ -233,7 +248,8 @@ class InternAttention(nn.Module):
|
|
233 |
|
234 |
def _flash_attn(self, x, key_padding_mask=None, need_weights=False):
|
235 |
qkv = self.qkv(x)
|
236 |
-
qkv = rearrange(qkv, 'b s (three h d) -> b s three h d',
|
|
|
237 |
|
238 |
if self.qk_normalization:
|
239 |
q, k, v = qkv.unbind(2)
|
@@ -249,7 +265,8 @@ class InternAttention(nn.Module):
|
|
249 |
return outs
|
250 |
|
251 |
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
252 |
-
x = self._naive_attn(
|
|
|
253 |
return x
|
254 |
|
255 |
|
@@ -277,13 +294,19 @@ class InternVisionEncoderLayer(nn.Module):
|
|
277 |
|
278 |
self.attn = InternAttention(config)
|
279 |
self.mlp = InternMLP(config)
|
280 |
-
self.norm1 = NORM2FN[self.norm_type](
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
self.
|
286 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
287 |
|
288 |
def forward(
|
289 |
self,
|
@@ -293,9 +316,11 @@ class InternVisionEncoderLayer(nn.Module):
|
|
293 |
Args:
|
294 |
hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
295 |
"""
|
296 |
-
hidden_states = hidden_states +
|
|
|
297 |
|
298 |
-
hidden_states = hidden_states +
|
|
|
299 |
|
300 |
return hidden_states
|
301 |
|
@@ -314,7 +339,8 @@ class InternVisionEncoder(nn.Module):
|
|
314 |
super().__init__()
|
315 |
self.config = config
|
316 |
# stochastic depth decay rule
|
317 |
-
dpr = [x.item() for x in torch.linspace(
|
|
|
318 |
self.layers = nn.ModuleList([
|
319 |
InternVisionEncoderLayer(config, dpr[idx]) for idx in range(config.num_hidden_layers)])
|
320 |
self.gradient_checkpointing = True
|
@@ -382,13 +408,17 @@ class InternVisionModel(PreTrainedModel):
|
|
382 |
pos_emb = self.embeddings.position_embedding
|
383 |
_, num_positions, embed_dim = pos_emb.shape
|
384 |
cls_emb = pos_emb[:, :1, :]
|
385 |
-
pos_emb = pos_emb[:, 1:, :].reshape(
|
386 |
-
|
387 |
-
pos_emb =
|
|
|
|
|
|
|
388 |
pos_emb = torch.cat([cls_emb, pos_emb], dim=1)
|
389 |
self.embeddings.position_embedding = nn.Parameter(pos_emb)
|
390 |
self.embeddings.image_size = new_size
|
391 |
-
logger.info('Resized position embeddings from {} to {}'.format(
|
|
|
392 |
|
393 |
def get_input_embeddings(self):
|
394 |
return self.embeddings
|
@@ -406,7 +436,8 @@ class InternVisionModel(PreTrainedModel):
|
|
406 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
407 |
|
408 |
if pixel_values is None and pixel_embeds is None:
|
409 |
-
raise ValueError(
|
|
|
410 |
|
411 |
if pixel_embeds is not None:
|
412 |
hidden_states = pixel_embeds
|
@@ -414,7 +445,8 @@ class InternVisionModel(PreTrainedModel):
|
|
414 |
if len(pixel_values.shape) == 4:
|
415 |
hidden_states = self.embeddings(pixel_values)
|
416 |
else:
|
417 |
-
raise ValueError(f'wrong pixel_values size: {
|
|
|
418 |
encoder_outputs = self.encoder(
|
419 |
inputs_embeds=hidden_states,
|
420 |
output_hidden_states=output_hidden_states,
|
|
|
20 |
from .configuration_intern_vit import InternVisionConfig
|
21 |
|
22 |
try:
|
23 |
+
from triton_flash_atn import _attention
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
+
from triton_bert_pading import pad_input, unpad_input
|
26 |
|
27 |
has_flash_attn = True
|
28 |
except:
|
|
|
69 |
max_s = seqlen
|
70 |
cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32,
|
71 |
device=qkv.device)
|
72 |
+
output = _attention.apply(
|
73 |
qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
74 |
+
sm_scale=self.softmax_scale, causal=causal
|
75 |
)
|
76 |
+
output = rearrange(
|
77 |
+
output, '(b s) ... -> b s ...', b=batch_size)
|
78 |
else:
|
79 |
nheads = qkv.shape[-2]
|
80 |
x = rearrange(qkv, 'b s three h d -> b s (three h d)')
|
81 |
+
x_unpad, indices, cu_seqlens, max_s = unpad_input(
|
82 |
+
x, key_padding_mask)
|
83 |
+
x_unpad = rearrange(
|
84 |
+
x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads)
|
85 |
+
output_unpad = _attention.apply(
|
86 |
x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
87 |
+
sm_scale=self.softmax_scale, causal=causal
|
88 |
)
|
89 |
output = rearrange(pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'),
|
90 |
indices, batch_size, seqlen),
|
91 |
'b s (h d) -> b s h d', h=nheads)
|
92 |
else:
|
93 |
assert max_s is not None
|
94 |
+
output = _attention.apply(
|
95 |
qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
|
96 |
+
sm_scale=self.softmax_scale, causal=causal
|
97 |
)
|
98 |
|
99 |
return output, None
|
|
|
109 |
input_dtype = hidden_states.dtype
|
110 |
hidden_states = hidden_states.to(torch.float32)
|
111 |
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
112 |
+
hidden_states = hidden_states * \
|
113 |
+
torch.rsqrt(variance + self.variance_epsilon)
|
114 |
return self.weight * hidden_states.to(input_dtype)
|
115 |
|
116 |
|
|
|
119 |
|
120 |
InternRMSNorm = FusedRMSNorm # noqa
|
121 |
|
122 |
+
logger.info(
|
123 |
+
'Discovered apex.normalization.FusedRMSNorm - will use it instead of InternRMSNorm')
|
124 |
except ImportError:
|
125 |
# using the normal InternRMSNorm
|
126 |
pass
|
127 |
except Exception:
|
128 |
+
logger.warning(
|
129 |
+
'discovered apex but it failed to load, falling back to InternRMSNorm')
|
130 |
pass
|
131 |
|
132 |
|
|
|
155 |
self.num_patches = (self.image_size // self.patch_size) ** 2
|
156 |
self.num_positions = self.num_patches + 1
|
157 |
|
158 |
+
self.position_embedding = nn.Parameter(
|
159 |
+
torch.randn(1, self.num_positions, self.embed_dim))
|
160 |
|
161 |
def _get_pos_embed(self, pos_embed, H, W):
|
162 |
target_dtype = pos_embed.dtype
|
|
|
168 |
|
169 |
def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
|
170 |
target_dtype = self.patch_embedding.weight.dtype
|
171 |
+
# shape = [*, channel, width, height]
|
172 |
+
patch_embeds = self.patch_embedding(pixel_values)
|
173 |
batch_size, _, height, width = patch_embeds.shape
|
174 |
patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
|
175 |
+
class_embeds = self.class_embedding.expand(
|
176 |
+
batch_size, 1, -1).to(target_dtype)
|
177 |
embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
|
178 |
position_embedding = torch.cat([
|
179 |
self.position_embedding[:, :1, :],
|
180 |
+
self._get_pos_embed(
|
181 |
+
self.position_embedding[:, 1:, :], height, width)
|
182 |
], dim=1)
|
183 |
embeddings = embeddings + position_embedding.to(target_dtype)
|
184 |
return embeddings
|
|
|
194 |
self.num_heads = config.num_attention_heads
|
195 |
self.use_flash_attn = config.use_flash_attn and has_flash_attn
|
196 |
if config.use_flash_attn and not has_flash_attn:
|
197 |
+
print(
|
198 |
+
'Warning: Flash Attention is not available, use_flash_attn is set to False.')
|
199 |
self.head_dim = self.embed_dim // self.num_heads
|
200 |
if self.head_dim * self.num_heads != self.embed_dim:
|
201 |
raise ValueError(
|
202 |
+
f'embed_dim must be divisible by num_heads (got `embed_dim`: {
|
203 |
+
self.embed_dim} and `num_heads`:'
|
204 |
f' {self.num_heads}).'
