take care of caching for simple gateloop

Author: lucidrains

gateloop_transformer/simplified_gate_loop.py

@@ -4,15 +4,18 @@
 
 from typing import Tuple
 
-from einops import rearrange
+from einops import rearrange, pack, unpack
 from einops.layers.torch import Rearrange
 
 from gateloop_transformer.gateloop_transformer import RMSNorm
 from gateloop_transformer.associative_scan import associative_scan
 
 # plain pytorch non-fused associative scan
 
-def gate_loop_operator(q, kv, a):
+def exists(v):
+    return v is not None
+
+def gate_loop_operator(q, kv, a, cache = None):
 
     @torch.jit.script
     def binary_operator(
@@ -23,9 +26,18 @@ def binary_operator(
         a_j, kv_j = b
         return a_j * a_i, torch.addcmul(kv_j, a_j, kv_i)
 
-    _, kv = associative_scan(binary_operator, (a, kv))
+    if exists(cache):
+        cache_a, cache_kv = cache
+        a, a_ps = pack([cache_a, a], 'b * d')
+        kv, kv_ps = pack([cache_kv, kv], 'b * d')
+
+    a, kv = associative_scan(binary_operator, (a, kv))
+
+    if exists(cache):
+        _, a = unpack(a, a_ps, 'b * d')
+        _, kv = unpack(kv, kv_ps, 'b * d')
 
-    return q * kv
+    return q * kv, (a[:, -1], kv[:, -1])
 
 # using jax associative scan
 
@@ -48,7 +60,7 @@ def binary_operator(e_i, e_j):
 
         return q * y
 
-    return jax2torch(jax_gate_loop_operator)
+    return jax2torch(jax_gate_loop_operator), None
 
 # simple gate loop layer
 
@@ -75,6 +87,8 @@ def __init__(
             Rearrange('b n (qkva d) -> qkva (b d) n 1', qkva = 3)
         )
 
+        self.use_jax = use_jax_associative_scan
+
         if use_jax_associative_scan:
             self.gate_loop_fn = get_jax_gate_loop_operator()
         else:
@@ -84,20 +98,30 @@ def __init__(
 
         self.reverse = reverse
 
-    def forward(self, x):
-
+    def forward(
+        self,
+        x,
+        cache = None,
+        return_cache = False
+    ):
         if self.reverse:
             x = torch.flip(x, dims = (-2,))
 
         x = self.norm(x)
 
         q, kv, a = self.to_qkva(x)
 
-        out = self.gate_loop_fn(q, kv, a.sigmoid())
+        out, cache = self.gate_loop_fn(q, kv, a.sigmoid(), cache = cache)
 
         out = self.split_heads(out)
 
         if self.reverse:
             out = torch.flip(out, dims = (-2,))
 
-        return out
+        if not return_cache:
+            return out
+
+        assert not self.reverse, 'caching only works with non-reversed seq'
+        assert not self.use_jax, 'jax associative scan does not have caching yet'
+
+        return out, cache

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'gateloop-transformer',
   packages = find_packages(exclude=[]),
-  version = '0.1.1',
+  version = '0.1.4',
   license='MIT',
   description = 'GateLoop Transformer',
   author = 'Phil Wang',