fix: synced_gpus must be passed to generate(), not GenerationConfig

transformers 4.40+ raises ValueError if synced_gpus is included in
GenerationConfig(**gen_kwargs). Pop it out before constructing the
config and pass it directly to every model.generate() call instead.

Affects all cl_trainer_*.py files in both root and improve codebases:
- cl_trainer_gainlora_inflora.py
- cl_trainer_inflora.py
- cl_trainer_olora.py
- cl_trainer_gainlora_olora_llama.py
- cl_trainer_gainlora_inflora_llama.py
- cl_trainer_inflora_llama.py
- cl_trainer_gainlora_olora.py

Does not affect T5-large scripts (same fix, backward compatible).

fix_synced_gpus.py

@@ -0,0 +1,50 @@
+import re, pathlib
+
+FILES = [
+    "root_gainlora/src/cl_trainer_inflora.py",
+    "root_gainlora/src/cl_trainer_olora.py",
+    "root_gainlora/src/cl_trainer_gainlora_olora_llama.py",
+    "root_gainlora/src/cl_trainer_gainlora_inflora_llama.py",
+    "improve_gainlora/src/cl_trainer_gainlora_inflora.py",
+    "improve_gainlora/src/cl_trainer_inflora.py",
+    "improve_gainlora/src/cl_trainer_inflora_llama.py",
+    "improve_gainlora/src/cl_trainer_gainlora_olora.py",
+    "improve_gainlora/src/cl_trainer_gainlora_olora_llama.py",
+    "improve_gainlora/src/cl_trainer_gainlora_inflora_llama.py",
+]
+
+for rel in FILES:
+    p = pathlib.Path(rel)
+    txt = p.read_text()
+
+    # Skip if already fixed
+    if 'gen_kwargs.pop("synced_gpus"' in txt:
+        pop_idx = txt.index('gen_kwargs.pop("synced_gpus"')
+        gc_idx  = txt.index('generation_config = GenerationConfig(**gen_kwargs)')
+        if pop_idx < gc_idx:
+            print(f"  ALREADY FIXED: {rel}")
+            continue
+
+    old_gc = '        generation_config = GenerationConfig(**gen_kwargs)'
+    new_gc = '        synced_gpus = gen_kwargs.pop("synced_gpus", False)\n        generation_config = GenerationConfig(**gen_kwargs)'
+    if old_gc not in txt:
+        print(f"  WARNING: GenerationConfig pattern not found in {rel}")
+        continue
+    txt = txt.replace(old_gc, new_gc, 1)
+
+    # Add synced_gpus= to each generate() call after generation_config= line
+    txt = re.sub(
+        r'(generation_config=generation_config,)\n(\s+\))',
+        r'\1\n\2    synced_gpus=synced_gpus,\n\2',
+        txt,
+    )
+    txt = re.sub(
+        r'(generation_config=generation_config)\n(\s+\))',
+        r'\1,\n\2    synced_gpus=synced_gpus,\n\2',
+        txt,
+    )
+
+    p.write_text(txt)
+    print(f"  FIXED: {rel}")
+
+print("Done")

fix_synced_gpus2.py

@@ -0,0 +1,60 @@
+"""Fix the broken synced_gpus insertion: convert
+    generation_config=generation_config,
+            )    synced_gpus=synced_gpus,
+            )
+to the correct form:
+    generation_config=generation_config,
+                synced_gpus=synced_gpus,
+            )
+
+This must be idempotent.
+"""
+import re, pathlib
+
+FILES = [
+    "root_gainlora/src/cl_trainer_inflora.py",
+    "root_gainlora/src/cl_trainer_olora.py",
+    "root_gainlora/src/cl_trainer_gainlora_olora_llama.py",
+    "root_gainlora/src/cl_trainer_gainlora_inflora_llama.py",
+    "root_gainlora/src/cl_trainer_gainlora_inflora.py",   # already done manually but check
+    "improve_gainlora/src/cl_trainer_gainlora_inflora.py",
+    "improve_gainlora/src/cl_trainer_inflora.py",
+    "improve_gainlora/src/cl_trainer_inflora_llama.py",
+    "improve_gainlora/src/cl_trainer_gainlora_olora.py",
+    "improve_gainlora/src/cl_trainer_gainlora_olora_llama.py",
+    "improve_gainlora/src/cl_trainer_gainlora_inflora_llama.py",
+]
+
+# Pattern produced by the broken regex:
+#   generation_config=generation_config,\n{ws})    synced_gpus=synced_gpus,\n{ws})
+BAD_PAT = re.compile(
+    r'(generation_config=generation_config,)\n'
+    r'( +)\)    synced_gpus=synced_gpus,\n'
+    r'( +\))'
+)
+
+# Correct form: synced_gpus goes INSIDE the call, before the closing paren
+# We add 4 extra spaces relative to the closing-paren indent level.
+def fix_bad(txt):
+    def replacer(m):
+        arg_line = m.group(1)            # generation_config=generation_config,
+        inner_ws = m.group(2)            # whitespace before the doubled )
+        close_paren = m.group(3)         # whitespace + )
+        # synced_gpus should be indented 4 more than the closing paren indent
+        synced_ws = inner_ws + "    "
+        return f"{arg_line}\n{synced_ws}synced_gpus=synced_gpus,\n{close_paren}"
+    return BAD_PAT.sub(replacer, txt)
+
+for rel in FILES:
+    p = pathlib.Path(rel)
+    txt = p.read_text()
+
+    if ")    synced_gpus=synced_gpus," in txt:
+        fixed = fix_bad(txt)
+        p.write_text(fixed)
+        n = txt.count(")    synced_gpus=synced_gpus,")
+        print(f"  FIXED ({n} occurrences): {rel}")
+    else:
+        print(f"  OK (no bad pattern): {rel}")
+
+print("Done")

