fix

Author: hjh0119

cookbook/legacy/grpo/lora_backup.py

@@ -44,7 +44,6 @@
 
 # SwanLab is optional - only used if SWANLAB_API_KEY is set
 USE_SWANLAB = True
-os.environ['SWANLAB_API_KEY'] = '3hVJrk0veNB2NCm72UdJg'
 if USE_SWANLAB:
     import swanlab
     if USE_SWANLAB:

src/twinkle/model/transformers/transformers.py

@@ -421,7 +421,7 @@ def calculate_loss(self, **kwargs):
             loss_value, counts = result
         else:
             loss_value = result
-            counts = torch.tensor(1, device=loss_value.device)
+            counts = torch.tensor(0, device=loss_value.device)
         optimizer_config = self.optimizer_group[adapter_name]
         optimizer_config.num_tokens += counts.item()
         if self.sp_strategy is not None and 'labels' in inputs:

src/twinkle/sampler/vllm_sampler/vllm_engine.py

@@ -94,6 +94,7 @@ def __init__(
 
         # Tokenizer is lazy loaded via get_tokenizer()
         self._tokenizer = None
+        breakpoint()
 
     def _create_engine(self):
         """Create and return the vLLM engine."""
@@ -258,6 +259,17 @@ async def sample(
             lora_request = await self._get_or_load_lora(adapter_path, adapter_user_id)
 
         # Generate
+        if lora_request is None:
+            from vllm.lora.request import LoRARequest
+            from twinkle.sampler.vllm_sampler.vllm_worker_extension import (
+                VLLM_LORA_INT_ID, VLLM_LORA_NAME, VLLM_LORA_PATH,
+            )
+            lora_request = LoRARequest(
+                lora_name=VLLM_LORA_NAME,
+                lora_int_id=VLLM_LORA_INT_ID,
+                lora_path=VLLM_LORA_PATH,
+            )
+            
         generator = self.engine.generate(
             prompt=prompt,
             sampling_params=vllm_params,