|
205 |
)
|
206 |
|
207 |
self.scale = self.head_dim ** -0.5
|
208 |
+
self.qkv = nn.Linear(self.embed_dim, 3 *
|
209 |
+
self.embed_dim, bias=config.qkv_bias)
|
210 |
self.attn_drop = nn.Dropout(config.attention_dropout)
|
211 |
self.proj_drop = nn.Dropout(config.dropout)
|
212 |
|
213 |
self.qk_normalization = config.qk_normalization
|
214 |
|
215 |
if self.qk_normalization:
|
216 |
+
self.q_norm = InternRMSNorm(
|
217 |
+
self.embed_dim, eps=config.layer_norm_eps)
|
218 |
+
self.k_norm = InternRMSNorm(
|
219 |
+
self.embed_dim, eps=config.layer_norm_eps)
|
220 |
|
221 |
if self.use_flash_attn:
|
222 |
+
self.inner_attn = FlashAttention(
|
223 |
+
attention_dropout=config.attention_dropout)
|
224 |
self.proj = nn.Linear(self.embed_dim, self.embed_dim)
|
225 |
|
226 |
def _naive_attn(self, x):
|
227 |
B, N, C = x.shape
|
228 |
+
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C //
|
229 |
+
self.num_heads).permute(2, 0, 3, 1, 4)
|
230 |
+
# make torchscript happy (cannot use tensor as tuple)
|
231 |
+
q, k, v = qkv.unbind(0)
|
232 |
|
233 |
if self.qk_normalization:
|
234 |
B_, H_, N_, D_ = q.shape
|
235 |
+
q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)
|
236 |
+
).view(B_, N_, H_, D_).transpose(1, 2)
|
237 |
+
k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)
|
238 |
+
).view(B_, N_, H_, D_).transpose(1, 2)
|
239 |
|
240 |
attn = ((q * self.scale) @ k.transpose(-2, -1))
|
241 |
attn = attn.softmax(dim=-1)
|
|
|
248 |
|
249 |
def _flash_attn(self, x, key_padding_mask=None, need_weights=False):
|
250 |
qkv = self.qkv(x)
|
251 |
+
qkv = rearrange(qkv, 'b s (three h d) -> b s three h d',
|
252 |
+
three=3, h=self.num_heads)
|
253 |
|
254 |
if self.qk_normalization:
|
255 |
q, k, v = qkv.unbind(2)
|
|
|
265 |
return outs
|
266 |
|
267 |
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
268 |
+
x = self._naive_attn(
|
269 |
+
hidden_states) if not self.use_flash_attn else self._flash_attn(hidden_states)
|
270 |
return x
|
271 |
|
272 |
|
|
|
294 |
|
295 |
self.attn = InternAttention(config)
|
296 |
self.mlp = InternMLP(config)
|
297 |
+
self.norm1 = NORM2FN[self.norm_type](
|
298 |
+
self.embed_dim, eps=config.layer_norm_eps)
|
299 |
+
self.norm2 = NORM2FN[self.norm_type](
|
300 |
+
self.embed_dim, eps=config.layer_norm_eps)
|
301 |
+
|
302 |
+
self.ls1 = nn.Parameter(
|
303 |
+
config.initializer_factor * torch.ones(self.embed_dim))
|
304 |
+
self.ls2 = nn.Parameter(
|
305 |
+
config.initializer_factor * torch.ones(self.embed_dim))
|
306 |
+
self.drop_path1 = DropPath(
|
307 |
+
drop_path_rate) if drop_path_rate > 0. else nn.Identity()
|
308 |
+
self.drop_path2 = DropPath(
|
309 |
+
drop_path_rate) if drop_path_rate > 0. else nn.Identity()
|
310 |
|
311 |
def forward(
|
312 |
self,
|
|
|
316 |
Args:
|
317 |
hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
318 |
"""
|
319 |
+
hidden_states = hidden_states + \
|
320 |
+
self.drop_path1(self.attn(self.norm1(hidden_states)) * self.ls1)
|
321 |
|
322 |
+
hidden_states = hidden_states + \
|
323 |
+
self.drop_path2(self.mlp(self.norm2(hidden_states)) * self.ls2)
|
324 |
|
325 |
return hidden_states
|
326 |
|
|
|
339 |
super().__init__()
|
340 |
self.config = config
|
341 |
# stochastic depth decay rule
|
342 |
+
dpr = [x.item() for x in torch.linspace(
|
343 |
+
0, config.drop_path_rate, config.num_hidden_layers)]
|
344 |
self.layers = nn.ModuleList([
|
345 |
InternVisionEncoderLayer(config, dpr[idx]) for idx in range(config.num_hidden_layers)])
|
346 |
self.gradient_checkpointing = True
|
|
|
408 |
pos_emb = self.embeddings.position_embedding
|
409 |
_, num_positions, embed_dim = pos_emb.shape
|
410 |
cls_emb = pos_emb[:, :1, :]
|
411 |
+
pos_emb = pos_emb[:, 1:, :].reshape(
|
412 |
+
1, old_size // patch_size, old_size // patch_size, -1).permute(0, 3, 1, 2)
|
413 |
+
pos_emb = F.interpolate(pos_emb.float(
|
414 |
+
), size=new_size // patch_size, mode='bicubic', align_corners=False)
|
415 |
+
pos_emb = pos_emb.to(cls_emb.dtype).reshape(
|
416 |
+
1, embed_dim, -1).permute(0, 2, 1)
|
417 |
pos_emb = torch.cat([cls_emb, pos_emb], dim=1)
|
418 |
self.embeddings.position_embedding = nn.Parameter(pos_emb)
|
419 |
self.embeddings.image_size = new_size
|
420 |
+
logger.info('Resized position embeddings from {} to {}'.format(
|
421 |
+
old_size, new_size))
|
422 |
|
423 |
def get_input_embeddings(self):
|
424 |
return self.embeddings
|
|
|
436 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
437 |
|
438 |
if pixel_values is None and pixel_embeds is None:
|
439 |
+
raise ValueError(
|
440 |
+
'You have to specify pixel_values or pixel_embeds')
|
441 |
|
442 |
if pixel_embeds is not None:
|
443 |
hidden_states = pixel_embeds
|
|
|
445 |
if len(pixel_values.shape) == 4:
|
446 |
hidden_states = self.embeddings(pixel_values)
|
447 |
else:
|
448 |
+
raise ValueError(f'wrong pixel_values size: {
|
449 |
+
pixel_values.shape}')
|
450 |
encoder_outputs = self.encoder(
|
451 |
inputs_embeds=hidden_states,
|
452 |
output_hidden_states=output_hidden_states,
|
triton.py → triton-test.py
RENAMED
File without changes
|
triton_bert_pading.py
ADDED
@@ -0,0 +1,224 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Adapted from https://github.com/mlcommons/training_results_v1.1/blob/main/NVIDIA/benchmarks/bert/implementations/pytorch/padding.py
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from einops import rearrange, repeat
|
6 |
+
|
7 |
+
|
8 |
+
class IndexFirstAxis(torch.autograd.Function):
|
9 |
+
@staticmethod
|
10 |
+
def forward(ctx, input, indices):
|
11 |
+
ctx.save_for_backward(indices)
|
12 |
+
assert input.ndim >= 2
|
13 |
+
ctx.first_axis_dim, other_shape = input.shape[0], input.shape[1:]
|
14 |
+
second_dim = other_shape.numel()
|
15 |
+
# TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
|
16 |
+
# return input[indices]
|
17 |
+
return torch.gather(
|
18 |
+
rearrange(input, "b ... -> b (...)"), 0, repeat(indices,
|
19 |
+
"z -> z d", d=second_dim)
|
20 |
+
).reshape(-1, *other_shape)
|
21 |
+
|
22 |
+
@staticmethod
|
23 |
+
def backward(ctx, grad_output):
|
24 |
+
(indices,) = ctx.saved_tensors
|
25 |
+
assert grad_output.ndim >= 2
|
26 |
+
other_shape = grad_output.shape[1:]
|
27 |
+
grad_output = rearrange(grad_output, "b ... -> b (...)")
|
28 |
+
grad_input = torch.zeros(
|
29 |
+
[ctx.first_axis_dim, grad_output.shape[1]],
|
30 |
+
device=grad_output.device,
|
31 |
+
dtype=grad_output.dtype,
|
32 |
+
)
|
33 |
+
# TD [2022-03-04] For some reason torch.scatter is a bit faster than indexing.
|
34 |
+
# grad_input[indices] = grad_output
|
35 |
+
grad_input.scatter_(0, repeat(indices, "z -> z d",
|
36 |
+
d=grad_output.shape[1]), grad_output)
|
37 |
+
return grad_input.reshape(ctx.first_axis_dim, *other_shape), None
|
38 |
+
|
39 |
+
|
40 |
+
index_first_axis = IndexFirstAxis.apply
|
41 |
+
|
42 |
+
|
43 |
+
class IndexPutFirstAxis(torch.autograd.Function):
|
44 |
+
@staticmethod
|
45 |
+
def forward(ctx, values, indices, first_axis_dim):
|
46 |
+
ctx.save_for_backward(indices)
|
47 |
+
assert indices.ndim == 1
|
48 |
+
assert values.ndim >= 2
|
49 |
+
output = torch.zeros(
|
50 |
+
first_axis_dim, *
|
51 |
+
values.shape[1:], device=values.device, dtype=values.dtype
|
52 |
+
)
|
53 |
+
# TD [2022-03-04] For some reason torch.scatter is a bit faster than indexing.
|
54 |
+
output[indices] = values
|
55 |
+
# output.scatter_(0, repeat(indices, 'z -> z d', d=values.shape[1]), values)
|
56 |
+
return output
|
57 |
+
|
58 |
+
@staticmethod
|
59 |
+
def backward(ctx, grad_output):
|
60 |
+
(indices,) = ctx.saved_tensors
|
61 |
+
# TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
|
62 |
+
grad_values = grad_output[indices]
|
63 |
+
# grad_values = torch.gather(grad_output, 0, repeat(indices, 'z -> z d', d=grad_output.shape[1]))
|
64 |
+
return grad_values, None, None
|
65 |
+
|
66 |
+
|
67 |
+
index_put_first_axis = IndexPutFirstAxis.apply
|
68 |
+
|
69 |
+
|
70 |
+
class IndexFirstAxisResidual(torch.autograd.Function):
|
71 |
+
@staticmethod
|
72 |
+
def forward(ctx, input, indices):
|
73 |
+
ctx.save_for_backward(indices)
|
74 |
+
assert input.ndim >= 2
|
75 |
+
ctx.first_axis_dim, other_shape = input.shape[0], input.shape[1:]
|
76 |
+
second_dim = other_shape.numel()
|
77 |
+
# TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
|
78 |
+
output = input[indices]
|
79 |
+
# We don't want to reshape input (b ... -> b (...)) since it could change the channel_last
|
80 |
+
# memory format to channel_first. In other words, input might not be contiguous.
|
81 |
+
# If we don't detach, Pytorch complains about output being a view and is being modified inplace
|
82 |
+
return output, input.detach()
|
83 |
+
|
84 |
+
@staticmethod
|
85 |
+
def backward(ctx, grad_output, grad_residual):
|
86 |
+
(indices,) = ctx.saved_tensors
|
87 |
+
assert grad_output.ndim >= 2
|
88 |
+
other_shape = grad_output.shape[1:]
|
89 |
+
assert grad_residual.shape[1:] == other_shape
|
90 |
+
grad_input = grad_residual
|
91 |
+
# grad_input[indices] += grad_output
|
92 |
+
indices = indices.reshape(
|
93 |
+
indices.shape[0], *((1,) * (grad_output.ndim - 1)))
|
94 |
+
indices = indices.expand_as(grad_output)
|
95 |
+
grad_input.scatter_add_(0, indices, grad_output)
|
96 |
+
return grad_input.reshape(ctx.first_axis_dim, *other_shape), None
|
97 |
+
|
98 |
+
|
99 |
+
index_first_axis_residual = IndexFirstAxisResidual.apply
|
100 |
+
|
101 |
+
|
102 |
+
def unpad_input(hidden_states, attention_mask):
|
103 |
+
"""
|
104 |
+
Arguments:
|
105 |
+
hidden_states: (batch, seqlen, ...)
|
106 |
+
attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
|
107 |
+
Return:
|
108 |
+
hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
|
109 |
+
indices: (total_nnz), the indices of non-masked tokens from the flattened input sequence.
|
110 |
+
cu_seqlens: (batch + 1), the cumulative sequence lengths, used to index into hidden_states.
