From 584cfbdc433cfb3d7852868db83060b6d9aa0edf Mon Sep 17 00:00:00 2001
From: Yuekai Zhang <zhangyuekai@foxmail.com>
Date: 星期一, 15 七月 2024 18:43:19 +0800
Subject: [PATCH] Add triton server for SenseVoice (#1901)
---
funasr/models/transformer/attention.py | 83 ++++++++++++++++-------------------------
1 files changed, 32 insertions(+), 51 deletions(-)
diff --git a/funasr/models/transformer/attention.py b/funasr/models/transformer/attention.py
index 695023d..6e6f754 100644
--- a/funasr/models/transformer/attention.py
+++ b/funasr/models/transformer/attention.py
@@ -17,6 +17,7 @@
from funasr.models.transformer.utils.nets_utils import make_pad_mask
import funasr.models.lora.layers as lora
+
class MultiHeadedAttention(nn.Module):
"""Multi-Head Attention layer.
@@ -81,9 +82,10 @@
n_batch = value.size(0)
if mask is not None:
mask = mask.unsqueeze(1).eq(0) # (batch, 1, *, time2)
- min_value = float(
- numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min
- )
+
+ min_value = -float(
+ "inf"
+ ) # min_value = float(np.finfo(torch.tensor(0, dtype=qk.dtype).numpy().dtype).min)
scores = scores.masked_fill(mask, min_value)
self.attn = torch.softmax(scores, dim=-1).masked_fill(
mask, 0.0
@@ -129,17 +131,17 @@
self.linear_out = model.linear_out
self.attn = None
self.all_head_size = self.h * self.d_k
-
+
def forward(self, query, key, value, mask):
q, k, v = self.forward_qkv(query, key, value)
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
return self.forward_attention(v, scores, mask)
-
+
def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
new_x_shape = x.size()[:-1] + (self.h, self.d_k)
x = x.view(new_x_shape)
return x.permute(0, 2, 1, 3)
-
+
def forward_qkv(self, query, key, value):
q = self.linear_q(query)
k = self.linear_k(key)
@@ -148,13 +150,13 @@
k = self.transpose_for_scores(k)
v = self.transpose_for_scores(v)
return q, k, v
-
+
def forward_attention(self, value, scores, mask):
scores = scores + mask
-
+
self.attn = torch.softmax(scores, dim=-1)
context_layer = torch.matmul(self.attn, value) # (batch, head, time1, d_k)
-
+
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
context_layer = context_layer.view(new_context_layer_shape)
@@ -167,51 +169,49 @@
self.linear_pos = model.linear_pos
self.pos_bias_u = model.pos_bias_u
self.pos_bias_v = model.pos_bias_v
-
+
def forward(self, query, key, value, pos_emb, mask):
q, k, v = self.forward_qkv(query, key, value)
q = q.transpose(1, 2) # (batch, time1, head, d_k)
-
+
p = self.transpose_for_scores(self.linear_pos(pos_emb)) # (batch, head, time1, d_k)
-
+
# (batch, head, time1, d_k)
q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
# (batch, head, time1, d_k)
q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
-
+
# compute attention score
# first compute matrix a and matrix c
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
# (batch, head, time1, time2)
matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))
-
+
# compute matrix b and matrix d
# (batch, head, time1, time1)
matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
matrix_bd = self.rel_shift(matrix_bd)
-
- scores = (matrix_ac + matrix_bd) / math.sqrt(
- self.d_k
- ) # (batch, head, time1, time2)
-
+
+ scores = (matrix_ac + matrix_bd) / math.sqrt(self.d_k) # (batch, head, time1, time2)
+
return self.forward_attention(v, scores, mask)
-
+
def rel_shift(self, x):
zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype)
x_padded = torch.cat([zero_pad, x], dim=-1)
-
+
x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
x = x_padded[:, :, 1:].view_as(x)[
:, :, :, : x.size(-1) // 2 + 1
- ] # only keep the positions from 0 to time2
+ ] # only keep the positions from 0 to time2
return x
-
+
def forward_attention(self, value, scores, mask):
scores = scores + mask
-
+
self.attn = torch.softmax(scores, dim=-1)
context_layer = torch.matmul(self.attn, value) # (batch, head, time1, d_k)
-
+
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
context_layer = context_layer.view(new_context_layer_shape)
@@ -306,9 +306,7 @@
matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
matrix_bd = self.rel_shift(matrix_bd)
- scores = (matrix_ac + matrix_bd) / math.sqrt(
- self.d_k
- ) # (batch, head, time1, time2)
+ scores = (matrix_ac + matrix_bd) / math.sqrt(self.d_k) # (batch, head, time1, time2)
return self.forward_attention(v, scores, mask)
@@ -358,7 +356,7 @@
x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
x = x_padded[:, :, 1:].view_as(x)[
:, :, :, : x.size(-1) // 2 + 1
- ] # only keep the positions from 0 to time2
+ ] # only keep the positions from 0 to time2
if self.zero_triu:
ones = torch.ones((x.size(2), x.size(3)), device=x.device)
@@ -405,9 +403,7 @@
matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
matrix_bd = self.rel_shift(matrix_bd)
- scores = (matrix_ac + matrix_bd) / math.sqrt(
- self.d_k
- ) # (batch, head, time1, time2)
+ scores = (matrix_ac + matrix_bd) / math.sqrt(self.d_k) # (batch, head, time1, time2)
return self.forward_attention(v, scores, mask)
@@ -552,21 +548,9 @@
"""
n_batch = query.size(0)
- q = (
- self.linear_q(query)
- .view(n_batch, -1, self.num_heads, self.d_k)
- .transpose(1, 2)
- )
- k = (
- self.linear_k(key)
- .view(n_batch, -1, self.num_heads, self.d_k)
- .transpose(1, 2)
- )
- v = (
- self.linear_v(value)
- .view(n_batch, -1, self.num_heads, self.d_k)
- .transpose(1, 2)
- )
+ q = self.linear_q(query).view(n_batch, -1, self.num_heads, self.d_k).transpose(1, 2)
+ k = self.linear_k(key).view(n_batch, -1, self.num_heads, self.d_k).transpose(1, 2)
+ v = self.linear_v(value).view(n_batch, -1, self.num_heads, self.d_k).transpose(1, 2)
return q, k, v
@@ -597,9 +581,7 @@
attn_output = torch.matmul(attn_output, value)
attn_output = self.linear_out(
- attn_output.transpose(1, 2)
- .contiguous()
- .view(batch_size, -1, self.num_heads * self.d_k)
+ attn_output.transpose(1, 2).contiguous().view(batch_size, -1, self.num_heads * self.d_k)
)
return attn_output
@@ -629,4 +611,3 @@
q, k, v = self.forward_qkv(query, key, value)
scores = self.compute_att_score(q, k, pos_enc, left_context=left_context)
return self.forward_attention(v, scores, mask, chunk_mask=chunk_mask)
-
--
Gitblit v1.9.1