update

ModelTC · Dec 13, 2024 · 2654cc7 · 2654cc7
1 parent e180c1f
commit 2654cc7
Show file tree

Hide file tree

Showing 4 changed files with 4 additions and 24 deletions.
diff --git a/llmc/eval/eval_base.py b/llmc/eval/eval_base.py
@@ -179,7 +179,7 @@ def register_hooks(self, model):
         return handles
 
     @torch.no_grad()
-    def eval(self, model_llmc, model_org=None, eval_pos=None):
+    def eval(self, model_llmc, eval_pos=None):
         handles, handles_org = [], []
         if self.inference_per_block:
             handles = self.register_hooks(model_llmc)
@@ -194,22 +194,7 @@ def eval(self, model_llmc, model_org=None, eval_pos=None):
         else:
             model_llmc.model.eval()
 
-        if model_org is not None:
-            if self.inference_per_block:
-                handles_org = self.register_hooks(model_org)
-            else:
-                if model_org.mm_model:
-                    model_org.mm_model.cuda()
-                else:
-                    model_org.model.cuda()
-
-            if model_org.mm_model:
-                model_org.mm_model.eval()
-            else:
-                model_org.model.eval()
-
         eval_res = self.eval_func(
-            model_org,
             model_llmc,
             self.testenc,
             self.seq_len,
@@ -224,11 +209,6 @@ def eval(self, model_llmc, model_org=None, eval_pos=None):
             model_llmc.mm_model.cpu()
         else:
             model_llmc.model.cpu()
-        if model_org is not None:
-            if model_org.mm_model:
-                model_org.mm_model.cpu()
-            else:
-                model_org.model.cpu()
 
         gc.collect()
         torch.cuda.empty_cache()

diff --git a/llmc/eval/eval_code.py b/llmc/eval/eval_code.py
@@ -22,7 +22,7 @@ def __init__(self, model, config):
         self.add_chat_temp = self.eval_cfg.get('add_chat_temp', False)
 
     @torch.no_grad()
-    def eval_func(self, org_model, model, testenc, seq_len, bs, eval_pos):
+    def eval_func(self, model, testenc, seq_len, bs, eval_pos):
         samples = []
         pbar = tqdm(total=len(testenc) * bs, dynamic_ncols=True, position=0, desc='Evaluating')
 

diff --git a/llmc/eval/eval_custom_generate.py b/llmc/eval/eval_custom_generate.py
@@ -16,7 +16,7 @@ def __init__(self, model, config):
         self.max_new_tokens = self.eval_cfg.get('max_new_tokens', 32)
 
     @torch.no_grad()
-    def eval_func(self, org_model, model, testenc, seq_len, bs, eval_pos):
+    def eval_func(self, model, testenc, seq_len, bs, eval_pos):
         responses = []
         for data in testenc:
             data = {

diff --git a/llmc/eval/eval_ppl.py b/llmc/eval/eval_ppl.py
@@ -12,7 +12,7 @@
 class PerplexityEval(BaseEval):
 
     @torch.no_grad()
-    def eval_func(self, org_model, model, testenc, seq_len, bs, eval_pos):
+    def eval_func(self, model, testenc, seq_len, bs, eval_pos):
         testenc = testenc.input_ids
         nsamples = testenc.numel() // seq_len