|
111 |
+
max_seqlen_in_batch: int
|
112 |
+
"""
|
113 |
+
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
|
114 |
+
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
|
115 |
+
max_seqlen_in_batch = seqlens_in_batch.max().item()
|
116 |
+
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0,
|
117 |
+
dtype=torch.torch.int32), (1, 0))
|
118 |
+
# TD [2022-03-04] We don't want to index with a bool mask, because Pytorch will expand the
|
119 |
+
# bool mask, then call nonzero to get the indices, then index with those. The indices is @dim
|
120 |
+
# times larger than it needs to be, wasting memory. It's faster and more memory-efficient to
|
121 |
+
# index with integer indices. Moreover, torch's index is a bit slower than it needs to be,
|
122 |
+
# so we write custom forward and backward to make it a bit faster.
|
123 |
+
return (
|
124 |
+
index_first_axis(
|
125 |
+
rearrange(hidden_states, "b s ... -> (b s) ..."), indices),
|
126 |
+
indices,
|
127 |
+
cu_seqlens,
|
128 |
+
max_seqlen_in_batch,
|
129 |
+
)
|
130 |
+
|
131 |
+
|
132 |
+
def unpad_input_for_concatenated_sequences(hidden_states, attention_mask_in_length):
|
133 |
+
"""
|
134 |
+
Supports concatenating short samples in one sequence. The attention_mask_in_length is utilized to mask other short samples. It helps efficient training of variant lengths-based samples (e.g., the supervised fine-tuning task in large language model).
|
135 |
+
The motivation for this function is explained [here](https://github.com/Dao-AILab/flash-attention/issues/432#issuecomment-1668822286).
|
136 |
+
|
137 |
+
For example, if batch = 3 and seqlen = 6, the attention_mask_in_length is:
|
138 |
+
```
|
139 |
+
[
|
140 |
+
[2, 3, 0, 0, 0, 0],
|
141 |
+
[3, 2, 0, 0, 0, 0],
|
142 |
+
[6, 0, 0, 0, 0, 0]
|
143 |
+
]
|
144 |
+
```
|
145 |
+
, which refers to the 3D-attention mask:
|
146 |
+
```
|
147 |
+
[
|
148 |
+
[
|
149 |
+
[1, 0, 0, 0, 0, 0],
|
150 |
+
[1, 1, 0, 0, 0, 0],
|
151 |
+
[0, 0, 1, 0, 0, 0],
|
152 |
+
[0, 0, 1, 1, 0, 0],
|
153 |
+
[0, 0, 1, 1, 1, 0],
|
154 |
+
[0, 0, 0, 0, 0, 1]
|
155 |
+
],
|
156 |
+
[
|
157 |
+
[1, 0, 0, 0, 0, 0],
|
158 |
+
[1, 1, 0, 0, 0, 0],
|
159 |
+
[1, 1, 1, 0, 0, 0],
|
160 |
+
[0, 0, 0, 1, 0, 0],
|
161 |
+
[0, 0, 0, 1, 1, 0],
|
162 |
+
[0, 0, 0, 0, 0, 1]
|
163 |
+
],
|
164 |
+
[
|
165 |
+
[1, 0, 0, 0, 0, 0],
|
166 |
+
[1, 1, 0, 0, 0, 0],
|
167 |
+
[1, 1, 1, 0, 0, 0],
|
168 |
+
[1, 1, 1, 1, 0, 0],
|
169 |
+
[1, 1, 1, 1, 1, 0],
|
170 |
+
[1, 1, 1, 1, 1, 1]
|
171 |
+
]
|
172 |
+
]
|
173 |
+
```.
|
174 |
+
|
175 |
+
Arguments:
|
176 |
+
hidden_states: (batch, seqlen, ...)
|
177 |
+
attention_mask_in_length: (batch, seqlen), int, a nonzero number (e.g., 1, 2, 3, etc.) means length of concatenated sequence in b-th batch, and 0 means none.
|
178 |
+
Return:
|
179 |
+
hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
|
180 |
+
indices: (total_nnz), the indices of non-masked tokens from the flattened input sequence.
|
181 |
+
cu_seqlens: (batch + 1), the cumulative sequence lengths, used to index into hidden_states.
|
182 |
+
max_seqlen_in_batch: int
|
183 |
+
"""
|
184 |
+
length = attention_mask_in_length.sum(dim=-1)
|
185 |
+
seqlen = attention_mask_in_length.size(-1)
|
186 |
+
attention_mask_2d = torch.arange(seqlen, device=length.device, dtype=length.dtype).expand(
|
187 |
+
len(length), seqlen) < length.unsqueeze(1)
|
188 |
+
real_indices_idx = torch.nonzero(
|
189 |
+
attention_mask_in_length.flatten(), as_tuple=False).flatten()
|
190 |
+
seqlens_in_batch = attention_mask_in_length.flatten()[real_indices_idx]
|
191 |
+
indices = torch.nonzero(attention_mask_2d.flatten(),
|
192 |
+
as_tuple=False).flatten()
|
193 |
+
max_seqlen_in_batch = seqlens_in_batch.max().item()
|
194 |
+
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0,
|
195 |
+
dtype=torch.torch.int32), (1, 0))
|
196 |
+
# TD [2022-03-04] We don't want to index with a bool mask, because Pytorch will expand the
|
197 |
+
# bool mask, then call nonzero to get the indices, then index with those. The indices is @dim
|
198 |
+
# times larger than it needs to be, wasting memory. It's faster and more memory-efficient to
|
199 |
+
# index with integer indices. Moreover, torch's index is a bit slower than it needs to be,
|
200 |
+
# so we write custom forward and backward to make it a bit faster.
|
201 |
+
return (
|
202 |
+
index_first_axis(
|
203 |
+
rearrange(hidden_states, "b s ... -> (b s) ..."), indices),
|
204 |
+
indices,
|
205 |
+
cu_seqlens,
|
206 |
+
max_seqlen_in_batch,
|
207 |
+
)
|
208 |
+
|
209 |
+
|
210 |
+
def pad_input(hidden_states, indices, batch, seqlen):
|
211 |
+
"""
|
212 |
+
Arguments:
|
213 |
+
hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
|
214 |
+
indices: (total_nnz), the indices that represent the non-masked tokens of the original padded input sequence.
|
215 |
+
batch: int, batch size for the padded sequence.
|
216 |
+
seqlen: int, maximum sequence length for the padded sequence.
|
217 |
+
Return:
|
218 |
+
hidden_states: (batch, seqlen, ...)
|
219 |
+
"""
|
220 |
+
dim = hidden_states.shape[-1]
|
221 |
+
# output = torch.zeros((batch * seqlen), dim, device=hidden_states.device, dtype=hidden_states.dtype)
|
222 |
+
# output[indices] = hidden_states
|
223 |
+
output = index_put_first_axis(hidden_states, indices, batch * seqlen)
|
224 |
+
return rearrange(output, "(b s) ... -> b s ...", b=batch)
|
triton_flash_atn.py
CHANGED
@@ -11,62 +11,66 @@ Extra Credits:
|
|
11 |
|
12 |
"""
|
13 |
|
|
|
14 |
import torch
|
15 |
|
16 |
import triton
|
17 |
import triton.language as tl
|
18 |
|
|
|
19 |
|
20 |
-
|
21 |
-
|
|
|
|
|
|
|
|
|
|
|
22 |
|
23 |
|
24 |
@triton.jit
|
25 |
-
def _attn_fwd_inner(acc, l_i, m_i, q,
|
26 |
-
K_block_ptr, V_block_ptr,
|
27 |
-
start_m,
|
28 |
-
BLOCK_M: tl.constexpr,
|
29 |
-
STAGE: tl.constexpr, offs_m: tl.constexpr, offs_n: tl.constexpr,
|
30 |
-
N_CTX
|
|
|
31 |
# range of values handled by this stage
|
32 |
if STAGE == 1:
|
33 |
lo, hi = 0, start_m * BLOCK_M
|
34 |
elif STAGE == 2:
|
35 |
lo, hi = start_m * BLOCK_M, (start_m + 1) * BLOCK_M
|
36 |
lo = tl.multiple_of(lo, BLOCK_M)
|
|
|
|
|
37 |
# causal = False
|
38 |
else:
|
39 |
lo, hi = 0, N_CTX
|
40 |
-
K_block_ptr = tl.advance(K_block_ptr, (0, lo))
|
41 |
-
V_block_ptr = tl.advance(V_block_ptr, (lo, 0))
|
42 |
# loop over k, v and update accumulator
|
43 |
for start_n in range(lo, hi, BLOCK_N):
|
44 |
start_n = tl.multiple_of(start_n, BLOCK_N)
|
45 |
# -- compute qk ----
|
46 |
k = tl.load(K_block_ptr)
|
47 |
-
|
|
|
|
|
48 |
if STAGE == 2:
|
49 |
mask = offs_m[:, None] >= (start_n + offs_n[None, :])
|
50 |
-
qk =
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
m_ij = tl.maximum(m_i, tl.max(qk, 1) * qk_scale)
|
55 |
-
qk = qk * qk_scale - m_ij[:, None]
|
56 |
p = tl.math.exp2(qk)
|
57 |
-
l_ij = tl.sum(p, 1)
|
58 |
-
# -- update m_i and l_i
|
59 |
-
alpha = tl.math.exp2(m_i - m_ij)
|
60 |
-
l_i = l_i * alpha + l_ij
|
61 |
# -- update output accumulator --
|
|
|
62 |
acc = acc * alpha[:, None]
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
acc = tl.dot(p, v, acc)
