rxnn 0.2.69__py3-none-any.whl → 0.2.71__py3-none-any.whl

rxnn/memory/attention.py

@@ -64,6 +64,8 @@ class StmMemoryAttention(nn.Module):
                 layer_stm = layer_stm.expand(x.size(0), -1, -1)
             encoded_layer_data = x[i]
             normalized_layer_stm = self.memory_norm_layers[i](layer_stm)
+            if torch.isnan(normalized_layer_stm).any():
+                print(f"NaN detected in {i} layer memory norm output")
 
             if self.debug_mode and self.training:
                 if self.debug_step != 0 and self.debug_step % self.debug_interval == 0:
@@ -72,7 +74,13 @@ class StmMemoryAttention(nn.Module):
                 else:
                     self.debug_step += 1
 
+            if torch.isnan(encoded_layer_data).any():
+                print(f"NaN detected in {i} layer encoded data input")
+
             new_layer_stm = self.attention_layers[i](normalized_layer_stm, encoded_layer_data, encoded_layer_data, mask=attention_mask)
+            if torch.isnan(new_layer_stm).any():
+                print(f"NaN detected in {i} layer memory attention output")
+
             if self.use_gated_residual:
                 new_stm[i] = self._residual_gate(self.gate[i], layer_stm, new_layer_stm) # gated residual
             else:

rxnn/training/dataset.py

@@ -4,7 +4,7 @@ from datasets import Dataset as HfDataset, load_dataset, concatenate_datasets
 from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
 from .tokenizer import load_tokenizer_from_hf_hub
 
-from typing import Union, TypedDict, Optional, TypeAlias, Any
+from typing import Union, TypedDict, Optional, TypeAlias, Any, Literal
 
 
 class BaseDataset(Dataset):
@@ -854,8 +854,8 @@ class EncoderSftDataset(BaseInteractionDataset):
             'labels': labels
         }
 
-
-MrlDataItem: TypeAlias = dict[str, Union[dict[str, torch.Tensor], list[dict[str, dict[str, torch.Tensor]]]]]
+ItemFields: TypeAlias = Literal['input_ids', 'attention_mask']
+MrlDataItem: TypeAlias = dict[str, Union[dict[ItemFields, torch.Tensor], list[dict[str, dict[ItemFields, torch.Tensor]]]]]
 
 
 class MrlCurriculumDataset(Dataset):
@@ -1031,7 +1031,7 @@ class MrlCurriculumDataset(Dataset):
         """Collate function for MRL curriculum dataset with nested interactions"""
 
         def collate_interaction_batch(interaction_batch: Union[list[dict[str, dict[str, torch.Tensor]]], tuple[Any]]) -> \
-        dict[str, dict[str, torch.Tensor]]:
+        dict[str, dict[ItemFields, torch.Tensor]]:
             """Helper to collate a batch of interactions"""
             return {
                 'query': {
@@ -1047,13 +1047,22 @@ class MrlCurriculumDataset(Dataset):
         batch_interactions = [x['interactions'] for x in batch]
         transposed_interactions = list(zip(*batch_interactions))
 
-        return {
+        def has_nans(tensor: dict[ItemFields, torch.Tensor]) -> bool:
+            return torch.isnan(tensor['input_ids']).any().item() or torch.isnan(tensor['attention_mask']).any().item()
+
+        results: MrlDataItem = {
             **collate_interaction_batch(batch),  # Collate initial query and answer
             'interactions': [
                 collate_interaction_batch(step_batch) for step_batch in transposed_interactions
             ]
         }
 
+        assert not has_nans(results['query']), "NaN in query"
+        assert not has_nans(results['answer']), "NaN in answer"
+        assert not any([(has_nans(item['query']) or has_nans(item['answer'])) for item in results['interactions']]), "NaN in interactions"
+
+        return results
+
 
 class MrlDatasetItem(TypedDict):
     steps: int

rxnn/training/models.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 from enum import Enum
-from typing import Literal, Iterator
+from typing import Literal, Iterator, Optional
 from huggingface_hub import PyTorchModelHubMixin
 from ..transformers.models import ReactiveTransformerEncoder, ReactiveTransformerDecoder
 from ..transformers.ff import GatedLinearUnit, get_activation_layer
@@ -188,7 +188,7 @@ class MrlActorModel(nn.Module):
                 list(self.memory_attention_parameters())
             ))
 
