x-transformers 1.39.4__py3-none-any.whl → 1.40.1__py3-none-any.whl

x_transformers/attend.py

@@ -109,12 +109,35 @@ def qk_l2_dist_squared(q, k):
 
 # one-hot straight through softmax
 
-def one_hot_straight_through(t, temperature = 1.):
-    one_hot_indices = t.argmax(dim = -1, keepdim = True)
-    one_hot = torch.zeros_like(t).scatter(-1, one_hot_indices, 1.)
+def one_hot_straight_through(logits, temperature = 1.):
+    one_hot_indices = logits.argmax(dim = -1, keepdim = True)
+    one_hot = torch.zeros_like(logits).scatter(-1, one_hot_indices, 1.)
 
-    t = (t / temperature).softmax(dim = -1)
-    return one_hot + t - t.detach()
+    soft_attn = (logits / temperature).softmax(dim = -1)
+    return one_hot + soft_attn - soft_attn.detach()
+
+# sparse topk attention - only keep topk attn logits for softmax
+# optional straight through with masked out logits by setting `attn_sparse_topk_straight_through = True`
+
+def sparse_topk_attn(
+    logits,
+    sparse_topk,
+    temperature = 1.,
+    straight_through = False
+):
+    orig_logits = logits
+
+    mask_value = -torch.finfo(logits.dtype).max
+    top_values, _ = logits.topk(sparse_topk, dim = -1)
+    sparse_topk_mask = (logits >= top_values[..., -1:]) & (logits > mask_value)
+    logits = logits.masked_fill(~sparse_topk_mask, mask_value)
+    topk_attn = logits.softmax(dim = -1)
+
+    if not straight_through:
+        return topk_attn
+
+    soft_attn = (orig_logits / temperature).softmax(dim = -1)
+    return topk_attn.detach() + soft_attn - soft_attn.detach()
 
 # functions for creating causal mask
 # need a special one for onnx cpu (no support for .triu)
@@ -141,6 +164,7 @@ class Attend(Module):
         post_talking_heads = False,
         pre_scale_post_talking_heads = False,
         sparse_topk = None,
+        sparse_topk_straight_through = False,
         scale = None,
         qk_norm = False,
         l2_distance = False,
@@ -152,7 +176,6 @@ class Attend(Module):
         add_zero_kv = False,
         selective = False,
         hard = False,
-        sigsoftmax = False,
         cope = None,
         onnxable = False,
         sdp_kwargs: dict = dict(
@@ -171,9 +194,13 @@ class Attend(Module):
 
         # attention type
 
+        is_sparse_topk_attn = exists(sparse_topk)
+
         assert not (flash and sigmoid), 'sigmoid attention not available for flash'
         assert not (flash and hard), 'hard attention not available for flash'
-        assert at_most_one_of(sigmoid, hard, l2_distance)
+        assert not (flash and is_sparse_topk_attn), 'topk attention not available for flash'
+
+        assert at_most_one_of(sigmoid, hard, l2_distance, is_sparse_topk_attn)
 
         if exists(custom_attn_fn):
             self.attn_fn = custom_attn_fn
@@ -181,6 +208,8 @@ class Attend(Module):
             self.attn_fn = F.sigmoid
         elif hard:
             self.attn_fn = one_hot_straight_through
+        elif is_sparse_topk_attn:
+            self.attn_fn = partial(sparse_topk_attn, sparse_topk = sparse_topk, straight_through = sparse_topk_straight_through)
         else:
             softmax_fn = partial(F.softmax, dim = -1)
             self.attn_fn = partial(softmax_fn, dtype = torch.float32) if not qk_norm else softmax_fn
@@ -214,16 +243,6 @@ class Attend(Module):
         assert not (selective and not causal), 'selective attention is designed for autoregressive'
         self.selective = selective
 
-        # sparse topk
-
-        assert not (flash and sparse_topk), 'sparse topk not compatible with flash attention'
-        self.sparse_topk = sparse_topk
-
-        # sig softmax
-
-        assert not (flash and sigsoftmax), 'sigsoftmax not available for flash attention'
-        self.sigsoftmax = sigsoftmax
-
         # l2 distance attention
 
         self.l2_distance = l2_distance
@@ -476,11 +495,6 @@ class Attend(Module):
             causal_mask = self.create_causal_mask(i, j, device = device)
             sim = sim.masked_fill(causal_mask, mask_value)
 