|
70 |
# update m_i and l_i
|
71 |
m_i = m_ij
|
72 |
V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
|
@@ -74,80 +78,77 @@ def _attn_fwd_inner(acc, l_i, m_i, q, #
|
|
74 |
return acc, l_i, m_i
|
75 |
|
76 |
|
77 |
-
# We don't run auto-tuning
|
78 |
# the code below and commenting out the equivalent parameters is convenient for
|
79 |
# re-tuning.
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
|
|
98 |
@triton.jit
|
99 |
-
def _attn_fwd(Q, K, V, sm_scale, M, Out,
|
100 |
-
stride_qz, stride_qh, stride_qm, stride_qk,
|
101 |
-
stride_kz, stride_kh, stride_kn, stride_kk,
|
102 |
-
stride_vz, stride_vh, stride_vk, stride_vn,
|
103 |
-
stride_oz, stride_oh, stride_om, stride_on,
|
104 |
-
Z, H,
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
|
|
|
|
109 |
):
|
110 |
-
tl.static_assert(BLOCK_N <= HEAD_DIM)
|
111 |
start_m = tl.program_id(0)
|
112 |
off_hz = tl.program_id(1)
|
113 |
-
|
114 |
-
off_h = off_hz % H
|
115 |
-
qvk_offset = off_z.to(tl.int64) * stride_qz + \
|
116 |
-
off_h.to(tl.int64) * stride_qh
|
117 |
|
118 |
# block pointers
|
119 |
Q_block_ptr = tl.make_block_ptr(
|
120 |
base=Q + qvk_offset,
|
121 |
-
shape=(N_CTX,
|
122 |
strides=(stride_qm, stride_qk),
|
123 |
offsets=(start_m * BLOCK_M, 0),
|
124 |
-
block_shape=(BLOCK_M,
|
125 |
order=(1, 0),
|
126 |
)
|
127 |
-
v_order: tl.constexpr = (
|
128 |
-
0, 1) if V.dtype.element_ty == tl.float8e5 else (1, 0)
|
129 |
V_block_ptr = tl.make_block_ptr(
|
130 |
base=V + qvk_offset,
|
131 |
-
shape=(N_CTX,
|
132 |
strides=(stride_vk, stride_vn),
|
133 |
offsets=(0, 0),
|
134 |
-
block_shape=(BLOCK_N,
|
135 |
-
order=
|
136 |
)
|
137 |
K_block_ptr = tl.make_block_ptr(
|
138 |
base=K + qvk_offset,
|
139 |
-
shape=(
|
140 |
strides=(stride_kk, stride_kn),
|
141 |
offsets=(0, 0),
|
142 |
-
block_shape=(
|
143 |
order=(0, 1),
|
144 |
)
|
145 |
O_block_ptr = tl.make_block_ptr(
|
146 |
base=Out + qvk_offset,
|
147 |
-
shape=(N_CTX,
|
148 |
strides=(stride_om, stride_on),
|
149 |
offsets=(start_m * BLOCK_M, 0),
|
150 |
-
block_shape=(BLOCK_M,
|
151 |
order=(1, 0),
|
152 |
)
|
153 |
# initialize offsets
|
@@ -156,82 +157,99 @@ def _attn_fwd(Q, K, V, sm_scale, M, Out, #
|
|
156 |
# initialize pointer to m and l
|
157 |
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
|
158 |
l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0
|
159 |
-
acc = tl.zeros([BLOCK_M,
|
160 |
-
#
|
161 |
-
|
162 |
-
|
163 |
-
|
|
|
164 |
q = tl.load(Q_block_ptr)
|
|
|
165 |
# stage 1: off-band
|
166 |
# For causal = True, STAGE = 3 and _attn_fwd_inner gets 1 as its STAGE
|
167 |
# For causal = False, STAGE = 1, and _attn_fwd_inner gets 3 as its STAGE
|
168 |
if STAGE & 1:
|
169 |
-
acc, l_i, m_i = _attn_fwd_inner(acc, l_i, m_i, q, K_block_ptr, V_block_ptr,
|
170 |
-
start_m,
|
171 |
-
BLOCK_M,
|
172 |
-
4 - STAGE, offs_m, offs_n, N_CTX,
|
|
|
173 |
)
|
174 |
# stage 2: on-band
|
175 |
if STAGE & 2:
|
176 |
# barrier makes it easier for compielr to schedule the
|
177 |
# two loops independently
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
|
|
|
|
182 |
)
|
183 |
# epilogue
|
184 |
-
|
185 |
acc = acc / l_i[:, None]
|
186 |
m_ptrs = M + off_hz * N_CTX + offs_m
|
187 |
-
tl.store(m_ptrs, m_i)
|
188 |
tl.store(O_block_ptr, acc.to(Out.type.element_ty))
|
189 |
|
190 |
|
191 |
@triton.jit
|
192 |
-
def _attn_bwd_preprocess(O, DO,
|
193 |
-
Delta,
|
194 |
-
Z, H, N_CTX,
|
195 |
-
BLOCK_M: tl.constexpr,
|
196 |
):
|
197 |
off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
|
198 |
off_hz = tl.program_id(1)
|
199 |
-
off_n = tl.arange(0,
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
off_m[:, None] * HEAD_DIM + off_n[None, :]).to(tl.float32)
|
205 |
delta = tl.sum(o * do, axis=1)
|
206 |
-
# write-back
|
207 |
tl.store(Delta + off_hz * N_CTX + off_m, delta)
|
208 |
|
209 |
|
210 |
# The main inner-loop logic for computing dK and dV.
|
211 |
@triton.jit
|
212 |
-
def _attn_bwd_dkdv(dk, dv,
|
213 |
-
Q, k, v, sm_scale,
|
214 |
-
DO,
|
215 |
-
M, D,
|
216 |
# shared by Q/K/V/DO.
|
217 |
-
stride_tok, stride_d,
|
218 |
-
H, N_CTX, BLOCK_M1: tl.constexpr,
|
219 |
-
BLOCK_N1: tl.constexpr,
|
220 |
-
|
221 |
# Filled in by the wrapper.
|
222 |
-
start_n, start_m, num_steps,
|
223 |
MASK: tl.constexpr):
|
224 |
offs_m = start_m + tl.arange(0, BLOCK_M1)
|
225 |
offs_n = start_n + tl.arange(0, BLOCK_N1)
|
226 |
-
offs_k = tl.arange(0,
|
227 |
-
|
228 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
229 |
# BLOCK_N1 must be a multiple of BLOCK_M1, otherwise the code wouldn't work.
|
230 |
tl.static_assert(BLOCK_N1 % BLOCK_M1 == 0)
|
231 |
curr_m = start_m
|
232 |
step_m = BLOCK_M1
|
233 |
for blk_idx in range(num_steps):
|
234 |
-
qT = tl.load(
|
235 |
# Load m before computing qk to reduce pipeline stall.
|
236 |
offs_m = curr_m + tl.arange(0, BLOCK_M1)
|
237 |
m = tl.load(M + offs_m)
|
@@ -241,7 +259,7 @@ def _attn_bwd_dkdv(dk, dv, #
|
|
241 |
if MASK:
|
242 |
mask = (offs_m[None, :] >= offs_n[:, None])
|
243 |
pT = tl.where(mask, pT, 0.0)
|
244 |
-
do = tl.load(
|
245 |
# Compute dV.
|
246 |
ppT = pT
|
247 |
ppT = ppT.to(tl.float16)
|
@@ -249,35 +267,49 @@ def _attn_bwd_dkdv(dk, dv, #
|
|
249 |
# D (= delta) is pre-divided by ds_scale.
|
250 |
Di = tl.load(D + offs_m)
|
251 |
# Compute dP and dS.
|
252 |
-
dpT = tl.dot(v, tl.trans(do))
|
253 |
dsT = pT * (dpT - Di[None, :])
|
254 |
dsT = dsT.to(tl.float16)
|
255 |
dk += tl.dot(dsT, tl.trans(qT))
|
256 |
# Increment pointers.
|
257 |
curr_m += step_m
|
258 |
-
|
259 |
-
|
260 |
return dk, dv
|
261 |
|
262 |
|
263 |
# the main inner-loop logic for computing dQ
|
264 |
@triton.jit
|
265 |
-
def _attn_bwd_dq(dq, q, K, V,
|
266 |
do, m, D,
|
267 |
# shared by Q/K/V/DO.
|
268 |
-
stride_tok, stride_d,
|
269 |
-
H, N_CTX,
|
270 |
-
BLOCK_M2: tl.constexpr,
|
271 |
-
BLOCK_N2: tl.constexpr,
|
272 |
-
|
273 |
# Filled in by the wrapper.
|
274 |
-
start_m, start_n, num_steps,
|
275 |
MASK: tl.constexpr):
|
276 |
offs_m = start_m + tl.arange(0, BLOCK_M2)
|
277 |
offs_n = start_n + tl.arange(0, BLOCK_N2)
|
278 |
-
offs_k = tl.arange(0,
|
279 |
-
|
280 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
281 |
# D (= delta) is pre-divided by ds_scale.
|
282 |
Di = tl.load(D + offs_m)
|
283 |
# BLOCK_M2 must be a multiple of BLOCK_N2, otherwise the code wouldn't work.
|
@@ -285,8 +317,7 @@ def _attn_bwd_dq(dq, q, K, V, #
|
|
285 |
curr_n = start_n
|
286 |
step_n = BLOCK_N2
|
287 |
for blk_idx in range(num_steps):
|
288 |
-
kT = tl.load(
|
289 |
-
vT = tl.load(vT_ptrs)
|
290 |
qk = tl.dot(q, kT)
|
291 |
p = tl.math.exp2(qk - m)
|
292 |
# Autoregressive masking.
|
@@ -295,6 +326,7 @@ def _attn_bwd_dq(dq, q, K, V, #
|
|
295 |
mask = (offs_m[:, None] >= offs_n[None, :])
|
296 |
p = tl.where(mask, p, 0.0)
|
297 |
# Compute dP and dS.
|
|
|
298 |
dp = tl.dot(do, vT).to(tl.float32)
|
299 |
ds = p * (dp - Di[:, None])
|
300 |
ds = ds.to(tl.float16)
|
@@ -303,25 +335,49 @@ def _attn_bwd_dq(dq, q, K, V, #
|
|
303 |
dq += tl.dot(ds, tl.trans(kT))
|
304 |
# Increment pointers.
|
305 |
curr_n += step_n
|
306 |
-
|
307 |
-
|
308 |
return dq
|
309 |
|
310 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
311 |
@triton.jit
|
312 |
-
def _attn_bwd(Q, K, V, sm_scale,
|
313 |
-
DO,
|
314 |
-
DQ, DK, DV,
|
315 |
M, D,
|
316 |
# shared by Q/K/V/DO.
|
317 |
-
stride_z, stride_h, stride_tok, stride_d,
|
318 |
-
H, N_CTX
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
323 |
-
|
324 |
-
|
|
|
325 |
LN2: tl.constexpr = 0.6931471824645996 # = ln(2)
|
326 |
|
327 |
bhid = tl.program_id(2)
|
@@ -340,58 +396,91 @@ def _attn_bwd(Q, K, V, sm_scale, #
|
|
340 |
M += off_chz
|
341 |
D += off_chz
|
342 |
|
343 |
-
|
344 |
-
offs_k = tl.arange(0, HEAD_DIM)
|
345 |
|
346 |
start_n = pid * BLOCK_N1
|
|
|
|
|
|
|
347 |
start_m = start_n
|
348 |
|
349 |
MASK_BLOCK_M1: tl.constexpr = BLOCK_M1 // BLK_SLICE_FACTOR
|
350 |
offs_n = start_n + tl.arange(0, BLOCK_N1)
|
351 |
|
352 |
-
dv = tl.zeros([BLOCK_N1,
|
353 |
-
dk = tl.zeros([BLOCK_N1,
|
354 |
|
355 |
-
|
356 |
-
|
357 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
|
359 |
num_steps = BLOCK_N1 // MASK_BLOCK_M1
|
360 |
|
361 |
-
dk, dv = _attn_bwd_dkdv(dk, dv,
|
362 |
-
Q, k, v, sm_scale,
|
363 |
-
DO,
|
364 |
-
M, D,
|
365 |
-
stride_tok, stride_d,
|
366 |
-
H, N_CTX,
|
367 |
-
MASK_BLOCK_M1, BLOCK_N1,
|
368 |
-
start_n, start_m, num_steps,
|
369 |
-
MASK=True
|
370 |
)