-    def moe_router_loss(self):
+    def moe_router_loss(self) -> Optional[torch.Tensor]:
         if self.encoder.model.use_moe and self.decoder.model.use_moe:
             return (self.encoder.model.moe_router_loss() + self.decoder.model.moe_router_loss()) / 2
         elif self.encoder.model.use_moe:

rxnn/training/mrl.py

@@ -591,6 +591,8 @@ class MRLTrainer:
         actor = next(self.actor.children()) if isinstance(self.actor, DistributedDataParallel) else self.actor
 
         router_loss = actor.moe_router_loss()
+        if torch.isnan(router_loss).any():
+            print("NaN detected in router loss")
         if router_loss is not None:
             return main_loss + self.moe_aux_loss_scale * router_loss
         else:
@@ -605,18 +607,38 @@ class MRLTrainer:
         print(f"Encoder grad norm - total: {encoder_total:.6f}, mean: {encoder_mean:.6f}")
         print(f"Decoder grad norm - total: {decoder_total:.6f}, mean: {decoder_mean:.6f}")
         print(f"Memory attention grad norm - total: {mem_att_total:.6f}, mean: {mem_att_mean:.6f}")
-        # decoder's cross att
-        dec_x_att_norms = [get_gradient_norms(layer.memory_cross_attention)[1] for layer in self.actor.decoder.model.layers]
-        print(f"Decoder cross-att mean norm: {(sum(dec_x_att_norms) / len(dec_x_att_norms)):.6f}, all: {dec_x_att_norms}")
 
+        dec_x_att_norms = [get_gradient_norms(layer.memory_cross_attention)[1] for layer in self.actor.decoder.model.layers]
         mem_att_norms = [get_gradient_norms(layer)[1] for layer in self.actor.memory_attention.model.attention_layers]
-        print(f"Memory attention layers mean norm: {(sum(mem_att_norms) / len(mem_att_norms)):.6f}, all: {mem_att_norms}")
-
         enc_ff_norms = [get_gradient_norms(layer.ff)[1] for layer in self.actor.encoder.model.layers]
-        print(f"Encoder ff mean norm: {(sum(enc_ff_norms) / len(enc_ff_norms)):.6f}, all: {enc_ff_norms}")
+        enc_self_att_norms = [get_gradient_norms(layer.attention)[1] for layer in self.actor.encoder.model.layers]
+        enc_x_att_norms = [get_gradient_norms(layer.memory_cross_attention)[1] for layer in
+                         self.actor.encoder.model.layers]
+
+        calc_mean = lambda x: sum(x) / len(x)
+
+        dec_x_att_norms_mean = calc_mean(dec_x_att_norms)
+        mem_att_norms_mean = calc_mean(mem_att_norms)
+        enc_ff_norms_mean = calc_mean(enc_ff_norms)
+        enc_self_att_norms_mean = calc_mean(enc_self_att_norms)
+        enc_x_att_norms_mean = calc_mean(enc_x_att_norms)
+
+        print(f"Decoder cross-att mean norm: {dec_x_att_norms_mean:.6f}, all: {dec_x_att_norms}")
+        print(f"Memory attention layers mean norm: {mem_att_norms_mean:.6f}, all: {mem_att_norms}")
+        print(f"Encoder ff mean norm: {enc_ff_norms_mean:.6f}, all: {enc_ff_norms}")
+        print(f"Encoder self-att mean norm: {enc_self_att_norms_mean:.6f}, all: {enc_self_att_norms}")
+        print(f"Encoder cross-att mean norm: {enc_x_att_norms_mean:.6f}, all: {enc_x_att_norms}")
+
+        if self.writer is not None:
+            self.writer.add_scalar('Gradient/encoder', encoder_mean, self.global_step['train'])
+            self.writer.add_scalar('Gradient/decoder', decoder_mean, self.global_step['train'])
+            self.writer.add_scalar('Gradient/mem-att', mem_att_mean, self.global_step['train'])
+            self.writer.add_scalar('Gradient/decoder x-att', dec_x_att_norms_mean, self.global_step['train'])
+            self.writer.add_scalar('Gradient/mem-att layers', mem_att_norms_mean, self.global_step['train'])
+            self.writer.add_scalar('Gradient/encoder ff', enc_ff_norms_mean, self.global_step['train'])
+            self.writer.add_scalar('Gradient/encoder self-att', enc_self_att_norms_mean, self.global_step['train'])
+            self.writer.add_scalar('Gradient/encoder x-att', enc_x_att_norms_mean, self.global_step['train'])
 
