Fix a shape annotation and typos in `mamba` slow forward (#30691)

* fix typos and one shape comment

* fix `intermediade` typo in jamba

src/transformers/models/jamba/modeling_jamba.py

@@ -962,15 +962,15 @@ class JambaMambaMixer(nn.Module):
         # 3.b. Discretization: B and C to [batch, seq_len, intermediate_size, ssm_state_size] (SRAM)
         A = -torch.exp(self.A_log.float())                                              # [intermediate_size, ssm_state_size]
         discrete_A = torch.exp(A[None, :, None, :] * discrete_time_step[:, :, :, None]) # [batch, intermediate_size, seq_len, ssm_state_size]
-        discrete_B = discrete_time_step[:, :, :, None] * B[:, None, :, :].float()       # [batch, intermediade_size, seq_len, ssm_state_size]
+        discrete_B = discrete_time_step[:, :, :, None] * B[:, None, :, :].float()       # [batch, intermediate_size, seq_len, ssm_state_size]
         deltaB_u = discrete_B * hidden_states[:, :, :, None].float()
         # 3.c perform the recurrence y ← SSM(A, B, C)(x)
         scan_outputs = []
         for i in range(seq_len):
-            ssm_state = discrete_A[:, :, i, :] * ssm_state + deltaB_u[:, :, i, :]      # [batch, intermediade_size, ssm_state]
-            scan_output = torch.matmul(ssm_state.to(dtype), C[:, i, :].unsqueeze(-1))  # [batch, intermediade_size, 1]
+            ssm_state = discrete_A[:, :, i, :] * ssm_state + deltaB_u[:, :, i, :]      # [batch, intermediate_size, ssm_state]
+            scan_output = torch.matmul(ssm_state.to(dtype), C[:, i, :].unsqueeze(-1))  # [batch, intermediate_size, 1]
             scan_outputs.append(scan_output[:, :, 0])
-        scan_output = torch.stack(scan_outputs, dim=-1)                                # [batch, intermediade_size, seq_len]
+        scan_output = torch.stack(scan_outputs, dim=-1)                                # [batch, intermediate_size, seq_len]
         scan_output = scan_output + (hidden_states * self.D[None, :, None])
         scan_output = (scan_output * self.act(gate))
 

src/transformers/models/mamba/modeling_mamba.py

@@ -279,16 +279,16 @@ class MambaMixer(nn.Module):
         # 3.b. Discretization: B and C to [batch, seq_len, intermediate_size, ssm_state_size] (SRAM)
         A = -torch.exp(self.A_log.float())                                              # [intermediate_size, ssm_state_size]
         discrete_A = torch.exp(A[None, :, None, :] * discrete_time_step[:, :, :, None]) # [batch, intermediate_size, seq_len, ssm_state_size]
-        discrete_B = discrete_time_step[:, :, :, None] * B[:, None, :, :].float()       # [batch, intermediade_size, seq_len, ssm_state_size]
+        discrete_B = discrete_time_step[:, :, :, None] * B[:, None, :, :].float()       # [batch, intermediate_size, seq_len, ssm_state_size]
         deltaB_u = discrete_B * hidden_states[:, :, :, None].float()
 
         # 3.c perform the recurrence y ← SSM(A, B, C)(x)
         scan_outputs = []
         for i in range(seq_len):
-            ssm_state = discrete_A[:, :, i, :] * ssm_state + deltaB_u[:, :, i, :]      # [batch, intermediade_size, ssm_state]
-            scan_output = torch.matmul(ssm_state.to(dtype), C[:, i, :].unsqueeze(-1))  # [batch, intermediade_size, 1]
+            ssm_state = discrete_A[:, :, i, :] * ssm_state + deltaB_u[:, :, i, :]      # [batch, intermediate_size, ssm_state]
+            scan_output = torch.matmul(ssm_state.to(dtype), C[:, i, :].unsqueeze(-1))  # [batch, intermediate_size, 1]
             scan_outputs.append(scan_output[:, :, 0])
-        scan_output = torch.stack(scan_outputs, dim=-1)                                # [batch, seq_len, intermediade_size]
+        scan_output = torch.stack(scan_outputs, dim=-1)                                # [batch, intermediate_size, seq_len]
         scan_output = scan_output + (hidden_states * self.D[None, :, None])
         scan_output = (scan_output * self.act(gate))