|
371 |
|
372 |
start_m += num_steps * MASK_BLOCK_M1
|
373 |
num_steps = (N_CTX - start_m) // BLOCK_M1
|
374 |
|
375 |
# Compute dK and dV for non-masked blocks.
|
376 |
-
dk, dv = _attn_bwd_dkdv(
|
377 |
-
dk, dv,
|
378 |
-
Q, k, v, sm_scale,
|
379 |
-
DO,
|
380 |
-
M, D,
|
381 |
-
stride_tok, stride_d,
|
382 |
-
H, N_CTX,
|
383 |
-
BLOCK_M1, BLOCK_N1,
|
384 |
-
start_n, start_m, num_steps,
|
385 |
-
MASK=False
|
386 |
)
|
387 |
|
388 |
-
|
389 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
390 |
|
391 |
# Write back dK.
|
392 |
dk *= sm_scale
|
393 |
-
|
394 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
395 |
|
396 |
# THIS BLOCK DOES DQ:
|
397 |
start_m = pid * BLOCK_M2
|
@@ -400,10 +489,26 @@ def _attn_bwd(Q, K, V, sm_scale, #
|
|
400 |
MASK_BLOCK_N2: tl.constexpr = BLOCK_N2 // BLK_SLICE_FACTOR
|
401 |
offs_m = start_m + tl.arange(0, BLOCK_M2)
|
402 |
|
403 |
-
|
404 |
-
|
405 |
-
|
406 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
407 |
|
408 |
m = tl.load(M + offs_m)
|
409 |
m = m[:, None]
|
@@ -414,29 +519,39 @@ def _attn_bwd(Q, K, V, sm_scale, #
|
|
414 |
# not due to anything important. I just wanted to reuse the loop
|
415 |
# structure for dK & dV above as much as possible.
|
416 |
num_steps = BLOCK_M2 // MASK_BLOCK_N2
|
417 |
-
dq = _attn_bwd_dq(dq, q, K, V,
|
418 |
-
do, m, D,
|
419 |
-
stride_tok, stride_d,
|
420 |
-
H, N_CTX,
|
421 |
-
BLOCK_M2, MASK_BLOCK_N2,
|
422 |
-
start_m, end_n - num_steps * MASK_BLOCK_N2, num_steps,
|
423 |
-
MASK=True
|
424 |
)
|
425 |
end_n -= num_steps * MASK_BLOCK_N2
|
426 |
# stage 2
|
427 |
num_steps = end_n // BLOCK_N2
|
428 |
-
dq = _attn_bwd_dq(dq, q, K, V,
|
429 |
-
do, m, D,
|
430 |
-
stride_tok, stride_d,
|
431 |
-
H, N_CTX,
|
432 |
-
BLOCK_M2, BLOCK_N2,
|
433 |
-
start_m, end_n - num_steps * BLOCK_N2, num_steps,
|
434 |
-
MASK=False
|
435 |
)
|
436 |
# Write back dQ.
|
437 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
438 |
dq *= LN2
|
439 |
-
tl.store(
|
|
|
|
|
|
|
440 |
|
441 |
|
442 |
class _attention(torch.autograd.Function):
|
@@ -444,45 +559,58 @@ class _attention(torch.autograd.Function):
|
|
444 |
@staticmethod
|
445 |
def forward(ctx, q, k, v, causal, sm_scale):
|
446 |
# shape constraints
|
447 |
-
|
448 |
-
|
449 |
-
|
450 |
-
|
451 |
-
|
452 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
453 |
stage = 3 if causal else 1
|
454 |
-
|
455 |
-
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
def grid(args): return (triton.cdiv(
|
462 |
-
q.shape[2], args["BLOCK_M"]), q.shape[0] * q.shape[1], 1)
|
463 |
-
M = torch.empty((q.shape[0], q.shape[1], q.shape[2]),
|
464 |
device=q.device, dtype=torch.float32)
|
465 |
_attn_fwd[grid](
|
466 |
-
q, k, v, sm_scale, M, o,
|
467 |
-
q.stride(0), q.stride(1), q.stride(2), q.stride(3),
|
468 |
-
k.stride(0), k.stride(1), k.stride(2), k.stride(3),
|
469 |
-
v.stride(0), v.stride(1), v.stride(2), v.stride(3),
|
470 |
-
o.stride(0), o.stride(1), o.stride(2), o.stride(3),
|
471 |
-
q.shape[0], q.shape[1],
|
472 |
-
N_CTX=q.shape[2],
|
473 |
-
|
474 |
-
STAGE=stage,
|
475 |
-
|
|
|
|
|
|
|
|
|
|
|
476 |
|
477 |
ctx.save_for_backward(q, k, v, o, M)
|
478 |
ctx.grid = grid
|
479 |
ctx.sm_scale = sm_scale
|
480 |
-
ctx.
|
481 |
ctx.causal = causal
|
482 |
return o
|
483 |
|
484 |
@staticmethod
|
485 |
def backward(ctx, do):
|
|
|
|
|
|
|
|
|
486 |
q, k, v, o, M = ctx.saved_tensors
|
487 |
assert do.is_contiguous()
|
488 |
assert q.stride() == k.stride() == v.stride() == o.stride() == do.stride()
|
@@ -491,34 +619,33 @@ class _attention(torch.autograd.Function):
|
|
491 |
dv = torch.empty_like(v)
|
492 |
BATCH, N_HEAD, N_CTX = q.shape[:3]
|
493 |
PRE_BLOCK = 128
|
494 |
-
NUM_WARPS, NUM_STAGES = 4,
|
495 |
-
BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 32,
|
496 |
BLK_SLICE_FACTOR = 2
|
497 |
RCP_LN2 = 1.4426950408889634 # = 1.0 / ln(2)
|
498 |
arg_k = k
|
499 |
arg_k = arg_k * (ctx.sm_scale * RCP_LN2)
|
500 |
-
PRE_BLOCK = 128
|
501 |
assert N_CTX % PRE_BLOCK == 0
|
502 |
pre_grid = (N_CTX // PRE_BLOCK, BATCH * N_HEAD)
|
503 |
delta = torch.empty_like(M)
|
504 |
_attn_bwd_preprocess[pre_grid](
|
505 |
-
o, do,
|
506 |
-
delta,
|
507 |
-
BATCH, N_HEAD, N_CTX,
|
508 |
-
BLOCK_M=PRE_BLOCK,
|
|
|
|
|
|
|
|
|
|
|
|
|
509 |
)
|
510 |
-
grid = (N_CTX // BLOCK_N1, 1, BATCH * N_HEAD)
|
511 |
_attn_bwd[grid](
|
512 |
-
q, arg_k, v, ctx.sm_scale, do, dq, dk, dv,
|
513 |
-
M, delta,
|
514 |
-
q.stride(0), q.stride(1), q.stride(2), q.stride(3),
|
515 |
-
N_HEAD, N_CTX,
|
516 |
-
|
517 |
-
BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2, #
|
518 |
-
BLK_SLICE_FACTOR=BLK_SLICE_FACTOR, #
|
519 |
-
HEAD_DIM=ctx.HEAD_DIM, #
|
520 |
-
num_warps=NUM_WARPS, #
|
521 |
-
num_stages=NUM_STAGES #
|
522 |
)
|
523 |
|
524 |
return dq, dk, dv, None, None
|
|
|
11 |
|
12 |
"""
|
13 |
|
14 |
+
import pytest
|
15 |
import torch
|
16 |
|
17 |
import triton
|
18 |
import triton.language as tl
|
19 |
|
20 |
+
# Pick the fp8 data type
|
21 |
|
22 |
+
# AMD E4M3B8
|
23 |
+
# Note: When picking this f8 data type, scaling is required when using f8
|
24 |
+
# for the second gemm
|
25 |
+
# TORCH_HAS_FP8E4B8 = hasattr(torch, 'float8_e4m3fnuz')
|
26 |
+
|
27 |
+
# AMD E5M2B16
|
28 |
+
TORCH_HAS_FP8E5B16 = hasattr(torch, 'float8_e5m2fnuz')
|
29 |
|
30 |
|
31 |
@triton.jit
|
32 |
+
def _attn_fwd_inner(acc, l_i, m_i, q,
|
33 |
+
K_block_ptr, V_block_ptr,
|
34 |
+
start_m,
|
35 |
+
BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr, BLOCK_N: tl.constexpr,
|
36 |
+
STAGE: tl.constexpr, offs_m: tl.constexpr, offs_n: tl.constexpr,
|
37 |
+
N_CTX,
|
38 |
+
pre_load_v: tl.constexpr):
|
39 |
# range of values handled by this stage
|
40 |
if STAGE == 1:
|
41 |
lo, hi = 0, start_m * BLOCK_M
|
42 |
elif STAGE == 2:
|
43 |
lo, hi = start_m * BLOCK_M, (start_m + 1) * BLOCK_M
|
44 |
lo = tl.multiple_of(lo, BLOCK_M)
|
45 |
+
K_block_ptr = tl.advance(K_block_ptr, (0, lo))
|
46 |
+
V_block_ptr = tl.advance(V_block_ptr, (lo, 0))
|
47 |
# causal = False
|
48 |
else:
|
49 |
lo, hi = 0, N_CTX
|
|
|
|
|
50 |
# loop over k, v and update accumulator
|
51 |
for start_n in range(lo, hi, BLOCK_N):
|
52 |
start_n = tl.multiple_of(start_n, BLOCK_N)
|
53 |
# -- compute qk ----
|
54 |
k = tl.load(K_block_ptr)
|
55 |
+
if pre_load_v:
|
56 |
+
v = tl.load(V_block_ptr)
|
57 |
+
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
|
58 |
if STAGE == 2:
|
59 |
mask = offs_m[:, None] >= (start_n + offs_n[None, :])
|
60 |
+
qk = tl.where(mask, qk, float("-inf"))
|
61 |
+
qk += tl.dot(q, k)
|
62 |
+
m_ij = tl.maximum(m_i, tl.max(qk, 1))
|
63 |
+
qk = qk - m_ij[:, None]
|
|
|
|
|
64 |
p = tl.math.exp2(qk)
|
|
|
|
|
|
|
|
|
65 |
# -- update output accumulator --
|
66 |
+
alpha = tl.math.exp2(m_i - m_ij)
|
67 |
acc = acc * alpha[:, None]
|
68 |
+
if not pre_load_v:
|
69 |
+
v = tl.load(V_block_ptr)
|
70 |
+
acc += tl.dot(p.to(v.dtype), v)
|
71 |
+
# -- update m_i and l_i
|
72 |
+
l_ij = tl.sum(p, 1)