improve_gainlora/src/cl_trainer_gainlora_inflora.py

@@ -909,6 +909,7 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -920,6 +921,7 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -929,12 +931,14 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

improve_gainlora/src/cl_trainer_gainlora_inflora_llama.py

@@ -940,6 +940,7 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -951,6 +952,7 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -960,12 +962,14 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

improve_gainlora/src/cl_trainer_gainlora_olora.py

@@ -948,6 +948,7 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -959,6 +960,7 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -968,12 +970,14 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

improve_gainlora/src/cl_trainer_gainlora_olora_llama.py

@@ -931,6 +931,7 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -942,6 +943,7 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -951,12 +953,14 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

improve_gainlora/src/cl_trainer_inflora.py

@@ -704,6 +704,7 @@ class InfLoRATrainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -715,6 +716,7 @@ class InfLoRATrainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -724,12 +726,14 @@ class InfLoRATrainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

improve_gainlora/src/cl_trainer_inflora_llama.py

@@ -731,6 +731,7 @@ class InfLoRATrainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -742,6 +743,7 @@ class InfLoRATrainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -751,12 +753,14 @@ class InfLoRATrainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

root_gainlora/src/cl_trainer_gainlora_inflora.py

@@ -933,7 +933,7 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
                     "pad_token_id": 0,
                 }
                 
-            gen_kwargs["synced_gpus"] = False
+            synced_gpus = gen_kwargs.pop("synced_gpus", False)
 
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
@@ -949,6 +949,7 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -958,12 +959,14 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

root_gainlora/src/cl_trainer_gainlora_inflora_llama.py

@@ -941,6 +941,7 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -952,6 +953,7 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -961,12 +963,14 @@ class GainLoRA_InfLoRA_Trainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

root_gainlora/src/cl_trainer_gainlora_olora_llama.py

@@ -932,6 +932,7 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -943,6 +944,7 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -952,12 +954,14 @@ class GainLoRA_OLoRA_Trainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

root_gainlora/src/cl_trainer_inflora.py

@@ -705,6 +705,7 @@ class InfLoRATrainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -716,6 +717,7 @@ class InfLoRATrainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -725,12 +727,14 @@ class InfLoRATrainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape

root_gainlora/src/cl_trainer_olora.py

@@ -547,6 +547,7 @@ class OLoRATrainer(Seq2SeqTrainer):
         if "attention_mask" in inputs:
             gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
 
+        synced_gpus = gen_kwargs.pop("synced_gpus", False)
         generation_config = GenerationConfig(**gen_kwargs)
 
         # prepare generation inputs
@@ -558,6 +559,7 @@ class OLoRATrainer(Seq2SeqTrainer):
             generated_tokens = self.model.generate(
                 input_ids=generation_inputs, 
                 generation_config=generation_config,
+                synced_gpus=synced_gpus,
             )
         else:
             generation_inputs = inputs[self.model.main_input_name]
@@ -567,12 +569,14 @@ class OLoRATrainer(Seq2SeqTrainer):
                     input_ids=generation_inputs,
                     input_ids_wo_label=inputs["input_ids_wo_label"],
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
             
             else:
                 generated_tokens = self.model.generate(
                     input_ids=generation_inputs,
                     generation_config=generation_config,
+                    synced_gpus=synced_gpus,
                 )
 
         bs, source_len = inputs['input_ids'].shape