-        enc_ff_norms = [get_gradient_norms(layer.memory_cross_attention)[1] for layer in self.actor.encoder.model.layers]
-        print(f"Encoder cross-att mean norm: {(sum(enc_ff_norms) / len(enc_ff_norms)):.6f}, all: {enc_ff_norms}")
 
     def update_actor(self, state: tuple[TokenizedDict, TokenizedDict, TokenizedDict], action: TokenizedDict,
                      advantages: torch.Tensor, old_log_probs: torch.Tensor, epoch: int) -> float:
@@ -649,7 +671,7 @@ class MRLTrainer:
             # 4.4 Unscale and clip gradient norms
             self.scaler.unscale_(self.optimizer)
             torch.nn.utils.clip_grad_norm_(self.actor.unique_parameters(), max_norm=1.0,
-                                           error_if_nonfinite=False)
+                                           error_if_nonfinite=self.debug_mode)
             if self.debug_mode and self.epoch_step['train'] % self.debug_interval == 0:
                 self._log_gradients(logits)
             # 4.5 Run scaled optimization step
@@ -670,7 +692,7 @@ class MRLTrainer:
             policy_loss.backward(retain_graph=True)
             # 4.4 Clip gradient norms
             torch.nn.utils.clip_grad_norm_(self.actor.unique_parameters(), max_norm=1.0,
-                                           error_if_nonfinite=False)
+                                           error_if_nonfinite=self.debug_mode)
             if self.debug_mode and self.epoch_step['train'] % self.debug_interval == 0:
                 self._log_gradients(logits)
             # 4.5 Run scaled optimization step

rxnn/transformers/layers.py

@@ -102,7 +102,11 @@ class ReactiveTransformerLayer(nn.Module):
         residual = x
         if not self.use_post_norm:
             x = self.norm1(x)
+            if torch.isnan(x).any():
+                print("NaN detected in pre-norm (self-attention) output")
         x = self.attention(x, x, x, mask=mask)
+        if torch.isnan(x).any():
+            print("NaN detected in self-attention output")
         x = residual + x
         if self.use_post_norm:
             x = self.norm1(x)
@@ -110,11 +114,18 @@ class ReactiveTransformerLayer(nn.Module):
         residual = x
         if not self.use_post_norm:
             x = self.norm2(x)
+            if torch.isnan(x).any():
+                print("NaN detected in pre-norm (cross-attention) output")
 
-        if mask is not None:
-            mem_mask = mask.squeeze(1).unsqueeze(-1).expand(-1, -1, -1, stm.size(1))
+        mem_mask = mask.squeeze(1).unsqueeze(-1).expand(-1, -1, -1, stm.size(1)) \
+            if mask is not None else None
+
+        if torch.isnan(stm).any():
+            print("NaN detected in STM cross-attention input")
 
         x = self.memory_cross_attention(x, stm, stm, mask=mem_mask)
+        if torch.isnan(x).any():
+            print("NaN detected in cross-attention output")
         x = residual + x
         if self.use_post_norm:
             x = self.norm2(x)
@@ -123,7 +134,11 @@ class ReactiveTransformerLayer(nn.Module):
         residual = x
         if not self.use_post_norm:
             x = self.norm3(x)
+            if torch.isnan(x).any():
+                print("NaN detected in pre-norm (ff) output")
         x = self.ff(x)
+        if torch.isnan(x).any():
+            print("NaN detected in ff output")
         x = residual + x
         if self.use_post_norm:
             x = self.norm3(x)

rxnn/transformers/models.py

@@ -58,6 +58,15 @@ class ReactiveTransformerBase(nn.Module):
         else:
             return None
 
+    def _handle_layer(self, i: int, x: torch.Tensor, mask: torch.Tensor = None, is_shared: bool = False):
+        stm_layer_idx = i if is_shared else i + self.num_shared_layers
+        layer_stm = self.stm(stm_layer_idx)
+        # expand layer STM to batch size, if it's not in batch mode
+        if layer_stm.size(0) == 1:
+            layer_stm = layer_stm.expand(x.size(0), -1, -1)
+        layer = self.shared_layers[i] if is_shared else self.layers[i]
+        return layer(x, layer_stm, mask=mask)
+
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         # Shared logic for encoders and decoders - apply embeddings and positional encoding
         x = self.embedding(x)
@@ -84,6 +93,8 @@ class ReactiveTransformerDecoder(ReactiveTransformerBase):
 