|
73 |
+
l_i = l_i * alpha + l_ij
|
|
|
74 |
# update m_i and l_i
|
75 |
m_i = m_ij
|
76 |
V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
|
|
|
78 |
return acc, l_i, m_i
|
79 |
|
80 |
|
81 |
+
# We don't run auto-tuning everytime to keep the tutorial fast. Uncommenting
|
82 |
# the code below and commenting out the equivalent parameters is convenient for
|
83 |
# re-tuning.
|
84 |
+
@triton.autotune(
|
85 |
+
configs=[
|
86 |
+
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 16, 'waves_per_eu': 2,
|
87 |
+
'slice_k_tile': 0, 'pre_load_v': False}, num_stages=1, num_warps=2),
|
88 |
+
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 16, 'waves_per_eu': 2,
|
89 |
+
'slice_k_tile': 32, 'pre_load_v': False}, num_stages=1, num_warps=2),
|
90 |
+
triton.Config({'BLOCK_M': 32, 'BLOCK_N': 32, 'waves_per_eu': 2,
|
91 |
+
'slice_k_tile': 0, 'pre_load_v': False}, num_stages=1, num_warps=1),
|
92 |
+
triton.Config({'BLOCK_M': 32, 'BLOCK_N': 32, 'waves_per_eu': 2,
|
93 |
+
'slice_k_tile': 32, 'pre_load_v': False}, num_stages=1, num_warps=1),
|
94 |
+
triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'waves_per_eu': 2,
|
95 |
+
'slice_k_tile': 0, 'pre_load_v': False}, num_stages=1, num_warps=2),
|
96 |
+
triton.Config({'BLOCK_M': 32, 'BLOCK_N': 16, 'waves_per_eu': 3,
|
97 |
+
'slice_k_tile': 0, 'pre_load_v': True}, num_stages=1, num_warps=1),
|
98 |
+
triton.Config({'BLOCK_M': 32, 'BLOCK_N': 16, 'waves_per_eu': 3,
|
99 |
+
'slice_k_tile': 0, 'pre_load_v': False}, num_stages=1, num_warps=1),
|
100 |
+
],
|
101 |
+
key=['Z', 'H', 'N_CTX', 'STAGE', 'BLOCK_DMODEL'],
|
102 |
+
)
|
103 |
@triton.jit
|
104 |
+
def _attn_fwd(Q, K, V, sm_scale, M, Out,
|
105 |
+
stride_qz, stride_qh, stride_qm, stride_qk,
|
106 |
+
stride_kz, stride_kh, stride_kn, stride_kk,
|
107 |
+
stride_vz, stride_vh, stride_vk, stride_vn,
|
108 |
+
stride_oz, stride_oh, stride_om, stride_on,
|
109 |
+
Z, H,
|
110 |
+
N_CTX,
|
111 |
+
BLOCK_DMODEL: tl.constexpr,
|
112 |
+
STAGE: tl.constexpr,
|
113 |
+
BLOCK_M: tl.constexpr,
|
114 |
+
BLOCK_N: tl.constexpr,
|
115 |
+
pre_load_v: tl.constexpr,
|
116 |
):
|
|
|
117 |
start_m = tl.program_id(0)
|
118 |
off_hz = tl.program_id(1)
|
119 |
+
qvk_offset = off_hz * stride_qh
|
|
|
|
|
|
|
120 |
|
121 |
# block pointers
|
122 |
Q_block_ptr = tl.make_block_ptr(
|
123 |
base=Q + qvk_offset,
|
124 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
125 |
strides=(stride_qm, stride_qk),
|
126 |
offsets=(start_m * BLOCK_M, 0),
|
127 |
+
block_shape=(BLOCK_M, BLOCK_DMODEL),
|
128 |
order=(1, 0),
|
129 |
)
|
|
|
|
|
130 |
V_block_ptr = tl.make_block_ptr(
|
131 |
base=V + qvk_offset,
|
132 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
133 |
strides=(stride_vk, stride_vn),
|
134 |
offsets=(0, 0),
|
135 |
+
block_shape=(BLOCK_N, BLOCK_DMODEL),
|
136 |
+
order=(1, 0),
|
137 |
)
|
138 |
K_block_ptr = tl.make_block_ptr(
|
139 |
base=K + qvk_offset,
|
140 |
+
shape=(BLOCK_DMODEL, N_CTX),
|
141 |
strides=(stride_kk, stride_kn),
|
142 |
offsets=(0, 0),
|
143 |
+
block_shape=(BLOCK_DMODEL, BLOCK_N),
|
144 |
order=(0, 1),
|
145 |
)
|
146 |
O_block_ptr = tl.make_block_ptr(
|
147 |
base=Out + qvk_offset,
|
148 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
149 |
strides=(stride_om, stride_on),
|
150 |
offsets=(start_m * BLOCK_M, 0),
|
151 |
+
block_shape=(BLOCK_M, BLOCK_DMODEL),
|
152 |
order=(1, 0),
|
153 |
)
|
154 |
# initialize offsets
|
|
|
157 |
# initialize pointer to m and l
|
158 |
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
|
159 |
l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0
|
160 |
+
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
|
161 |
+
# scale sm_scale by log_2(e) and use
|
162 |
+
# 2^x instead of exp in the loop because CSE and LICM
|
163 |
+
# don't work as expected with `exp` in the loop
|
164 |
+
qk_scale = sm_scale * 1.44269504
|
165 |
+
# load q: it will stay in SRAM throughout on NV GPUs but in VGPRs on AMD GPUs
|
166 |
q = tl.load(Q_block_ptr)
|
167 |
+
q = (q * qk_scale).to(q.dtype)
|
168 |
# stage 1: off-band
|
169 |
# For causal = True, STAGE = 3 and _attn_fwd_inner gets 1 as its STAGE
|
170 |
# For causal = False, STAGE = 1, and _attn_fwd_inner gets 3 as its STAGE
|
171 |
if STAGE & 1:
|
172 |
+
acc, l_i, m_i = _attn_fwd_inner(acc, l_i, m_i, q, K_block_ptr, V_block_ptr,
|
173 |
+
start_m,
|
174 |
+
BLOCK_M, BLOCK_DMODEL, BLOCK_N,
|
175 |
+
4 - STAGE, offs_m, offs_n, N_CTX,
|
176 |
+
pre_load_v,
|
177 |
)
|
178 |
# stage 2: on-band
|
179 |
if STAGE & 2:
|
180 |
# barrier makes it easier for compielr to schedule the
|
181 |
# two loops independently
|
182 |
+
tl.debug_barrier()
|
183 |
+
acc, l_i, m_i = _attn_fwd_inner(acc, l_i, m_i, q, K_block_ptr, V_block_ptr,
|
184 |
+
start_m,
|
185 |
+
BLOCK_M, BLOCK_DMODEL, BLOCK_N,
|
186 |
+
2, offs_m, offs_n, N_CTX,
|
187 |
+
pre_load_v,
|
188 |
)
|
189 |
# epilogue
|
190 |
+
# write back m
|
191 |
acc = acc / l_i[:, None]
|
192 |
m_ptrs = M + off_hz * N_CTX + offs_m
|
193 |
+
tl.store(m_ptrs, m_i + tl.math.log2(l_i))
|
194 |
tl.store(O_block_ptr, acc.to(Out.type.element_ty))
|
195 |
|
196 |
|
197 |
@triton.jit
|
198 |
+
def _attn_bwd_preprocess(O, DO,
|
199 |
+
Delta,
|
200 |
+
Z, H, N_CTX,
|
201 |
+
BLOCK_M: tl.constexpr, D_HEAD: tl.constexpr
|
202 |
):
|
203 |
off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
|
204 |
off_hz = tl.program_id(1)
|
205 |
+
off_n = tl.arange(0, D_HEAD)
|
206 |
+
o = tl.load(O + off_hz * D_HEAD * N_CTX +
|
207 |
+
off_m[:, None] * D_HEAD + off_n[None, :])
|
208 |
+
do = tl.load(DO + off_hz * D_HEAD * N_CTX +
|
209 |
+
off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
|
|
|
210 |
delta = tl.sum(o * do, axis=1)
|
|
|
211 |
tl.store(Delta + off_hz * N_CTX + off_m, delta)
|
212 |
|
213 |
|
214 |
# The main inner-loop logic for computing dK and dV.
|
215 |
@triton.jit
|
216 |
+
def _attn_bwd_dkdv(dk, dv,
|
217 |
+
Q, k, v, sm_scale,
|
218 |
+
DO,
|
219 |
+
M, D,
|
220 |
# shared by Q/K/V/DO.
|
221 |
+
stride_tok, stride_d,
|
222 |
+
H, N_CTX, BLOCK_M1: tl.constexpr,
|
223 |
+
BLOCK_N1: tl.constexpr,
|
224 |
+
BLOCK_DMODEL: tl.constexpr,
|
225 |
# Filled in by the wrapper.
|
226 |
+
start_n, start_m, num_steps,
|
227 |
MASK: tl.constexpr):
|
228 |
offs_m = start_m + tl.arange(0, BLOCK_M1)
|
229 |
offs_n = start_n + tl.arange(0, BLOCK_N1)
|
230 |
+
offs_k = tl.arange(0, BLOCK_DMODEL)
|
231 |
+
QT_block_ptr = tl.make_block_ptr(
|
232 |
+
base=Q,
|
233 |
+
shape=(BLOCK_DMODEL, N_CTX),
|
234 |
+
strides=(stride_d, stride_tok),
|
235 |
+
offsets=(0, start_m),
|
236 |
+
block_shape=(BLOCK_DMODEL, BLOCK_M1),
|
237 |
+
order=(0, 1)
|
238 |
+
)
|
239 |
+
DO_block_ptr = tl.make_block_ptr(
|
240 |
+
base=DO,
|
241 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
242 |
+
strides=(stride_tok, stride_d),
|
243 |
+
offsets=(start_m, 0),
|
244 |
+
block_shape=(BLOCK_M1, BLOCK_DMODEL),
|
245 |
+
order=(1, 0)
|
246 |
+
)
|
247 |
# BLOCK_N1 must be a multiple of BLOCK_M1, otherwise the code wouldn't work.
|
248 |
tl.static_assert(BLOCK_N1 % BLOCK_M1 == 0)
|
249 |
curr_m = start_m
|
250 |
step_m = BLOCK_M1
|
251 |
for blk_idx in range(num_steps):
|
252 |
+
qT = tl.load(QT_block_ptr)
