rxnn 0.1.35__py3-none-any.whl → 0.1.37__py3-none-any.whl

rxnn/experimental/attention.py

@@ -614,7 +614,7 @@ def init_moe_attention(
         num_query_experts: int = None,
         num_query_groups: int = None,
 ) -> GroupedQueryAttention:
-    assert attention_type in ['gma', 'dma', 'gma_v', 'dma_v'], "Error, attention type should be one of: 'gma', 'dma', 'gma_v', 'dma_v'"
+    assert attention_type in ['gma', 'dma', 'gma_s', 'dma_s'], "Error, attention type should be one of: 'gma', 'dma', 'gma_s', 'dma_s'"
 
     if attention_type == "gma":
         return GroupedMoeAttention(
@@ -648,8 +648,8 @@ def init_moe_attention(
             num_query_experts=num_query_experts,
             num_query_groups=num_query_groups,
         )
-    elif attention_type == "gma_v":
-        return GroupedMoeAttentionVectorized(
+    elif attention_type == "gma_s":
+        return GroupedMoeAttentionSimplified(
             embed_dim,
             num_heads,
             gqa_groups,
@@ -664,7 +664,7 @@ def init_moe_attention(
             num_experts=num_experts,
         )
     else:
-        return DeepMoeAttentionVectorized(
+        return DeepMoeAttentionSimplified(
             embed_dim,
             num_heads,
             gqa_groups,

rxnn/experimental/models.py

@@ -65,7 +65,7 @@ class MoeAttentionTransformer(nn.Module, PyTorchModelHubMixin, pipeline_tag="tex
         assert ff_activation in ['relu', 'gelu',
                                  'swish', 'silu', 'linear',
                                  'sigmoid'], 'Feed-forward activation could be "relu", "gelu", "swish", "silu", "linear", "sigmoid".'
-        assert att_type in ['mha', 'gqa', 'mqa', 'gma', 'dma', 'gma_v', 'dma_v'], 'Self-attention type could be "mha", "gqa", "mqa", "gma", "dma", "gma_v", "dma_v".'
+        assert att_type in ['mha', 'gqa', 'mqa', 'gma', 'dma', 'gma_s', 'dma_s'], 'Self-attention type could be "mha", "gqa", "mqa", "gma", "dma", "gma_s", "dma_s".'
 
         embedding = nn.Embedding(vocab_size, embed_dim)
         rope = RotaryPositionalEmbedding(embed_dim // att_heads, seq_len)

rxnn/transformers/attention.py

@@ -102,8 +102,12 @@ class MultiHeadAttention(nn.Module):
 
     def _flash_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, b: int, t: int, d: int,
                          mask: torch.Tensor = None, enable_gqa: bool = False):
-        with sdpa_kernel(backends=[SDPBackend.FLASH_ATTENTION]):
-            return self._torch_attention(q, k, v, b, t, d, mask=mask, enable_gqa=enable_gqa)
+        # After ~6h of fighthing, PyTorch based is still now working so I decided to use FlashAttention directly
+        # with sdpa_kernel(backends=[SDPBackend.FLASH_ATTENTION]):
+        #     return self._torch_attention(q, k, v, b, t, d, mask=mask, enable_gqa=enable_gqa)
+        from flash_attn import flash_attn_func
+        attn_output = flash_attn_func(q, k, v, dropout_p=self.dropout.p if self.training else 0.0, is_causal=self.is_causal)
+        return self._transpose_output(attn_output, b, t, d)
 
     def _torch_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, b: int, t: int, d: int,
                          mask: torch.Tensor = None, enable_gqa: bool = False):

{rxnn-0.1.35.dist-info → rxnn-0.1.37.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.1.35
+Version: 0.1.37
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.1.35.dist-info → rxnn-0.1.37.dist-info}/RECORD

@@ -1,7 +1,7 @@
 rxnn/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 rxnn/experimental/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-rxnn/experimental/attention.py,sha256=GxbLmOTBvUiYU0Rc_0ju1n_ocJciHC6i3neDGe-rZZc,29426
-rxnn/experimental/models.py,sha256=QEuFBB9iEg5AbKQLwGJkAwPjMfaVeTqazhKDWPRkm7o,4598
+rxnn/experimental/attention.py,sha256=PjmVwNeJXDy72LJr5cl9JD1oqjlwYK-Ahx1K1gLQgf8,29426
+rxnn/experimental/models.py,sha256=IzUVc5s-cA__8jsG2mVvzUDmzPRcfBcI5btaOjnPYhA,4598
 rxnn/experimental/moe.py,sha256=jHZ1QhpWiVQOswVpFmuH7b2IUOPf0Uuf-I2Ddwsd7Us,6140
 rxnn/memory/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 rxnn/memory/norm.py,sha256=Ofl8Q5NYEF9GQeO0bhM43tkTW91J0y6TSvTAOYMgloM,6278
@@ -16,7 +16,7 @@ rxnn/training/dataset.py,sha256=vQ5mDF3bA0HXya474n4D4iL8Mn3AEpJukgzFNVkxjGU,5106
 rxnn/training/scheduler.py,sha256=ow6oALzWjWQmHSpcJEjv6tg4g4CDMvr73TypxfcefMc,712
 rxnn/training/tokenizer.py,sha256=4Y41f07uo2KPA_7bp3FCcwGKbXoS2hsckOoXUsXfQxY,8052
 rxnn/transformers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-rxnn/transformers/attention.py,sha256=zv0uH3_L39tVmpiwNdmEf6Cp602uqdbr3UQj8Z3hIIk,15349
+rxnn/transformers/attention.py,sha256=bsuAXCKR0WbOxgu-IkJHgn7jUu2CK4hqNw60IZbGTEE,15698
 rxnn/transformers/ff.py,sha256=jJnuBDsnnX5uYC_WZH8cXAYrMnz0P-iX7MwcPivjRtI,2533
 rxnn/transformers/layers.py,sha256=n_jZTqEF_vLkF31AkB5XGErfm2sQFd9CRqJUHKRFkKI,6956
 rxnn/transformers/mask.py,sha256=J0cfLVLt3SzS2ra3KcY4khrkhI975Dw4CjpUi3Sn25s,419
@@ -25,7 +25,7 @@ rxnn/transformers/moe.py,sha256=6Cffyo0QjmEWc4rK1ncOmLRCQbY0OpQJ4D7xH_4nTN4,4738
 rxnn/transformers/positional.py,sha256=2l38RS0Dini3f6Z3LUHr3XwWzg1UK7fO2C6wazWDAYU,4292
 rxnn/transformers/sampler.py,sha256=poWBpxg1iuK5gEJtxHkk5VVfS9V48hs2Olqdhy_Gw8c,6548
 rxnn/utils.py,sha256=d5U8i5ukovgDyqiycc2AoxObTz_eF_bgo2MKvdtJ98s,467
-rxnn-0.1.35.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
-rxnn-0.1.35.dist-info/METADATA,sha256=aziCzqOeetdE3gMV2i15QoB5O31bGpiZgzcpGM97QPk,16627
-rxnn-0.1.35.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
-rxnn-0.1.35.dist-info/RECORD,,
+rxnn-0.1.37.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
+rxnn-0.1.37.dist-info/METADATA,sha256=9cKmCtODY8tw_VimbuCN9787asAvNbNylZBK2gOBzLE,16627
+rxnn-0.1.37.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
+rxnn-0.1.37.dist-info/RECORD,,