     def forward(self, x: torch.Tensor, attention_mask: torch.Tensor = None) -> torch.Tensor:
         x = super().forward(x)  # apply embeddings
+        if torch.isnan(x).any():
+            print("NaN detected in decoder embedding output")
         seq_len = x.size(1)
         if not self.use_flash_attention and self.use_relative_embedding:
             mask = create_causal_mask(seq_len, device=x.device)
@@ -96,18 +107,12 @@ class ReactiveTransformerDecoder(ReactiveTransformerBase):
         # Process shared layers
         if self.shared_layers is not None:
             for i in range(self.num_shared_layers):
-                layer_stm = self.stm(i)
-                # expand layer STM to batch size, if it's not in batch mode
-                if layer_stm.size(0) == 1:
-                    layer_stm = layer_stm.expand(x.size(0), -1, -1)
-                x = self.shared_layers[i](x, layer_stm, mask=mask)
+                x = self._handle_layer(i, x, mask=mask, is_shared=True)
         # Process own layers
         for i in range(self.num_own_layers):
-            layer_stm = self.stm(i)
-            # expand layer STM to batch size, if it's not in batch mode
-            if layer_stm.size(0) == 1:
-                layer_stm = layer_stm.expand(x.size(0), -1, -1)
-            x = self.layers[i](x, layer_stm, mask=mask)
+            x = self._handle_layer(i, x, mask=mask)
+            if torch.isnan(x).any():
+                print(f"NaN detected in {i}. decoder layer output")
         return self.head(self.head_norm(x) if self.use_head_norm else x)
 
 
@@ -116,6 +121,8 @@ class ReactiveTransformerEncoder(ReactiveTransformerBase):
 
     def forward(self, x: torch.Tensor, attention_mask: torch.Tensor = None) -> tuple[torch.Tensor, torch.Tensor]:
         x = super().forward(x)  # apply embeddings
+        if torch.isnan(x).any():
+            print("NaN detected in encoder embedding output")
         if attention_mask is not None:
             attention_mask = attention_mask.unsqueeze(1).unsqueeze(1).bool()
 
@@ -123,19 +130,13 @@ class ReactiveTransformerEncoder(ReactiveTransformerBase):
         # Process shared layers
         if self.shared_layers is not None:
             for i in range(self.num_shared_layers):
-                layer_stm = self.stm(i)
-                # expand layer STM to batch size, if it's not in batch mode
-                if layer_stm.size(0) == 1:
-                    layer_stm = layer_stm.expand(x.size(0), -1, -1)
-                x = self.shared_layers[i](x, layer_stm, mask=attention_mask)
+                x = self._handle_layer(i, x, mask=attention_mask, is_shared=True)
                 hidden_states.append(x)
         # Process own layers
         for i in range(self.num_own_layers):
-            layer_stm = self.stm(i)
-            # expand layer STM to batch size, if it's not in batch mode
-            if layer_stm.size(0) == 1:
-                layer_stm = layer_stm.expand(x.size(0), -1, -1)
-            x = self.layers[i](x, layer_stm, mask=attention_mask)
+            x = self._handle_layer(i, x, mask=attention_mask)
+            if torch.isnan(x).any():
+                print(f"NaN detected in {i}. encoder layer output")
             hidden_states.append(x)
         return x, torch.stack(hidden_states)
 

{rxnn-0.2.69.dist-info → rxnn-0.2.71.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.2.69
+Version: 0.2.71
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.2.69.dist-info → rxnn-0.2.71.dist-info}/RECORD