|
253 |
# Load m before computing qk to reduce pipeline stall.
|
254 |
offs_m = curr_m + tl.arange(0, BLOCK_M1)
|
255 |
m = tl.load(M + offs_m)
|
|
|
259 |
if MASK:
|
260 |
mask = (offs_m[None, :] >= offs_n[:, None])
|
261 |
pT = tl.where(mask, pT, 0.0)
|
262 |
+
do = tl.load(DO_block_ptr)
|
263 |
# Compute dV.
|
264 |
ppT = pT
|
265 |
ppT = ppT.to(tl.float16)
|
|
|
267 |
# D (= delta) is pre-divided by ds_scale.
|
268 |
Di = tl.load(D + offs_m)
|
269 |
# Compute dP and dS.
|
270 |
+
dpT = tl.dot(v, tl.trans(do))
|
271 |
dsT = pT * (dpT - Di[None, :])
|
272 |
dsT = dsT.to(tl.float16)
|
273 |
dk += tl.dot(dsT, tl.trans(qT))
|
274 |
# Increment pointers.
|
275 |
curr_m += step_m
|
276 |
+
QT_block_ptr = tl.advance(QT_block_ptr, (0, step_m))
|
277 |
+
DO_block_ptr = tl.advance(DO_block_ptr, (step_m, 0))
|
278 |
return dk, dv
|
279 |
|
280 |
|
281 |
# the main inner-loop logic for computing dQ
|
282 |
@triton.jit
|
283 |
+
def _attn_bwd_dq(dq, q, K, V,
|
284 |
do, m, D,
|
285 |
# shared by Q/K/V/DO.
|
286 |
+
stride_tok, stride_d,
|
287 |
+
H, N_CTX,
|
288 |
+
BLOCK_M2: tl.constexpr,
|
289 |
+
BLOCK_N2: tl.constexpr,
|
290 |
+
BLOCK_DMODEL: tl.constexpr,
|
291 |
# Filled in by the wrapper.
|
292 |
+
start_m, start_n, num_steps,
|
293 |
MASK: tl.constexpr):
|
294 |
offs_m = start_m + tl.arange(0, BLOCK_M2)
|
295 |
offs_n = start_n + tl.arange(0, BLOCK_N2)
|
296 |
+
offs_k = tl.arange(0, BLOCK_DMODEL)
|
297 |
+
KT_block_ptr = tl.make_block_ptr(
|
298 |
+
base=K,
|
299 |
+
shape=(BLOCK_DMODEL, N_CTX),
|
300 |
+
strides=(stride_d, stride_tok),
|
301 |
+
offsets=(0, start_n),
|
302 |
+
block_shape=(BLOCK_DMODEL, BLOCK_N2),
|
303 |
+
order=(0, 1)
|
304 |
+
)
|
305 |
+
VT_block_ptr = tl.make_block_ptr(
|
306 |
+
base=V,
|
307 |
+
shape=(BLOCK_DMODEL, N_CTX),
|
308 |
+
strides=(stride_d, stride_tok),
|
309 |
+
offsets=(0, start_n),
|
310 |
+
block_shape=(BLOCK_DMODEL, BLOCK_N2),
|
311 |
+
order=(0, 1)
|
312 |
+
)
|
313 |
# D (= delta) is pre-divided by ds_scale.
|
314 |
Di = tl.load(D + offs_m)
|
315 |
# BLOCK_M2 must be a multiple of BLOCK_N2, otherwise the code wouldn't work.
|
|
|
317 |
curr_n = start_n
|
318 |
step_n = BLOCK_N2
|
319 |
for blk_idx in range(num_steps):
|
320 |
+
kT = tl.load(KT_block_ptr)
|
|
|
321 |
qk = tl.dot(q, kT)
|
322 |
p = tl.math.exp2(qk - m)
|
323 |
# Autoregressive masking.
|
|
|
326 |
mask = (offs_m[:, None] >= offs_n[None, :])
|
327 |
p = tl.where(mask, p, 0.0)
|
328 |
# Compute dP and dS.
|
329 |
+
vT = tl.load(VT_block_ptr)
|
330 |
dp = tl.dot(do, vT).to(tl.float32)
|
331 |
ds = p * (dp - Di[:, None])
|
332 |
ds = ds.to(tl.float16)
|
|
|
335 |
dq += tl.dot(ds, tl.trans(kT))
|
336 |
# Increment pointers.
|
337 |
curr_n += step_n
|
338 |
+
KT_block_ptr = tl.advance(KT_block_ptr, (0, step_n))
|
339 |
+
VT_block_ptr = tl.advance(VT_block_ptr, (0, step_n))
|
340 |
return dq
|
341 |
|
342 |
|
343 |
+
@triton.autotune(
|
344 |
+
configs=[
|
345 |
+
triton.Config({'BLOCK_M1': 32, 'BLOCK_N1': 64, 'BLOCK_M2': 64, 'BLOCK_N2': 32, 'BLK_SLICE_FACTOR': 1},
|
346 |
+
num_stages=1, num_warps=4),
|
347 |
+
triton.Config({'BLOCK_M1': 32, 'BLOCK_N1': 64, 'BLOCK_M2': 64, 'BLOCK_N2': 32, 'BLK_SLICE_FACTOR': 2},
|
348 |
+
num_stages=1, num_warps=4),
|
349 |
+
triton.Config({'BLOCK_M1': 64, 'BLOCK_N1': 128, 'BLOCK_M2': 128, 'BLOCK_N2': 64, 'BLK_SLICE_FACTOR': 1},
|
350 |
+
num_stages=1, num_warps=4),
|
351 |
+
triton.Config({'BLOCK_M1': 64, 'BLOCK_N1': 128, 'BLOCK_M2': 128, 'BLOCK_N2': 64, 'BLK_SLICE_FACTOR': 2},
|
352 |
+
num_stages=1, num_warps=4),
|
353 |
+
triton.Config({'BLOCK_M1': 64, 'BLOCK_N1': 64, 'BLOCK_M2': 64, 'BLOCK_N2': 64, 'BLK_SLICE_FACTOR': 1},
|
354 |
+
num_stages=1, num_warps=4),
|
355 |
+
triton.Config({'BLOCK_M1': 64, 'BLOCK_N1': 64, 'BLOCK_M2': 64, 'BLOCK_N2': 64, 'BLK_SLICE_FACTOR': 2},
|
356 |
+
num_stages=1, num_warps=4),
|
357 |
+
triton.Config({'BLOCK_M1': 32, 'BLOCK_N1': 128, 'BLOCK_M2': 128, 'BLOCK_N2': 32, 'BLK_SLICE_FACTOR': 1},
|
358 |
+
num_stages=1, num_warps=4),
|
359 |
+
triton.Config({'BLOCK_M1': 32, 'BLOCK_N1': 128, 'BLOCK_M2': 128, 'BLOCK_N2': 32, 'BLK_SLICE_FACTOR': 2},
|
360 |
+
num_stages=1, num_warps=4),
|
361 |
+
triton.Config({'BLOCK_M1': 32, 'BLOCK_N1': 128, 'BLOCK_M2': 128, 'BLOCK_N2': 32, 'BLK_SLICE_FACTOR': 2},
|
362 |
+
num_stages=1, num_warps=8),
|
363 |
+
],
|
364 |
+
key=['H', 'N_CTX', 'BLOCK_DMODEL'],
|
365 |
+
)
|
366 |
@triton.jit
|
367 |
+
def _attn_bwd(Q, K, V, sm_scale,
|
368 |
+
DO,
|
369 |
+
DQ, DK, DV,
|
370 |
M, D,
|
371 |
# shared by Q/K/V/DO.
|
372 |
+
stride_z, stride_h, stride_tok, stride_d,
|
373 |
+
# H = 16, N_CTX = 1024
|
374 |
+
H, N_CTX,
|
375 |
+
BLOCK_DMODEL: tl.constexpr,
|
376 |
+
BLOCK_M1: tl.constexpr,
|
377 |
+
BLOCK_N1: tl.constexpr,
|
378 |
+
BLOCK_M2: tl.constexpr,
|
379 |
+
BLOCK_N2: tl.constexpr,
|
380 |
+
BLK_SLICE_FACTOR: tl.constexpr):
|
381 |
LN2: tl.constexpr = 0.6931471824645996 # = ln(2)
|
382 |
|
383 |
bhid = tl.program_id(2)
|
|
|
396 |
M += off_chz
|
397 |
D += off_chz
|
398 |
|
399 |
+
offs_k = tl.arange(0, BLOCK_DMODEL)
|
|
|
400 |
|
401 |
start_n = pid * BLOCK_N1
|
402 |
+
# This assignment is important. It is what allows us to pick the diagonal
|
403 |
+
# blocks. Later, when we want to do the lower triangular, we update start_m
|
404 |
+
# after the first dkdv call.
|
405 |
start_m = start_n
|
406 |
|
407 |
MASK_BLOCK_M1: tl.constexpr = BLOCK_M1 // BLK_SLICE_FACTOR
|
408 |
offs_n = start_n + tl.arange(0, BLOCK_N1)
|
409 |
|
410 |
+
dv = tl.zeros([BLOCK_N1, BLOCK_DMODEL], dtype=tl.float32)
|
411 |
+
dk = tl.zeros([BLOCK_N1, BLOCK_DMODEL], dtype=tl.float32)
|
412 |
|
413 |
+
K_block_ptr = tl.make_block_ptr(
|
414 |
+
base=K,
|
415 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
416 |
+
strides=(stride_tok, stride_d),
|
417 |
+
offsets=(start_n, 0),
|
418 |
+
block_shape=(BLOCK_N1, BLOCK_DMODEL),
|
419 |
+
order=(1, 0),
|
420 |
+
)
|
421 |
+
V_block_ptr = tl.make_block_ptr(
|
422 |
+
base=V,
|
423 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
424 |
+
strides=(stride_tok, stride_d),
|
425 |
+
offsets=(start_n, 0),
|
426 |
+
block_shape=(BLOCK_N1, BLOCK_DMODEL),
|
427 |
+
order=(1, 0),
|
428 |
+
)
|
429 |
+
|
430 |
+
# load K and V: they stay in SRAM throughout the inner loop for dkdv.
|
431 |
+
k = tl.load(K_block_ptr)
|
432 |
+
v = tl.load(V_block_ptr)
|
433 |
|
434 |
num_steps = BLOCK_N1 // MASK_BLOCK_M1
|
435 |
|
436 |
+
dk, dv = _attn_bwd_dkdv(dk, dv,
|
437 |
+
Q, k, v, sm_scale,
|
438 |
+
DO,
|
439 |
+
M, D,
|
440 |
+
stride_tok, stride_d,
|
441 |
+
H, N_CTX,
|
442 |
+
MASK_BLOCK_M1, BLOCK_N1, BLOCK_DMODEL,
|
443 |
+
start_n, start_m, num_steps,
|
444 |
+
MASK=True
|
445 |
)
