upload

config/arch/hrm.yaml

@@ -0,0 +1,24 @@
+name: recursive_reasoning.hrm@HierarchicalReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 2
+L_cycles: 2
+
+H_layers: 4
+L_layers: 4
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
+forward_dtype: bfloat16
+
+mlp_t: False # use mlp on L instead of transformer

config/arch/transformers_baseline.yaml

@@ -0,0 +1,18 @@
+name: recursive_reasoning.transformers_baseline@Model_ACTV2
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 1  # kept for compatibility
+H_layers: 8 
+
+hidden_size: 512  
+num_heads: 12
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope

config/arch/trm.yaml

@@ -0,0 +1,26 @@
+name: recursive_reasoning.trm@TinyRecursiveReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 3
+L_cycles: 6
+
+H_layers: 0
+L_layers: 2
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
+forward_dtype: bfloat16
+
+mlp_t: False # use mlp on L instead of transformer
+puzzle_emb_len: 16 # if non-zero, its specified to this value
+no_ACT_continue: True # No continue ACT loss, only use the sigmoid of the halt which makes much more sense

config/arch/trm_hier6.yaml

@@ -0,0 +1,26 @@
+name: recursive_reasoning.trm_hier6@TinyRecursiveReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 3
+L_cycles: 6
+
+H_layers: 0
+L_layers: 2
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
+forward_dtype: bfloat16
+
+mlp_t: False # use mlp on L instead of transformer
+puzzle_emb_len: 16 # if non-zero, its specified to this value
+no_ACT_continue: True # No continue ACT loss, only use the sigmoid of the halt which makes much more sense

config/arch/trm_singlez.yaml

@@ -0,0 +1,26 @@
+name: recursive_reasoning.trm_singlez@TinyRecursiveReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 3
+L_cycles: 6
+
+H_layers: 0
+L_layers: 2
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
+forward_dtype: bfloat16
+
+mlp_t: False # use mlp on L instead of transformer
+puzzle_emb_len: 16 # if non-zero, its specified to this value
+no_ACT_continue: True # No continue ACT loss, only use the sigmoid of the halt which makes much more sense