@@ -5,7 +5,7 @@ rxnn/experimental/attention.py,sha256=jlNS82INjycNEfmk3HtkIacUvT_ELhaCO2g-kZTvhX
 rxnn/experimental/models.py,sha256=KheR1zSNJIaeVvpVAkEJwcuM5nOqQP0ZF08XhrtGJ8E,5387
 rxnn/experimental/moe.py,sha256=jHZ1QhpWiVQOswVpFmuH7b2IUOPf0Uuf-I2Ddwsd7Us,6140
 rxnn/memory/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-rxnn/memory/attention.py,sha256=lSniKrf_skiM1V1zbfmV84PbKoQ-t_fVcKfwNKW3_OY,3844
+rxnn/memory/attention.py,sha256=O4ycW3KKP5hFYadgVh47LvGWJn9zNHz8vh9E9okC0h8,4223
 rxnn/memory/norm.py,sha256=cVjjhCLqR5K6-321SP_ObG17y-ddlcTJeCTXvW4vpk0,6675
 rxnn/memory/stm.py,sha256=jv57gsH9XW19sLbxpRDqsp1yfsii_4Ef4Ncr_ztk-i4,3937
 rxnn/rxt/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -14,10 +14,10 @@ rxnn/training/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 rxnn/training/base.py,sha256=CqaArEZYOdH64nmKfx28U3GI46TzO4oNkjf_hrF23Cw,11835
 rxnn/training/bml.py,sha256=hw6gLpLkGvqLzxIvBg4MvCc5r8cHpEm2RDyh7nH6CtE,16914
 rxnn/training/callbacks.py,sha256=rS8leuVFPVVfE5Zc8DMkUZhRIPN-vpPbUjowXE5TSBw,36779
-rxnn/training/dataset.py,sha256=tbtOSYldHnQB6SWgee_yUj9zTbgoEoLFNa6wvUS6Apg,51292
+rxnn/training/dataset.py,sha256=ruU6k33pQmpTqhxpjLFNdDJnCjcrBcGeFOzJqFahJDM,51880
 rxnn/training/ddp.py,sha256=VsNBjn3cY-uUj8hbsW7oKvb0_ZKnXnJ2KgObm-Mr9i4,836
-rxnn/training/models.py,sha256=KIiOCW0VgKtMA4EMQ---xsVExdI1mBsgWjtRSmJpecA,9033
-rxnn/training/mrl.py,sha256=c_7P_DhroK3pQLubfmlVryWBSwlZ0BssU8zZ6UhjOaI,65919
+rxnn/training/models.py,sha256=ILkcqBV1MImnULnq-YDSSEf8cUdEbUgQaH0FRTsa4LA,9069
+rxnn/training/mrl.py,sha256=Ntkti6DDKipKa-AwTvo1WDOdIXOL3uXOhT-Xx29wR-w,67369
 rxnn/training/reward.py,sha256=uiSsBXmjMw2yv-1Bssy3RTlpU6zP8ape3490Sl-aT0M,16144
 rxnn/training/rl.py,sha256=hWtExxY-_pAmTOGYxyCNounUbaGWvLDVltC4sRC7MN4,7175
 rxnn/training/scheduler.py,sha256=LcjU35mEwz2U5x3U6tLfeeYlBqMxbFSxYzJYuXkWbSY,1408
@@ -26,14 +26,14 @@ rxnn/training/utils.py,sha256=ngDCm654NL3UsPy190Er4XPc9HI-OyEV6tDLMgEEvQc,6219
 rxnn/transformers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 rxnn/transformers/attention.py,sha256=KRnKT6XUqAXElxV9y72mSpdTeiMgCKCCLqqxCFNTHmA,16372
 rxnn/transformers/ff.py,sha256=WDjO-H9XWInoWnUnxiseIH6Kx5GlHP0zGJygwhcb1gc,2589
-rxnn/transformers/layers.py,sha256=OlbqD5kKygn5WZziLbU3jZjhr8hBrxLpqlCjJ_BNCW0,8119
+rxnn/transformers/layers.py,sha256=bcDP8vZ5dpTWWqMCkzrPG8yQA0D0G5VjnV2Nq9IO8Dc,8816
 rxnn/transformers/mask.py,sha256=J0cfLVLt3SzS2ra3KcY4khrkhI975Dw4CjpUi3Sn25s,419
-rxnn/transformers/models.py,sha256=7xKixlBN5uZckcOXfukZeA4f7R34A35Gk98lNzew37o,11559
+rxnn/transformers/models.py,sha256=r4vNldYqCIpwMpXkFZvYbw0UBK3NE75qH7bc6OZ8YjE,11587
 rxnn/transformers/moe.py,sha256=j6jEx6Ip0zttlUZKKn82azxo95lkLZs-H2GLSMD88hY,5859
 rxnn/transformers/positional.py,sha256=1PjcJybUzeQlIKJI4tahAGZcYgCRCL0otxs7mpsNuzM,4410
 rxnn/transformers/sampler.py,sha256=t6iiQTdLQ0TakUWnnhKkb5DKF2F_9-thXHBydDF3fxg,17389
 rxnn/utils.py,sha256=ihb6OTyDtPiocB_lOvnq7eOkjjpCkgs8wxvXUBNQ7mM,996
-rxnn-0.2.69.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
-rxnn-0.2.69.dist-info/METADATA,sha256=YcmghdF8ypeyOCmglJaws18cDtTqSIE8P-gReGIMzsU,60420
-rxnn-0.2.69.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
-rxnn-0.2.69.dist-info/RECORD,,
+rxnn-0.2.71.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
+rxnn-0.2.71.dist-info/METADATA,sha256=7BHHcFtImjPB57X2eRLgO4IFOSBNb7GOR5ytMaCttkI,60420
+rxnn-0.2.71.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
+rxnn-0.2.71.dist-info/RECORD,,