fix spacing and fast rms not for training

onmt/inputters/text_utils.py

@@ -14,7 +14,7 @@ def parse_features(line, n_feats=0, defaults=None):
     text, feats = [], [[] for _ in range(n_feats)]
     check, count = 0, 0
     for token in line.split(" "):
-        tok, *fts = token.strip().split("￨")
+        tok, *fts = token.strip("\n").split("￨")
         check += len(fts)
         count += 1
         if not fts and defaults is not None:

onmt/modules/rmsnorm.py

@@ -24,16 +24,17 @@ def __init__(self, hidden_size: int, eps: float = 1e-6):
         self.weight = nn.Parameter(torch.ones(hidden_size))
 
     def forward(self, hidden_states):
-        if AWQ_INFERENCE_ENGINE:
-            output = torch.empty_like(hidden_states)
+        if AWQ_INFERENCE_ENGINE and not self.training:
+            inp_type = hidden_states.dtype
+            output = torch.empty_like(hidden_states).to(inp_type)
             if hidden_states.dim() == 2:  # patch for multi experts
                 hidden_states = hidden_states.unsqueeze(0)
             awq_inference_engine.layernorm_forward_cuda(
-                hidden_states, self.weight, output, self.eps
+                hidden_states.half(), self.weight.half(), output.half(), self.eps
             )
             if hidden_states.dim() == 2:  # patch for multi experts
                 output = output.unsqueeze(0)
-            return output
+            return output.to(inp_type)
         else:
             hidden_states = hidden_states.to(torch.float32)
             variance = hidden_states.pow(2).mean(-1, keepdim=True)