|
446 |
|
447 |
start_m += num_steps * MASK_BLOCK_M1
|
448 |
num_steps = (N_CTX - start_m) // BLOCK_M1
|
449 |
|
450 |
# Compute dK and dV for non-masked blocks.
|
451 |
+
dk, dv = _attn_bwd_dkdv(
|
452 |
+
dk, dv,
|
453 |
+
Q, k, v, sm_scale,
|
454 |
+
DO,
|
455 |
+
M, D,
|
456 |
+
stride_tok, stride_d,
|
457 |
+
H, N_CTX,
|
458 |
+
BLOCK_M1, BLOCK_N1, BLOCK_DMODEL,
|
459 |
+
start_n, start_m, num_steps,
|
460 |
+
MASK=False
|
461 |
)
|
462 |
|
463 |
+
DV_block_ptrs = tl.make_block_ptr(
|
464 |
+
base=DV,
|
465 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
466 |
+
strides=(stride_tok, stride_d),
|
467 |
+
offsets=(start_n, 0),
|
468 |
+
block_shape=(BLOCK_N1, BLOCK_DMODEL),
|
469 |
+
order=(1, 0)
|
470 |
+
)
|
471 |
+
tl.store(DV_block_ptrs, dv.to(tl.float16))
|
472 |
|
473 |
# Write back dK.
|
474 |
dk *= sm_scale
|
475 |
+
DK_block_ptrs = tl.make_block_ptr(
|
476 |
+
base=DK,
|
477 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
478 |
+
strides=(stride_tok, stride_d),
|
479 |
+
offsets=(start_n, 0),
|
480 |
+
block_shape=(BLOCK_N1, BLOCK_DMODEL),
|
481 |
+
order=(1, 0)
|
482 |
+
)
|
483 |
+
tl.store(DK_block_ptrs, dk.to(tl.float16))
|
484 |
|
485 |
# THIS BLOCK DOES DQ:
|
486 |
start_m = pid * BLOCK_M2
|
|
|
489 |
MASK_BLOCK_N2: tl.constexpr = BLOCK_N2 // BLK_SLICE_FACTOR
|
490 |
offs_m = start_m + tl.arange(0, BLOCK_M2)
|
491 |
|
492 |
+
Q_block_ptr = tl.make_block_ptr(
|
493 |
+
base=Q,
|
494 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
495 |
+
strides=(stride_tok, stride_d),
|
496 |
+
offsets=(start_m, 0),
|
497 |
+
block_shape=(BLOCK_M2, BLOCK_DMODEL),
|
498 |
+
order=(1, 0)
|
499 |
+
)
|
500 |
+
|
501 |
+
DO_block_ptr = tl.make_block_ptr(
|
502 |
+
base=DO,
|
503 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
504 |
+
strides=(stride_tok, stride_d),
|
505 |
+
offsets=(start_m, 0),
|
506 |
+
block_shape=(BLOCK_M2, BLOCK_DMODEL),
|
507 |
+
order=(1, 0)
|
508 |
+
)
|
509 |
+
q = tl.load(Q_block_ptr)
|
510 |
+
do = tl.load(DO_block_ptr)
|
511 |
+
dq = tl.zeros([BLOCK_M2, BLOCK_DMODEL], dtype=tl.float32)
|
512 |
|
513 |
m = tl.load(M + offs_m)
|
514 |
m = m[:, None]
|
|
|
519 |
# not due to anything important. I just wanted to reuse the loop
|
520 |
# structure for dK & dV above as much as possible.
|
521 |
num_steps = BLOCK_M2 // MASK_BLOCK_N2
|
522 |
+
dq = _attn_bwd_dq(dq, q, K, V,
|
523 |
+
do, m, D,
|
524 |
+
stride_tok, stride_d,
|
525 |
+
H, N_CTX,
|
526 |
+
BLOCK_M2, MASK_BLOCK_N2, BLOCK_DMODEL,
|
527 |
+
start_m, end_n - num_steps * MASK_BLOCK_N2, num_steps,
|
528 |
+
MASK=True
|
529 |
)
|
530 |
end_n -= num_steps * MASK_BLOCK_N2
|
531 |
# stage 2
|
532 |
num_steps = end_n // BLOCK_N2
|
533 |
+
dq = _attn_bwd_dq(dq, q, K, V,
|
534 |
+
do, m, D,
|
535 |
+
stride_tok, stride_d,
|
536 |
+
H, N_CTX,
|
537 |
+
BLOCK_M2, BLOCK_N2, BLOCK_DMODEL,
|
538 |
+
start_m, end_n - num_steps * BLOCK_N2, num_steps,
|
539 |
+
MASK=False
|
540 |
)
|
541 |
# Write back dQ.
|
542 |
+
DQ_block_ptr = tl.make_block_ptr(
|
543 |
+
base=DQ,
|
544 |
+
shape=(N_CTX, BLOCK_DMODEL),
|
545 |
+
strides=(stride_tok, stride_d),
|
546 |
+
offsets=(start_m, 0),
|
547 |
+
block_shape=(BLOCK_M2, BLOCK_DMODEL),
|
548 |
+
order=(1, 0)
|
549 |
+
)
|
550 |
dq *= LN2
|
551 |
+
tl.store(DQ_block_ptr, dq.to(tl.float16))
|
552 |
+
|
553 |
+
|
554 |
+
empty = torch.empty(128, device="cuda")
|
555 |
|
556 |
|
557 |
class _attention(torch.autograd.Function):
|
|
|
559 |
@staticmethod
|
560 |
def forward(ctx, q, k, v, causal, sm_scale):
|
561 |
# shape constraints
|
562 |
+
Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
|
563 |
+
assert Lq == Lk and Lk == Lv
|
564 |
+
assert Lk in {16, 32, 64, 128}
|
565 |
+
o = torch.empty_like(q, dtype=v.dtype)
|
566 |
+
if torch.version.hip is None:
|
567 |
+
BLOCK_M = 128
|
568 |
+
BLOCK_N = 64 if Lk <= 64 else 32
|
569 |
+
num_stages = 4 if Lk <= 64 else 3
|
570 |
+
num_warps = 4 if Lk <= 64 else 8
|
571 |
+
# Tuning for H100
|
572 |
+
if torch.cuda.get_device_capability()[0] == 9:
|
573 |
+
num_warps = 8
|
574 |
+
num_stages = 7 if Lk >= 64 else 3
|
575 |
stage = 3 if causal else 1
|
576 |
+
|
577 |
+
def grid(META): return (
|
578 |
+
triton.cdiv(q.shape[2], META['BLOCK_M']),
|
579 |
+
q.shape[0] * q.shape[1],
|
580 |
+
1
|
581 |
+
)
|
582 |
+
M = torch.empty((q.shape[0] * q.shape[1], q.shape[2]),
|
|
|
|
|
|
|
583 |
device=q.device, dtype=torch.float32)
|
584 |
_attn_fwd[grid](
|
585 |
+
q, k, v, sm_scale, M, o,
|
586 |
+
q.stride(0), q.stride(1), q.stride(2), q.stride(3),
|
587 |
+
k.stride(0), k.stride(1), k.stride(2), k.stride(3),
|
588 |
+
v.stride(0), v.stride(1), v.stride(2), v.stride(3),
|
589 |
+
o.stride(0), o.stride(1), o.stride(2), o.stride(3),
|
590 |
+
q.shape[0], q.shape[1],
|
591 |
+
N_CTX=q.shape[2],
|
592 |
+
BLOCK_DMODEL=Lk,
|
593 |
+
STAGE=stage,
|
594 |
+
)
|
595 |
+
|
596 |
+
# restore the grid for bwd kernel
|
597 |
+
best_config = _attn_fwd.get_best_config()
|
598 |
+
block_m = int(best_config.__str__().split(",")[0].split("BLOCK_M:")[1])
|
599 |
+
grid = (triton.cdiv(q.shape[2], block_m), q.shape[0] * q.shape[1], 1)
|
600 |
|
601 |
ctx.save_for_backward(q, k, v, o, M)
|
602 |
ctx.grid = grid
|
603 |
ctx.sm_scale = sm_scale
|
604 |
+
ctx.BLOCK_DMODEL = Lk
|
605 |
ctx.causal = causal
|
606 |
return o
|
607 |
|
608 |
@staticmethod
|
609 |
def backward(ctx, do):
|
610 |
+
if torch.version.hip is not None:
|
611 |
+
BLOCK = 64
|
612 |
+
else:
|
613 |
+
BLOCK = 128
|
614 |
q, k, v, o, M = ctx.saved_tensors
|
615 |
assert do.is_contiguous()
|
616 |
assert q.stride() == k.stride() == v.stride() == o.stride() == do.stride()
|
|
|
619 |
dv = torch.empty_like(v)
|
620 |
BATCH, N_HEAD, N_CTX = q.shape[:3]
|
621 |
PRE_BLOCK = 128
|
622 |
+
NUM_WARPS, NUM_STAGES = 4, 1
|
623 |
+
BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 32, 64, 64, 32
|
624 |
BLK_SLICE_FACTOR = 2
|
625 |
RCP_LN2 = 1.4426950408889634 # = 1.0 / ln(2)
|
626 |
arg_k = k
|
627 |
arg_k = arg_k * (ctx.sm_scale * RCP_LN2)
|
|
|
628 |
assert N_CTX % PRE_BLOCK == 0
|
629 |
pre_grid = (N_CTX // PRE_BLOCK, BATCH * N_HEAD)
|
630 |
delta = torch.empty_like(M)
|
631 |
_attn_bwd_preprocess[pre_grid](
|
632 |
+
o, do,
|
633 |
+
delta,
|
634 |
+
BATCH, N_HEAD, N_CTX,
|
635 |
+
BLOCK_M=PRE_BLOCK, D_HEAD=ctx.BLOCK_DMODEL
|
636 |
+
)
|
637 |
+
|
638 |
+
def grid(META): return (
|
639 |
+
triton.cdiv(N_CTX, META['BLOCK_N1']),
|
640 |
+
1,
|
641 |
+
BATCH * N_HEAD
|
642 |
)
|
|
|
643 |
_attn_bwd[grid](
|
644 |
+
q, arg_k, v, ctx.sm_scale, do, dq, dk, dv,
|
645 |
+
M, delta,
|
646 |
+
q.stride(0), q.stride(1), q.stride(2), q.stride(3),
|
647 |
+
N_HEAD, N_CTX,
|
648 |
+
BLOCK_DMODEL=ctx.BLOCK_DMODEL
|
|
|
|
|
|
|
|
|
|
|
649 |
)
|
650 |
|
651 |
return dq, dk, dv, None, None
|