-        if exists(self.sparse_topk):
-            top_values, _ = sim.topk(self.sparse_topk, dim = -1)
-            sparse_topk_mask = (sim >= top_values[..., -1:]) & (sim > mask_value)
-            sim = sim.masked_fill(~sparse_topk_mask, mask_value)
-
         row_is_entirely_masked = None
 
         if exists(mask):
@@ -494,9 +508,6 @@ class Attend(Module):
 
         pre_softmax_attn = sim
 
-        if self.sigsoftmax:
-            sim = sim + sim.sigmoid().log()
-
         attn = self.attn_fn(sim)
 
         attn = attn.type(dtype)

x_transformers/x_transformers.py

@@ -912,6 +912,7 @@ class Attention(Module):
         pre_scale_post_talking_heads = False,
         head_scale = False,
         sparse_topk = None,
+        sparse_topk_straight_through = False,
         num_mem_kv = 0,
         dropout = 0.,
         on_attn = False,
@@ -920,7 +921,6 @@ class Attention(Module):
         gate_values = False,
         zero_init_output = False,
         hard = False,
-        sigsoftmax = False,
         max_attend_past = None,
         qk_norm = False,
         qk_norm_groups = 1,
@@ -1044,6 +1044,7 @@ class Attention(Module):
             pre_scale_post_talking_heads = pre_scale_post_talking_heads,
             dropout = dropout,
             sparse_topk = sparse_topk,
+            sparse_topk_straight_through = sparse_topk_straight_through,
             hard = hard,
             qk_norm = qk_norm,
             scale = qk_norm_scale if qk_norm else self.scale,
@@ -1054,7 +1055,6 @@ class Attention(Module):
             add_zero_kv = add_zero_kv,
             flash = flash,
             softclamp_logits = softclamp_logits,
-            sigsoftmax = sigsoftmax,
             logit_softclamp_value = logit_softclamp_value,
             cope = cope,
             onnxable = onnxable

{x_transformers-1.39.4.dist-info → x_transformers-1.40.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.39.4
+Version: 1.40.1
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.39.4.dist-info → x_transformers-1.40.1.dist-info}/RECORD

@@ -1,15 +1,15 @@
 x_transformers/__init__.py,sha256=-MkQrSc37cTVDX7AOykxunYnqVtFlQ7lb0Cse5dsGWU,793
-x_transformers/attend.py,sha256=013qsFwoABVbyc-1L3RZTRCWo6BW9fAD8IVnC_qALGk,16708
+x_transformers/attend.py,sha256=VbB0fi-ETgAF4dc2a_Meaqvt14LMaRVIjZ8NexUX8F0,17239
 x_transformers/autoregressive_wrapper.py,sha256=DOJJCMMDOqDYKWy_IaG5IyKsXD3AW6amzfUgdAADOLY,10500
 x_transformers/continuous.py,sha256=cIVEdhfei258__ziV7kQBrJMxCel54bExBTDrO9rfCI,6450
 x_transformers/dpo.py,sha256=LjvWgCkqTl-UuehrzQ8nkX5guLr4whYwsmm7SKSwdls,3450
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
-x_transformers/x_transformers.py,sha256=3JrVqwYVrd5UVf2esdunTcer7QL72H7VF4mL3UsCWOI,84508
+x_transformers/x_transformers.py,sha256=TGXJZXCWR5BiMkS5Kx-JhFQ85AxkiJabLiHnrCTC874,84562
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
-x_transformers-1.39.4.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.39.4.dist-info/METADATA,sha256=2KawHim0IOdlRjbRJCVsELM10T7nojxnMy6WrWtG0UE,661
-x_transformers-1.39.4.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
-x_transformers-1.39.4.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.39.4.dist-info/RECORD,,
+x_transformers-1.40.1.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.40.1.dist-info/METADATA,sha256=WouMl3Ld1llknOwj7BcKi-_YZ9Hx9RZ-ni-eGCP_uQY,661
+x_transformers-1.40.1.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
+x_transformers-1.40.1.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.40.1.dist-info/RECORD,,