fx k,v history after project
Browse files- audiocraft/lm.py +1 -1
- audiocraft/transformer.py +20 -16
audiocraft/lm.py
CHANGED
@@ -147,7 +147,7 @@ class LMModel(nn.Module):
|
|
147 |
super().__init__()
|
148 |
self.cfg_coef = cfg_coef
|
149 |
|
150 |
-
self.n_draw =
|
151 |
self.condition_provider = condition_provider
|
152 |
self.fuser = fuser
|
153 |
self.card = card # 2048 ?
|
|
|
147 |
super().__init__()
|
148 |
self.cfg_coef = cfg_coef
|
149 |
|
150 |
+
self.n_draw = 2
|
151 |
self.condition_provider = condition_provider
|
152 |
self.fuser = fuser
|
153 |
self.card = card # 2048 ?
|
audiocraft/transformer.py
CHANGED
@@ -194,21 +194,12 @@ class StreamingMultiheadAttention(nn.Module):
|
|
194 |
|
195 |
q, k, v = [rearrange(x, f"b t (h d) -> {layout}", h=self.num_heads) for x in [q, k, v]]
|
196 |
else:
|
197 |
-
|
|
|
|
|
|
|
198 |
# HISTORY - DIFFERENT FOR EACH TRANSF LAYER
|
199 |
-
|
200 |
-
#
|
201 |
-
# pk.shape=torch.Size([2, 24, 3, 64]) k.shape=torch.Size([2, 24, 1, 64]) CONCAT
|
202 |
-
# has to be 4D with batch 1 due to single condition 3=seqlen
|
203 |
-
# 24 heads 64 dimofh
|
204 |
-
self.k_history = torch.cat([self.k_history, query], 2)
|
205 |
-
self.v_history = torch.cat([self.v_history, query], 2)
|
206 |
-
else:
|
207 |
-
# init on 1st token (for all 47 transf layers)
|
208 |
-
self.k_history = query
|
209 |
-
self.v_history = query
|
210 |
-
|
211 |
-
|
212 |
projected = nn.functional.linear(query, self.in_proj_weight, self.in_proj_bias)
|
213 |
if self.kv_repeat == 1:
|
214 |
# if time_dim == 2:
|
@@ -217,7 +208,21 @@ class StreamingMultiheadAttention(nn.Module):
|
|
217 |
# bound_layout = "b t p h d"
|
218 |
packed = rearrange(projected, f"b t (p h d) -> {bound_layout}", p=3, h=self.num_heads)
|
219 |
q, k, v = ops.unbind(packed, dim=2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
220 |
|
|
|
|
|
221 |
|
222 |
|
223 |
|
@@ -235,8 +240,7 @@ class StreamingMultiheadAttention(nn.Module):
|
|
235 |
# k, v = self._complete_kv(k, v)
|
236 |
# print(k.sum(), v.sum(), k.shape, v.shape,'ATTNext')
|
237 |
|
238 |
-
|
239 |
-
q, k, v = [x.float() for x in [q, k, v]]
|
240 |
if self.memory_efficient:
|
241 |
# print('EVER IN MEMORY EFFICIENT A')
|
242 |
|
|
|
194 |
|
195 |
q, k, v = [rearrange(x, f"b t (h d) -> {layout}", h=self.num_heads) for x in [q, k, v]]
|
196 |
else:
|
197 |
+
# 1st projected makes k,v (instantaneous)
|
198 |
+
# 2nd cat
|
199 |
+
|
200 |
+
|
201 |
# HISTORY - DIFFERENT FOR EACH TRANSF LAYER
|
202 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
203 |
projected = nn.functional.linear(query, self.in_proj_weight, self.in_proj_bias)
|
204 |
if self.kv_repeat == 1:
|
205 |
# if time_dim == 2:
|
|
|
208 |
# bound_layout = "b t p h d"
|
209 |
packed = rearrange(projected, f"b t (p h d) -> {bound_layout}", p=3, h=self.num_heads)
|
210 |
q, k, v = ops.unbind(packed, dim=2)
|
211 |
+
|
212 |
+
if self.k_history is not None:
|
213 |
+
#
|
214 |
+
# pk.shape=torch.Size([2, 24, 3, 64]) k.shape=torch.Size([2, 24, 1, 64]) CONCAT
|
215 |
+
# has to be 4D with batch 1 due to single condition 3=seqlen
|
216 |
+
# 24 heads 64 dimofh
|
217 |
+
self.k_history = torch.cat([self.k_history, k], 2)
|
218 |
+
self.v_history = torch.cat([self.v_history, v], 2)
|
219 |
+
else:
|
220 |
+
# init on 1st token (for all 47 transf layers)
|
221 |
+
self.k_history = k
|
222 |
+
self.v_history = v
|
223 |
|
224 |
+
k = self.k_history
|
225 |
+
v = self.v_history
|
226 |
|
227 |
|
228 |
|
|
|
240 |
# k, v = self._complete_kv(k, v)
|
241 |
# print(k.sum(), v.sum(), k.shape, v.shape,'ATTNext')
|
242 |
|
243 |
+
print(f'{self.attention_as_float32=}')
|
|
|
244 |
if self.memory_efficient:
|
245 |
# print('EVER IN MEMORY EFFICIENT A')
|
246 |
|